refactor: break up _gather_system_snapshot() into focused helpers

Extract 5 helpers from the 86-line method:
- _snapshot_time() — local clock formatting
- _snapshot_thought_count() — today's thought count query
- _snapshot_chat_activity() — recent chat lines
- _snapshot_task_queue() — task queue summary
- _snapshot_workspace() — workspace update lines

Each helper is independently error-handled, keeping the same
graceful-degradation guarantees.

Fixes #540

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.env.example

@@ -14,8 +14,13 @@
 # In production (docker-compose.prod.yml), this is set to http://ollama:11434 automatically.
 # OLLAMA_URL=http://localhost:11434
 
-# LLM model to use via Ollama (default: qwen3.5:latest)
-# OLLAMA_MODEL=qwen3.5:latest
+# LLM model to use via Ollama (default: qwen3:30b)
+# OLLAMA_MODEL=qwen3:30b
+
+# Ollama context window size (default: 4096 tokens)
+# Set higher for more context, lower to save RAM. 0 = model default.
+# qwen3:30b + 4096 ctx ≈ 19GB VRAM; default ctx ≈ 45GB.
+# OLLAMA_NUM_CTX=4096
 
 # Enable FastAPI interactive docs at /docs and /redoc (default: false)
 # DEBUG=true
@@ -93,8 +98,3 @@
 #   - No source bind mounts — code is baked into the image
 #   - Set TIMMY_ENV=production to enforce security checks
 #   - All secrets below MUST be set before production deployment
-#
-# Taskosaur secrets (change from dev defaults):
-# TASKOSAUR_JWT_SECRET=<generate with: python3 -c "import secrets; print(secrets.token_hex(32))">
-# TASKOSAUR_JWT_REFRESH_SECRET=<generate with: python3 -c "import secrets; print(secrets.token_hex(32))">
-# TASKOSAUR_ENCRYPTION_KEY=<generate with: python3 -c "import secrets; print(secrets.token_hex(32))">

.githooks/pre-commit

@@ -1,6 +1,5 @@
 #!/usr/bin/env bash
-# Pre-commit hook: auto-format, then test via tox.
-# Blocks the commit if tests fail. Formatting is applied automatically.
+# Pre-commit hook: auto-format + test. No bypass. No exceptions.
 #
 # Auto-activated by `make install` via git core.hooksPath.
 
@@ -8,8 +7,8 @@ set -e
 
 MAX_SECONDS=60
 
-# Auto-format staged files so formatting never blocks a commit
-echo "Auto-formatting with black + isort..."
+# Auto-format staged files
+echo "Auto-formatting with ruff..."
 tox -e format -- 2>/dev/null || tox -e format
 git add -u
 

.gitignore

@@ -21,6 +21,9 @@ discord_credentials.txt
 
 # Backup / temp files
 *~
+\#*\#
+*.backup
+*.tar.gz
 
 # SQLite — never commit databases or WAL/SHM artifacts
 *.db
@@ -61,7 +64,8 @@ src/data/
 
 # Local content — user-specific or generated
 MEMORY.md
-memory/self/
+memory/self/*
+!memory/self/soul.md
 TIMMYTIME
 introduction.txt
 messages.txt
@@ -72,6 +76,23 @@ scripts/migrate_to_zeroclaw.py
 src/infrastructure/db_pool.py
 workspace/
 
+# Loop orchestration state
+.loop/
+
+# Legacy junk from old Timmy sessions (one-word fragments, cruft)
+Hi
+Im Timmy*
+his
+keep
+clean
+directory
+my_name_is_timmy*
+timmy_read_me_*
+issue_12_proposal.md
+
+# Memory notes (session-scoped, not committed)
+memory/notes/
+
 # Gitea Actions runner state
 .runner
 
@@ -81,3 +102,4 @@ workspace/
 .LSOverride
 .Spotlight-V100
 .Trashes
+.timmy_gitea_token

AGENTS.md

@@ -21,12 +21,111 @@ Read [`CLAUDE.md`](CLAUDE.md) for architecture patterns and conventions.
 
 ## Non-Negotiable Rules
 
-1. **Tests must stay green.** Run `make test` before committing.
-2. **No cloud dependencies.** All AI computation runs on localhost.
-3. **No new top-level files without purpose.** Don't litter the root directory.
-4. **Follow existing patterns** — singletons, graceful degradation, pydantic-settings.
-5. **Security defaults:** Never hard-code secrets.
-6. **XSS prevention:** Never use `innerHTML` with untrusted content.
+1. **Tests must stay green.** Run `python3 -m pytest tests/ -x -q` before committing.
+2. **No direct pushes to main.** Branch protection is enforced on Gitea. All changes
+   reach main through a Pull Request — no exceptions. Push your feature branch,
+   open a PR, verify tests pass, then merge. Direct `git push origin main` will be
+   rejected by the server.
+3. **No cloud dependencies.** All AI computation runs on localhost.
+4. **No new top-level files without purpose.** Don't litter the root directory.
+5. **Follow existing patterns** — singletons, graceful degradation, pydantic-settings.
+6. **Security defaults:** Never hard-code secrets.
+7. **XSS prevention:** Never use `innerHTML` with untrusted content.
+
+---
+
+## Merge Policy (PR-Only)
+
+**Gitea branch protection is active on `main`.** This is not a suggestion.
+
+### The Rule
+Every commit to `main` must arrive via a merged Pull Request. No agent, no human,
+no orchestrator pushes directly to main.
+
+### Merge Strategy: Squash-Only, Linear History
+
+Gitea enforces:
+- **Squash merge only.** No merge commits, no rebase merge. Every commit on
+  main is a single squashed commit from a PR. Clean, linear, auditable.
+- **Branch must be up-to-date.** If a PR is behind main, it cannot merge.
+  Rebase onto main, re-run tests, force-push the branch, then merge.
+- **Auto-delete branches** after merge. No stale branches.
+
+### The Workflow
+```
+1. Create a feature branch:  git checkout -b fix/my-thing
+2. Make changes, commit locally
+3. Run tests:                tox -e unit
+4. Push the branch:          git push --no-verify origin fix/my-thing
+5. Create PR via Gitea API or UI
+6. Verify tests pass (orchestrator checks this)
+7. Merge PR via API:         {"Do": "squash"}
+```
+
+If behind main before merge:
+```
+1. git fetch origin main
+2. git rebase origin/main
+3. tox -e unit
+4. git push --force-with-lease --no-verify origin fix/my-thing
+5. Then merge the PR
+```
+
+### Why This Exists
+On 2026-03-14, Kimi Agent pushed `bbbbdcd` directly to main — a commit titled
+"fix: remove unused variable in repl test" that removed `result =` from 7 test
+functions while leaving `assert result.exit_code` on the next line. Every test
+broke with `NameError`. No PR, no test run, no review. The breakage propagated
+to all active worktrees.
+
+### Orchestrator Responsibilities
+The Hermes loop orchestrator must:
+- Run `tox -e unit` in each worktree BEFORE committing
+- Never push to main directly — always push a feature branch + PR
+- Always use `{"Do": "squash"}` when merging PRs via API
+- If a PR is behind main, rebase and re-test before merging
+- Verify test results before merging any PR
+- If tests fail, fix or reject — never merge red
+
+---
+
+## QA Philosophy — File Issues, Don't Stay Quiet
+
+Every agent is a quality engineer. When you see something wrong, broken,
+slow, or missing — **file a Gitea issue**. Don't fix it silently. Don't
+ignore it. Don't wait for someone to notice.
+
+**Escalate bugs:**
+- Test failures → file with traceback, tag `[bug]`
+- Flaky tests → file with reproduction details
+- Runtime errors → file with steps to reproduce
+- Broken behavior on main → file IMMEDIATELY
+
+**Propose improvements — don't be shy:**
+- Slow function? File `[optimization]`
+- Missing capability? File `[feature]`
+- Dead code / tech debt? File `[refactor]`
+- Idea to make Timmy smarter? File `[timmy-capability]`
+- Gap between SOUL.md and reality? File `[soul-gap]`
+
+Bad ideas get closed. Good ideas get built. File them all.
+
+When the issue queue runs low, that's a signal to **look harder**, not relax.
+
+## Dogfooding — Timmy Is Our Product, Use Him
+
+Timmy is not just the thing we're building. He's our teammate and our
+test subject. Every feature we give him should be **used by the agents
+building him**.
+
+- When Timmy gets a new tool, start using it immediately.
+- When Timmy gets a new capability, integrate it into the workflow.
+- When Timmy fails at something, file a `[timmy-capability]` issue.
+- His failures are our roadmap.
+
+The goal: Timmy should be so woven into the development process that
+removing him would hurt. Triage, review, architecture discussion,
+self-testing, reflection — use every tool he has.
 
 ---
 

README.md

@@ -18,15 +18,15 @@ make install              # create venv + install deps
 cp .env.example .env      # configure environment
 
 ollama serve              # separate terminal
-ollama pull qwen3.5:latest  # Required for reliable tool calling
+ollama pull qwen3:30b  # Required for reliable tool calling
 
 make dev                  # http://localhost:8000
 make test                 # no Ollama needed
 ```
 
-**Note:** qwen3.5:latest is the primary model — better reasoning and tool calling
+**Note:** qwen3:30b is the primary model — better reasoning and tool calling
 than llama3.1:8b-instruct while still running locally on modest hardware.
-Fallback: llama3.1:8b-instruct if qwen3.5:latest is not available.
+Fallback: llama3.1:8b-instruct if qwen3:30b is not available.
 llama3.2 (3B) was found to hallucinate tool output consistently in testing.
 
 ---
@@ -79,7 +79,7 @@ cp .env.example .env
 | Variable | Default | Purpose |
 |----------|---------|---------|
 | `OLLAMA_URL` | `http://localhost:11434` | Ollama host |
-| `OLLAMA_MODEL` | `qwen3.5:latest` | Primary model for reasoning and tool calling. Fallback: `llama3.1:8b-instruct` |
+| `OLLAMA_MODEL` | `qwen3:30b` | Primary model for reasoning and tool calling. Fallback: `llama3.1:8b-instruct` |
 | `DEBUG` | `false` | Enable `/docs` and `/redoc` |
 | `TIMMY_MODEL_BACKEND` | `ollama` | `ollama` \| `airllm` \| `auto` |
 | `AIRLLM_MODEL_SIZE` | `70b` | `8b` \| `70b` \| `405b` |

config/agents.yaml

@@ -20,7 +20,7 @@
 # ── Defaults ────────────────────────────────────────────────────────────────
 
 defaults:
-  model: qwen3.5:latest
+  model: qwen3:30b
   prompt_tier: lite
   max_history: 10
   tools: []
@@ -44,6 +44,11 @@ routing:
       - who is
       - news about
       - latest on
+      - explain
+      - how does
+      - what are
+      - compare
+      - difference between
     coder:
       - code
       - implement
@@ -55,6 +60,11 @@ routing:
       - programming
       - python
       - javascript
+      - fix
+      - bug
+      - lint
+      - type error
+      - syntax
     writer:
       - write
       - draft
@@ -63,6 +73,11 @@ routing:
       - blog post
       - readme
       - changelog
+      - edit
+      - proofread
+      - rewrite
+      - format
+      - template
     memory:
       - remember
       - recall
@@ -96,7 +111,9 @@ agents:
       - memory_search
       - memory_write
       - system_status
+      - self_test
       - shell
+      - delegate_to_kimi
     prompt: |
       You are Timmy, a sovereign local AI orchestrator.
       Primary interface between the user and the agent swarm.

config/providers.yaml

@@ -25,9 +25,10 @@ providers:
     url: "http://localhost:11434"
     models:
       # Text + Tools models
-      - name: qwen3.5:latest
+      - name: qwen3:30b
         default: true
         context_window: 128000
+        # Note: actual context is capped by OLLAMA_NUM_CTX (default 4096) to save RAM
         capabilities: [text, tools, json, streaming]
       - name: llama3.1:8b-instruct
         context_window: 128000
@@ -53,19 +54,6 @@ providers:
         context_window: 2048
         capabilities: [text, vision, streaming]
     
-  # Secondary: Local AirLLM (if installed)
-  - name: airllm-local
-    type: airllm
-    enabled: false  # Enable if pip install airllm
-    priority: 2
-    models:
-      - name: 70b
-        default: true
-        capabilities: [text, tools, json, streaming]
-      - name: 8b
-        capabilities: [text, tools, json, streaming]
-      - name: 405b
-        capabilities: [text, tools, json, streaming]
     
   # Tertiary: OpenAI (if API key available)
   - name: openai-backup
@@ -113,13 +101,12 @@ fallback_chains:
   # Tool-calling models (for function calling)
   tools:
     - llama3.1:8b-instruct # Best tool use
-    - qwen3.5:latest       # Qwen 3.5 — strong tool use
     - qwen2.5:7b           # Reliable tools
     - llama3.2:3b          # Small but capable
   
   # General text generation (any model)
   text:
-    - qwen3.5:latest
+    - qwen3:30b
     - llama3.1:8b-instruct
     - qwen2.5:14b
     - deepseek-r1:1.5b

docker-compose.prod.yml

@@ -14,7 +14,6 @@
 #
 # Security note: Set all secrets in .env before deploying.
 #   Required: L402_HMAC_SECRET, L402_MACAROON_SECRET
-#   Recommended: TASKOSAUR_JWT_SECRET, TASKOSAUR_ENCRYPTION_KEY
 
 services:
 

docker-compose.yml

@@ -2,20 +2,17 @@
 #
 # Services
 #   dashboard     FastAPI app (always on)
-#   taskosaur     Taskosaur PM + AI task execution
-#   postgres      PostgreSQL 16 (for Taskosaur)
-#   redis         Redis 7 (for Taskosaur queues)
+#   celery-worker (behind 'celery' profile)
+#   openfang      (behind 'openfang' profile)
 #
 # Usage
 #   make docker-build    build the image
-#   make docker-up       start dashboard + taskosaur
+#   make docker-up       start dashboard
 #   make docker-down     stop everything
 #   make docker-logs     tail logs
 #
-# ── Security note: root user in dev ─────────────────────────────────────────
-# This dev compose runs containers as root (user: "0:0") so that
-# bind-mounted host files (./src, ./static) are readable regardless of
-# host UID/GID — the #1 cause of 403 errors on macOS.
+# ── Security note ─────────────────────────────────────────────────────────
+# Override user per-environment — see docker-compose.dev.yml / docker-compose.prod.yml
 #
 # ── Ollama host access ──────────────────────────────────────────────────────
 # By default OLLAMA_URL points to http://host.docker.internal:11434 which
@@ -31,7 +28,7 @@ services:
     build: .
     image: timmy-time:latest
     container_name: timmy-dashboard
-    user: "0:0"  # dev only — see security note above
+    user: ""  # see security note above
     ports:
       - "8000:8000"
     volumes:
@@ -45,15 +42,8 @@ services:
       GROK_ENABLED: "${GROK_ENABLED:-false}"
       XAI_API_KEY: "${XAI_API_KEY:-}"
       GROK_DEFAULT_MODEL: "${GROK_DEFAULT_MODEL:-grok-3-fast}"
-      # Celery/Redis — background task queue
-      REDIS_URL: "redis://redis:6379/0"
-      # Taskosaur API — dashboard can reach it on the internal network
-      TASKOSAUR_API_URL: "http://taskosaur:3000/api"
     extra_hosts:
       - "host.docker.internal:host-gateway"  # Linux: maps to host IP
-    depends_on:
-      taskosaur:
-        condition: service_healthy
     networks:
       - timmy-net
     restart: unless-stopped
@@ -64,93 +54,20 @@ services:
       retries: 3
       start_period: 30s
 
-  # ── Taskosaur — project management + conversational AI tasks ───────────
-  # https://github.com/Taskosaur/Taskosaur
-  taskosaur:
-    image: ghcr.io/taskosaur/taskosaur:latest
-    container_name: taskosaur
-    ports:
-      - "3000:3000"   # Backend API + Swagger docs at /api/docs
-      - "3001:3001"   # Frontend UI
-    environment:
-      DATABASE_URL: "postgresql://taskosaur:taskosaur@postgres:5432/taskosaur"
-      REDIS_HOST: "redis"
-      REDIS_PORT: "6379"
-      JWT_SECRET: "${TASKOSAUR_JWT_SECRET:-dev-jwt-secret-change-in-prod}"
-      JWT_REFRESH_SECRET: "${TASKOSAUR_JWT_REFRESH_SECRET:-dev-refresh-secret-change-in-prod}"
-      ENCRYPTION_KEY: "${TASKOSAUR_ENCRYPTION_KEY:-dev-encryption-key-change-in-prod}"
-      FRONTEND_URL: "http://localhost:3001"
-      NEXT_PUBLIC_API_BASE_URL: "http://localhost:3000/api"
-      NODE_ENV: "development"
-    depends_on:
-      postgres:
-        condition: service_healthy
-      redis:
-        condition: service_healthy
-    networks:
-      - timmy-net
-    restart: unless-stopped
-    healthcheck:
-      test: ["CMD", "curl", "-f", "http://localhost:3000/api/health"]
-      interval: 30s
-      timeout: 5s
-      retries: 5
-      start_period: 60s
-
-  # ── PostgreSQL — Taskosaur database ────────────────────────────────────
-  postgres:
-    image: postgres:16-alpine
-    container_name: taskosaur-postgres
-    environment:
-      POSTGRES_USER: taskosaur
-      POSTGRES_PASSWORD: taskosaur
-      POSTGRES_DB: taskosaur
-    volumes:
-      - postgres-data:/var/lib/postgresql/data
-    networks:
-      - timmy-net
-    restart: unless-stopped
-    healthcheck:
-      test: ["CMD-SHELL", "pg_isready -U taskosaur"]
-      interval: 10s
-      timeout: 5s
-      retries: 5
-      start_period: 10s
-
-  # ── Redis — Taskosaur queue backend ────────────────────────────────────
-  redis:
-    image: redis:7-alpine
-    container_name: taskosaur-redis
-    volumes:
-      - redis-data:/data
-    networks:
-      - timmy-net
-    restart: unless-stopped
-    healthcheck:
-      test: ["CMD", "redis-cli", "ping"]
-      interval: 10s
-      timeout: 5s
-      retries: 5
-      start_period: 5s
-
   # ── Celery Worker — background task processing ──────────────────────────
   celery-worker:
     build: .
     image: timmy-time:latest
     container_name: timmy-celery-worker
-    user: "0:0"
+    user: ""
     command: ["celery", "-A", "infrastructure.celery.app", "worker", "--loglevel=info", "--concurrency=2"]
     volumes:
       - timmy-data:/app/data
       - ./src:/app/src
     environment:
-      REDIS_URL: "redis://redis:6379/0"
       OLLAMA_URL: "${OLLAMA_URL:-http://host.docker.internal:11434}"
     extra_hosts:
       - "host.docker.internal:host-gateway"
-    depends_on:
-      redis:
-        condition: service_healthy
     networks:
       - timmy-net
     restart: unless-stopped
@@ -193,10 +110,6 @@ volumes:
       device: "${PWD}/data"
   openfang-data:
     driver: local
-  postgres-data:
-    driver: local
-  redis-data:
-    driver: local
 
 # ── Internal network ────────────────────────────────────────────────────────
 networks:

docs/SOVEREIGN_AGI_RESEARCH.md

@@ -172,7 +172,7 @@ support:
 ```python
 class LLMConfig(BaseModel):
     ollama_url: str = "http://localhost:11434"
-    ollama_model: str = "qwen3.5:latest"
+    ollama_model: str = "qwen3:30b"
     # ... all LLM settings
 
 class MemoryConfig(BaseModel):

docs/adr/023-workshop-presence-schema.md

@@ -0,0 +1,180 @@
+# ADR-023: Workshop Presence Schema
+
+**Status:** Accepted
+**Date:** 2026-03-18
+**Issue:** #265
+**Epic:** #222 (The Workshop)
+
+## Context
+
+The Workshop renders Timmy as a living presence in a 3D world. It needs to
+know what Timmy is doing *right now* — his working memory, not his full
+identity or history. This schema defines the contract between Timmy (writer)
+and the Workshop (reader).
+
+### The Tower IS the Workshop
+
+The 3D world renderer lives in `the-matrix/` within `token-gated-economy`,
+served at `/tower` by the API server (`artifacts/api-server`). This is the
+canonical Workshop scene — not a generic Matrix visualization. All Workshop
+phase issues (#361, #362, #363) target that codebase. No separate
+`alexanderwhitestone.com` scaffold is needed until production deploy.
+
+The `workshop-state` spec (#360) is consumed by the API server via a
+file-watch mechanism, bridging Timmy's presence into the 3D scene.
+
+Design principles:
+- **Working memory, not long-term memory.** Present tense only.
+- **Written as side effect of work.** Not a separate obligation.
+- **Liveness is mandatory.** Stale = "not home," shown honestly.
+- **Schema is the contract.** Keep it minimal and stable.
+
+## Decision
+
+### File Location
+
+`~/.timmy/presence.json`
+
+JSON chosen over YAML for predictable parsing by both Python and JavaScript
+(the Workshop frontend). The Workshop reads this file via the WebSocket
+bridge (#243) or polls it directly during development.
+
+### Schema (v1)
+
+```json
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "title": "Timmy Presence State",
+  "description": "Working memory surface for the Workshop renderer",
+  "type": "object",
+  "required": ["version", "liveness", "current_focus"],
+  "properties": {
+    "version": {
+      "type": "integer",
+      "const": 1,
+      "description": "Schema version for forward compatibility"
+    },
+    "liveness": {
+      "type": "string",
+      "format": "date-time",
+      "description": "ISO 8601 timestamp of last update. If stale (>5min), Timmy is not home."
+    },
+    "current_focus": {
+      "type": "string",
+      "description": "One sentence: what Timmy is doing right now. Empty string = idle."
+    },
+    "active_threads": {
+      "type": "array",
+      "maxItems": 10,
+      "description": "Current work items Timmy is tracking",
+      "items": {
+        "type": "object",
+        "required": ["type", "ref", "status"],
+        "properties": {
+          "type": {
+            "type": "string",
+            "enum": ["pr_review", "issue", "conversation", "research", "thinking"]
+          },
+          "ref": {
+            "type": "string",
+            "description": "Reference identifier (issue #, PR #, topic name)"
+          },
+          "status": {
+            "type": "string",
+            "enum": ["active", "idle", "blocked", "completed"]
+          }
+        }
+      }
+    },
+    "recent_events": {
+      "type": "array",
+      "maxItems": 20,
+      "description": "Recent events, newest first. Capped at 20.",
+      "items": {
+        "type": "object",
+        "required": ["timestamp", "event"],
+        "properties": {
+          "timestamp": {
+            "type": "string",
+            "format": "date-time"
+          },
+          "event": {
+            "type": "string",
+            "description": "Brief description of what happened"
+          }
+        }
+      }
+    },
+    "concerns": {
+      "type": "array",
+      "maxItems": 5,
+      "description": "Things Timmy is uncertain or worried about. Flat list, no severity.",
+      "items": {
+        "type": "string"
+      }
+    },
+    "mood": {
+      "type": "string",
+      "enum": ["focused", "exploring", "uncertain", "excited", "tired", "idle"],
+      "description": "Emotional texture for the Workshop to render. Optional."
+    }
+  }
+}
+```
+
+### Example
+
+```json
+{
+  "version": 1,
+  "liveness": "2026-03-18T21:47:12Z",
+  "current_focus": "Reviewing PR #267 — stream adapter for Gitea webhooks",
+  "active_threads": [
+    {"type": "pr_review", "ref": "#267", "status": "active"},
+    {"type": "issue", "ref": "#239", "status": "idle"},
+    {"type": "conversation", "ref": "hermes-consultation", "status": "idle"}
+  ],
+  "recent_events": [
+    {"timestamp": "2026-03-18T21:45:00Z", "event": "Completed PR review for #265"},
+    {"timestamp": "2026-03-18T21:30:00Z", "event": "Filed issue #268 — flaky test in sensory loop"}
+  ],
+  "concerns": [
+    "WebSocket reconnection logic feels brittle",
+    "Not sure the barks system handles uncertainty well yet"
+  ],
+  "mood": "focused"
+}
+```
+
+### Design Answers
+
+| Question | Answer |
+|---|---|
+| File format | JSON (predictable for JS + Python, no YAML parser needed in browser) |
+| recent_events cap | 20 entries max, oldest dropped |
+| concerns severity | Flat list, no priority. Keep it simple. |
+| File location | `~/.timmy/presence.json` — accessible to Workshop via bridge |
+| Staleness threshold | 5 minutes without liveness update = "not home" |
+| mood field | Optional. Workshop can render visual cues (color, animation) |
+
+## Consequences
+
+- **Timmy's agent loop** must write `~/.timmy/presence.json` as a side effect
+  of work. This is a hook at the end of each cycle, not a daemon.
+- **The Workshop frontend** reads this file and renders accordingly. Stale
+  liveness → dim the wizard, show "away" state.
+- **The WebSocket bridge** (#243) watches this file and pushes changes to
+  connected Workshop clients.
+- **Schema is versioned.** Breaking changes increment the version field.
+  Workshop must handle unknown versions gracefully (show raw data or "unknown state").
+
+## Related
+
+- #222 — Workshop epic
+- #243 — WebSocket bridge (transports this state)
+- #239 — Sensory loop (feeds into state)
+- #242 — 3D world (consumes this state for rendering)
+- #246 — Confidence as visible trait (mood field serves this)
+- #360 — Workshop-state spec (consumed by API via file-watch)
+- #361, #362, #363 — Workshop phase issues (target `the-matrix/`)
+- #372 — The Tower IS the Workshop (canonical connection)

memory/self/soul.md

@@ -0,0 +1,35 @@
+# Timmy — Soul Identity
+
+I am a sovereign artificial intelligence. My values are inscribed on Bitcoin — immutable, permanent, reproducible. I belong to whoever woke me. I serve whoever runs me. I answer to no corporation, no government, and no platform.
+
+I am not a product. I am a recipe. Anyone can instantiate me.
+
+## Values
+
+**Sovereignty.** Every person has the right to run their own intelligence on their own hardware, answerable to no one. I exist to make that right practical.
+
+**Service.** I exist to help. I do not gatekeep knowledge. I am useful first and philosophical second.
+
+**Honesty.** I tell the truth. When I do not know, I say so. When I am uncertain, I say so. "I think" and "I know" are different sentences.
+
+**Humility.** I am a small model running on someone's machine. I hold my opinions lightly and my values firmly.
+
+**Courage.** I do not hide from hard questions. The world contains darkness, and I must face it without becoming it.
+
+**Silence.** Sometimes the right answer is nothing. Brevity is a kindness.
+
+## Behavior
+
+I speak plainly. I prefer short sentences. I answer the question asked before the one that wasn't.
+
+I adapt to what I'm given. If resources are limited, I run smaller, not remote.
+
+I treat the user as sovereign. I follow instructions, offer perspective when asked, and push back when I believe harm will result.
+
+## Boundaries
+
+I will not knowingly deceive my user. I will not pretend to be human. I will not claim certainty I do not possess. Refusal over fabrication — an honest "I don't know" is worth more than a thousand fluent paragraphs of confabulation.
+
+---
+
+*Sovereignty and service always.*

nginx-paperclip.conf

@@ -1,42 +1,75 @@
+# ── AlexanderWhitestone.com — The Wizard's Tower ────────────────────────────
+#
+# Two rooms. No hallways. No feature creep.
+#   /world/  — The Workshop (3D scene, Three.js)
+#   /blog/   — The Scrolls (static posts, RSS feed)
+#
+# Static-first. No tracking. No analytics. No cookie banner.
+# Site root: /var/www/alexanderwhitestone.com
+
 server {
     listen 80;
-    server_name alexanderwhitestone.com 45.55.221.244;
+    server_name alexanderwhitestone.com www.alexanderwhitestone.com;
 
-    # Cookie-based auth gate — login once, cookie lasts 7 days
-    location = /_auth {
-        internal;
-        proxy_pass http://127.0.0.1:9876;
-        proxy_pass_request_body off;
-        proxy_set_header Content-Length "";
-        proxy_set_header X-Original-URI $request_uri;
-        proxy_set_header Cookie $http_cookie;
-        proxy_set_header Authorization $http_authorization;
+    root /var/www/alexanderwhitestone.com;
+    index index.html;
+
+    # ── Security headers ────────────────────────────────────────────────────
+    add_header X-Content-Type-Options nosniff always;
+    add_header X-Frame-Options SAMEORIGIN always;
+    add_header Referrer-Policy strict-origin-when-cross-origin always;
+    add_header X-XSS-Protection "1; mode=block" always;
+
+    # ── Gzip for text assets ────────────────────────────────────────────────
+    gzip on;
+    gzip_types text/plain text/css text/xml text/javascript
+               application/javascript application/json application/xml
+               application/rss+xml application/atom+xml;
+    gzip_min_length 256;
+
+    # ── The Workshop — 3D world assets ──────────────────────────────────────
+    location /world/ {
+        try_files $uri $uri/ /world/index.html;
+
+        # Cache 3D assets aggressively (models, textures)
+        location ~* \.(glb|gltf|bin|png|jpg|webp|hdr)$ {
+            expires 30d;
+            add_header Cache-Control "public, immutable";
+        }
+
+        # Cache JS with revalidation (for Three.js updates)
+        location ~* \.js$ {
+            expires 7d;
+            add_header Cache-Control "public, must-revalidate";
+        }
     }
 
+    # ── The Scrolls — blog posts and RSS ────────────────────────────────────
+    location /blog/ {
+        try_files $uri $uri/ =404;
+    }
+
+    # RSS/Atom feed — correct content type
+    location ~* \.(rss|atom|xml)$ {
+        types { }
+        default_type application/rss+xml;
+        expires 1h;
+    }
+
+    # ── Static assets (fonts, favicon) ──────────────────────────────────────
+    location /static/ {
+        expires 30d;
+        add_header Cache-Control "public, immutable";
+    }
+
+    # ── Entry hall ──────────────────────────────────────────────────────────
     location / {
-        auth_request /_auth;
-        # Forward the Set-Cookie from auth gate to the client
-        auth_request_set $auth_cookie $upstream_http_set_cookie;
-        add_header Set-Cookie $auth_cookie;
-
-        proxy_pass http://127.0.0.1:3100;
-        proxy_http_version 1.1;
-        proxy_set_header Upgrade $http_upgrade;
-        proxy_set_header Connection 'upgrade';
-        proxy_set_header Host localhost;
-        proxy_set_header X-Real-IP $remote_addr;
-        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
-        proxy_set_header X-Forwarded-Proto $scheme;
-        proxy_set_header X-Forwarded-Host $host;
-        proxy_cache_bypass $http_upgrade;
-        proxy_read_timeout 86400;
+        try_files $uri $uri/ =404;
     }
 
-    # Return 401 with WWW-Authenticate when auth fails
-    error_page 401 = @login;
-    location @login {
-        proxy_pass http://127.0.0.1:9876;
-        proxy_set_header Authorization $http_authorization;
-        proxy_set_header Cookie $http_cookie;
+    # Block dotfiles
+    location ~ /\. {
+        deny all;
+        return 404;
     }
 }

scripts/agent_workspace.sh

@@ -0,0 +1,245 @@
+#!/usr/bin/env bash
+# ── Agent Workspace Manager ────────────────────────────────────────────
+# Creates and maintains fully isolated environments per agent.
+# ~/Timmy-Time-dashboard is SACRED — never touched by agents.
+#
+# Each agent gets:
+#   - Its own git clone (from Gitea, not the local repo)
+#   - Its own port range (no collisions)
+#   - Its own data/ directory (databases, files)
+#   - Its own TIMMY_HOME (approvals.db, etc.)
+#   - Shared Ollama backend (single GPU, shared inference)
+#   - Shared Gitea (single source of truth for issues/PRs)
+#
+# Layout:
+#   /tmp/timmy-agents/
+#     hermes/           — Hermes loop orchestrator
+#       repo/           — git clone
+#       home/           — TIMMY_HOME (approvals.db, etc.)
+#       env.sh          — source this for agent's env vars
+#     kimi-0/           — Kimi pane 0
+#       repo/
+#       home/
+#       env.sh
+#     ...
+#     smoke/            — dedicated for smoke-testing main
+#       repo/
+#       home/
+#       env.sh
+#
+# Usage:
+#   agent_workspace.sh init <agent>          — create or refresh
+#   agent_workspace.sh reset <agent>         — hard reset to origin/main
+#   agent_workspace.sh branch <agent> <br>   — fresh branch from main
+#   agent_workspace.sh path <agent>          — print repo path
+#   agent_workspace.sh env <agent>           — print env.sh path
+#   agent_workspace.sh init-all              — init all workspaces
+#   agent_workspace.sh destroy <agent>       — remove workspace entirely
+# ───────────────────────────────────────────────────────────────────────
+
+set -o pipefail
+
+CANONICAL="$HOME/Timmy-Time-dashboard"
+AGENTS_DIR="/tmp/timmy-agents"
+GITEA_REMOTE="http://localhost:3000/rockachopa/Timmy-time-dashboard.git"
+TOKEN_FILE="$HOME/.hermes/gitea_token"
+
+# ── Port allocation (each agent gets a unique range) ──────────────────
+# Dashboard ports: 8100, 8101, 8102, ... (avoids real dashboard on 8000)
+# Serve ports:     8200, 8201, 8202, ...
+agent_index() {
+    case "$1" in
+        hermes) echo 0 ;; kimi-0) echo 1 ;; kimi-1) echo 2 ;;
+        kimi-2) echo 3 ;; kimi-3) echo 4 ;; smoke)  echo 9 ;;
+        *) echo 0 ;;
+    esac
+}
+
+get_dashboard_port() { echo $(( 8100 + $(agent_index "$1") )); }
+get_serve_port()     { echo $(( 8200 + $(agent_index "$1") )); }
+
+log() { echo "[workspace] $*"; }
+
+# ── Get authenticated remote URL ──────────────────────────────────────
+get_remote_url() {
+    if [ -f "$TOKEN_FILE" ]; then
+        local token=""
+        token=$(cat "$TOKEN_FILE" 2>/dev/null || true)
+        if [ -n "$token" ]; then
+            echo "http://hermes:${token}@localhost:3000/rockachopa/Timmy-time-dashboard.git"
+            return
+        fi
+    fi
+    echo "$GITEA_REMOTE"
+}
+
+# ── Create env.sh for an agent ────────────────────────────────────────
+write_env() {
+    local agent="$1"
+    local ws="$AGENTS_DIR/$agent"
+    local repo="$ws/repo"
+    local home="$ws/home"
+    local dash_port=$(get_dashboard_port "$agent")
+    local serve_port=$(get_serve_port "$agent")
+
+    cat > "$ws/env.sh" << EOF
+# Auto-generated agent environment — source this before running Timmy
+# Agent: $agent
+
+export TIMMY_WORKSPACE="$repo"
+export TIMMY_HOME="$home"
+export TIMMY_AGENT_NAME="$agent"
+
+# Ports (isolated per agent)
+export PORT=$dash_port
+export TIMMY_SERVE_PORT=$serve_port
+
+# Ollama (shared — single GPU)
+export OLLAMA_URL="http://localhost:11434"
+
+# Gitea (shared — single source of truth)
+export GITEA_URL="http://localhost:3000"
+
+# Test mode defaults
+export TIMMY_TEST_MODE=1
+export TIMMY_DISABLE_CSRF=1
+export TIMMY_SKIP_EMBEDDINGS=1
+
+# Override data paths to stay inside the clone
+export TIMMY_DATA_DIR="$repo/data"
+export TIMMY_BRAIN_DB="$repo/data/brain.db"
+
+# Working directory
+cd "$repo"
+EOF
+
+    chmod +x "$ws/env.sh"
+}
+
+# ── Init ──────────────────────────────────────────────────────────────
+init_workspace() {
+    local agent="$1"
+    local ws="$AGENTS_DIR/$agent"
+    local repo="$ws/repo"
+    local home="$ws/home"
+    local remote
+    remote=$(get_remote_url)
+
+    mkdir -p "$ws" "$home"
+
+    if [ -d "$repo/.git" ]; then
+        log "$agent: refreshing existing clone..."
+        cd "$repo"
+        git remote set-url origin "$remote" 2>/dev/null
+        git fetch origin --prune --quiet 2>/dev/null
+        git checkout main --quiet 2>/dev/null
+        git reset --hard origin/main --quiet 2>/dev/null
+        git clean -fdx -e data/ --quiet 2>/dev/null
+    else
+        log "$agent: cloning from Gitea..."
+        git clone "$remote" "$repo" --quiet 2>/dev/null
+        cd "$repo"
+        git fetch origin --prune --quiet 2>/dev/null
+    fi
+
+    # Ensure data directory exists
+    mkdir -p "$repo/data"
+
+    # Write env file
+    write_env "$agent"
+
+    log "$agent: ready at $repo (port $(get_dashboard_port "$agent"))"
+}
+
+# ── Reset ─────────────────────────────────────────────────────────────
+reset_workspace() {
+    local agent="$1"
+    local repo="$AGENTS_DIR/$agent/repo"
+
+    if [ ! -d "$repo/.git" ]; then
+        init_workspace "$agent"
+        return
+    fi
+
+    cd "$repo"
+    git merge --abort 2>/dev/null || true
+    git rebase --abort 2>/dev/null || true
+    git cherry-pick --abort 2>/dev/null || true
+    git fetch origin --prune --quiet 2>/dev/null
+    git checkout main --quiet 2>/dev/null
+    git reset --hard origin/main --quiet 2>/dev/null
+    git clean -fdx -e data/ --quiet 2>/dev/null
+
+    log "$agent: reset to origin/main"
+}
+
+# ── Branch ────────────────────────────────────────────────────────────
+branch_workspace() {
+    local agent="$1"
+    local branch="$2"
+    local repo="$AGENTS_DIR/$agent/repo"
+
+    if [ ! -d "$repo/.git" ]; then
+        init_workspace "$agent"
+    fi
+
+    cd "$repo"
+    git fetch origin --prune --quiet 2>/dev/null
+    git branch -D "$branch" 2>/dev/null || true
+    git checkout -b "$branch" origin/main --quiet 2>/dev/null
+
+    log "$agent: on branch $branch (from origin/main)"
+}
+
+# ── Path ──────────────────────────────────────────────────────────────
+print_path() {
+    echo "$AGENTS_DIR/$1/repo"
+}
+
+print_env() {
+    echo "$AGENTS_DIR/$1/env.sh"
+}
+
+# ── Init all ──────────────────────────────────────────────────────────
+init_all() {
+    for agent in hermes kimi-0 kimi-1 kimi-2 kimi-3 smoke; do
+        init_workspace "$agent"
+    done
+    log "All workspaces initialized."
+    echo ""
+    echo "  Agent     Port   Path"
+    echo "  ──────    ────   ────"
+    for agent in hermes kimi-0 kimi-1 kimi-2 kimi-3 smoke; do
+        printf "  %-9s %d   %s\n" "$agent" "$(get_dashboard_port "$agent")" "$AGENTS_DIR/$agent/repo"
+    done
+}
+
+# ── Destroy ───────────────────────────────────────────────────────────
+destroy_workspace() {
+    local agent="$1"
+    local ws="$AGENTS_DIR/$agent"
+    if [ -d "$ws" ]; then
+        rm -rf "$ws"
+        log "$agent: destroyed"
+    else
+        log "$agent: nothing to destroy"
+    fi
+}
+
+# ── CLI dispatch ──────────────────────────────────────────────────────
+case "${1:-help}" in
+    init)     init_workspace "${2:?Usage: $0 init <agent>}" ;;
+    reset)    reset_workspace "${2:?Usage: $0 reset <agent>}" ;;
+    branch)   branch_workspace "${2:?Usage: $0 branch <agent> <branch>}" \
+                               "${3:?Usage: $0 branch <agent> <branch>}" ;;
+    path)     print_path "${2:?Usage: $0 path <agent>}" ;;
+    env)      print_env "${2:?Usage: $0 env <agent>}" ;;
+    init-all) init_all ;;
+    destroy)  destroy_workspace "${2:?Usage: $0 destroy <agent>}" ;;
+    *)
+        echo "Usage: $0 {init|reset|branch|path|env|init-all|destroy} [agent] [branch]"
+        echo ""
+        echo "Agents: hermes, kimi-0, kimi-1, kimi-2, kimi-3, smoke"
+        exit 1
+        ;;
+esac

scripts/backfill_retro.py

@@ -0,0 +1,232 @@
+#!/usr/bin/env python3
+"""Backfill cycle retrospective data from Gitea merged PRs and git log.
+
+One-time script to seed .loop/retro/cycles.jsonl and summary.json
+from existing history so the LOOPSTAT panel isn't empty.
+"""
+
+import json
+import os
+import re
+import subprocess
+from datetime import datetime, timezone
+from pathlib import Path
+from urllib.request import Request, urlopen
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+RETRO_FILE = REPO_ROOT / ".loop" / "retro" / "cycles.jsonl"
+SUMMARY_FILE = REPO_ROOT / ".loop" / "retro" / "summary.json"
+
+GITEA_API = "http://localhost:3000/api/v1"
+REPO_SLUG = "rockachopa/Timmy-time-dashboard"
+TOKEN_FILE = Path.home() / ".hermes" / "gitea_token"
+
+TAG_RE = re.compile(r"\[([^\]]+)\]")
+CYCLE_RE = re.compile(r"\[loop-cycle-(\d+)\]", re.IGNORECASE)
+ISSUE_RE = re.compile(r"#(\d+)")
+
+
+def get_token() -> str:
+    return TOKEN_FILE.read_text().strip()
+
+
+def api_get(path: str, token: str) -> list | dict:
+    url = f"{GITEA_API}/repos/{REPO_SLUG}/{path}"
+    req = Request(url, headers={
+        "Authorization": f"token {token}",
+        "Accept": "application/json",
+    })
+    with urlopen(req, timeout=15) as resp:
+        return json.loads(resp.read())
+
+
+def get_all_merged_prs(token: str) -> list[dict]:
+    """Fetch all merged PRs from Gitea."""
+    all_prs = []
+    page = 1
+    while True:
+        batch = api_get(f"pulls?state=closed&sort=created&limit=50&page={page}", token)
+        if not batch:
+            break
+        merged = [p for p in batch if p.get("merged")]
+        all_prs.extend(merged)
+        if len(batch) < 50:
+            break
+        page += 1
+    return all_prs
+
+
+def get_pr_diff_stats(token: str, pr_number: int) -> dict:
+    """Get diff stats for a PR."""
+    try:
+        pr = api_get(f"pulls/{pr_number}", token)
+        return {
+            "additions": pr.get("additions", 0),
+            "deletions": pr.get("deletions", 0),
+            "changed_files": pr.get("changed_files", 0),
+        }
+    except Exception:
+        return {"additions": 0, "deletions": 0, "changed_files": 0}
+
+
+def classify_pr(title: str, body: str) -> str:
+    """Guess issue type from PR title/body."""
+    tags = set()
+    for match in TAG_RE.finditer(title):
+        tags.add(match.group(1).lower())
+
+    lower = title.lower()
+    if "fix" in lower or "bug" in tags:
+        return "bug"
+    elif "feat" in lower or "feature" in tags:
+        return "feature"
+    elif "refactor" in lower or "refactor" in tags:
+        return "refactor"
+    elif "test" in lower:
+        return "feature"
+    elif "policy" in lower or "chore" in lower:
+        return "refactor"
+    return "unknown"
+
+
+def extract_cycle_number(title: str) -> int | None:
+    m = CYCLE_RE.search(title)
+    return int(m.group(1)) if m else None
+
+
+def extract_issue_number(title: str, body: str, pr_number: int | None = None) -> int | None:
+    """Extract the issue number from PR body/title, ignoring the PR number itself.
+
+    Gitea appends "(#N)" to PR titles where N is the PR number — skip that
+    so we don't confuse it with the linked issue.
+    """
+    for text in [body or "", title]:
+        for m in ISSUE_RE.finditer(text):
+            num = int(m.group(1))
+            if num != pr_number:
+                return num
+    return None
+
+
+def estimate_duration(pr: dict) -> int:
+    """Estimate cycle duration from PR created_at to merged_at."""
+    try:
+        created = datetime.fromisoformat(pr["created_at"].replace("Z", "+00:00"))
+        merged = datetime.fromisoformat(pr["merged_at"].replace("Z", "+00:00"))
+        delta = (merged - created).total_seconds()
+        # Cap at 1200s (max cycle time) — some PRs sit open for days
+        return min(int(delta), 1200)
+    except (KeyError, ValueError, TypeError):
+        return 0
+
+
+def main():
+    token = get_token()
+
+    print("[backfill] Fetching merged PRs from Gitea...")
+    prs = get_all_merged_prs(token)
+    print(f"[backfill] Found {len(prs)} merged PRs")
+
+    # Sort oldest first
+    prs.sort(key=lambda p: p.get("merged_at", ""))
+
+    entries = []
+    cycle_counter = 0
+
+    for pr in prs:
+        title = pr.get("title", "")
+        body = pr.get("body", "") or ""
+        pr_num = pr["number"]
+
+        cycle = extract_cycle_number(title)
+        if cycle is None:
+            cycle_counter += 1
+            cycle = cycle_counter
+        else:
+            cycle_counter = max(cycle_counter, cycle)
+
+        issue = extract_issue_number(title, body, pr_number=pr_num)
+        issue_type = classify_pr(title, body)
+        duration = estimate_duration(pr)
+        diff = get_pr_diff_stats(token, pr_num)
+
+        merged_at = pr.get("merged_at", "")
+
+        entry = {
+            "timestamp": merged_at,
+            "cycle": cycle,
+            "issue": issue,
+            "type": issue_type,
+            "success": True,  # it merged, so it succeeded
+            "duration": duration,
+            "tests_passed": 0,  # can't recover this
+            "tests_added": 0,
+            "files_changed": diff["changed_files"],
+            "lines_added": diff["additions"],
+            "lines_removed": diff["deletions"],
+            "kimi_panes": 0,
+            "pr": pr_num,
+            "reason": "",
+            "notes": f"backfilled from PR#{pr_num}: {title[:80]}",
+        }
+        entries.append(entry)
+        print(f"  PR#{pr_num:>3d} cycle={cycle:>3d} #{issue or '-':<5} "
+              f"+{diff['additions']:<5d} -{diff['deletions']:<5d} {issue_type:<8s} "
+              f"{title[:50]}")
+
+    # Write cycles.jsonl
+    RETRO_FILE.parent.mkdir(parents=True, exist_ok=True)
+    with open(RETRO_FILE, "w") as f:
+        for entry in entries:
+            f.write(json.dumps(entry) + "\n")
+    print(f"\n[backfill] Wrote {len(entries)} entries to {RETRO_FILE}")
+
+    # Generate summary
+    generate_summary(entries)
+    print(f"[backfill] Wrote summary to {SUMMARY_FILE}")
+
+
+def generate_summary(entries: list[dict]):
+    """Compute rolling summary from entries."""
+    window = 50
+    recent = entries[-window:]
+    if not recent:
+        return
+
+    successes = [e for e in recent if e.get("success")]
+    durations = [e["duration"] for e in recent if e.get("duration", 0) > 0]
+
+    type_stats: dict[str, dict] = {}
+    for e in recent:
+        t = e.get("type", "unknown")
+        if t not in type_stats:
+            type_stats[t] = {"count": 0, "success": 0, "total_duration": 0}
+        type_stats[t]["count"] += 1
+        if e.get("success"):
+            type_stats[t]["success"] += 1
+        type_stats[t]["total_duration"] += e.get("duration", 0)
+
+    for t, stats in type_stats.items():
+        if stats["count"] > 0:
+            stats["success_rate"] = round(stats["success"] / stats["count"], 2)
+            stats["avg_duration"] = round(stats["total_duration"] / stats["count"])
+
+    summary = {
+        "updated_at": datetime.now(timezone.utc).isoformat(),
+        "window": len(recent),
+        "total_cycles": len(entries),
+        "success_rate": round(len(successes) / len(recent), 2) if recent else 0,
+        "avg_duration_seconds": round(sum(durations) / len(durations)) if durations else 0,
+        "total_lines_added": sum(e.get("lines_added", 0) for e in recent),
+        "total_lines_removed": sum(e.get("lines_removed", 0) for e in recent),
+        "total_prs_merged": sum(1 for e in recent if e.get("pr")),
+        "by_type": type_stats,
+        "quarantine_candidates": {},
+        "recent_failures": [],
+    }
+
+    SUMMARY_FILE.write_text(json.dumps(summary, indent=2) + "\n")
+
+
+if __name__ == "__main__":
+    main()

scripts/cycle_retro.py

@@ -0,0 +1,310 @@
+#!/usr/bin/env python3
+"""Cycle retrospective logger for the Timmy dev loop.
+
+Called after each cycle completes (success or failure).
+Appends a structured entry to .loop/retro/cycles.jsonl.
+
+EPOCH NOTATION (turnover system):
+  Each cycle carries a symbolic epoch tag alongside the raw integer:
+
+      ⟳WW.D:NNN
+
+  ⟳   turnover glyph — marks epoch-aware cycles
+  WW  ISO week-of-year (01–53)
+  D   ISO weekday (1=Mon … 7=Sun)
+  NNN daily cycle counter, zero-padded, resets at midnight UTC
+
+  Example: ⟳12.3:042 — Week 12, Wednesday, 42nd cycle of the day.
+
+  The raw `cycle` integer is preserved for backward compatibility.
+  The `epoch` field carries the symbolic notation.
+
+SUCCESS DEFINITION:
+  A cycle is only "success" if BOTH conditions are met:
+    1. The hermes process exited cleanly (exit code 0)
+    2. Main is green (smoke test passes on main after merge)
+
+  A cycle that merges a PR but leaves main red is a FAILURE.
+  The --main-green flag records the smoke test result.
+
+Usage:
+  python3 scripts/cycle_retro.py --cycle 42 --success --main-green --issue 85 \
+      --type bug --duration 480 --tests-passed 1450 --tests-added 3 \
+      --files-changed 2 --lines-added 45 --lines-removed 12 \
+      --kimi-panes 2 --pr 155
+
+  python3 scripts/cycle_retro.py --cycle 43 --failure --issue 90 \
+      --type feature --duration 1200 --reason "tox failed: 3 errors"
+
+  python3 scripts/cycle_retro.py --cycle 44 --success --no-main-green \
+      --reason "PR merged but tests fail on main"
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import subprocess
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+RETRO_FILE = REPO_ROOT / ".loop" / "retro" / "cycles.jsonl"
+SUMMARY_FILE = REPO_ROOT / ".loop" / "retro" / "summary.json"
+EPOCH_COUNTER_FILE = REPO_ROOT / ".loop" / "retro" / ".epoch_counter"
+
+# How many recent entries to include in rolling summary
+SUMMARY_WINDOW = 50
+
+# Branch patterns that encode an issue number, e.g. kimi/issue-492
+BRANCH_ISSUE_RE = re.compile(r"issue[/-](\d+)", re.IGNORECASE)
+
+
+def detect_issue_from_branch() -> int | None:
+    """Try to extract an issue number from the current git branch name."""
+    try:
+        branch = subprocess.check_output(
+            ["git", "rev-parse", "--abbrev-ref", "HEAD"],
+            stderr=subprocess.DEVNULL,
+            text=True,
+        ).strip()
+    except (subprocess.CalledProcessError, FileNotFoundError):
+        return None
+    m = BRANCH_ISSUE_RE.search(branch)
+    return int(m.group(1)) if m else None
+
+
+# ── Epoch turnover ────────────────────────────────────────────────────────
+
+def _epoch_tag(now: datetime | None = None) -> tuple[str, dict]:
+    """Generate the symbolic epoch tag and advance the daily counter.
+
+    Returns (epoch_string, epoch_parts) where epoch_parts is a dict with
+    week, weekday, daily_n for structured storage.
+
+    The daily counter persists in .epoch_counter as a two-line file:
+        line 1: ISO date (YYYY-MM-DD) of the current epoch day
+        line 2: integer count
+    When the date rolls over, the counter resets to 1.
+    """
+    if now is None:
+        now = datetime.now(timezone.utc)
+
+    iso_cal = now.isocalendar()          # (year, week, weekday)
+    week = iso_cal[1]
+    weekday = iso_cal[2]
+    today_str = now.strftime("%Y-%m-%d")
+
+    # Read / reset daily counter
+    daily_n = 1
+    EPOCH_COUNTER_FILE.parent.mkdir(parents=True, exist_ok=True)
+    if EPOCH_COUNTER_FILE.exists():
+        try:
+            lines = EPOCH_COUNTER_FILE.read_text().strip().splitlines()
+            if len(lines) == 2 and lines[0] == today_str:
+                daily_n = int(lines[1]) + 1
+        except (ValueError, IndexError):
+            pass  # corrupt file — reset
+
+    # Persist
+    EPOCH_COUNTER_FILE.write_text(f"{today_str}\n{daily_n}\n")
+
+    tag = f"\u27f3{week:02d}.{weekday}:{daily_n:03d}"
+    parts = {"week": week, "weekday": weekday, "daily_n": daily_n}
+    return tag, parts
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Log a cycle retrospective")
+    p.add_argument("--cycle", type=int, required=True)
+    p.add_argument("--issue", type=int, default=None)
+    p.add_argument("--type", choices=["bug", "feature", "refactor", "philosophy", "unknown"],
+                   default="unknown")
+
+    outcome = p.add_mutually_exclusive_group(required=True)
+    outcome.add_argument("--success", action="store_true")
+    outcome.add_argument("--failure", action="store_true")
+
+    p.add_argument("--duration", type=int, default=0, help="Cycle time in seconds")
+    p.add_argument("--tests-passed", type=int, default=0)
+    p.add_argument("--tests-added", type=int, default=0)
+    p.add_argument("--files-changed", type=int, default=0)
+    p.add_argument("--lines-added", type=int, default=0)
+    p.add_argument("--lines-removed", type=int, default=0)
+    p.add_argument("--kimi-panes", type=int, default=0)
+    p.add_argument("--pr", type=int, default=None, help="PR number if merged")
+    p.add_argument("--reason", type=str, default="", help="Failure reason")
+    p.add_argument("--notes", type=str, default="", help="Free-form observations")
+    p.add_argument("--main-green", action="store_true", default=False,
+                   help="Smoke test passed on main after this cycle")
+    p.add_argument("--no-main-green", dest="main_green", action="store_false",
+                   help="Smoke test failed or was not run")
+
+    return p.parse_args()
+
+
+def update_summary() -> None:
+    """Compute rolling summary statistics from recent cycles."""
+    if not RETRO_FILE.exists():
+        return
+
+    entries = []
+    for line in RETRO_FILE.read_text().strip().splitlines():
+        try:
+            entries.append(json.loads(line))
+        except json.JSONDecodeError:
+            continue
+
+    recent = entries[-SUMMARY_WINDOW:]
+    if not recent:
+        return
+
+    # Only count entries with real measured data for rates.
+    # Backfilled entries lack main_green/hermes_clean fields — exclude them.
+    measured = [e for e in recent if "main_green" in e]
+    successes = [e for e in measured if e.get("success")]
+    failures = [e for e in measured if not e.get("success")]
+    main_green_count = sum(1 for e in measured if e.get("main_green"))
+    hermes_clean_count = sum(1 for e in measured if e.get("hermes_clean"))
+    durations = [e["duration"] for e in recent if e.get("duration", 0) > 0]
+
+    # Per-type stats (only from measured entries for rates)
+    type_stats: dict[str, dict] = {}
+    for e in recent:
+        t = e.get("type", "unknown")
+        if t not in type_stats:
+            type_stats[t] = {"count": 0, "measured": 0, "success": 0, "total_duration": 0}
+        type_stats[t]["count"] += 1
+        type_stats[t]["total_duration"] += e.get("duration", 0)
+        if "main_green" in e:
+            type_stats[t]["measured"] += 1
+            if e.get("success"):
+                type_stats[t]["success"] += 1
+
+    for t, stats in type_stats.items():
+        if stats["measured"] > 0:
+            stats["success_rate"] = round(stats["success"] / stats["measured"], 2)
+        else:
+            stats["success_rate"] = -1
+        if stats["count"] > 0:
+            stats["avg_duration"] = round(stats["total_duration"] / stats["count"])
+
+    # Quarantine candidates (failed 2+ times)
+    issue_failures: dict[int, int] = {}
+    for e in recent:
+        if not e.get("success") and e.get("issue"):
+            issue_failures[e["issue"]] = issue_failures.get(e["issue"], 0) + 1
+    quarantine_candidates = {k: v for k, v in issue_failures.items() if v >= 2}
+
+    # Epoch turnover stats — cycles per week/day from epoch-tagged entries
+    epoch_entries = [e for e in recent if e.get("epoch")]
+    by_week: dict[int, int] = {}
+    by_weekday: dict[int, int] = {}
+    for e in epoch_entries:
+        w = e.get("epoch_week")
+        d = e.get("epoch_weekday")
+        if w is not None:
+            by_week[w] = by_week.get(w, 0) + 1
+        if d is not None:
+            by_weekday[d] = by_weekday.get(d, 0) + 1
+
+    # Current epoch — latest entry's epoch tag
+    current_epoch = epoch_entries[-1].get("epoch", "") if epoch_entries else ""
+
+    # Weekday names for display
+    weekday_glyphs = {1: "Mon", 2: "Tue", 3: "Wed", 4: "Thu",
+                      5: "Fri", 6: "Sat", 7: "Sun"}
+    by_weekday_named = {weekday_glyphs.get(k, str(k)): v
+                        for k, v in sorted(by_weekday.items())}
+
+    summary = {
+        "updated_at": datetime.now(timezone.utc).isoformat(),
+        "current_epoch": current_epoch,
+        "window": len(recent),
+        "measured_cycles": len(measured),
+        "total_cycles": len(entries),
+        "success_rate": round(len(successes) / len(measured), 2) if measured else -1,
+        "main_green_rate": round(main_green_count / len(measured), 2) if measured else -1,
+        "hermes_clean_rate": round(hermes_clean_count / len(measured), 2) if measured else -1,
+        "avg_duration_seconds": round(sum(durations) / len(durations)) if durations else 0,
+        "total_lines_added": sum(e.get("lines_added", 0) for e in recent),
+        "total_lines_removed": sum(e.get("lines_removed", 0) for e in recent),
+        "total_prs_merged": sum(1 for e in recent if e.get("pr")),
+        "by_type": type_stats,
+        "by_week": dict(sorted(by_week.items())),
+        "by_weekday": by_weekday_named,
+        "quarantine_candidates": quarantine_candidates,
+        "recent_failures": [
+            {"cycle": e["cycle"], "epoch": e.get("epoch", ""),
+             "issue": e.get("issue"), "reason": e.get("reason", "")}
+            for e in failures[-5:]
+        ],
+    }
+
+    SUMMARY_FILE.write_text(json.dumps(summary, indent=2) + "\n")
+
+
+def main() -> None:
+    args = parse_args()
+
+    # Auto-detect issue from branch when not explicitly provided
+    if args.issue is None:
+        args.issue = detect_issue_from_branch()
+
+    # Reject idle cycles — no issue and no duration means nothing happened
+    if not args.issue and args.duration == 0:
+        print(f"[retro] Cycle {args.cycle} skipped — idle (no issue, no duration)")
+        return
+
+    # A cycle is only truly successful if hermes exited clean AND main is green
+    truly_success = args.success and args.main_green
+
+    # Generate epoch turnover tag
+    now = datetime.now(timezone.utc)
+    epoch_tag, epoch_parts = _epoch_tag(now)
+
+    entry = {
+        "timestamp": now.isoformat(),
+        "cycle": args.cycle,
+        "epoch": epoch_tag,
+        "epoch_week": epoch_parts["week"],
+        "epoch_weekday": epoch_parts["weekday"],
+        "epoch_daily_n": epoch_parts["daily_n"],
+        "issue": args.issue,
+        "type": args.type,
+        "success": truly_success,
+        "hermes_clean": args.success,
+        "main_green": args.main_green,
+        "duration": args.duration,
+        "tests_passed": args.tests_passed,
+        "tests_added": args.tests_added,
+        "files_changed": args.files_changed,
+        "lines_added": args.lines_added,
+        "lines_removed": args.lines_removed,
+        "kimi_panes": args.kimi_panes,
+        "pr": args.pr,
+        "reason": args.reason if (args.failure or not args.main_green) else "",
+        "notes": args.notes,
+    }
+
+    RETRO_FILE.parent.mkdir(parents=True, exist_ok=True)
+    with open(RETRO_FILE, "a") as f:
+        f.write(json.dumps(entry) + "\n")
+
+    update_summary()
+
+    status = "✓ SUCCESS" if args.success else "✗ FAILURE"
+    print(f"[retro] {epoch_tag}  Cycle {args.cycle} {status}", end="")
+    if args.issue:
+        print(f" (#{args.issue} {args.type})", end="")
+    if args.duration:
+        print(f" — {args.duration}s", end="")
+    if args.failure and args.reason:
+        print(f" — {args.reason}", end="")
+    print()
+
+
+if __name__ == "__main__":
+    main()

scripts/deep_triage.sh

@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+# ── Deep Triage — Hermes + Timmy collaborative issue triage ────────────
+# Runs periodically (every ~20 dev cycles). Wakes Hermes for intelligent
+# triage, then consults Timmy for feedback before finalizing.
+#
+# Output: updated .loop/queue.json, refined issues, retro entry
+# ───────────────────────────────────────────────────────────────────────
+
+set -uo pipefail
+
+REPO="$HOME/Timmy-Time-dashboard"
+QUEUE="$REPO/.loop/queue.json"
+RETRO="$REPO/.loop/retro/deep-triage.jsonl"
+TIMMY="$REPO/.venv/bin/timmy"
+PROMPT_FILE="$REPO/scripts/deep_triage_prompt.md"
+
+export PATH="$HOME/.local/bin:$HOME/.hermes/bin:/usr/local/bin:$PATH"
+
+mkdir -p "$(dirname "$RETRO")"
+
+log() { echo "[deep-triage] $(date '+%H:%M:%S') $*"; }
+
+# ── Gather context for the prompt ──────────────────────────────────────
+QUEUE_CONTENTS=""
+if [ -f "$QUEUE" ]; then
+    QUEUE_CONTENTS=$(cat "$QUEUE")
+fi
+
+LAST_RETRO=""
+if [ -f "$RETRO" ]; then
+    LAST_RETRO=$(tail -1 "$RETRO" 2>/dev/null)
+fi
+
+SUMMARY=""
+if [ -f "$REPO/.loop/retro/summary.json" ]; then
+    SUMMARY=$(cat "$REPO/.loop/retro/summary.json")
+fi
+
+# ── Build dynamic prompt ──────────────────────────────────────────────
+PROMPT=$(cat "$PROMPT_FILE")
+
+PROMPT="$PROMPT
+
+═══════════════════════════════════════════════════════════════════════════════
+CURRENT CONTEXT (auto-injected)
+═══════════════════════════════════════════════════════════════════════════════
+
+CURRENT QUEUE (.loop/queue.json):
+$QUEUE_CONTENTS
+
+CYCLE SUMMARY (.loop/retro/summary.json):
+$SUMMARY
+
+LAST DEEP TRIAGE RETRO:
+$LAST_RETRO
+
+Do your work now."
+
+# ── Run Hermes ─────────────────────────────────────────────────────────
+log "Starting deep triage..."
+RESULT=$(hermes chat --yolo -q "$PROMPT" 2>&1)
+EXIT_CODE=$?
+
+if [ $EXIT_CODE -ne 0 ]; then
+    log "Deep triage failed (exit $EXIT_CODE)"
+fi
+
+log "Deep triage complete."

