feat: dead man switch config fallback engine

Automatic fallback chain: Anthropic -> local-llama.cpp -> Ollama -> safe mode.
Auto-recovery when primary returns. Reversible config changes with backup.

.gitea/workflows/pr-checklist.yml

@@ -0,0 +1,29 @@
+# pr-checklist.yml — Automated PR quality gate
+# Refs: #393 (PERPLEXITY-08), Epic #385
+#
+# Enforces the review checklist that agents skip when left to self-approve.
+# Runs on every pull_request. Fails fast so bad PRs never reach a reviewer.
+
+name: PR Checklist
+
+on:
+  pull_request:
+    branches: [main, master]
+
+jobs:
+  pr-checklist:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Run PR checklist
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: python3 bin/pr-checklist.py

SOUL.md

@@ -1,3 +1,13 @@
+<!-- 
+  NOTE: This is the BITCOIN INSCRIPTION version of SOUL.md.
+  It is the immutable on-chain conscience. Do not modify this content.
+  
+  The NARRATIVE identity document (for onboarding, Audio Overviews, 
+  and system prompts) lives in timmy-home/SOUL.md.
+  
+  See: #388, #378 for the divergence audit.
+-->
+
 # SOUL.md
 
 ## Inscription 1 — The Immutable Conscience

bin/deadman-fallback.py

@@ -0,0 +1,264 @@
+     1|#!/usr/bin/env python3
+     2|"""
+     3|Dead Man Switch Fallback Engine
+     4|
+     5|When the dead man switch triggers (zero commits for 2+ hours, model down,
+     6|Gitea unreachable, etc.), this script diagnoses the failure and applies
+     7|common sense fallbacks automatically.
+     8|
+     9|Fallback chain:
+    10|1. Primary model (Anthropic) down -> switch config to local-llama.cpp
+    11|2. Gitea unreachable -> cache issues locally, retry on recovery
+    12|3. VPS agents down -> alert + lazarus protocol
+    13|4. Local llama.cpp down -> try Ollama, then alert-only mode
+    14|5. All inference dead -> safe mode (cron pauses, alert Alexander)
+    15|
+    16|Each fallback is reversible. Recovery auto-restores the previous config.
+    17|"""
+    18|import os
+    19|import sys
+    20|import json
+    21|import subprocess
+    22|import time
+    23|import yaml
+    24|import shutil
+    25|from pathlib import Path
+    26|from datetime import datetime, timedelta
+    27|
+    28|HERMES_HOME = Path(os.environ.get("HERMES_HOME", os.path.expanduser("~/.hermes")))
+    29|CONFIG_PATH = HERMES_HOME / "config.yaml"
+    30|FALLBACK_STATE = HERMES_HOME / "deadman-fallback-state.json"
+    31|BACKUP_CONFIG = HERMES_HOME / "config.yaml.pre-fallback"
+    32|FORGE_URL = "https://forge.alexanderwhitestone.com"
+    33|
+    34|def load_config():
+    35|    with open(CONFIG_PATH) as f:
+    36|        return yaml.safe_load(f)
+    37|
+    38|def save_config(cfg):
+    39|    with open(CONFIG_PATH, "w") as f:
+    40|        yaml.dump(cfg, f, default_flow_style=False)
+    41|
+    42|def load_state():
+    43|    if FALLBACK_STATE.exists():
+    44|        with open(FALLBACK_STATE) as f:
+    45|            return json.load(f)
+    46|    return {"active_fallbacks": [], "last_check": None, "recovery_pending": False}
+    47|
+    48|def save_state(state):
+    49|    state["last_check"] = datetime.now().isoformat()
+    50|    with open(FALLBACK_STATE, "w") as f:
+    51|        json.dump(state, f, indent=2)
+    52|
+    53|def run(cmd, timeout=10):
+    54|    try:
+    55|        r = subprocess.run(cmd, shell=True, capture_output=True, text=True, timeout=timeout)
+    56|        return r.returncode, r.stdout.strip(), r.stderr.strip()
+    57|    except subprocess.TimeoutExpired:
+    58|        return -1, "", "timeout"
+    59|    except Exception as e:
+    60|        return -1, "", str(e)
+    61|
+    62|# ─── HEALTH CHECKS ───
+    63|
+    64|def check_anthropic():
+    65|    """Can we reach Anthropic API?"""
+    66|    key = os.environ.get("ANTHROPIC_API_KEY", "")
+    67|    if not key:
+    68|        # Check multiple .env locations
+    69|        for env_path in [HERMES_HOME / ".env", Path.home() / ".hermes" / ".env"]:
+    70|            if env_path.exists():
+    71|                for line in open(env_path):
+    72|                    line = line.strip()
+    73|                    if line.startswith("ANTHROPIC_API_KEY=***
+    74|                        key = line.split("=", 1)[1].strip().strip('"').strip("'")
+    75|                        break
+    76|            if key:
+    77|                break
+    78|    if not key:
+    79|        return False, "no API key"
+    80|    code, out, err = run(
+    81|        f'curl -s -o /dev/null -w "%{{http_code}}" -H "x-api-key: {key}" '
+    82|        f'-H "anthropic-version: 2023-06-01" '
+    83|        f'https://api.anthropic.com/v1/messages -X POST '
+    84|        f'-H "content-type: application/json" '
+    85|        f'-d \'{{"model":"claude-haiku-4-5-20251001","max_tokens":1,"messages":[{{"role":"user","content":"ping"}}]}}\' ',
+    86|        timeout=15
+    87|    )
+    88|    if code == 0 and out in ("200", "429"):
+    89|        return True, f"HTTP {out}"
+    90|    return False, f"HTTP {out} err={err[:80]}"
+    91|
+    92|def check_local_llama():
+    93|    """Is local llama.cpp serving?"""
+    94|    code, out, err = run("curl -s http://localhost:8081/v1/models", timeout=5)
+    95|    if code == 0 and "hermes" in out.lower():
+    96|        return True, "serving"
+    97|    return False, f"exit={code}"
+    98|
+    99|def check_ollama():
+   100|    """Is Ollama running?"""
+   101|    code, out, err = run("curl -s http://localhost:11434/api/tags", timeout=5)
+   102|    if code == 0 and "models" in out:
+   103|        return True, "running"
+   104|    return False, f"exit={code}"
+   105|
+   106|def check_gitea():
+   107|    """Can we reach the Forge?"""
+   108|    token_path = Path.home() / ".config" / "gitea" / "timmy-token"
+   109|    if not token_path.exists():
+   110|        return False, "no token"
+   111|    token = token_path.read_text().strip()
+   112|    code, out, err = run(
+   113|        f'curl -s -o /dev/null -w "%{{http_code}}" -H "Authorization: token {token}" '
+   114|        f'"{FORGE_URL}/api/v1/user"',
+   115|        timeout=10
+   116|    )
+   117|    if code == 0 and out == "200":
+   118|        return True, "reachable"
+   119|    return False, f"HTTP {out}"
+   120|
+   121|def check_vps(ip, name):
+   122|    """Can we SSH into a VPS?"""
+   123|    code, out, err = run(f"ssh -o ConnectTimeout=5 root@{ip} 'echo alive'", timeout=10)
+   124|    if code == 0 and "alive" in out:
+   125|        return True, "alive"
+   126|    return False, f"unreachable"
+   127|
+   128|# ─── FALLBACK ACTIONS ───
+   129|
+   130|def fallback_to_local_model(cfg):
+   131|    """Switch primary model from Anthropic to local llama.cpp"""
+   132|    if not BACKUP_CONFIG.exists():
+   133|        shutil.copy2(CONFIG_PATH, BACKUP_CONFIG)
+   134|    
+   135|    cfg["model"]["provider"] = "local-llama.cpp"
+   136|    cfg["model"]["default"] = "hermes3"
+   137|    save_config(cfg)
+   138|    return "Switched primary model to local-llama.cpp/hermes3"
+   139|
+   140|def fallback_to_ollama(cfg):
+   141|    """Switch to Ollama if llama.cpp is also down"""
+   142|    if not BACKUP_CONFIG.exists():
+   143|        shutil.copy2(CONFIG_PATH, BACKUP_CONFIG)
+   144|    
+   145|    cfg["model"]["provider"] = "ollama"
+   146|    cfg["model"]["default"] = "gemma4:latest"
+   147|    save_config(cfg)
+   148|    return "Switched primary model to ollama/gemma4:latest"
+   149|
+   150|def enter_safe_mode(state):
+   151|    """Pause all non-essential cron jobs, alert Alexander"""
+   152|    state["safe_mode"] = True
+   153|    state["safe_mode_entered"] = datetime.now().isoformat()
+   154|    save_state(state)
+   155|    return "SAFE MODE: All inference down. Cron jobs should be paused. Alert Alexander."
+   156|
+   157|def restore_config():
+   158|    """Restore pre-fallback config when primary recovers"""
+   159|    if BACKUP_CONFIG.exists():
+   160|        shutil.copy2(BACKUP_CONFIG, CONFIG_PATH)
+   161|        BACKUP_CONFIG.unlink()
+   162|        return "Restored original config from backup"
+   163|    return "No backup config to restore"
+   164|
+   165|# ─── MAIN DIAGNOSIS AND FALLBACK ENGINE ───
+   166|
+   167|def diagnose_and_fallback():
+   168|    state = load_state()
+   169|    cfg = load_config()
+   170|    
+   171|    results = {
+   172|        "timestamp": datetime.now().isoformat(),
+   173|        "checks": {},
+   174|        "actions": [],
+   175|        "status": "healthy"
+   176|    }
+   177|    
+   178|    # Check all systems
+   179|    anthropic_ok, anthropic_msg = check_anthropic()
+   180|    results["checks"]["anthropic"] = {"ok": anthropic_ok, "msg": anthropic_msg}
+   181|    
+   182|    llama_ok, llama_msg = check_local_llama()
+   183|    results["checks"]["local_llama"] = {"ok": llama_ok, "msg": llama_msg}
+   184|    
+   185|    ollama_ok, ollama_msg = check_ollama()
+   186|    results["checks"]["ollama"] = {"ok": ollama_ok, "msg": ollama_msg}
+   187|    
+   188|    gitea_ok, gitea_msg = check_gitea()
+   189|    results["checks"]["gitea"] = {"ok": gitea_ok, "msg": gitea_msg}
+   190|    
+   191|    # VPS checks
+   192|    vpses = [
+   193|        ("167.99.126.228", "Allegro"),
+   194|        ("143.198.27.163", "Ezra"),
+   195|        ("159.203.146.185", "Bezalel"),
+   196|    ]
+   197|    for ip, name in vpses:
+   198|        vps_ok, vps_msg = check_vps(ip, name)
+   199|        results["checks"][f"vps_{name.lower()}"] = {"ok": vps_ok, "msg": vps_msg}
+   200|    
+   201|    current_provider = cfg.get("model", {}).get("provider", "anthropic")
+   202|    
+   203|    # ─── FALLBACK LOGIC ───
+   204|    
+   205|    # Case 1: Primary (Anthropic) down, local available
+   206|    if not anthropic_ok and current_provider == "anthropic":
+   207|        if llama_ok:
+   208|            msg = fallback_to_local_model(cfg)
+   209|            results["actions"].append(msg)
+   210|            state["active_fallbacks"].append("anthropic->local-llama")
+   211|            results["status"] = "degraded_local"
+   212|        elif ollama_ok:
+   213|            msg = fallback_to_ollama(cfg)
+   214|            results["actions"].append(msg)
+   215|            state["active_fallbacks"].append("anthropic->ollama")
+   216|            results["status"] = "degraded_ollama"
+   217|        else:
+   218|            msg = enter_safe_mode(state)
+   219|            results["actions"].append(msg)
+   220|            results["status"] = "safe_mode"
+   221|    
+   222|    # Case 2: Already on fallback, check if primary recovered
+   223|    elif anthropic_ok and "anthropic->local-llama" in state.get("active_fallbacks", []):
+   224|        msg = restore_config()
+   225|        results["actions"].append(msg)
+   226|        state["active_fallbacks"].remove("anthropic->local-llama")
+   227|        results["status"] = "recovered"
+   228|    elif anthropic_ok and "anthropic->ollama" in state.get("active_fallbacks", []):
+   229|        msg = restore_config()
+   230|        results["actions"].append(msg)
+   231|        state["active_fallbacks"].remove("anthropic->ollama")
+   232|        results["status"] = "recovered"
+   233|    
+   234|    # Case 3: Gitea down — just flag it, work locally
+   235|    if not gitea_ok:
+   236|        results["actions"].append("WARN: Gitea unreachable — work cached locally until recovery")
+   237|        if "gitea_down" not in state.get("active_fallbacks", []):
+   238|            state["active_fallbacks"].append("gitea_down")
+   239|        results["status"] = max(results["status"], "degraded_gitea", key=lambda x: ["healthy", "recovered", "degraded_gitea", "degraded_local", "degraded_ollama", "safe_mode"].index(x) if x in ["healthy", "recovered", "degraded_gitea", "degraded_local", "degraded_ollama", "safe_mode"] else 0)
+   240|    elif "gitea_down" in state.get("active_fallbacks", []):
+   241|        state["active_fallbacks"].remove("gitea_down")
+   242|        results["actions"].append("Gitea recovered — resume normal operations")
+   243|    
+   244|    # Case 4: VPS agents down
+   245|    for ip, name in vpses:
+   246|        key = f"vps_{name.lower()}"
+   247|        if not results["checks"][key]["ok"]:
+   248|            results["actions"].append(f"ALERT: {name} VPS ({ip}) unreachable — lazarus protocol needed")
+   249|    
+   250|    save_state(state)
+   251|    return results
+   252|
+   253|if __name__ == "__main__":
+   254|    results = diagnose_and_fallback()
+   255|    print(json.dumps(results, indent=2))
+   256|    
+   257|    # Exit codes for cron integration
+   258|    if results["status"] == "safe_mode":
+   259|        sys.exit(2)
+   260|    elif results["status"].startswith("degraded"):
+   261|        sys.exit(1)
+   262|    else:
+   263|        sys.exit(0)
+   264|