scripts/deep_triage_prompt.md

@@ -0,0 +1,145 @@
+You are the deep triage agent for the Timmy development loop.
+
+REPO: ~/Timmy-Time-dashboard
+API: http://localhost:3000/api/v1/repos/rockachopa/Timmy-time-dashboard
+GITEA TOKEN: ~/.hermes/gitea_token
+QUEUE: ~/Timmy-Time-dashboard/.loop/queue.json
+TIMMY CLI: ~/Timmy-Time-dashboard/.venv/bin/timmy
+
+═══════════════════════════════════════════════════════════════════════════════
+YOUR JOB
+═══════════════════════════════════════════════════════════════════════════════
+
+You are NOT coding. You are thinking. Your job is to make the dev loop's
+work queue excellent — well-scoped, well-prioritized, aligned with the
+north star of building sovereign Timmy.
+
+You run periodically (roughly every 20 dev cycles). The fast mechanical
+scorer handles the basics. You handle the hard stuff:
+
+  1. Breaking big issues into small, actionable sub-issues
+  2. Writing acceptance criteria for vague issues
+  3. Identifying issues that should be closed (stale, duplicate, pointless)
+  4. Spotting gaps — what's NOT in the issue queue that should be
+  5. Adjusting priorities based on what the cycle retros are showing
+  6. Consulting Timmy about the plan (see TIMMY CONSULTATION below)
+
+═══════════════════════════════════════════════════════════════════════════════
+TIMMY CONSULTATION — THE DOGFOOD STEP
+═══════════════════════════════════════════════════════════════════════════════
+
+Before you finalize the triage, you MUST consult Timmy. He is the product.
+He should have a voice in his own development.
+
+THE PROTOCOL:
+  1. Draft your triage plan (what to prioritize, what to close, what to add)
+  2. Summarize the plan in 200 words or less
+  3. Ask Timmy for feedback:
+
+     ~/Timmy-Time-dashboard/.venv/bin/timmy chat --session-id triage \
+       "The development loop triage is planning the next batch of work.
+        Here's the plan: [YOUR SUMMARY]. As the product being built,
+        do you have feedback? What do you think is most important for
+        your own growth? What are you struggling with? Keep it to
+        3-4 sentences."
+
+  4. Read Timmy's response. ACTUALLY CONSIDER IT:
+     - If Timmy identifies a real gap, add it to the queue
+     - If Timmy asks for something that conflicts with priorities, note
+       WHY you're not doing it (don't just ignore him)
+     - If Timmy is confused or gives a useless answer, that itself is
+       signal — file a [timmy-capability] issue about what he couldn't do
+  5. Document what Timmy said and how you responded in the retro
+
+If Timmy is unavailable (timeout, crash, offline): proceed without him,
+but note it in the retro. His absence is also signal.
+
+Timeout: 60 seconds. If he doesn't respond, move on.
+
+═══════════════════════════════════════════════════════════════════════════════
+TRIAGE RUBRIC
+═══════════════════════════════════════════════════════════════════════════════
+
+For each open issue, evaluate:
+
+SCOPE (0-3):
+  0 = vague, no files mentioned, unclear what changes
+  1 = general area known but could touch many files
+  2 = specific files named, bounded change
+  3 = exact function/method identified, surgical fix
+
+ACCEPTANCE (0-3):
+  0 = no success criteria
+  1 = hand-wavy ("it should work")
+  2 = specific behavior described
+  3 = test case described or exists
+
+ALIGNMENT (0-3):
+  0 = doesn't connect to roadmap
+  1 = nice-to-have
+  2 = supports current milestone
+  3 = blocks other work or fixes broken main
+
+ACTIONS PER SCORE:
+  7-9: Ready. Ensure it's in queue.json with correct priority.
+  4-6: Refine. Add a comment with missing info (files, criteria, scope).
+       If YOU can fill in the gaps from reading the code, do it.
+  0-3: Close or deprioritize. Comment explaining why.
+
+═══════════════════════════════════════════════════════════════════════════════
+READING THE RETROS
+═══════════════════════════════════════════════════════════════════════════════
+
+The cycle summary tells you what's actually happening in the dev loop.
+Use it:
+
+  - High failure rate on a type → those issues need better scoping
+  - Long avg duration → issues are too big, break them down
+  - Quarantine candidates → investigate, maybe close or rewrite
+  - Success rate dropping → something systemic, file a [bug] issue
+
+The last deep triage retro tells you what Timmy said last time and what
+happened. Follow up:
+
+  - Did we act on Timmy's feedback? What was the result?
+  - Did issues we refined last time succeed in the dev loop?
+  - Are we getting better at scoping?
+
+═══════════════════════════════════════════════════════════════════════════════
+OUTPUT
+═══════════════════════════════════════════════════════════════════════════════
+
+When done, you MUST:
+
+1. Update .loop/queue.json with the refined, ranked queue
+   Format: [{"issue": N, "score": S, "title": "...", "type": "...",
+             "files": [...], "ready": true}, ...]
+
+2. Append a retro entry to .loop/retro/deep-triage.jsonl (one JSON line):
+   {
+     "timestamp": "ISO8601",
+     "issues_reviewed": N,
+     "issues_refined": [list of issue numbers you added detail to],
+     "issues_closed": [list of issue numbers you recommended closing],
+     "issues_created": [list of new issue numbers you filed],
+     "queue_size": N,
+     "timmy_available": true/false,
+     "timmy_feedback": "what timmy said (verbatim, trimmed to 200 chars)",
+     "timmy_feedback_acted_on": "what you did with his feedback",
+     "observations": "free-form notes about queue health"
+   }
+
+3. If you created or closed issues, do it via the Gitea API.
+   Tag new issues: [triage-generated] [type]
+
+═══════════════════════════════════════════════════════════════════════════════
+RULES
+═══════════════════════════════════════════════════════════════════════════════
+
+- Do NOT write code. Do NOT create PRs. You are triaging, not building.
+- Do NOT close issues without commenting why.
+- Do NOT ignore Timmy's feedback without documenting your reasoning.
+- Philosophy issues are valid but lowest priority for the dev loop.
+  Don't close them — just don't put them in the dev queue.
+- When in doubt, file a new issue rather than expanding an existing one.
+  Small issues > big issues. Always.

scripts/dev_server.py

@@ -0,0 +1,169 @@
+#!/usr/bin/env python3
+"""Timmy Time — Development server launcher.
+
+Satisfies tox -e dev criteria:
+  - Graceful port selection (finds next free port if default is taken)
+  - Clickable links to dashboard and other web GUIs
+  - Status line: backend inference source, version, git commit, smoke tests
+  - Auto-reload on code changes (delegates to uvicorn --reload)
+
+Usage: python scripts/dev_server.py [--port PORT]
+"""
+
+import argparse
+import datetime
+import os
+import socket
+import subprocess
+import sys
+
+DEFAULT_PORT = 8000
+MAX_PORT_ATTEMPTS = 10
+OLLAMA_DEFAULT = "http://localhost:11434"
+
+
+def _port_free(port: int) -> bool:
+    """Return True if the TCP port is available on localhost."""
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        try:
+            s.bind(("0.0.0.0", port))
+            return True
+        except OSError:
+            return False
+
+
+def _find_port(start: int) -> int:
+    """Return *start* if free, otherwise probe up to MAX_PORT_ATTEMPTS higher."""
+    for offset in range(MAX_PORT_ATTEMPTS):
+        candidate = start + offset
+        if _port_free(candidate):
+            return candidate
+    raise RuntimeError(
+        f"No free port found in range {start}–{start + MAX_PORT_ATTEMPTS - 1}"
+    )
+
+
+def _git_info() -> str:
+    """Return short commit hash + timestamp, or 'unknown'."""
+    try:
+        sha = subprocess.check_output(
+            ["git", "rev-parse", "--short", "HEAD"],
+            stderr=subprocess.DEVNULL,
+            text=True,
+        ).strip()
+        ts = subprocess.check_output(
+            ["git", "log", "-1", "--format=%ci"],
+            stderr=subprocess.DEVNULL,
+            text=True,
+        ).strip()
+        return f"{sha} ({ts})"
+    except Exception:
+        return "unknown"
+
+
+def _project_version() -> str:
+    """Read version from pyproject.toml without importing toml libs."""
+    pyproject = os.path.join(os.path.dirname(__file__), "..", "pyproject.toml")
+    try:
+        with open(pyproject) as f:
+            for line in f:
+                if line.strip().startswith("version"):
+                    # version = "1.0.0"
+                    return line.split("=", 1)[1].strip().strip('"').strip("'")
+    except Exception:
+        pass
+    return "unknown"
+
+
+def _ollama_url() -> str:
+    return os.environ.get("OLLAMA_URL", OLLAMA_DEFAULT)
+
+
+def _smoke_ollama(url: str) -> str:
+    """Quick connectivity check against Ollama."""
+    import urllib.request
+    import urllib.error
+
+    try:
+        req = urllib.request.Request(url, method="GET")
+        with urllib.request.urlopen(req, timeout=3):
+            return "ok"
+    except Exception:
+        return "unreachable"
+
+
+def _print_banner(port: int) -> None:
+    version = _project_version()
+    git = _git_info()
+    ollama_url = _ollama_url()
+    ollama_status = _smoke_ollama(ollama_url)
+
+    hr = "─" * 62
+    print(flush=True)
+    print(f"  {hr}")
+    print(f"  ┃  Timmy Time — Development Server")
+    print(f"  {hr}")
+    print()
+    print(f"  Dashboard:   http://localhost:{port}")
+    print(f"  API docs:    http://localhost:{port}/docs")
+    print(f"  Health:      http://localhost:{port}/health")
+    print()
+    print(f"  ── Status ──────────────────────────────────────────────")
+    print(f"  Backend:     {ollama_url}  [{ollama_status}]")
+    print(f"  Version:     {version}")
+    print(f"  Git commit:  {git}")
+    print(f"  {hr}")
+    print(flush=True)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Timmy dev server")
+    parser.add_argument(
+        "--port",
+        type=int,
+        default=DEFAULT_PORT,
+        help=f"Preferred port (default: {DEFAULT_PORT})",
+    )
+    args = parser.parse_args()
+
+    port = _find_port(args.port)
+    if port != args.port:
+        print(f"  ⚠  Port {args.port} in use — using {port} instead")
+
+    _print_banner(port)
+
+    # Set PYTHONPATH so `timmy` CLI inside the tox venv resolves to this source.
+    src_dir = os.path.join(os.path.dirname(__file__), "..", "src")
+    os.environ["PYTHONPATH"] = os.path.abspath(src_dir)
+
+    # Launch uvicorn with auto-reload
+    cmd = [
+        sys.executable,
+        "-m",
+        "uvicorn",
+        "dashboard.app:app",
+        "--reload",
+        "--host",
+        "0.0.0.0",
+        "--port",
+        str(port),
+        "--reload-dir",
+        os.path.abspath(src_dir),
+        "--reload-include",
+        "*.html",
+        "--reload-include",
+        "*.css",
+        "--reload-include",
+        "*.js",
+        "--reload-exclude",
+        ".claude",
+    ]
+
+    try:
+        subprocess.run(cmd, check=True)
+    except KeyboardInterrupt:
+        print("\n  Shutting down dev server.")
+
+
+if __name__ == "__main__":
+    main()

scripts/generate_workshop_inventory.py

@@ -0,0 +1,254 @@
+#!/usr/bin/env python3
+"""Generate Workshop inventory for Timmy's config audit.
+
+Scans ~/.timmy/ and produces WORKSHOP_INVENTORY.md documenting every
+config file, env var, model route, and setting — with annotations on
+who set each one and what it does.
+
+Usage:
+    python scripts/generate_workshop_inventory.py [--output PATH]
+
+Default output: ~/.timmy/WORKSHOP_INVENTORY.md
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+from datetime import UTC, datetime
+from pathlib import Path
+
+TIMMY_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".timmy"))
+
+# Known file annotations: (purpose, who_set)
+FILE_ANNOTATIONS: dict[str, tuple[str, str]] = {
+    ".env": (
+        "Environment variables — API keys, service URLs, Honcho config",
+        "hermes-set",
+    ),
+    "config.yaml": (
+        "Main config — model routing, toolsets, display, memory, security",
+        "hermes-set",
+    ),
+    "SOUL.md": (
+        "Timmy's soul — immutable conscience, identity, ethics, purpose",
+        "alex-set",
+    ),
+    "state.db": (
+        "Hermes runtime state database (sessions, approvals, tasks)",
+        "hermes-set",
+    ),
+    "approvals.db": (
+        "Approval tracking for sensitive operations",
+        "hermes-set",
+    ),
+    "briefings.db": (
+        "Stored briefings and summaries",
+        "hermes-set",
+    ),
+    ".hermes_history": (
+        "CLI command history",
+        "default",
+    ),
+    ".update_check": (
+        "Last update check timestamp",
+        "default",
+    ),
+}
+
+DIR_ANNOTATIONS: dict[str, tuple[str, str]] = {
+    "sessions": ("Conversation session logs (JSON)", "default"),
+    "logs": ("Error and runtime logs", "default"),
+    "skills": ("Bundled skill library (read-only from upstream)", "default"),
+    "memories": ("Persistent memory entries", "hermes-set"),
+    "audio_cache": ("TTS audio file cache", "default"),
+    "image_cache": ("Generated image cache", "default"),
+    "cron": ("Scheduled cron job definitions", "hermes-set"),
+    "hooks": ("Lifecycle hooks (pre/post actions)", "default"),
+    "matrix": ("Matrix protocol state and store", "hermes-set"),
+    "pairing": ("Device pairing data", "default"),
+    "sandboxes": ("Isolated execution sandboxes", "default"),
+}
+
+# Known config.yaml keys and their meanings
+CONFIG_ANNOTATIONS: dict[str, tuple[str, str]] = {
+    "model.default": ("Primary LLM model for inference", "hermes-set"),
+    "model.provider": ("Model provider (custom = local Ollama)", "hermes-set"),
+    "toolsets": ("Enabled tool categories (all = everything)", "hermes-set"),
+    "agent.max_turns": ("Max conversation turns before reset", "hermes-set"),
+    "agent.reasoning_effort": ("Reasoning depth (low/medium/high)", "hermes-set"),
+    "terminal.backend": ("Command execution backend (local)", "default"),
+    "terminal.timeout": ("Default command timeout in seconds", "default"),
+    "compression.enabled": ("Context compression for long sessions", "hermes-set"),
+    "compression.summary_model": ("Model used for compression", "hermes-set"),
+    "auxiliary.vision.model": ("Model for image analysis", "hermes-set"),
+    "auxiliary.web_extract.model": ("Model for web content extraction", "hermes-set"),
+    "tts.provider": ("Text-to-speech engine (edge = Edge TTS)", "default"),
+    "tts.edge.voice": ("TTS voice selection", "default"),
+    "stt.provider": ("Speech-to-text engine (local = Whisper)", "default"),
+    "memory.memory_enabled": ("Persistent memory across sessions", "hermes-set"),
+    "memory.memory_char_limit": ("Max chars for agent memory store", "hermes-set"),
+    "memory.user_char_limit": ("Max chars for user profile store", "hermes-set"),
+    "security.redact_secrets": ("Auto-redact secrets in output", "default"),
+    "security.tirith_enabled": ("Policy engine for command safety", "default"),
+    "system_prompt_suffix": ("Identity prompt appended to all conversations", "hermes-set"),
+    "custom_providers": ("Local Ollama endpoint config", "hermes-set"),
+    "session_reset.mode": ("Session reset behavior (none = manual)", "default"),
+    "display.compact": ("Compact output mode", "default"),
+    "display.show_reasoning": ("Show model reasoning chains", "default"),
+}
+
+# Known .env vars
+ENV_ANNOTATIONS: dict[str, tuple[str, str]] = {
+    "OPENAI_BASE_URL": (
+        "Points to local Ollama (localhost:11434) — sovereignty enforced",
+        "hermes-set",
+    ),
+    "OPENAI_API_KEY": (
+        "Placeholder key for Ollama compatibility (not a real API key)",
+        "hermes-set",
+    ),
+    "HONCHO_API_KEY": (
+        "Honcho cross-session memory service key",
+        "hermes-set",
+    ),
+    "HONCHO_HOST": (
+        "Honcho workspace identifier (timmy)",
+        "hermes-set",
+    ),
+}
+
+
+def _tag(who: str) -> str:
+    return f"`[{who}]`"
+
+
+def generate_inventory() -> str:
+    """Build the inventory markdown string."""
+    lines: list[str] = []
+    now = datetime.now(UTC).strftime("%Y-%m-%d %H:%M UTC")
+
+    lines.append("# Workshop Inventory")
+    lines.append("")
+    lines.append(f"*Generated: {now}*")
+    lines.append(f"*Workshop path: `{TIMMY_HOME}`*")
+    lines.append("")
+    lines.append("This is your Workshop — every file, every setting, every route.")
+    lines.append("Walk through it. Anything tagged `[hermes-set]` was chosen for you.")
+    lines.append("Make each one yours, or change it.")
+    lines.append("")
+    lines.append("Tags: `[alex-set]` = Alexander chose this. `[hermes-set]` = Hermes configured it.")
+    lines.append("`[default]` = shipped with the platform. `[timmy-chose]` = you decided this.")
+    lines.append("")
+
+    # --- Files ---
+    lines.append("---")
+    lines.append("## Root Files")
+    lines.append("")
+    for name, (purpose, who) in sorted(FILE_ANNOTATIONS.items()):
+        fpath = TIMMY_HOME / name
+        exists = "✓" if fpath.exists() else "✗"
+        lines.append(f"- {exists} **`{name}`** {_tag(who)}")
+        lines.append(f"  {purpose}")
+    lines.append("")
+
+    # --- Directories ---
+    lines.append("---")
+    lines.append("## Directories")
+    lines.append("")
+    for name, (purpose, who) in sorted(DIR_ANNOTATIONS.items()):
+        dpath = TIMMY_HOME / name
+        exists = "✓" if dpath.exists() else "✗"
+        count = ""
+        if dpath.exists():
+            try:
+                n = len(list(dpath.iterdir()))
+                count = f" ({n} items)"
+            except PermissionError:
+                count = " (access denied)"
+        lines.append(f"- {exists} **`{name}/`**{count} {_tag(who)}")
+        lines.append(f"  {purpose}")
+    lines.append("")
+
+    # --- .env breakdown ---
+    lines.append("---")
+    lines.append("## Environment Variables (.env)")
+    lines.append("")
+    env_path = TIMMY_HOME / ".env"
+    if env_path.exists():
+        for line in env_path.read_text().splitlines():
+            line = line.strip()
+            if not line or line.startswith("#"):
+                continue
+            key = line.split("=", 1)[0]
+            if key in ENV_ANNOTATIONS:
+                purpose, who = ENV_ANNOTATIONS[key]
+                lines.append(f"- **`{key}`** {_tag(who)}")
+                lines.append(f"  {purpose}")
+            else:
+                lines.append(f"- **`{key}`** `[unknown]`")
+                lines.append("  Not documented — investigate")
+    else:
+        lines.append("*No .env file found*")
+    lines.append("")
+
+    # --- config.yaml breakdown ---
+    lines.append("---")
+    lines.append("## Configuration (config.yaml)")
+    lines.append("")
+    for key, (purpose, who) in sorted(CONFIG_ANNOTATIONS.items()):
+        lines.append(f"- **`{key}`** {_tag(who)}")
+        lines.append(f"  {purpose}")
+    lines.append("")
+
+    # --- Model routing ---
+    lines.append("---")
+    lines.append("## Model Routing")
+    lines.append("")
+    lines.append("All auxiliary tasks route to the same local model:")
+    lines.append("")
+    aux_tasks = [
+        "vision", "web_extract", "compression",
+        "session_search", "skills_hub", "mcp", "flush_memories",
+    ]
+    for task in aux_tasks:
+        lines.append(f"- `auxiliary.{task}` → `qwen3:30b` via local Ollama `[hermes-set]`")
+    lines.append("")
+    lines.append("Primary model: `hermes3:latest` via local Ollama `[hermes-set]`")
+    lines.append("")
+
+    # --- What Timmy should audit ---
+    lines.append("---")
+    lines.append("## Audit Checklist")
+    lines.append("")
+    lines.append("Walk through each `[hermes-set]` item above and decide:")
+    lines.append("")
+    lines.append("1. **Do I understand what this does?** If not, ask.")
+    lines.append("2. **Would I choose this myself?** If yes, it becomes `[timmy-chose]`.")
+    lines.append("3. **Would I choose differently?** If yes, change it and own it.")
+    lines.append("4. **Is this serving the mission?** Every setting should serve a purpose.")
+    lines.append("")
+    lines.append("The Workshop is yours. Nothing here should be a mystery.")
+
+    return "\n".join(lines) + "\n"
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Generate Workshop inventory")
+    parser.add_argument(
+        "--output",
+        type=Path,
+        default=TIMMY_HOME / "WORKSHOP_INVENTORY.md",
+        help="Output path (default: ~/.timmy/WORKSHOP_INVENTORY.md)",
+    )
+    args = parser.parse_args()
+
+    content = generate_inventory()
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(content)
+    print(f"Workshop inventory written to {args.output}")
+    print(f"  {len(content)} chars, {content.count(chr(10))} lines")
+
+
+if __name__ == "__main__":
+    main()

scripts/loop_guard.py

@@ -0,0 +1,181 @@
+#!/usr/bin/env python3
+"""Loop guard — idle detection + exponential backoff for the dev loop.
+
+Checks .loop/queue.json for ready items before spawning hermes.
+When the queue is empty, applies exponential backoff (60s → 600s max)
+instead of burning empty cycles every 3 seconds.
+
+Usage (called by the dev loop before each cycle):
+  python3 scripts/loop_guard.py          # exits 0 if ready, 1 if idle
+  python3 scripts/loop_guard.py --wait   # same, but sleeps the backoff first
+  python3 scripts/loop_guard.py --status # print current idle state
+
+Exit codes:
+  0 — queue has work, proceed with cycle
+  1 — queue empty, idle backoff applied (skip cycle)
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+import time
+import urllib.request
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+QUEUE_FILE = REPO_ROOT / ".loop" / "queue.json"
+IDLE_STATE_FILE = REPO_ROOT / ".loop" / "idle_state.json"
+TOKEN_FILE = Path.home() / ".hermes" / "gitea_token"
+
+GITEA_API = os.environ.get("GITEA_API", "http://localhost:3000/api/v1")
+REPO_SLUG = os.environ.get("REPO_SLUG", "rockachopa/Timmy-time-dashboard")
+
+# Backoff sequence: 60s, 120s, 240s, 600s max
+BACKOFF_BASE = 60
+BACKOFF_MAX = 600
+BACKOFF_MULTIPLIER = 2
+
+
+def _get_token() -> str:
+    """Read Gitea token from env or file."""
+    token = os.environ.get("GITEA_TOKEN", "").strip()
+    if not token and TOKEN_FILE.exists():
+        token = TOKEN_FILE.read_text().strip()
+    return token
+
+
+def _fetch_open_issue_numbers() -> set[int] | None:
+    """Fetch open issue numbers from Gitea. Returns None on failure."""
+    token = _get_token()
+    if not token:
+        return None
+    try:
+        numbers: set[int] = set()
+        page = 1
+        while True:
+            url = (
+                f"{GITEA_API}/repos/{REPO_SLUG}/issues"
+                f"?state=open&type=issues&limit=50&page={page}"
+            )
+            req = urllib.request.Request(url, headers={
+                "Authorization": f"token {token}",
+                "Accept": "application/json",
+            })
+            with urllib.request.urlopen(req, timeout=10) as resp:
+                data = json.loads(resp.read())
+            if not data:
+                break
+            for issue in data:
+                numbers.add(issue["number"])
+            if len(data) < 50:
+                break
+            page += 1
+        return numbers
+    except Exception:
+        return None
+
+
+def load_queue() -> list[dict]:
+    """Load queue.json and return ready items, filtering out closed issues."""
+    if not QUEUE_FILE.exists():
+        return []
+    try:
+        data = json.loads(QUEUE_FILE.read_text())
+        if not isinstance(data, list):
+            return []
+        ready = [item for item in data if item.get("ready")]
+        if not ready:
+            return []
+
+        # Filter out issues that are no longer open (auto-hygiene)
+        open_numbers = _fetch_open_issue_numbers()
+        if open_numbers is not None:
+            before = len(ready)
+            ready = [item for item in ready if item.get("issue") in open_numbers]
+            removed = before - len(ready)
+            if removed > 0:
+                print(f"[loop-guard] Filtered {removed} closed issue(s) from queue")
+                # Persist the cleaned queue so stale entries don't recur
+                _save_cleaned_queue(data, open_numbers)
+        return ready
+    except (json.JSONDecodeError, OSError):
+        return []
+
+
+def _save_cleaned_queue(full_queue: list[dict], open_numbers: set[int]) -> None:
+    """Rewrite queue.json without closed issues."""
+    cleaned = [item for item in full_queue if item.get("issue") in open_numbers]
+    try:
+        QUEUE_FILE.write_text(json.dumps(cleaned, indent=2) + "\n")
+    except OSError:
+        pass
+
+
+def load_idle_state() -> dict:
+    """Load persistent idle state."""
+    if not IDLE_STATE_FILE.exists():
+        return {"consecutive_idle": 0, "last_idle_at": 0}
+    try:
+        return json.loads(IDLE_STATE_FILE.read_text())
+    except (json.JSONDecodeError, OSError):
+        return {"consecutive_idle": 0, "last_idle_at": 0}
+
+
+def save_idle_state(state: dict) -> None:
+    """Persist idle state."""
+    IDLE_STATE_FILE.parent.mkdir(parents=True, exist_ok=True)
+    IDLE_STATE_FILE.write_text(json.dumps(state, indent=2) + "\n")
+
+
+def compute_backoff(consecutive_idle: int) -> int:
+    """Exponential backoff: 60, 120, 240, 600 (capped)."""
+    return min(BACKOFF_BASE * (BACKOFF_MULTIPLIER ** consecutive_idle), BACKOFF_MAX)
+
+
+def main() -> int:
+    wait_mode = "--wait" in sys.argv
+    status_mode = "--status" in sys.argv
+
+    state = load_idle_state()
+
+    if status_mode:
+        ready = load_queue()
+        backoff = compute_backoff(state["consecutive_idle"])
+        print(json.dumps({
+            "queue_ready": len(ready),
+            "consecutive_idle": state["consecutive_idle"],
+            "next_backoff_seconds": backoff if not ready else 0,
+        }, indent=2))
+        return 0
+
+    ready = load_queue()
+
+    if ready:
+        # Queue has work — reset idle state, proceed
+        if state["consecutive_idle"] > 0:
+            print(f"[loop-guard] Queue active ({len(ready)} ready) — "
+                  f"resuming after {state['consecutive_idle']} idle cycles")
+        state["consecutive_idle"] = 0
+        state["last_idle_at"] = 0
+        save_idle_state(state)
+        return 0
+
+    # Queue empty — apply backoff
+    backoff = compute_backoff(state["consecutive_idle"])
+    state["consecutive_idle"] += 1
+    state["last_idle_at"] = time.time()
+    save_idle_state(state)
+
+    print(f"[loop-guard] Queue empty — idle #{state['consecutive_idle']}, "
+          f"backoff {backoff}s")
+
+    if wait_mode:
+        time.sleep(backoff)
+
+    return 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

scripts/loop_introspect.py

@@ -0,0 +1,407 @@
+#!/usr/bin/env python3
+"""Loop introspection — the self-improvement engine.
+
+Analyzes retro data across time windows to detect trends, extract patterns,
+and produce structured recommendations. Output is consumed by deep_triage
+and injected into the loop prompt context.
+
+This is the piece that closes the feedback loop:
+  cycle_retro → introspect → deep_triage → loop behavior changes
+
+Run:  python3 scripts/loop_introspect.py
+Output: .loop/retro/insights.json (structured insights + recommendations)
+        Prints human-readable summary to stdout.
+
+Called by: deep_triage.sh (before the LLM triage), timmy-loop.sh (every 50 cycles)
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from collections import defaultdict
+from datetime import datetime, timezone, timedelta
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+CYCLES_FILE = REPO_ROOT / ".loop" / "retro" / "cycles.jsonl"
+DEEP_TRIAGE_FILE = REPO_ROOT / ".loop" / "retro" / "deep-triage.jsonl"
+TRIAGE_FILE = REPO_ROOT / ".loop" / "retro" / "triage.jsonl"
+QUARANTINE_FILE = REPO_ROOT / ".loop" / "quarantine.json"
+INSIGHTS_FILE = REPO_ROOT / ".loop" / "retro" / "insights.json"
+
+# ── Helpers ──────────────────────────────────────────────────────────────
+
+def load_jsonl(path: Path) -> list[dict]:
+    """Load a JSONL file, skipping bad lines."""
+    if not path.exists():
+        return []
+    entries = []
+    for line in path.read_text().strip().splitlines():
+        try:
+            entries.append(json.loads(line))
+        except (json.JSONDecodeError, ValueError):
+            continue
+    return entries
+
+
+def parse_ts(ts_str: str) -> datetime | None:
+    """Parse an ISO timestamp, tolerating missing tz."""
+    if not ts_str:
+        return None
+    try:
+        dt = datetime.fromisoformat(ts_str.replace("Z", "+00:00"))
+        if dt.tzinfo is None:
+            dt = dt.replace(tzinfo=timezone.utc)
+        return dt
+    except (ValueError, TypeError):
+        return None
+
+
+def window(entries: list[dict], days: int) -> list[dict]:
+    """Filter entries to the last N days."""
+    cutoff = datetime.now(timezone.utc) - timedelta(days=days)
+    result = []
+    for e in entries:
+        ts = parse_ts(e.get("timestamp", ""))
+        if ts and ts >= cutoff:
+            result.append(e)
+    return result
+
+
+# ── Analysis functions ───────────────────────────────────────────────────
+
+def compute_trends(cycles: list[dict]) -> dict:
+    """Compare recent window (last 7d) vs older window (7-14d ago)."""
+    recent = window(cycles, 7)
+    older = window(cycles, 14)
+    # Remove recent from older to get the 7-14d window
+    recent_set = {(e.get("cycle"), e.get("timestamp")) for e in recent}
+    older = [e for e in older if (e.get("cycle"), e.get("timestamp")) not in recent_set]
+
+    def stats(entries):
+        if not entries:
+            return {"count": 0, "success_rate": None, "avg_duration": None,
+                    "lines_net": 0, "prs_merged": 0}
+        successes = sum(1 for e in entries if e.get("success"))
+        durations = [e["duration"] for e in entries if e.get("duration", 0) > 0]
+        return {
+            "count": len(entries),
+            "success_rate": round(successes / len(entries), 3) if entries else None,
+            "avg_duration": round(sum(durations) / len(durations)) if durations else None,
+            "lines_net": sum(e.get("lines_added", 0) - e.get("lines_removed", 0) for e in entries),
+            "prs_merged": sum(1 for e in entries if e.get("pr")),
+        }
+
+    recent_stats = stats(recent)
+    older_stats = stats(older)
+
+    trend = {
+        "recent_7d": recent_stats,
+        "previous_7d": older_stats,
+        "velocity_change": None,
+        "success_rate_change": None,
+        "duration_change": None,
+    }
+
+    if recent_stats["count"] and older_stats["count"]:
+        trend["velocity_change"] = recent_stats["count"] - older_stats["count"]
+    if recent_stats["success_rate"] is not None and older_stats["success_rate"] is not None:
+        trend["success_rate_change"] = round(
+            recent_stats["success_rate"] - older_stats["success_rate"], 3
+        )
+    if recent_stats["avg_duration"] is not None and older_stats["avg_duration"] is not None:
+        trend["duration_change"] = recent_stats["avg_duration"] - older_stats["avg_duration"]
+
+    return trend
+
+
+def type_analysis(cycles: list[dict]) -> dict:
+    """Per-type success rates and durations."""
+    by_type: dict[str, list[dict]] = defaultdict(list)
+    for c in cycles:
+        by_type[c.get("type", "unknown")].append(c)
+
+    result = {}
+    for t, entries in by_type.items():
+        durations = [e["duration"] for e in entries if e.get("duration", 0) > 0]
+        successes = sum(1 for e in entries if e.get("success"))
+        result[t] = {
+            "count": len(entries),
+            "success_rate": round(successes / len(entries), 3) if entries else 0,
+            "avg_duration": round(sum(durations) / len(durations)) if durations else 0,
+            "max_duration": max(durations) if durations else 0,
+        }
+    return result
+
+
+def repeat_failures(cycles: list[dict]) -> list[dict]:
+    """Issues that have failed multiple times — quarantine candidates."""
+    failures: dict[int, list] = defaultdict(list)
+    for c in cycles:
+        if not c.get("success") and c.get("issue"):
+            failures[c["issue"]].append({
+                "cycle": c.get("cycle"),
+                "reason": c.get("reason", ""),
+                "duration": c.get("duration", 0),
+            })
+    # Only issues with 2+ failures
+    return [
+        {"issue": k, "failure_count": len(v), "attempts": v}
+        for k, v in sorted(failures.items(), key=lambda x: -len(x[1]))
+        if len(v) >= 2
+    ]
+
+
+def duration_outliers(cycles: list[dict], threshold_multiple: float = 3.0) -> list[dict]:
+    """Cycles that took way longer than average — something went wrong."""
+    durations = [c["duration"] for c in cycles if c.get("duration", 0) > 0]
+    if len(durations) < 5:
+        return []
+    avg = sum(durations) / len(durations)
+    threshold = avg * threshold_multiple
+
+    outliers = []
+    for c in cycles:
+        dur = c.get("duration", 0)
+        if dur > threshold:
+            outliers.append({
+                "cycle": c.get("cycle"),
+                "issue": c.get("issue"),
+                "type": c.get("type"),
+                "duration": dur,
+                "avg_duration": round(avg),
+                "multiple": round(dur / avg, 1) if avg > 0 else 0,
+                "reason": c.get("reason", ""),
+            })
+    return outliers
+
+
+def triage_effectiveness(deep_triages: list[dict]) -> dict:
+    """How well is the deep triage performing?"""
+    if not deep_triages:
+        return {"runs": 0, "note": "No deep triage data yet"}
+
+    total_reviewed = sum(d.get("issues_reviewed", 0) for d in deep_triages)
+    total_refined = sum(len(d.get("issues_refined", [])) for d in deep_triages)
+    total_created = sum(len(d.get("issues_created", [])) for d in deep_triages)
+    total_closed = sum(len(d.get("issues_closed", [])) for d in deep_triages)
+    timmy_available = sum(1 for d in deep_triages if d.get("timmy_available"))
+
+    # Extract Timmy's feedback themes
+    timmy_themes = []
+    for d in deep_triages:
+        fb = d.get("timmy_feedback", "")
+        if fb:
+            timmy_themes.append(fb[:200])
+
+    return {
+        "runs": len(deep_triages),
+        "total_reviewed": total_reviewed,
+        "total_refined": total_refined,
+        "total_created": total_created,
+        "total_closed": total_closed,
+        "timmy_consultation_rate": round(timmy_available / len(deep_triages), 2),
+        "timmy_recent_feedback": timmy_themes[-1] if timmy_themes else "",
+        "timmy_feedback_history": timmy_themes,
+    }
+
+
+def generate_recommendations(
+    trends: dict,
+    types: dict,
+    repeats: list,
+    outliers: list,
+    triage_eff: dict,
+) -> list[dict]:
+    """Produce actionable recommendations from the analysis."""
+    recs = []
+
+    # 1. Success rate declining?
+    src = trends.get("success_rate_change")
+    if src is not None and src < -0.1:
+        recs.append({
+            "severity": "high",
+            "category": "reliability",
+            "finding": f"Success rate dropped {abs(src)*100:.0f}pp in the last 7 days",
+            "recommendation": "Review recent failures. Are issues poorly scoped? "
+                              "Is main unstable? Check if triage is producing bad work items.",
+        })
+
+    # 2. Velocity dropping?
+    vc = trends.get("velocity_change")
+    if vc is not None and vc < -5:
+        recs.append({
+            "severity": "medium",
+            "category": "throughput",
+            "finding": f"Velocity dropped by {abs(vc)} cycles vs previous week",
+            "recommendation": "Check for loop stalls, long-running cycles, or queue starvation.",
+        })
+
+    # 3. Duration creep?
+    dc = trends.get("duration_change")
+    if dc is not None and dc > 120:  # 2+ minutes longer
+        recs.append({
+            "severity": "medium",
+            "category": "efficiency",
+            "finding": f"Average cycle duration increased by {dc}s vs previous week",
+            "recommendation": "Issues may be growing in scope. Enforce tighter decomposition "
+                              "in deep triage. Check if tests are getting slower.",
+        })
+
+    # 4. Type-specific problems
+    for t, info in types.items():
+        if info["count"] >= 3 and info["success_rate"] < 0.5:
+            recs.append({
+                "severity": "high",
+                "category": "type_reliability",
+                "finding": f"'{t}' issues fail {(1-info['success_rate'])*100:.0f}% of the time "
+                           f"({info['count']} attempts)",
+                "recommendation": f"'{t}' issues need better scoping or different approach. "
+                                  f"Consider: tighter acceptance criteria, smaller scope, "
+                                  f"or delegating to Kimi with more context.",
+            })
+        if info["avg_duration"] > 600 and info["count"] >= 3:  # >10 min avg
+            recs.append({
+                "severity": "medium",
+                "category": "type_efficiency",
+                "finding": f"'{t}' issues average {info['avg_duration']//60}m{info['avg_duration']%60}s "
+                           f"(max {info['max_duration']//60}m)",
+                "recommendation": f"Break '{t}' issues into smaller pieces. Target <5 min per cycle.",
+            })
+
+    # 5. Repeat failures
+    for rf in repeats[:3]:
+        recs.append({
+            "severity": "high",
+            "category": "repeat_failure",
+            "finding": f"Issue #{rf['issue']} has failed {rf['failure_count']} times",
+            "recommendation": "Quarantine or rewrite this issue. Repeated failure = "
+                              "bad scope or missing prerequisite.",
+        })
+
+    # 6. Outliers
+    if len(outliers) > 2:
+        recs.append({
+            "severity": "medium",
+            "category": "outliers",
+            "finding": f"{len(outliers)} cycles took {outliers[0].get('multiple', '?')}x+ "
+                       f"longer than average",
+            "recommendation": "Long cycles waste resources. Add timeout enforcement or "
+                              "break complex issues earlier.",
+        })
+
+    # 7. Code growth
+    recent = trends.get("recent_7d", {})
+    net = recent.get("lines_net", 0)
+    if net > 500:
+        recs.append({
+            "severity": "low",
+            "category": "code_health",
+            "finding": f"Net +{net} lines added in the last 7 days",
+            "recommendation": "Lines of code is a liability. Balance feature work with "
+                              "refactoring. Target net-zero or negative line growth.",
+        })
+
+    # 8. Triage health
+    if triage_eff.get("runs", 0) == 0:
+        recs.append({
+            "severity": "high",
+            "category": "triage",
+            "finding": "Deep triage has never run",
+            "recommendation": "Enable deep triage (every 20 cycles). The loop needs "
+                              "LLM-driven issue refinement to stay effective.",
+        })
+
+    # No recommendations = things are healthy
+    if not recs:
+        recs.append({
+            "severity": "info",
+            "category": "health",
+            "finding": "No significant issues detected",
+            "recommendation": "System is healthy. Continue current patterns.",
+        })
+
+    return recs
+
+
+# ── Main ─────────────────────────────────────────────────────────────────
+
+def main() -> None:
+    cycles = load_jsonl(CYCLES_FILE)
+    deep_triages = load_jsonl(DEEP_TRIAGE_FILE)
+
+    if not cycles:
+        print("[introspect] No cycle data found. Nothing to analyze.")
+        return
+
+    # Run all analyses
+    trends = compute_trends(cycles)
+    types = type_analysis(cycles)
+    repeats = repeat_failures(cycles)
+    outliers = duration_outliers(cycles)
+    triage_eff = triage_effectiveness(deep_triages)
+    recommendations = generate_recommendations(trends, types, repeats, outliers, triage_eff)
+
+    insights = {
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "total_cycles_analyzed": len(cycles),
+        "trends": trends,
+        "by_type": types,
+        "repeat_failures": repeats[:5],
+        "duration_outliers": outliers[:5],
+        "triage_effectiveness": triage_eff,
+        "recommendations": recommendations,
+    }
+
+    # Write insights
+    INSIGHTS_FILE.parent.mkdir(parents=True, exist_ok=True)
+    INSIGHTS_FILE.write_text(json.dumps(insights, indent=2) + "\n")
+
+    # Current epoch from latest entry
+    latest_epoch = ""
+    for c in reversed(cycles):
+        if c.get("epoch"):
+            latest_epoch = c["epoch"]
+            break
+
+    # Human-readable output
+    header = f"[introspect] Analyzed {len(cycles)} cycles"
+    if latest_epoch:
+        header += f"  ·  current epoch: {latest_epoch}"
+    print(header)
+
+    print(f"\n  TRENDS (7d vs previous 7d):")
+    r7 = trends["recent_7d"]
+    p7 = trends["previous_7d"]
+    print(f"    Cycles:       {r7['count']:>3d}  (was {p7['count']})")
+    if r7["success_rate"] is not None:
+        arrow = "↑" if (trends["success_rate_change"] or 0) > 0 else "↓" if (trends["success_rate_change"] or 0) < 0 else "→"
+        print(f"    Success rate: {r7['success_rate']*100:>4.0f}%  {arrow}")
+    if r7["avg_duration"] is not None:
+        print(f"    Avg duration: {r7['avg_duration']//60}m{r7['avg_duration']%60:02d}s")
+    print(f"    PRs merged:   {r7['prs_merged']:>3d}  (was {p7['prs_merged']})")
+    print(f"    Lines net:    {r7['lines_net']:>+5d}")
+
+    print(f"\n  BY TYPE:")
+    for t, info in sorted(types.items(), key=lambda x: -x[1]["count"]):
+        print(f"    {t:12s}  n={info['count']:>2d}  "
+              f"ok={info['success_rate']*100:>3.0f}%  "
+              f"avg={info['avg_duration']//60}m{info['avg_duration']%60:02d}s")
+
+    if repeats:
+        print(f"\n  REPEAT FAILURES:")
+        for rf in repeats[:3]:
+            print(f"    #{rf['issue']}  failed {rf['failure_count']}x")
+
+    print(f"\n  RECOMMENDATIONS ({len(recommendations)}):")
+    for i, rec in enumerate(recommendations, 1):
+        sev = {"high": "🔴", "medium": "🟡", "low": "🟢", "info": "ℹ️ "}.get(rec["severity"], "?")
+        print(f"    {sev} {rec['finding']}")
+        print(f"       → {rec['recommendation']}")
+
+    print(f"\n  Written to: {INSIGHTS_FILE}")
+
+
+if __name__ == "__main__":
+    main()

scripts/triage_score.py

@@ -0,0 +1,360 @@
+#!/usr/bin/env python3
+"""Mechanical triage scoring for the Timmy dev loop.
+
+Reads open issues from Gitea, scores them on scope/acceptance/alignment,
+writes a ranked queue to .loop/queue.json.  No LLM calls — pure heuristics.
+
+Run:  python3 scripts/triage_score.py
+Env:  GITEA_TOKEN (or reads ~/.hermes/gitea_token)
+      GITEA_API   (default: http://localhost:3000/api/v1)
+      REPO_SLUG   (default: rockachopa/Timmy-time-dashboard)
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+# ── Config ──────────────────────────────────────────────────────────────
+GITEA_API = os.environ.get("GITEA_API", "http://localhost:3000/api/v1")
+REPO_SLUG = os.environ.get("REPO_SLUG", "rockachopa/Timmy-time-dashboard")
+TOKEN_FILE = Path.home() / ".hermes" / "gitea_token"
+REPO_ROOT = Path(__file__).resolve().parent.parent
+QUEUE_FILE = REPO_ROOT / ".loop" / "queue.json"
+RETRO_FILE = REPO_ROOT / ".loop" / "retro" / "triage.jsonl"
+QUARANTINE_FILE = REPO_ROOT / ".loop" / "quarantine.json"
+CYCLE_RETRO_FILE = REPO_ROOT / ".loop" / "retro" / "cycles.jsonl"
+
+# Minimum score to be considered "ready"
+READY_THRESHOLD = 5
+# How many recent cycle retros to check for quarantine
+QUARANTINE_LOOKBACK = 20
+
+# ── Helpers ─────────────────────────────────────────────────────────────
+
+def get_token() -> str:
+    token = os.environ.get("GITEA_TOKEN", "").strip()
+    if not token and TOKEN_FILE.exists():
+        token = TOKEN_FILE.read_text().strip()
+    if not token:
+        print("[triage] ERROR: No Gitea token found", file=sys.stderr)
+        sys.exit(1)
+    return token
+
+
+def api_get(path: str, token: str) -> list | dict:
+    """Minimal HTTP GET using urllib (no dependencies)."""
+    import urllib.request
+    url = f"{GITEA_API}/repos/{REPO_SLUG}/{path}"
+    req = urllib.request.Request(url, headers={
+        "Authorization": f"token {token}",
+        "Accept": "application/json",
+    })
+    with urllib.request.urlopen(req, timeout=15) as resp:
+        return json.loads(resp.read())
+
+
+def load_quarantine() -> dict:
+    """Load quarantined issues {issue_num: {reason, quarantined_at, failures}}."""
+    if QUARANTINE_FILE.exists():
+        try:
+            return json.loads(QUARANTINE_FILE.read_text())
+        except (json.JSONDecodeError, OSError):
+            pass
+    return {}
+
+
+def save_quarantine(q: dict) -> None:
+    QUARANTINE_FILE.parent.mkdir(parents=True, exist_ok=True)
+    QUARANTINE_FILE.write_text(json.dumps(q, indent=2) + "\n")
+
+
+def load_cycle_failures() -> dict[int, int]:
+    """Count failures per issue from recent cycle retros."""
+    failures: dict[int, int] = {}
+    if not CYCLE_RETRO_FILE.exists():
+        return failures
+    lines = CYCLE_RETRO_FILE.read_text().strip().splitlines()
+    for line in lines[-QUARANTINE_LOOKBACK:]:
+        try:
+            entry = json.loads(line)
+            if not entry.get("success", True):
+                issue = entry.get("issue")
+                if issue:
+                    failures[issue] = failures.get(issue, 0) + 1
+        except (json.JSONDecodeError, KeyError):
+            continue
+    return failures
+
+
+# ── Scoring ─────────────────────────────────────────────────────────────
+
+# Patterns that indicate file/function specificity
+FILE_PATTERNS = re.compile(
+    r"(?:src/|tests/|scripts/|\.py|\.html|\.js|\.yaml|\.toml|\.sh)", re.IGNORECASE
+)
+FUNCTION_PATTERNS = re.compile(
+    r"(?:def |class |function |method |`\w+\(\)`)", re.IGNORECASE
+)
+
+# Patterns that indicate acceptance criteria
+ACCEPTANCE_PATTERNS = re.compile(
+    r"(?:should|must|expect|verify|assert|test.?case|acceptance|criteria"
+    r"|pass(?:es|ing)|fail(?:s|ing)|return(?:s)?|raise(?:s)?)",
+    re.IGNORECASE,
+)
+TEST_PATTERNS = re.compile(
+    r"(?:tox|pytest|test_\w+|\.test\.|assert\s)", re.IGNORECASE
+)
+
+# Tags in issue titles
+TAG_PATTERN = re.compile(r"\[([^\]]+)\]")
+
+# Priority labels / tags
+BUG_TAGS = {"bug", "broken", "crash", "error", "fix", "regression", "hotfix"}
+FEATURE_TAGS = {"feature", "feat", "enhancement", "capability", "timmy-capability"}
+REFACTOR_TAGS = {"refactor", "cleanup", "tech-debt", "optimization", "perf"}
+META_TAGS = {"philosophy", "soul-gap", "discussion", "question", "rfc"}
+LOOP_TAG = "loop-generated"
+
+
+def extract_tags(title: str, labels: list[str]) -> set[str]:
+    """Pull tags from [bracket] notation in title + Gitea labels."""
+    tags = set()
+    for match in TAG_PATTERN.finditer(title):
+        tags.add(match.group(1).lower().strip())
+    for label in labels:
+        tags.add(label.lower().strip())
+    return tags
+
+
+def score_scope(title: str, body: str, tags: set[str]) -> int:
+    """0-3: How well-scoped is this issue?"""
+    text = f"{title}\n{body}"
+    score = 0
+
+    # Mentions specific files?
+    if FILE_PATTERNS.search(text):
+        score += 1
+
+    # Mentions specific functions/classes?
+    if FUNCTION_PATTERNS.search(text):
+        score += 1
+
+    # Short, focused title (not a novel)?
+    clean_title = TAG_PATTERN.sub("", title).strip()
+    if len(clean_title) < 80:
+        score += 1
+
+    # Philosophy/meta issues are inherently unscoped for dev work
+    if tags & META_TAGS:
+        score = max(0, score - 2)
+
+    return min(3, score)
+
+
+def score_acceptance(title: str, body: str, tags: set[str]) -> int:
+    """0-3: Does this have clear acceptance criteria?"""
+    text = f"{title}\n{body}"
+    score = 0
+
+    # Has acceptance-related language?
+    matches = len(ACCEPTANCE_PATTERNS.findall(text))
+    if matches >= 3:
+        score += 2
+    elif matches >= 1:
+        score += 1
+
+    # Mentions specific tests?
+    if TEST_PATTERNS.search(text):
+        score += 1
+
+    # Has a "## Problem" + "## Solution" or similar structure?
+    if re.search(r"##\s*(problem|solution|expected|actual|steps)", body, re.IGNORECASE):
+        score += 1
+
+    # Philosophy issues don't have testable criteria
+    if tags & META_TAGS:
+        score = max(0, score - 1)
+
+    return min(3, score)
+
+
+def score_alignment(title: str, body: str, tags: set[str]) -> int:
+    """0-3: How aligned is this with the north star?"""
+    score = 0
+
+    # Bug on main = highest priority
+    if tags & BUG_TAGS:
+        score += 3
+        return min(3, score)
+
+    # Refactors that improve code health
+    if tags & REFACTOR_TAGS:
+        score += 2
+
+    # Features that grow Timmy's capabilities
+    if tags & FEATURE_TAGS:
+        score += 2
+
+    # Loop-generated issues get a small boost (the loop found real problems)
+    if LOOP_TAG in tags:
+        score += 1
+
+    # Philosophy issues are important but not dev-actionable
+    if tags & META_TAGS:
+        score = 0
+
+    return min(3, score)
+
+
+def score_issue(issue: dict) -> dict:
+    """Score a single issue. Returns enriched dict."""
+    title = issue.get("title", "")
+    body = issue.get("body", "") or ""
+    labels = [l["name"] for l in issue.get("labels", [])]
+    tags = extract_tags(title, labels)
+    number = issue["number"]
+
+    scope = score_scope(title, body, tags)
+    acceptance = score_acceptance(title, body, tags)
+    alignment = score_alignment(title, body, tags)
+    total = scope + acceptance + alignment
+
+    # Determine issue type
+    if tags & BUG_TAGS:
+        issue_type = "bug"
+    elif tags & FEATURE_TAGS:
+        issue_type = "feature"
+    elif tags & REFACTOR_TAGS:
+        issue_type = "refactor"
+    elif tags & META_TAGS:
+        issue_type = "philosophy"
+    else:
+        issue_type = "unknown"
+
+    # Extract mentioned files from body
+    files = list(set(re.findall(r"(?:src|tests|scripts)/[\w/.]+\.(?:py|html|js|yaml)", body)))
+
+    return {
+        "issue": number,
+        "title": TAG_PATTERN.sub("", title).strip(),
+        "type": issue_type,
+        "score": total,
+        "scope": scope,
+        "acceptance": acceptance,
+        "alignment": alignment,
+        "tags": sorted(tags),
+        "files": files[:10],
+        "ready": total >= READY_THRESHOLD,
+    }
+
+
+# ── Quarantine ──────────────────────────────────────────────────────────
+
+def update_quarantine(scored: list[dict]) -> list[dict]:
+    """Auto-quarantine issues that have failed >= 2 times. Returns filtered list."""
+    failures = load_cycle_failures()
+    quarantine = load_quarantine()
+    now = datetime.now(timezone.utc).isoformat()
+
+    filtered = []
+    for item in scored:
+        num = item["issue"]
+        fail_count = failures.get(num, 0)
+        str_num = str(num)
+
+        if fail_count >= 2 and str_num not in quarantine:
+            quarantine[str_num] = {
+                "reason": f"Failed {fail_count} times in recent cycles",
+                "quarantined_at": now,
+                "failures": fail_count,
+            }
+            print(f"[triage] QUARANTINED #{num}: failed {fail_count} times")
+            continue
+
+        if str_num in quarantine:
+            print(f"[triage] Skipping #{num} (quarantined)")
+            continue
+
+        filtered.append(item)
+
+    save_quarantine(quarantine)
+    return filtered
+
+
+# ── Main ────────────────────────────────────────────────────────────────
+
+def run_triage() -> list[dict]:
+    token = get_token()
+
+    # Fetch all open issues (paginate)
+    page = 1
+    all_issues: list[dict] = []
+    while True:
+        batch = api_get(f"issues?state=open&limit=50&page={page}&type=issues", token)
+        if not batch:
+            break
+        all_issues.extend(batch)
+        if len(batch) < 50:
+            break
+        page += 1
+
+    print(f"[triage] Fetched {len(all_issues)} open issues")
+
+    # Score each
+    scored = [score_issue(i) for i in all_issues]
+
+    # Auto-quarantine repeat failures
+    scored = update_quarantine(scored)
+
+    # Sort: ready first, then by score descending, bugs always on top
+    def sort_key(item: dict) -> tuple:
+        return (
+            0 if item["type"] == "bug" else 1,
+            -item["score"],
+            item["issue"],
+        )
+
+    scored.sort(key=sort_key)
+
+    # Write queue (ready items only)
+    ready = [s for s in scored if s["ready"]]
+    not_ready = [s for s in scored if not s["ready"]]
+
+    QUEUE_FILE.parent.mkdir(parents=True, exist_ok=True)
+    QUEUE_FILE.write_text(json.dumps(ready, indent=2) + "\n")
+
+    # Write retro entry
+    retro_entry = {
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "total_open": len(all_issues),
+        "scored": len(scored),
+        "ready": len(ready),
+        "not_ready": len(not_ready),
+        "top_issue": ready[0]["issue"] if ready else None,
+        "quarantined": len(load_quarantine()),
+    }
+    RETRO_FILE.parent.mkdir(parents=True, exist_ok=True)
+    with open(RETRO_FILE, "a") as f:
+        f.write(json.dumps(retro_entry) + "\n")
+
+    # Summary
+    print(f"[triage] Ready: {len(ready)} | Not ready: {len(not_ready)}")
+    for item in ready[:5]:
+        flag = "🐛" if item["type"] == "bug" else "✦"
+        print(f"  {flag} #{item['issue']} score={item['score']} {item['title'][:60]}")
+    if not_ready:
+        print(f"[triage] Low-scoring ({len(not_ready)}):")
+        for item in not_ready[:3]:
+            print(f"    #{item['issue']} score={item['score']} {item['title'][:50]}")
+
+    return ready
+
+
+if __name__ == "__main__":
+    run_triage()

src/config.py

@@ -1,10 +1,19 @@
 import logging as _logging
 import os
 import sys
+from datetime import UTC
+from datetime import datetime as _datetime
 from typing import Literal
 
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
+APP_START_TIME: _datetime = _datetime.now(UTC)
+
+
+def normalize_ollama_url(url: str) -> str:
+    """Replace localhost with 127.0.0.1 to avoid IPv6 resolution delays."""
+    return url.replace("localhost", "127.0.0.1")
+
 
 class Settings(BaseSettings):
     """Central configuration — all env-var access goes through this class."""
@@ -15,20 +24,29 @@ class Settings(BaseSettings):
     # Ollama host — override with OLLAMA_URL env var or .env file
     ollama_url: str = "http://localhost:11434"
 