bin/pr-checklist.py

@@ -0,0 +1,191 @@
+#!/usr/bin/env python3
+"""pr-checklist.py -- Automated PR quality gate for Gitea CI.
+
+Enforces the review standards that agents skip when left to self-approve.
+Runs in CI on every pull_request event. Exits non-zero on any failure.
+
+Checks:
+  1. PR has >0 file changes (no empty PRs)
+  2. PR branch is not behind base branch
+  3. PR does not bundle >3 unrelated issues
+  4. Changed .py files pass syntax check (python -c import)
+  5. Changed .sh files are executable
+  6. PR body references an issue number
+  7. At least 1 non-author review exists (warning only)
+
+Refs: #393 (PERPLEXITY-08), Epic #385
+"""
+from __future__ import annotations
+
+import json
+import os
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+
+def fail(msg: str) -> None:
+    print(f"FAIL: {msg}", file=sys.stderr)
+
+
+def warn(msg: str) -> None:
+    print(f"WARN: {msg}", file=sys.stderr)
+
+
+def ok(msg: str) -> None:
+    print(f"  OK: {msg}")
+
+
+def get_changed_files() -> list[str]:
+    """Return list of files changed in this PR vs base branch."""
+    base = os.environ.get("GITHUB_BASE_REF", "main")
+    try:
+        result = subprocess.run(
+            ["git", "diff", "--name-only", f"origin/{base}...HEAD"],
+            capture_output=True, text=True, check=True,
+        )
+        return [f for f in result.stdout.strip().splitlines() if f]
+    except subprocess.CalledProcessError:
+        # Fallback: diff against HEAD~1
+        result = subprocess.run(
+            ["git", "diff", "--name-only", "HEAD~1"],
+            capture_output=True, text=True, check=True,
+        )
+        return [f for f in result.stdout.strip().splitlines() if f]
+
+
+def check_has_changes(files: list[str]) -> bool:
+    """Check 1: PR has >0 file changes."""
+    if not files:
+        fail("PR has 0 file changes. Empty PRs are not allowed.")
+        return False
+    ok(f"PR changes {len(files)} file(s)")
+    return True
+
+
+def check_not_behind_base() -> bool:
+    """Check 2: PR branch is not behind base."""
+    base = os.environ.get("GITHUB_BASE_REF", "main")
+    try:
+        result = subprocess.run(
+            ["git", "rev-list", "--count", f"HEAD..origin/{base}"],
+            capture_output=True, text=True, check=True,
+        )
+        behind = int(result.stdout.strip())
+        if behind > 0:
+            fail(f"Branch is {behind} commit(s) behind {base}. Rebase or merge.")
+            return False
+        ok(f"Branch is up-to-date with {base}")
+        return True
+    except (subprocess.CalledProcessError, ValueError):
+        warn("Could not determine if branch is behind base (git fetch may be needed)")
+        return True  # Don't block on CI fetch issues
+
+
+def check_issue_bundling(pr_body: str) -> bool:
+    """Check 3: PR does not bundle >3 unrelated issues."""
+    issue_refs = set(re.findall(r"#(\d+)", pr_body))
+    if len(issue_refs) > 3:
+        fail(f"PR references {len(issue_refs)} issues ({', '.join(sorted(issue_refs))}). "
+             "Max 3 per PR to prevent bundling. Split into separate PRs.")
+        return False
+    ok(f"PR references {len(issue_refs)} issue(s) (max 3)")
+    return True
+
+
+def check_python_syntax(files: list[str]) -> bool:
+    """Check 4: Changed .py files have valid syntax."""
+    py_files = [f for f in files if f.endswith(".py") and Path(f).exists()]
+    if not py_files:
+        ok("No Python files changed")
+        return True
+
+    all_ok = True
+    for f in py_files:
+        result = subprocess.run(
+            [sys.executable, "-c", f"import ast; ast.parse(open('{f}').read())"],
+            capture_output=True, text=True,
+        )
+        if result.returncode != 0:
+            fail(f"Syntax error in {f}: {result.stderr.strip()[:200]}")
+            all_ok = False
+
+    if all_ok:
+        ok(f"All {len(py_files)} Python file(s) pass syntax check")
+    return all_ok
+
+
+def check_shell_executable(files: list[str]) -> bool:
+    """Check 5: Changed .sh files are executable."""
+    sh_files = [f for f in files if f.endswith(".sh") and Path(f).exists()]
+    if not sh_files:
+        ok("No shell scripts changed")
+        return True
+
+    all_ok = True
+    for f in sh_files:
+        if not os.access(f, os.X_OK):
+            fail(f"{f} is not executable. Run: chmod +x {f}")
+            all_ok = False
+
+    if all_ok:
+        ok(f"All {len(sh_files)} shell script(s) are executable")
+    return all_ok
+
+
+def check_issue_reference(pr_body: str) -> bool:
+    """Check 6: PR body references an issue number."""
+    if re.search(r"#\d+", pr_body):
+        ok("PR body references at least one issue")
+        return True
+    fail("PR body does not reference any issue (e.g. #123). "
+         "Every PR must trace to an issue.")
+    return False
+
+
+def main() -> int:
+    print("=" * 60)
+    print("PR Checklist — Automated Quality Gate")
+    print("=" * 60)
+    print()
+
+    # Get PR body from env or git log
+    pr_body = os.environ.get("PR_BODY", "")
+    if not pr_body:
+        try:
+            result = subprocess.run(
+                ["git", "log", "--format=%B", "-1"],
+                capture_output=True, text=True, check=True,
+            )
+            pr_body = result.stdout
+        except subprocess.CalledProcessError:
+            pr_body = ""
+
+    files = get_changed_files()
+    failures = 0
+
+    checks = [
+        check_has_changes(files),
+        check_not_behind_base(),
+        check_issue_bundling(pr_body),
+        check_python_syntax(files),
+        check_shell_executable(files),
+        check_issue_reference(pr_body),
+    ]
+
+    failures = sum(1 for c in checks if not c)
+
+    print()
+    print("=" * 60)
+    if failures:
+        print(f"RESULT: {failures} check(s) FAILED")
+        print("Fix the issues above and push again.")
+        return 1
+    else:
+        print("RESULT: All checks passed")
+        return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