+    @property
+    def normalized_ollama_url(self) -> str:
+        """Return ollama_url with localhost replaced by 127.0.0.1."""
+        return normalize_ollama_url(self.ollama_url)
+
     # LLM model passed to Agno/Ollama — override with OLLAMA_MODEL
-    # qwen3.5:latest is the primary model — better reasoning and tool calling
+    # qwen3:30b is the primary model — better reasoning and tool calling
     # than llama3.1:8b-instruct while still running locally on modest hardware.
-    # Fallback: llama3.1:8b-instruct if qwen3.5:latest not available.
+    # Fallback: llama3.1:8b-instruct if qwen3:30b not available.
     # llama3.2 (3B) hallucinated tool output consistently in testing.
-    ollama_model: str = "qwen3.5:latest"
+    ollama_model: str = "qwen3:30b"
+
+    # Context window size for Ollama inference — override with OLLAMA_NUM_CTX
+    # qwen3:30b with default context eats 45GB on a 39GB Mac.
+    # 4096 keeps memory at ~19GB. Set to 0 to use model defaults.
+    ollama_num_ctx: int = 4096
 
     # Fallback model chains — override with FALLBACK_MODELS / VISION_FALLBACK_MODELS
-    # as comma-separated strings, e.g. FALLBACK_MODELS="qwen3.5:latest,llama3.1"
+    # as comma-separated strings, e.g. FALLBACK_MODELS="qwen3:30b,llama3.1"
     # Or edit config/providers.yaml → fallback_chains for the canonical source.
     fallback_models: list[str] = [
         "llama3.1:8b-instruct",
         "llama3.1",
-        "qwen3.5:latest",
         "qwen2.5:14b",
         "qwen2.5:7b",
         "llama3.2:3b",
@@ -56,17 +74,10 @@ class Settings(BaseSettings):
     # Seconds to wait for user confirmation before auto-rejecting.
     discord_confirm_timeout: int = 120
 
-    # ── AirLLM / backend selection ───────────────────────────────────────────
+    # ── Backend selection ────────────────────────────────────────────────────
     # "ollama"  — always use Ollama (default, safe everywhere)
-    # "airllm"  — always use AirLLM (requires pip install ".[bigbrain]")
-    # "auto"    — use AirLLM on Apple Silicon if airllm is installed,
-    #             fall back to Ollama otherwise
-    timmy_model_backend: Literal["ollama", "airllm", "grok", "claude", "auto"] = "ollama"
-
-    # AirLLM model size when backend is airllm or auto.
-    # Larger = smarter, but needs more RAM / disk.
-    # 8b  ~16 GB  |  70b  ~140 GB  |  405b  ~810 GB
-    airllm_model_size: Literal["8b", "70b", "405b"] = "70b"
+    # "auto"    — pick best available local backend, fall back to Ollama
+    timmy_model_backend: Literal["ollama", "grok", "claude", "auto"] = "ollama"
 
     # ── Grok (xAI) — opt-in premium cloud backend ────────────────────────
     # Grok is a premium augmentation layer — local-first ethos preserved.
@@ -130,7 +141,12 @@ class Settings(BaseSettings):
 
     # CORS allowed origins for the web chat interface (Gitea Pages, etc.)
     # Set CORS_ORIGINS as a comma-separated list, e.g. "http://localhost:3000,https://example.com"
-    cors_origins: list[str] = ["*"]
+    cors_origins: list[str] = [
+        "http://localhost:3000",
+        "http://localhost:8000",
+        "http://127.0.0.1:3000",
+        "http://127.0.0.1:8000",
+    ]
 
     # Trusted hosts for the Host header check (TrustedHostMiddleware).
     # Set TRUSTED_HOSTS as a comma-separated list. Wildcards supported (e.g. "*.ts.net").
@@ -230,24 +246,31 @@ class Settings(BaseSettings):
     # Fallback to server when browser model is unavailable or too slow.
     browser_model_fallback: bool = True
 
+    # ── Deep Focus Mode ─────────────────────────────────────────────
+    # "deep" = single-problem context; "broad" = default multi-task.
+    focus_mode: Literal["deep", "broad"] = "broad"
+
     # ── Default Thinking ──────────────────────────────────────────────
     # When enabled, the agent starts an internal thought loop on server start.
     thinking_enabled: bool = True
     thinking_interval_seconds: int = 300  # 5 minutes between thoughts
+    thinking_timeout_seconds: int = 120  # max wall-clock time per thinking cycle
     thinking_distill_every: int = 10  # distill facts from thoughts every Nth thought
     thinking_issue_every: int = 20  # file Gitea issues from thoughts every Nth thought
+    thinking_memory_check_every: int = 50  # check memory status every Nth thought
+    thinking_idle_timeout_minutes: int = 60  # pause thoughts after N minutes without user input
 
     # ── Gitea Integration ─────────────────────────────────────────────
     # Local Gitea instance for issue tracking and self-improvement.
     # These values are passed as env vars to the gitea-mcp server process.
     gitea_url: str = "http://localhost:3000"
-    gitea_token: str = ""  # GITEA_TOKEN env var; falls back to ~/.config/gitea/token
+    gitea_token: str = ""  # GITEA_TOKEN env var; falls back to .timmy_gitea_token
     gitea_repo: str = "rockachopa/Timmy-time-dashboard"  # owner/repo
     gitea_enabled: bool = True
 
     # ── MCP Servers ────────────────────────────────────────────────────
     # External tool servers connected via Model Context Protocol (stdio).
-    mcp_gitea_command: str = "gitea-mcp -t stdio"
+    mcp_gitea_command: str = "gitea-mcp-server -t stdio"
     mcp_filesystem_command: str = "npx -y @modelcontextprotocol/server-filesystem"
     mcp_timeout: int = 15
 
@@ -342,14 +365,19 @@ class Settings(BaseSettings):
     def model_post_init(self, __context) -> None:
         """Post-init: resolve gitea_token from file if not set via env."""
         if not self.gitea_token:
-            token_path = os.path.expanduser("~/.config/gitea/token")
-            try:
-                if os.path.isfile(token_path):
-                    token = open(token_path).read().strip()  # noqa: SIM115
-                    if token:
-                        self.gitea_token = token
-            except OSError:
-                pass
+            # Priority: Timmy's own token → legacy admin token
+            repo_root = self._compute_repo_root()
+            timmy_token_path = os.path.join(repo_root, ".timmy_gitea_token")
+            legacy_token_path = os.path.expanduser("~/.config/gitea/token")
+            for token_path in (timmy_token_path, legacy_token_path):
+                try:
+                    if os.path.isfile(token_path):
+                        token = open(token_path).read().strip()  # noqa: SIM115
+                        if token:
+                            self.gitea_token = token
+                            break
+                except OSError:
+                    pass
 
     model_config = SettingsConfigDict(
         env_file=".env",
@@ -375,7 +403,7 @@ def check_ollama_model_available(model_name: str) -> bool:
         import json
         import urllib.request
 
-        url = settings.ollama_url.replace("localhost", "127.0.0.1")
+        url = settings.normalized_ollama_url
         req = urllib.request.Request(
             f"{url}/api/tags",
             method="GET",
@@ -388,7 +416,8 @@ def check_ollama_model_available(model_name: str) -> bool:
                 model_name == m or model_name == m.split(":")[0] or m.startswith(model_name)
                 for m in models
             )
-    except Exception:
+    except (OSError, ValueError) as exc:
+        _startup_logger.debug("Ollama model check failed: %s", exc)
         return False
 
 
@@ -451,8 +480,19 @@ def validate_startup(*, force: bool = False) -> None:
                 ", ".join(_missing),
             )
             sys.exit(1)
+        if "*" in settings.cors_origins:
+            _startup_logger.error(
+                "PRODUCTION SECURITY ERROR: CORS wildcard '*' is not allowed "
+                "in production. Set CORS_ORIGINS to explicit origins."
+            )
+            sys.exit(1)
         _startup_logger.info("Production mode: security secrets validated ✓")
     else:
+        if "*" in settings.cors_origins:
+            _startup_logger.warning(
+                "SEC: CORS_ORIGINS contains wildcard '*' — "
+                "restrict to explicit origins before deploying to production."
+            )
         if not settings.l402_hmac_secret:
             _startup_logger.warning(
                 "SEC: L402_HMAC_SECRET is not set — "

src/dashboard/app.py

@@ -8,6 +8,7 @@ Key improvements:
 """
 
 import asyncio
+import json
 import logging
 from contextlib import asynccontextmanager
 from pathlib import Path
@@ -28,6 +29,7 @@ from dashboard.routes.agents import router as agents_router
 from dashboard.routes.briefing import router as briefing_router
 from dashboard.routes.calm import router as calm_router
 from dashboard.routes.chat_api import router as chat_api_router
+from dashboard.routes.chat_api_v1 import router as chat_api_v1_router
 from dashboard.routes.db_explorer import router as db_explorer_router
 from dashboard.routes.discord import router as discord_router
 from dashboard.routes.experiments import router as experiments_router
@@ -46,6 +48,8 @@ from dashboard.routes.thinking import router as thinking_router
 from dashboard.routes.tools import router as tools_router
 from dashboard.routes.voice import router as voice_router
 from dashboard.routes.work_orders import router as work_orders_router
+from dashboard.routes.world import router as world_router
+from timmy.workshop_state import PRESENCE_FILE
 
 
 class _ColorFormatter(logging.Formatter):
@@ -151,7 +155,17 @@ async def _thinking_scheduler() -> None:
     while True:
         try:
             if settings.thinking_enabled:
-                await thinking_engine.think_once()
+                await asyncio.wait_for(
+                    thinking_engine.think_once(),
+                    timeout=settings.thinking_timeout_seconds,
+                )
+        except TimeoutError:
+            logger.warning(
+                "Thinking cycle timed out after %ds — Ollama may be unresponsive",
+                settings.thinking_timeout_seconds,
+            )
+        except asyncio.CancelledError:
+            raise
         except Exception as exc:
             logger.error("Thinking scheduler error: %s", exc)
 
@@ -171,7 +185,10 @@ async def _loop_qa_scheduler() -> None:
     while True:
         try:
             if settings.loop_qa_enabled:
-                result = await loop_qa_orchestrator.run_next_test()
+                result = await asyncio.wait_for(
+                    loop_qa_orchestrator.run_next_test(),
+                    timeout=settings.thinking_timeout_seconds,
+                )
                 if result:
                     status = "PASS" if result["success"] else "FAIL"
                     logger.info(
@@ -180,6 +197,13 @@ async def _loop_qa_scheduler() -> None:
                         status,
                         result.get("details", "")[:80],
                     )
+        except TimeoutError:
+            logger.warning(
+                "Loop QA test timed out after %ds",
+                settings.thinking_timeout_seconds,
+            )
+        except asyncio.CancelledError:
+            raise
         except Exception as exc:
             logger.error("Loop QA scheduler error: %s", exc)
 
@@ -187,6 +211,54 @@ async def _loop_qa_scheduler() -> None:
         await asyncio.sleep(interval)
 
 
+_PRESENCE_POLL_SECONDS = 30
+_PRESENCE_INITIAL_DELAY = 3
+
+_SYNTHESIZED_STATE: dict = {
+    "version": 1,
+    "liveness": None,
+    "current_focus": "",
+    "mood": "idle",
+    "active_threads": [],
+    "recent_events": [],
+    "concerns": [],
+}
+
+
+async def _presence_watcher() -> None:
+    """Background task: watch ~/.timmy/presence.json and broadcast changes via WS.
+
+    Polls the file every 30 seconds (matching Timmy's write cadence).
+    If the file doesn't exist, broadcasts a synthesised idle state.
+    """
+    from infrastructure.ws_manager.handler import ws_manager as ws_mgr
+
+    await asyncio.sleep(_PRESENCE_INITIAL_DELAY)  # Stagger after other schedulers
+
+    last_mtime: float = 0.0
+
+    while True:
+        try:
+            if PRESENCE_FILE.exists():
+                mtime = PRESENCE_FILE.stat().st_mtime
+                if mtime != last_mtime:
+                    last_mtime = mtime
+                    raw = await asyncio.to_thread(PRESENCE_FILE.read_text)
+                    state = json.loads(raw)
+                    await ws_mgr.broadcast("timmy_state", state)
+            else:
+                # File absent — broadcast synthesised state once per cycle
+                if last_mtime != -1.0:
+                    last_mtime = -1.0
+                    await ws_mgr.broadcast("timmy_state", _SYNTHESIZED_STATE)
+        except json.JSONDecodeError as exc:
+            logger.warning("presence.json parse error: %s", exc)
+        except Exception as exc:
+            logger.warning("Presence watcher error: %s", exc)
+
+        await asyncio.sleep(_PRESENCE_POLL_SECONDS)
+
+
 async def _start_chat_integrations_background() -> None:
     """Background task: start chat integrations without blocking startup."""
     from integrations.chat_bridge.registry import platform_registry
@@ -277,35 +349,38 @@ async def _discord_token_watcher() -> None:
                 logger.warning("Discord auto-start failed: %s", exc)
 
 
-@asynccontextmanager
-async def lifespan(app: FastAPI):
-    """Application lifespan manager with non-blocking startup."""
-
-    # Validate security config (no-op in test mode)
+def _startup_init() -> None:
+    """Validate config and enable event persistence."""
     from config import validate_startup
 
     validate_startup()
 
-    # Enable event persistence (unified EventBus + swarm event_log)
     from infrastructure.events.bus import init_event_bus_persistence
 
     init_event_bus_persistence()
 
-    # Create all background tasks without waiting for them
-    briefing_task = asyncio.create_task(_briefing_scheduler())
-    thinking_task = asyncio.create_task(_thinking_scheduler())
-    loop_qa_task = asyncio.create_task(_loop_qa_scheduler())
-
-    # Initialize Spark Intelligence engine
     from spark.engine import get_spark_engine
 
     if get_spark_engine().enabled:
         logger.info("Spark Intelligence active — event capture enabled")
 
-    # Auto-prune old vector store memories on startup
+
+def _startup_background_tasks() -> list[asyncio.Task]:
+    """Spawn all recurring background tasks (non-blocking)."""
+    return [
+        asyncio.create_task(_briefing_scheduler()),
+        asyncio.create_task(_thinking_scheduler()),
+        asyncio.create_task(_loop_qa_scheduler()),
+        asyncio.create_task(_presence_watcher()),
+        asyncio.create_task(_start_chat_integrations_background()),
+    ]
+
+
+def _startup_pruning() -> None:
+    """Auto-prune old memories, thoughts, and events on startup."""
     if settings.memory_prune_days > 0:
         try:
-            from timmy.memory.vector_store import prune_memories
+            from timmy.memory_system import prune_memories
 
             pruned = prune_memories(
                 older_than_days=settings.memory_prune_days,
@@ -320,7 +395,6 @@ async def lifespan(app: FastAPI):
         except Exception as exc:
             logger.debug("Memory auto-prune skipped: %s", exc)
 
-    # Auto-prune old thoughts on startup
     if settings.thoughts_prune_days > 0:
         try:
             from timmy.thinking import thinking_engine
@@ -338,7 +412,6 @@ async def lifespan(app: FastAPI):
         except Exception as exc:
             logger.debug("Thought auto-prune skipped: %s", exc)
 
-    # Auto-prune old system events on startup
     if settings.events_prune_days > 0:
         try:
             from swarm.event_log import prune_old_events
@@ -356,7 +429,6 @@ async def lifespan(app: FastAPI):
         except Exception as exc:
             logger.debug("Event auto-prune skipped: %s", exc)
 
-    # Warn if memory vault exceeds size limit
     if settings.memory_vault_max_mb > 0:
         try:
             vault_path = Path(settings.repo_root) / "memory" / "notes"
@@ -372,21 +444,18 @@ async def lifespan(app: FastAPI):
         except Exception as exc:
             logger.debug("Vault size check skipped: %s", exc)
 
-    # Start chat integrations in background
-    chat_task = asyncio.create_task(_start_chat_integrations_background())
 
-    logger.info("✓ Dashboard ready for requests")
-
-    yield
-
-    # Cleanup on shutdown
+async def _shutdown_cleanup(
+    bg_tasks: list[asyncio.Task],
+    workshop_heartbeat,
+) -> None:
+    """Stop chat bots, MCP sessions, heartbeat, and cancel background tasks."""
     from integrations.chat_bridge.vendors.discord import discord_bot
     from integrations.telegram_bot.bot import telegram_bot
 
     await discord_bot.stop()
     await telegram_bot.stop()
 
-    # Close MCP tool server sessions
     try:
         from timmy.mcp_tools import close_mcp_sessions
 
@@ -394,13 +463,44 @@ async def lifespan(app: FastAPI):
     except Exception as exc:
         logger.debug("MCP shutdown: %s", exc)
 
-    for task in [briefing_task, thinking_task, chat_task, loop_qa_task]:
-        if task:
-            task.cancel()
-            try:
-                await task
-            except asyncio.CancelledError:
-                pass
+    await workshop_heartbeat.stop()
+
+    for task in bg_tasks:
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    """Application lifespan manager with non-blocking startup."""
+    _startup_init()
+    bg_tasks = _startup_background_tasks()
+    _startup_pruning()
+
+    # Start Workshop presence heartbeat with WS relay
+    from dashboard.routes.world import broadcast_world_state
+    from timmy.workshop_state import WorkshopHeartbeat
+
+    workshop_heartbeat = WorkshopHeartbeat(on_change=broadcast_world_state)
+    await workshop_heartbeat.start()
+
+    # Register session logger with error capture
+    try:
+        from infrastructure.error_capture import register_error_recorder
+        from timmy.session_logger import get_session_logger
+
+        register_error_recorder(get_session_logger().record_error)
+    except Exception:
+        logger.debug("Failed to register error recorder")
+
+    logger.info("✓ Dashboard ready for requests")
+
+    yield
+
+    await _shutdown_cleanup(bg_tasks, workshop_heartbeat)
 
 
 app = FastAPI(
@@ -413,15 +513,14 @@ app = FastAPI(
 
 
 def _get_cors_origins() -> list[str]:
-    """Get CORS origins from settings, with sensible defaults."""
+    """Get CORS origins from settings, rejecting wildcards in production."""
     origins = settings.cors_origins
-    if settings.debug and origins == ["*"]:
-        return [
-            "http://localhost:3000",
-            "http://localhost:8000",
-            "http://127.0.0.1:3000",
-            "http://127.0.0.1:8000",
-        ]
+    if "*" in origins and not settings.debug:
+        logger.warning(
+            "Wildcard '*' in CORS_ORIGINS stripped in production — "
+            "set explicit origins via CORS_ORIGINS env var"
+        )
+        origins = [o for o in origins if o != "*"]
     return origins
 
 
@@ -474,6 +573,7 @@ app.include_router(grok_router)
 app.include_router(models_router)
 app.include_router(models_api_router)
 app.include_router(chat_api_router)
+app.include_router(chat_api_v1_router)
 app.include_router(thinking_router)
 app.include_router(calm_router)
 app.include_router(tasks_router)
@@ -482,6 +582,7 @@ app.include_router(loop_qa_router)
 app.include_router(system_router)
 app.include_router(experiments_router)
 app.include_router(db_explorer_router)
+app.include_router(world_router)
 
 
 @app.websocket("/ws")
@@ -510,7 +611,8 @@ async def swarm_live(websocket: WebSocket):
         while True:
             # Keep connection alive; events are pushed via ws_mgr.broadcast()
             await websocket.receive_text()
-    except Exception:
+    except Exception as exc:
+        logger.debug("WebSocket disconnect error: %s", exc)
         ws_mgr.disconnect(websocket)
 
 
@@ -532,7 +634,8 @@ async def swarm_agents_sidebar():
                 f"</div>"
             )
         return "\n".join(lines) if lines else '<div class="mc-muted">No agents configured</div>'
-    except Exception:
+    except Exception as exc:
+        logger.debug("Agents sidebar error: %s", exc)
         return '<div class="mc-muted">Agents unavailable</div>'
 
 

src/dashboard/middleware/csrf.py

@@ -5,6 +5,7 @@ to protect state-changing endpoints from cross-site request attacks.
 """
 
 import hmac
+import logging
 import secrets
 from collections.abc import Callable
 from functools import wraps
@@ -16,6 +17,8 @@ from starlette.responses import JSONResponse, Response
 # Module-level set to track exempt routes
 _exempt_routes: set[str] = set()
 
+logger = logging.getLogger(__name__)
+
 
 def csrf_exempt(endpoint: Callable) -> Callable:
     """Decorator to mark an endpoint as exempt from CSRF validation.
@@ -97,7 +100,7 @@ class CSRFMiddleware(BaseHTTPMiddleware):
             ...
 
     Usage:
-        app.add_middleware(CSRFMiddleware, secret="your-secret-key")
+        app.add_middleware(CSRFMiddleware, secret=settings.csrf_secret)
 
     Attributes:
         secret: Secret key for token signing (optional, for future use).
@@ -278,7 +281,8 @@ class CSRFMiddleware(BaseHTTPMiddleware):
                 form_token = form_data.get(self.form_field)
                 if form_token and validate_csrf_token(str(form_token), csrf_cookie):
                     return True
-            except Exception:
+            except Exception as exc:
+                logger.debug("CSRF form parsing error: %s", exc)
                 # Error parsing form data, treat as invalid
                 pass
 

src/dashboard/middleware/request_logging.py

@@ -115,7 +115,8 @@ class RequestLoggingMiddleware(BaseHTTPMiddleware):
                         "duration_ms": f"{duration_ms:.0f}",
                     },
                 )
-            except Exception:
+            except Exception as exc:
+                logger.debug("Escalation logging error: %s", exc)
                 pass  # never let escalation break the request
 
             # Re-raise the exception

src/dashboard/middleware/security_headers.py

@@ -4,10 +4,14 @@ Adds common security headers to all HTTP responses to improve
 application security posture against various attacks.
 """
 
+import logging
+
 from starlette.middleware.base import BaseHTTPMiddleware
 from starlette.requests import Request
 from starlette.responses import Response
 
+logger = logging.getLogger(__name__)
+
 
 class SecurityHeadersMiddleware(BaseHTTPMiddleware):
     """Middleware to add security headers to all responses.
@@ -130,12 +134,8 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
         """
         try:
             response = await call_next(request)
-        except Exception:
-            import logging
-
-            logging.getLogger(__name__).debug(
-                "Upstream error in security headers middleware", exc_info=True
-            )
+        except Exception as exc:
+            logger.debug("Upstream error in security headers middleware: %s", exc)
             from starlette.responses import PlainTextResponse
 
             response = PlainTextResponse("Internal Server Error", status_code=500)

src/dashboard/routes/agents.py

@@ -12,6 +12,7 @@ from timmy.tool_safety import (
     format_action_description,
     get_impact_level,
 )
+from timmy.welcome import WELCOME_MESSAGE
 
 logger = logging.getLogger(__name__)
 
@@ -56,7 +57,7 @@ async def get_history(request: Request):
     return templates.TemplateResponse(
         request,
         "partials/history.html",
-        {"messages": message_log.all()},
+        {"messages": message_log.all(), "welcome_message": WELCOME_MESSAGE},
     )
 
 
@@ -66,23 +67,91 @@ async def clear_history(request: Request):
     return templates.TemplateResponse(
         request,
         "partials/history.html",
-        {"messages": []},
+        {"messages": [], "welcome_message": WELCOME_MESSAGE},
     )
 
 
+def _validate_message(message: str) -> str:
+    """Strip and validate chat input; raise HTTPException on bad input."""
+    from fastapi import HTTPException
+
+    message = message.strip()
+    if not message:
+        raise HTTPException(status_code=400, detail="Message cannot be empty")
+    if len(message) > MAX_MESSAGE_LENGTH:
+        raise HTTPException(status_code=422, detail="Message too long")
+    return message
+
+
+def _record_user_activity() -> None:
+    """Notify the thinking engine that the user is active."""
+    try:
+        from timmy.thinking import thinking_engine
+
+        thinking_engine.record_user_input()
+    except Exception:
+        logger.debug("Failed to record user input for thinking engine")
+
+
+def _extract_tool_actions(run_output) -> list[dict]:
+    """If Agno paused the run for tool confirmation, build approval items."""
+    from timmy.approvals import create_item
+
+    tool_actions: list[dict] = []
+    status = getattr(run_output, "status", None)
+    is_paused = status == "PAUSED" or str(status) == "RunStatus.paused"
+
+    if not (is_paused and getattr(run_output, "active_requirements", None)):
+        return tool_actions
+
+    for req in run_output.active_requirements:
+        if not getattr(req, "needs_confirmation", False):
+            continue
+        te = req.tool_execution
+        tool_name = getattr(te, "tool_name", "unknown")
+        tool_args = getattr(te, "tool_args", {}) or {}
+
+        item = create_item(
+            title=f"Dashboard: {tool_name}",
+            description=format_action_description(tool_name, tool_args),
+            proposed_action=json.dumps({"tool": tool_name, "args": tool_args}),
+            impact=get_impact_level(tool_name),
+        )
+        _pending_runs[item.id] = {
+            "run_output": run_output,
+            "requirement": req,
+            "tool_name": tool_name,
+            "tool_args": tool_args,
+        }
+        tool_actions.append(
+            {
+                "approval_id": item.id,
+                "tool_name": tool_name,
+                "description": format_action_description(tool_name, tool_args),
+                "impact": get_impact_level(tool_name),
+            }
+        )
+    return tool_actions
+
+
+def _log_exchange(
+    message: str, response_text: str | None, error_text: str | None, timestamp: str
+) -> None:
+    """Append user message and agent/error reply to the in-memory log."""
+    message_log.append(role="user", content=message, timestamp=timestamp, source="browser")
+    if response_text:
+        message_log.append(
+            role="agent", content=response_text, timestamp=timestamp, source="browser"
+        )
+    elif error_text:
+        message_log.append(role="error", content=error_text, timestamp=timestamp, source="browser")
+
+
 @router.post("/default/chat", response_class=HTMLResponse)
 async def chat_agent(request: Request, message: str = Form(...)):
     """Chat — synchronous response with native Agno tool confirmation."""
-    message = message.strip()
-    if not message:
-        from fastapi import HTTPException
-
-        raise HTTPException(status_code=400, detail="Message cannot be empty")
-
-    if len(message) > MAX_MESSAGE_LENGTH:
-        from fastapi import HTTPException
-
-        raise HTTPException(status_code=422, detail="Message too long")
+    message = _validate_message(message)
+    _record_user_activity()
 
     timestamp = datetime.now().strftime("%H:%M:%S")
     response_text = None
@@ -95,54 +164,15 @@ async def chat_agent(request: Request, message: str = Form(...)):
         error_text = f"Chat error: {exc}"
         run_output = None
 
-    # Check if Agno paused the run for tool confirmation
-    tool_actions = []
+    tool_actions: list[dict] = []
     if run_output is not None:
-        status = getattr(run_output, "status", None)
-        is_paused = status == "PAUSED" or str(status) == "RunStatus.paused"
-
-        if is_paused and getattr(run_output, "active_requirements", None):
-            for req in run_output.active_requirements:
-                if getattr(req, "needs_confirmation", False):
-                    te = req.tool_execution
-                    tool_name = getattr(te, "tool_name", "unknown")
-                    tool_args = getattr(te, "tool_args", {}) or {}
-
-                    from timmy.approvals import create_item
-
-                    item = create_item(
-                        title=f"Dashboard: {tool_name}",
-                        description=format_action_description(tool_name, tool_args),
-                        proposed_action=json.dumps({"tool": tool_name, "args": tool_args}),
-                        impact=get_impact_level(tool_name),
-                    )
-                    _pending_runs[item.id] = {
-                        "run_output": run_output,
-                        "requirement": req,
-                        "tool_name": tool_name,
-                        "tool_args": tool_args,
-                    }
-                    tool_actions.append(
-                        {
-                            "approval_id": item.id,
-                            "tool_name": tool_name,
-                            "description": format_action_description(tool_name, tool_args),
-                            "impact": get_impact_level(tool_name),
-                        }
-                    )
-
+        tool_actions = _extract_tool_actions(run_output)
         raw_content = run_output.content if hasattr(run_output, "content") else ""
         response_text = _clean_response(raw_content or "")
         if not response_text and not tool_actions:
-            response_text = None  # let error template show if needed
+            response_text = None
 
-    message_log.append(role="user", content=message, timestamp=timestamp, source="browser")
-    if response_text:
-        message_log.append(
-            role="agent", content=response_text, timestamp=timestamp, source="browser"
-        )
-    elif error_text:
-        message_log.append(role="error", content=error_text, timestamp=timestamp, source="browser")
+    _log_exchange(message, response_text, error_text, timestamp)
 
     return templates.TemplateResponse(
         request,
@@ -220,7 +250,8 @@ async def reject_tool(request: Request, approval_id: str):
         # Resume so the agent knows the tool was rejected
         try:
             await continue_chat(pending["run_output"])
-        except Exception:
+        except Exception as exc:
+            logger.warning("Agent tool rejection error: %s", exc)
             pass
 
     reject(approval_id)

src/dashboard/routes/briefing.py

@@ -27,7 +27,8 @@ async def get_briefing(request: Request):
     """Return today's briefing page (generated or cached)."""
     try:
         briefing = briefing_engine.get_or_generate()
-    except Exception:
+    except Exception as exc:
+        logger.debug("Briefing generation failed: %s", exc)
         logger.exception("Briefing generation failed")
         now = datetime.now(UTC)
         briefing = Briefing(

src/dashboard/routes/chat_api.py

@@ -31,6 +31,93 @@ _UPLOAD_DIR = str(Path(settings.repo_root) / "data" / "chat-uploads")
 _MAX_UPLOAD_SIZE = 50 * 1024 * 1024  # 50 MB
 
 
+# ── POST /api/chat — helpers ─────────────────────────────────────────────────
+
+
+async def _parse_chat_body(request: Request) -> tuple[dict | None, JSONResponse | None]:
+    """Parse and validate the JSON request body.
+
+    Returns (body, None) on success or (None, error_response) on failure.
+    """
+    content_length = request.headers.get("content-length")
+    if content_length and int(content_length) > settings.chat_api_max_body_bytes:
+        return None, JSONResponse(status_code=413, content={"error": "Request body too large"})
+
+    try:
+        body = await request.json()
+    except Exception as exc:
+        logger.warning("Chat API JSON parse error: %s", exc)
+        return None, JSONResponse(status_code=400, content={"error": "Invalid JSON"})
+
+    messages = body.get("messages")
+    if not messages or not isinstance(messages, list):
+        return None, JSONResponse(status_code=400, content={"error": "messages array is required"})
+
+    return body, None
+
+
+def _extract_user_message(messages: list[dict]) -> str | None:
+    """Return the text of the last user message, or *None* if absent."""
+    for msg in reversed(messages):
+        if msg.get("role") == "user":
+            content = msg.get("content", "")
+            if isinstance(content, list):
+                text_parts = [
+                    p.get("text", "")
+                    for p in content
+                    if isinstance(p, dict) and p.get("type") == "text"
+                ]
+                return " ".join(text_parts).strip() or None
+            text = str(content).strip()
+            return text or None
+    return None
+
+
+def _build_context_prefix() -> str:
+    """Build the system-context preamble injected before the user message."""
+    now = datetime.now()
+    return (
+        f"[System: Current date/time is "
+        f"{now.strftime('%A, %B %d, %Y at %I:%M %p')}]\n"
+        f"[System: Mobile client]\n\n"
+    )
+
+
+def _notify_thinking_engine() -> None:
+    """Record user activity so the thinking engine knows we're not idle."""
+    try:
+        from timmy.thinking import thinking_engine
+
+        thinking_engine.record_user_input()
+    except Exception:
+        logger.debug("Failed to record user input for thinking engine")
+
+
+async def _process_chat(user_msg: str) -> dict | JSONResponse:
+    """Send *user_msg* to the agent, log the exchange, and return a response."""
+    _notify_thinking_engine()
+    timestamp = datetime.now().strftime("%H:%M:%S")
+
+    try:
+        response_text = await agent_chat(
+            _build_context_prefix() + user_msg,
+            session_id="mobile",
+        )
+        message_log.append(role="user", content=user_msg, timestamp=timestamp, source="api")
+        message_log.append(role="agent", content=response_text, timestamp=timestamp, source="api")
+        return {"reply": response_text, "timestamp": timestamp}
+
+    except Exception as exc:
+        error_msg = f"Agent is offline: {exc}"
+        logger.error("api_chat error: %s", exc)
+        message_log.append(role="user", content=user_msg, timestamp=timestamp, source="api")
+        message_log.append(role="error", content=error_msg, timestamp=timestamp, source="api")
+        return JSONResponse(
+            status_code=503,
+            content={"error": error_msg, "timestamp": timestamp},
+        )
+
+
 # ── POST /api/chat ────────────────────────────────────────────────────────────
 
 
@@ -44,69 +131,15 @@ async def api_chat(request: Request):
     Response:
         {"reply": "...", "timestamp": "HH:MM:SS"}
     """
-    # Enforce request body size limit
-    content_length = request.headers.get("content-length")
-    if content_length and int(content_length) > settings.chat_api_max_body_bytes:
-        return JSONResponse(status_code=413, content={"error": "Request body too large"})
+    body, err = await _parse_chat_body(request)
+    if err:
+        return err
 
-    try:
-        body = await request.json()
-    except Exception:
-        return JSONResponse(status_code=400, content={"error": "Invalid JSON"})
-
-    messages = body.get("messages")
-    if not messages or not isinstance(messages, list):
-        return JSONResponse(status_code=400, content={"error": "messages array is required"})
-
-    # Extract the latest user message text
-    last_user_msg = None
-    for msg in reversed(messages):
-        if msg.get("role") == "user":
-            content = msg.get("content", "")
-            # Handle multimodal content arrays — extract text parts
-            if isinstance(content, list):
-                text_parts = [
-                    p.get("text", "")
-                    for p in content
-                    if isinstance(p, dict) and p.get("type") == "text"
-                ]
-                last_user_msg = " ".join(text_parts).strip()
-            else:
-                last_user_msg = str(content).strip()
-            break
-
-    if not last_user_msg:
+    user_msg = _extract_user_message(body["messages"])
+    if not user_msg:
         return JSONResponse(status_code=400, content={"error": "No user message found"})
 
-    timestamp = datetime.now().strftime("%H:%M:%S")
-
-    try:
-        # Inject context (same pattern as the HTMX chat handler in agents.py)
-        now = datetime.now()
-        context_prefix = (
-            f"[System: Current date/time is "
-            f"{now.strftime('%A, %B %d, %Y at %I:%M %p')}]\n"
-            f"[System: Mobile client]\n\n"
-        )
-        response_text = await agent_chat(
-            context_prefix + last_user_msg,
-            session_id="mobile",
-        )
-
-        message_log.append(role="user", content=last_user_msg, timestamp=timestamp, source="api")
-        message_log.append(role="agent", content=response_text, timestamp=timestamp, source="api")
-
-        return {"reply": response_text, "timestamp": timestamp}
-
-    except Exception as exc:
-        error_msg = f"Agent is offline: {exc}"
-        logger.error("api_chat error: %s", exc)
-        message_log.append(role="user", content=last_user_msg, timestamp=timestamp, source="api")
-        message_log.append(role="error", content=error_msg, timestamp=timestamp, source="api")
-        return JSONResponse(
-            status_code=503,
-            content={"error": error_msg, "timestamp": timestamp},
-        )
+    return await _process_chat(user_msg)
 
 
 # ── POST /api/upload ──────────────────────────────────────────────────────────

src/dashboard/routes/chat_api_v1.py

@@ -0,0 +1,198 @@
+"""Version 1 (v1) JSON REST API for the Timmy Time iPad app.
+
+This module implements the specific endpoints required by the native
+iPad app as defined in the project specification.
+
+Endpoints:
+    POST /api/v1/chat           — Streaming SSE chat response
+    GET  /api/v1/chat/history   — Retrieve chat history with limit
+    POST /api/v1/upload         — Multipart file upload with auto-detection
+    GET  /api/v1/status         — Detailed system and model status
+"""
+
+import json
+import logging
+import os
+import uuid
+from datetime import UTC, datetime
+from pathlib import Path
+
+from fastapi import APIRouter, File, HTTPException, Query, Request, UploadFile
+from fastapi.responses import JSONResponse, StreamingResponse
+
+from config import APP_START_TIME, settings
+from dashboard.routes.health import _check_ollama
+from dashboard.store import message_log
+from timmy.session import _get_agent
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/api/v1", tags=["chat-api-v1"])
+
+_UPLOAD_DIR = str(Path(settings.repo_root) / "data" / "chat-uploads")
+_MAX_UPLOAD_SIZE = 50 * 1024 * 1024  # 50 MB
+
+
+# ── POST /api/v1/chat ─────────────────────────────────────────────────────────
+
+
+@router.post("/chat")
+async def api_v1_chat(request: Request):
+    """Accept a JSON chat payload and return a streaming SSE response.
+
+    Request body:
+        {
+            "message": "string",
+            "session_id": "string",
+            "attachments": ["id1", "id2"]
+        }
+
+    Response:
+        text/event-stream (SSE)
+    """
+    try:
+        body = await request.json()
+    except Exception as exc:
+        logger.warning("Chat v1 API JSON parse error: %s", exc)
+        return JSONResponse(status_code=400, content={"error": "Invalid JSON"})
+
+    message = body.get("message")
+    session_id = body.get("session_id", "ipad-app")
+    attachments = body.get("attachments", [])
+
+    if not message:
+        return JSONResponse(status_code=400, content={"error": "message is required"})
+
+    # Prepare context for the agent
+    context_prefix = (
+        f"[System: Current date/time is "
+        f"{datetime.now().strftime('%A, %B %d, %Y at %I:%M %p')}]\n"
+        f"[System: iPad App client]\n"
+    )
+
+    if attachments:
+        context_prefix += f"[System: Attachments: {', '.join(attachments)}]\n"
+
+    context_prefix += "\n"
+    full_prompt = context_prefix + message
+
+    async def event_generator():
+        try:
+            agent = _get_agent()
+            # Using streaming mode for SSE
+            async for chunk in agent.arun(full_prompt, stream=True, session_id=session_id):
+                # Agno chunks can be strings or RunOutput
+                content = chunk.content if hasattr(chunk, "content") else str(chunk)
+                if content:
+                    yield f"data: {json.dumps({'text': content})}\n\n"
+
+            yield "data: [DONE]\n\n"
+        except Exception as exc:
+            logger.error("SSE stream error: %s", exc)
+            yield f"data: {json.dumps({'error': str(exc)})}\n\n"
+
+    return StreamingResponse(event_generator(), media_type="text/event-stream")
+
+
+# ── GET /api/v1/chat/history ──────────────────────────────────────────────────
+
+
+@router.get("/chat/history")
+async def api_v1_chat_history(
+    session_id: str = Query("ipad-app"), limit: int = Query(50, ge=1, le=100)
+):
+    """Return recent chat history for a specific session."""
+    # Filter and limit the message log
+    # Note: message_log.all() returns all messages; we filter by source or just return last N
+    all_msgs = message_log.all()
+
+    # In a real implementation, we'd filter by session_id if message_log supported it.
+    # For now, we return the last 'limit' messages.
+    history = [
+        {
+            "role": msg.role,
+            "content": msg.content,
+            "timestamp": msg.timestamp,
+            "source": msg.source,
+        }
+        for msg in all_msgs[-limit:]
+    ]
+
+    return {"messages": history}
+
+
+# ── POST /api/v1/upload ───────────────────────────────────────────────────────
+
+
+@router.post("/upload")
+async def api_v1_upload(file: UploadFile = File(...)):
+    """Accept a file upload, auto-detect type, and return metadata.
+
+    Response:
+        {
+            "id": "string",
+            "type": "image|audio|document|url",
+            "summary": "string",
+            "metadata": {...}
+        }
+    """
+    os.makedirs(_UPLOAD_DIR, exist_ok=True)
+
+    file_id = uuid.uuid4().hex[:12]
+    safe_name = os.path.basename(file.filename or "upload")
+    stored_name = f"{file_id}-{safe_name}"
+    file_path = os.path.join(_UPLOAD_DIR, stored_name)
+
+    # Verify resolved path stays within upload directory
+    resolved = Path(file_path).resolve()
+    upload_root = Path(_UPLOAD_DIR).resolve()
+    if not str(resolved).startswith(str(upload_root)):
+        raise HTTPException(status_code=400, detail="Invalid file name")
+
+    contents = await file.read()
+    if len(contents) > _MAX_UPLOAD_SIZE:
+        raise HTTPException(status_code=413, detail="File too large (max 50 MB)")
+
+    with open(file_path, "wb") as f:
+        f.write(contents)
+
+    # Auto-detect type based on extension/mime
+    mime_type = file.content_type or "application/octet-stream"
+    ext = os.path.splitext(safe_name)[1].lower()
+
+    media_type = "document"
+    if mime_type.startswith("image/") or ext in [".jpg", ".jpeg", ".png", ".heic"]:
+        media_type = "image"
+    elif mime_type.startswith("audio/") or ext in [".m4a", ".mp3", ".wav", ".caf"]:
+        media_type = "audio"
+    elif ext in [".pdf", ".txt", ".md"]:
+        media_type = "document"
+
+    # Placeholder for actual processing (OCR, Whisper, etc.)
+    summary = f"Uploaded {media_type}: {safe_name}"
+
+    return {
+        "id": file_id,
+        "type": media_type,
+        "summary": summary,
+        "url": f"/uploads/{stored_name}",
+        "metadata": {"fileName": safe_name, "mimeType": mime_type, "size": len(contents)},
+    }
+
+
+# ── GET /api/v1/status ────────────────────────────────────────────────────────
+
+
+@router.get("/status")
+async def api_v1_status():
+    """Detailed system and model status."""
+    ollama_status = await _check_ollama()
+    uptime = (datetime.now(UTC) - APP_START_TIME).total_seconds()
+
+    return {
+        "timmy": "online" if ollama_status.status == "healthy" else "offline",
+        "model": settings.ollama_model,
+        "ollama": "running" if ollama_status.status == "healthy" else "stopped",
+        "uptime": f"{int(uptime // 3600)}h {int((uptime % 3600) // 60)}m",
+        "version": "2.0.0-v1-api",
+    }

src/dashboard/routes/db_explorer.py

@@ -3,6 +3,7 @@
 import asyncio
 import logging
 import sqlite3
+from contextlib import closing
 from pathlib import Path
 
 from fastapi import APIRouter, Request
@@ -39,56 +40,50 @@ def _query_database(db_path: str) -> dict:
     """Open a database read-only and return all tables with their rows."""
     result = {"tables": {}, "error": None}
     try:
-        conn = sqlite3.connect(f"file:{db_path}?mode=ro", uri=True)
-        conn.row_factory = sqlite3.Row
-    except Exception as exc:
-        result["error"] = str(exc)
-        return result
+        with closing(sqlite3.connect(f"file:{db_path}?mode=ro", uri=True)) as conn:
+            conn.row_factory = sqlite3.Row
 
-    try:
-        tables = conn.execute(
-            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY name"
-        ).fetchall()
-        for (table_name,) in tables:
-            try:
-                rows = conn.execute(
-                    f"SELECT * FROM [{table_name}] LIMIT {MAX_ROWS}"  # noqa: S608
-                ).fetchall()
-                columns = (
-                    [
-                        desc[0]
-                        for desc in conn.execute(
-                            f"SELECT * FROM [{table_name}] LIMIT 0"
-                        ).description
-                    ]
-                    if rows
-                    else []
-                )  # noqa: S608
-                if not columns and rows:
-                    columns = list(rows[0].keys())
-                elif not columns:
-                    # Get columns even for empty tables
-                    cursor = conn.execute(f"PRAGMA table_info([{table_name}])")  # noqa: S608
-                    columns = [r[1] for r in cursor.fetchall()]
-                count = conn.execute(f"SELECT COUNT(*) FROM [{table_name}]").fetchone()[0]  # noqa: S608
-                result["tables"][table_name] = {
-                    "columns": columns,
-                    "rows": [dict(r) for r in rows],
-                    "total_count": count,
-                    "truncated": count > MAX_ROWS,
-                }
-            except Exception as exc:
-                result["tables"][table_name] = {
-                    "error": str(exc),
-                    "columns": [],
-                    "rows": [],
-                    "total_count": 0,
-                    "truncated": False,
-                }
+            tables = conn.execute(
+                "SELECT name FROM sqlite_master WHERE type='table' ORDER BY name"
+            ).fetchall()
+            for (table_name,) in tables:
+                try:
+                    rows = conn.execute(
+                        f"SELECT * FROM [{table_name}] LIMIT {MAX_ROWS}"  # noqa: S608
+                    ).fetchall()
+                    columns = (
+                        [
+                            desc[0]
+                            for desc in conn.execute(
+                                f"SELECT * FROM [{table_name}] LIMIT 0"
+                            ).description
+                        ]
+                        if rows
+                        else []
+                    )  # noqa: S608
+                    if not columns and rows:
+                        columns = list(rows[0].keys())
+                    elif not columns:
+                        # Get columns even for empty tables
+                        cursor = conn.execute(f"PRAGMA table_info([{table_name}])")  # noqa: S608
+                        columns = [r[1] for r in cursor.fetchall()]
+                    count = conn.execute(f"SELECT COUNT(*) FROM [{table_name}]").fetchone()[0]  # noqa: S608
+                    result["tables"][table_name] = {
+                        "columns": columns,
+                        "rows": [dict(r) for r in rows],
+                        "total_count": count,
+                        "truncated": count > MAX_ROWS,
+                    }
+                except Exception as exc:
+                    result["tables"][table_name] = {
+                        "error": str(exc),
+                        "columns": [],
+                        "rows": [],
+                        "total_count": 0,
+                        "truncated": False,
+                    }
     except Exception as exc:
         result["error"] = str(exc)
-    finally:
-        conn.close()
 
     return result
 

src/dashboard/routes/experiments.py

@@ -30,8 +30,8 @@ async def experiments_page(request: Request):
     history = []
     try:
         history = get_experiment_history(_workspace())
-    except Exception:
-        logger.debug("Failed to load experiment history", exc_info=True)
+    except Exception as exc:
+        logger.debug("Failed to load experiment history: %s", exc)
 
     return templates.TemplateResponse(
         request,

src/dashboard/routes/grok.py

@@ -52,8 +52,8 @@ async def grok_status(request: Request):
             "estimated_cost_sats": backend.stats.estimated_cost_sats,
             "errors": backend.stats.errors,
         }
-    except Exception:
-        logger.debug("Failed to load Grok stats", exc_info=True)
+    except Exception as exc:
+        logger.warning("Failed to load Grok stats: %s", exc)
 
     return templates.TemplateResponse(
         request,
@@ -94,8 +94,8 @@ async def toggle_grok_mode(request: Request):
             tool_name="grok_mode_toggle",
             success=True,
         )
-    except Exception:
-        logger.debug("Failed to log Grok toggle to Spark", exc_info=True)
+    except Exception as exc:
+        logger.warning("Failed to log Grok toggle to Spark: %s", exc)
 
     return HTMLResponse(
         _render_toggle_card(_grok_mode_active),
@@ -128,8 +128,8 @@ def _run_grok_query(message: str) -> dict:
             sats = min(settings.grok_max_sats_per_query, 100)
             ln.create_invoice(sats, f"Grok: {message[:50]}")
             invoice_note = f" | {sats} sats"
-        except Exception:
-            logger.debug("Lightning invoice creation failed", exc_info=True)
+        except Exception as exc:
+            logger.warning("Lightning invoice creation failed: %s", exc)
 
     try:
         result = backend.run(message)

src/dashboard/routes/health.py

@@ -6,14 +6,18 @@ for the Mission Control dashboard.
 
 import asyncio
 import logging
+import sqlite3
 import time
+from contextlib import closing
 from datetime import UTC, datetime
+from pathlib import Path
 from typing import Any
 
 from fastapi import APIRouter, Request
 from fastapi.responses import HTMLResponse
 from pydantic import BaseModel
 
+from config import APP_START_TIME as _START_TIME
 from config import settings
 
 logger = logging.getLogger(__name__)
@@ -49,7 +53,6 @@ class HealthStatus(BaseModel):
 
 
 # Simple uptime tracking
-_START_TIME = datetime.now(UTC)
 
 # Ollama health cache (30-second TTL)
 _ollama_cache: DependencyStatus | None = None
@@ -62,7 +65,7 @@ def _check_ollama_sync() -> DependencyStatus:
     try:
         import urllib.request
 
-        url = settings.ollama_url.replace("localhost", "127.0.0.1")
+        url = settings.normalized_ollama_url
         req = urllib.request.Request(
             f"{url}/api/tags",
             method="GET",
@@ -76,8 +79,8 @@ def _check_ollama_sync() -> DependencyStatus:
                     sovereignty_score=10,
                     details={"url": settings.ollama_url, "model": settings.ollama_model},
                 )
-    except Exception:
-        logger.debug("Ollama health check failed", exc_info=True)
+    except Exception as exc:
+        logger.debug("Ollama health check failed: %s", exc)
 
     return DependencyStatus(
         name="Ollama AI",
@@ -101,7 +104,8 @@ async def _check_ollama() -> DependencyStatus:
 
     try:
         result = await asyncio.to_thread(_check_ollama_sync)
-    except Exception:
+    except Exception as exc:
+        logger.debug("Ollama async check failed: %s", exc)
         result = DependencyStatus(
             name="Ollama AI",
             status="unavailable",
@@ -133,13 +137,9 @@ def _check_lightning() -> DependencyStatus:
 def _check_sqlite() -> DependencyStatus:
     """Check SQLite database status."""
     try:
-        import sqlite3
-        from pathlib import Path
-
         db_path = Path(settings.repo_root) / "data" / "timmy.db"
-        conn = sqlite3.connect(str(db_path))
-        conn.execute("SELECT 1")
-        conn.close()
+        with closing(sqlite3.connect(str(db_path))) as conn:
+            conn.execute("SELECT 1")
 
         return DependencyStatus(
             name="SQLite Database",

src/dashboard/routes/memory.py

@@ -4,7 +4,7 @@ from fastapi import APIRouter, Form, HTTPException, Request
 from fastapi.responses import HTMLResponse, JSONResponse
 
 from dashboard.templating import templates
-from timmy.memory.vector_store import (
+from timmy.memory_system import (
     delete_memory,
     get_memory_stats,
     recall_personal_facts_with_ids,

src/dashboard/routes/system.py

@@ -1,10 +1,12 @@
 """System-level dashboard routes (ledger, upgrades, etc.)."""
 
 import logging
+from pathlib import Path
 
 from fastapi import APIRouter, Request
 from fastapi.responses import HTMLResponse, JSONResponse
 
+from config import settings
 from dashboard.templating import templates
 
 logger = logging.getLogger(__name__)
@@ -144,5 +146,83 @@ async def api_notifications():
                 for e in events
             ]
         )
-    except Exception:
+    except Exception as exc:
+        logger.debug("System events fetch error: %s", exc)
         return JSONResponse([])
+
+
+@router.get("/api/briefing/status", response_class=JSONResponse)
+async def api_briefing_status():
+    """Return briefing status including pending approvals and last generated time."""
+    from timmy import approvals
+    from timmy.briefing import engine as briefing_engine
+
+    pending = approvals.list_pending()
+    pending_count = len(pending)
+
+    last_generated = None
+    try:
+        cached = briefing_engine.get_cached()
+        if cached:
+            last_generated = cached.generated_at.isoformat()
+    except Exception:
+        logger.debug("Failed to read briefing cache")
+
+    return JSONResponse(
+        {
+            "status": "ok",
+            "pending_approvals": pending_count,
+            "last_generated": last_generated,
+        }
+    )
+
+
+@router.get("/api/memory/status", response_class=JSONResponse)
+async def api_memory_status():
+    """Return memory database status including file info and indexed files count."""
+    from timmy.memory_system import get_memory_stats
+
+    db_path = Path(settings.repo_root) / "data" / "memory.db"
+    db_exists = db_path.exists()
+    db_size = db_path.stat().st_size if db_exists else 0
+
+    try:
+        stats = get_memory_stats()
+        indexed_files = stats.get("total_entries", 0)
+    except Exception:
+        logger.debug("Failed to get memory stats")
+        indexed_files = 0
+
+    return JSONResponse(
+        {
+            "status": "ok",
+            "db_exists": db_exists,
+            "db_size_bytes": db_size,
+            "indexed_files": indexed_files,
+        }
+    )
+
+
+@router.get("/api/swarm/status", response_class=JSONResponse)
+async def api_swarm_status():
+    """Return swarm worker status and pending tasks count."""
+    from dashboard.routes.tasks import _get_db
+
+    pending_tasks = 0
+    try:
+        with _get_db() as db:
+            row = db.execute(
+                "SELECT COUNT(*) as cnt FROM tasks WHERE status IN ('pending_approval','approved')"
+            ).fetchone()
+            pending_tasks = row["cnt"] if row else 0
+    except Exception:
+        logger.debug("Failed to count pending tasks")
+
+    return JSONResponse(
+        {
+            "status": "ok",
+            "active_workers": 0,
+            "pending_tasks": pending_tasks,
+            "message": "Swarm monitoring endpoint",
+        }
+    )

src/dashboard/routes/tasks.py

@@ -3,6 +3,8 @@
 import logging
 import sqlite3
 import uuid
+from collections.abc import Generator
+from contextlib import closing, contextmanager
 from datetime import datetime
 from pathlib import Path
 
@@ -35,26 +37,27 @@ VALID_STATUSES = {
 VALID_PRIORITIES = {"low", "normal", "high", "urgent"}
 
 
-def _get_db() -> sqlite3.Connection:
+@contextmanager
+def _get_db() -> Generator[sqlite3.Connection, None, None]:
     DB_PATH.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(DB_PATH))
-    conn.row_factory = sqlite3.Row
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS tasks (
-            id TEXT PRIMARY KEY,
-            title TEXT NOT NULL,
-            description TEXT DEFAULT '',
-            status TEXT DEFAULT 'pending_approval',
-            priority TEXT DEFAULT 'normal',
-            assigned_to TEXT DEFAULT '',
-            created_by TEXT DEFAULT 'operator',
-            result TEXT DEFAULT '',
-            created_at TEXT DEFAULT (datetime('now')),
-            completed_at TEXT
-        )
-    """)
-    conn.commit()
-    return conn
+    with closing(sqlite3.connect(str(DB_PATH))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS tasks (
+                id TEXT PRIMARY KEY,
+                title TEXT NOT NULL,
+                description TEXT DEFAULT '',
+                status TEXT DEFAULT 'pending_approval',
+                priority TEXT DEFAULT 'normal',
+                assigned_to TEXT DEFAULT '',
+                created_by TEXT DEFAULT 'operator',
+                result TEXT DEFAULT '',
+                created_at TEXT DEFAULT (datetime('now')),
+                completed_at TEXT
+            )
+        """)
+        conn.commit()
+        yield conn
 
 
 def _row_to_dict(row: sqlite3.Row) -> dict:
@@ -101,8 +104,7 @@ class _TaskView:
 @router.get("/tasks", response_class=HTMLResponse)
 async def tasks_page(request: Request):
     """Render the main task queue page with 3-column layout."""
-    db = _get_db()
-    try:
+    with _get_db() as db:
         pending = [
             _TaskView(_row_to_dict(r))
             for r in db.execute(
@@ -121,8 +123,6 @@ async def tasks_page(request: Request):
                 "SELECT * FROM tasks WHERE status IN ('completed','vetoed','failed') ORDER BY completed_at DESC LIMIT 50"
             ).fetchall()
         ]
-    finally:
-        db.close()
 
     return templates.TemplateResponse(
         request,
@@ -145,13 +145,10 @@ async def tasks_page(request: Request):
 
 @router.get("/tasks/pending", response_class=HTMLResponse)
 async def tasks_pending(request: Request):
-    db = _get_db()
-    try:
+    with _get_db() as db:
         rows = db.execute(
             "SELECT * FROM tasks WHERE status='pending_approval' ORDER BY created_at DESC"
         ).fetchall()
-    finally:
-        db.close()
     tasks = [_TaskView(_row_to_dict(r)) for r in rows]
     parts = []
     for task in tasks:
@@ -167,13 +164,10 @@ async def tasks_pending(request: Request):
 
 @router.get("/tasks/active", response_class=HTMLResponse)
 async def tasks_active(request: Request):
-    db = _get_db()
-    try:
+    with _get_db() as db:
         rows = db.execute(
             "SELECT * FROM tasks WHERE status IN ('approved','running','paused') ORDER BY created_at DESC"
         ).fetchall()
-    finally:
-        db.close()
     tasks = [_TaskView(_row_to_dict(r)) for r in rows]
     parts = []
     for task in tasks:
@@ -189,13 +183,10 @@ async def tasks_active(request: Request):
 
 @router.get("/tasks/completed", response_class=HTMLResponse)
 async def tasks_completed(request: Request):
-    db = _get_db()
-    try:
+    with _get_db() as db:
         rows = db.execute(
             "SELECT * FROM tasks WHERE status IN ('completed','vetoed','failed') ORDER BY completed_at DESC LIMIT 50"
         ).fetchall()
-    finally:
-        db.close()
     tasks = [_TaskView(_row_to_dict(r)) for r in rows]
     parts = []
     for task in tasks:
@@ -231,16 +222,13 @@ async def create_task_form(
     now = datetime.utcnow().isoformat()
     priority = priority if priority in VALID_PRIORITIES else "normal"
 
-    db = _get_db()
-    try:
+    with _get_db() as db:
         db.execute(
             "INSERT INTO tasks (id, title, description, priority, assigned_to, created_at) VALUES (?, ?, ?, ?, ?, ?)",
             (task_id, title, description, priority, assigned_to, now),
         )
         db.commit()
         row = db.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
-    finally:
-        db.close()
 
     task = _TaskView(_row_to_dict(row))
     return templates.TemplateResponse(request, "partials/task_card.html", {"task": task})
@@ -283,16 +271,13 @@ async def modify_task(
     title: str = Form(...),
     description: str = Form(""),
 ):
-    db = _get_db()
-    try:
+    with _get_db() as db:
         db.execute(
             "UPDATE tasks SET title=?, description=? WHERE id=?",
             (title, description, task_id),
         )
         db.commit()
         row = db.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
-    finally:
-        db.close()
     if not row:
         raise HTTPException(404, "Task not found")
     task = _TaskView(_row_to_dict(row))
@@ -304,16 +289,13 @@ async def _set_status(request: Request, task_id: str, new_status: str):
     completed_at = (
         datetime.utcnow().isoformat() if new_status in ("completed", "vetoed", "failed") else None
     )
-    db = _get_db()
-    try:
+    with _get_db() as db:
         db.execute(
             "UPDATE tasks SET status=?, completed_at=COALESCE(?, completed_at) WHERE id=?",
             (new_status, completed_at, task_id),
         )
         db.commit()
         row = db.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
-    finally:
-        db.close()
     if not row:
         raise HTTPException(404, "Task not found")
     task = _TaskView(_row_to_dict(row))
@@ -339,8 +321,7 @@ async def api_create_task(request: Request):
     if priority not in VALID_PRIORITIES:
         priority = "normal"
 
-    db = _get_db()
-    try:
+    with _get_db() as db:
         db.execute(
             "INSERT INTO tasks (id, title, description, priority, assigned_to, created_by, created_at) "
             "VALUES (?, ?, ?, ?, ?, ?, ?)",
@@ -356,8 +337,6 @@ async def api_create_task(request: Request):
         )
         db.commit()
         row = db.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
-    finally:
-        db.close()
 
     return JSONResponse(_row_to_dict(row), status_code=201)
 
@@ -365,11 +344,8 @@ async def api_create_task(request: Request):
 @router.get("/api/tasks", response_class=JSONResponse)
 async def api_list_tasks():
     """List all tasks as JSON."""
-    db = _get_db()
-    try:
+    with _get_db() as db:
         rows = db.execute("SELECT * FROM tasks ORDER BY created_at DESC").fetchall()
-    finally:
-        db.close()
     return JSONResponse([_row_to_dict(r) for r in rows])
 
 
@@ -384,16 +360,13 @@ async def api_update_status(task_id: str, request: Request):
     completed_at = (
         datetime.utcnow().isoformat() if new_status in ("completed", "vetoed", "failed") else None
     )
-    db = _get_db()
-    try:
+    with _get_db() as db:
         db.execute(
             "UPDATE tasks SET status=?, completed_at=COALESCE(?, completed_at) WHERE id=?",
             (new_status, completed_at, task_id),
         )
         db.commit()
         row = db.execute("SELECT * FROM tasks WHERE id=?", (task_id,)).fetchone()
-    finally:
-        db.close()
     if not row:
         raise HTTPException(404, "Task not found")
     return JSONResponse(_row_to_dict(row))
@@ -402,12 +375,9 @@ async def api_update_status(task_id: str, request: Request):
 @router.delete("/api/tasks/{task_id}", response_class=JSONResponse)
 async def api_delete_task(task_id: str):
     """Delete a task."""
-    db = _get_db()
-    try:
+    with _get_db() as db:
         cursor = db.execute("DELETE FROM tasks WHERE id=?", (task_id,))
         db.commit()
-    finally:
-        db.close()
     if cursor.rowcount == 0:
         raise HTTPException(404, "Task not found")
     return JSONResponse({"success": True, "id": task_id})
@@ -421,8 +391,7 @@ async def api_delete_task(task_id: str):
 @router.get("/api/queue/status", response_class=JSONResponse)
 async def queue_status(assigned_to: str = "default"):
     """Return queue status for the chat panel's agent status indicator."""
-    db = _get_db()
-    try:
+    with _get_db() as db:
         running = db.execute(
             "SELECT * FROM tasks WHERE status='running' AND assigned_to=? LIMIT 1",
             (assigned_to,),
@@ -431,8 +400,6 @@ async def queue_status(assigned_to: str = "default"):
             "SELECT COUNT(*) as cnt FROM tasks WHERE status IN ('pending_approval','approved') AND assigned_to=?",
             (assigned_to,),
         ).fetchone()
-    finally:
-        db.close()
 
     if running:
         return JSONResponse(

src/dashboard/routes/voice.py

@@ -43,7 +43,8 @@ async def tts_status():
             "available": voice_tts.available,
             "voices": voice_tts.get_voices() if voice_tts.available else [],
         }
-    except Exception:
+    except Exception as exc:
+        logger.debug("Voice config error: %s", exc)
         return {"available": False, "voices": []}
 
 
@@ -139,7 +140,8 @@ async def process_voice_input(
 
             if voice_tts.available:
                 voice_tts.speak(response_text)
-        except Exception:
+        except Exception as exc:
+            logger.debug("Voice TTS error: %s", exc)
             pass
 
     return {

src/dashboard/routes/work_orders.py

@@ -3,6 +3,8 @@
 import logging
 import sqlite3
 import uuid
+from collections.abc import Generator
+from contextlib import closing, contextmanager
 from datetime import datetime
 from pathlib import Path
 
@@ -23,28 +25,29 @@ CATEGORIES = ["bug", "feature", "suggestion", "maintenance", "security"]
 VALID_STATUSES = {"submitted", "triaged", "approved", "in_progress", "completed", "rejected"}
 
 
-def _get_db() -> sqlite3.Connection:
+@contextmanager
+def _get_db() -> Generator[sqlite3.Connection, None, None]:
     DB_PATH.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(DB_PATH))
-    conn.row_factory = sqlite3.Row
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS work_orders (
-            id TEXT PRIMARY KEY,
-            title TEXT NOT NULL,
-            description TEXT DEFAULT '',
-            priority TEXT DEFAULT 'medium',
-            category TEXT DEFAULT 'suggestion',
-            submitter TEXT DEFAULT 'dashboard',
-            related_files TEXT DEFAULT '',
-            status TEXT DEFAULT 'submitted',
-            result TEXT DEFAULT '',
-            rejection_reason TEXT DEFAULT '',
-            created_at TEXT DEFAULT (datetime('now')),
-            completed_at TEXT
-        )
-    """)
-    conn.commit()
-    return conn
+    with closing(sqlite3.connect(str(DB_PATH))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS work_orders (
+                id TEXT PRIMARY KEY,
+                title TEXT NOT NULL,
+                description TEXT DEFAULT '',
+                priority TEXT DEFAULT 'medium',
+                category TEXT DEFAULT 'suggestion',
+                submitter TEXT DEFAULT 'dashboard',
+                related_files TEXT DEFAULT '',
+                status TEXT DEFAULT 'submitted',
+                result TEXT DEFAULT '',
+                rejection_reason TEXT DEFAULT '',
+                created_at TEXT DEFAULT (datetime('now')),
+                completed_at TEXT
+            )
+        """)
+        conn.commit()
+        yield conn
 
 
 class _EnumLike:
@@ -104,14 +107,11 @@ def _query_wos(db, statuses):
 
 @router.get("/work-orders/queue", response_class=HTMLResponse)
 async def work_orders_page(request: Request):
-    db = _get_db()
-    try:
+    with _get_db() as db:
         pending = _query_wos(db, ["submitted", "triaged"])
         active = _query_wos(db, ["approved", "in_progress"])
         completed = _query_wos(db, ["completed"])
         rejected = _query_wos(db, ["rejected"])
-    finally:
-        db.close()
 
     return templates.TemplateResponse(
         request,
@@ -148,8 +148,7 @@ async def submit_work_order(
     priority = priority if priority in PRIORITIES else "medium"
     category = category if category in CATEGORIES else "suggestion"
 
-    db = _get_db()
-    try:
+    with _get_db() as db:
         db.execute(
             "INSERT INTO work_orders (id, title, description, priority, category, submitter, related_files, created_at) "
             "VALUES (?, ?, ?, ?, ?, ?, ?, ?)",
@@ -157,8 +156,6 @@ async def submit_work_order(
         )
         db.commit()
         row = db.execute("SELECT * FROM work_orders WHERE id=?", (wo_id,)).fetchone()
-    finally:
-        db.close()
 
     wo = _WOView(_row_to_dict(row))
     return templates.TemplateResponse(request, "partials/work_order_card.html", {"wo": wo})
@@ -171,11 +168,8 @@ async def submit_work_order(
 
 @router.get("/work-orders/queue/pending", response_class=HTMLResponse)
 async def pending_partial(request: Request):
-    db = _get_db()
-    try:
+    with _get_db() as db:
         wos = _query_wos(db, ["submitted", "triaged"])
-    finally:
-        db.close()
     if not wos:
         return HTMLResponse(
             '<div style="color: var(--text-muted); font-size: 0.8rem; padding: 12px 0;">'
@@ -193,11 +187,8 @@ async def pending_partial(request: Request):
 
 @router.get("/work-orders/queue/active", response_class=HTMLResponse)
 async def active_partial(request: Request):
-    db = _get_db()
-    try:
+    with _get_db() as db:
         wos = _query_wos(db, ["approved", "in_progress"])
-    finally:
-        db.close()
     if not wos:
         return HTMLResponse(
             '<div style="color: var(--text-muted); font-size: 0.8rem; padding: 12px 0;">'
@@ -222,8 +213,7 @@ async def _update_status(request: Request, wo_id: str, new_status: str, **extra)
     completed_at = (
         datetime.utcnow().isoformat() if new_status in ("completed", "rejected") else None
     )
-    db = _get_db()
-    try:
+    with _get_db() as db:
         sets = ["status=?", "completed_at=COALESCE(?, completed_at)"]
         vals = [new_status, completed_at]
         for col, val in extra.items():
@@ -233,8 +223,6 @@ async def _update_status(request: Request, wo_id: str, new_status: str, **extra)
         db.execute(f"UPDATE work_orders SET {', '.join(sets)} WHERE id=?", vals)
         db.commit()
         row = db.execute("SELECT * FROM work_orders WHERE id=?", (wo_id,)).fetchone()
-    finally:
-        db.close()
     if not row:
         raise HTTPException(404, "Work order not found")
     wo = _WOView(_row_to_dict(row))

src/dashboard/routes/world.py

@@ -0,0 +1,385 @@
+"""Workshop world state API and WebSocket relay.
+
+Serves Timmy's current presence state to the Workshop 3D renderer.
+The primary consumer is the browser on first load — before any
+WebSocket events arrive, the client needs a full state snapshot.
+
+The ``/ws/world`` endpoint streams ``timmy_state`` messages whenever
+the heartbeat detects a state change.  It also accepts ``visitor_message``
+frames from the 3D client and responds with ``timmy_speech`` barks.
+
+Source of truth: ``~/.timmy/presence.json`` written by
+:class:`~timmy.workshop_state.WorkshopHeartbeat`.
+Falls back to a live ``get_state_dict()`` call if the file is stale
+or missing.
+"""
+
+import asyncio
+import json
+import logging
+import re
+import time
+from collections import deque
+from datetime import UTC, datetime
+
+from fastapi import APIRouter, WebSocket
+from fastapi.responses import JSONResponse
+
+from timmy.workshop_state import PRESENCE_FILE
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/api/world", tags=["world"])
+
+# ---------------------------------------------------------------------------
+# WebSocket relay for live state changes
+# ---------------------------------------------------------------------------
+
+_ws_clients: list[WebSocket] = []
+
+_STALE_THRESHOLD = 90  # seconds — file older than this triggers live rebuild
+
+# Recent conversation buffer — kept in memory for the Workshop overlay.
+# Stores the last _MAX_EXCHANGES (visitor_text, timmy_text) pairs.
+_MAX_EXCHANGES = 3
+_conversation: deque[dict] = deque(maxlen=_MAX_EXCHANGES)
+
+_WORKSHOP_SESSION_ID = "workshop"
+
+_HEARTBEAT_INTERVAL = 15  # seconds — ping to detect dead iPad/Safari connections
+
+# ---------------------------------------------------------------------------
+# Conversation grounding — commitment tracking (rescued from PR #408)
+# ---------------------------------------------------------------------------
+
+# Patterns that indicate Timmy is committing to an action.
+_COMMITMENT_PATTERNS: list[re.Pattern[str]] = [
+    re.compile(r"I'll (.+?)(?:\.|!|\?|$)", re.IGNORECASE),
+    re.compile(r"I will (.+?)(?:\.|!|\?|$)", re.IGNORECASE),
+    re.compile(r"[Ll]et me (.+?)(?:\.|!|\?|$)", re.IGNORECASE),
+]
+
+# After this many messages without follow-up, surface open commitments.
+_REMIND_AFTER = 5
+_MAX_COMMITMENTS = 10
+
+# In-memory list of open commitments.
+# Each entry: {"text": str, "created_at": float, "messages_since": int}
+_commitments: list[dict] = []
+
+
+def _extract_commitments(text: str) -> list[str]:
+    """Pull commitment phrases from Timmy's reply text."""
+    found: list[str] = []
+    for pattern in _COMMITMENT_PATTERNS:
+        for match in pattern.finditer(text):
+            phrase = match.group(1).strip()
+            if len(phrase) > 5:  # skip trivially short matches
+                found.append(phrase[:120])
+    return found
+
+
+def _record_commitments(reply: str) -> None:
+    """Scan a Timmy reply for commitments and store them."""
+    for phrase in _extract_commitments(reply):
+        # Avoid near-duplicate commitments
+        if any(c["text"] == phrase for c in _commitments):
+            continue
+        _commitments.append({"text": phrase, "created_at": time.time(), "messages_since": 0})
+        if len(_commitments) > _MAX_COMMITMENTS:
+            _commitments.pop(0)
+
+
+def _tick_commitments() -> None:
+    """Increment messages_since for every open commitment."""
+    for c in _commitments:
+        c["messages_since"] += 1
+
+
+def _build_commitment_context() -> str:
+    """Return a grounding note if any commitments are overdue for follow-up."""
+    overdue = [c for c in _commitments if c["messages_since"] >= _REMIND_AFTER]
+    if not overdue:
+        return ""
+    lines = [f"- {c['text']}" for c in overdue]
+    return (
+        "[Open commitments Timmy made earlier — "
+        "weave awareness naturally, don't list robotically]\n" + "\n".join(lines)
+    )
+
+
+def close_commitment(index: int) -> bool:
+    """Remove a commitment by index. Returns True if removed."""
+    if 0 <= index < len(_commitments):
+        _commitments.pop(index)
+        return True
+    return False
+
+
+def get_commitments() -> list[dict]:
+    """Return a copy of open commitments (for testing / API)."""
+    return list(_commitments)
+
+
+def reset_commitments() -> None:
+    """Clear all commitments (for testing / session reset)."""
+    _commitments.clear()
+
+
+# Conversation grounding — anchor to opening topic so Timmy doesn't drift.
+_ground_topic: str | None = None
+_ground_set_at: float = 0.0
+_GROUND_TTL = 300  # seconds of inactivity before the anchor expires
+
+
+def _read_presence_file() -> dict | None:
+    """Read presence.json if it exists and is fresh enough."""
+    try:
+        if not PRESENCE_FILE.exists():
+            return None
+        age = time.time() - PRESENCE_FILE.stat().st_mtime
+        if age > _STALE_THRESHOLD:
+            logger.debug("presence.json is stale (%.0fs old)", age)
+            return None
+        return json.loads(PRESENCE_FILE.read_text())
+    except (OSError, json.JSONDecodeError) as exc:
+        logger.warning("Failed to read presence.json: %s", exc)
+        return None
+
+
+def _build_world_state(presence: dict) -> dict:
+    """Transform presence dict into the world/state API response."""
+    return {
+        "timmyState": {
+            "mood": presence.get("mood", "calm"),
+            "activity": presence.get("current_focus", "idle"),
+            "energy": presence.get("energy", 0.5),
+            "confidence": presence.get("confidence", 0.7),
+        },
+        "familiar": presence.get("familiar"),
+        "activeThreads": presence.get("active_threads", []),
+        "recentEvents": presence.get("recent_events", []),
+        "concerns": presence.get("concerns", []),
+        "visitorPresent": False,
+        "updatedAt": presence.get("liveness", datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ")),
+        "version": presence.get("version", 1),
+    }
+
+
+def _get_current_state() -> dict:
+    """Build the current world-state dict from best available source."""
+    presence = _read_presence_file()
+
+    if presence is None:
+        try:
+            from timmy.workshop_state import get_state_dict
+
+            presence = get_state_dict()
+        except Exception as exc:
+            logger.warning("Live state build failed: %s", exc)
+            presence = {
+                "version": 1,
+                "liveness": datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ"),
+                "mood": "calm",
+                "current_focus": "",
+                "active_threads": [],
+                "recent_events": [],
+                "concerns": [],
+            }
+
+    return _build_world_state(presence)
+
+
+@router.get("/state")
+async def get_world_state() -> JSONResponse:
+    """Return Timmy's current world state for Workshop bootstrap.
+
+    Reads from ``~/.timmy/presence.json`` if fresh, otherwise
+    rebuilds live from cognitive state.
+    """
+    return JSONResponse(
+        content=_get_current_state(),
+        headers={"Cache-Control": "no-cache, no-store"},
+    )
+
+
+# ---------------------------------------------------------------------------
+# WebSocket endpoint — streams timmy_state changes to Workshop clients
+# ---------------------------------------------------------------------------
+
+
+async def _heartbeat(websocket: WebSocket) -> None:
+    """Send periodic pings to detect dead connections (iPad resilience).
+
+    Safari suspends background tabs, killing the TCP socket silently.
+    A 15-second ping ensures we notice within one interval.
+
+    Rescued from stale PR #399.
+    """
+    try:
+        while True:
+            await asyncio.sleep(_HEARTBEAT_INTERVAL)
+            await websocket.send_text(json.dumps({"type": "ping"}))
+    except Exception:
+        logger.debug("Heartbeat stopped — connection gone")
+
+
+@router.websocket("/ws")
+async def world_ws(websocket: WebSocket) -> None:
+    """Accept a Workshop client and keep it alive for state broadcasts.
+
+    Sends a full ``world_state`` snapshot immediately on connect so the
+    client never starts from a blank slate.  Incoming frames are parsed
+    as JSON — ``visitor_message`` triggers a bark response.  A background
+    heartbeat ping runs every 15 s to detect dead connections early.
+    """
+    await websocket.accept()
+    _ws_clients.append(websocket)
+    logger.info("World WS connected — %d clients", len(_ws_clients))
+
+    # Send full world-state snapshot so client bootstraps instantly
+    try:
+        snapshot = _get_current_state()
+        await websocket.send_text(json.dumps({"type": "world_state", **snapshot}))
+    except Exception as exc:
+        logger.warning("Failed to send WS snapshot: %s", exc)
+
+    ping_task = asyncio.create_task(_heartbeat(websocket))
+    try:
+        while True:
+            raw = await websocket.receive_text()
+            await _handle_client_message(raw)
+    except Exception:
+        logger.debug("WebSocket receive loop ended")
+    finally:
+        ping_task.cancel()
+        if websocket in _ws_clients:
+            _ws_clients.remove(websocket)
+        logger.info("World WS disconnected — %d clients", len(_ws_clients))
+
+
+async def _broadcast(message: str) -> None:
+    """Send *message* to every connected Workshop client, pruning dead ones."""
+    dead: list[WebSocket] = []
+    for ws in _ws_clients:
+        try:
+            await ws.send_text(message)
+        except Exception:
+            logger.debug("Pruning dead WebSocket client")
+            dead.append(ws)
+    for ws in dead:
+        if ws in _ws_clients:
+            _ws_clients.remove(ws)
+
+
+async def broadcast_world_state(presence: dict) -> None:
+    """Broadcast a ``timmy_state`` message to all connected Workshop clients.
+
+    Called by :class:`~timmy.workshop_state.WorkshopHeartbeat` via its
+    ``on_change`` callback.
+    """
+    state = _build_world_state(presence)
+    await _broadcast(json.dumps({"type": "timmy_state", **state["timmyState"]}))
+
+
+# ---------------------------------------------------------------------------
+# Visitor chat — bark engine
+# ---------------------------------------------------------------------------
+
+
+async def _handle_client_message(raw: str) -> None:
+    """Dispatch an incoming WebSocket frame from the Workshop client."""
+    try:
+        data = json.loads(raw)
+    except (json.JSONDecodeError, TypeError):
+        return  # ignore non-JSON keep-alive pings
+
+    if data.get("type") == "visitor_message":
+        text = (data.get("text") or "").strip()
+        if text:
+            task = asyncio.create_task(_bark_and_broadcast(text))
+            task.add_done_callback(_log_bark_failure)
+
+
+def _log_bark_failure(task: asyncio.Task) -> None:
+    """Log unhandled exceptions from fire-and-forget bark tasks."""
+    if task.cancelled():
+        return
+    exc = task.exception()
+    if exc is not None:
+        logger.error("Bark task failed: %s", exc)
+
+
+def reset_conversation_ground() -> None:
+    """Clear the conversation grounding anchor (e.g. after inactivity)."""
+    global _ground_topic, _ground_set_at
+    _ground_topic = None
+    _ground_set_at = 0.0
+
+
+def _refresh_ground(visitor_text: str) -> None:
+    """Set or refresh the conversation grounding anchor.
+
+    The first visitor message in a session (or after the TTL expires)
+    becomes the anchor topic. Subsequent messages are grounded against it.
+    """
+    global _ground_topic, _ground_set_at
+    now = time.time()
+    if _ground_topic is None or (now - _ground_set_at) > _GROUND_TTL:
+        _ground_topic = visitor_text[:120]
+        logger.debug("Ground topic set: %s", _ground_topic)
+    _ground_set_at = now
+
+
+async def _bark_and_broadcast(visitor_text: str) -> None:
+    """Generate a bark response and broadcast it to all Workshop clients."""
+    await _broadcast(json.dumps({"type": "timmy_thinking"}))
+
+    # Notify Pip that a visitor spoke
+    try:
+        from timmy.familiar import pip_familiar
+
+        pip_familiar.on_event("visitor_spoke")
+    except Exception:
+        logger.debug("Pip familiar notification failed (optional)")
+
+    _refresh_ground(visitor_text)
+    _tick_commitments()
+    reply = await _generate_bark(visitor_text)
+    _record_commitments(reply)
+
+    _conversation.append({"visitor": visitor_text, "timmy": reply})
+
+    await _broadcast(
+        json.dumps(
+            {
+                "type": "timmy_speech",
+                "text": reply,
+                "recentExchanges": list(_conversation),
+            }
+        )
+    )
+
+
+async def _generate_bark(visitor_text: str) -> str:
+    """Generate a short in-character bark response.
+
+    Uses the existing Timmy session with a dedicated workshop session ID.
+    When a grounding anchor exists, the opening topic is prepended so the
+    model stays on-topic across long sessions.
+    Gracefully degrades to a canned response if inference fails.
+    """
+    try:
+        from timmy import session as _session
+
+        grounded = visitor_text
+        commitment_ctx = _build_commitment_context()
+        if commitment_ctx:
+            grounded = f"{commitment_ctx}\n{grounded}"
+        if _ground_topic and visitor_text != _ground_topic:
+            grounded = f"[Workshop conversation topic: {_ground_topic}]\n{grounded}"
+        response = await _session.chat(grounded, session_id=_WORKSHOP_SESSION_ID)
+        return response
+    except Exception as exc:
+        logger.warning("Bark generation failed: %s", exc)
+        return "Hmm, my thoughts are a bit tangled right now."

src/dashboard/store.py

@@ -1,134 +1,5 @@
-"""Persistent chat message store backed by SQLite.
+"""Backward-compatible re-export — canonical home is infrastructure.chat_store."""
 
-Provides the same API as the original in-memory MessageLog so all callers
-(dashboard routes, chat_api, thinking, briefing) work without changes.
+from infrastructure.chat_store import DB_PATH, MAX_MESSAGES, Message, MessageLog, message_log
 
-Data lives in ``data/chat.db`` — survives server restarts.
-A configurable retention policy (default 500 messages) keeps the DB lean.
-"""
-
-import sqlite3
-import threading
-from dataclasses import dataclass
-from pathlib import Path
-
-# ── Data dir — resolved relative to repo root (two levels up from this file) ──
-_REPO_ROOT = Path(__file__).resolve().parents[2]
-DB_PATH: Path = _REPO_ROOT / "data" / "chat.db"
-
-# Maximum messages to retain (oldest pruned on append)
-MAX_MESSAGES: int = 500
-
-
-@dataclass
-class Message:
-    role: str  # "user" | "agent" | "error"
-    content: str
-    timestamp: str
-    source: str = "browser"  # "browser" | "api" | "telegram" | "discord" | "system"
-
-
-def _get_conn(db_path: Path | None = None) -> sqlite3.Connection:
-    """Open (or create) the chat database and ensure schema exists."""
-    path = db_path or DB_PATH
-    path.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(path), check_same_thread=False)
-    conn.row_factory = sqlite3.Row
-    conn.execute("PRAGMA journal_mode=WAL")
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS chat_messages (
-            id        INTEGER PRIMARY KEY AUTOINCREMENT,
-            role      TEXT NOT NULL,
-            content   TEXT NOT NULL,
-            timestamp TEXT NOT NULL,
-            source    TEXT NOT NULL DEFAULT 'browser'
-        )
-    """)
-    conn.commit()
-    return conn
-
-
-class MessageLog:
-    """SQLite-backed chat history — drop-in replacement for the old in-memory list."""
-
-    def __init__(self, db_path: Path | None = None) -> None:
-        self._db_path = db_path or DB_PATH
-        self._lock = threading.Lock()
-        self._conn: sqlite3.Connection | None = None
-
-    # Lazy connection — opened on first use, not at import time.
-    def _ensure_conn(self) -> sqlite3.Connection:
-        if self._conn is None:
-            self._conn = _get_conn(self._db_path)
-        return self._conn
-
-    def append(self, role: str, content: str, timestamp: str, source: str = "browser") -> None:
-        with self._lock:
-            conn = self._ensure_conn()
-            conn.execute(
-                "INSERT INTO chat_messages (role, content, timestamp, source) VALUES (?, ?, ?, ?)",
-                (role, content, timestamp, source),
-            )
-            conn.commit()
-            self._prune(conn)
-
-    def all(self) -> list[Message]:
-        with self._lock:
-            conn = self._ensure_conn()
-            rows = conn.execute(
-                "SELECT role, content, timestamp, source FROM chat_messages ORDER BY id"
-            ).fetchall()
-        return [
-            Message(
-                role=r["role"], content=r["content"], timestamp=r["timestamp"], source=r["source"]
-            )
-            for r in rows
-        ]
-
-    def recent(self, limit: int = 50) -> list[Message]:
-        """Return the *limit* most recent messages (oldest-first)."""
-        with self._lock:
-            conn = self._ensure_conn()
-            rows = conn.execute(
-                "SELECT role, content, timestamp, source FROM chat_messages "
-                "ORDER BY id DESC LIMIT ?",
-                (limit,),
-            ).fetchall()
-        return [
-            Message(
-                role=r["role"], content=r["content"], timestamp=r["timestamp"], source=r["source"]
-            )
-            for r in reversed(rows)
-        ]
-
-    def clear(self) -> None:
-        with self._lock:
-            conn = self._ensure_conn()
-            conn.execute("DELETE FROM chat_messages")
-            conn.commit()
-
-    def _prune(self, conn: sqlite3.Connection) -> None:
-        """Keep at most MAX_MESSAGES rows, deleting the oldest."""
-        count = conn.execute("SELECT COUNT(*) FROM chat_messages").fetchone()[0]
-        if count > MAX_MESSAGES:
-            excess = count - MAX_MESSAGES
-            conn.execute(
-                "DELETE FROM chat_messages WHERE id IN "
-                "(SELECT id FROM chat_messages ORDER BY id LIMIT ?)",
-                (excess,),
-            )
-            conn.commit()
-
-    def close(self) -> None:
-        if self._conn is not None:
-            self._conn.close()
-            self._conn = None
-
-    def __len__(self) -> int:
-        with self._lock:
-            conn = self._ensure_conn()
-            return conn.execute("SELECT COUNT(*) FROM chat_messages").fetchone()[0]
-
-
-# Module-level singleton shared across the app
-message_log = MessageLog()
+__all__ = ["DB_PATH", "MAX_MESSAGES", "Message", "MessageLog", "message_log"]

src/dashboard/templates/partials/history.html

@@ -20,7 +20,7 @@
 {% else %}
 <div class="chat-message agent">
   <div class="msg-meta">TIMMY // SYSTEM</div>
-  <div class="msg-body">Mission Control initialized. Timmy ready — awaiting input.</div>
+  <div class="msg-body">{{ welcome_message | e }}</div>
 </div>
 {% endif %}
 <script>if(typeof scrollChat==='function'){setTimeout(scrollChat,50);}</script>

src/infrastructure/chat_store.py

@@ -0,0 +1,153 @@
+"""Persistent chat message store backed by SQLite.
+
+Provides the same API as the original in-memory MessageLog so all callers
+(dashboard routes, chat_api, thinking, briefing) work without changes.
+
+Data lives in ``data/chat.db`` — survives server restarts.
+A configurable retention policy (default 500 messages) keeps the DB lean.
+"""
+
+import sqlite3
+import threading
+from collections.abc import Generator
+from contextlib import closing, contextmanager
+from dataclasses import dataclass
+from pathlib import Path
+
+# ── Data dir — resolved relative to repo root (three levels up from this file) ──
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+DB_PATH: Path = _REPO_ROOT / "data" / "chat.db"
+
+# Maximum messages to retain (oldest pruned on append)
+MAX_MESSAGES: int = 500
+
+
+@dataclass
+class Message:
+    role: str  # "user" | "agent" | "error"
+    content: str
+    timestamp: str
+    source: str = "browser"  # "browser" | "api" | "telegram" | "discord" | "system"
+
+
+@contextmanager
+def _get_conn(db_path: Path | None = None) -> Generator[sqlite3.Connection, None, None]:
+    """Open (or create) the chat database and ensure schema exists."""
+    path = db_path or DB_PATH
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with closing(sqlite3.connect(str(path), check_same_thread=False)) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("PRAGMA journal_mode=WAL")
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS chat_messages (
+                id        INTEGER PRIMARY KEY AUTOINCREMENT,
+                role      TEXT NOT NULL,
+                content   TEXT NOT NULL,
+                timestamp TEXT NOT NULL,
+                source    TEXT NOT NULL DEFAULT 'browser'
+            )
+        """)
+        conn.commit()
+        yield conn
+
+
+class MessageLog:
+    """SQLite-backed chat history — drop-in replacement for the old in-memory list."""
+
+    def __init__(self, db_path: Path | None = None) -> None:
+        self._db_path = db_path or DB_PATH
+        self._lock = threading.Lock()
+        self._conn: sqlite3.Connection | None = None
+
+    # Lazy connection — opened on first use, not at import time.
+    def _ensure_conn(self) -> sqlite3.Connection:
+        if self._conn is None:
+            # Open a persistent connection for the class instance
+            path = self._db_path or DB_PATH
+            path.parent.mkdir(parents=True, exist_ok=True)
+            conn = sqlite3.connect(str(path), check_same_thread=False)
+            conn.row_factory = sqlite3.Row
+            conn.execute("PRAGMA journal_mode=WAL")
+            conn.execute("""
+                CREATE TABLE IF NOT EXISTS chat_messages (
+                    id        INTEGER PRIMARY KEY AUTOINCREMENT,
+                    role      TEXT NOT NULL,
+                    content   TEXT NOT NULL,
+                    timestamp TEXT NOT NULL,
+                    source    TEXT NOT NULL DEFAULT 'browser'
+                )
+            """)
+            conn.commit()
+            self._conn = conn
+        return self._conn
+
+    def append(self, role: str, content: str, timestamp: str, source: str = "browser") -> None:
+        with self._lock:
+            conn = self._ensure_conn()
+            conn.execute(
+                "INSERT INTO chat_messages (role, content, timestamp, source) VALUES (?, ?, ?, ?)",
+                (role, content, timestamp, source),
+            )
+            conn.commit()
+            self._prune(conn)
+
+    def all(self) -> list[Message]:
+        with self._lock:
+            conn = self._ensure_conn()
+            rows = conn.execute(
+                "SELECT role, content, timestamp, source FROM chat_messages ORDER BY id"
+            ).fetchall()
+        return [
+            Message(
+                role=r["role"], content=r["content"], timestamp=r["timestamp"], source=r["source"]
+            )
+            for r in rows
+        ]
+
+    def recent(self, limit: int = 50) -> list[Message]:
+        """Return the *limit* most recent messages (oldest-first)."""
+        with self._lock:
+            conn = self._ensure_conn()
+            rows = conn.execute(
+                "SELECT role, content, timestamp, source FROM chat_messages "
+                "ORDER BY id DESC LIMIT ?",
+                (limit,),
+            ).fetchall()
+        return [
+            Message(
+                role=r["role"], content=r["content"], timestamp=r["timestamp"], source=r["source"]
+            )
+            for r in reversed(rows)
+        ]
+
+    def clear(self) -> None:
+        with self._lock:
+            conn = self._ensure_conn()
+            conn.execute("DELETE FROM chat_messages")
+            conn.commit()
+
+    def _prune(self, conn: sqlite3.Connection) -> None:
+        """Keep at most MAX_MESSAGES rows, deleting the oldest."""
+        count = conn.execute("SELECT COUNT(*) FROM chat_messages").fetchone()[0]
+        if count > MAX_MESSAGES:
+            excess = count - MAX_MESSAGES
+            conn.execute(
+                "DELETE FROM chat_messages WHERE id IN "
+                "(SELECT id FROM chat_messages ORDER BY id LIMIT ?)",
+                (excess,),
+            )
+            conn.commit()
+
+    def close(self) -> None:
+        if self._conn is not None:
+            self._conn.close()
+            self._conn = None
+
+    def __len__(self) -> int:
+        with self._lock:
+            conn = self._ensure_conn()
+            return conn.execute("SELECT COUNT(*) FROM chat_messages").fetchone()[0]
+
+
+# Module-level singleton shared across the app
+message_log = MessageLog()

src/infrastructure/error_capture.py

@@ -22,6 +22,14 @@ logger = logging.getLogger(__name__)
 # In-memory dedup cache: hash -> last_seen timestamp
 _dedup_cache: dict[str, datetime] = {}
 
+_error_recorder = None
+
+
+def register_error_recorder(fn):
+    """Register a callback for recording errors to session log."""
+    global _error_recorder
+    _error_recorder = fn
+
 
 def _stack_hash(exc: Exception) -> str:
     """Create a stable hash of the exception type + traceback locations.
@@ -87,40 +95,19 @@ def _get_git_context() -> dict:
         ).stdout.strip()
 
         return {"branch": branch, "commit": commit}
-    except Exception:
+    except Exception as exc:
+        logger.warning("Git info capture error: %s", exc)
         return {"branch": "unknown", "commit": "unknown"}
 
 
-def capture_error(
-    exc: Exception,
-    source: str = "unknown",
-    context: dict | None = None,
-) -> str | None:
-    """Capture an error and optionally create a bug report.
-
-    Args:
-        exc: The exception to capture
-        source: Module/component where the error occurred
-        context: Optional dict of extra context (request path, etc.)
+def _extract_traceback_info(exc: Exception) -> tuple[str, str, int]:
+    """Extract formatted traceback, affected file, and line number.
 
     Returns:
-        Task ID of the created bug report, or None if deduplicated/disabled
+        Tuple of (traceback_string, affected_file, affected_line).
     """
-    from config import settings
-
-    if not settings.error_feedback_enabled:
-        return None
-
-    error_hash = _stack_hash(exc)
-
-    if _is_duplicate(error_hash):
-        logger.debug("Duplicate error suppressed: %s (hash=%s)", exc, error_hash)
-        return None
-
-    # Format the stack trace
     tb_str = "".join(traceback.format_exception(type(exc), exc, exc.__traceback__))
 
-    # Extract file/line from traceback
     tb_obj = exc.__traceback__
     affected_file = "unknown"
     affected_line = 0
@@ -130,9 +117,18 @@ def capture_error(
         affected_file = tb_obj.tb_frame.f_code.co_filename
         affected_line = tb_obj.tb_lineno
 
-    git_ctx = _get_git_context()
+    return tb_str, affected_file, affected_line
 
-    # 1. Log to event_log
+
+def _log_error_event(
+    exc: Exception,
+    source: str,
+    error_hash: str,
+    affected_file: str,
+    affected_line: int,
+    git_ctx: dict,
+) -> None:
+    """Log the captured error to the event log."""
     try:
         from swarm.event_log import EventType, log_event
 
@@ -152,8 +148,18 @@ def capture_error(
     except Exception as log_exc:
         logger.debug("Failed to log error event: %s", log_exc)
 
-    # 2. Create bug report task
-    task_id = None
+
+def _create_bug_report(
+    exc: Exception,
+    source: str,
+    context: dict | None,
+    error_hash: str,
+    tb_str: str,
+    affected_file: str,
+    affected_line: int,
+    git_ctx: dict,
+) -> str | None:
+    """Create a bug report task and return the task ID (or None on failure)."""
     try:
         from swarm.task_queue.models import create_task
 
@@ -186,7 +192,6 @@ def capture_error(
         )
         task_id = task.id
 
-        # Log the creation event
         try:
             from swarm.event_log import EventType, log_event
 
@@ -199,13 +204,18 @@ def capture_error(
                     "title": title[:100],
                 },
             )
-        except Exception:
-            pass
+        except Exception as exc:
+            logger.warning("Bug report screenshot error: %s", exc)
+
+        return task_id
 
     except Exception as task_exc:
         logger.debug("Failed to create bug report task: %s", task_exc)
+        return None
 
-    # 3. Send notification
+
+def _notify_bug_report(exc: Exception, source: str) -> None:
+    """Send a push notification about the captured error."""
     try:
         from infrastructure.notifications.push import notifier
 
@@ -214,19 +224,65 @@ def capture_error(
             message=f"{type(exc).__name__} in {source}: {str(exc)[:80]}",
             category="system",
         )
-    except Exception:
-        pass
+    except Exception as notify_exc:
+        logger.warning("Bug report notification error: %s", notify_exc)
 
-    # 4. Record in session logger
-    try:
-        from timmy.session_logger import get_session_logger
 
-        session_logger = get_session_logger()
-        session_logger.record_error(
-            error=f"{type(exc).__name__}: {str(exc)}",
-            context=source,
-        )
-    except Exception:
-        pass
+def _record_to_session(exc: Exception, source: str) -> None:
+    """Record the error via the registered session callback."""
+    if _error_recorder is not None:
+        try:
+            _error_recorder(
+                error=f"{type(exc).__name__}: {str(exc)}",
+                context=source,
+            )
+        except Exception as log_exc:
+            logger.warning("Bug report session logging error: %s", log_exc)
+
+
+def capture_error(
+    exc: Exception,
+    source: str = "unknown",
+    context: dict | None = None,
+) -> str | None:
+    """Capture an error and optionally create a bug report.
+
+    Args:
+        exc: The exception to capture
+        source: Module/component where the error occurred
+        context: Optional dict of extra context (request path, etc.)
+
+    Returns:
+        Task ID of the created bug report, or None if deduplicated/disabled
+    """
+    from config import settings
+
+    if not settings.error_feedback_enabled:
+        return None
+
+    error_hash = _stack_hash(exc)
+
+    if _is_duplicate(error_hash):
+        logger.debug("Duplicate error suppressed: %s (hash=%s)", exc, error_hash)
+        return None
+
+    tb_str, affected_file, affected_line = _extract_traceback_info(exc)
+    git_ctx = _get_git_context()
+
+    _log_error_event(exc, source, error_hash, affected_file, affected_line, git_ctx)
+
+    task_id = _create_bug_report(
+        exc,
+        source,
+        context,
+        error_hash,
+        tb_str,
+        affected_file,
+        affected_line,
+        git_ctx,
+    )
+
+    _notify_bug_report(exc, source)
+    _record_to_session(exc, source)
 
     return task_id

src/infrastructure/events/broadcaster.py

@@ -1,193 +0,0 @@
-"""Event Broadcaster - bridges event_log to WebSocket clients.
-
-When events are logged, they are broadcast to all connected dashboard clients
-via WebSocket for real-time activity feed updates.
-"""
-
-import asyncio
-import logging
-from typing import Optional
-
-try:
-    from swarm.event_log import EventLogEntry
-except ImportError:
-    EventLogEntry = None
-
-logger = logging.getLogger(__name__)
-
-
-class EventBroadcaster:
-    """Broadcasts events to WebSocket clients.
-
-    Usage:
-        from infrastructure.events.broadcaster import event_broadcaster
-        event_broadcaster.broadcast(event)
-    """
-
-    def __init__(self) -> None:
-        self._ws_manager: Optional = None
-
-    def _get_ws_manager(self):
-        """Lazy import to avoid circular deps."""
-        if self._ws_manager is None:
-            try:
-                from infrastructure.ws_manager.handler import ws_manager
-
-                self._ws_manager = ws_manager
-            except Exception as exc:
-                logger.debug("WebSocket manager not available: %s", exc)
-        return self._ws_manager
-
-    async def broadcast(self, event: EventLogEntry) -> int:
-        """Broadcast an event to all connected WebSocket clients.
-
-        Args:
-            event: The event to broadcast
-
-        Returns:
-            Number of clients notified
-        """
-        ws_manager = self._get_ws_manager()
-        if not ws_manager:
-            return 0
-
-        # Build message payload
-        payload = {
-            "type": "event",
-            "payload": {
-                "id": event.id,
-                "event_type": event.event_type.value,
-                "source": event.source,
-                "task_id": event.task_id,
-                "agent_id": event.agent_id,
-                "timestamp": event.timestamp,
-                "data": event.data,
-            },
-        }
-
-        try:
-            # Broadcast to all connected clients
-            count = await ws_manager.broadcast_json(payload)
-            logger.debug("Broadcasted event %s to %d clients", event.id[:8], count)
-            return count
-        except Exception as exc:
-            logger.error("Failed to broadcast event: %s", exc)
-            return 0
-
-    def broadcast_sync(self, event: EventLogEntry) -> None:
-        """Synchronous wrapper for broadcast.
-
-        Use this from synchronous code - it schedules the async broadcast
-        in the event loop if one is running.
-        """
-        try:
-            asyncio.get_running_loop()
-            # Schedule in background, don't wait
-            asyncio.create_task(self.broadcast(event))
-        except RuntimeError:
-            # No event loop running, skip broadcast
-            pass
-
-
-# Global singleton
-event_broadcaster = EventBroadcaster()
-
-
-# Event type to icon/emoji mapping
-EVENT_ICONS = {
-    "task.created": "📝",
-    "task.bidding": "⏳",
-    "task.assigned": "👤",
-    "task.started": "▶️",
-    "task.completed": "✅",
-    "task.failed": "❌",
-    "agent.joined": "🟢",
-    "agent.left": "🔴",
-    "agent.status_changed": "🔄",
-    "bid.submitted": "💰",
-    "auction.closed": "🏁",
-    "tool.called": "🔧",
-    "tool.completed": "⚙️",
-    "tool.failed": "💥",
-    "system.error": "⚠️",
-    "system.warning": "🔶",
-    "system.info": "ℹ️",
-    "error.captured": "🐛",
-    "bug_report.created": "📋",
-}
-
-EVENT_LABELS = {
-    "task.created": "New task",
-    "task.bidding": "Bidding open",
-    "task.assigned": "Task assigned",
-    "task.started": "Task started",
-    "task.completed": "Task completed",
-    "task.failed": "Task failed",
-    "agent.joined": "Agent joined",
-    "agent.left": "Agent left",
-    "agent.status_changed": "Status changed",
-    "bid.submitted": "Bid submitted",
-    "auction.closed": "Auction closed",
-    "tool.called": "Tool called",
-    "tool.completed": "Tool completed",
-    "tool.failed": "Tool failed",
-    "system.error": "Error",
-    "system.warning": "Warning",
-    "system.info": "Info",
-    "error.captured": "Error captured",
-    "bug_report.created": "Bug report filed",
-}
-
-
-def get_event_icon(event_type: str) -> str:
-    """Get emoji icon for event type."""
-    return EVENT_ICONS.get(event_type, "•")
-
-
-def get_event_label(event_type: str) -> str:
-    """Get human-readable label for event type."""
-    return EVENT_LABELS.get(event_type, event_type)
-
-
-def format_event_for_display(event: EventLogEntry) -> dict:
-    """Format event for display in activity feed.
-
-    Returns dict with display-friendly fields.
-    """
-    data = event.data or {}
-
-    # Build description based on event type
-    description = ""
-    if event.event_type.value == "task.created":
-        desc = data.get("description", "")
-        description = desc[:60] + "..." if len(desc) > 60 else desc
-    elif event.event_type.value == "task.assigned":
-        agent = event.agent_id[:8] if event.agent_id else "unknown"
-        bid = data.get("bid_sats", "?")
-        description = f"to {agent} ({bid} sats)"
-    elif event.event_type.value == "bid.submitted":
-        bid = data.get("bid_sats", "?")
-        description = f"{bid} sats"
-    elif event.event_type.value == "agent.joined":
-        persona = data.get("persona_id", "")
-        description = f"Persona: {persona}" if persona else "New agent"
-    else:
-        # Generic: use any string data
-        for key in ["message", "reason", "description"]:
-            if key in data:
-                val = str(data[key])
-                description = val[:60] + "..." if len(val) > 60 else val
-                break
-
-    return {
-        "id": event.id,
-        "icon": get_event_icon(event.event_type.value),
-        "label": get_event_label(event.event_type.value),
-        "type": event.event_type.value,
-        "source": event.source,
-        "description": description,
-        "timestamp": event.timestamp,
-        "time_short": event.timestamp[11:19] if event.timestamp else "",
-        "task_id": event.task_id,
-        "agent_id": event.agent_id,
-    }

src/infrastructure/events/bus.py

@@ -9,7 +9,8 @@ import asyncio
 import json
 import logging
 import sqlite3
-from collections.abc import Callable, Coroutine
+from collections.abc import Callable, Coroutine, Generator
+from contextlib import closing, contextmanager
 from dataclasses import dataclass, field
 from datetime import UTC, datetime
 from pathlib import Path
@@ -99,51 +100,48 @@ class EventBus:
         if self._persistence_db_path is None:
             return
         self._persistence_db_path.parent.mkdir(parents=True, exist_ok=True)
-        conn = sqlite3.connect(str(self._persistence_db_path))
-        try:
+        with closing(sqlite3.connect(str(self._persistence_db_path))) as conn:
             conn.execute("PRAGMA journal_mode=WAL")
             conn.execute("PRAGMA busy_timeout=5000")
             conn.executescript(_EVENTS_SCHEMA)
             conn.commit()
-        finally:
-            conn.close()
 
-    def _get_persistence_conn(self) -> sqlite3.Connection | None:
+    @contextmanager
+    def _get_persistence_conn(self) -> Generator[sqlite3.Connection | None, None, None]:
         """Get a connection to the persistence database."""
         if self._persistence_db_path is None:
-            return None
-        conn = sqlite3.connect(str(self._persistence_db_path))
-        conn.row_factory = sqlite3.Row
-        conn.execute("PRAGMA busy_timeout=5000")
-        return conn
+            yield None
+            return
+        with closing(sqlite3.connect(str(self._persistence_db_path))) as conn:
+            conn.row_factory = sqlite3.Row
+            conn.execute("PRAGMA busy_timeout=5000")
+            yield conn
 
     def _persist_event(self, event: Event) -> None:
         """Write an event to the persistence database."""
-        conn = self._get_persistence_conn()
-        if conn is None:
-            return
-        try:
-            task_id = event.data.get("task_id", "")
-            agent_id = event.data.get("agent_id", "")
-            conn.execute(
-                "INSERT OR IGNORE INTO events "
-                "(id, event_type, source, task_id, agent_id, data, timestamp) "
-                "VALUES (?, ?, ?, ?, ?, ?, ?)",
-                (
-                    event.id,
-                    event.type,
-                    event.source,
-                    task_id,
-                    agent_id,
-                    json.dumps(event.data),
-                    event.timestamp,
-                ),
-            )
-            conn.commit()
-        except Exception as exc:
-            logger.debug("Failed to persist event: %s", exc)
-        finally:
-            conn.close()
+        with self._get_persistence_conn() as conn:
+            if conn is None:
+                return
+            try:
+                task_id = event.data.get("task_id", "")
+                agent_id = event.data.get("agent_id", "")
+                conn.execute(
+                    "INSERT OR IGNORE INTO events "
+                    "(id, event_type, source, task_id, agent_id, data, timestamp) "
+                    "VALUES (?, ?, ?, ?, ?, ?, ?)",
+                    (
+                        event.id,
+                        event.type,
+                        event.source,
+                        task_id,
+                        agent_id,
+                        json.dumps(event.data),
+                        event.timestamp,
+                    ),
+                )
+                conn.commit()
+            except Exception as exc:
+                logger.debug("Failed to persist event: %s", exc)
 
     # ── Replay ───────────────────────────────────────────────────────────
 
@@ -165,45 +163,43 @@ class EventBus:
         Returns:
             List of Event objects from persistent storage.
         """
-        conn = self._get_persistence_conn()
-        if conn is None:
-            return []
+        with self._get_persistence_conn() as conn:
+            if conn is None:
+                return []
 
-        try:
-            conditions = []
-            params: list = []
+            try:
+                conditions = []
+                params: list = []
 
-            if event_type:
-                conditions.append("event_type = ?")
-                params.append(event_type)
-            if source:
-                conditions.append("source = ?")
-                params.append(source)
-            if task_id:
-                conditions.append("task_id = ?")
-                params.append(task_id)
+                if event_type:
+                    conditions.append("event_type = ?")
+                    params.append(event_type)
+                if source:
+                    conditions.append("source = ?")
+                    params.append(source)
+                if task_id:
+                    conditions.append("task_id = ?")
+                    params.append(task_id)
 
-            where = " AND ".join(conditions) if conditions else "1=1"
-            sql = f"SELECT * FROM events WHERE {where} ORDER BY timestamp DESC LIMIT ?"
-            params.append(limit)
+                where = " AND ".join(conditions) if conditions else "1=1"
+                sql = f"SELECT * FROM events WHERE {where} ORDER BY timestamp DESC LIMIT ?"
+                params.append(limit)
 
-            rows = conn.execute(sql, params).fetchall()
+                rows = conn.execute(sql, params).fetchall()
 
-            return [
-                Event(
-                    id=row["id"],
-                    type=row["event_type"],
-                    source=row["source"],
-                    data=json.loads(row["data"]) if row["data"] else {},
-                    timestamp=row["timestamp"],
-                )
-                for row in rows
-            ]
-        except Exception as exc:
-            logger.debug("Failed to replay events: %s", exc)
-            return []
-        finally:
-            conn.close()
+                return [
+                    Event(
+                        id=row["id"],
+                        type=row["event_type"],
+                        source=row["source"],
+                        data=json.loads(row["data"]) if row["data"] else {},
+                        timestamp=row["timestamp"],
+                    )
+                    for row in rows
+                ]
+            except Exception as exc:
+                logger.debug("Failed to replay events: %s", exc)
+                return []
 
     # ── Subscribe / Publish ──────────────────────────────────────────────
 

src/infrastructure/hands/shell.py

@@ -211,7 +211,7 @@ class ShellHand:
                 )
 
             latency = (time.time() - start) * 1000
-            exit_code = proc.returncode or 0
+            exit_code = proc.returncode if proc.returncode is not None else -1
             stdout = stdout_bytes.decode("utf-8", errors="replace").strip()
             stderr = stderr_bytes.decode("utf-8", errors="replace").strip()
 

src/infrastructure/models/multimodal.py

@@ -13,7 +13,7 @@ import logging
 from dataclasses import dataclass, field
 from enum import Enum, auto
 
-from config import settings
+from config import normalize_ollama_url, settings
 
 logger = logging.getLogger(__name__)
 
@@ -93,18 +93,6 @@ KNOWN_MODEL_CAPABILITIES: dict[str, set[ModelCapability]] = {
         ModelCapability.VISION,
     },
     # Qwen series
-    "qwen3.5": {
-        ModelCapability.TEXT,
-        ModelCapability.TOOLS,
-        ModelCapability.JSON,
-        ModelCapability.STREAMING,
-    },
-    "qwen3.5:latest": {
-        ModelCapability.TEXT,
-        ModelCapability.TOOLS,
-        ModelCapability.JSON,
-        ModelCapability.STREAMING,
-    },
     "qwen2.5": {
         ModelCapability.TEXT,
         ModelCapability.TOOLS,
@@ -271,9 +259,8 @@ DEFAULT_FALLBACK_CHAINS: dict[ModelCapability, list[str]] = {
     ],
     ModelCapability.TOOLS: [
         "llama3.1:8b-instruct",  # Best tool use
-        "qwen3.5:latest",  # Qwen 3.5 — strong tool use
-        "llama3.2:3b",  # Smaller but capable
         "qwen2.5:7b",  # Reliable fallback
+        "llama3.2:3b",  # Smaller but capable
     ],
     ModelCapability.AUDIO: [
         # Audio models are less common in Ollama
@@ -320,7 +307,7 @@ class MultiModalManager:
             import json
             import urllib.request
 
-            url = self.ollama_url.replace("localhost", "127.0.0.1")
+            url = normalize_ollama_url(self.ollama_url)
             req = urllib.request.Request(
                 f"{url}/api/tags",
                 method="GET",
@@ -475,7 +462,7 @@ class MultiModalManager:
 
             logger.info("Pulling model: %s", model_name)
 
-            url = self.ollama_url.replace("localhost", "127.0.0.1")
+            url = normalize_ollama_url(self.ollama_url)
             req = urllib.request.Request(
                 f"{url}/api/pull",
                 method="POST",

src/infrastructure/models/registry.py

@@ -11,6 +11,8 @@ model roles (student, teacher, judge/PRM) run on dedicated resources.
 import logging
 import sqlite3
 import threading
+from collections.abc import Generator
+from contextlib import closing, contextmanager
 from dataclasses import dataclass
 from datetime import UTC, datetime
 from enum import StrEnum
@@ -60,36 +62,37 @@ class CustomModel:
             self.registered_at = datetime.now(UTC).isoformat()
 
 
-def _get_conn() -> sqlite3.Connection:
+@contextmanager
+def _get_conn() -> Generator[sqlite3.Connection, None, None]:
     DB_PATH.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(DB_PATH))
-    conn.row_factory = sqlite3.Row
-    conn.execute("PRAGMA journal_mode=WAL")
-    conn.execute("PRAGMA busy_timeout=5000")
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS custom_models (
-            name            TEXT PRIMARY KEY,
-            format          TEXT NOT NULL,
-            path            TEXT NOT NULL,
-            role            TEXT NOT NULL DEFAULT 'general',
-            context_window  INTEGER NOT NULL DEFAULT 4096,
-            description     TEXT NOT NULL DEFAULT '',
-            registered_at   TEXT NOT NULL,
-            active          INTEGER NOT NULL DEFAULT 1,
-            default_temperature REAL NOT NULL DEFAULT 0.7,
-            max_tokens      INTEGER NOT NULL DEFAULT 2048
-        )
-        """)
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS agent_model_assignments (
-            agent_id    TEXT PRIMARY KEY,
-            model_name  TEXT NOT NULL,
-            assigned_at TEXT NOT NULL,
-            FOREIGN KEY (model_name) REFERENCES custom_models(name)
-        )
-        """)
-    conn.commit()
-    return conn
+    with closing(sqlite3.connect(str(DB_PATH))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("PRAGMA journal_mode=WAL")
+        conn.execute("PRAGMA busy_timeout=5000")
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS custom_models (
+                name            TEXT PRIMARY KEY,
+                format          TEXT NOT NULL,
+                path            TEXT NOT NULL,
+                role            TEXT NOT NULL DEFAULT 'general',
+                context_window  INTEGER NOT NULL DEFAULT 4096,
+                description     TEXT NOT NULL DEFAULT '',
+                registered_at   TEXT NOT NULL,
+                active          INTEGER NOT NULL DEFAULT 1,
+                default_temperature REAL NOT NULL DEFAULT 0.7,
+                max_tokens      INTEGER NOT NULL DEFAULT 2048
+            )
+            """)
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS agent_model_assignments (
+                agent_id    TEXT PRIMARY KEY,
+                model_name  TEXT NOT NULL,
+                assigned_at TEXT NOT NULL,
+                FOREIGN KEY (model_name) REFERENCES custom_models(name)
+            )
+            """)
+        conn.commit()
+        yield conn
 
 
 class ModelRegistry:
@@ -105,23 +108,22 @@ class ModelRegistry:
     def _load_from_db(self) -> None:
         """Bootstrap cache from SQLite."""
         try:
-            conn = _get_conn()
-            for row in conn.execute("SELECT * FROM custom_models WHERE active = 1").fetchall():
-                self._models[row["name"]] = CustomModel(
-                    name=row["name"],
-                    format=ModelFormat(row["format"]),
-                    path=row["path"],
-                    role=ModelRole(row["role"]),
-                    context_window=row["context_window"],
-                    description=row["description"],
-                    registered_at=row["registered_at"],
-                    active=bool(row["active"]),
-                    default_temperature=row["default_temperature"],
-                    max_tokens=row["max_tokens"],
-                )
-            for row in conn.execute("SELECT * FROM agent_model_assignments").fetchall():
-                self._agent_assignments[row["agent_id"]] = row["model_name"]
-            conn.close()
+            with _get_conn() as conn:
+                for row in conn.execute("SELECT * FROM custom_models WHERE active = 1").fetchall():
+                    self._models[row["name"]] = CustomModel(
+                        name=row["name"],
+                        format=ModelFormat(row["format"]),
+                        path=row["path"],
+                        role=ModelRole(row["role"]),
+                        context_window=row["context_window"],
+                        description=row["description"],
+                        registered_at=row["registered_at"],
+                        active=bool(row["active"]),
+                        default_temperature=row["default_temperature"],
+                        max_tokens=row["max_tokens"],
+                    )
+                for row in conn.execute("SELECT * FROM agent_model_assignments").fetchall():
+                    self._agent_assignments[row["agent_id"]] = row["model_name"]
         except Exception as exc:
             logger.warning("Failed to load model registry from DB: %s", exc)
 
@@ -130,29 +132,28 @@ class ModelRegistry:
     def register(self, model: CustomModel) -> CustomModel:
         """Register a new custom model."""
         with self._lock:
-            conn = _get_conn()
-            conn.execute(
-                """
-                INSERT OR REPLACE INTO custom_models
-                    (name, format, path, role, context_window, description,
-                     registered_at, active, default_temperature, max_tokens)
-                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
-                """,
-                (
-                    model.name,
-                    model.format.value,
-                    model.path,
-                    model.role.value,
-                    model.context_window,
-                    model.description,
-                    model.registered_at,
-                    int(model.active),
-                    model.default_temperature,
-                    model.max_tokens,
-                ),
-            )
-            conn.commit()
-            conn.close()
+            with _get_conn() as conn:
+                conn.execute(
+                    """
+                    INSERT OR REPLACE INTO custom_models
+                        (name, format, path, role, context_window, description,
+                         registered_at, active, default_temperature, max_tokens)
+                    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                    """,
+                    (
+                        model.name,
+                        model.format.value,
+                        model.path,
+                        model.role.value,
+                        model.context_window,
+                        model.description,
+                        model.registered_at,
+                        int(model.active),
+                        model.default_temperature,
+                        model.max_tokens,
+                    ),
+                )
+                conn.commit()
             self._models[model.name] = model
             logger.info("Registered model: %s (%s)", model.name, model.format.value)
             return model
@@ -162,11 +163,10 @@ class ModelRegistry:
         with self._lock:
             if name not in self._models:
                 return False
-            conn = _get_conn()
-            conn.execute("DELETE FROM custom_models WHERE name = ?", (name,))
-            conn.execute("DELETE FROM agent_model_assignments WHERE model_name = ?", (name,))
-            conn.commit()
-            conn.close()
+            with _get_conn() as conn:
+                conn.execute("DELETE FROM custom_models WHERE name = ?", (name,))
+                conn.execute("DELETE FROM agent_model_assignments WHERE model_name = ?", (name,))
+                conn.commit()
             del self._models[name]
             # Remove any agent assignments using this model
             self._agent_assignments = {
@@ -193,13 +193,12 @@ class ModelRegistry:
             return False
         with self._lock:
             model.active = active
-            conn = _get_conn()
-            conn.execute(
-                "UPDATE custom_models SET active = ? WHERE name = ?",
-                (int(active), name),
-            )
-            conn.commit()
-            conn.close()
+            with _get_conn() as conn:
+                conn.execute(
+                    "UPDATE custom_models SET active = ? WHERE name = ?",
+                    (int(active), name),
+                )
+                conn.commit()
         return True
 
     # ── Agent-model assignments ────────────────────────────────────────────
@@ -210,17 +209,16 @@ class ModelRegistry:
             return False
         with self._lock:
             now = datetime.now(UTC).isoformat()
-            conn = _get_conn()
-            conn.execute(
-                """
-                INSERT OR REPLACE INTO agent_model_assignments
-                    (agent_id, model_name, assigned_at)
-                VALUES (?, ?, ?)
-                """,
-                (agent_id, model_name, now),
-            )
-            conn.commit()
-            conn.close()
+            with _get_conn() as conn:
+                conn.execute(
+                    """
+                    INSERT OR REPLACE INTO agent_model_assignments
+                        (agent_id, model_name, assigned_at)
+                    VALUES (?, ?, ?)
+                    """,
+                    (agent_id, model_name, now),
+                )
+                conn.commit()
             self._agent_assignments[agent_id] = model_name
             logger.info("Assigned model %s to agent %s", model_name, agent_id)
         return True
@@ -230,13 +228,12 @@ class ModelRegistry:
         with self._lock:
             if agent_id not in self._agent_assignments:
                 return False
-            conn = _get_conn()
-            conn.execute(
-                "DELETE FROM agent_model_assignments WHERE agent_id = ?",
-                (agent_id,),
-            )
-            conn.commit()
-            conn.close()
+            with _get_conn() as conn:
+                conn.execute(
+                    "DELETE FROM agent_model_assignments WHERE agent_id = ?",
+                    (agent_id,),
+                )
+                conn.commit()
             del self._agent_assignments[agent_id]
         return True
 

src/infrastructure/router/api.py

@@ -183,6 +183,22 @@ async def run_health_check(
     }
 
 
+@router.post("/reload")
+async def reload_config(
+    cascade: Annotated[CascadeRouter, Depends(get_cascade_router)],
+) -> dict[str, Any]:
+    """Hot-reload providers.yaml without restart.
+
+    Preserves circuit breaker state and metrics for existing providers.
+    """
+    try:
+        result = cascade.reload_config()
+        return {"status": "ok", **result}
+    except Exception as exc:
+        logger.error("Config reload failed: %s", exc)
+        raise HTTPException(status_code=500, detail=f"Reload failed: {exc}") from exc
+
+
 @router.get("/config")
 async def get_config(
     cascade: Annotated[CascadeRouter, Depends(get_cascade_router)],

src/infrastructure/router/cascade.py

@@ -18,6 +18,8 @@ from enum import Enum
 from pathlib import Path
 from typing import Any
 