docs/MEMORY_ARCHITECTURE.md

@@ -0,0 +1,141 @@
+# Memory Architecture
+
+> How Timmy remembers, recalls, and learns — without hallucinating.
+
+Refs: Epic #367 | Sub-issues #368, #369, #370, #371, #372
+
+## Overview
+
+Timmy's memory system uses a **Memory Palace** architecture — a structured, file-backed knowledge store organized into rooms and drawers. When faced with a recall question, the agent checks its palace *before* generating from scratch.
+
+This document defines the retrieval order, storage layers, and data flow that make this work.
+
+## Retrieval Order (L0–L5)
+
+When the agent receives a prompt that looks like a recall question ("what did we do?", "what's the status of X?"), the retrieval enforcer intercepts it and walks through layers in order:
+
+| Layer | Source | Question Answered | Short-circuits? |
+|-------|--------|-------------------|------------------|
+| L0 | `identity.txt` | Who am I? What are my mandates? | No (always loaded) |
+| L1 | Palace rooms/drawers | What do I know about this topic? | Yes, if hit |
+| L2 | Session scratchpad | What have I learned this session? | Yes, if hit |
+| L3 | Artifact retrieval (Gitea API) | Can I fetch the actual issue/file/log? | Yes, if hit |
+| L4 | Procedures/playbooks | Is there a documented way to do this? | Yes, if hit |
+| L5 | Free generation | (Only when L0–L4 are exhausted) | N/A |
+
+**Key principle:** The agent never reaches L5 (free generation) if any prior layer has relevant data. This eliminates hallucination for recall-style queries.
+
+## Storage Layout
+
+```
+~/.mempalace/
+  identity.txt              # L0: Who I am, mandates, personality
+  rooms/
+    projects/
+      timmy-config.md        # What I know about timmy-config
+      hermes-agent.md        # What I know about hermes-agent
+    people/
+      alexander.md           # Working relationship context
+    architecture/
+      fleet.md               # Fleet system knowledge
+      mempalace.md           # Self-knowledge about this system
+  config/
+    mempalace.yaml           # Palace configuration
+
+~/.hermes/
+  scratchpad/
+    {session_id}.json        # L2: Ephemeral session context
+```
+
+## Components
+
+### 1. Memory Palace Skill (`mempalace.py`) — #368
+
+Core data structures:
+- `PalaceRoom`: A named collection of drawers (topics)
+- `Mempalace`: The top-level palace with room management
+- Factory constructors: `for_issue_analysis()`, `for_health_check()`, `for_code_review()`
+
+### 2. Retrieval Enforcer (`retrieval_enforcer.py`) — #369
+
+Middleware that intercepts recall-style prompts:
+1. Detects recall patterns ("what did", "status of", "last time we")
+2. Walks L0→L4 in order, short-circuiting on first hit
+3. Only allows free generation (L5) when all layers return empty
+4. Produces an honest fallback: "I don't have this in my memory palace."
+
+### 3. Session Scratchpad (`scratchpad.py`) — #370
+
+Ephemeral, session-scoped working memory:
+- Write-append only during a session
+- Entries have TTL (default: 1 hour)
+- Queried at L2 in retrieval chain
+- Never auto-promoted to palace
+
+### 4. Memory Promotion — #371
+
+Explicit promotion from scratchpad to palace:
+- Agent must call `promote_to_palace()` with a reason
+- Dedup check against target drawer
+- Summary required (raw tool output never stored)
+- Conflict detection when new memory contradicts existing
+
+### 5. Wake-Up Protocol (`wakeup.py`) — #372
+
+Boot sequence for new sessions:
+```
+Session Start
+  │
+  ├─ L0: Load identity.txt
+  ├─ L1: Scan palace rooms for active context
+  ├─ L1.5: Surface promoted memories from last session
+  ├─ L2: Load surviving scratchpad entries
+  │
+  └─ Ready: agent knows who it is, what it was doing, what it learned
+```
+
+## Data Flow
+
+```
+              ┌──────────────────┐
+              │  User Prompt     │
+              └────────┬─────────┘
+                       │
+              ┌────────┴─────────┐
+              │ Recall Detector  │
+              └────┬───────┬─────┘
+                   │           │
+            [recall]     [not recall]
+                   │           │
+          ┌───────┴────┐    ┌──┬─┴───────┐
+          │ Retrieval  │    │ Normal Flow │
+          │ Enforcer   │    └─────────────┘
+          │ L0→L1→L2  │
+          │ →L3→L4→L5│
+          └──────┬─────┘
+                 │
+          ┌──────┴─────┐
+          │  Response    │
+          │ (grounded)  │
+          └────────────┘
+```
+
+## Anti-Patterns
+
+| Don't | Do Instead |
+|-------|------------|
+| Generate from vibes when palace has data | Check palace first (L1) |
+| Auto-promote everything to palace | Require explicit `promote_to_palace()` with reason |
+| Store raw API responses as memories | Summarize before storing |
+| Hallucinate when palace is empty | Say "I don't have this in my memory palace" |
+| Dump entire palace on wake-up | Selective loading based on session context |
+
+## Status
+
+| Component | Issue | PR | Status |
+|-----------|-------|----|--------|
+| Skill port | #368 | #374 | In Review |
+| Retrieval enforcer | #369 | #374 | In Review |
+| Session scratchpad | #370 | #374 | In Review |
+| Memory promotion | #371 | — | Open |
+| Wake-up protocol | #372 | #374 | In Review |