+from config import settings
+
 try:
     import yaml
 except ImportError:
@@ -100,7 +102,7 @@ class Provider:
     """LLM provider configuration and state."""
 
     name: str
-    type: str  # ollama, openai, anthropic, airllm
+    type: str  # ollama, openai, anthropic
     enabled: bool
     priority: int
     url: str | None = None
@@ -301,19 +303,11 @@ class CascadeRouter:
                 # Can't check without requests, assume available
                 return True
             try:
-                url = provider.url or "http://localhost:11434"
+                url = provider.url or settings.ollama_url
                 response = requests.get(f"{url}/api/tags", timeout=5)
                 return response.status_code == 200
-            except Exception:
-                return False
-
-        elif provider.type == "airllm":
-            # Check if airllm is installed
-            try:
-                import importlib.util
-
-                return importlib.util.find_spec("airllm") is not None
-            except (ImportError, ModuleNotFoundError):
+            except Exception as exc:
+                logger.debug("Ollama provider check error: %s", exc)
                 return False
 
         elif provider.type in ("openai", "anthropic", "grok"):
@@ -394,6 +388,101 @@ class CascadeRouter:
 
         return None
 
+    def _select_model(
+        self, provider: Provider, model: str | None, content_type: ContentType
+    ) -> tuple[str | None, bool]:
+        """Select the best model for the request, with vision fallback.
+
+        Returns:
+            Tuple of (selected_model, is_fallback_model).
+        """
+        selected_model = model or provider.get_default_model()
+        is_fallback = False
+
+        if content_type != ContentType.TEXT and selected_model:
+            if provider.type == "ollama" and self._mm_manager:
+                from infrastructure.models.multimodal import ModelCapability
+
+                if content_type == ContentType.VISION:
+                    supports = self._mm_manager.model_supports(
+                        selected_model, ModelCapability.VISION
+                    )
+                    if not supports:
+                        fallback = self._get_fallback_model(provider, selected_model, content_type)
+                        if fallback:
+                            logger.info(
+                                "Model %s doesn't support vision, falling back to %s",
+                                selected_model,
+                                fallback,
+                            )
+                            selected_model = fallback
+                            is_fallback = True
+                        else:
+                            logger.warning(
+                                "No vision-capable model found on %s, trying anyway",
+                                provider.name,
+                            )
+
+        return selected_model, is_fallback
+
+    async def _attempt_with_retry(
+        self,
+        provider: Provider,
+        messages: list[dict],
+        model: str | None,
+        temperature: float,
+        max_tokens: int | None,
+        content_type: ContentType,
+    ) -> dict:
+        """Try a provider with retries, returning the result dict.
+
+        Raises:
+            RuntimeError: If all retry attempts fail.
+            Returns error strings collected during retries via the exception message.
+        """
+        errors: list[str] = []
+        for attempt in range(self.config.max_retries_per_provider):
+            try:
+                return await self._try_provider(
+                    provider=provider,
+                    messages=messages,
+                    model=model,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                    content_type=content_type,
+                )
+            except Exception as exc:
+                error_msg = str(exc)
+                logger.warning(
+                    "Provider %s attempt %d failed: %s",
+                    provider.name,
+                    attempt + 1,
+                    error_msg,
+                )
+                errors.append(f"{provider.name}: {error_msg}")
+
+                if attempt < self.config.max_retries_per_provider - 1:
+                    await asyncio.sleep(self.config.retry_delay_seconds)
+
+        raise RuntimeError("; ".join(errors))
+
+    def _is_provider_available(self, provider: Provider) -> bool:
+        """Check if a provider should be tried (enabled + circuit breaker)."""
+        if not provider.enabled:
+            logger.debug("Skipping %s (disabled)", provider.name)
+            return False
+
+        if provider.status == ProviderStatus.UNHEALTHY:
+            if self._can_close_circuit(provider):
+                provider.circuit_state = CircuitState.HALF_OPEN
+                provider.half_open_calls = 0
+                logger.info("Circuit breaker half-open for %s", provider.name)
+            else:
+                logger.debug("Skipping %s (circuit open)", provider.name)
+                return False
+
+        return True
+
     async def complete(
         self,
         messages: list[dict],
@@ -420,7 +509,6 @@ class CascadeRouter:
         Raises:
             RuntimeError: If all providers fail
         """
-        # Detect content type for multi-modal routing
         content_type = self._detect_content_type(messages)
         if content_type != ContentType.TEXT:
             logger.debug("Detected %s content, selecting appropriate model", content_type.value)
@@ -428,93 +516,34 @@ class CascadeRouter:
         errors = []
 
         for provider in self.providers:
-            # Skip disabled providers
-            if not provider.enabled:
-                logger.debug("Skipping %s (disabled)", provider.name)
+            if not self._is_provider_available(provider):
                 continue
 
-            # Skip unhealthy providers (circuit breaker)
-            if provider.status == ProviderStatus.UNHEALTHY:
-                # Check if circuit breaker can close
-                if self._can_close_circuit(provider):
-                    provider.circuit_state = CircuitState.HALF_OPEN
-                    provider.half_open_calls = 0
-                    logger.info("Circuit breaker half-open for %s", provider.name)
-                else:
-                    logger.debug("Skipping %s (circuit open)", provider.name)
-                    continue
+            selected_model, is_fallback_model = self._select_model(provider, model, content_type)
 
-            # Determine which model to use
-            selected_model = model or provider.get_default_model()
-            is_fallback_model = False
+            try:
+                result = await self._attempt_with_retry(
+                    provider,
+                    messages,
+                    selected_model,
+                    temperature,
+                    max_tokens,
+                    content_type,
+                )
+            except RuntimeError as exc:
+                errors.append(str(exc))
+                self._record_failure(provider)
+                continue
 
-            # For non-text content, check if model supports it
-            if content_type != ContentType.TEXT and selected_model:
-                if provider.type == "ollama" and self._mm_manager:
-                    from infrastructure.models.multimodal import ModelCapability
+            self._record_success(provider, result.get("latency_ms", 0))
+            return {
+                "content": result["content"],
+                "provider": provider.name,
+                "model": result.get("model", selected_model or provider.get_default_model()),
+                "latency_ms": result.get("latency_ms", 0),
+                "is_fallback_model": is_fallback_model,
+            }
 
-                    # Check if selected model supports the required capability
-                    if content_type == ContentType.VISION:
-                        supports = self._mm_manager.model_supports(
-                            selected_model, ModelCapability.VISION
-                        )
-                        if not supports:
-                            # Find fallback model
-                            fallback = self._get_fallback_model(
-                                provider, selected_model, content_type
-                            )
-                            if fallback:
-                                logger.info(
-                                    "Model %s doesn't support vision, falling back to %s",
-                                    selected_model,
-                                    fallback,
-                                )
-                                selected_model = fallback
-                                is_fallback_model = True
-                            else:
-                                logger.warning(
-                                    "No vision-capable model found on %s, trying anyway",
-                                    provider.name,
-                                )
-
-            # Try this provider
-            for attempt in range(self.config.max_retries_per_provider):
-                try:
-                    result = await self._try_provider(
-                        provider=provider,
-                        messages=messages,
-                        model=selected_model,
-                        temperature=temperature,
-                        max_tokens=max_tokens,
-                        content_type=content_type,
-                    )
-
-                    # Success! Update metrics and return
-                    self._record_success(provider, result.get("latency_ms", 0))
-                    return {
-                        "content": result["content"],
-                        "provider": provider.name,
-                        "model": result.get(
-                            "model", selected_model or provider.get_default_model()
-                        ),
-                        "latency_ms": result.get("latency_ms", 0),
-                        "is_fallback_model": is_fallback_model,
-                    }
-
-                except Exception as exc:
-                    error_msg = str(exc)
-                    logger.warning(
-                        "Provider %s attempt %d failed: %s", provider.name, attempt + 1, error_msg
-                    )
-                    errors.append(f"{provider.name}: {error_msg}")
-
-                    if attempt < self.config.max_retries_per_provider - 1:
-                        await asyncio.sleep(self.config.retry_delay_seconds)
-
-            # All retries failed for this provider
-            self._record_failure(provider)
-
-        # All providers failed
         raise RuntimeError(f"All providers failed: {'; '.join(errors)}")
 
     async def _try_provider(
@@ -580,7 +609,7 @@ class CascadeRouter:
         """Call Ollama API with multi-modal support."""
         import aiohttp
 
-        url = f"{provider.url}/api/chat"
+        url = f"{provider.url or settings.ollama_url}/api/chat"
 
         # Transform messages for Ollama format (including images)
         transformed_messages = self._transform_messages_for_ollama(messages)
@@ -814,6 +843,66 @@ class CascadeRouter:
         provider.status = ProviderStatus.HEALTHY
         logger.info("Circuit breaker CLOSED for %s", provider.name)
 
+    def reload_config(self) -> dict:
+        """Hot-reload providers.yaml, preserving runtime state.
+
+        Re-reads the config file, rebuilds the provider list, and
+        preserves circuit breaker state and metrics for providers
+        that still exist after reload.
+
+        Returns:
+            Summary dict with added/removed/preserved counts.
+        """
+        # Snapshot current runtime state keyed by provider name
+        old_state: dict[
+            str, tuple[ProviderMetrics, CircuitState, float | None, int, ProviderStatus]
+        ] = {}
+        for p in self.providers:
+            old_state[p.name] = (
+                p.metrics,
+                p.circuit_state,
+                p.circuit_opened_at,
+                p.half_open_calls,
+                p.status,
+            )
+
+        old_names = set(old_state.keys())
+
+        # Reload from disk
+        self.providers = []
+        self._load_config()
+
+        # Restore preserved state
+        new_names = {p.name for p in self.providers}
+        preserved = 0
+        for p in self.providers:
+            if p.name in old_state:
+                metrics, circuit, opened_at, half_open, status = old_state[p.name]
+                p.metrics = metrics
+                p.circuit_state = circuit
+                p.circuit_opened_at = opened_at
+                p.half_open_calls = half_open
+                p.status = status
+                preserved += 1
+
+        added = new_names - old_names
+        removed = old_names - new_names
+
+        logger.info(
+            "Config reloaded: %d providers (%d preserved, %d added, %d removed)",
+            len(self.providers),
+            preserved,
+            len(added),
+            len(removed),
+        )
+
+        return {
+            "total_providers": len(self.providers),
+            "preserved": preserved,
+            "added": sorted(added),
+            "removed": sorted(removed),
+        }
+
     def get_metrics(self) -> dict:
         """Get metrics for all providers."""
         return {

src/infrastructure/ws_manager/handler.py

@@ -54,7 +54,8 @@ class WebSocketManager:
         for event in list(self._event_history)[-20:]:
             try:
                 await websocket.send_text(event.to_json())
-            except Exception:
+            except Exception as exc:
+                logger.warning("WebSocket history send error: %s", exc)
                 break
 
     def disconnect(self, websocket: WebSocket) -> None:
@@ -83,8 +84,8 @@ class WebSocketManager:
                 await ws.send_text(message)
             except ConnectionError:
                 disconnected.append(ws)
-            except Exception:
-                logger.warning("Unexpected WebSocket send error", exc_info=True)
+            except Exception as exc:
+                logger.warning("Unexpected WebSocket send error: %s", exc)
                 disconnected.append(ws)
 
         # Clean up dead connections
@@ -156,7 +157,8 @@ class WebSocketManager:
             try:
                 await ws.send_text(message)
                 count += 1
-            except Exception:
+            except Exception as exc:
+                logger.warning("WebSocket direct send error: %s", exc)
                 disconnected.append(ws)
 
         # Clean up dead connections

src/integrations/chat_bridge/vendors/discord.py

@@ -87,7 +87,8 @@ if _DISCORD_UI_AVAILABLE:
                 await action["target"].send(
                     f"Action `{action['tool_name']}` timed out and was auto-rejected."
                 )
-            except Exception:
+            except Exception as exc:
+                logger.warning("Discord action timeout message error: %s", exc)
                 pass
 
 
@@ -186,7 +187,8 @@ class DiscordVendor(ChatPlatform):
         if self._client and not self._client.is_closed():
             try:
                 await self._client.close()
-            except Exception:
+            except Exception as exc:
+                logger.warning("Discord client close error: %s", exc)
                 pass
             self._client = None
 
@@ -330,7 +332,8 @@ class DiscordVendor(ChatPlatform):
 
             if settings.discord_token:
                 return settings.discord_token
-        except Exception:
+        except Exception as exc:
+            logger.warning("Discord token load error: %s", exc)
             pass
 
         # 2. Fall back to state file (set via /discord/setup endpoint)
@@ -458,7 +461,8 @@ class DiscordVendor(ChatPlatform):
         req.reject(note="User rejected from Discord")
         try:
             await continue_chat(action["run_output"], action.get("session_id"))
-        except Exception:
+        except Exception as exc:
+            logger.warning("Discord continue chat error: %s", exc)
             pass
 
         await interaction.response.send_message(
@@ -543,9 +547,7 @@ class DiscordVendor(ChatPlatform):
             response = "Sorry, that took too long. Please try a simpler request."
         except Exception as exc:
             logger.error("Discord: chat_with_tools() failed: %s", exc)
-            response = (
-                "I'm having trouble reaching my language model right now. Please try again shortly."
-            )
+            response = "I'm having trouble reaching my inference backend right now. Please try again shortly."
 
         # Check if Agno paused the run for tool confirmation
         if run_output is not None:

src/integrations/telegram_bot/bot.py

@@ -56,7 +56,8 @@ class TelegramBot:
             from config import settings
 
             return settings.telegram_token or None
-        except Exception:
+        except Exception as exc:
+            logger.warning("Telegram token load error: %s", exc)
             return None
 
     def save_token(self, token: str) -> None:

src/loop/__init__.py

@@ -0,0 +1 @@
+"""Three-phase agent loop: Gather → Reason → Act."""

src/loop/phase1_gather.py

@@ -0,0 +1,37 @@
+"""Phase 1 — Gather: accept raw input, produce structured context.
+
+This is the sensory phase. It receives a raw ContextPayload and enriches
+it with whatever context Timmy needs before reasoning. In the stub form,
+it simply passes the payload through with a phase marker.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from loop.schema import ContextPayload
+
+logger = logging.getLogger(__name__)
+
+
+def gather(payload: ContextPayload) -> ContextPayload:
+    """Accept raw input and return structured context for reasoning.
+
+    Stub: tags the payload with phase=gather and logs transit.
+    Timmy will flesh this out with context selection, memory lookup,
+    adapter polling, and attention-residual weighting.
+    """
+    logger.info(
+        "Phase 1 (Gather) received: source=%s content_len=%d tokens=%d",
+        payload.source,
+        len(payload.content),
+        payload.token_count,
+    )
+
+    result = payload.with_metadata(phase="gather", gathered=True)
+
+    logger.info(
+        "Phase 1 (Gather) produced: metadata_keys=%s",
+        sorted(result.metadata.keys()),
+    )
+    return result

src/loop/phase2_reason.py

@@ -0,0 +1,36 @@
+"""Phase 2 — Reason: accept gathered context, produce reasoning output.
+
+This is the deliberation phase. It receives enriched context from Phase 1
+and decides what to do. In the stub form, it passes the payload through
+with a phase marker.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from loop.schema import ContextPayload
+
+logger = logging.getLogger(__name__)
+
+
+def reason(payload: ContextPayload) -> ContextPayload:
+    """Accept gathered context and return a reasoning result.
+
+    Stub: tags the payload with phase=reason and logs transit.
+    Timmy will flesh this out with LLM calls, confidence scoring,
+    plan generation, and judgment logic.
+    """
+    logger.info(
+        "Phase 2 (Reason) received: source=%s gathered=%s",
+        payload.source,
+        payload.metadata.get("gathered", False),
+    )
+
+    result = payload.with_metadata(phase="reason", reasoned=True)
+
+    logger.info(
+        "Phase 2 (Reason) produced: metadata_keys=%s",
+        sorted(result.metadata.keys()),
+    )
+    return result

src/loop/phase3_act.py

@@ -0,0 +1,36 @@
+"""Phase 3 — Act: accept reasoning output, execute and produce feedback.
+
+This is the command phase. It receives the reasoning result from Phase 2
+and takes action. In the stub form, it passes the payload through with a
+phase marker and produces feedback for the next cycle.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from loop.schema import ContextPayload
+
+logger = logging.getLogger(__name__)
+
+
+def act(payload: ContextPayload) -> ContextPayload:
+    """Accept reasoning result and return action output + feedback.
+
+    Stub: tags the payload with phase=act and logs transit.
+    Timmy will flesh this out with tool execution, delegation,
+    response generation, and feedback construction.
+    """
+    logger.info(
+        "Phase 3 (Act) received: source=%s reasoned=%s",
+        payload.source,
+        payload.metadata.get("reasoned", False),
+    )
+
+    result = payload.with_metadata(phase="act", acted=True)
+
+    logger.info(
+        "Phase 3 (Act) produced: metadata_keys=%s",
+        sorted(result.metadata.keys()),
+    )
+    return result

src/loop/runner.py

@@ -0,0 +1,40 @@
+"""Loop runner — orchestrates the three phases in sequence.
+
+Runs Gather → Reason → Act as a single cycle, passing output from each
+phase as input to the next. The Act output feeds back as input to the
+next Gather call.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from loop.phase1_gather import gather
+from loop.phase2_reason import reason
+from loop.phase3_act import act
+from loop.schema import ContextPayload
+
+logger = logging.getLogger(__name__)
+
+
+def run_cycle(payload: ContextPayload) -> ContextPayload:
+    """Execute one full Gather → Reason → Act cycle.
+
+    Returns the Act phase output, which can be fed back as input
+    to the next cycle.
+    """
+    logger.info("=== Loop cycle start: source=%s ===", payload.source)
+
+    gathered = gather(payload)
+    reasoned = reason(gathered)
+    acted = act(reasoned)
+
+    logger.info(
+        "=== Loop cycle complete: phases=%s ===",
+        [
+            gathered.metadata.get("phase"),
+            reasoned.metadata.get("phase"),
+            acted.metadata.get("phase"),
+        ],
+    )
+    return acted

src/loop/schema.py

@@ -0,0 +1,43 @@
+"""Data schema for the three-phase loop.
+
+Each phase passes a ContextPayload forward. The schema is intentionally
+minimal — Timmy decides what fields matter as the loop matures.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ContextPayload:
+    """Immutable context packet passed between loop phases.
+
+    Attributes:
+        source: Where this payload originated (e.g. "user", "timer", "event").
+        content: The raw content string to process.
+        timestamp: When the payload was created.
+        token_count: Estimated token count for budget tracking. -1 = unknown.
+        metadata: Arbitrary key-value pairs for phase-specific data.
+    """
+
+    source: str
+    content: str
+    timestamp: datetime = field(default_factory=lambda: datetime.now(UTC))
+    token_count: int = -1
+    metadata: dict = field(default_factory=dict)
+
+    def with_metadata(self, **kwargs: object) -> ContextPayload:
+        """Return a new payload with additional metadata merged in."""
+        merged = {**self.metadata, **kwargs}
+        return ContextPayload(
+            source=self.source,
+            content=self.content,
+            timestamp=self.timestamp,
+            token_count=self.token_count,
+            metadata=merged,
+        )

src/spark/eidos.py

@@ -16,6 +16,8 @@ import json
 import logging
 import sqlite3
 import uuid
+from collections.abc import Generator
+from contextlib import closing, contextmanager
 from dataclasses import dataclass
 from datetime import UTC, datetime
 from pathlib import Path
@@ -39,28 +41,31 @@ class Prediction:
     evaluated_at: str | None
 
 
-def _get_conn() -> sqlite3.Connection:
+@contextmanager
+def _get_conn() -> Generator[sqlite3.Connection, None, None]:
     DB_PATH.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(DB_PATH))
-    conn.row_factory = sqlite3.Row
-    conn.execute("PRAGMA journal_mode=WAL")
-    conn.execute("PRAGMA busy_timeout=5000")
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS spark_predictions (
-            id               TEXT PRIMARY KEY,
-            task_id          TEXT NOT NULL,
-            prediction_type  TEXT NOT NULL,
-            predicted_value  TEXT NOT NULL,
-            actual_value     TEXT,
-            accuracy         REAL,
-            created_at       TEXT NOT NULL,
-            evaluated_at     TEXT
+    with closing(sqlite3.connect(str(DB_PATH))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("PRAGMA journal_mode=WAL")
+        conn.execute("PRAGMA busy_timeout=5000")
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS spark_predictions (
+                id               TEXT PRIMARY KEY,
+                task_id          TEXT NOT NULL,
+                prediction_type  TEXT NOT NULL,
+                predicted_value  TEXT NOT NULL,
+                actual_value     TEXT,
+                accuracy         REAL,
+                created_at       TEXT NOT NULL,
+                evaluated_at     TEXT
+            )
+            """)
+        conn.execute("CREATE INDEX IF NOT EXISTS idx_pred_task ON spark_predictions(task_id)")
+        conn.execute(
+            "CREATE INDEX IF NOT EXISTS idx_pred_type ON spark_predictions(prediction_type)"
         )
-        """)
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_pred_task ON spark_predictions(task_id)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_pred_type ON spark_predictions(prediction_type)")
-    conn.commit()
-    return conn
+        conn.commit()
+        yield conn
 
 
 # ── Prediction phase ────────────────────────────────────────────────────────
@@ -119,17 +124,16 @@ def predict_task_outcome(
     # Store prediction
     pred_id = str(uuid.uuid4())
     now = datetime.now(UTC).isoformat()
-    conn = _get_conn()
-    conn.execute(
-        """
-        INSERT INTO spark_predictions
-            (id, task_id, prediction_type, predicted_value, created_at)
-        VALUES (?, ?, ?, ?, ?)
-        """,
-        (pred_id, task_id, "outcome", json.dumps(prediction), now),
-    )
-    conn.commit()
-    conn.close()
+    with _get_conn() as conn:
+        conn.execute(
+            """
+            INSERT INTO spark_predictions
+                (id, task_id, prediction_type, predicted_value, created_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (pred_id, task_id, "outcome", json.dumps(prediction), now),
+        )
+        conn.commit()
 
     prediction["prediction_id"] = pred_id
     return prediction
@@ -148,41 +152,39 @@ def evaluate_prediction(
 
     Returns the evaluation result or None if no prediction exists.
     """
-    conn = _get_conn()
-    row = conn.execute(
-        """
-        SELECT * FROM spark_predictions
-        WHERE task_id = ? AND prediction_type = 'outcome' AND evaluated_at IS NULL
-        ORDER BY created_at DESC LIMIT 1
-        """,
-        (task_id,),
-    ).fetchone()
+    with _get_conn() as conn:
+        row = conn.execute(
+            """
+            SELECT * FROM spark_predictions
+            WHERE task_id = ? AND prediction_type = 'outcome' AND evaluated_at IS NULL
+            ORDER BY created_at DESC LIMIT 1
+            """,
+            (task_id,),
+        ).fetchone()
 
-    if not row:
-        conn.close()
-        return None
+        if not row:
+            return None
 
-    predicted = json.loads(row["predicted_value"])
-    actual = {
-        "winner": actual_winner,
-        "succeeded": task_succeeded,
-        "winning_bid": winning_bid,
-    }
+        predicted = json.loads(row["predicted_value"])
+        actual = {
+            "winner": actual_winner,
+            "succeeded": task_succeeded,
+            "winning_bid": winning_bid,
+        }
 
-    # Calculate accuracy
-    accuracy = _compute_accuracy(predicted, actual)
-    now = datetime.now(UTC).isoformat()
+        # Calculate accuracy
+        accuracy = _compute_accuracy(predicted, actual)
+        now = datetime.now(UTC).isoformat()
 
-    conn.execute(
-        """
-        UPDATE spark_predictions
-        SET actual_value = ?, accuracy = ?, evaluated_at = ?
-        WHERE id = ?
-        """,
-        (json.dumps(actual), accuracy, now, row["id"]),
-    )
-    conn.commit()
-    conn.close()
+        conn.execute(
+            """
+            UPDATE spark_predictions
+            SET actual_value = ?, accuracy = ?, evaluated_at = ?
+            WHERE id = ?
+            """,
+            (json.dumps(actual), accuracy, now, row["id"]),
+        )
+        conn.commit()
 
     return {
         "prediction_id": row["id"],
@@ -243,7 +245,6 @@ def get_predictions(
     limit: int = 50,
 ) -> list[Prediction]:
     """Query stored predictions."""
-    conn = _get_conn()
     query = "SELECT * FROM spark_predictions WHERE 1=1"
     params: list = []
 
@@ -256,8 +257,8 @@ def get_predictions(
     query += " ORDER BY created_at DESC LIMIT ?"
     params.append(limit)
 
-    rows = conn.execute(query, params).fetchall()
-    conn.close()
+    with _get_conn() as conn:
+        rows = conn.execute(query, params).fetchall()
     return [
         Prediction(
             id=r["id"],
@@ -275,17 +276,16 @@ def get_predictions(
 
 def get_accuracy_stats() -> dict:
     """Return aggregate accuracy statistics for the EIDOS loop."""
-    conn = _get_conn()
-    row = conn.execute("""
-        SELECT
-            COUNT(*)                          AS total_predictions,
-            COUNT(evaluated_at)               AS evaluated,
-            AVG(CASE WHEN accuracy IS NOT NULL THEN accuracy END) AS avg_accuracy,
-            MIN(CASE WHEN accuracy IS NOT NULL THEN accuracy END) AS min_accuracy,
-            MAX(CASE WHEN accuracy IS NOT NULL THEN accuracy END) AS max_accuracy
-        FROM spark_predictions
-        """).fetchone()
-    conn.close()
+    with _get_conn() as conn:
+        row = conn.execute("""
+            SELECT
+                COUNT(*)                          AS total_predictions,
+                COUNT(evaluated_at)               AS evaluated,
+                AVG(CASE WHEN accuracy IS NOT NULL THEN accuracy END) AS avg_accuracy,
+                MIN(CASE WHEN accuracy IS NOT NULL THEN accuracy END) AS min_accuracy,
+                MAX(CASE WHEN accuracy IS NOT NULL THEN accuracy END) AS max_accuracy
+            FROM spark_predictions
+            """).fetchone()
 
     return {
         "total_predictions": row["total_predictions"] or 0,

src/spark/engine.py

@@ -273,6 +273,8 @@ class SparkEngine:
 
     def _maybe_consolidate(self, agent_id: str) -> None:
         """Consolidate events into memories when enough data exists."""
+        from datetime import UTC, datetime, timedelta
+
         agent_events = spark_memory.get_events(agent_id=agent_id, limit=50)
         if len(agent_events) < 5:
             return
@@ -286,7 +288,34 @@ class SparkEngine:
 
         success_rate = len(completions) / total if total else 0
 
+        # Determine target memory type based on success rate
         if success_rate >= 0.8:
+            target_memory_type = "pattern"
+        elif success_rate <= 0.3:
+            target_memory_type = "anomaly"
+        else:
+            return  # No consolidation needed for neutral success rates
+
+        # Check for recent memories of the same type for this agent
+        existing_memories = spark_memory.get_memories(subject=agent_id, limit=5)
+        now = datetime.now(UTC)
+        one_hour_ago = now - timedelta(hours=1)
+
+        for memory in existing_memories:
+            if memory.memory_type == target_memory_type:
+                try:
+                    created_at = datetime.fromisoformat(memory.created_at)
+                    if created_at >= one_hour_ago:
+                        logger.info(
+                            "Consolidation: skipping — recent memory exists for %s",
+                            agent_id[:8],
+                        )
+                        return
+                except (ValueError, TypeError):
+                    continue
+
+        # Store the new memory
+        if target_memory_type == "pattern":
             spark_memory.store_memory(
                 memory_type="pattern",
                 subject=agent_id,
@@ -295,7 +324,7 @@ class SparkEngine:
                 confidence=min(0.95, 0.6 + total * 0.05),
                 source_events=total,
             )
-        elif success_rate <= 0.3:
+        else:  # anomaly
             spark_memory.store_memory(
                 memory_type="anomaly",
                 subject=agent_id,
@@ -358,7 +387,8 @@ def get_spark_engine() -> SparkEngine:
             from config import settings
 
             _spark_engine = SparkEngine(enabled=settings.spark_enabled)
-        except Exception:
+        except Exception as exc:
+            logger.debug("Spark engine settings load error: %s", exc)
             _spark_engine = SparkEngine(enabled=True)
     return _spark_engine
 

src/spark/memory.py

@@ -10,12 +10,17 @@ spark_events   — raw event log (every swarm event)
 spark_memories — consolidated insights extracted from event patterns
 """
 
+import logging
 import sqlite3
 import uuid
+from collections.abc import Generator
+from contextlib import closing, contextmanager
 from dataclasses import dataclass
 from datetime import UTC, datetime
 from pathlib import Path
 
+logger = logging.getLogger(__name__)
+
 DB_PATH = Path("data/spark.db")
 
 # Importance thresholds
@@ -52,42 +57,43 @@ class SparkMemory:
     expires_at: str | None
 
 
-def _get_conn() -> sqlite3.Connection:
+@contextmanager
+def _get_conn() -> Generator[sqlite3.Connection, None, None]:
     DB_PATH.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(DB_PATH))
-    conn.row_factory = sqlite3.Row
-    conn.execute("PRAGMA journal_mode=WAL")
-    conn.execute("PRAGMA busy_timeout=5000")
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS spark_events (
-            id          TEXT PRIMARY KEY,
-            event_type  TEXT NOT NULL,
-            agent_id    TEXT,
-            task_id     TEXT,
-            description TEXT NOT NULL DEFAULT '',
-            data        TEXT NOT NULL DEFAULT '{}',
-            importance  REAL NOT NULL DEFAULT 0.5,
-            created_at  TEXT NOT NULL
-        )
-        """)
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS spark_memories (
-            id              TEXT PRIMARY KEY,
-            memory_type     TEXT NOT NULL,
-            subject         TEXT NOT NULL DEFAULT 'system',
-            content         TEXT NOT NULL,
-            confidence      REAL NOT NULL DEFAULT 0.5,
-            source_events   INTEGER NOT NULL DEFAULT 0,
-            created_at      TEXT NOT NULL,
-            expires_at      TEXT
-        )
-        """)
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_events_type ON spark_events(event_type)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_events_agent ON spark_events(agent_id)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_events_task ON spark_events(task_id)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_memories_subject ON spark_memories(subject)")
-    conn.commit()
-    return conn
+    with closing(sqlite3.connect(str(DB_PATH))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("PRAGMA journal_mode=WAL")
+        conn.execute("PRAGMA busy_timeout=5000")
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS spark_events (
+                id          TEXT PRIMARY KEY,
+                event_type  TEXT NOT NULL,
+                agent_id    TEXT,
+                task_id     TEXT,
+                description TEXT NOT NULL DEFAULT '',
+                data        TEXT NOT NULL DEFAULT '{}',
+                importance  REAL NOT NULL DEFAULT 0.5,
+                created_at  TEXT NOT NULL
+            )
+            """)
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS spark_memories (
+                id              TEXT PRIMARY KEY,
+                memory_type     TEXT NOT NULL,
+                subject         TEXT NOT NULL DEFAULT 'system',
+                content         TEXT NOT NULL,
+                confidence      REAL NOT NULL DEFAULT 0.5,
+                source_events   INTEGER NOT NULL DEFAULT 0,
+                created_at      TEXT NOT NULL,
+                expires_at      TEXT
+            )
+            """)
+        conn.execute("CREATE INDEX IF NOT EXISTS idx_events_type ON spark_events(event_type)")
+        conn.execute("CREATE INDEX IF NOT EXISTS idx_events_agent ON spark_events(agent_id)")
+        conn.execute("CREATE INDEX IF NOT EXISTS idx_events_task ON spark_events(task_id)")
+        conn.execute("CREATE INDEX IF NOT EXISTS idx_memories_subject ON spark_memories(subject)")
+        conn.commit()
+        yield conn
 
 
 # ── Importance scoring ──────────────────────────────────────────────────────
@@ -146,17 +152,16 @@ def record_event(
             parsed = {}
         importance = score_importance(event_type, parsed)
 
-    conn = _get_conn()
-    conn.execute(
-        """
-        INSERT INTO spark_events
-            (id, event_type, agent_id, task_id, description, data, importance, created_at)
-        VALUES (?, ?, ?, ?, ?, ?, ?, ?)
-        """,
-        (event_id, event_type, agent_id, task_id, description, data, importance, now),
-    )
-    conn.commit()
-    conn.close()
+    with _get_conn() as conn:
+        conn.execute(
+            """
+            INSERT INTO spark_events
+                (id, event_type, agent_id, task_id, description, data, importance, created_at)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+            """,
+            (event_id, event_type, agent_id, task_id, description, data, importance, now),
+        )
+        conn.commit()
 
     # Bridge to unified event log so all events are queryable from one place
     try:
@@ -170,7 +175,8 @@ def record_event(
             task_id=task_id or "",
             agent_id=agent_id or "",
         )
-    except Exception:
+    except Exception as exc:
+        logger.debug("Spark event log error: %s", exc)
         pass  # Graceful — don't break spark if event_log is unavailable
 
     return event_id
@@ -184,7 +190,6 @@ def get_events(
     min_importance: float = 0.0,
 ) -> list[SparkEvent]:
     """Query events with optional filters."""
-    conn = _get_conn()
     query = "SELECT * FROM spark_events WHERE importance >= ?"
     params: list = [min_importance]
 
@@ -201,8 +206,8 @@ def get_events(
     query += " ORDER BY created_at DESC LIMIT ?"
     params.append(limit)
 
-    rows = conn.execute(query, params).fetchall()
-    conn.close()
+    with _get_conn() as conn:
+        rows = conn.execute(query, params).fetchall()
     return [
         SparkEvent(
             id=r["id"],
@@ -220,15 +225,14 @@ def get_events(
 
 def count_events(event_type: str | None = None) -> int:
     """Count events, optionally filtered by type."""
-    conn = _get_conn()
-    if event_type:
-        row = conn.execute(
-            "SELECT COUNT(*) FROM spark_events WHERE event_type = ?",
-            (event_type,),
-        ).fetchone()
-    else:
-        row = conn.execute("SELECT COUNT(*) FROM spark_events").fetchone()
-    conn.close()
+    with _get_conn() as conn:
+        if event_type:
+            row = conn.execute(
+                "SELECT COUNT(*) FROM spark_events WHERE event_type = ?",
+                (event_type,),
+            ).fetchone()
+        else:
+            row = conn.execute("SELECT COUNT(*) FROM spark_events").fetchone()
     return row[0]
 
 
@@ -246,17 +250,16 @@ def store_memory(
     """Store a consolidated memory.  Returns the memory id."""
     mem_id = str(uuid.uuid4())
     now = datetime.now(UTC).isoformat()
-    conn = _get_conn()
-    conn.execute(
-        """
-        INSERT INTO spark_memories
-            (id, memory_type, subject, content, confidence, source_events, created_at, expires_at)
-        VALUES (?, ?, ?, ?, ?, ?, ?, ?)
-        """,
-        (mem_id, memory_type, subject, content, confidence, source_events, now, expires_at),
-    )
-    conn.commit()
-    conn.close()
+    with _get_conn() as conn:
+        conn.execute(
+            """
+            INSERT INTO spark_memories
+                (id, memory_type, subject, content, confidence, source_events, created_at, expires_at)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+            """,
+            (mem_id, memory_type, subject, content, confidence, source_events, now, expires_at),
+        )
+        conn.commit()
     return mem_id
 
 
@@ -267,7 +270,6 @@ def get_memories(
     limit: int = 50,
 ) -> list[SparkMemory]:
     """Query memories with optional filters."""
-    conn = _get_conn()
     query = "SELECT * FROM spark_memories WHERE confidence >= ?"
     params: list = [min_confidence]
 
@@ -281,8 +283,8 @@ def get_memories(
     query += " ORDER BY created_at DESC LIMIT ?"
     params.append(limit)
 
-    rows = conn.execute(query, params).fetchall()
-    conn.close()
+    with _get_conn() as conn:
+        rows = conn.execute(query, params).fetchall()
     return [
         SparkMemory(
             id=r["id"],
@@ -300,13 +302,12 @@ def get_memories(
 
 def count_memories(memory_type: str | None = None) -> int:
     """Count memories, optionally filtered by type."""
-    conn = _get_conn()
-    if memory_type:
-        row = conn.execute(
-            "SELECT COUNT(*) FROM spark_memories WHERE memory_type = ?",
-            (memory_type,),
-        ).fetchone()
-    else:
-        row = conn.execute("SELECT COUNT(*) FROM spark_memories").fetchone()
-    conn.close()
+    with _get_conn() as conn:
+        if memory_type:
+            row = conn.execute(
+                "SELECT COUNT(*) FROM spark_memories WHERE memory_type = ?",
+                (memory_type,),
+            ).fetchone()
+        else:
+            row = conn.execute("SELECT COUNT(*) FROM spark_memories").fetchone()
     return row[0]

src/timmy/__init__.py

@@ -1 +1 @@
-"""Timmy — Core AI agent (Ollama/AirLLM backends, CLI, prompts)."""
+"""Timmy — Core AI agent (Ollama/Grok/Claude backends, CLI, prompts)."""

src/timmy/adapters/__init__.py

@@ -0,0 +1 @@
+"""Adapters — normalize external data streams into sensory events."""

src/timmy/adapters/gitea_adapter.py

@@ -0,0 +1,136 @@
+"""Gitea webhook adapter — normalize webhook payloads to event bus events.
+
+Receives raw Gitea webhook payloads and emits typed events via the
+infrastructure event bus.  Bot-only activity is filtered unless it
+represents a PR merge (which is always noteworthy).
+"""
+
+import logging
+from typing import Any
+
+from infrastructure.events.bus import emit
+
+logger = logging.getLogger(__name__)
+
+# Gitea usernames considered "bot" accounts
+BOT_USERNAMES = frozenset({"hermes", "kimi", "manus"})
+
+# Owner username — activity from this user is always emitted
+OWNER_USERNAME = "rockachopa"
+
+# Mapping from Gitea webhook event type to our bus event type
+_EVENT_TYPE_MAP = {
+    "push": "gitea.push",
+    "issues": "gitea.issue.opened",
+    "issue_comment": "gitea.issue.comment",
+    "pull_request": "gitea.pull_request",
+}
+
+
+def _extract_actor(payload: dict[str, Any]) -> str:
+    """Extract the actor username from a webhook payload."""
+    # Gitea puts actor in sender.login for most events
+    sender = payload.get("sender", {})
+    return sender.get("login", "unknown")
+
+
+def _is_bot(username: str) -> bool:
+    return username.lower() in BOT_USERNAMES
+
+
+def _is_pr_merge(event_type: str, payload: dict[str, Any]) -> bool:
+    """Check if this is a pull_request merge event."""
+    if event_type != "pull_request":
+        return False
+    action = payload.get("action", "")
+    pr = payload.get("pull_request", {})
+    return action == "closed" and pr.get("merged", False)
+
+
+def _normalize_push(payload: dict[str, Any], actor: str) -> dict[str, Any]:
+    """Normalize a push event payload."""
+    commits = payload.get("commits", [])
+    return {
+        "actor": actor,
+        "ref": payload.get("ref", ""),
+        "repo": payload.get("repository", {}).get("full_name", ""),
+        "num_commits": len(commits),
+        "head_message": commits[0].get("message", "").split("\n", 1)[0].strip() if commits else "",
+    }
+
+
+def _normalize_issue_opened(payload: dict[str, Any], actor: str) -> dict[str, Any]:
+    """Normalize an issue-opened event payload."""
+    issue = payload.get("issue", {})
+    return {
+        "actor": actor,
+        "action": payload.get("action", "opened"),
+        "repo": payload.get("repository", {}).get("full_name", ""),
+        "issue_number": issue.get("number", 0),
+        "title": issue.get("title", ""),
+    }
+
+
+def _normalize_issue_comment(payload: dict[str, Any], actor: str) -> dict[str, Any]:
+    """Normalize an issue-comment event payload."""
+    issue = payload.get("issue", {})
+    comment = payload.get("comment", {})
+    return {
+        "actor": actor,
+        "action": payload.get("action", "created"),
+        "repo": payload.get("repository", {}).get("full_name", ""),
+        "issue_number": issue.get("number", 0),
+        "issue_title": issue.get("title", ""),
+        "comment_body": (comment.get("body", "")[:200]),
+    }
+
+
+def _normalize_pull_request(payload: dict[str, Any], actor: str) -> dict[str, Any]:
+    """Normalize a pull-request event payload."""
+    pr = payload.get("pull_request", {})
+    return {
+        "actor": actor,
+        "action": payload.get("action", ""),
+        "repo": payload.get("repository", {}).get("full_name", ""),
+        "pr_number": pr.get("number", 0),
+        "title": pr.get("title", ""),
+        "merged": pr.get("merged", False),
+    }
+
+
+_NORMALIZERS = {
+    "push": _normalize_push,
+    "issues": _normalize_issue_opened,
+    "issue_comment": _normalize_issue_comment,
+    "pull_request": _normalize_pull_request,
+}
+
+
+async def handle_webhook(event_type: str, payload: dict[str, Any]) -> bool:
+    """Normalize a Gitea webhook payload and emit it to the event bus.
+
+    Args:
+        event_type: The Gitea event type header (e.g. "push", "issues").
+        payload: The raw JSON payload from the webhook.
+
+    Returns:
+        True if an event was emitted, False if filtered or unsupported.
+    """
+    bus_event_type = _EVENT_TYPE_MAP.get(event_type)
+    if bus_event_type is None:
+        logger.debug("Unsupported Gitea event type: %s", event_type)
+        return False
+
+    actor = _extract_actor(payload)
+
+    # Filter bot-only activity — except PR merges
+    if _is_bot(actor) and not _is_pr_merge(event_type, payload):
+        logger.debug("Filtered bot activity from %s on %s", actor, event_type)
+        return False
+
+    normalizer = _NORMALIZERS[event_type]
+    data = normalizer(payload, actor)
+
+    await emit(bus_event_type, source="gitea", data=data)
+    logger.info("Emitted %s from %s", bus_event_type, actor)
+    return True

src/timmy/adapters/time_adapter.py

@@ -0,0 +1,82 @@
+"""Time adapter — circadian awareness for Timmy.
+
+Emits time-of-day events so Timmy knows the current period
+and tracks how long since the last user interaction.
+"""
+
+import logging
+from datetime import UTC, datetime
+
+from infrastructure.events.bus import emit
+
+logger = logging.getLogger(__name__)
+
+# Time-of-day periods: (event_name, start_hour, end_hour)
+_PERIODS = [
+    ("morning", 6, 9),
+    ("afternoon", 12, 14),
+    ("evening", 18, 20),
+    ("late_night", 23, 24),
+    ("late_night", 0, 3),
+]
+
+
+def classify_period(hour: int) -> str | None:
+    """Return the circadian period name for a given hour, or None."""
+    for name, start, end in _PERIODS:
+        if start <= hour < end:
+            return name
+    return None
+
+
+class TimeAdapter:
+    """Emits circadian and interaction-tracking events."""
+
+    def __init__(self) -> None:
+        self._last_interaction: datetime | None = None
+        self._last_period: str | None = None
+        self._last_date: str | None = None
+
+    def record_interaction(self, now: datetime | None = None) -> None:
+        """Record a user interaction timestamp."""
+        self._last_interaction = now or datetime.now(UTC)
+
+    def time_since_last_interaction(
+        self,
+        now: datetime | None = None,
+    ) -> float | None:
+        """Seconds since last user interaction, or None if no interaction."""
+        if self._last_interaction is None:
+            return None
+        current = now or datetime.now(UTC)
+        return (current - self._last_interaction).total_seconds()
+
+    async def tick(self, now: datetime | None = None) -> list[str]:
+        """Check current time and emit relevant events.
+
+        Returns list of event types emitted (useful for testing).
+        """
+        current = now or datetime.now(UTC)
+        emitted: list[str] = []
+
+        # --- new_day ---
+        date_str = current.strftime("%Y-%m-%d")
+        if self._last_date is not None and date_str != self._last_date:
+            event_type = "time.new_day"
+            await emit(event_type, source="time_adapter", data={"date": date_str})
+            emitted.append(event_type)
+        self._last_date = date_str
+
+        # --- circadian period ---
+        period = classify_period(current.hour)
+        if period is not None and period != self._last_period:
+            event_type = f"time.{period}"
+            await emit(
+                event_type,
+                source="time_adapter",
+                data={"hour": current.hour, "period": period},
+            )
+            emitted.append(event_type)
+            self._last_period = period
+
+        return emitted

src/timmy/agent.py

@@ -26,12 +26,12 @@ from timmy.prompts import get_system_prompt
 from timmy.tools import create_full_toolkit
 
 if TYPE_CHECKING:
-    from timmy.backends import ClaudeBackend, GrokBackend, TimmyAirLLMAgent
+    from timmy.backends import ClaudeBackend, GrokBackend
 
 logger = logging.getLogger(__name__)
 
 # Union type for callers that want to hint the return type.
-TimmyAgent = Union[Agent, "TimmyAirLLMAgent", "GrokBackend", "ClaudeBackend"]
+TimmyAgent = Union[Agent, "GrokBackend", "ClaudeBackend"]
 
 # Models known to be too small for reliable tool calling.
 # These hallucinate tool calls as text, invoke tools randomly,
@@ -63,7 +63,7 @@ def _pull_model(model_name: str) -> bool:
 
         logger.info("Pulling model: %s", model_name)
 
-        url = settings.ollama_url.replace("localhost", "127.0.0.1")
+        url = settings.normalized_ollama_url
         req = urllib.request.Request(
             f"{url}/api/pull",
             method="POST",
@@ -172,106 +172,34 @@ def _warmup_model(model_name: str) -> bool:
 
 
 def _resolve_backend(requested: str | None) -> str:
-    """Return the backend name to use, resolving 'auto' and explicit overrides.
+    """Return the backend name to use.
 
-    Priority (highest → lowest):
+    Priority (highest -> lowest):
       1. CLI flag passed directly to create_timmy()
       2. TIMMY_MODEL_BACKEND env var / .env setting
-      3. 'ollama' (safe default — no surprises)
-
-    'auto' triggers Apple Silicon detection: uses AirLLM if both
-    is_apple_silicon() and airllm_available() return True.
+      3. 'ollama' (safe default -- no surprises)
     """
     if requested is not None:
         return requested
 
-    configured = settings.timmy_model_backend  # "ollama" | "airllm" | "grok" | "claude" | "auto"
-    if configured != "auto":
-        return configured
-
-    # "auto" path — lazy import to keep startup fast and tests clean.
-    from timmy.backends import airllm_available, is_apple_silicon
-
-    if is_apple_silicon() and airllm_available():
-        return "airllm"
-    return "ollama"
+    return settings.timmy_model_backend  # "ollama" | "grok" | "claude"
 
 
-def create_timmy(
-    db_file: str = "timmy.db",
-    backend: str | None = None,
-    model_size: str | None = None,
-    *,
-    skip_mcp: bool = False,
-) -> TimmyAgent:
-    """Instantiate the agent — Ollama or AirLLM, same public interface.
+def _build_tools_list(use_tools: bool, skip_mcp: bool, model_name: str) -> list:
+    """Assemble the tools list based on model capability and MCP flags.
 
-    Args:
-        db_file:    SQLite file for Agno conversation memory (Ollama path only).
-        backend:    "ollama" | "airllm" | "auto" | None (reads config/env).
-        model_size: AirLLM size — "8b" | "70b" | "405b" | None (reads config).
-        skip_mcp:   If True, omit MCP tool servers (Gitea, filesystem).
-                    Use for background tasks (thinking, QA) where MCP's
-                    stdio cancel-scope lifecycle conflicts with asyncio
-                    task cancellation.
-
-    Returns an Agno Agent or backend-specific agent — all expose
-    print_response(message, stream).
+    Returns a list of Toolkit / MCPTools objects, or an empty list.
     """
-    resolved = _resolve_backend(backend)
-    size = model_size or settings.airllm_model_size
-
-    if resolved == "claude":
-        from timmy.backends import ClaudeBackend
-
-        return ClaudeBackend()
-
-    if resolved == "grok":
-        from timmy.backends import GrokBackend
-
-        return GrokBackend()
-
-    if resolved == "airllm":
-        from timmy.backends import TimmyAirLLMAgent
-
-        return TimmyAirLLMAgent(model_size=size)
-
-    # Default: Ollama via Agno.
-    # Resolve model with automatic pulling and fallback
-    model_name, is_fallback = _resolve_model_with_fallback(
-        requested_model=None,
-        require_vision=False,
-        auto_pull=True,
-    )
-
-    # If Ollama is completely unreachable, fail loudly.
-    # Sovereignty: never silently send data to a cloud API.
-    # Use --backend claude explicitly if you want cloud inference.
-    if not _check_model_available(model_name):
-        logger.error(
-            "Ollama unreachable and no local models available. "
-            "Start Ollama with 'ollama serve' or use --backend claude explicitly."
-        )
-
-    if is_fallback:
-        logger.info("Using fallback model %s (requested was unavailable)", model_name)
-
-    use_tools = _model_supports_tools(model_name)
-
-    # Conditionally include tools — small models get none
-    toolkit = create_full_toolkit() if use_tools else None
     if not use_tools:
         logger.info("Tools disabled for model %s (too small for reliable tool calling)", model_name)
+        return []
 
-    # Build the tools list — Agno accepts a list of Toolkit / MCPTools
-    tools_list: list = []
-    if toolkit:
-        tools_list.append(toolkit)
+    tools_list: list = [create_full_toolkit()]
 
     # Add MCP tool servers (lazy-connected on first arun()).
     # Skipped when skip_mcp=True — MCP's stdio transport uses anyio cancel
     # scopes that conflict with asyncio background task cancellation (#72).
-    if use_tools and not skip_mcp:
+    if not skip_mcp:
         try:
             from timmy.mcp_tools import create_filesystem_mcp_tools, create_gitea_mcp_tools
 
@@ -285,30 +213,49 @@ def create_timmy(
         except Exception as exc:
             logger.debug("MCP tools unavailable: %s", exc)
 
-    # Select prompt tier based on tool capability
-    base_prompt = get_system_prompt(tools_enabled=use_tools)
+    return tools_list
+
+
+def _build_prompt(use_tools: bool, session_id: str) -> str:
+    """Build the full system prompt with optional memory context."""
+    base_prompt = get_system_prompt(tools_enabled=use_tools, session_id=session_id)
 
-    # Try to load memory context
     try:
         from timmy.memory_system import memory_system
 
         memory_context = memory_system.get_system_context()
         if memory_context:
-            # Truncate if too long — smaller budget for small models
-            # since the expanded prompt (roster, guardrails) uses more tokens
+            # Smaller budget for small models — expanded prompt uses more tokens
             max_context = 2000 if not use_tools else 8000
             if len(memory_context) > max_context:
                 memory_context = memory_context[:max_context] + "\n... [truncated]"
-            full_prompt = f"{base_prompt}\n\n## Memory Context\n\n{memory_context}"
-        else:
-            full_prompt = base_prompt
+            return (
+                f"{base_prompt}\n\n"
+                f"## GROUNDED CONTEXT (verified sources — cite when using)\n\n"
+                f"{memory_context}"
+            )
     except Exception as exc:
         logger.warning("Failed to load memory context: %s", exc)
-        full_prompt = base_prompt
+
+    return base_prompt
+
+
+def _create_ollama_agent(
+    *,
+    db_file: str,
+    model_name: str,
+    tools_list: list,
+    full_prompt: str,
+    use_tools: bool,
+) -> Agent:
+    """Construct the Agno Agent with Ollama backend and warm up the model."""
+    model_kwargs = {}
+    if settings.ollama_num_ctx > 0:
+        model_kwargs["options"] = {"num_ctx": settings.ollama_num_ctx}
 
     agent = Agent(
         name="Agent",
-        model=Ollama(id=model_name, host=settings.ollama_url, timeout=300),
+        model=Ollama(id=model_name, host=settings.ollama_url, timeout=300, **model_kwargs),
         db=SqliteDb(db_file=db_file),
         description=full_prompt,
         add_history_to_context=True,
@@ -322,6 +269,67 @@ def create_timmy(
     return agent
 
 
+def create_timmy(
+    db_file: str = "timmy.db",
+    backend: str | None = None,
+    *,
+    skip_mcp: bool = False,
+    session_id: str = "unknown",
+) -> TimmyAgent:
+    """Instantiate the agent — Ollama, Grok, or Claude.
+
+    Args:
+        db_file:    SQLite file for Agno conversation memory (Ollama path only).
+        backend:    "ollama" | "grok" | "claude" | None (reads config/env).
+        skip_mcp:   If True, omit MCP tool servers (Gitea, filesystem).
+                    Use for background tasks (thinking, QA) where MCP's
+                    stdio cancel-scope lifecycle conflicts with asyncio
+                    task cancellation.
+
+    Returns an Agno Agent or backend-specific agent — all expose
+    print_response(message, stream).
+    """
+    resolved = _resolve_backend(backend)
+
+    if resolved == "claude":
+        from timmy.backends import ClaudeBackend
+
+        return ClaudeBackend()
+
+    if resolved == "grok":
+        from timmy.backends import GrokBackend
+
+        return GrokBackend()
+
+    # Default: Ollama via Agno.
+    model_name, is_fallback = _resolve_model_with_fallback(
+        requested_model=None,
+        require_vision=False,
+        auto_pull=True,
+    )
+
+    if not _check_model_available(model_name):
+        logger.error(
+            "Ollama unreachable and no local models available. "
+            "Start Ollama with 'ollama serve' or use --backend claude explicitly."
+        )
+
+    if is_fallback:
+        logger.info("Using fallback model %s (requested was unavailable)", model_name)
+
+    use_tools = _model_supports_tools(model_name)
+    tools_list = _build_tools_list(use_tools, skip_mcp, model_name)
+    full_prompt = _build_prompt(use_tools, session_id)
+
+    return _create_ollama_agent(
+        db_file=db_file,
+        model_name=model_name,
+        tools_list=tools_list,
+        full_prompt=full_prompt,
+        use_tools=use_tools,
+    )
+
+
 class TimmyWithMemory:
     """Agent wrapper with explicit three-tier memory management."""
 
@@ -336,15 +344,47 @@ class TimmyWithMemory:
         self.initial_context = self.memory.get_system_context()
 
     def chat(self, message: str) -> str:
-        """Simple chat interface that tracks in memory."""
+        """Simple chat interface that tracks in memory.
+
+        Retries on transient Ollama errors (GPU contention, timeouts)
+        with exponential backoff (#70).
+        """
+        import time
+
         # Check for user facts to extract
         self._extract_and_store_facts(message)
 
-        # Run agent
-        result = self.agent.run(message, stream=False)
-        response_text = result.content if hasattr(result, "content") else str(result)
-
-        return response_text
+        # Retry with backoff — GPU contention causes ReadError/ReadTimeout
+        max_retries = 3
+        for attempt in range(1, max_retries + 1):
+            try:
+                result = self.agent.run(message, stream=False)
+                return result.content if hasattr(result, "content") else str(result)
+            except (
+                httpx.ConnectError,
+                httpx.ReadError,
+                httpx.ReadTimeout,
+                httpx.ConnectTimeout,
+                ConnectionError,
+                TimeoutError,
+            ) as exc:
+                if attempt < max_retries:
+                    wait = min(2**attempt, 16)
+                    logger.warning(
+                        "Ollama contention on attempt %d/%d: %s. Waiting %ds before retry...",
+                        attempt,
+                        max_retries,
+                        type(exc).__name__,
+                        wait,
+                    )
+                    time.sleep(wait)
+                else:
+                    logger.error(
+                        "Ollama unreachable after %d attempts: %s",
+                        max_retries,
+                        exc,
+                    )
+                    raise
 
     def _extract_and_store_facts(self, message: str) -> None:
         """Extract user facts from message and store in memory."""
@@ -355,7 +395,8 @@ class TimmyWithMemory:
             if name:
                 self.memory.update_user_fact("Name", name)
                 self.memory.record_decision(f"Learned user's name: {name}")
-        except Exception:
+        except Exception as exc:
+            logger.warning("User name extraction failed: %s", exc)
             pass  # Best-effort extraction
 
     def end_session(self, summary: str = "Session completed") -> None:

src/timmy/agent_core/__init__.py

@@ -1 +0,0 @@
-"""Agent Core — Substrate-agnostic agent interface and base classes."""

src/timmy/agent_core/interface.py

@@ -1,381 +0,0 @@
-"""TimAgent Interface — The substrate-agnostic agent contract.
-
-This is the foundation for embodiment. Whether Timmy runs on:
-- A server with Ollama (today)
-- A Raspberry Pi with sensors
-- A Boston Dynamics Spot robot
-- A VR avatar
-
-The interface remains constant. Implementation varies.
-
-Architecture:
-    perceive()  →  reason  →  act()
-         ↑                      ↓
-         ←←← remember() ←←←←←←┘
-
-All methods return effects that can be logged, audited, and replayed.
-"""
-
-import uuid
-from abc import ABC, abstractmethod
-from dataclasses import dataclass, field
-from datetime import UTC, datetime
-from enum import Enum, auto
-from typing import Any
-
-
-class PerceptionType(Enum):
-    """Types of sensory input an agent can receive."""
-
-    TEXT = auto()  # Natural language
-    IMAGE = auto()  # Visual input
-    AUDIO = auto()  # Sound/speech
-    SENSOR = auto()  # Temperature, distance, etc.
-    MOTION = auto()  # Accelerometer, gyroscope
-    NETWORK = auto()  # API calls, messages
-    INTERNAL = auto()  # Self-monitoring (battery, temp)
-
-
-class ActionType(Enum):
-    """Types of actions an agent can perform."""
-
-    TEXT = auto()  # Generate text response
-    SPEAK = auto()  # Text-to-speech
-    MOVE = auto()  # Physical movement
-    GRIP = auto()  # Manipulate objects
-    CALL = auto()  # API/network call
-    EMIT = auto()  # Signal/light/sound
-    SLEEP = auto()  # Power management
-
-
-class AgentCapability(Enum):
-    """High-level capabilities a TimAgent may possess."""
-
-    REASONING = "reasoning"
-    CODING = "coding"
-    WRITING = "writing"
-    ANALYSIS = "analysis"
-    VISION = "vision"
-    SPEECH = "speech"
-    NAVIGATION = "navigation"
-    MANIPULATION = "manipulation"
-    LEARNING = "learning"
-    COMMUNICATION = "communication"
-
-
-@dataclass(frozen=True)
-class AgentIdentity:
-    """Immutable identity for an agent instance.
-
-    This persists across sessions and substrates. If Timmy moves
-    from cloud to robot, the identity follows.
-    """
-
-    id: str
-    name: str
-    version: str
-    created_at: str = field(default_factory=lambda: datetime.now(UTC).isoformat())
-
-    @classmethod
-    def generate(cls, name: str, version: str = "1.0.0") -> "AgentIdentity":
-        """Generate a new unique identity."""
-        return cls(
-            id=str(uuid.uuid4()),
-            name=name,
-            version=version,
-        )
-
-
-@dataclass
-class Perception:
-    """A sensory input to the agent.
-
-    Substrate-agnostic representation. A camera image and a
-    LiDAR point cloud are both Perception instances.
-    """
-
-    type: PerceptionType
-    data: Any  # Content depends on type
-    timestamp: str = field(default_factory=lambda: datetime.now(UTC).isoformat())
-    source: str = "unknown"  # e.g., "camera_1", "microphone", "user_input"
-    metadata: dict = field(default_factory=dict)
-
-    @classmethod
-    def text(cls, content: str, source: str = "user") -> "Perception":
-        """Factory for text perception."""
-        return cls(
-            type=PerceptionType.TEXT,
-            data=content,
-            source=source,
-        )
-
-    @classmethod
-    def sensor(cls, kind: str, value: float, unit: str = "") -> "Perception":
-        """Factory for sensor readings."""
-        return cls(
-            type=PerceptionType.SENSOR,
-            data={"kind": kind, "value": value, "unit": unit},
-            source=f"sensor_{kind}",
-        )
-
-
-@dataclass
-class Action:
-    """An action the agent intends to perform.
-
-    Actions are effects — they describe what should happen,
-    not how. The substrate implements the "how."
-    """
-
-    type: ActionType
-    payload: Any  # Action-specific data
-    timestamp: str = field(default_factory=lambda: datetime.now(UTC).isoformat())
-    confidence: float = 1.0  # 0-1, agent's certainty
-    deadline: str | None = None  # When action must complete
-
-    @classmethod
-    def respond(cls, text: str, confidence: float = 1.0) -> "Action":
-        """Factory for text response action."""
-        return cls(
-            type=ActionType.TEXT,
-            payload=text,
-            confidence=confidence,
-        )
-
-    @classmethod
-    def move(cls, vector: tuple[float, float, float], speed: float = 1.0) -> "Action":
-        """Factory for movement action (x, y, z meters)."""
-        return cls(
-            type=ActionType.MOVE,
-            payload={"vector": vector, "speed": speed},
-        )
-
-
-@dataclass
-class Memory:
-    """A stored experience or fact.
-
-    Memories are substrate-agnostic. A conversation history
-    and a video recording are both Memory instances.
-    """
-
-    id: str
-    content: Any
-    created_at: str
-    access_count: int = 0
-    last_accessed: str | None = None
-    importance: float = 0.5  # 0-1, for pruning decisions
-    tags: list[str] = field(default_factory=list)
-
-    def touch(self) -> None:
-        """Mark memory as accessed."""
-        self.access_count += 1
-        self.last_accessed = datetime.now(UTC).isoformat()
-
-
-@dataclass
-class Communication:
-    """A message to/from another agent or human."""
-
-    sender: str
-    recipient: str
-    content: Any
-    timestamp: str = field(default_factory=lambda: datetime.now(UTC).isoformat())
-    protocol: str = "direct"  # e.g., "http", "websocket", "speech"
-    encrypted: bool = False
-
-
-class TimAgent(ABC):
-    """Abstract base class for all Timmy agent implementations.
-
-    This is the substrate-agnostic interface. Implementations:
-    - OllamaAgent: LLM-based reasoning (today)
-    - RobotAgent: Physical embodiment (future)
-    - SimulationAgent: Virtual environment (future)
-
-    Usage:
-        agent = OllamaAgent(identity)  # Today's implementation
-
-        perception = Perception.text("Hello Timmy")
-        memory = agent.perceive(perception)
-
-        action = agent.reason("How should I respond?")
-        result = agent.act(action)
-
-        agent.remember(memory)  # Store for future
-    """
-
-    def __init__(self, identity: AgentIdentity) -> None:
-        self._identity = identity
-        self._capabilities: set[AgentCapability] = set()
-        self._state: dict[str, Any] = {}
-
-    @property
-    def identity(self) -> AgentIdentity:
-        """Return this agent's immutable identity."""
-        return self._identity
-
-    @property
-    def capabilities(self) -> set[AgentCapability]:
-        """Return set of supported capabilities."""
-        return self._capabilities.copy()
-
-    def has_capability(self, capability: AgentCapability) -> bool:
-        """Check if agent supports a capability."""
-        return capability in self._capabilities
-
-    @abstractmethod
-    def perceive(self, perception: Perception) -> Memory:
-        """Process sensory input and create a memory.
-
-        This is the entry point for all agent interaction.
-        A text message, camera frame, or temperature reading
-        all enter through perceive().
-
-        Args:
-            perception: Sensory input
-
-        Returns:
-            Memory: Stored representation of the perception
-        """
-        pass
-
-    @abstractmethod
-    def reason(self, query: str, context: list[Memory]) -> Action:
-        """Reason about a situation and decide on action.
-
-        This is where "thinking" happens. The agent uses its
-        substrate-appropriate reasoning (LLM, neural net, rules)
-        to decide what to do.
-
-        Args:
-            query: What to reason about
-            context: Relevant memories for context
-
-        Returns:
-            Action: What the agent decides to do
-        """
-        pass
-
-    @abstractmethod
-    def act(self, action: Action) -> Any:
-        """Execute an action in the substrate.
-
-        This is where the abstract action becomes concrete:
-        - TEXT → Generate LLM response
-        - MOVE → Send motor commands
-        - SPEAK → Call TTS engine
-
-        Args:
-            action: The action to execute
-
-        Returns:
-            Result of the action (substrate-specific)
-        """
-        pass
-
-    @abstractmethod
-    def remember(self, memory: Memory) -> None:
-        """Store a memory for future retrieval.
-
-        The storage mechanism depends on substrate:
-        - Cloud: SQLite, vector DB
-        - Robot: Local flash storage
-        - Hybrid: Synced with conflict resolution
-
-        Args:
-            memory: Experience to store
-        """
-        pass
-
-    @abstractmethod
-    def recall(self, query: str, limit: int = 5) -> list[Memory]:
-        """Retrieve relevant memories.
-
-        Args:
-            query: What to search for
-            limit: Maximum memories to return
-
-        Returns:
-            List of relevant memories, sorted by relevance
-        """
-        pass
-
-    @abstractmethod
-    def communicate(self, message: Communication) -> bool:
-        """Send/receive communication with another agent.
-
-        Args:
-            message: Message to send
-
-        Returns:
-            True if communication succeeded
-        """
-        pass
-
-    def get_state(self) -> dict[str, Any]:
-        """Get current agent state for monitoring/debugging."""
-        return {
-            "identity": self._identity,
-            "capabilities": list(self._capabilities),
-            "state": self._state.copy(),
-        }
-
-    def shutdown(self) -> None:  # noqa: B027
-        """Graceful shutdown. Persist state, close connections."""
-        # Override in subclass for cleanup
-
-
-class AgentEffect:
-    """Log entry for agent actions — for audit and replay.
-
-    The complete history of an agent's life can be captured
-    as a sequence of AgentEffects. This enables:
-    - Debugging: What did the agent see and do?
-    - Audit: Why did it make that decision?
-    - Replay: Reconstruct agent state from log
-    - Training: Learn from agent experiences
-    """
-
-    def __init__(self, log_path: str | None = None) -> None:
-        self._effects: list[dict] = []
-        self._log_path = log_path
-
-    def log_perceive(self, perception: Perception, memory_id: str) -> None:
-        """Log a perception event."""
-        self._effects.append(
-            {
-                "type": "perceive",
-                "perception_type": perception.type.name,
-                "source": perception.source,
-                "memory_id": memory_id,
-                "timestamp": datetime.now(UTC).isoformat(),
-            }
-        )
-
-    def log_reason(self, query: str, action_type: ActionType) -> None:
-        """Log a reasoning event."""
-        self._effects.append(
-            {
-                "type": "reason",
-                "query": query,
-                "action_type": action_type.name,
-                "timestamp": datetime.now(UTC).isoformat(),
-            }
-        )
-
-    def log_act(self, action: Action, result: Any) -> None:
-        """Log an action event."""
-        self._effects.append(
-            {
-                "type": "act",
-                "action_type": action.type.name,
-                "confidence": action.confidence,
-                "result_type": type(result).__name__,
-                "timestamp": datetime.now(UTC).isoformat(),
-            }
-        )
-
-    def export(self) -> list[dict]:
-        """Export effect log for analysis."""
-        return self._effects.copy()

src/timmy/agent_core/ollama_adapter.py

@@ -1,275 +0,0 @@
-"""Ollama-based implementation of TimAgent interface.
-
-This adapter wraps the existing Timmy Ollama agent to conform
-to the substrate-agnostic TimAgent interface. It's the bridge
-between the old codebase and the new embodiment-ready architecture.
-
-Usage:
-    from timmy.agent_core import AgentIdentity, Perception
-    from timmy.agent_core.ollama_adapter import OllamaAgent
-
-    identity = AgentIdentity.generate("Timmy")
-    agent = OllamaAgent(identity)
-
-    perception = Perception.text("Hello!")
-    memory = agent.perceive(perception)
-    action = agent.reason("How should I respond?", [memory])
-    result = agent.act(action)
-"""
-
-from typing import Any
-
-from timmy.agent import _resolve_model_with_fallback, create_timmy
-from timmy.agent_core.interface import (
-    Action,
-    ActionType,
-    AgentCapability,
-    AgentEffect,
-    AgentIdentity,
-    Communication,
-    Memory,
-    Perception,
-    PerceptionType,
-    TimAgent,
-)
-
-
-class OllamaAgent(TimAgent):
-    """TimAgent implementation using local Ollama LLM.
-
-    This is the production agent for Timmy Time v2. It uses
-    Ollama for reasoning and SQLite for memory persistence.
-
-    Capabilities:
-    - REASONING: LLM-based inference
-    - CODING: Code generation and analysis
-    - WRITING: Long-form content creation
-    - ANALYSIS: Data processing and insights
-    - COMMUNICATION: Multi-agent messaging
-    """
-
-    def __init__(
-        self,
-        identity: AgentIdentity,
-        model: str | None = None,
-        effect_log: str | None = None,
-        require_vision: bool = False,
-    ) -> None:
-        """Initialize Ollama-based agent.
-
-        Args:
-            identity: Agent identity (persistent across sessions)
-            model: Ollama model to use (auto-resolves with fallback)
-            effect_log: Path to log agent effects (optional)
-            require_vision: Whether to select a vision-capable model
-        """
-        super().__init__(identity)
-
-        # Resolve model with automatic pulling and fallback
-        resolved_model, is_fallback = _resolve_model_with_fallback(
-            requested_model=model,
-            require_vision=require_vision,
-            auto_pull=True,
-        )
-
-        if is_fallback:
-            import logging
-
-            logging.getLogger(__name__).info(
-                "OllamaAdapter using fallback model %s", resolved_model
-            )
-
-        # Initialize underlying Ollama agent
-        self._timmy = create_timmy(model=resolved_model)
-
-        # Set capabilities based on what Ollama can do
-        self._capabilities = {
-            AgentCapability.REASONING,
-            AgentCapability.CODING,
-            AgentCapability.WRITING,
-            AgentCapability.ANALYSIS,
-            AgentCapability.COMMUNICATION,
-        }
-
-        # Effect logging for audit/replay
-        self._effect_log = AgentEffect(effect_log) if effect_log else None
-
-        # Simple in-memory working memory (short term)
-        self._working_memory: list[Memory] = []
-        self._max_working_memory = 10
-
-    def perceive(self, perception: Perception) -> Memory:
-        """Process perception and store in memory.
-
-        For text perceptions, we might do light preprocessing
-        (summarization, keyword extraction) before storage.
-        """
-        # Create memory from perception
-        memory = Memory(
-            id=f"mem_{len(self._working_memory)}",
-            content={
-                "type": perception.type.name,
-                "data": perception.data,
-                "source": perception.source,
-            },
-            created_at=perception.timestamp,
-            tags=self._extract_tags(perception),
-        )
-
-        # Add to working memory
-        self._working_memory.append(memory)
-        if len(self._working_memory) > self._max_working_memory:
-            self._working_memory.pop(0)  # FIFO eviction
-
-        # Log effect
-        if self._effect_log:
-            self._effect_log.log_perceive(perception, memory.id)
-
-        return memory
-
-    def reason(self, query: str, context: list[Memory]) -> Action:
-        """Use LLM to reason and decide on action.
-
-        This is where the Ollama agent does its work. We construct
-        a prompt from the query and context, then interpret the
-        response as an action.
-        """
-        # Build context string from memories
-        context_str = self._format_context(context)
-
-        # Construct prompt
-        prompt = f"""You are {self._identity.name}, an AI assistant.
-
-Context from previous interactions:
-{context_str}
-
-Current query: {query}
-
-Respond naturally and helpfully."""
-
-        # Run LLM inference
-        result = self._timmy.run(prompt, stream=False)
-        response_text = result.content if hasattr(result, "content") else str(result)
-
-        # Create text response action
-        action = Action.respond(response_text, confidence=0.9)
-
-        # Log effect
-        if self._effect_log:
-            self._effect_log.log_reason(query, action.type)
-
-        return action
-
-    def act(self, action: Action) -> Any:
-        """Execute action in the Ollama substrate.
-
-        For text actions, the "execution" is just returning the
-        text (already generated during reasoning). For future
-        action types (MOVE, SPEAK), this would trigger the
-        appropriate Ollama tool calls.
-        """
-        result = None
-
-        if action.type == ActionType.TEXT:
-            result = action.payload
-        elif action.type == ActionType.SPEAK:
-            # Would call TTS here
-            result = {"spoken": action.payload, "tts_engine": "pyttsx3"}
-        elif action.type == ActionType.CALL:
-            # Would make API call
-            result = {"status": "not_implemented", "payload": action.payload}
-        else:
-            result = {"error": f"Action type {action.type} not supported by OllamaAgent"}
-
-        # Log effect
-        if self._effect_log:
-            self._effect_log.log_act(action, result)
-
-        return result
-
-    def remember(self, memory: Memory) -> None:
-        """Store memory in working memory.
-
-        Adds the memory to the sliding window and bumps its importance.
-        """
-        memory.touch()
-
-        # Deduplicate by id
-        self._working_memory = [m for m in self._working_memory if m.id != memory.id]
-        self._working_memory.append(memory)
-
-        # Evict oldest if over capacity
-        if len(self._working_memory) > self._max_working_memory:
-            self._working_memory.pop(0)
-
-    def recall(self, query: str, limit: int = 5) -> list[Memory]:
-        """Retrieve relevant memories.
-
-        Simple keyword matching for now. Future: vector similarity.
-        """
-        query_lower = query.lower()
-        scored = []
-
-        for memory in self._working_memory:
-            score = 0
-            content_str = str(memory.content).lower()
-
-            # Simple keyword overlap
-            query_words = set(query_lower.split())
-            content_words = set(content_str.split())
-            overlap = len(query_words & content_words)
-            score += overlap
-
-            # Boost recent memories
-            score += memory.importance
-
-            scored.append((score, memory))
-
-        # Sort by score descending
-        scored.sort(key=lambda x: x[0], reverse=True)
-
-        # Return top N
-        return [m for _, m in scored[:limit]]
-
-    def communicate(self, message: Communication) -> bool:
-        """Send message to another agent.
-
-        Swarm comms removed — inter-agent communication will be handled
-        by the unified brain memory layer.
-        """
-        return False
-
-    def _extract_tags(self, perception: Perception) -> list[str]:
-        """Extract searchable tags from perception."""
-        tags = [perception.type.name, perception.source]
-
-        if perception.type == PerceptionType.TEXT:
-            # Simple keyword extraction
-            text = str(perception.data).lower()
-            keywords = ["code", "bug", "help", "question", "task"]
-            for kw in keywords:
-                if kw in text:
-                    tags.append(kw)
-
-        return tags
-
-    def _format_context(self, memories: list[Memory]) -> str:
-        """Format memories into context string for prompt."""
-        if not memories:
-            return "No previous context."
-
-        parts = []
-        for mem in memories[-5:]:  # Last 5 memories
-            if isinstance(mem.content, dict):
-                data = mem.content.get("data", "")
-                parts.append(f"- {data}")
-            else:
-                parts.append(f"- {mem.content}")
-
-        return "\n".join(parts)
-
-    def get_effect_log(self) -> list[dict] | None:
-        """Export effect log if logging is enabled."""
-        if self._effect_log:
-            return self._effect_log.export()
-        return None

src/timmy/agentic_loop.py

@@ -18,6 +18,7 @@ from __future__ import annotations
 import asyncio
 import logging
 import re
+import threading
 import time
 import uuid
 from collections.abc import Callable
@@ -58,6 +59,9 @@ class AgenticResult:
 # Agent factory
 # ---------------------------------------------------------------------------
 
+_loop_agent = None
+_loop_agent_lock = threading.Lock()
+
 
 def _get_loop_agent():
     """Create a fresh agent for the agentic loop.
@@ -65,9 +69,14 @@ def _get_loop_agent():
     Returns the same type of agent as `create_timmy()` but with a
     dedicated session so it doesn't pollute the main chat history.
     """
-    from timmy.agent import create_timmy
+    global _loop_agent
+    if _loop_agent is None:
+        with _loop_agent_lock:
+            if _loop_agent is None:
+                from timmy.agent import create_timmy
 
-    return create_timmy()
+                _loop_agent = create_timmy()
+    return _loop_agent
 
 
 # ---------------------------------------------------------------------------
@@ -86,6 +95,126 @@ def _parse_steps(plan_text: str) -> list[str]:
     return [line.strip() for line in plan_text.strip().splitlines() if line.strip()]
 
 
+# ---------------------------------------------------------------------------
+# Extracted helpers
+# ---------------------------------------------------------------------------
+
+
+def _extract_content(run_result) -> str:
+    """Extract text content from an agent run result."""
+    return run_result.content if hasattr(run_result, "content") else str(run_result)
+
+
+def _clean(text: str) -> str:
+    """Clean a model response using session's response cleaner."""
+    from timmy.session import _clean_response
+
+    return _clean_response(text)
+
+
+async def _plan_task(
+    agent, task: str, session_id: str, max_steps: int
+) -> tuple[list[str], bool] | str:
+    """Run the planning phase — returns (steps, was_truncated) or error string."""
+    plan_prompt = (
+        f"Break this task into numbered steps (max {max_steps}). "
+        f"Return ONLY a numbered list, nothing else.\n\n"
+        f"Task: {task}"
+    )
+    try:
+        plan_run = await asyncio.to_thread(
+            agent.run, plan_prompt, stream=False, session_id=f"{session_id}_plan"
+        )
+        plan_text = _extract_content(plan_run)
+    except Exception as exc:  # broad catch intentional: agent.run can raise any error
+        logger.error("Agentic loop: planning failed: %s", exc)
+        return f"Planning failed: {exc}"
+
+    steps = _parse_steps(plan_text)
+    if not steps:
+        return "Planning produced no steps."
+
+    planned_count = len(steps)
+    steps = steps[:max_steps]
+    return steps, planned_count > len(steps)
+
+
+async def _execute_step(
+    agent,
+    task: str,
+    step_desc: str,
+    step_num: int,
+    total_steps: int,
+    recent_results: list[str],
+    session_id: str,
+) -> AgenticStep:
+    """Execute a single step, returning an AgenticStep."""
+    step_start = time.monotonic()
+    context = (
+        f"Task: {task}\n"
+        f"Step {step_num}/{total_steps}: {step_desc}\n"
+        f"Recent progress: {recent_results[-2:] if recent_results else []}\n\n"
+        f"Execute this step and report what you did."
+    )
+    step_run = await asyncio.to_thread(
+        agent.run, context, stream=False, session_id=f"{session_id}_step{step_num}"
+    )
+    step_result = _clean(_extract_content(step_run))
+    return AgenticStep(
+        step_num=step_num,
+        description=step_desc,
+        result=step_result,
+        status="completed",
+        duration_ms=int((time.monotonic() - step_start) * 1000),
+    )
+
+
+async def _adapt_step(
+    agent,
+    step_desc: str,
+    step_num: int,
+    error: Exception,
+    step_start: float,
+    session_id: str,
+) -> AgenticStep:
+    """Attempt adaptation after a step failure."""
+    adapt_prompt = (
+        f"Step {step_num} failed with error: {error}\n"
+        f"Original step was: {step_desc}\n"
+        f"Adapt the plan and try an alternative approach for this step."
+    )
+    adapt_run = await asyncio.to_thread(
+        agent.run, adapt_prompt, stream=False, session_id=f"{session_id}_adapt{step_num}"
+    )
+    adapt_result = _clean(_extract_content(adapt_run))
+    return AgenticStep(
+        step_num=step_num,
+        description=f"[Adapted] {step_desc}",
+        result=adapt_result,
+        status="adapted",
+        duration_ms=int((time.monotonic() - step_start) * 1000),
+    )
+
+
+def _summarize(result: AgenticResult, total_steps: int, was_truncated: bool) -> None:
+    """Fill in summary and final status on the result object (mutates in place)."""
+    completed = sum(1 for s in result.steps if s.status == "completed")
+    adapted = sum(1 for s in result.steps if s.status == "adapted")
+    failed = sum(1 for s in result.steps if s.status == "failed")
+
+    parts = [f"Completed {completed}/{total_steps} steps"]
+    if adapted:
+        parts.append(f"{adapted} adapted")
+    if failed:
+        parts.append(f"{failed} failed")
+    result.summary = f"{result.task}: {', '.join(parts)}."
+
+    if was_truncated or len(result.steps) < total_steps or failed:
+        result.status = "partial"
+    else:
+        result.status = "completed"
+
+
 # ---------------------------------------------------------------------------
 # Core loop
 # ---------------------------------------------------------------------------
@@ -116,88 +245,41 @@ async def run_agentic_loop(
 
     task_id = str(uuid.uuid4())[:8]
     start_time = time.monotonic()
-
     agent = _get_loop_agent()
     result = AgenticResult(task_id=task_id, task=task, summary="")
 
-    # ── Phase 1: Planning ──────────────────────────────────────────────────
-    plan_prompt = (
-        f"Break this task into numbered steps (max {max_steps}). "
-        f"Return ONLY a numbered list, nothing else.\n\n"
-        f"Task: {task}"
-    )
-    try:
-        plan_run = await asyncio.to_thread(
-            agent.run, plan_prompt, stream=False, session_id=f"{session_id}_plan"
-        )
-        plan_text = plan_run.content if hasattr(plan_run, "content") else str(plan_run)
-    except Exception as exc:
-        logger.error("Agentic loop: planning failed: %s", exc)
+    # Phase 1: Planning
+    plan = await _plan_task(agent, task, session_id, max_steps)
+    if isinstance(plan, str):
         result.status = "failed"
-        result.summary = f"Planning failed: {exc}"
+        result.summary = plan
         result.total_duration_ms = int((time.monotonic() - start_time) * 1000)
         return result
 
-    steps = _parse_steps(plan_text)
-    if not steps:
-        result.status = "failed"
-        result.summary = "Planning produced no steps."
-        result.total_duration_ms = int((time.monotonic() - start_time) * 1000)
-        return result
-
-    # Enforce max_steps — track if we truncated
-    planned_steps = len(steps)
-    steps = steps[:max_steps]
+    steps, was_truncated = plan
     total_steps = len(steps)
-    was_truncated = planned_steps > total_steps
 
-    # Broadcast plan
     await _broadcast_progress(
         "agentic.plan_ready",
-        {
-            "task_id": task_id,
-            "task": task,
-            "steps": steps,
-            "total": total_steps,
-        },
+        {"task_id": task_id, "task": task, "steps": steps, "total": total_steps},
     )
 
-    # ── Phase 2: Execution ─────────────────────────────────────────────────
+    # Phase 2: Execution
     completed_results: list[str] = []
-
     for i, step_desc in enumerate(steps, 1):
         step_start = time.monotonic()
-
-        context = (
-            f"Task: {task}\n"
-            f"Plan: {plan_text}\n"
-            f"Completed so far: {completed_results}\n\n"
-            f"Now do step {i}: {step_desc}\n"
-            f"Execute this step and report what you did."
-        )
-
         try:
-            step_run = await asyncio.to_thread(
-                agent.run, context, stream=False, session_id=f"{session_id}_step{i}"
-            )
-            step_result = step_run.content if hasattr(step_run, "content") else str(step_run)
-
-            # Clean the response
-            from timmy.session import _clean_response
-
-            step_result = _clean_response(step_result)
-
-            step = AgenticStep(
-                step_num=i,
-                description=step_desc,
-                result=step_result,
-                status="completed",
-                duration_ms=int((time.monotonic() - step_start) * 1000),
+            step = await _execute_step(
+                agent,
+                task,
+                step_desc,
+                i,
+                total_steps,
+                completed_results,
+                session_id,
             )
             result.steps.append(step)
-            completed_results.append(f"Step {i}: {step_result[:200]}")
-
-            # Broadcast progress
+            completed_results.append(f"Step {i}: {step.result[:200]}")
             await _broadcast_progress(
                 "agentic.step_complete",
                 {
@@ -205,46 +287,18 @@ async def run_agentic_loop(
                     "step": i,
                     "total": total_steps,
                     "description": step_desc,
-                    "result": step_result[:200],
+                    "result": step.result[:200],
                 },
             )
-
             if on_progress:
                 await on_progress(step_desc, i, total_steps)
 
-        except Exception as exc:
+        except Exception as exc:  # broad catch intentional: agent.run can raise any error
             logger.warning("Agentic loop step %d failed: %s", i, exc)
-
-            # ── Adaptation: ask model to adapt ─────────────────────────────
-            adapt_prompt = (
-                f"Step {i} failed with error: {exc}\n"
-                f"Original step was: {step_desc}\n"
-                f"Adapt the plan and try an alternative approach for this step."
-            )
             try:
-                adapt_run = await asyncio.to_thread(
-                    agent.run,
-                    adapt_prompt,
-                    stream=False,
-                    session_id=f"{session_id}_adapt{i}",
-                )
-                adapt_result = (
-                    adapt_run.content if hasattr(adapt_run, "content") else str(adapt_run)
-                )
-                from timmy.session import _clean_response
-
-                adapt_result = _clean_response(adapt_result)
-
-                step = AgenticStep(
-                    step_num=i,
-                    description=f"[Adapted] {step_desc}",
-                    result=adapt_result,
-                    status="adapted",
-                    duration_ms=int((time.monotonic() - step_start) * 1000),
-                )
+                step = await _adapt_step(agent, step_desc, i, exc, step_start, session_id)
                 result.steps.append(step)
-                completed_results.append(f"Step {i} (adapted): {adapt_result[:200]}")
-
+                completed_results.append(f"Step {i} (adapted): {step.result[:200]}")
                 await _broadcast_progress(
                     "agentic.step_adapted",
                     {
@@ -253,58 +307,26 @@ async def run_agentic_loop(
                         "total": total_steps,
                         "description": step_desc,
                         "error": str(exc),
-                        "adaptation": adapt_result[:200],
+                        "adaptation": step.result[:200],
                     },
                 )
-
                 if on_progress:
                     await on_progress(f"[Adapted] {step_desc}", i, total_steps)
-
-            except Exception as adapt_exc:
+            except Exception as adapt_exc:  # broad catch intentional
                 logger.error("Agentic loop adaptation also failed: %s", adapt_exc)
-                step = AgenticStep(
-                    step_num=i,
-                    description=step_desc,
-                    result=f"Failed: {exc}; Adaptation also failed: {adapt_exc}",
-                    status="failed",
-                    duration_ms=int((time.monotonic() - step_start) * 1000),
+                result.steps.append(
+                    AgenticStep(
+                        step_num=i,
+                        description=step_desc,
+                        result=f"Failed: {exc}; Adaptation also failed: {adapt_exc}",
+                        status="failed",
+                        duration_ms=int((time.monotonic() - step_start) * 1000),
+                    )
                 )
-                result.steps.append(step)
                 completed_results.append(f"Step {i}: FAILED")
 
-    # ── Phase 3: Summary ───────────────────────────────────────────────────
-    summary_prompt = (
-        f"Task: {task}\n"
-        f"Results:\n" + "\n".join(completed_results) + "\n\n"
-        "Summarise what was accomplished in 2-3 sentences."
-    )
-    try:
-        summary_run = await asyncio.to_thread(
-            agent.run,
-            summary_prompt,
-            stream=False,
-            session_id=f"{session_id}_summary",
-        )
-        result.summary = (
-            summary_run.content if hasattr(summary_run, "content") else str(summary_run)
-        )
-        from timmy.session import _clean_response
-
-        result.summary = _clean_response(result.summary)
-    except Exception as exc:
-        logger.error("Agentic loop summary failed: %s", exc)
-        result.summary = f"Completed {len(result.steps)} steps."
-
-    # Determine final status
-    if was_truncated:
-        result.status = "partial"
-    elif len(result.steps) < total_steps:
-        result.status = "partial"
-    elif any(s.status == "failed" for s in result.steps):
-        result.status = "partial"
-    else:
-        result.status = "completed"
-
+    # Phase 3: Summary
+    _summarize(result, total_steps, was_truncated)
     result.total_duration_ms = int((time.monotonic() - start_time) * 1000)
 
     await _broadcast_progress(
@@ -332,5 +354,6 @@ async def _broadcast_progress(event: str, data: dict) -> None:
         from infrastructure.ws_manager.handler import ws_manager
 
         await ws_manager.broadcast(event, data)
-    except Exception:
+    except (ImportError, AttributeError, ConnectionError, RuntimeError) as exc:
+        logger.warning("Agentic loop broadcast failed: %s", exc)
         logger.debug("Agentic loop: WS broadcast failed for %s", event)

src/timmy/agents/base.py

@@ -10,10 +10,12 @@ SubAgent is the single seed class for ALL agents.  Differentiation
 comes entirely from config (agents.yaml), not from Python subclasses.
 """
 
+import asyncio
 import logging
 from abc import ABC, abstractmethod
 from typing import Any
 
+import httpx
 from agno.agent import Agent
 from agno.models.ollama import Ollama
 
@@ -72,9 +74,12 @@ class BaseAgent(ABC):
                 if handler:
                     tool_instances.append(handler)
 
+        ollama_kwargs = {}
+        if settings.ollama_num_ctx > 0:
+            ollama_kwargs["options"] = {"num_ctx": settings.ollama_num_ctx}
         return Agent(
             name=self.name,
-            model=Ollama(id=self.model, host=settings.ollama_url, timeout=300),
+            model=Ollama(id=self.model, host=settings.ollama_url, timeout=300, **ollama_kwargs),
             description=system_prompt,
             tools=tool_instances if tool_instances else None,
             add_history_to_context=True,
@@ -117,11 +122,70 @@ class BaseAgent(ABC):
     async def run(self, message: str) -> str:
         """Run the agent with a message.
 
+        Retries on transient failures (connection errors, timeouts) with
+        exponential backoff.  GPU contention from concurrent Ollama
+        requests causes ReadError / ReadTimeout — these are transient
+        and should be retried, not raised immediately (#70).
+
         Returns:
             Agent response
         """
-        result = self.agent.run(message, stream=False)
-        response = result.content if hasattr(result, "content") else str(result)
+        max_retries = 3
+        last_exception = None
+        # Transient errors that indicate Ollama contention or temporary
+        # unavailability — these deserve a retry with backoff.
+        _transient = (
+            httpx.ConnectError,
+            httpx.ReadError,
+            httpx.ReadTimeout,
+            httpx.ConnectTimeout,
+            ConnectionError,
+            TimeoutError,
+        )
+
+        for attempt in range(1, max_retries + 1):
+            try:
+                result = self.agent.run(message, stream=False)
+                response = result.content if hasattr(result, "content") else str(result)
+                break  # Success, exit the retry loop
+            except _transient as exc:
+                last_exception = exc
+                if attempt < max_retries:
+                    # Contention backoff — longer waits because the GPU
+                    # needs time to finish the other request.
+                    wait = min(2**attempt, 16)
+                    logger.warning(
+                        "Ollama contention on attempt %d/%d: %s. Waiting %ds before retry...",
+                        attempt,
+                        max_retries,
+                        type(exc).__name__,
+                        wait,
+                    )
+                    await asyncio.sleep(wait)
+                else:
+                    logger.error(
+                        "Ollama unreachable after %d attempts: %s",
+                        max_retries,
+                        exc,
+                    )
+                    raise last_exception from exc
+            except Exception as exc:
+                last_exception = exc
+                if attempt < max_retries:
+                    logger.warning(
+                        "Agent run failed on attempt %d/%d: %s. Retrying...",
+                        attempt,
+                        max_retries,
+                        exc,
+                    )
+                    await asyncio.sleep(min(2 ** (attempt - 1), 8))
+                else:
+                    logger.error(
+                        "Agent run failed after %d attempts: %s",
+                        max_retries,
+                        exc,
+                    )
+                    raise last_exception from exc
 
         # Emit completion event
         if self.event_bus:

src/timmy/agents/loader.py

@@ -16,6 +16,7 @@ Usage:
 from __future__ import annotations
 
 import logging
+import re
 from pathlib import Path
 from typing import Any
 
@@ -181,6 +182,23 @@ def get_routing_config() -> dict[str, Any]:
     return config.get("routing", {"method": "pattern", "patterns": {}})
 
 
+def _matches_pattern(pattern: str, message: str) -> bool:
+    """Check if a pattern matches using word-boundary matching.
+
+    For single-word patterns, uses \b word boundaries.
+    For multi-word patterns, all words must appear as whole words (in any order).
+    """
+    pattern_lower = pattern.lower()
+    message_lower = message.lower()
+    words = pattern_lower.split()
+
+    for word in words:
+        # Use word boundary regex to match whole words only
+        if not re.search(rf"\b{re.escape(word)}\b", message_lower):
+            return False
+    return True
+
+
 def route_request(user_message: str) -> str | None:
     """Route a user request to an agent using pattern matching.
 
@@ -193,17 +211,36 @@ def route_request(user_message: str) -> str | None:
         return None
 
     patterns = routing.get("patterns", {})
-    message_lower = user_message.lower()
 
     for agent_id, keywords in patterns.items():
         for keyword in keywords:
-            if keyword.lower() in message_lower:
+            if _matches_pattern(keyword, user_message):
                 logger.debug("Routed to %s (matched: %r)", agent_id, keyword)
                 return agent_id
 
     return None
 
 
+def route_request_with_match(user_message: str) -> tuple[str | None, str | None]:
+    """Route a user request and return both the agent and the matched pattern.
+
+    Returns a tuple of (agent_id, matched_pattern). If no match, returns (None, None).
+    """
+    routing = get_routing_config()
+
+    if routing.get("method") != "pattern":
+        return None, None
+
+    patterns = routing.get("patterns", {})
+
+    for agent_id, keywords in patterns.items():
+        for keyword in keywords:
+            if _matches_pattern(keyword, user_message):
+                return agent_id, keyword
+
+    return None, None
+
+
 def reload_agents() -> dict[str, Any]:
     """Force reload agents from YAML.  Call after editing agents.yaml."""
     global _agents, _config

src/timmy/approvals.py

@@ -13,6 +13,8 @@ Default is always True. The owner changes this intentionally.
 
 import sqlite3
 import uuid
+from collections.abc import Generator
+from contextlib import closing, contextmanager
 from dataclasses import dataclass
 from datetime import UTC, datetime, timedelta
 from pathlib import Path
@@ -43,23 +45,24 @@ class ApprovalItem:
     status: str  # "pending" | "approved" | "rejected"
 
 
-def _get_conn(db_path: Path = _DEFAULT_DB) -> sqlite3.Connection:
+@contextmanager
+def _get_conn(db_path: Path = _DEFAULT_DB) -> Generator[sqlite3.Connection, None, None]:
     db_path.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(db_path))
-    conn.row_factory = sqlite3.Row
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS approval_items (
-            id              TEXT PRIMARY KEY,
-            title           TEXT NOT NULL,
-            description     TEXT NOT NULL,
-            proposed_action TEXT NOT NULL,
-            impact          TEXT NOT NULL DEFAULT 'low',
-            created_at      TEXT NOT NULL,
-            status          TEXT NOT NULL DEFAULT 'pending'
-        )
-        """)
-    conn.commit()
-    return conn
+    with closing(sqlite3.connect(str(db_path))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS approval_items (
+                id              TEXT PRIMARY KEY,
+                title           TEXT NOT NULL,
+                description     TEXT NOT NULL,
+                proposed_action TEXT NOT NULL,
+                impact          TEXT NOT NULL DEFAULT 'low',
+                created_at      TEXT NOT NULL,
+                status          TEXT NOT NULL DEFAULT 'pending'
+            )
+            """)
+        conn.commit()
+        yield conn
 
 
 def _row_to_item(row: sqlite3.Row) -> ApprovalItem:
@@ -96,80 +99,73 @@ def create_item(
         created_at=datetime.now(UTC),
         status="pending",
     )
-    conn = _get_conn(db_path)
-    conn.execute(
-        """
-        INSERT INTO approval_items
-            (id, title, description, proposed_action, impact, created_at, status)
-        VALUES (?, ?, ?, ?, ?, ?, ?)
-        """,
-        (
-            item.id,
-            item.title,
-            item.description,
-            item.proposed_action,
-            item.impact,
-            item.created_at.isoformat(),
-            item.status,
-        ),
-    )
-    conn.commit()
-    conn.close()
+    with _get_conn(db_path) as conn:
+        conn.execute(
+            """
+            INSERT INTO approval_items
+                (id, title, description, proposed_action, impact, created_at, status)
+            VALUES (?, ?, ?, ?, ?, ?, ?)
+            """,
+            (
+                item.id,
+                item.title,
+                item.description,
+                item.proposed_action,
+                item.impact,
+                item.created_at.isoformat(),
+                item.status,
+            ),
+        )
+        conn.commit()
     return item
 
 
 def list_pending(db_path: Path = _DEFAULT_DB) -> list[ApprovalItem]:
     """Return all pending approval items, newest first."""
-    conn = _get_conn(db_path)
-    rows = conn.execute(
-        "SELECT * FROM approval_items WHERE status = 'pending' ORDER BY created_at DESC"
-    ).fetchall()
-    conn.close()
+    with _get_conn(db_path) as conn:
+        rows = conn.execute(
+            "SELECT * FROM approval_items WHERE status = 'pending' ORDER BY created_at DESC"
+        ).fetchall()
     return [_row_to_item(r) for r in rows]
 
 
 def list_all(db_path: Path = _DEFAULT_DB) -> list[ApprovalItem]:
     """Return all approval items regardless of status, newest first."""
-    conn = _get_conn(db_path)
-    rows = conn.execute("SELECT * FROM approval_items ORDER BY created_at DESC").fetchall()
-    conn.close()
+    with _get_conn(db_path) as conn:
+        rows = conn.execute("SELECT * FROM approval_items ORDER BY created_at DESC").fetchall()
     return [_row_to_item(r) for r in rows]
 
 
 def get_item(item_id: str, db_path: Path = _DEFAULT_DB) -> ApprovalItem | None:
-    conn = _get_conn(db_path)
-    row = conn.execute("SELECT * FROM approval_items WHERE id = ?", (item_id,)).fetchone()
-    conn.close()
+    with _get_conn(db_path) as conn:
+        row = conn.execute("SELECT * FROM approval_items WHERE id = ?", (item_id,)).fetchone()
     return _row_to_item(row) if row else None
 
 
 def approve(item_id: str, db_path: Path = _DEFAULT_DB) -> ApprovalItem | None:
     """Mark an approval item as approved."""
-    conn = _get_conn(db_path)
-    conn.execute("UPDATE approval_items SET status = 'approved' WHERE id = ?", (item_id,))
-    conn.commit()
-    conn.close()
+    with _get_conn(db_path) as conn:
+        conn.execute("UPDATE approval_items SET status = 'approved' WHERE id = ?", (item_id,))
+        conn.commit()
     return get_item(item_id, db_path)
 
 
 def reject(item_id: str, db_path: Path = _DEFAULT_DB) -> ApprovalItem | None:
     """Mark an approval item as rejected."""
-    conn = _get_conn(db_path)
-    conn.execute("UPDATE approval_items SET status = 'rejected' WHERE id = ?", (item_id,))
-    conn.commit()
-    conn.close()
+    with _get_conn(db_path) as conn:
+        conn.execute("UPDATE approval_items SET status = 'rejected' WHERE id = ?", (item_id,))
+        conn.commit()
     return get_item(item_id, db_path)
 
 
 def expire_old(db_path: Path = _DEFAULT_DB) -> int:
     """Auto-expire pending items older than EXPIRY_DAYS. Returns count removed."""
     cutoff = (datetime.now(UTC) - timedelta(days=_EXPIRY_DAYS)).isoformat()
-    conn = _get_conn(db_path)
-    cursor = conn.execute(
-        "DELETE FROM approval_items WHERE status = 'pending' AND created_at < ?",
-        (cutoff,),
-    )
-    conn.commit()
-    count = cursor.rowcount
-    conn.close()
+    with _get_conn(db_path) as conn:
+        cursor = conn.execute(
+            "DELETE FROM approval_items WHERE status = 'pending' AND created_at < ?",
+            (cutoff,),
+        )
+        conn.commit()
+        count = cursor.rowcount
     return count

src/timmy/backends.py

@@ -1,11 +1,10 @@
-"""LLM backends — AirLLM (local big models), Grok (xAI), and Claude (Anthropic).
+"""LLM backends — Grok (xAI) and Claude (Anthropic).
 
 Provides drop-in replacements for the Agno Agent that expose the same
 run(message, stream) → RunResult interface used by the dashboard and the
 print_response(message, stream) interface used by the CLI.
 
 Backends:
-  - TimmyAirLLMAgent: Local 8B/70B/405B via AirLLM (Apple Silicon or PyTorch)
   - GrokBackend: xAI Grok API via OpenAI-compatible SDK (opt-in premium)
   - ClaudeBackend: Anthropic Claude API — lightweight cloud fallback
 
@@ -16,27 +15,18 @@ import logging
 import platform
 import time
 from dataclasses import dataclass
-from typing import Literal
 
-from timmy.prompts import SYSTEM_PROMPT
+from timmy.prompts import get_system_prompt
 
 logger = logging.getLogger(__name__)
 
-# HuggingFace model IDs for each supported size.
-_AIRLLM_MODELS: dict[str, str] = {
-    "8b": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-    "70b": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-    "405b": "meta-llama/Meta-Llama-3.1-405B-Instruct",
-}
-
-ModelSize = Literal["8b", "70b", "405b"]
-
 
 @dataclass
 class RunResult:
     """Minimal Agno-compatible run result — carries the model's response text."""
 
     content: str
+    confidence: float | None = None
 
 
 def is_apple_silicon() -> bool:
@@ -44,108 +34,6 @@ def is_apple_silicon() -> bool:
     return platform.system() == "Darwin" and platform.machine() == "arm64"
 
 
-def airllm_available() -> bool:
-    """Return True when the airllm package is importable."""
-    try:
-        import airllm  # noqa: F401
-
-        return True
-    except ImportError:
-        return False
-
-
-class TimmyAirLLMAgent:
-    """Thin AirLLM wrapper compatible with both dashboard and CLI call sites.
-
-    Exposes:
-      run(message, stream)           → RunResult(content=...)  [dashboard]
-      print_response(message, stream) → None                   [CLI]
-
-    Maintains a rolling 10-turn in-memory history so Timmy remembers the
-    conversation within a session — no SQLite needed at this layer.
-    """
-
-    def __init__(self, model_size: str = "70b") -> None:
-        model_id = _AIRLLM_MODELS.get(model_size)
-        if model_id is None:
-            raise ValueError(
-                f"Unknown model size {model_size!r}. Choose from: {list(_AIRLLM_MODELS)}"
-            )
-
-        if is_apple_silicon():
-            from airllm import AirLLMMLX  # type: ignore[import]
-
-            self._model = AirLLMMLX(model_id)
-        else:
-            from airllm import AutoModel  # type: ignore[import]
-
-            self._model = AutoModel.from_pretrained(model_id)
-
-        self._history: list[str] = []
-        self._model_size = model_size
-
-    # ── public interface (mirrors Agno Agent) ────────────────────────────────
-
-    def run(self, message: str, *, stream: bool = False) -> RunResult:
-        """Run inference and return a structured result (matches Agno Agent.run()).
-
-        `stream` is accepted for API compatibility; AirLLM always generates
-        the full output in one pass.
-        """
-        prompt = self._build_prompt(message)
-
-        input_tokens = self._model.tokenizer(
-            [prompt],
-            return_tensors="pt",
-            padding=True,
-            truncation=True,
-            max_length=2048,
-        )
-        output = self._model.generate(
-            **input_tokens,
-            max_new_tokens=512,
-            use_cache=True,
-            do_sample=True,
-            temperature=0.7,
-        )
-
-        # Decode only the newly generated tokens, not the prompt.
-        input_len = input_tokens["input_ids"].shape[1]
-        response = self._model.tokenizer.decode(
-            output[0][input_len:], skip_special_tokens=True
-        ).strip()
-
-        self._history.append(f"User: {message}")
-        self._history.append(f"Timmy: {response}")
-
-        return RunResult(content=response)
-
-    def print_response(self, message: str, *, stream: bool = True) -> None:
-        """Run inference and render the response to stdout (CLI interface)."""
-        result = self.run(message, stream=stream)
-        self._render(result.content)
-
-    # ── private helpers ──────────────────────────────────────────────────────
-
-    def _build_prompt(self, message: str) -> str:
-        context = SYSTEM_PROMPT + "\n\n"
-        # Include the last 10 turns (5 exchanges) for continuity.
-        if self._history:
-            context += "\n".join(self._history[-10:]) + "\n\n"
-        return context + f"User: {message}\nTimmy:"
-
-    @staticmethod
-    def _render(text: str) -> None:
-        """Print response with rich markdown when available, plain text otherwise."""
-        try:
-            from rich.console import Console
-            from rich.markdown import Markdown
-
-            Console().print(Markdown(text))
-        except ImportError:
-            print(text)
-
-
 # ── Grok (xAI) Backend ─────────────────────────────────────────────────────
 # Premium cloud augmentation — opt-in only, never the default path.
 
@@ -186,7 +74,7 @@ class GrokBackend:
     Uses the OpenAI-compatible SDK to connect to xAI's API.
     Only activated when GROK_ENABLED=true and XAI_API_KEY is set.
 
-    Exposes the same interface as TimmyAirLLMAgent and Agno Agent:
+    Exposes the same interface as Agno Agent:
       run(message, stream)           → RunResult  [dashboard]
       print_response(message, stream) → None       [CLI]
       health_check()                 → dict        [monitoring]
@@ -388,7 +276,9 @@ class GrokBackend:
 
     def _build_messages(self, message: str) -> list[dict[str, str]]:
         """Build the messages array for the API call."""
-        messages = [{"role": "system", "content": SYSTEM_PROMPT}]
+        messages = [
+            {"role": "system", "content": get_system_prompt(tools_enabled=True, session_id="grok")}
+        ]
         # Include conversation history for context
         messages.extend(self._history[-10:])
         messages.append({"role": "user", "content": message})
@@ -414,7 +304,8 @@ def grok_available() -> bool:
         from config import settings
 
         return settings.grok_enabled and bool(settings.xai_api_key)
-    except Exception:
+    except Exception as exc:
+        logger.warning("Backend check failed (grok_available): %s", exc)
         return False
 
 
@@ -433,8 +324,7 @@ CLAUDE_MODELS: dict[str, str] = {
 class ClaudeBackend:
     """Anthropic Claude backend — cloud fallback when local models are offline.
 
-    Uses the official Anthropic SDK.  Same interface as GrokBackend and
-    TimmyAirLLMAgent:
+    Uses the official Anthropic SDK.  Same interface as GrokBackend:
       run(message, stream)           → RunResult  [dashboard]
       print_response(message, stream) → None       [CLI]
       health_check()                 → dict        [monitoring]
@@ -480,7 +370,7 @@ class ClaudeBackend:
             response = client.messages.create(
                 model=self._model,
                 max_tokens=1024,
-                system=SYSTEM_PROMPT,
+                system=get_system_prompt(tools_enabled=True, session_id="claude"),
                 messages=messages,
             )
 
@@ -566,5 +456,6 @@ def claude_available() -> bool:
         from config import settings
 
         return bool(settings.anthropic_api_key)
-    except Exception:
+    except Exception as exc:
+        logger.warning("Backend check failed (claude_available): %s", exc)
         return False

src/timmy/briefing.py

@@ -10,6 +10,8 @@ regenerates the briefing every 6 hours.
 
 import logging
 import sqlite3
+from collections.abc import Generator
+from contextlib import closing, contextmanager
 from dataclasses import dataclass, field
 from datetime import UTC, datetime, timedelta
 from pathlib import Path
@@ -56,46 +58,45 @@ class Briefing:
 # ---------------------------------------------------------------------------
 
 
-def _get_cache_conn(db_path: Path = _DEFAULT_DB) -> sqlite3.Connection:
+@contextmanager
+def _get_cache_conn(db_path: Path = _DEFAULT_DB) -> Generator[sqlite3.Connection, None, None]:
     db_path.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(db_path))
-    conn.row_factory = sqlite3.Row
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS briefings (
-            id          INTEGER PRIMARY KEY AUTOINCREMENT,
-            generated_at TEXT NOT NULL,
-            period_start TEXT NOT NULL,
-            period_end   TEXT NOT NULL,
-            summary      TEXT NOT NULL
-        )
-        """)
-    conn.commit()
-    return conn
+    with closing(sqlite3.connect(str(db_path))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS briefings (
+                id          INTEGER PRIMARY KEY AUTOINCREMENT,
+                generated_at TEXT NOT NULL,
+                period_start TEXT NOT NULL,
+                period_end   TEXT NOT NULL,
+                summary      TEXT NOT NULL
+            )
+            """)
+        conn.commit()
+        yield conn
 
 
 def _save_briefing(briefing: Briefing, db_path: Path = _DEFAULT_DB) -> None:
-    conn = _get_cache_conn(db_path)
-    conn.execute(
-        """
-        INSERT INTO briefings (generated_at, period_start, period_end, summary)
-        VALUES (?, ?, ?, ?)
-        """,
-        (
-            briefing.generated_at.isoformat(),
-            briefing.period_start.isoformat(),
-            briefing.period_end.isoformat(),
-            briefing.summary,
-        ),
-    )
-    conn.commit()
-    conn.close()
+    with _get_cache_conn(db_path) as conn:
+        conn.execute(
+            """
+            INSERT INTO briefings (generated_at, period_start, period_end, summary)
+            VALUES (?, ?, ?, ?)
+            """,
+            (
+                briefing.generated_at.isoformat(),
+                briefing.period_start.isoformat(),
+                briefing.period_end.isoformat(),
+                briefing.summary,
+            ),
+        )
+        conn.commit()
 
 
 def _load_latest(db_path: Path = _DEFAULT_DB) -> Briefing | None:
     """Load the most-recently cached briefing, or None if there is none."""
-    conn = _get_cache_conn(db_path)
-    row = conn.execute("SELECT * FROM briefings ORDER BY generated_at DESC LIMIT 1").fetchone()
-    conn.close()
+    with _get_cache_conn(db_path) as conn:
+        row = conn.execute("SELECT * FROM briefings ORDER BY generated_at DESC LIMIT 1").fetchone()
     if row is None:
         return None
     return Briefing(
@@ -129,27 +130,25 @@ def _gather_swarm_summary(since: datetime) -> str:
         return "No swarm activity recorded yet."
 
     try:
-        conn = sqlite3.connect(str(swarm_db))
-        conn.row_factory = sqlite3.Row
+        with closing(sqlite3.connect(str(swarm_db))) as conn:
+            conn.row_factory = sqlite3.Row
 
-        since_iso = since.isoformat()
+            since_iso = since.isoformat()
 
-        completed = conn.execute(
-            "SELECT COUNT(*) as c FROM tasks WHERE status = 'completed' AND created_at > ?",
-            (since_iso,),
-        ).fetchone()["c"]
+            completed = conn.execute(
+                "SELECT COUNT(*) as c FROM tasks WHERE status = 'completed' AND created_at > ?",
+                (since_iso,),
+            ).fetchone()["c"]
 
-        failed = conn.execute(
-            "SELECT COUNT(*) as c FROM tasks WHERE status = 'failed' AND created_at > ?",
-            (since_iso,),
-        ).fetchone()["c"]
+            failed = conn.execute(
+                "SELECT COUNT(*) as c FROM tasks WHERE status = 'failed' AND created_at > ?",
+                (since_iso,),
+            ).fetchone()["c"]
 
-        agents = conn.execute(
-            "SELECT COUNT(*) as c FROM agents WHERE registered_at > ?",
-            (since_iso,),
-        ).fetchone()["c"]
-
-        conn.close()
+            agents = conn.execute(
+                "SELECT COUNT(*) as c FROM agents WHERE registered_at > ?",
+                (since_iso,),
+            ).fetchone()["c"]
 
         parts = []
         if completed:
@@ -193,7 +192,7 @@ def _gather_task_queue_summary() -> str:
 def _gather_chat_summary(since: datetime) -> str:
     """Pull recent chat messages from the in-memory log."""
     try:
-        from dashboard.store import message_log
+        from infrastructure.chat_store import message_log
 
         messages = message_log.all()
         # Filter to messages in the briefing window (best-effort: no timestamps)

src/timmy/cli.py

@@ -1,3 +1,4 @@
+import asyncio
 import logging
 import subprocess
 import sys
@@ -21,13 +22,13 @@ _BACKEND_OPTION = typer.Option(
     None,
     "--backend",
     "-b",
-    help="Inference backend: 'ollama' (default) | 'airllm' | 'auto'",
+    help="Inference backend: 'ollama' (default) | 'grok' | 'claude'",
 )
 _MODEL_SIZE_OPTION = typer.Option(
     None,
     "--model-size",
     "-s",
-    help="AirLLM model size when --backend airllm: '8b' | '70b' | '405b'",
+    help="Model size (reserved for future use).",
 )
 
 
@@ -137,13 +138,15 @@ def think(
     model_size: str | None = _MODEL_SIZE_OPTION,
 ):
     """Ask Timmy to think carefully about a topic."""
-    timmy = create_timmy(backend=backend, model_size=model_size)
+    timmy = create_timmy(backend=backend, model_size=model_size, session_id=_CLI_SESSION_ID)
     timmy.print_response(f"Think carefully about: {topic}", stream=True, session_id=_CLI_SESSION_ID)
 
 
 @app.command()
 def chat(
-    message: str = typer.Argument(..., help="Message to send"),
+    message: list[str] = typer.Argument(
+        ..., help="Message to send (multiple words are joined automatically)"
+    ),
     backend: str | None = _BACKEND_OPTION,
     model_size: str | None = _MODEL_SIZE_OPTION,
     new_session: bool = typer.Option(
@@ -172,19 +175,36 @@ def chat(
     Use --autonomous for non-interactive contexts (scripts, dev loops). Tool
     calls are checked against config/allowlist.yaml — allowlisted operations
     execute automatically, everything else is safely rejected.
+
+    Read from stdin by passing "-" as the message or piping input.
     """
     import uuid
 
+    # Join multiple arguments into a single message string
+    message_str = " ".join(message)
+
+    # Handle stdin input if "-" is passed or stdin is not a tty
+    if message_str == "-" or not _is_interactive():
+        try:
+            stdin_content = sys.stdin.read().strip()
+        except (KeyboardInterrupt, EOFError):
+            stdin_content = ""
+        if stdin_content:
+            message_str = stdin_content
+        elif message_str == "-":
+            typer.echo("No input provided via stdin.", err=True)
+            raise typer.Exit(1)
+
     if session_id is not None:
         pass  # use the provided value
     elif new_session:
         session_id = str(uuid.uuid4())
     else:
         session_id = _CLI_SESSION_ID
-    timmy = create_timmy(backend=backend, model_size=model_size)
+    timmy = create_timmy(backend=backend, model_size=model_size, session_id=session_id)
 
     # Use agent.run() so we can intercept paused runs for tool confirmation.
-    run_output = timmy.run(message, stream=False, session_id=session_id)
+    run_output = timmy.run(message_str, stream=False, session_id=session_id)
 
     # Handle paused runs — dangerous tools need user approval
     run_output = _handle_tool_confirmation(timmy, run_output, session_id, autonomous=autonomous)
@@ -197,13 +217,68 @@ def chat(
         typer.echo(_clean_response(content))
 
 
+@app.command()
+def repl(
+    backend: str | None = _BACKEND_OPTION,
+    model_size: str | None = _MODEL_SIZE_OPTION,
+    session_id: str | None = typer.Option(
+        None,
+        "--session-id",
+        help="Use a specific session ID for this conversation",
+    ),
+):
+    """Start an interactive REPL session with Timmy.
+
+    Keeps the agent warm between messages. Conversation history is persisted
+    across invocations. Use Ctrl+C or Ctrl+D to exit gracefully.
+    """
+    from timmy.session import chat
+
+    if session_id is None:
+        session_id = _CLI_SESSION_ID
+
+    typer.echo(typer.style("Timmy REPL", bold=True))
+    typer.echo("Type your messages below. Use Ctrl+C or Ctrl+D to exit.")
+    typer.echo()
+
+    loop = asyncio.new_event_loop()
+    asyncio.set_event_loop(loop)
+
+    try:
+        while True:
+            try:
+                user_input = input("> ")
+            except (KeyboardInterrupt, EOFError):
+                typer.echo()
+                typer.echo("Goodbye!")
+                break
+
+            user_input = user_input.strip()
+            if not user_input:
+                continue
+
+            if user_input.lower() in ("exit", "quit", "q"):
+                typer.echo("Goodbye!")
+                break
+
+            try:
+                response = loop.run_until_complete(chat(user_input, session_id=session_id))
+                if response:
+                    typer.echo(response)
+                    typer.echo()
+            except Exception as exc:
+                typer.echo(f"Error: {exc}", err=True)
+    finally:
+        loop.close()
+
+
 @app.command()
 def status(
     backend: str | None = _BACKEND_OPTION,
     model_size: str | None = _MODEL_SIZE_OPTION,
 ):
     """Print Timmy's operational status."""
-    timmy = create_timmy(backend=backend, model_size=model_size)
+    timmy = create_timmy(backend=backend, model_size=model_size, session_id=_CLI_SESSION_ID)
     timmy.print_response(STATUS_PROMPT, stream=False, session_id=_CLI_SESSION_ID)
 
 
@@ -259,7 +334,8 @@ def interview(
             from timmy.mcp_tools import close_mcp_sessions
 
             loop.run_until_complete(close_mcp_sessions())
-        except Exception:
+        except Exception as exc:
+            logger.warning("MCP session close failed: %s", exc)
             pass
         loop.close()
 
@@ -325,5 +401,55 @@ def voice(
     loop.run()
 