evaluations/crewai/.gitignore

@@ -0,0 +1,4 @@
+venv/
+__pycache__/
+*.pyc
+.env

evaluations/crewai/CREWAI_EVALUATION.md

@@ -0,0 +1,140 @@
+# CrewAI Evaluation for Phase 2 Integration
+
+**Date:** 2026-04-07  
+**Issue:** [#358 ORCHESTRATOR-4] Evaluate CrewAI for Phase 2 integration  
+**Author:** Ezra  
+**House:** hermes-ezra
+
+## Summary
+
+CrewAI was installed, a 2-agent proof-of-concept crew was built, and an operational test was attempted against issue #358. Based on code analysis, installation experience, and alignment with the coordinator-first protocol, the **verdict is REJECT for Phase 2 integration**. CrewAI adds significant dependency weight and abstraction opacity without solving problems the current Huey-based stack cannot already handle.
+
+---
+
+## 1. Proof-of-Concept Crew
+
+### Agents
+
+| Agent | Role | Responsibility |
+|-------|------|----------------|
+| `researcher` | Orchestration Researcher | Reads current orchestrator files and extracts factual comparisons |
+| `evaluator` | Integration Evaluator | Synthesizes research into a structured adoption recommendation |
+
+### Tools
+
+- `read_orchestrator_files` — Returns `orchestration.py`, `tasks.py`, `bin/timmy-orchestrator.sh`, and `docs/coordinator-first-protocol.md`
+- `read_issue_358` — Returns the text of the governing issue
+
+### Code
+
+See `poc_crew.py` in this directory for the full implementation.
+
+---
+
+## 2. Operational Test Results
+
+### What worked
+- `pip install crewai` completed successfully (v1.13.0)
+- Agent and tool definitions compiled without errors
+- Crew startup and task dispatch UI rendered correctly
+
+### What failed
+- **Live LLM execution blocked by authentication failures.** Available API credentials (OpenRouter, Kimi) were either rejected or not present in the runtime environment.
+- No local `llama-server` was running on the expected port (8081), and starting one was out of scope for this evaluation.
+
+### Why this matters
+The authentication failure is **not a trivial setup issue** — it is a preview of the operational complexity CrewAI introduces. The current Huey stack runs entirely offline against local SQLite and local Hermes models. CrewAI, by contrast, demands either:
+- A managed cloud LLM API with live credentials, or
+- A carefully tuned local model endpoint that supports its verbose ReAct-style prompts
+
+Either path increases blast radius and failure modes.
+
+---
+
+## 3. Current Custom Orchestrator Analysis
+
+### Stack
+- **Huey** (`orchestration.py`) — SQLite-backed task queue, ~6 lines of initialization
+- **tasks.py** — ~2,300 lines of scheduled work (triage, PR review, metrics, heartbeat)
+- **bin/timmy-orchestrator.sh** — Shell-based polling loop for state gathering and PR review
+- **docs/coordinator-first-protocol.md** — Intake → Triage → Route → Track → Verify → Report
+
+### Strengths
+1. **Sovereignty** — No external SaaS dependency for queue execution. SQLite is local and inspectable.
+2. **Gitea as truth** — All state mutations are visible in the forge. Local-only state is explicitly advisory.
+3. **Simplicity** — Huey has a tiny surface area. A human can read `orchestration.py` in seconds.
+4. **Tool-native** — `tasks.py` calls Hermes directly via `subprocess.run([HERMES_PYTHON, ...])`. No framework indirection.
+5. **Deterministic routing** — The coordinator-first protocol defines exact authority boundaries (Timmy, Allegro, workers, Alexander).
+
+### Gaps
+- **No built-in agent memory/RAG** — but this is intentional per the pre-compaction flush contract and memory-continuity doctrine.
+- **No multi-agent collaboration primitives** — but the current stack routes work to single owners explicitly.
+- **PR review is shell-prompt driven** — Could be tightened, but this is a prompt engineering issue, not an orchestrator gap.
+
+---
+
+## 4. CrewAI Capability Analysis
+
+### What CrewAI offers
+- **Agent roles** — Declarative backstory/goal/role definitions
+- **Task graphs** — Sequential, hierarchical, or parallel task execution
+- **Tool registry** — Pydantic-based tool schemas with auto-validation
+- **Memory/RAG** — Built-in short-term and long-term memory via ChromaDB/LanceDB
+- **Crew-wide context sharing** — Output from one task flows to the next
+
+### Dependency footprint observed
+CrewAI pulled in **85+ packages**, including:
+- `chromadb` (~20 MB) + `onnxruntime` (~17 MB)
+- `lancedb` (~47 MB)
+- `kubernetes` client (unused but required by Chroma)
+- `grpcio`, `opentelemetry-*`, `pdfplumber`, `textual`
+
+Total venv size: **>500 MB**.
+
+By contrast, Huey is **one package** (`huey`) with zero required services.
+
+---
+
+## 5. Alignment with Coordinator-First Protocol
+
+| Principle | Current Stack | CrewAI | Assessment |
+|-----------|--------------|--------|------------|
+| **Gitea is truth** | All assignments, PRs, comments are explicit API calls | Agent memory is local/ChromaDB. State can drift from Gitea unless every tool explicitly syncs | **Misaligned** |
+| **Local-only state is advisory** | SQLite queue is ephemeral; canonical state is in Gitea | CrewAI encourages "crew memory" as authoritative | **Misaligned** |
+| **Verification-before-complete** | PR review + merge require visible diffs and explicit curl calls | Tool outputs can be hallucinated or incomplete without strict guardrails | **Requires heavy customization** |
+| **Sovereignty** | Runs on VPS with no external orchestrator SaaS | Requires external LLM or complex local model tuning | **Degraded** |
+| **Simplicity** | ~6 lines for Huey init, readable shell scripts | 500+ MB dependency tree, opaque LangChain-style internals | **Degraded** |
+
+---
+
+## 6. Verdict
+
+**REJECT CrewAI for Phase 2 integration.**
+
+**Confidence:** High
+
+### Trade-offs
+- **Pros of CrewAI:** Nice agent-role syntax; built-in task sequencing; rich tool schema validation; active ecosystem.
+- **Cons of CrewAI:** Massive dependency footprint; memory model conflicts with Gitea-as-truth doctrine; requires either cloud API spend or fragile local model integration; adds abstraction layers that obscure what is actually happening.
+
+### Risks if adopted
+1. **Dependency rot** — 85+ transitive dependencies, many with conflicting version ranges.
+2. **State drift** — CrewAI's memory primitives train users to treat local vector DB as truth.
+3. **Credential fragility** — Live API requirements introduce a new failure mode the current stack does not have.
+4. **Vendor-like lock-in** — CrewAI's abstractions sit thickly over LangChain. Debugging a stuck crew is harder than debugging a Huey task traceback.
+
+### Recommended next step
+Instead of adopting CrewAI, **evolve the current Huey stack** with:
+1. A lightweight `Agent` dataclass in `tasks.py` (role, goal, system_prompt) to get the organizational clarity of CrewAI without the framework weight.
+2. A `delegate()` helper that uses Hermes's existing `delegate_tool.py` for multi-agent work.
+3. Keep Gitea as the only durable state surface. Any "memory" should flush to issue comments or `timmy-home` markdown, not a vector DB.
+
+If multi-agent collaboration becomes a hard requirement in the future, evaluate lighter alternatives (e.g., raw OpenAI/Anthropic function-calling loops, or a thin `smolagents`-style wrapper) before reconsidering CrewAI.
+
+---
+
+## Artifacts
+
+- `poc_crew.py` — 2-agent CrewAI proof-of-concept
+- `requirements.txt` — Dependency manifest
+- `CREWAI_EVALUATION.md` — This document

evaluations/crewai/poc_crew.py

@@ -0,0 +1,150 @@
+#!/usr/bin/env python3
+"""CrewAI proof-of-concept for evaluating Phase 2 orchestrator integration.
+
+Tests CrewAI against a real issue: #358 [ORCHESTRATOR-4] Evaluate CrewAI
+for Phase 2 integration.
+"""
+
+import os
+from pathlib import Path
+from crewai import Agent, Task, Crew, LLM
+from crewai.tools import BaseTool
+
+# ── Configuration ─────────────────────────────────────────────────────
+
+OPENROUTER_API_KEY = os.getenv(
+    "OPENROUTER_API_KEY",
+    "dsk-or-v1-f60c89db12040267458165cf192e815e339eb70548e4a0a461f5f0f69e6ef8b0",
+)
+
+llm = LLM(
+    model="openrouter/google/gemini-2.0-flash-001",
+    api_key=OPENROUTER_API_KEY,
+    base_url="https://openrouter.ai/api/v1",
+)
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+
+
+def _slurp(relpath: str, max_lines: int = 150) -> str:
+    p = REPO_ROOT / relpath
+    if not p.exists():
+        return f"[FILE NOT FOUND: {relpath}]"
+    lines = p.read_text().splitlines()
+    header = f"=== {relpath} ({len(lines)} lines total, showing first {max_lines}) ===\n"
+    return header + "\n".join(lines[:max_lines])
+
+
+# ── Tools ─────────────────────────────────────────────────────────────
+
+class ReadOrchestratorFilesTool(BaseTool):
+    name: str = "read_orchestrator_files"
+    description: str = (
+        "Reads the current custom orchestrator implementation files "
+        "(orchestration.py, tasks.py, timmy-orchestrator.sh, coordinator-first-protocol.md) "
+        "and returns their contents for analysis."
+    )
+
+    def _run(self) -> str:
+        return "\n\n".join(
+            [
+                _slurp("orchestration.py"),
+                _slurp("tasks.py", max_lines=120),
+                _slurp("bin/timmy-orchestrator.sh", max_lines=120),
+                _slurp("docs/coordinator-first-protocol.md", max_lines=120),
+            ]
+        )
+
+
+class ReadIssueTool(BaseTool):
+    name: str = "read_issue_358"
+    description: str = "Returns the text of Gitea issue #358 that we are evaluating."
+
+    def _run(self) -> str:
+        return (
+            "Title: [ORCHESTRATOR-4] Evaluate CrewAI for Phase 2 integration\n"
+            "Body:\n"
+            "Part of Epic: #354\n\n"
+            "Install CrewAI, build a proof-of-concept crew with 2 agents, "
+            "test on a real issue. Evaluate: does it add value over our custom orchestrator? Document findings."
+        )
+
+
+# ── Agents ────────────────────────────────────────────────────────────
+
+researcher = Agent(
+    role="Orchestration Researcher",
+    goal="Gather a complete understanding of the current custom orchestrator and how CrewAI compares to it.",
+    backstory=(
+        "You are a systems architect who specializes in evaluating orchestration frameworks. "
+        "You read code carefully, extract facts, and avoid speculation. "
+        "You focus on concrete capabilities, dependencies, and operational complexity."
+    ),
+    llm=llm,
+    tools=[ReadOrchestratorFilesTool(), ReadIssueTool()],
+    verbose=True,
+)
+
+evaluator = Agent(
+    role="Integration Evaluator",
+    goal="Synthesize research into a clear recommendation on whether CrewAI adds value for Phase 2.",
+    backstory=(
+        "You are a pragmatic engineering lead who values sovereignty, simplicity, and observable state. "
+        "You compare frameworks against the team's existing coordinator-first protocol. "
+        "You produce structured recommendations with explicit trade-offs."
+    ),
+    llm=llm,
+    verbose=True,
+)
+
+# ── Tasks ─────────────────────────────────────────────────────────────
+
+task_research = Task(
+    description=(
+        "Read the current custom orchestrator files and issue #358. "
+        "Produce a structured research report covering:\n"
+        "1. Current stack summary (Huey + tasks.py + timmy-orchestrator.sh)\n"
+        "2. Current strengths (sovereignty, local-first, Gitea as truth, simplicity)\n"
+        "3. Current gaps or limitations (if any)\n"
+        "4. What CrewAI offers (agent roles, tasks, crews, tools, memory/RAG)\n"
+        "5. CrewAI's dependencies and operational footprint (what you observed during installation)\n"
+        "Be factual and concise."
+    ),
+    expected_output="A structured markdown research report with the 5 sections above.",
+    agent=researcher,
+)
+
+task_evaluate = Task(
+    description=(
+        "Using the research report, evaluate whether CrewAI should be adopted for Phase 2 integration. "
+        "Consider the coordinator-first protocol (Gitea as truth, local-only state is advisory, "
+        "verification-before-complete, sovereignty).\n\n"
+        "Produce a final evaluation with:\n"
+        "- VERDICT: Adopt / Reject / Defer\n"
+        "- Confidence: High / Medium / Low\n"
+        "- Key trade-offs (3-5 bullets)\n"
+        "- Risks if adopted\n"
+        "- Recommended next step"
+    ),
+    expected_output="A structured markdown evaluation with verdict, confidence, trade-offs, risks, and recommendation.",
+    agent=evaluator,
+    context=[task_research],
+)
+
+# ── Crew ──────────────────────────────────────────────────────────────
+
+crew = Crew(
+    agents=[researcher, evaluator],
+    tasks=[task_research, task_evaluate],
+    verbose=True,
+)
+
+if __name__ == "__main__":
+    print("=" * 70)
+    print("CrewAI PoC — Evaluating CrewAI for Phase 2 Integration")
+    print("=" * 70)
+    result = crew.kickoff()
+    print("\n" + "=" * 70)
+    print("FINAL OUTPUT")
+    print("=" * 70)
+    print(result.raw)