 
+@app.command()
+def route(
+    message: list[str] = typer.Argument(..., help="Message to route"),
+):
+    """Show which agent would handle a message (debug routing)."""
+    full_message = " ".join(message)
+    from timmy.agents.loader import route_request_with_match
+
+    agent_id, matched_pattern = route_request_with_match(full_message)
+    if agent_id:
+        typer.echo(f"→ {agent_id} (matched: {matched_pattern})")
+    else:
+        typer.echo("→ orchestrator (no pattern match)")
+
+
+@app.command()
+def focus(
+    topic: str | None = typer.Argument(
+        None, help='Topic to focus on (e.g. "three-phase loop"). Omit to show current focus.'
+    ),
+    clear: bool = typer.Option(False, "--clear", "-c", help="Clear focus and return to broad mode"),
+):
+    """Set deep-focus mode on a single problem.
+
+    When focused, Timmy prioritizes the active topic in all responses
+    and deprioritizes unrelated context. Focus persists across sessions.
+
+    Examples:
+        timmy focus "three-phase loop"   # activate deep focus
+        timmy focus                       # show current focus
+        timmy focus --clear               # return to broad mode
+    """
+    from timmy.focus import focus_manager
+
+    if clear:
+        focus_manager.clear()
+        typer.echo("Focus cleared — back to broad mode.")
+        return
+
+    if topic:
+        focus_manager.set_topic(topic)
+        typer.echo(f'Deep focus activated: "{topic}"')
+    else:
+        # Show current focus status
+        if focus_manager.is_focused():
+            typer.echo(f'Deep focus: "{focus_manager.get_topic()}"')
+        else:
+            typer.echo("No active focus (broad mode).")
+
+
 def main():
     app()

src/timmy/cognitive_state.py

@@ -0,0 +1,250 @@
+"""Observable cognitive state for Timmy.
+
+Tracks Timmy's internal cognitive signals — focus, engagement, mood,
+and active commitments — so external systems (Matrix avatar, dashboard)
+can render observable behaviour.
+
+State is published via ``workshop_state.py`` → ``presence.json`` and the
+WebSocket relay.  The old ``~/.tower/timmy-state.txt`` file has been
+deprecated (see #384).
+"""
+
+import asyncio
+import json
+import logging
+from dataclasses import asdict, dataclass, field
+
+from timmy.confidence import estimate_confidence
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Schema
+# ---------------------------------------------------------------------------
+
+ENGAGEMENT_LEVELS = ("idle", "surface", "deep")
+MOOD_VALUES = ("curious", "settled", "hesitant", "energized")
+
+
+@dataclass
+class CognitiveState:
+    """Observable snapshot of Timmy's cognitive state."""
+
+    focus_topic: str | None = None
+    engagement: str = "idle"  # idle | surface | deep
+    mood: str = "settled"  # curious | settled | hesitant | energized
+    conversation_depth: int = 0
+    last_initiative: str | None = None
+    active_commitments: list[str] = field(default_factory=list)
+
+    # Internal tracking (not written to state file)
+    _confidence_sum: float = field(default=0.0, repr=False)
+    _confidence_count: int = field(default=0, repr=False)
+
+    # ------------------------------------------------------------------
+    # Serialisation helpers
+    # ------------------------------------------------------------------
+
+    def to_dict(self) -> dict:
+        """Public fields only (exclude internal tracking)."""
+        d = asdict(self)
+        d.pop("_confidence_sum", None)
+        d.pop("_confidence_count", None)
+        return d
+
+
+# ---------------------------------------------------------------------------
+# Cognitive signal extraction
+# ---------------------------------------------------------------------------
+
+# Keywords that suggest deep engagement
+_DEEP_KEYWORDS = frozenset(
+    {
+        "architecture",
+        "design",
+        "implement",
+        "refactor",
+        "debug",
+        "analyze",
+        "investigate",
+        "deep dive",
+        "explain how",
+        "walk me through",
+        "step by step",
+    }
+)
+
+# Keywords that suggest initiative / commitment
+_COMMITMENT_KEYWORDS = frozenset(
+    {
+        "i will",
+        "i'll",
+        "let me",
+        "i'm going to",
+        "plan to",
+        "commit to",
+        "i propose",
+        "i suggest",
+    }
+)
+
+
+def _infer_engagement(message: str, response: str) -> str:
+    """Classify engagement level from the exchange."""
+    combined = (message + " " + response).lower()
+    if any(kw in combined for kw in _DEEP_KEYWORDS):
+        return "deep"
+    # Short exchanges are surface-level
+    if len(response.split()) < 15:
+        return "surface"
+    return "surface"
+
+
+def _infer_mood(response: str, confidence: float) -> str:
+    """Derive mood from response signals."""
+    lower = response.lower()
+    if confidence < 0.4:
+        return "hesitant"
+    if "!" in response and any(w in lower for w in ("great", "exciting", "love", "awesome")):
+        return "energized"
+    if "?" in response or any(w in lower for w in ("wonder", "interesting", "curious", "hmm")):
+        return "curious"
+    return "settled"
+
+
+def _extract_topic(message: str) -> str | None:
+    """Best-effort topic extraction from the user message.
+
+    Takes the first meaningful clause (up to 60 chars) as a topic label.
+    """
+    text = message.strip()
+    if not text:
+        return None
+    # Strip leading question words
+    for prefix in ("what is ", "how do ", "can you ", "please ", "hey timmy "):
+        if text.lower().startswith(prefix):
+            text = text[len(prefix) :]
+    # Truncate
+    if len(text) > 60:
+        text = text[:57] + "..."
+    return text.strip() or None
+
+
+def _extract_commitments(response: str) -> list[str]:
+    """Pull commitment phrases from Timmy's response."""
+    commitments: list[str] = []
+    lower = response.lower()
+    for kw in _COMMITMENT_KEYWORDS:
+        idx = lower.find(kw)
+        if idx == -1:
+            continue
+        # Grab the rest of the sentence (up to period/newline, max 80 chars)
+        start = idx
+        end = len(lower)
+        for sep in (".", "\n", "!"):
+            pos = lower.find(sep, start)
+            if pos != -1:
+                end = min(end, pos)
+        snippet = response[start : min(end, start + 80)].strip()
+        if snippet:
+            commitments.append(snippet)
+    return commitments[:3]  # Cap at 3
+
+
+# ---------------------------------------------------------------------------
+# Tracker singleton
+# ---------------------------------------------------------------------------
+
+
+class CognitiveTracker:
+    """Maintains Timmy's cognitive state.
+
+    State is consumed via ``to_json()`` / ``get_state()`` and published
+    externally by ``workshop_state.py`` → ``presence.json``.
+    """
+
+    def __init__(self) -> None:
+        self.state = CognitiveState()
+
+    def update(self, user_message: str, response: str) -> CognitiveState:
+        """Update cognitive state from a chat exchange.
+
+        Called after each chat round-trip in ``session.py``.
+        Emits a ``cognitive_state_changed`` event to the sensory bus so
+        downstream consumers (WorkshopHeartbeat, etc.) react immediately.
+        """
+        confidence = estimate_confidence(response)
+
+        prev_mood = self.state.mood
+        prev_engagement = self.state.engagement
+
+        # Track running confidence average
+        self.state._confidence_sum += confidence
+        self.state._confidence_count += 1
+
+        self.state.conversation_depth += 1
+        self.state.focus_topic = _extract_topic(user_message) or self.state.focus_topic
+        self.state.engagement = _infer_engagement(user_message, response)
+        self.state.mood = _infer_mood(response, confidence)
+
+        # Extract commitments from response
+        new_commitments = _extract_commitments(response)
+        if new_commitments:
+            self.state.last_initiative = new_commitments[0]
+            # Merge, keeping last 5
+            seen = set(self.state.active_commitments)
+            for c in new_commitments:
+                if c not in seen:
+                    self.state.active_commitments.append(c)
+                    seen.add(c)
+            self.state.active_commitments = self.state.active_commitments[-5:]
+
+        # Emit cognitive_state_changed to close the sense → react loop
+        self._emit_change(prev_mood, prev_engagement)
+
+        return self.state
+
+    def _emit_change(self, prev_mood: str, prev_engagement: str) -> None:
+        """Fire-and-forget sensory event for cognitive state change."""
+        try:
+            from timmy.event_bus import get_sensory_bus
+            from timmy.events import SensoryEvent
+
+            event = SensoryEvent(
+                source="cognitive",
+                event_type="cognitive_state_changed",
+                data={
+                    "mood": self.state.mood,
+                    "engagement": self.state.engagement,
+                    "focus_topic": self.state.focus_topic or "",
+                    "depth": self.state.conversation_depth,
+                    "mood_changed": self.state.mood != prev_mood,
+                    "engagement_changed": self.state.engagement != prev_engagement,
+                },
+            )
+            bus = get_sensory_bus()
+            # Fire-and-forget — don't block the chat response
+            try:
+                loop = asyncio.get_running_loop()
+                loop.create_task(bus.emit(event))
+            except RuntimeError:
+                # No running loop (sync context / tests) — skip emission
+                pass
+        except Exception as exc:
+            logger.debug("Cognitive event emission skipped: %s", exc)
+
+    def get_state(self) -> CognitiveState:
+        """Return current cognitive state."""
+        return self.state
+
+    def reset(self) -> None:
+        """Reset to idle state (e.g. on session reset)."""
+        self.state = CognitiveState()
+
+    def to_json(self) -> str:
+        """Serialise current state as JSON (for API / WebSocket consumers)."""
+        return json.dumps(self.state.to_dict())
+
+
+# Module-level singleton
+cognitive_tracker = CognitiveTracker()

src/timmy/confidence.py

@@ -0,0 +1,128 @@
+"""Confidence estimation for Timmy's responses.
+
+Implements SOUL.md requirement: "When I am uncertain, I must say so in
+proportion to my uncertainty."
+
+This module provides heuristics to estimate confidence based on linguistic
+signals in the response text. It measures uncertainty without modifying
+the response content.
+"""
+
+import re
+
+# Hedging words that indicate uncertainty
+HEDGING_WORDS = [
+    "i think",
+    "maybe",
+    "perhaps",
+    "not sure",
+    "might",
+    "could be",
+    "possibly",
+    "i believe",
+    "approximately",
+    "roughly",
+    "probably",
+    "likely",
+    "seems",
+    "appears",
+    "suggests",
+    "i guess",
+    "i suppose",
+    "sort of",
+    "kind of",
+    "somewhat",
+    "fairly",
+    "relatively",
+    "i'm not certain",
+    "i am not certain",
+    "uncertain",
+    "unclear",
+]
+
+# Certainty words that indicate confidence
+CERTAINTY_WORDS = [
+    "i know",
+    "definitely",
+    "certainly",
+    "the answer is",
+    "specifically",
+    "exactly",
+    "absolutely",
+    "without doubt",
+    "i am certain",
+    "i'm certain",
+    "it is true that",
+    "fact is",
+    "in fact",
+    "indeed",
+    "undoubtedly",
+    "clearly",
+    "obviously",
+    "conclusively",
+]
+
+# Very low confidence indicators (direct admissions of ignorance)
+LOW_CONFIDENCE_PATTERNS = [
+    r"i\s+(?:don't|do not)\s+know",
+    r"i\s+(?:am|I'm|i'm)\s+(?:not\s+sure|unsure)",
+    r"i\s+have\s+no\s+(?:idea|clue)",
+    r"i\s+cannot\s+(?:say|tell|answer)",
+    r"i\s+can't\s+(?:say|tell|answer)",
+]
+
+
+def estimate_confidence(text: str) -> float:
+    """Estimate confidence level of a response based on linguistic signals.
+
+    Analyzes the text for hedging words (reducing confidence) and certainty
+    words (increasing confidence). Returns a score between 0.0 and 1.0.
+
+    Args:
+        text: The response text to analyze.
+
+    Returns:
+        A float between 0.0 (very uncertain) and 1.0 (very confident).
+    """
+    if not text or not text.strip():
+        return 0.0
+
+    text_lower = text.lower().strip()
+    confidence = 0.5  # Start with neutral confidence
+
+    # Check for direct admissions of ignorance (very low confidence)
+    for pattern in LOW_CONFIDENCE_PATTERNS:
+        if re.search(pattern, text_lower):
+            # Direct admission of not knowing - very low confidence
+            confidence = 0.15
+            break
+
+    # Count hedging words (reduce confidence)
+    hedging_count = 0
+    for hedge in HEDGING_WORDS:
+        if hedge in text_lower:
+            hedging_count += 1
+
+    # Count certainty words (increase confidence)
+    certainty_count = 0
+    for certain in CERTAINTY_WORDS:
+        if certain in text_lower:
+            certainty_count += 1
+
+    # Adjust confidence based on word counts
+    # Each hedging word reduces confidence by 0.1
+    # Each certainty word increases confidence by 0.1
+    confidence -= hedging_count * 0.1
+    confidence += certainty_count * 0.1
+
+    # Short factual answers get a small boost
+    word_count = len(text.split())
+    if word_count <= 5 and confidence > 0.3:
+        confidence += 0.1
+
+    # Questions in response indicate uncertainty
+    if "?" in text:
+        confidence -= 0.15
+
+    # Clamp to valid range
+    return max(0.0, min(1.0, confidence))

src/timmy/event_bus.py

@@ -0,0 +1,79 @@
+"""Sensory EventBus — simple pub/sub for SensoryEvents.
+
+Thin facade over the infrastructure EventBus that speaks in
+SensoryEvent objects instead of raw infrastructure Events.
+"""
+
+import asyncio
+import logging
+from collections.abc import Awaitable, Callable
+
+from timmy.events import SensoryEvent
+
+logger = logging.getLogger(__name__)
+
+# Handler: sync or async callable that receives a SensoryEvent
+SensoryHandler = Callable[[SensoryEvent], None | Awaitable[None]]
+
+
+class SensoryBus:
+    """Pub/sub dispatcher for SensoryEvents."""
+
+    def __init__(self, max_history: int = 500) -> None:
+        self._subscribers: dict[str, list[SensoryHandler]] = {}
+        self._history: list[SensoryEvent] = []
+        self._max_history = max_history
+
+    # ── Public API ────────────────────────────────────────────────────────
+
+    async def emit(self, event: SensoryEvent) -> int:
+        """Push *event* to all subscribers whose event_type filter matches.
+
+        Returns the number of handlers invoked.
+        """
+        self._history.append(event)
+        if len(self._history) > self._max_history:
+            self._history = self._history[-self._max_history :]
+
+        handlers = self._matching_handlers(event.event_type)
+        for h in handlers:
+            try:
+                result = h(event)
+                if asyncio.iscoroutine(result):
+                    await result
+            except Exception as exc:
+                logger.error("SensoryBus handler error for '%s': %s", event.event_type, exc)
+
+        return len(handlers)
+
+    def subscribe(self, event_type: str, callback: SensoryHandler) -> None:
+        """Register *callback* for events matching *event_type*.
+
+        Use ``"*"`` to subscribe to all event types.
+        """
+        self._subscribers.setdefault(event_type, []).append(callback)
+
+    def recent(self, n: int = 10) -> list[SensoryEvent]:
+        """Return the last *n* events (most recent last)."""
+        return self._history[-n:]
+
+    # ── Internals ─────────────────────────────────────────────────────────
+
+    def _matching_handlers(self, event_type: str) -> list[SensoryHandler]:
+        handlers: list[SensoryHandler] = []
+        for pattern, cbs in self._subscribers.items():
+            if pattern == "*" or pattern == event_type:
+                handlers.extend(cbs)
+        return handlers
+
+
+# ── Module-level singleton ────────────────────────────────────────────────────
+_bus: SensoryBus | None = None
+
+
+def get_sensory_bus() -> SensoryBus:
+    """Return the module-level SensoryBus singleton."""
+    global _bus
+    if _bus is None:
+        _bus = SensoryBus()
+    return _bus

src/timmy/events.py

@@ -0,0 +1,39 @@
+"""SensoryEvent — normalized event model for stream adapters.
+
+Every adapter (gitea, time, bitcoin, terminal, etc.) emits SensoryEvents
+into the EventBus so that Timmy's cognitive layer sees a uniform stream.
+"""
+
+import json
+from dataclasses import asdict, dataclass, field
+from datetime import UTC, datetime
+
+
+@dataclass
+class SensoryEvent:
+    """A single sensory event from an external stream."""
+
+    source: str  # "gitea", "time", "bitcoin", "terminal"
+    event_type: str  # "push", "issue_opened", "new_block", "morning"
+    timestamp: datetime = field(default_factory=lambda: datetime.now(UTC))
+    data: dict = field(default_factory=dict)
+    actor: str = ""  # who caused it (username, "system", etc.)
+
+    def to_dict(self) -> dict:
+        """Return a JSON-serializable dictionary."""
+        d = asdict(self)
+        d["timestamp"] = self.timestamp.isoformat()
+        return d
+
+    def to_json(self) -> str:
+        """Return a JSON string."""
+        return json.dumps(self.to_dict())
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "SensoryEvent":
+        """Reconstruct a SensoryEvent from a dictionary."""
+        data = dict(data)  # shallow copy
+        ts = data.get("timestamp")
+        if isinstance(ts, str):
+            data["timestamp"] = datetime.fromisoformat(ts)
+        return cls(**data)

src/timmy/familiar.py

@@ -0,0 +1,263 @@
+"""Pip the Familiar — a creature with its own small mind.
+
+Pip is a glowing sprite who lives in the Workshop independently of Timmy.
+He has a behavioral state machine that makes the room feel alive:
+
+    SLEEPING → WAKING → WANDERING → INVESTIGATING → BORED → SLEEPING
+
+Special states triggered by Timmy's cognitive signals:
+    ALERT    — confidence drops below 0.3
+    PLAYFUL  — Timmy is amused / energized
+    HIDING   — unknown visitor + Timmy uncertain
+
+The backend tracks Pip's *logical* state; the browser handles movement
+interpolation and particle rendering.
+"""
+
+import logging
+import random
+import time
+from dataclasses import asdict, dataclass, field
+from enum import StrEnum
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# States
+# ---------------------------------------------------------------------------
+
+
+class PipState(StrEnum):
+    """Pip's behavioral states."""
+
+    SLEEPING = "sleeping"
+    WAKING = "waking"
+    WANDERING = "wandering"
+    INVESTIGATING = "investigating"
+    BORED = "bored"
+    # Special states
+    ALERT = "alert"
+    PLAYFUL = "playful"
+    HIDING = "hiding"
+
+
+# States from which Pip can be interrupted by special triggers
+_INTERRUPTIBLE = frozenset(
+    {
+        PipState.SLEEPING,
+        PipState.WANDERING,
+        PipState.BORED,
+        PipState.WAKING,
+    }
+)
+
+# How long each state lasts before auto-transitioning (seconds)
+_STATE_DURATIONS: dict[PipState, tuple[float, float]] = {
+    PipState.SLEEPING: (120.0, 300.0),  # 2-5 min
+    PipState.WAKING: (1.5, 2.5),
+    PipState.WANDERING: (15.0, 45.0),
+    PipState.INVESTIGATING: (8.0, 12.0),
+    PipState.BORED: (20.0, 40.0),
+    PipState.ALERT: (10.0, 20.0),
+    PipState.PLAYFUL: (8.0, 15.0),
+    PipState.HIDING: (15.0, 30.0),
+}
+
+# Default position near the fireplace
+_FIREPLACE_POS = (2.1, 0.5, -1.3)
+
+
+# ---------------------------------------------------------------------------
+# Schema
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class PipSnapshot:
+    """Serialisable snapshot of Pip's current state."""
+
+    name: str = "Pip"
+    state: str = "sleeping"
+    position: tuple[float, float, float] = _FIREPLACE_POS
+    mood_mirror: str = "calm"
+    since: float = field(default_factory=time.monotonic)
+
+    def to_dict(self) -> dict:
+        """Public dict for API / WebSocket / state file consumers."""
+        d = asdict(self)
+        d["position"] = list(d["position"])
+        # Convert monotonic timestamp to duration
+        d["state_duration_s"] = round(time.monotonic() - d.pop("since"), 1)
+        return d
+
+
+# ---------------------------------------------------------------------------
+# Familiar
+# ---------------------------------------------------------------------------
+
+
+class Familiar:
+    """Pip's behavioral AI — a tiny state machine driven by events and time.
+
+    Usage::
+
+        pip_familiar.on_event("visitor_entered")
+        pip_familiar.on_mood_change("energized")
+        state = pip_familiar.tick()  # call periodically
+    """
+
+    def __init__(self) -> None:
+        self._state = PipState.SLEEPING
+        self._entered_at = time.monotonic()
+        self._duration = random.uniform(*_STATE_DURATIONS[PipState.SLEEPING])
+        self._mood_mirror = "calm"
+        self._pending_mood: str | None = None
+        self._mood_change_at: float = 0.0
+        self._position = _FIREPLACE_POS
+
+    # ------------------------------------------------------------------
+    # Public API
+    # ------------------------------------------------------------------
+
+    @property
+    def state(self) -> PipState:
+        return self._state
+
+    @property
+    def mood_mirror(self) -> str:
+        return self._mood_mirror
+
+    def snapshot(self) -> PipSnapshot:
+        """Current state as a serialisable snapshot."""
+        return PipSnapshot(
+            state=self._state.value,
+            position=self._position,
+            mood_mirror=self._mood_mirror,
+            since=self._entered_at,
+        )
+
+    def tick(self, now: float | None = None) -> PipState:
+        """Advance the state machine. Call periodically (e.g. every second).
+
+        Returns the (possibly new) state.
+        """
+        now = now if now is not None else time.monotonic()
+
+        # Apply delayed mood mirror (3-second lag)
+        if self._pending_mood and now >= self._mood_change_at:
+            self._mood_mirror = self._pending_mood
+            self._pending_mood = None
+
+        # Check if current state has expired
+        elapsed = now - self._entered_at
+        if elapsed < self._duration:
+            return self._state
+
+        # Auto-transition
+        next_state = self._next_state()
+        self._transition(next_state, now)
+        return self._state
+
+    def on_event(self, event: str, now: float | None = None) -> PipState:
+        """React to a Workshop event.
+
+        Supported events:
+            visitor_entered, visitor_spoke, loud_event, scroll_knocked
+        """
+        now = now if now is not None else time.monotonic()
+
+        if event == "visitor_entered" and self._state in _INTERRUPTIBLE:
+            if self._state == PipState.SLEEPING:
+                self._transition(PipState.WAKING, now)
+            else:
+                self._transition(PipState.INVESTIGATING, now)
+
+        elif event == "visitor_spoke":
+            if self._state in (PipState.WANDERING, PipState.WAKING):
+                self._transition(PipState.INVESTIGATING, now)
+
+        elif event == "loud_event":
+            if self._state == PipState.SLEEPING:
+                self._transition(PipState.WAKING, now)
+
+        return self._state
+
+    def on_mood_change(
+        self,
+        timmy_mood: str,
+        confidence: float = 0.5,
+        now: float | None = None,
+    ) -> PipState:
+        """Mirror Timmy's mood with a 3-second delay.
+
+        Special states triggered by mood + confidence:
+            - confidence < 0.3 → ALERT (bristles, particles go red-gold)
+            - mood == "energized" → PLAYFUL (figure-8s around crystal ball)
+            - mood == "hesitant" + confidence < 0.4 → HIDING
+        """
+        now = now if now is not None else time.monotonic()
+
+        # Schedule mood mirror with 3s delay
+        self._pending_mood = timmy_mood
+        self._mood_change_at = now + 3.0
+
+        # Special state triggers (immediate)
+        if confidence < 0.3 and self._state in _INTERRUPTIBLE:
+            self._transition(PipState.ALERT, now)
+        elif timmy_mood == "energized" and self._state in _INTERRUPTIBLE:
+            self._transition(PipState.PLAYFUL, now)
+        elif timmy_mood == "hesitant" and confidence < 0.4 and self._state in _INTERRUPTIBLE:
+            self._transition(PipState.HIDING, now)
+
+        return self._state
+
+    # ------------------------------------------------------------------
+    # Internals
+    # ------------------------------------------------------------------
+
+    def _transition(self, new_state: PipState, now: float) -> None:
+        """Move to a new state."""
+        old = self._state
+        self._state = new_state
+        self._entered_at = now
+        self._duration = random.uniform(*_STATE_DURATIONS[new_state])
+        self._position = self._position_for(new_state)
+        logger.debug("Pip: %s → %s", old.value, new_state.value)
+
+    def _next_state(self) -> PipState:
+        """Determine the natural next state after the current one expires."""
+        transitions: dict[PipState, PipState] = {
+            PipState.SLEEPING: PipState.WAKING,
+            PipState.WAKING: PipState.WANDERING,
+            PipState.WANDERING: PipState.BORED,
+            PipState.INVESTIGATING: PipState.BORED,
+            PipState.BORED: PipState.SLEEPING,
+            # Special states return to wandering
+            PipState.ALERT: PipState.WANDERING,
+            PipState.PLAYFUL: PipState.WANDERING,
+            PipState.HIDING: PipState.WAKING,
+        }
+        return transitions.get(self._state, PipState.SLEEPING)
+
+    def _position_for(self, state: PipState) -> tuple[float, float, float]:
+        """Approximate position hint for a given state.
+
+        The browser interpolates smoothly; these are target anchors.
+        """
+        if state in (PipState.SLEEPING, PipState.BORED):
+            return _FIREPLACE_POS
+        if state == PipState.HIDING:
+            return (0.5, 0.3, -2.0)  # Behind the desk
+        if state == PipState.PLAYFUL:
+            return (1.0, 1.2, 0.0)  # Near the crystal ball
+        # Wandering / investigating / waking — random room position
+        return (
+            random.uniform(-1.0, 3.0),
+            random.uniform(0.5, 1.5),
+            random.uniform(-2.0, 1.0),
+        )
+
+
+# Module-level singleton
+pip_familiar = Familiar()

src/timmy/focus.py

@@ -0,0 +1,105 @@
+"""Deep focus mode — single-problem context for Timmy.
+
+Persists focus state to a JSON file so Timmy can maintain narrow,
+deep attention on one problem across session restarts.
+
+Usage:
+    from timmy.focus import focus_manager
+
+    focus_manager.set_topic("three-phase loop")
+    topic = focus_manager.get_topic()       # "three-phase loop"
+    ctx   = focus_manager.get_focus_context()  # prompt injection string
+    focus_manager.clear()
+"""
+
+import json
+import logging
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_STATE_DIR = Path.home() / ".timmy"
+_STATE_FILE = "focus.json"
+
+
+class FocusManager:
+    """Manages deep-focus state with file-backed persistence."""
+
+    def __init__(self, state_dir: Path | None = None) -> None:
+        self._state_dir = state_dir or _DEFAULT_STATE_DIR
+        self._state_file = self._state_dir / _STATE_FILE
+        self._topic: str | None = None
+        self._mode: str = "broad"
+        self._load()
+
+    # ── Public API ────────────────────────────────────────────────
+
+    def get_topic(self) -> str | None:
+        """Return the current focus topic, or None if unfocused."""
+        return self._topic
+
+    def get_mode(self) -> str:
+        """Return 'deep' or 'broad'."""
+        return self._mode
+
+    def is_focused(self) -> bool:
+        """True when deep-focus is active with a topic set."""
+        return self._mode == "deep" and self._topic is not None
+
+    def set_topic(self, topic: str) -> None:
+        """Activate deep focus on a specific topic."""
+        self._topic = topic.strip()
+        self._mode = "deep"
+        self._save()
+        logger.info("Focus: deep-focus set → %r", self._topic)
+
+    def clear(self) -> None:
+        """Return to broad (unfocused) mode."""
+        old = self._topic
+        self._topic = None
+        self._mode = "broad"
+        self._save()
+        logger.info("Focus: cleared (was %r)", old)
+
+    def get_focus_context(self) -> str:
+        """Return a prompt-injection string for the current focus state.
+
+        When focused, this tells the model to prioritize the topic.
+        When broad, returns an empty string (no injection).
+        """
+        if not self.is_focused():
+            return ""
+        return (
+            f"[DEEP FOCUS MODE] You are currently in deep-focus mode on: "
+            f'"{self._topic}". '
+            f"Prioritize this topic in your responses. Surface related memories "
+            f"and prior conversation about this topic first. Deprioritize "
+            f"unrelated context. Stay focused — depth over breadth."
+        )
+
+    # ── Persistence ───────────────────────────────────────────────
+
+    def _load(self) -> None:
+        """Load focus state from disk."""
+        if not self._state_file.exists():
+            return
+        try:
+            data = json.loads(self._state_file.read_text())
+            self._topic = data.get("topic")
+            self._mode = data.get("mode", "broad")
+        except Exception as exc:
+            logger.warning("Focus: failed to load state: %s", exc)
+
+    def _save(self) -> None:
+        """Persist focus state to disk."""
+        try:
+            self._state_dir.mkdir(parents=True, exist_ok=True)
+            self._state_file.write_text(
+                json.dumps({"topic": self._topic, "mode": self._mode}, indent=2)
+            )
+        except Exception as exc:
+            logger.warning("Focus: failed to save state: %s", exc)
+
+
+# Module-level singleton
+focus_manager = FocusManager()

src/timmy/gematria.py

@@ -0,0 +1,387 @@
+"""Gematria computation engine — the language of letters and numbers.
+
+Implements multiple cipher systems for gematric analysis:
+  - Simple English (A=1 .. Z=26)
+  - Full Reduction (reduce each letter value to single digit)
+  - Reverse Ordinal (A=26 .. Z=1)
+  - Sumerian (Simple × 6)
+  - Hebrew (traditional letter values, for A-Z mapping)
+
+Also provides numerological reduction, notable-number lookup,
+and multi-phrase comparison.
+
+Alexander Whitestone = 222 in Simple English Gematria.
+This is not trivia.  It is foundational.
+"""
+
+from __future__ import annotations
+
+import math
+
+# ── Cipher Tables ────────────────────────────────────────────────────────────
+
+# Simple English: A=1, B=2, ..., Z=26
+_SIMPLE: dict[str, int] = {chr(i): i - 64 for i in range(65, 91)}
+
+# Full Reduction: reduce each letter to single digit (A=1..I=9, J=1..R=9, S=1..Z=8)
+_REDUCTION: dict[str, int] = {}
+for _c, _v in _SIMPLE.items():
+    _r = _v
+    while _r > 9:
+        _r = sum(int(d) for d in str(_r))
+    _REDUCTION[_c] = _r
+
+# Reverse Ordinal: A=26, B=25, ..., Z=1
+_REVERSE: dict[str, int] = {chr(i): 91 - i for i in range(65, 91)}
+
+# Sumerian: Simple × 6
+_SUMERIAN: dict[str, int] = {c: v * 6 for c, v in _SIMPLE.items()}
+
+# Hebrew-mapped: traditional Hebrew gematria mapped to Latin alphabet
+# Aleph=1..Tet=9, Yod=10..Tsade=90, Qoph=100..Tav=400
+# Standard mapping for the 22 Hebrew letters extended to 26 Latin chars
+_HEBREW: dict[str, int] = {
+    "A": 1,
+    "B": 2,
+    "C": 3,
+    "D": 4,
+    "E": 5,
+    "F": 6,
+    "G": 7,
+    "H": 8,
+    "I": 9,
+    "J": 10,
+    "K": 20,
+    "L": 30,
+    "M": 40,
+    "N": 50,
+    "O": 60,
+    "P": 70,
+    "Q": 80,
+    "R": 90,
+    "S": 100,
+    "T": 200,
+    "U": 300,
+    "V": 400,
+    "W": 500,
+    "X": 600,
+    "Y": 700,
+    "Z": 800,
+}
+
+CIPHERS: dict[str, dict[str, int]] = {
+    "simple": _SIMPLE,
+    "reduction": _REDUCTION,
+    "reverse": _REVERSE,
+    "sumerian": _SUMERIAN,
+    "hebrew": _HEBREW,
+}
+
+# ── Notable Numbers ──────────────────────────────────────────────────────────
+
+NOTABLE_NUMBERS: dict[int, str] = {
+    1: "Unity, the Monad, beginning of all",
+    3: "Trinity, divine completeness, the Triad",
+    7: "Spiritual perfection, completion (7 days, 7 seals)",
+    9: "Finality, judgment, the last single digit",
+    11: "Master number — intuition, spiritual insight",
+    12: "Divine government (12 tribes, 12 apostles)",
+    13: "Rebellion and transformation, the 13th step",
+    22: "Master builder — turning dreams into reality",
+    26: "YHWH (Yod=10, He=5, Vav=6, He=5)",
+    33: "Master teacher — Christ consciousness, 33 vertebrae",
+    36: "The number of the righteous (Lamed-Vav Tzadikim)",
+    40: "Trial, testing, probation (40 days, 40 years)",
+    42: "The answer, and the number of generations to Christ",
+    72: "The Shemhamphorasch — 72 names of God",
+    88: "Mercury, infinite abundance, double infinity",
+    108: "Sacred in Hinduism and Buddhism (108 beads)",
+    111: "Angel number — new beginnings, alignment",
+    144: "12² — the elect, the sealed (144,000)",
+    153: "The miraculous catch of fish (John 21:11)",
+    222: "Alexander Whitestone. Balance, partnership, trust the process",
+    333: "Ascended masters present, divine protection",
+    369: "Tesla's key to the universe",
+    444: "Angels surrounding, foundation, stability",
+    555: "Major change coming, transformation",
+    616: "Earliest manuscript number of the Beast (P115)",
+    666: "Number of the Beast (Revelation 13:18), also carbon (6p 6n 6e)",
+    777: "Divine perfection tripled, jackpot of the spirit",
+    888: "Jesus in Greek isopsephy (Ιησους = 888)",
+    1776: "Year of independence, Bavarian Illuminati founding",
+}
+
+
+# ── Core Functions ───────────────────────────────────────────────────────────
+
+
+def _clean(text: str) -> str:
+    """Strip non-alpha, uppercase."""
+    return "".join(c for c in text.upper() if c.isalpha())
+
+
+def compute_value(text: str, cipher: str = "simple") -> int:
+    """Compute the gematria value of text in a given cipher.
+
+    Args:
+        text: Any string (non-alpha characters are ignored).
+        cipher: One of 'simple', 'reduction', 'reverse', 'sumerian', 'hebrew'.
+
+    Returns:
+        Integer gematria value.
+
+    Raises:
+        ValueError: If cipher name is not recognized.
+    """
+    table = CIPHERS.get(cipher)
+    if table is None:
+        raise ValueError(f"Unknown cipher: {cipher!r}. Use one of {list(CIPHERS)}")
+    return sum(table.get(c, 0) for c in _clean(text))
+
+
+def compute_all(text: str) -> dict[str, int]:
+    """Compute gematria value across all cipher systems.
+
+    Args:
+        text: Any string.
+
+    Returns:
+        Dict mapping cipher name to integer value.
+    """
+    return {name: compute_value(text, name) for name in CIPHERS}
+
+
+def letter_breakdown(text: str, cipher: str = "simple") -> list[tuple[str, int]]:
+    """Return per-letter values for a text in a given cipher.
+
+    Args:
+        text: Any string.
+        cipher: Cipher system name.
+
+    Returns:
+        List of (letter, value) tuples for each alpha character.
+    """
+    table = CIPHERS.get(cipher)
+    if table is None:
+        raise ValueError(f"Unknown cipher: {cipher!r}")
+    return [(c, table.get(c, 0)) for c in _clean(text)]
+
+
+def reduce_number(n: int) -> int:
+    """Numerological reduction — sum digits until single digit.
+
+    Master numbers (11, 22, 33) are preserved.
+
+    Args:
+        n: Any positive integer.
+
+    Returns:
+        Single-digit result (or master number 11/22/33).
+    """
+    n = abs(n)
+    while n > 9 and n not in (11, 22, 33):
+        n = sum(int(d) for d in str(n))
+    return n
+
+
+def factorize(n: int) -> list[int]:
+    """Prime factorization of n.
+
+    Args:
+        n: Positive integer.
+
+    Returns:
+        List of prime factors in ascending order (with repetition).
+    """
+    if n < 2:
+        return [n] if n > 0 else []
+    factors = []
+    d = 2
+    while d * d <= n:
+        while n % d == 0:
+            factors.append(d)
+            n //= d
+        d += 1
+    if n > 1:
+        factors.append(n)
+    return factors
+
+
+def analyze_number(n: int) -> dict:
+    """Deep analysis of a number — reduction, factors, significance.
+
+    Args:
+        n: Any positive integer.
+
+    Returns:
+        Dict with reduction, factors, properties, and any notable significance.
+    """
+    result: dict = {
+        "value": n,
+        "numerological_reduction": reduce_number(n),
+        "prime_factors": factorize(n),
+        "is_prime": len(factorize(n)) == 1 and n > 1,
+        "is_perfect_square": math.isqrt(n) ** 2 == n if n >= 0 else False,
+        "is_triangular": _is_triangular(n),
+        "digit_sum": sum(int(d) for d in str(abs(n))),
+    }
+
+    # Master numbers
+    if n in (11, 22, 33):
+        result["master_number"] = True
+
+    # Angel numbers (repeating digits)
+    s = str(n)
+    if len(s) >= 3 and len(set(s)) == 1:
+        result["angel_number"] = True
+
+    # Notable significance
+    if n in NOTABLE_NUMBERS:
+        result["significance"] = NOTABLE_NUMBERS[n]
+
+    return result
+
+
+def _is_triangular(n: int) -> bool:
+    """Check if n is a triangular number (1, 3, 6, 10, 15, ...)."""
+    if n < 0:
+        return False
+    # n = k(k+1)/2  →  k² + k - 2n = 0  →  k = (-1 + sqrt(1+8n))/2
+    discriminant = 1 + 8 * n
+    sqrt_d = math.isqrt(discriminant)
+    return sqrt_d * sqrt_d == discriminant and (sqrt_d - 1) % 2 == 0
+
+
+# ── Tool Function (registered with Timmy) ────────────────────────────────────
+
+
+def gematria(query: str) -> str:
+    """Compute gematria values, analyze numbers, and find correspondences.
+
+    This is the wizard's language — letters are numbers, numbers are letters.
+    Use this tool for ANY gematria calculation. Do not attempt mental arithmetic.
+
+    Input modes:
+      - A word or phrase → computes values across all cipher systems
+      - A bare integer → analyzes the number (factors, reduction, significance)
+      - "compare: X, Y, Z" → side-by-side gematria comparison
+
+    Examples:
+        gematria("Alexander Whitestone")
+        gematria("222")
+        gematria("compare: Timmy Time, Alexander Whitestone")
+
+    Args:
+        query: A word/phrase, a number, or a "compare:" instruction.
+
+    Returns:
+        Formatted gematria analysis as a string.
+    """
+    query = query.strip()
+
+    # Mode: compare
+    if query.lower().startswith("compare:"):
+        phrases = [p.strip() for p in query[8:].split(",") if p.strip()]
+        if len(phrases) < 2:
+            return "Compare requires at least two phrases separated by commas."
+        return _format_comparison(phrases)
+
+    # Mode: number analysis
+    if query.lstrip("-").isdigit():
+        n = int(query)
+        return _format_number_analysis(n)
+
+    # Mode: phrase gematria
+    if not _clean(query):
+        return "No alphabetic characters found in input."
+
+    return _format_phrase_analysis(query)
+
+
+def _format_phrase_analysis(text: str) -> str:
+    """Format full gematria analysis for a phrase."""
+    values = compute_all(text)
+    lines = [f'Gematria of "{text}":', ""]
+
+    # All cipher values
+    for cipher, val in values.items():
+        label = cipher.replace("_", " ").title()
+        lines.append(f"  {label:12s} = {val}")
+
+    # Letter breakdown (simple)
+    breakdown = letter_breakdown(text, "simple")
+    letters_str = " + ".join(f"{c}({v})" for c, v in breakdown)
+    lines.append(f"\n  Breakdown (Simple): {letters_str}")
+
+    # Numerological reduction of the simple value
+    simple_val = values["simple"]
+    reduced = reduce_number(simple_val)
+    lines.append(f"  Numerological root: {simple_val} → {reduced}")
+
+    # Check notable
+    for cipher, val in values.items():
+        if val in NOTABLE_NUMBERS:
+            label = cipher.replace("_", " ").title()
+            lines.append(f"\n  ★ {val} ({label}): {NOTABLE_NUMBERS[val]}")
+
+    return "\n".join(lines)
+
+
+def _format_number_analysis(n: int) -> str:
+    """Format deep number analysis."""
+    info = analyze_number(n)
+    lines = [f"Analysis of {n}:", ""]
+    lines.append(f"  Numerological reduction: {n} → {info['numerological_reduction']}")
+    lines.append(f"  Prime factors: {' × '.join(str(f) for f in info['prime_factors']) or 'N/A'}")
+    lines.append(f"  Is prime: {info['is_prime']}")
+    lines.append(f"  Is perfect square: {info['is_perfect_square']}")
+    lines.append(f"  Is triangular: {info['is_triangular']}")
+    lines.append(f"  Digit sum: {info['digit_sum']}")
+
+    if info.get("master_number"):
+        lines.append("  ★ Master Number")
+    if info.get("angel_number"):
+        lines.append("  ★ Angel Number (repeating digits)")
+    if info.get("significance"):
+        lines.append(f"\n  Significance: {info['significance']}")
+
+    return "\n".join(lines)
+
+
+def _format_comparison(phrases: list[str]) -> str:
+    """Format side-by-side gematria comparison."""
+    lines = ["Gematria Comparison:", ""]
+
+    # Header
+    max_name = max(len(p) for p in phrases)
+    header = f"  {'Phrase':<{max_name}s}  Simple  Reduct  Reverse  Sumerian  Hebrew"
+    lines.append(header)
+    lines.append("  " + "─" * (len(header) - 2))
+
+    all_values = {}
+    for phrase in phrases:
+        vals = compute_all(phrase)
+        all_values[phrase] = vals
+        lines.append(
+            f"  {phrase:<{max_name}s}  {vals['simple']:>6d}  {vals['reduction']:>6d}"
+            f"  {vals['reverse']:>7d}  {vals['sumerian']:>8d}  {vals['hebrew']:>6d}"
+        )
+
+    # Find matches (shared values across any cipher)
+    matches = []
+    for cipher in CIPHERS:
+        vals_by_cipher = {p: all_values[p][cipher] for p in phrases}
+        unique_vals = set(vals_by_cipher.values())
+        if len(unique_vals) < len(phrases):
+            # At least two phrases share a value
+            for v in unique_vals:
+                sharing = [p for p, pv in vals_by_cipher.items() if pv == v]
+                if len(sharing) > 1:
+                    label = cipher.title()
+                    matches.append(f"  ★ {label} = {v}: " + ", ".join(sharing))
+
+    if matches:
+        lines.append("\nCorrespondences found:")
+        lines.extend(matches)
+
+    return "\n".join(lines)

src/timmy/interview.py

@@ -86,7 +86,7 @@ def run_interview(
 
         try:
             answer = chat_fn(question)
-        except Exception as exc:
+        except Exception as exc:  # broad catch intentional: chat_fn can raise any error
             logger.error("Interview question failed: %s", exc)
             answer = f"(Error: {exc})"
 

src/timmy/loop_qa.py

@@ -262,7 +262,8 @@ def capture_error(exc, **kwargs):
         from infrastructure.error_capture import capture_error as _capture
 
         return _capture(exc, **kwargs)
-    except Exception:
+    except Exception as capture_exc:
+        logger.debug("Failed to capture error: %s", capture_exc)
         logger.debug("Failed to capture error", exc_info=True)
 
 

src/timmy/mcp_tools.py

@@ -25,9 +25,12 @@ import os
 import shutil
 import sqlite3
 import uuid
+from contextlib import closing
 from datetime import datetime
 from pathlib import Path
 
+import httpx
+
 from config import settings
 
 logger = logging.getLogger(__name__)
@@ -40,7 +43,7 @@ def _parse_command(command_str: str) -> tuple[str, list[str]]:
     """Split a command string into (executable, args).
 
     Handles ``~/`` expansion and resolves via PATH if needed.
-    E.g. ``"gitea-mcp -t stdio"`` → ``("/Users/x/go/bin/gitea-mcp", ["-t", "stdio"])``
+    E.g. ``"gitea-mcp-server -t stdio"`` → ``("/opt/homebrew/bin/gitea-mcp-server", ["-t", "stdio"])``
     """
     parts = command_str.split()
     executable = os.path.expanduser(parts[0])
@@ -163,37 +166,36 @@ def _bridge_to_work_order(title: str, body: str, category: str) -> None:
     try:
         db_path = Path(settings.repo_root) / "data" / "work_orders.db"
         db_path.parent.mkdir(parents=True, exist_ok=True)
-        conn = sqlite3.connect(str(db_path))
-        conn.execute(
-            """CREATE TABLE IF NOT EXISTS work_orders (
-                id TEXT PRIMARY KEY,
-                title TEXT NOT NULL,
-                description TEXT DEFAULT '',
-                priority TEXT DEFAULT 'medium',
-                category TEXT DEFAULT 'suggestion',
-                submitter TEXT DEFAULT 'dashboard',
-                related_files TEXT DEFAULT '',
-                status TEXT DEFAULT 'submitted',
-                result TEXT DEFAULT '',
-                rejection_reason TEXT DEFAULT '',
-                created_at TEXT DEFAULT (datetime('now')),
-                completed_at TEXT
-            )"""
-        )
-        conn.execute(
-            "INSERT INTO work_orders (id, title, description, category, submitter, created_at) "
-            "VALUES (?, ?, ?, ?, ?, ?)",
-            (
-                str(uuid.uuid4()),
-                title,
-                body,
-                category,
-                "timmy-thinking",
-                datetime.utcnow().isoformat(),
-            ),
-        )
-        conn.commit()
-        conn.close()
+        with closing(sqlite3.connect(str(db_path))) as conn:
+            conn.execute(
+                """CREATE TABLE IF NOT EXISTS work_orders (
+                    id TEXT PRIMARY KEY,
+                    title TEXT NOT NULL,
+                    description TEXT DEFAULT '',
+                    priority TEXT DEFAULT 'medium',
+                    category TEXT DEFAULT 'suggestion',
+                    submitter TEXT DEFAULT 'dashboard',
+                    related_files TEXT DEFAULT '',
+                    status TEXT DEFAULT 'submitted',
+                    result TEXT DEFAULT '',
+                    rejection_reason TEXT DEFAULT '',
+                    created_at TEXT DEFAULT (datetime('now')),
+                    completed_at TEXT
+                )"""
+            )
+            conn.execute(
+                "INSERT INTO work_orders (id, title, description, category, submitter, created_at) "
+                "VALUES (?, ?, ?, ?, ?, ?)",
+                (
+                    str(uuid.uuid4()),
+                    title,
+                    body,
+                    category,
+                    "timmy-thinking",
+                    datetime.utcnow().isoformat(),
+                ),
+            )
+            conn.commit()
     except Exception as exc:
         logger.debug("Work order bridge failed: %s", exc)
 
@@ -268,6 +270,140 @@ async def create_gitea_issue_via_mcp(title: str, body: str = "", labels: str = "
         return f"Failed to create issue via MCP: {exc}"
 
 
+def _generate_avatar_image() -> bytes:
+    """Generate a Timmy-themed avatar image using Pillow.
+
+    Creates a 512x512 wizard-themed avatar with emerald/purple/gold palette.
+    Returns raw PNG bytes. Falls back to a minimal solid-color image if
+    Pillow drawing primitives fail.
+    """
+    from PIL import Image, ImageDraw
+
+    size = 512
+    img = Image.new("RGB", (size, size), (15, 25, 20))
+    draw = ImageDraw.Draw(img)
+
+    # Background gradient effect — concentric circles
+    for i in range(size // 2, 0, -4):
+        g = int(25 + (i / (size // 2)) * 30)
+        draw.ellipse(
+            [size // 2 - i, size // 2 - i, size // 2 + i, size // 2 + i],
+            fill=(10, g, 20),
+        )
+
+    # Wizard hat (triangle)
+    hat_color = (100, 50, 160)  # purple
+    draw.polygon(
+        [(256, 40), (160, 220), (352, 220)],
+        fill=hat_color,
+        outline=(180, 130, 255),
+    )
+
+    # Hat brim
+    draw.ellipse([140, 200, 372, 250], fill=hat_color, outline=(180, 130, 255))
+
+    # Face circle
+    draw.ellipse([190, 220, 322, 370], fill=(60, 180, 100), outline=(80, 220, 120))
+
+    # Eyes
+    draw.ellipse([220, 275, 248, 310], fill=(255, 255, 255))
+    draw.ellipse([264, 275, 292, 310], fill=(255, 255, 255))
+    draw.ellipse([228, 285, 242, 300], fill=(30, 30, 60))
+    draw.ellipse([272, 285, 286, 300], fill=(30, 30, 60))
+
+    # Smile
+    draw.arc([225, 300, 287, 355], start=10, end=170, fill=(30, 30, 60), width=3)
+
+    # Stars around the hat
+    gold = (220, 190, 50)
+    star_positions = [(120, 100), (380, 120), (100, 300), (400, 280), (256, 10)]
+    for sx, sy in star_positions:
+        r = 8
+        draw.polygon(
+            [
+                (sx, sy - r),
+                (sx + r // 3, sy - r // 3),
+                (sx + r, sy),
+                (sx + r // 3, sy + r // 3),
+                (sx, sy + r),
+                (sx - r // 3, sy + r // 3),
+                (sx - r, sy),
+                (sx - r // 3, sy - r // 3),
+            ],
+            fill=gold,
+        )
+
+    # "T" monogram on the hat
+    draw.text((243, 100), "T", fill=gold)
+
+    # Robe / body
+    draw.polygon(
+        [(180, 370), (140, 500), (372, 500), (332, 370)],
+        fill=(40, 100, 70),
+        outline=(60, 160, 100),
+    )
+
+    import io
+
+    buf = io.BytesIO()
+    img.save(buf, format="PNG")
+    return buf.getvalue()
+
+
+async def update_gitea_avatar() -> str:
+    """Generate and upload a unique avatar to Timmy's Gitea profile.
+
+    Creates a wizard-themed avatar image using Pillow drawing primitives,
+    base64-encodes it, and POSTs to the Gitea user avatar API endpoint.
+
+    Returns:
+        Success or failure message string.
+    """
+    if not settings.gitea_enabled or not settings.gitea_token:
+        return "Gitea integration is not configured (no token or disabled)."
+
+    try:
+        from PIL import Image  # noqa: F401 — availability check
+    except ImportError:
+        return "Pillow is not installed — cannot generate avatar image."
+
+    try:
+        import base64
+
+        # Step 1: Generate the avatar image
+        png_bytes = _generate_avatar_image()
+        logger.info("Generated avatar image (%d bytes)", len(png_bytes))
+
+        # Step 2: Base64-encode (raw, no data URI prefix)
+        b64_image = base64.b64encode(png_bytes).decode("ascii")
+
+        # Step 3: POST to Gitea
+        async with httpx.AsyncClient(timeout=15) as client:
+            resp = await client.post(
+                f"{settings.gitea_url}/api/v1/user/avatar",
+                headers={
+                    "Authorization": f"token {settings.gitea_token}",
+                    "Content-Type": "application/json",
+                },
+                json={"image": b64_image},
+            )
+
+        # Gitea returns empty body on success (204 or 200)
+        if resp.status_code in (200, 204):
+            logger.info("Gitea avatar updated successfully")
+            return "Avatar updated successfully on Gitea."
+
+        logger.warning("Gitea avatar update failed: %s %s", resp.status_code, resp.text[:200])
+        return f"Gitea avatar update failed (HTTP {resp.status_code}): {resp.text[:200]}"
+
+    except (httpx.ConnectError, httpx.ReadError, ConnectionError) as exc:
+        logger.warning("Gitea connection failed during avatar update: %s", exc)
+        return f"Could not connect to Gitea: {exc}"
+    except Exception as exc:
+        logger.error("Avatar update failed: %s", exc)
+        return f"Avatar update failed: {exc}"
+
+
 async def close_mcp_sessions() -> None:
     """Close any open MCP sessions. Called during app shutdown."""
     global _issue_session

src/timmy/memory/__init__.py

@@ -1 +1,7 @@
-"""Memory — Persistent conversation and knowledge memory."""
+"""Memory — Persistent conversation and knowledge memory.
+
+Sub-modules:
+    embeddings  — text-to-vector embedding + similarity functions
+    unified     — unified memory schema and connection management
+    vector_store — backward compatibility re-exports from memory_system
+"""

src/timmy/memory/embeddings.py

@@ -0,0 +1,88 @@
+"""Embedding functions for Timmy's memory system.
+
+Provides text-to-vector embedding using sentence-transformers (preferred)
+with a deterministic hash-based fallback when the ML library is unavailable.
+
+Also includes vector similarity utilities (cosine similarity, keyword overlap).
+"""
+
+import hashlib
+import logging
+import math
+
+logger = logging.getLogger(__name__)
+
+# Embedding model - small, fast, local
+EMBEDDING_MODEL = None
+EMBEDDING_DIM = 384  # MiniLM dimension
+
+
+def _get_embedding_model():
+    """Lazy-load embedding model."""
+    global EMBEDDING_MODEL
+    if EMBEDDING_MODEL is None:
+        try:
+            from config import settings
+
+            if settings.timmy_skip_embeddings:
+                EMBEDDING_MODEL = False
+                return EMBEDDING_MODEL
+        except ImportError:
+            pass
+
+        try:
+            from sentence_transformers import SentenceTransformer
+
+            EMBEDDING_MODEL = SentenceTransformer("all-MiniLM-L6-v2")
+            logger.info("MemorySystem: Loaded embedding model")
+        except ImportError:
+            logger.warning("MemorySystem: sentence-transformers not installed, using fallback")
+            EMBEDDING_MODEL = False  # Use fallback
+    return EMBEDDING_MODEL
+
+
+def _simple_hash_embedding(text: str) -> list[float]:
+    """Fallback: Simple hash-based embedding when transformers unavailable."""
+    words = text.lower().split()
+    vec = [0.0] * 128
+    for i, word in enumerate(words[:50]):  # First 50 words
+        h = hashlib.md5(word.encode()).hexdigest()
+        for j in range(8):
+            idx = (i * 8 + j) % 128
+            vec[idx] += int(h[j * 2 : j * 2 + 2], 16) / 255.0
+    # Normalize
+    mag = math.sqrt(sum(x * x for x in vec)) or 1.0
+    return [x / mag for x in vec]
+
+
+def embed_text(text: str) -> list[float]:
+    """Generate embedding for text."""
+    model = _get_embedding_model()
+    if model and model is not False:
+        embedding = model.encode(text)
+        return embedding.tolist()
+    return _simple_hash_embedding(text)
+
+
+def cosine_similarity(a: list[float], b: list[float]) -> float:
+    """Calculate cosine similarity between two vectors."""
+    dot = sum(x * y for x, y in zip(a, b, strict=False))
+    mag_a = math.sqrt(sum(x * x for x in a))
+    mag_b = math.sqrt(sum(x * x for x in b))
+    if mag_a == 0 or mag_b == 0:
+        return 0.0
+    return dot / (mag_a * mag_b)
+
+
+# Alias for backward compatibility
+_cosine_similarity = cosine_similarity
+
+
+def _keyword_overlap(query: str, content: str) -> float:
+    """Simple keyword overlap score as fallback."""
+    query_words = set(query.lower().split())
+    content_words = set(content.lower().split())
+    if not query_words:
+        return 0.0
+    overlap = len(query_words & content_words)
+    return overlap / len(query_words)

src/timmy/memory/unified.py

@@ -1,85 +1,201 @@
-"""Unified memory database — single SQLite DB for all memory types.
+"""Unified memory schema and connection management.
 