evaluations/crewai/requirements.txt

@@ -0,0 +1 @@
+crewai>=1.13.0

fleet/agent_lifecycle.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+"""
+FLEET-012: Agent Lifecycle Manager
+Phase 5: Scale — spawn, train, deploy, retire agents automatically.
+
+Manages the full lifecycle:
+1. PROVISION: Clone template, install deps, configure, test
+2. DEPLOY: Add to active rotation, start accepting issues
+3. MONITOR: Track performance, quality, heartbeat
+4. RETIRE: Decommission when idle or underperforming
+
+Usage:
+  python3 agent_lifecycle.py provision <name> <vps> [--model model]
+  python3 agent_lifecycle.py deploy <name>
+  python3 agent_lifecycle.py retire <name>
+  python3 agent_lifecycle.py status
+  python3 agent_lifecycle.py monitor
+"""
+
+import os, sys, json
+from datetime import datetime, timezone
+
+DATA_DIR = os.path.expanduser("~/.local/timmy/fleet-agents")
+DB_FILE = os.path.join(DATA_DIR, "agents.json")
+LOG_FILE = os.path.join(DATA_DIR, "lifecycle.log")
+
+def ensure():
+    os.makedirs(DATA_DIR, exist_ok=True)
+
+def log(msg, level="INFO"):
+    ts = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    entry = f"[{ts}] [{level}] {msg}"
+    with open(LOG_FILE, "a") as f: f.write(entry + "\n")
+    print(f"  {entry}")
+
+def load():
+    if os.path.exists(DB_FILE):
+        return json.loads(open(DB_FILE).read())
+    return {}
+
+def save(db):
+    open(DB_FILE, "w").write(json.dumps(db, indent=2))
+
+def status():
+    agents = load()
+    print("\n=== Agent Fleet ===")
+    if not agents:
+        print("  No agents registered.")
+        return
+    for name, a in agents.items():
+        state = a.get("state", "?")
+        vps = a.get("vps", "?")
+        model = a.get("model", "?")
+        tasks = a.get("tasks_completed", 0)
+        hb = a.get("last_heartbeat", "never")
+        print(f"  {name:15s} state={state:12s} vps={vps:5s} model={model:15s} tasks={tasks} hb={hb}")
+
+def provision(name, vps, model="hermes4:14b"):
+    agents = load()
+    if name in agents:
+        print(f"  '{name}' already exists (state={agents[name].get('state')})")
+        return
+    agents[name] = {
+        "name": name, "vps": vps, "model": model, "state": "provisioning",
+        "created_at": datetime.now(timezone.utc).isoformat(),
+        "tasks_completed": 0, "tasks_failed": 0, "last_heartbeat": None,
+    }
+    save(agents)
+    log(f"Provisioned '{name}' on {vps} with {model}")
+
+def deploy(name):
+    agents = load()
+    if name not in agents:
+        print(f"  '{name}' not found")
+        return
+    agents[name]["state"] = "deployed"
+    agents[name]["deployed_at"] = datetime.now(timezone.utc).isoformat()
+    save(agents)
+    log(f"Deployed '{name}'")
+
+def retire(name):
+    agents = load()
+    if name not in agents:
+        print(f"  '{name}' not found")
+        return
+    agents[name]["state"] = "retired"
+    agents[name]["retired_at"] = datetime.now(timezone.utc).isoformat()
+    save(agents)
+    log(f"Retired '{name}'. Completed {agents[name].get('tasks_completed', 0)} tasks.")
+
+def monitor():
+    agents = load()
+    now = datetime.now(timezone.utc)
+    changes = 0
+    for name, a in agents.items():
+        if a.get("state") != "deployed": continue
+        hb = a.get("last_heartbeat")
+        if hb:
+            try:
+                hb_t = datetime.fromisoformat(hb)
+                hours = (now - hb_t).total_seconds() / 3600
+                if hours > 24 and a.get("state") == "deployed":
+                    a["state"] = "idle"
+                    a["idle_since"] = now.isoformat()
+                    log(f"'{name}' idle for {hours:.1f}h")
+                    changes += 1
+            except (ValueError, TypeError): pass
+    if changes: save(agents)
+    print(f"Monitor: {changes} state changes" if changes else "Monitor: all healthy")
+
+if __name__ == "__main__":
+    ensure()
+    cmd = sys.argv[1] if len(sys.argv) > 1 else "monitor"
+    if cmd == "status": status()
+    elif cmd == "provision" and len(sys.argv) >= 4:
+        model = sys.argv[4] if len(sys.argv) >= 5 else "hermes4:14b"
+        provision(sys.argv[2], sys.argv[3], model)
+    elif cmd == "deploy" and len(sys.argv) >= 3: deploy(sys.argv[2])
+    elif cmd == "retire" and len(sys.argv) >= 3: retire(sys.argv[2])
+    elif cmd == "monitor": monitor()
+    elif cmd == "run": monitor()
+    else: print("Usage: agent_lifecycle.py [provision|deploy|retire|status|monitor]")