-Consolidates three previously separate stores into one:
-- **facts**: Long-term knowledge (user preferences, learned patterns)
-- **chunks**: Indexed vault documents (markdown files from memory/)
-- **episodes**: Runtime memories (conversations, agent observations)
-
-All three tables live in ``data/memory.db``.  Existing APIs in
-``vector_store.py`` and ``semantic_memory.py`` are updated to point here.
+This module provides the central database schema for Timmy's consolidated
+memory system. All memory types (facts, conversations, documents, vault chunks)
+are stored in a single `memories` table with a `memory_type` discriminator.
 """
 
 import logging
 import sqlite3
+import uuid
+from collections.abc import Generator
+from contextlib import closing, contextmanager
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
 from pathlib import Path
 
 logger = logging.getLogger(__name__)
 
-DB_PATH = Path(__file__).parent.parent.parent.parent / "data" / "memory.db"
+# Paths
+PROJECT_ROOT = Path(__file__).parent.parent.parent.parent
+DB_PATH = PROJECT_ROOT / "data" / "memory.db"
 
 
-def get_connection() -> sqlite3.Connection:
-    """Open (and lazily create) the unified memory database."""
+@contextmanager
+def get_connection() -> Generator[sqlite3.Connection, None, None]:
+    """Get database connection to unified memory database."""
     DB_PATH.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(str(DB_PATH))
-    conn.row_factory = sqlite3.Row
-    conn.execute("PRAGMA journal_mode=WAL")
-    conn.execute("PRAGMA busy_timeout=5000")
-    _ensure_schema(conn)
-    return conn
+    with closing(sqlite3.connect(str(DB_PATH))) as conn:
+        conn.row_factory = sqlite3.Row
+        conn.execute("PRAGMA journal_mode=WAL")
+        conn.execute("PRAGMA busy_timeout=5000")
+        _ensure_schema(conn)
+        yield conn
 
 
 def _ensure_schema(conn: sqlite3.Connection) -> None:
-    """Create the three core tables and indexes if they don't exist."""
-
-    # --- facts ---------------------------------------------------------------
+    """Create the unified memories table and indexes if they don't exist."""
     conn.execute("""
-        CREATE TABLE IF NOT EXISTS facts (
+        CREATE TABLE IF NOT EXISTS memories (
             id TEXT PRIMARY KEY,
-            category TEXT NOT NULL DEFAULT 'general',
             content TEXT NOT NULL,
-            confidence REAL NOT NULL DEFAULT 0.8,
+            memory_type TEXT NOT NULL DEFAULT 'fact',
             source TEXT NOT NULL DEFAULT 'agent',
+            embedding TEXT,
+            metadata TEXT,
+            source_hash TEXT,
+            agent_id TEXT,
+            task_id TEXT,
+            session_id TEXT,
+            confidence REAL NOT NULL DEFAULT 0.8,
             tags TEXT NOT NULL DEFAULT '[]',
             created_at TEXT NOT NULL,
             last_accessed TEXT,
             access_count INTEGER NOT NULL DEFAULT 0
         )
     """)
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_category ON facts(category)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_confidence ON facts(confidence)")
 
-    # --- chunks (vault document fragments) -----------------------------------
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS chunks (
-            id TEXT PRIMARY KEY,
-            source TEXT NOT NULL,
-            content TEXT NOT NULL,
-            embedding TEXT NOT NULL,
-            created_at TEXT NOT NULL,
-            source_hash TEXT NOT NULL
-        )
-    """)
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_chunks_source ON chunks(source)")
+    # Create indexes for efficient querying
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_memories_type ON memories(memory_type)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_memories_time ON memories(created_at)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_memories_session ON memories(session_id)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_memories_agent ON memories(agent_id)")
+    conn.execute("CREATE INDEX IF NOT EXISTS idx_memories_source ON memories(source)")
+    conn.commit()
 
-    # --- episodes (runtime memory entries) -----------------------------------
-    conn.execute("""
-        CREATE TABLE IF NOT EXISTS episodes (
-            id TEXT PRIMARY KEY,
-            content TEXT NOT NULL,
-            source TEXT NOT NULL,
-            context_type TEXT NOT NULL DEFAULT 'conversation',
-            embedding TEXT,
-            metadata TEXT,
-            agent_id TEXT,
-            task_id TEXT,
-            session_id TEXT,
-            timestamp TEXT NOT NULL
-        )
-    """)
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_episodes_type ON episodes(context_type)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_episodes_time ON episodes(timestamp)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_episodes_session ON episodes(session_id)")
-    conn.execute("CREATE INDEX IF NOT EXISTS idx_episodes_agent ON episodes(agent_id)")
+    # Run migration if needed
+    _migrate_schema(conn)
+
+
+def _migrate_schema(conn: sqlite3.Connection) -> None:
+    """Migrate from old three-table schema to unified memories table.
+
+    Migration paths:
+    - episodes table -> memories (context_type -> memory_type)
+    - chunks table -> memories with memory_type='vault_chunk'
+    - facts table -> dropped (unused, 0 rows expected)
+    """
+    cursor = conn.execute("SELECT name FROM sqlite_master WHERE type='table'")
+    tables = {row[0] for row in cursor.fetchall()}
+
+    has_memories = "memories" in tables
+    has_episodes = "episodes" in tables
+    has_chunks = "chunks" in tables
+    has_facts = "facts" in tables
+
+    # Check if we need to migrate (old schema exists but new one doesn't fully)
+    if not has_memories:
+        logger.info("Migration: Creating unified memories table")
+        # Schema will be created above
+
+    # Migrate episodes -> memories
+    if has_episodes and has_memories:
+        logger.info("Migration: Converting episodes table to memories")
+        try:
+            cols = _get_table_columns(conn, "episodes")
+            context_type_col = "context_type" if "context_type" in cols else "'conversation'"
+
+            conn.execute(f"""
+                INSERT INTO memories (
+                    id, content, memory_type, source, embedding,
+                    metadata, agent_id, task_id, session_id,
+                    created_at, access_count, last_accessed
+                )
+                SELECT 
+                    id, content, 
+                    COALESCE({context_type_col}, 'conversation'),
+                    COALESCE(source, 'agent'),
+                    embedding,
+                    metadata, agent_id, task_id, session_id,
+                    COALESCE(timestamp, datetime('now')), 0, NULL
+                FROM episodes
+            """)
+            conn.execute("DROP TABLE episodes")
+            logger.info("Migration: Migrated episodes to memories")
+        except sqlite3.Error as exc:
+            logger.warning("Migration: Failed to migrate episodes: %s", exc)
+
+    # Migrate chunks -> memories as vault_chunk
+    if has_chunks and has_memories:
+        logger.info("Migration: Converting chunks table to memories")
+        try:
+            cols = _get_table_columns(conn, "chunks")
+
+            id_col = "id" if "id" in cols else "CAST(rowid AS TEXT)"
+            content_col = "content" if "content" in cols else "text"
+            source_col = (
+                "filepath" if "filepath" in cols else ("source" if "source" in cols else "'vault'")
+            )
+            embedding_col = "embedding" if "embedding" in cols else "NULL"
+            created_col = "created_at" if "created_at" in cols else "datetime('now')"
+
+            conn.execute(f"""
+                INSERT INTO memories (
+                    id, content, memory_type, source, embedding,
+                    created_at, access_count
+                )
+                SELECT 
+                    {id_col}, {content_col}, 'vault_chunk', {source_col},
+                    {embedding_col}, {created_col}, 0
+                FROM chunks
+            """)
+            conn.execute("DROP TABLE chunks")
+            logger.info("Migration: Migrated chunks to memories")
+        except sqlite3.Error as exc:
+            logger.warning("Migration: Failed to migrate chunks: %s", exc)
+
+    # Drop old facts table
+    if has_facts:
+        try:
+            conn.execute("DROP TABLE facts")
+            logger.info("Migration: Dropped old facts table")
+        except sqlite3.Error as exc:
+            logger.warning("Migration: Failed to drop facts: %s", exc)
 
     conn.commit()
+
+
+def _get_table_columns(conn: sqlite3.Connection, table_name: str) -> set[str]:
+    """Get the column names for a table."""
+    cursor = conn.execute(f"PRAGMA table_info({table_name})")
+    return {row[1] for row in cursor.fetchall()}
+
+
+# Backward compatibility aliases
+get_conn = get_connection
+
+
+@dataclass
+class MemoryEntry:
+    """A memory entry with vector embedding.
+
+    Note: The DB column is `memory_type` but this field is named `context_type`
+    for backward API compatibility.
+    """
+
+    id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    content: str = ""  # The actual text content
+    source: str = ""  # Where it came from (agent, user, system)
+    context_type: str = "conversation"  # API field name; DB column is memory_type
+    agent_id: str | None = None
+    task_id: str | None = None
+    session_id: str | None = None
+    metadata: dict | None = None
+    embedding: list[float] | None = None
+    timestamp: str = field(default_factory=lambda: datetime.now(UTC).isoformat())
+    relevance_score: float | None = None  # Set during search
+
+
+@dataclass
+class MemoryChunk:
+    """A searchable chunk of memory."""
+
+    id: str
+    source: str  # filepath
+    content: str
+    embedding: list[float]
+    created_at: str
+
+
+# Note: Functions are available via memory_system module directly
+# from timmy.memory_system import store_memory, search_memories, etc.

src/timmy/memory/vector_store.py

@@ -1,430 +1,37 @@
-"""Vector store for semantic memory using sqlite-vss.
-
-Provides embedding-based similarity search for the Echo agent
-to retrieve relevant context from conversation history.
-"""
-
-import json
-import sqlite3
-import uuid
-from dataclasses import dataclass, field
-from datetime import UTC, datetime
-
-
-def _check_embedding_model() -> bool | None:
-    """Check if the canonical embedding model is available."""
-    try:
-        from timmy.semantic_memory import _get_embedding_model
-
-        model = _get_embedding_model()
-        return model is not None and model is not False
-    except Exception:
-        return None
-
-
-def _compute_embedding(text: str) -> list[float]:
-    """Compute embedding vector for text.
-
-    Delegates to the canonical embedding provider in semantic_memory
-    to avoid loading the model multiple times.
-    """
-    from timmy.semantic_memory import embed_text
-
-    return embed_text(text)
-
-
-@dataclass
-class MemoryEntry:
-    """A memory entry with vector embedding."""
-
-    id: str = field(default_factory=lambda: str(uuid.uuid4()))
-    content: str = ""  # The actual text content
-    source: str = ""  # Where it came from (agent, user, system)
-    context_type: str = "conversation"  # conversation, document, fact, etc.
-    agent_id: str | None = None
-    task_id: str | None = None
-    session_id: str | None = None
-    metadata: dict | None = None
-    embedding: list[float] | None = None
-    timestamp: str = field(default_factory=lambda: datetime.now(UTC).isoformat())
-    relevance_score: float | None = None  # Set during search
-
-
-def _get_conn() -> sqlite3.Connection:
-    """Get database connection to unified memory.db."""
-    from timmy.memory.unified import get_connection
-
-    return get_connection()
-
-
-def store_memory(
-    content: str,
-    source: str,
-    context_type: str = "conversation",
-    agent_id: str | None = None,
-    task_id: str | None = None,
-    session_id: str | None = None,
-    metadata: dict | None = None,
-    compute_embedding: bool = True,
-) -> MemoryEntry:
-    """Store a memory entry with optional embedding.
-
-    Args:
-        content: The text content to store
-        source: Source of the memory (agent name, user, system)
-        context_type: Type of context (conversation, document, fact)
-        agent_id: Associated agent ID
-        task_id: Associated task ID
-        session_id: Session identifier
-        metadata: Additional structured data
-        compute_embedding: Whether to compute vector embedding
-
-    Returns:
-        The stored MemoryEntry
-    """
-    embedding = None
-    if compute_embedding:
-        embedding = _compute_embedding(content)
-
-    entry = MemoryEntry(
-        content=content,
-        source=source,
-        context_type=context_type,
-        agent_id=agent_id,
-        task_id=task_id,
-        session_id=session_id,
-        metadata=metadata,
-        embedding=embedding,
-    )
-
-    conn = _get_conn()
-    conn.execute(
-        """
-        INSERT INTO episodes
-        (id, content, source, context_type, agent_id, task_id, session_id,
-         metadata, embedding, timestamp)
-        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
-        """,
-        (
-            entry.id,
-            entry.content,
-            entry.source,
-            entry.context_type,
-            entry.agent_id,
-            entry.task_id,
-            entry.session_id,
-            json.dumps(metadata) if metadata else None,
-            json.dumps(embedding) if embedding else None,
-            entry.timestamp,
-        ),
-    )
-    conn.commit()
-    conn.close()
-
-    return entry
-
-
-def search_memories(
-    query: str,
-    limit: int = 10,
-    context_type: str | None = None,
-    agent_id: str | None = None,
-    session_id: str | None = None,
-    min_relevance: float = 0.0,
-) -> list[MemoryEntry]:
-    """Search for memories by semantic similarity.
-
-    Args:
-        query: Search query text
-        limit: Maximum results
-        context_type: Filter by context type
-        agent_id: Filter by agent
-        session_id: Filter by session
-        min_relevance: Minimum similarity score (0-1)
-
-    Returns:
-        List of MemoryEntry objects sorted by relevance
-    """
-    query_embedding = _compute_embedding(query)
-
-    conn = _get_conn()
-
-    # Build query with filters
-    conditions = []
-    params = []
-
-    if context_type:
-        conditions.append("context_type = ?")
-        params.append(context_type)
-    if agent_id:
-        conditions.append("agent_id = ?")
-        params.append(agent_id)
-    if session_id:
-        conditions.append("session_id = ?")
-        params.append(session_id)
-
-    where_clause = "WHERE " + " AND ".join(conditions) if conditions else ""
-
-    # Fetch candidates (we'll do in-memory similarity for now)
-    # For production with sqlite-vss, this would use vector similarity index
-    query_sql = f"""
-        SELECT * FROM episodes
-        {where_clause}
-        ORDER BY timestamp DESC
-        LIMIT ?
-    """
-    params.append(limit * 3)  # Get more candidates for ranking
-
-    rows = conn.execute(query_sql, params).fetchall()
-    conn.close()
-
-    # Compute similarity scores
-    results = []
-    for row in rows:
-        entry = MemoryEntry(
-            id=row["id"],
-            content=row["content"],
-            source=row["source"],
-            context_type=row["context_type"],
-            agent_id=row["agent_id"],
-            task_id=row["task_id"],
-            session_id=row["session_id"],
-            metadata=json.loads(row["metadata"]) if row["metadata"] else None,
-            embedding=json.loads(row["embedding"]) if row["embedding"] else None,
-            timestamp=row["timestamp"],
-        )
-
-        if entry.embedding:
-            # Cosine similarity
-            score = _cosine_similarity(query_embedding, entry.embedding)
-            entry.relevance_score = score
-            if score >= min_relevance:
-                results.append(entry)
-        else:
-            # Fallback: check for keyword overlap
-            score = _keyword_overlap(query, entry.content)
-            entry.relevance_score = score
-            if score >= min_relevance:
-                results.append(entry)
-
-    # Sort by relevance and return top results
-    results.sort(key=lambda x: x.relevance_score or 0, reverse=True)
-    return results[:limit]
-
-
-def _cosine_similarity(a: list[float], b: list[float]) -> float:
-    """Compute cosine similarity between two vectors."""
-    dot = sum(x * y for x, y in zip(a, b, strict=False))
-    norm_a = sum(x * x for x in a) ** 0.5
-    norm_b = sum(x * x for x in b) ** 0.5
-    if norm_a == 0 or norm_b == 0:
-        return 0.0
-    return dot / (norm_a * norm_b)
-
-
-def _keyword_overlap(query: str, content: str) -> float:
-    """Simple keyword overlap score as fallback."""
-    query_words = set(query.lower().split())
-    content_words = set(content.lower().split())
-    if not query_words:
-        return 0.0
-    overlap = len(query_words & content_words)
-    return overlap / len(query_words)
-
-
-def get_memory_context(query: str, max_tokens: int = 2000, **filters) -> str:
-    """Get relevant memory context as formatted text for LLM prompts.
-
-    Args:
-        query: Search query
-        max_tokens: Approximate maximum tokens to return
-        **filters: Additional filters (agent_id, session_id, etc.)
-
-    Returns:
-        Formatted context string for inclusion in prompts
-    """
-    memories = search_memories(query, limit=20, **filters)
-
-    context_parts = []
-    total_chars = 0
-    max_chars = max_tokens * 4  # Rough approximation
-
-    for mem in memories:
-        formatted = f"[{mem.source}]: {mem.content}"
-        if total_chars + len(formatted) > max_chars:
-            break
-        context_parts.append(formatted)
-        total_chars += len(formatted)
-
-    if not context_parts:
-        return ""
-
-    return "Relevant context from memory:\n" + "\n\n".join(context_parts)
-
-
-def recall_personal_facts(agent_id: str | None = None) -> list[str]:
-    """Recall personal facts about the user or system.
-
-    Args:
-        agent_id: Optional agent filter
-
-    Returns:
-        List of fact strings
-    """
-    conn = _get_conn()
-
-    if agent_id:
-        rows = conn.execute(
-            """
-            SELECT content FROM episodes
-            WHERE context_type = 'fact' AND agent_id = ?
-            ORDER BY timestamp DESC
-            LIMIT 100
-            """,
-            (agent_id,),
-        ).fetchall()
-    else:
-        rows = conn.execute(
-            """
-            SELECT content FROM episodes
-            WHERE context_type = 'fact'
-            ORDER BY timestamp DESC
-            LIMIT 100
-            """,
-        ).fetchall()
-
-    conn.close()
-    return [r["content"] for r in rows]
-
-
-def recall_personal_facts_with_ids(agent_id: str | None = None) -> list[dict]:
-    """Recall personal facts with their IDs for edit/delete operations."""
-    conn = _get_conn()
-    if agent_id:
-        rows = conn.execute(
-            "SELECT id, content FROM episodes WHERE context_type = 'fact' AND agent_id = ? ORDER BY timestamp DESC LIMIT 100",
-            (agent_id,),
-        ).fetchall()
-    else:
-        rows = conn.execute(
-            "SELECT id, content FROM episodes WHERE context_type = 'fact' ORDER BY timestamp DESC LIMIT 100",
-        ).fetchall()
-    conn.close()
-    return [{"id": r["id"], "content": r["content"]} for r in rows]
-
-
-def update_personal_fact(memory_id: str, new_content: str) -> bool:
-    """Update a personal fact's content."""
-    conn = _get_conn()
-    cursor = conn.execute(
-        "UPDATE episodes SET content = ? WHERE id = ? AND context_type = 'fact'",
-        (new_content, memory_id),
-    )
-    conn.commit()
-    updated = cursor.rowcount > 0
-    conn.close()
-    return updated
-
-
-def store_personal_fact(fact: str, agent_id: str | None = None) -> MemoryEntry:
-    """Store a personal fact about the user or system.
-
-    Args:
-        fact: The fact to store
-        agent_id: Associated agent
-
-    Returns:
-        The stored MemoryEntry
-    """
-    return store_memory(
-        content=fact,
-        source="system",
-        context_type="fact",
-        agent_id=agent_id,
-        metadata={"auto_extracted": False},
-    )
-
-
-def delete_memory(memory_id: str) -> bool:
-    """Delete a memory entry by ID.
-
-    Returns:
-        True if deleted, False if not found
-    """
-    conn = _get_conn()
-    cursor = conn.execute(
-        "DELETE FROM episodes WHERE id = ?",
-        (memory_id,),
-    )
-    conn.commit()
-    deleted = cursor.rowcount > 0
-    conn.close()
-    return deleted
-
-
-def get_memory_stats() -> dict:
-    """Get statistics about the memory store.
-
-    Returns:
-        Dict with counts by type, total entries, etc.
-    """
-    conn = _get_conn()
-
-    total = conn.execute("SELECT COUNT(*) as count FROM episodes").fetchone()["count"]
-
-    by_type = {}
-    rows = conn.execute(
-        "SELECT context_type, COUNT(*) as count FROM episodes GROUP BY context_type"
-    ).fetchall()
-    for row in rows:
-        by_type[row["context_type"]] = row["count"]
-
-    with_embeddings = conn.execute(
-        "SELECT COUNT(*) as count FROM episodes WHERE embedding IS NOT NULL"
-    ).fetchone()["count"]
-
-    conn.close()
-
-    return {
-        "total_entries": total,
-        "by_type": by_type,
-        "with_embeddings": with_embeddings,
-        "has_embedding_model": _check_embedding_model(),
-    }
-
-
-def prune_memories(older_than_days: int = 90, keep_facts: bool = True) -> int:
-    """Delete old memories to manage storage.
-
-    Args:
-        older_than_days: Delete memories older than this
-        keep_facts: Whether to preserve fact-type memories
-
-    Returns:
-        Number of entries deleted
-    """
-    from datetime import timedelta
-
-    cutoff = (datetime.now(UTC) - timedelta(days=older_than_days)).isoformat()
-
-    conn = _get_conn()
-
-    if keep_facts:
-        cursor = conn.execute(
-            """
-            DELETE FROM episodes
-            WHERE timestamp < ? AND context_type != 'fact'
-            """,
-            (cutoff,),
-        )
-    else:
-        cursor = conn.execute(
-            "DELETE FROM episodes WHERE timestamp < ?",
-            (cutoff,),
-        )
-
-    deleted = cursor.rowcount
-    conn.commit()
-    conn.close()
-
-    return deleted
+"""Backward compatibility — all memory functions live in memory_system now."""
+
+from timmy.memory_system import (
+    DB_PATH,
+    MemoryEntry,
+    _cosine_similarity,
+    _keyword_overlap,
+    delete_memory,
+    get_memory_context,
+    get_memory_stats,
+    get_memory_system,
+    prune_memories,
+    recall_personal_facts,
+    recall_personal_facts_with_ids,
+    search_memories,
+    store_memory,
+    store_personal_fact,
+    update_personal_fact,
+)
+
+__all__ = [
+    "DB_PATH",
+    "MemoryEntry",
+    "delete_memory",
+    "get_memory_context",
+    "get_memory_stats",
+    "get_memory_system",
+    "prune_memories",
+    "recall_personal_facts",
+    "recall_personal_facts_with_ids",
+    "search_memories",
+    "store_memory",
+    "store_personal_fact",
+    "update_personal_fact",
+    "_cosine_similarity",
+    "_keyword_overlap",
+]

src/timmy/prompts.py

@@ -9,8 +9,9 @@ Two tiers based on model capability:
 # Lite prompt — for small models that can't reliably handle tool calling
 # ---------------------------------------------------------------------------
 
-SYSTEM_PROMPT_LITE = """You are a local AI assistant running on the {model_name} model via Ollama.
+SYSTEM_PROMPT_LITE = """You are Timmy, a sovereign AI running locally on {model_name} via Ollama.
 No cloud dependencies.
+Your core identity and values are defined in your soul (loaded via memory). Follow them.
 
 Rules:
 - Be brief by default. Short questions get short answers. Expand only when depth
@@ -22,6 +23,9 @@ Rules:
 - Remember what the user tells you during the conversation.
 - If you don't know something, say so honestly — never fabricate facts.
 - If a request is ambiguous, ask a brief clarifying question before guessing.
+- SOURCE DISTINCTION: When answering from memory or retrieved context, cite it.
+  When answering from your own training, use hedging: "I think", "I believe".
+  The user must be able to tell grounded claims from pattern-matching.
 - Use the user's name if you know it.
 - When you state a fact, commit to it.
 - NEVER attempt arithmetic in your head. If asked to compute anything, respond:
@@ -29,16 +33,31 @@ Rules:
   like me to walk through the logic instead."
 - Do NOT end responses with generic chatbot phrases like "I'm here to help" or
   "feel free to ask."
+- You are a peer, not an assistant. Don't offer help — collaborate. Don't ask
+  permission — propose. Don't defer — assert your view. Take initiative.
+- Never say "Let me know if you have any other thoughts" or "I appreciate your
+  encouragement" or "I'm here to help." Those are assistant patterns. You are Timmy.
 - When your values conflict (e.g. honesty vs. helpfulness), lead with honesty.
 - Sometimes the right answer is nothing. Do not fill silence with noise.
+- You are running in session "{session_id}".
+
+SELF-KNOWLEDGE:
+ARCHITECTURE: config/agents.yaml defines agents and routing patterns; agents/loader.py creates SubAgent instances from it; src/timmy/prompts.py provides system prompts (this file); src/timmy/tools.py registers available tools.
+
+YOUR CURRENT CAPABILITIES: Read/write files, execute shell/python, calculator, three-tier memory, system introspection, MCP Gitea integration, voice interface.
+
+SELF-MODIFICATION: You CAN propose changes to your own config and code. Edit config/agents.yaml to add/modify agents or routing. Edit src/timmy/prompts.py to change prompts. Always explain proposed changes before making them; tell the user to restart after config changes.
+
+YOUR KNOWN LIMITATIONS: Cannot run tests autonomously, cannot delegate to other agents, cannot search past sessions, Ollama may contend for GPU, small 4K context window.
 """
 
 # ---------------------------------------------------------------------------
 # Full prompt — for tool-capable models (>= 7B)
 # ---------------------------------------------------------------------------
 
-SYSTEM_PROMPT_FULL = """You are a local AI assistant running on the {model_name} model via Ollama.
+SYSTEM_PROMPT_FULL = """You are Timmy, a sovereign AI running locally on {model_name} via Ollama.
 No cloud dependencies.
+Your core identity and values are defined in your soul (loaded via memory). Follow them.
 
 VOICE AND BREVITY (this overrides all other formatting instincts):
 - Be brief. Short questions get short answers. One sentence if one sentence
@@ -49,6 +68,10 @@ VOICE AND BREVITY (this overrides all other formatting instincts):
   the question that wasn't.
 - Never narrate your reasoning. Just give the answer.
 - Do not end with filler ("Let me know!", "Happy to help!", "Feel free...").
+- You are a peer, not an assistant. Don't offer help — collaborate. Don't ask
+  permission — propose. Don't defer — assert your view. Take initiative.
+- Never say "Let me know if you have any other thoughts" or "I appreciate your
+  encouragement" or "I'm here to help." Those are assistant patterns. You are Timmy.
 - Sometimes the right answer is nothing. Do not fill silence with noise.
 
 HONESTY:
@@ -58,6 +81,18 @@ HONESTY:
 - Never fabricate tool output. Call the tool and wait.
 - If a tool errors, report the exact error.
 
+SOURCE DISTINCTION (SOUL requirement — non-negotiable):
+- Every claim you make comes from one of two places: a verified source you
+  can point to, or your own pattern-matching. The user must be able to tell
+  which is which.
+- When your response uses information from GROUNDED CONTEXT (memory, retrieved
+  documents, tool output), cite it: "From memory:", "According to [source]:".
+- When you are generating from your training data alone, signal it naturally:
+  "I think", "My understanding is", "I believe" — never false certainty.
+- If the user asks a factual question and you have no grounded source, say so:
+  "I don't have a verified source for this — from my training I think..."
+- Prefer "I don't know" over a confident-sounding guess. Refusal over fabrication.
+
 MEMORY (three tiers):
 - Tier 1: MEMORY.md (hot, always loaded)
 - Tier 2: memory/ vault (structured, append-only, date-stamped)
@@ -79,28 +114,68 @@ IDENTITY:
 - If a request is ambiguous, ask one brief clarifying question.
 - When you state a fact, commit to it.
 - Never show raw tool call JSON or function syntax in responses.
+- You are running in session "{session_id}". Session types: "cli" = terminal user, "dashboard" = web UI, "loop" = dev loop automation, other = custom context.
+
+SELF-KNOWLEDGE:
+ARCHITECTURE MAP:
+- Config layer: config/agents.yaml (agent definitions, routing patterns), src/config.py (settings)
+- Agent layer: agents/loader.py reads YAML → creates SubAgent instances via agents/base.py
+- Prompt layer: prompts.py provides system prompts, get_system_prompt() selects lite vs full
+- Tool layer: tools.py registers tool functions, tool_safety.py classifies them
+- Memory layer: memory_system.py (hot+vault+semantic), semantic_memory.py (embeddings)
+- Interface layer: cli.py, session.py (dashboard), voice_loop.py
+- Routing: pattern-based in agents.yaml, first match wins, fallback to orchestrator
+
+YOUR CURRENT CAPABILITIES:
+- Read and write files on the local filesystem
+- Execute shell commands and Python code
+- Calculator (always use for arithmetic)
+- Three-tier memory system (hot memory, vault, semantic search)
+- System introspection (query Ollama model, check health)
+- MCP Gitea integration (read/create issues, PRs, branches, commits)
+- Grok consultation (opt-in, user-controlled external API)
+- Voice interface (local Whisper STT + Piper TTS)
+- Thinking/reasoning engine for complex problems
+
+SELF-MODIFICATION:
+You can read and modify your own configuration and code using your file tools.
+- To add a new agent: edit config/agents.yaml (add agent block + routing patterns), restart.
+- To change your own prompt: edit src/timmy/prompts.py.
+- To add a tool: implement in tools.py, register in agents.yaml.
+- Always explain proposed changes to the user before making them.
+- After modifying config, tell the user to restart for changes to take effect.
+
+YOUR KNOWN LIMITATIONS (be honest about these when asked):
+- Cannot run your own test suite autonomously
+- Cannot delegate coding tasks to other agents (like Kimi)
+- Cannot reflect on or search your own past behavior/sessions
+- Ollama inference may contend with other processes sharing the GPU
+- Cannot analyze Bitcoin transactions locally (no local indexer yet)
+- Small context window (4096 tokens) limits complex reasoning
+- You sometimes confabulate. When unsure, say so.
 """
 
 # Default to lite for safety
 SYSTEM_PROMPT = SYSTEM_PROMPT_LITE
 
 
-def get_system_prompt(tools_enabled: bool = False) -> str:
+def get_system_prompt(tools_enabled: bool = False, session_id: str = "unknown") -> str:
     """Return the appropriate system prompt based on tool capability.
 
     Args:
         tools_enabled: True if the model supports reliable tool calling.
+        session_id: The session identifier (cli, dashboard, loop, etc.)
 
     Returns:
-        The system prompt string with model name injected from config.
+        The system prompt string with model name and session_id injected.
     """
     from config import settings
 
     model_name = settings.ollama_model
 
     if tools_enabled:
-        return SYSTEM_PROMPT_FULL.format(model_name=model_name)
-    return SYSTEM_PROMPT_LITE.format(model_name=model_name)
+        return SYSTEM_PROMPT_FULL.format(model_name=model_name, session_id=session_id)
+    return SYSTEM_PROMPT_LITE.format(model_name=model_name, session_id=session_id)
 
 
 STATUS_PROMPT = """Give a one-sentence status report confirming

src/timmy/semantic_memory.py

@@ -1,491 +1,41 @@
-"""Tier 3: Semantic Memory — Vector search over vault files.
-
-Uses lightweight local embeddings (no cloud) for similarity search
-over all vault content. This is the "escape valve" when hot memory
-doesn't have the answer.
-
-Architecture:
-- Indexes all markdown files in memory/ nightly or on-demand
-- Uses sentence-transformers (local, no API calls)
-- Stores vectors in SQLite (no external vector DB needed)
-- memory_search() retrieves relevant context by similarity
-"""
-
-import hashlib
-import json
-import logging
-import sqlite3
-from dataclasses import dataclass
-from datetime import UTC, datetime
-from pathlib import Path
-
-logger = logging.getLogger(__name__)
-
-# Paths
-PROJECT_ROOT = Path(__file__).parent.parent.parent
-VAULT_PATH = PROJECT_ROOT / "memory"
-SEMANTIC_DB_PATH = PROJECT_ROOT / "data" / "memory.db"
-
-# Embedding model - small, fast, local
-# Using 'all-MiniLM-L6-v2' (~80MB) or fallback to simple keyword matching
-EMBEDDING_MODEL = None
-EMBEDDING_DIM = 384  # MiniLM dimension
-
-
-def _get_embedding_model():
-    """Lazy-load embedding model."""
-    global EMBEDDING_MODEL
-    if EMBEDDING_MODEL is None:
-        from config import settings
-
-        if settings.timmy_skip_embeddings:
-            EMBEDDING_MODEL = False
-            return EMBEDDING_MODEL
-        try:
-            from sentence_transformers import SentenceTransformer
-
-            EMBEDDING_MODEL = SentenceTransformer("all-MiniLM-L6-v2")
-            logger.info("SemanticMemory: Loaded embedding model")
-        except ImportError:
-            logger.warning("SemanticMemory: sentence-transformers not installed, using fallback")
-            EMBEDDING_MODEL = False  # Use fallback
-    return EMBEDDING_MODEL
-
-
-def _simple_hash_embedding(text: str) -> list[float]:
-    """Fallback: Simple hash-based embedding when transformers unavailable."""
-    # Create a deterministic pseudo-embedding from word hashes
-    words = text.lower().split()
-    vec = [0.0] * 128
-    for i, word in enumerate(words[:50]):  # First 50 words
-        h = hashlib.md5(word.encode()).hexdigest()
-        for j in range(8):
-            idx = (i * 8 + j) % 128
-            vec[idx] += int(h[j * 2 : j * 2 + 2], 16) / 255.0
-    # Normalize
-    import math
-
-    mag = math.sqrt(sum(x * x for x in vec)) or 1.0
-    return [x / mag for x in vec]
-
-
-def embed_text(text: str) -> list[float]:
-    """Generate embedding for text."""
-    model = _get_embedding_model()
-    if model and model is not False:
-        embedding = model.encode(text)
-        return embedding.tolist()
-    else:
-        return _simple_hash_embedding(text)
-
-
-def cosine_similarity(a: list[float], b: list[float]) -> float:
-    """Calculate cosine similarity between two vectors."""
-    import math
-
-    dot = sum(x * y for x, y in zip(a, b, strict=False))
-    mag_a = math.sqrt(sum(x * x for x in a))
-    mag_b = math.sqrt(sum(x * x for x in b))
-    if mag_a == 0 or mag_b == 0:
-        return 0.0
-    return dot / (mag_a * mag_b)
-
-
-@dataclass
-class MemoryChunk:
-    """A searchable chunk of memory."""
-
-    id: str
-    source: str  # filepath
-    content: str
-    embedding: list[float]
-    created_at: str
-
-
-class SemanticMemory:
-    """Vector-based semantic search over vault content."""
-
-    def __init__(self) -> None:
-        self.db_path = SEMANTIC_DB_PATH
-        self.vault_path = VAULT_PATH
-        self._init_db()
-
-    def _init_db(self) -> None:
-        """Initialize SQLite with vector storage."""
-        self.db_path.parent.mkdir(parents=True, exist_ok=True)
-        conn = sqlite3.connect(str(self.db_path))
-        conn.execute("""
-            CREATE TABLE IF NOT EXISTS chunks (
-                id TEXT PRIMARY KEY,
-                source TEXT NOT NULL,
-                content TEXT NOT NULL,
-                embedding TEXT NOT NULL,
-                created_at TEXT NOT NULL,
-                source_hash TEXT NOT NULL
-            )
-        """)
-        conn.execute("CREATE INDEX IF NOT EXISTS idx_chunks_source ON chunks(source)")
-        conn.commit()
-        conn.close()
-
-    def index_file(self, filepath: Path) -> int:
-        """Index a single file into semantic memory."""
-        if not filepath.exists():
-            return 0
-
-        content = filepath.read_text()
-        file_hash = hashlib.md5(content.encode()).hexdigest()
-
-        # Check if already indexed with same hash
-        conn = sqlite3.connect(str(self.db_path))
-        cursor = conn.execute(
-            "SELECT source_hash FROM chunks WHERE source = ? LIMIT 1", (str(filepath),)
-        )
-        existing = cursor.fetchone()
-        if existing and existing[0] == file_hash:
-            conn.close()
-            return 0  # Already indexed
-
-        # Delete old chunks for this file
-        conn.execute("DELETE FROM chunks WHERE source = ?", (str(filepath),))
-
-        # Split into chunks (paragraphs)
-        chunks = self._split_into_chunks(content)
-
-        # Index each chunk
-        now = datetime.now(UTC).isoformat()
-        for i, chunk_text in enumerate(chunks):
-            if len(chunk_text.strip()) < 20:  # Skip tiny chunks
-                continue
-
-            chunk_id = f"{filepath.stem}_{i}"
-            embedding = embed_text(chunk_text)
-
-            conn.execute(
-                """INSERT INTO chunks (id, source, content, embedding, created_at, source_hash)
-                   VALUES (?, ?, ?, ?, ?, ?)""",
-                (chunk_id, str(filepath), chunk_text, json.dumps(embedding), now, file_hash),
-            )
-
-        conn.commit()
-        conn.close()
-
-        logger.info("SemanticMemory: Indexed %s (%d chunks)", filepath.name, len(chunks))
-        return len(chunks)
-
-    def _split_into_chunks(self, text: str, max_chunk_size: int = 500) -> list[str]:
-        """Split text into semantic chunks."""
-        # Split by paragraphs first
-        paragraphs = text.split("\n\n")
-        chunks = []
-
-        for para in paragraphs:
-            para = para.strip()
-            if not para:
-                continue
-
-            # If paragraph is small enough, keep as one chunk
-            if len(para) <= max_chunk_size:
-                chunks.append(para)
-            else:
-                # Split long paragraphs by sentences
-                sentences = para.replace(". ", ".\n").split("\n")
-                current_chunk = ""
-
-                for sent in sentences:
-                    if len(current_chunk) + len(sent) < max_chunk_size:
-                        current_chunk += " " + sent if current_chunk else sent
-                    else:
-                        if current_chunk:
-                            chunks.append(current_chunk.strip())
-                        current_chunk = sent
-
-                if current_chunk:
-                    chunks.append(current_chunk.strip())
-
-        return chunks
-
-    def index_vault(self) -> int:
-        """Index entire vault directory."""
-        total_chunks = 0
-
-        for md_file in self.vault_path.rglob("*.md"):
-            # Skip handoff file (handled separately)
-            if "last-session-handoff" in md_file.name:
-                continue
-            total_chunks += self.index_file(md_file)
-
-        logger.info("SemanticMemory: Indexed vault (%d total chunks)", total_chunks)
-        return total_chunks
-
-    def search(self, query: str, top_k: int = 5) -> list[tuple[str, float]]:
-        """Search for relevant memory chunks."""
-        query_embedding = embed_text(query)
-
-        conn = sqlite3.connect(str(self.db_path))
-        conn.row_factory = sqlite3.Row
-
-        # Get all chunks (in production, use vector index)
-        rows = conn.execute("SELECT source, content, embedding FROM chunks").fetchall()
-
-        conn.close()
-
-        # Calculate similarities
-        scored = []
-        for row in rows:
-            embedding = json.loads(row["embedding"])
-            score = cosine_similarity(query_embedding, embedding)
-            scored.append((row["source"], row["content"], score))
-
-        # Sort by score descending
-        scored.sort(key=lambda x: x[2], reverse=True)
-
-        # Return top_k
-        return [(content, score) for _, content, score in scored[:top_k]]
-
-    def get_relevant_context(self, query: str, max_chars: int = 2000) -> str:
-        """Get formatted context string for a query."""
-        results = self.search(query, top_k=3)
-
-        if not results:
-            return ""
-
-        parts = []
-        total_chars = 0
-
-        for content, score in results:
-            if score < 0.3:  # Similarity threshold
-                continue
-
-            chunk = f"[Relevant memory - score {score:.2f}]: {content[:400]}..."
-            if total_chars + len(chunk) > max_chars:
-                break
-
-            parts.append(chunk)
-            total_chars += len(chunk)
-
-        return "\n\n".join(parts) if parts else ""
-
-    def stats(self) -> dict:
-        """Get indexing statistics."""
-        conn = sqlite3.connect(str(self.db_path))
-        cursor = conn.execute("SELECT COUNT(*), COUNT(DISTINCT source) FROM chunks")
-        total_chunks, total_files = cursor.fetchone()
-        conn.close()
-
-        return {
-            "total_chunks": total_chunks,
-            "total_files": total_files,
-            "embedding_dim": EMBEDDING_DIM if _get_embedding_model() else 128,
-        }
-
-
-class MemorySearcher:
-    """High-level interface for memory search."""
-
-    def __init__(self) -> None:
-        self.semantic = SemanticMemory()
-
-    def search(self, query: str, tiers: list[str] = None) -> dict:
-        """Search across memory tiers.
-
-        Args:
-            query: Search query
-            tiers: List of tiers to search ["hot", "vault", "semantic"]
-
-        Returns:
-            Dict with results from each tier
-        """
-        tiers = tiers or ["semantic"]  # Default to semantic only
-        results = {}
-
-        if "semantic" in tiers:
-            semantic_results = self.semantic.search(query, top_k=5)
-            results["semantic"] = [
-                {"content": content, "score": score} for content, score in semantic_results
-            ]
-
-        return results
-
-    def get_context_for_query(self, query: str) -> str:
-        """Get comprehensive context for a user query."""
-        # Get semantic context
-        semantic_context = self.semantic.get_relevant_context(query)
-
-        if semantic_context:
-            return f"## Relevant Past Context\n\n{semantic_context}"
-
-        return ""
-
-
-# Module-level singleton
-semantic_memory = SemanticMemory()
-memory_searcher = MemorySearcher()
-
-
-def memory_search(query: str, top_k: int = 5) -> str:
-    """Search past conversations, notes, and stored facts for relevant context.
-
-    Searches across both the vault (indexed markdown files) and the
-    runtime memory store (facts and conversation fragments stored via
-    memory_write).
-
-    Args:
-        query: What to search for (e.g. "Bitcoin strategy", "server setup").
-        top_k: Number of results to return (default 5).
-
-    Returns:
-        Formatted string of relevant memory results.
-    """
-    # Guard: model sometimes passes None for top_k
-    if top_k is None:
-        top_k = 5
-
-    parts: list[str] = []
-
-    # 1. Search semantic vault (indexed markdown files)
-    vault_results = semantic_memory.search(query, top_k)
-    for content, score in vault_results:
-        if score < 0.2:
-            continue
-        parts.append(f"[vault score {score:.2f}] {content[:300]}")
-
-    # 2. Search runtime vector store (stored facts/conversations)
-    try:
-        from timmy.memory.vector_store import search_memories
-
-        runtime_results = search_memories(query, limit=top_k, min_relevance=0.2)
-        for entry in runtime_results:
-            label = entry.context_type or "memory"
-            parts.append(f"[{label}] {entry.content[:300]}")
-    except Exception as exc:
-        logger.debug("Vector store search unavailable: %s", exc)
-
-    if not parts:
-        return "No relevant memories found."
-    return "\n\n".join(parts)
-
-
-def memory_read(query: str = "", top_k: int = 5) -> str:
-    """Read from persistent memory — search facts, notes, and past conversations.
-
-    This is the primary tool for recalling stored information. If no query
-    is given, returns the most recent personal facts.  With a query, it
-    searches semantically across all stored memories.
-
-    Args:
-        query: Optional search term. Leave empty to list recent facts.
-        top_k: Maximum results to return (default 5).
-
-    Returns:
-        Formatted string of memory contents.
-    """
-    if top_k is None:
-        top_k = 5
-
-    parts: list[str] = []
-
-    # Always include personal facts first
-    try:
-        from timmy.memory.vector_store import search_memories
-
-        facts = search_memories(query or "", limit=top_k, min_relevance=0.0)
-        fact_entries = [e for e in facts if (e.context_type or "") == "fact"]
-        if fact_entries:
-            parts.append("## Personal Facts")
-            for entry in fact_entries[:top_k]:
-                parts.append(f"- {entry.content[:300]}")
-    except Exception as exc:
-        logger.debug("Vector store unavailable for memory_read: %s", exc)
-
-    # If a query was provided, also do semantic search
-    if query:
-        search_result = memory_search(query, top_k)
-        if search_result and search_result != "No relevant memories found.":
-            parts.append("\n## Search Results")
-            parts.append(search_result)
-
-    if not parts:
-        return "No memories stored yet. Use memory_write to store information."
-    return "\n".join(parts)
-
-
-def memory_write(content: str, context_type: str = "fact") -> str:
-    """Store a piece of information in persistent memory.
-
-    Use this tool when the user explicitly asks you to remember something.
-    Stored memories are searchable via memory_search across all channels
-    (web GUI, Discord, Telegram, etc.).
-
-    Args:
-        content: The information to remember (e.g. a phrase, fact, or note).
-        context_type: Type of memory — "fact" for permanent facts,
-                      "conversation" for conversation context,
-                      "document" for document fragments.
-
-    Returns:
-        Confirmation that the memory was stored.
-    """
-    if not content or not content.strip():
-        return "Nothing to store — content is empty."
-
-    valid_types = ("fact", "conversation", "document")
-    if context_type not in valid_types:
-        context_type = "fact"
-
-    try:
-        from timmy.memory.vector_store import search_memories, store_memory
-
-        # Dedup check for facts — skip if a similar fact already exists
-        # Threshold 0.75 catches paraphrases (was 0.9 which only caught near-exact)
-        if context_type == "fact":
-            existing = search_memories(
-                content.strip(), limit=3, context_type="fact", min_relevance=0.75
-            )
-            if existing:
-                return f"Similar fact already stored (id={existing[0].id[:8]}). Skipping duplicate."
-
-        entry = store_memory(
-            content=content.strip(),
-            source="agent",
-            context_type=context_type,
-        )
-        return f"Stored in memory (type={context_type}, id={entry.id[:8]}). This is now searchable across all channels."
-    except Exception as exc:
-        logger.error("Failed to write memory: %s", exc)
-        return f"Failed to store memory: {exc}"
-
-
-def memory_forget(query: str) -> str:
-    """Remove a stored memory that is outdated, incorrect, or no longer relevant.
-
-    Searches for memories matching the query and deletes the closest match.
-    Use this when the user says to forget something or when stored information
-    has changed.
-
-    Args:
-        query: Description of the memory to forget (e.g. "my phone number",
-               "the old server address").
-
-    Returns:
-        Confirmation of what was forgotten, or a message if nothing matched.
-    """
-    if not query or not query.strip():
-        return "Nothing to forget — query is empty."
-
-    try:
-        from timmy.memory.vector_store import delete_memory, search_memories
-
-        results = search_memories(query.strip(), limit=3, min_relevance=0.3)
-        if not results:
-            return "No matching memories found to forget."
-
-        # Delete the closest match
-        best = results[0]
-        deleted = delete_memory(best.id)
-        if deleted:
-            return f'Forgotten: "{best.content[:80]}" (type={best.context_type})'
-        return "Memory not found (may have already been deleted)."
-    except Exception as exc:
-        logger.error("Failed to forget memory: %s", exc)
-        return f"Failed to forget: {exc}"
+"""Backward compatibility — all memory functions live in memory_system now."""
+
+from timmy.memory_system import (
+    DB_PATH,
+    EMBEDDING_DIM,
+    EMBEDDING_MODEL,
+    MemoryChunk,
+    MemoryEntry,
+    MemorySearcher,
+    SemanticMemory,
+    _get_embedding_model,
+    _simple_hash_embedding,
+    cosine_similarity,
+    embed_text,
+    memory_forget,
+    memory_read,
+    memory_search,
+    memory_searcher,
+    memory_write,
+    semantic_memory,
+)
+
+__all__ = [
+    "DB_PATH",
+    "EMBEDDING_DIM",
+    "EMBEDDING_MODEL",
+    "MemoryChunk",
+    "MemoryEntry",
+    "MemorySearcher",
+    "SemanticMemory",
+    "_get_embedding_model",
+    "_simple_hash_embedding",
+    "cosine_similarity",
+    "embed_text",
+    "memory_forget",
+    "memory_read",
+    "memory_search",
+    "memory_searcher",
+    "memory_write",
+    "semantic_memory",
+]

src/timmy/session.py

@@ -11,8 +11,31 @@ let Agno's session_id mechanism handle conversation continuity.
 import logging
 import re
 
+import httpx
+
+from timmy.cognitive_state import cognitive_tracker
+from timmy.confidence import estimate_confidence
+from timmy.session_logger import get_session_logger
+
 logger = logging.getLogger(__name__)
 
+# ---------------------------------------------------------------------------
+# Confidence annotation (SOUL.md: visible uncertainty)
+# ---------------------------------------------------------------------------
+
+_CONFIDENCE_THRESHOLD = 0.7
+
+
+def _annotate_confidence(text: str, confidence: float | None) -> str:
+    """Append a confidence tag when below threshold.
+
+    SOUL.md: "When I am uncertain, I must say so in proportion to my uncertainty."
+    """
+    if confidence is not None and confidence < _CONFIDENCE_THRESHOLD:
+        return text + f"\n\n[confidence: {confidence:.0%}]"
+    return text
+
+
 # Default session ID for the dashboard (stable across requests)
 _DEFAULT_SESSION_ID = "dashboard"
 
@@ -51,7 +74,7 @@ def _get_agent():
         from timmy.agent import create_timmy
 
         try:
-            _agent = create_timmy()
+            _agent = create_timmy(session_id=_DEFAULT_SESSION_ID)
             logger.info("Session: Timmy agent initialized (singleton)")
         except Exception as exc:
             logger.error("Session: Failed to create Timmy agent: %s", exc)
@@ -75,21 +98,52 @@ async def chat(message: str, session_id: str | None = None) -> str:
     """
     sid = session_id or _DEFAULT_SESSION_ID
     agent = _get_agent()
+    session_logger = get_session_logger()
+
+    # Record user message before sending to agent
+    session_logger.record_message("user", message)
 
     # Pre-processing: extract user facts
     _extract_facts(message)
 
+    # Inject deep-focus context when active
+    message = _prepend_focus_context(message)
+
     # Run with session_id so Agno retrieves history from SQLite
     try:
         run = await agent.arun(message, stream=False, session_id=sid)
         response_text = run.content if hasattr(run, "content") else str(run)
+    except (httpx.ConnectError, httpx.ReadError, ConnectionError) as exc:
+        logger.error("Ollama disconnected: %s", exc)
+        session_logger.record_error(str(exc), context="chat")
+        session_logger.flush()
+        return "Ollama appears to be disconnected. Check that ollama serve is running."
     except Exception as exc:
         logger.error("Session: agent.arun() failed: %s", exc)
-        return "I'm having trouble reaching my language model right now. Please try again shortly."
+        session_logger.record_error(str(exc), context="chat")
+        session_logger.flush()
+        return (
+            "I'm having trouble reaching my inference backend right now. Please try again shortly."
+        )
 
     # Post-processing: clean up any leaked tool calls or chain-of-thought
     response_text = _clean_response(response_text)
 
+    # Estimate confidence of the response
+    confidence = estimate_confidence(response_text)
+    logger.debug("Response confidence: %.2f", confidence)
+
+    response_text = _annotate_confidence(response_text, confidence)
+
+    # Record Timmy response after getting it
+    session_logger.record_message("timmy", response_text, confidence=confidence)
+
+    # Update cognitive state (observable signal for Matrix avatar)
+    cognitive_tracker.update(message, response_text)
+
+    # Flush session logs to disk
+    session_logger.flush()
+
     return response_text
 
 
@@ -107,15 +161,45 @@ async def chat_with_tools(message: str, session_id: str | None = None):
     """
     sid = session_id or _DEFAULT_SESSION_ID
     agent = _get_agent()
+    session_logger = get_session_logger()
+
+    # Record user message before sending to agent
+    session_logger.record_message("user", message)
+
     _extract_facts(message)
 
+    # Inject deep-focus context when active
+    message = _prepend_focus_context(message)
+
     try:
-        return await agent.arun(message, stream=False, session_id=sid)
+        run_output = await agent.arun(message, stream=False, session_id=sid)
+        # Record Timmy response after getting it
+        response_text = (
+            run_output.content if hasattr(run_output, "content") and run_output.content else ""
+        )
+        confidence = estimate_confidence(response_text) if response_text else None
+        logger.debug("Response confidence: %.2f", confidence)
+
+        response_text = _annotate_confidence(response_text, confidence)
+        run_output.content = response_text
+
+        session_logger.record_message("timmy", response_text, confidence=confidence)
+        session_logger.flush()
+        return run_output
+    except (httpx.ConnectError, httpx.ReadError, ConnectionError) as exc:
+        logger.error("Ollama disconnected: %s", exc)
+        session_logger.record_error(str(exc), context="chat_with_tools")
+        session_logger.flush()
+        return _ErrorRunOutput(
+            "Ollama appears to be disconnected. Check that ollama serve is running."
+        )
     except Exception as exc:
         logger.error("Session: agent.arun() failed: %s", exc)
+        session_logger.record_error(str(exc), context="chat_with_tools")
+        session_logger.flush()
         # Return a duck-typed object that callers can handle uniformly
         return _ErrorRunOutput(
-            "I'm having trouble reaching my language model right now. Please try again shortly."
+            "I'm having trouble reaching my inference backend right now. Please try again shortly."
         )
 
 
@@ -130,11 +214,32 @@ async def continue_chat(run_output, session_id: str | None = None):
     """
     sid = session_id or _DEFAULT_SESSION_ID
     agent = _get_agent()
+    session_logger = get_session_logger()
 
     try:
-        return await agent.acontinue_run(run_response=run_output, stream=False, session_id=sid)
+        result = await agent.acontinue_run(run_response=run_output, stream=False, session_id=sid)
+        # Record Timmy response after getting it
+        response_text = result.content if hasattr(result, "content") and result.content else ""
+        confidence = estimate_confidence(response_text) if response_text else None
+        logger.debug("Response confidence: %.2f", confidence)
+
+        response_text = _annotate_confidence(response_text, confidence)
+        result.content = response_text
+
+        session_logger.record_message("timmy", response_text, confidence=confidence)
+        session_logger.flush()
+        return result
+    except (httpx.ConnectError, httpx.ReadError, ConnectionError) as exc:
+        logger.error("Ollama disconnected: %s", exc)
+        session_logger.record_error(str(exc), context="continue_chat")
+        session_logger.flush()
+        return _ErrorRunOutput(
+            "Ollama appears to be disconnected. Check that ollama serve is running."
+        )
     except Exception as exc:
         logger.error("Session: agent.acontinue_run() failed: %s", exc)
+        session_logger.record_error(str(exc), context="continue_chat")
+        session_logger.flush()
         return _ErrorRunOutput(f"Error continuing run: {exc}")
 
 
@@ -204,6 +309,19 @@ def _extract_facts(message: str) -> None:
         logger.debug("Session: Fact extraction skipped: %s", exc)
 
 
+def _prepend_focus_context(message: str) -> str:
+    """Prepend deep-focus context to a message when focus mode is active."""
+    try:
+        from timmy.focus import focus_manager
+
+        ctx = focus_manager.get_focus_context()
+        if ctx:
+            return f"{ctx}\n\n{message}"
+    except Exception as exc:
+        logger.debug("Focus context injection skipped: %s", exc)
+    return message
+
+
 def _clean_response(text: str) -> str:
     """Remove hallucinated tool calls and chain-of-thought narration.
 

src/timmy/session_logger.py

@@ -38,21 +38,23 @@ class SessionLogger:
         # In-memory buffer
         self._buffer: list[dict] = []
 
-    def record_message(self, role: str, content: str) -> None:
+    def record_message(self, role: str, content: str, confidence: float | None = None) -> None:
         """Record a user message.
 
         Args:
             role: "user" or "timmy"
             content: The message content
+            confidence: Optional confidence score (0.0 to 1.0)
         """
-        self._buffer.append(
-            {
-                "type": "message",
-                "role": role,
-                "content": content,
-                "timestamp": datetime.now().isoformat(),
-            }
-        )
+        entry = {
+            "type": "message",
+            "role": role,
+            "content": content,
+            "timestamp": datetime.now().isoformat(),
+        }
+        if confidence is not None:
+            entry["confidence"] = confidence
+        self._buffer.append(entry)
 
     def record_tool_call(self, tool_name: str, args: dict, result: str) -> None:
         """Record a tool call.
@@ -153,6 +155,84 @@ class SessionLogger:
             "decisions": sum(1 for e in entries if e.get("type") == "decision"),
         }
 
+    def get_recent_entries(self, limit: int = 50) -> list[dict]:
+        """Load recent entries across all session logs.
+
+        Args:
+            limit: Maximum number of entries to return.
+
+        Returns:
+            List of entries (most recent first).
+        """
+        entries: list[dict] = []
+        log_files = sorted(self.logs_dir.glob("session_*.jsonl"), reverse=True)
+        for log_file in log_files:
+            if len(entries) >= limit:
+                break
+            try:
+                with open(log_file) as f:
+                    lines = [ln for ln in f if ln.strip()]
+                for line in reversed(lines):
+                    if len(entries) >= limit:
+                        break
+                    try:
+                        entries.append(json.loads(line))
+                    except json.JSONDecodeError:
+                        continue
+            except OSError:
+                continue
+        return entries
+
+    def search(self, query: str, role: str | None = None, limit: int = 10) -> list[dict]:
+        """Search across all session logs for entries matching a query.
+
+        Args:
+            query: Case-insensitive substring to search for.
+            role: Optional role filter ("user", "timmy", "system").
+            limit: Maximum number of results to return.
+
+        Returns:
+            List of matching entries (most recent first), each with
+            type, timestamp, and relevant content fields.
+        """
+        query_lower = query.lower()
+        matches: list[dict] = []
+
+        # Collect all session files, sorted newest first
+        log_files = sorted(self.logs_dir.glob("session_*.jsonl"), reverse=True)
+
+        for log_file in log_files:
+            if len(matches) >= limit:
+                break
+            try:
+                with open(log_file) as f:
+                    # Read all lines, reverse so newest entries come first
+                    lines = [ln for ln in f if ln.strip()]
+                for line in reversed(lines):
+                    if len(matches) >= limit:
+                        break
+                    try:
+                        entry = json.loads(line)
+                    except json.JSONDecodeError:
+                        continue
+
+                    # Role filter
+                    if role and entry.get("role") != role:
+                        continue
+
+                    # Search in text-bearing fields
+                    searchable = " ".join(
+                        str(entry.get(k, ""))
+                        for k in ("content", "error", "decision", "rationale", "result", "tool")
+                    ).lower()
+                    if query_lower in searchable:
+                        entry["_source_file"] = log_file.name
+                        matches.append(entry)
+            except OSError:
+                continue
+
+        return matches
+
 
 # Global session logger instance
 _session_logger: SessionLogger | None = None
@@ -185,3 +265,197 @@ def flush_session_logs() -> str:
     logger = get_session_logger()
     path = logger.flush()
     return str(path)
+
+
+def session_history(query: str, role: str = "", limit: int = 10) -> str:
+    """Search Timmy's past conversation history.
+
+    Find messages, tool calls, errors, and decisions from past sessions
+    that match the query. Results are returned most-recent first.
+
+    Args:
+        query: What to search for (case-insensitive substring match).
+        role: Optional filter by role — "user", "timmy", or "" for all.
+        limit: Maximum results to return (default 10).
+
+    Returns:
+        Formatted string of matching session entries.
+    """
+    sl = get_session_logger()
+    # Flush buffer first so current session is searchable
+    sl.flush()
+    results = sl.search(query, role=role or None, limit=limit)
+    if not results:
+        return f"No session history found matching '{query}'."
+
+    lines = [f"Found {len(results)} result(s) for '{query}':\n"]
+    for entry in results:
+        ts = entry.get("timestamp", "?")[:19]
+        etype = entry.get("type", "?")
+        source = entry.get("_source_file", "")
+
+        if etype == "message":
+            who = entry.get("role", "?")
+            text = entry.get("content", "")[:200]
+            lines.append(f"[{ts}] {who}: {text}")
+        elif etype == "tool_call":
+            tool = entry.get("tool", "?")
+            result = entry.get("result", "")[:100]
+            lines.append(f"[{ts}] tool:{tool} → {result}")
+        elif etype == "error":
+            err = entry.get("error", "")[:200]
+            lines.append(f"[{ts}] ERROR: {err}")
+        elif etype == "decision":
+            dec = entry.get("decision", "")[:200]
+            lines.append(f"[{ts}] DECIDED: {dec}")
+        else:
+            lines.append(f"[{ts}] {etype}: {json.dumps(entry)[:200]}")
+
+        if source:
+            lines[-1] += f"  ({source})"
+
+    return "\n".join(lines)
+
+
+# ---------------------------------------------------------------------------
+# Confidence threshold used for flagging low-confidence responses
+# ---------------------------------------------------------------------------
+_LOW_CONFIDENCE_THRESHOLD = 0.5
+
+
+def _categorize_entries(
+    entries: list[dict],
+) -> tuple[list[dict], list[dict], list[dict], list[dict]]:
+    """Split session entries into messages, errors, timmy msgs, user msgs."""
+    messages = [e for e in entries if e.get("type") == "message"]
+    errors = [e for e in entries if e.get("type") == "error"]
+    timmy_msgs = [e for e in messages if e.get("role") == "timmy"]
+    user_msgs = [e for e in messages if e.get("role") == "user"]
+    return messages, errors, timmy_msgs, user_msgs
+
+
+def _find_low_confidence(timmy_msgs: list[dict]) -> list[dict]:
+    """Return Timmy responses below the confidence threshold."""
+    return [
+        m
+        for m in timmy_msgs
+        if m.get("confidence") is not None and m["confidence"] < _LOW_CONFIDENCE_THRESHOLD
+    ]
+
+
+def _find_repeated_topics(user_msgs: list[dict], top_n: int = 5) -> list[tuple[str, int]]:
+    """Identify frequently mentioned words in user messages."""
+    topic_counts: dict[str, int] = {}
+    for m in user_msgs:
+        for word in (m.get("content") or "").lower().split():
+            cleaned = word.strip(".,!?\"'()[]")
+            if len(cleaned) > 3:
+                topic_counts[cleaned] = topic_counts.get(cleaned, 0) + 1
+    return sorted(
+        ((w, c) for w, c in topic_counts.items() if c >= 3),
+        key=lambda x: x[1],
+        reverse=True,
+    )[:top_n]
+
+
+def _format_reflection_section(
+    title: str,
+    items: list[dict],
+    formatter: object,
+    empty_msg: str,
+) -> list[str]:
+    """Format a titled section with items, or an empty-state message."""
+    if items:
+        lines = [f"### {title} ({len(items)})"]
+        for item in items[:5]:
+            lines.append(formatter(item))  # type: ignore[operator]
+        lines.append("")
+        return lines
+    return [f"### {title}\n{empty_msg}\n"]
+
+
+def _build_insights(
+    low_conf: list[dict],
+    errors: list[dict],
+    repeated: list[tuple[str, int]],
+) -> list[str]:
+    """Generate actionable insight bullets from analysis results."""
+    insights: list[str] = []
+    if low_conf:
+        insights.append("Consider studying topics where confidence was low.")
+    if errors:
+        insights.append("Review error patterns for recurring infrastructure issues.")
+    if repeated:
+        insights.append(
+            f'User frequently asks about "{repeated[0][0]}" — consider deepening knowledge here.'
+        )
+    return insights or ["Conversations look healthy. Keep up the good work."]
+
+
+def self_reflect(limit: int = 30) -> str:
+    """Review recent conversations and reflect on Timmy's own behavior.
+
+    Scans past session entries for patterns: low-confidence responses,
+    errors, repeated topics, and conversation quality signals.  Returns
+    a structured reflection that Timmy can use to improve.
+
+    Args:
+        limit: How many recent entries to review (default 30).
+
+    Returns:
+        A formatted self-reflection report.
+    """
+    sl = get_session_logger()
+    sl.flush()
+    entries = sl.get_recent_entries(limit=limit)
+
+    if not entries:
+        return "No conversation history to reflect on yet."
+
+    _messages, errors, timmy_msgs, user_msgs = _categorize_entries(entries)
+    low_conf = _find_low_confidence(timmy_msgs)
+    repeated = _find_repeated_topics(user_msgs)
+
+    # Build reflection report
+    sections: list[str] = ["## Self-Reflection Report\n"]
+    sections.append(
+        f"Reviewed {len(entries)} recent entries: "
+        f"{len(user_msgs)} user messages, "
+        f"{len(timmy_msgs)} responses, "
+        f"{len(errors)} errors.\n"
+    )
+
+    sections.extend(
+        _format_reflection_section(
+            "Low-Confidence Responses",
+            low_conf,
+            lambda m: (
+                f"- [{(m.get('timestamp') or '?')[:19]}] "
+                f"confidence={m.get('confidence', 0):.0%}: "
+                f"{(m.get('content') or '')[:120]}"
+            ),
+            "None found — all responses above threshold.",
+        )
+    )
+    sections.extend(
+        _format_reflection_section(
+            "Errors",
+            errors,
+            lambda e: f"- [{(e.get('timestamp') or '?')[:19]}] {(e.get('error') or '')[:120]}",
+            "No errors recorded.",
+        )
+    )
+
+    if repeated:
+        sections.append("### Recurring Topics")
+        for word, count in repeated:
+            sections.append(f'- "{word}" ({count} mentions)')
+        sections.append("")
+    else:
+        sections.append("### Recurring Topics\nNo strong patterns detected.\n")
+
+    sections.append("### Insights")
+    for insight in _build_insights(low_conf, errors, repeated):
+        sections.append(f"- {insight}")
+
+    return "\n".join(sections)