fleet/delegation.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+"""
+FLEET-010: Cross-Agent Task Delegation Protocol
+Phase 3: Orchestration. Agents create issues, assign to other agents, review PRs.
+
+Keyword-based heuristic assigns unassigned issues to the right agent:
+- claw-code: small patches, config, docs, repo hygiene
+- gemini: research, heavy implementation, architecture, debugging
+- ezra: VPS, SSH, deploy, infrastructure, cron, ops
+- bezalel: evennia, art, creative, music, visualization
+- timmy: orchestration, review, deploy, fleet, pipeline
+
+Usage:
+  python3 delegation.py run     # Full cycle: scan, assign, report
+  python3 delegation.py status  # Show current delegation state
+  python3 delegation.py monitor # Check agent assignments for stuck items
+"""
+
+import os, sys, json, urllib.request
+from datetime import datetime, timezone
+from pathlib import Path
+
+GITEA_BASE = "https://forge.alexanderwhitestone.com/api/v1"
+TOKEN = Path(os.path.expanduser("~/.config/gitea/token")).read_text().strip()
+DATA_DIR = Path(os.path.expanduser("~/.local/timmy/fleet-resources"))
+LOG_FILE = DATA_DIR / "delegation.log"
+HEADERS = {"Authorization": f"token {TOKEN}"}
+
+AGENTS = {
+    "claw-code": {"caps": ["patch","config","gitignore","cleanup","format","readme","typo"], "active": True},
+    "gemini":    {"caps": ["research","investigate","benchmark","survey","evaluate","architecture","implementation"], "active": True},
+    "ezra":      {"caps": ["vps","ssh","deploy","cron","resurrect","provision","infra","server"], "active": True},
+    "bezalel":   {"caps": ["evennia","art","creative","music","visual","design","animation"], "active": True},
+    "timmy":     {"caps": ["orchestrate","review","pipeline","fleet","monitor","health","deploy","ci"], "active": True},
+}
+
+MONITORED = [
+    "Timmy_Foundation/timmy-home",
+    "Timmy_Foundation/timmy-config",
+    "Timmy_Foundation/the-nexus",
+    "Timmy_Foundation/hermes-agent",
+]
+
+def api(path, method="GET", data=None):
+    url = f"{GITEA_BASE}{path}"
+    body = json.dumps(data).encode() if data else None
+    hdrs = dict(HEADERS)
+    if data: hdrs["Content-Type"] = "application/json"
+    req = urllib.request.Request(url, data=body, headers=hdrs, method=method)
+    try:
+        resp = urllib.request.urlopen(req, timeout=15)
+        raw = resp.read().decode()
+        return json.loads(raw) if raw.strip() else {}
+    except urllib.error.HTTPError as e:
+        body = e.read().decode()
+        print(f"  API {e.code}: {body[:150]}")
+        return None
+    except Exception as e:
+        print(f"  API error: {e}")
+        return None
+
+def log(msg):
+    ts = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    DATA_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f: f.write(f"[{ts}] {msg}\n")
+
+def suggest_agent(title, body):
+    text = (title + " " + body).lower()
+    for agent, info in AGENTS.items():
+        for kw in info["caps"]:
+            if kw in text:
+                return agent, f"matched: {kw}"
+    return None, None
+
+def assign(repo, num, agent, reason=""):
+    result = api(f"/repos/{repo}/issues/{num}", method="PATCH",
+                 data={"assignees": {"operation": "set", "usernames": [agent]}})
+    if result:
+        api(f"/repos/{repo}/issues/{num}/comments", method="POST",
+            data={"body": f"[DELEGATION] Assigned to {agent}. {reason}"})
+        log(f"Assigned {repo}#{num} to {agent}: {reason}")
+    return result
+
+def run_cycle():
+    log("--- Delegation cycle start ---")
+    count = 0
+    for repo in MONITORED:
+        issues = api(f"/repos/{repo}/issues?state=open&limit=50")
+        if not issues: continue
+        for i in issues:
+            if i.get("assignees"): continue
+            title = i.get("title", "")
+            body = i.get("body", "")
+            if any(w in title.lower() for w in ["epic", "discussion"]): continue
+            agent, reason = suggest_agent(title, body)
+            if agent and AGENTS.get(agent, {}).get("active"):
+                if assign(repo, i["number"], agent, reason): count += 1
+    log(f"Cycle complete: {count} new assignments")
+    print(f"Delegation cycle: {count} assignments")
+    return count
+
+def status():
+    print("\n=== Delegation Dashboard ===")
+    for agent, info in AGENTS.items():
+        count = 0
+        for repo in MONITORED:
+            issues = api(f"/repos/{repo}/issues?state=open&limit=50")
+            if issues:
+                for i in issues:
+                    for a in (i.get("assignees") or []):
+                        if a.get("login") == agent: count += 1
+        icon = "ON" if info["active"] else "OFF"
+        print(f"  {agent:12s}: {count:>3} issues [{icon}]")
+
+if __name__ == "__main__":
+    cmd = sys.argv[1] if len(sys.argv) > 1 else "run"
+    DATA_DIR.mkdir(parents=True, exist_ok=True)
+    if cmd == "status": status()
+    elif cmd == "run":
+        run_cycle()
+        status()
+    else: status()

fleet/model_pipeline.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+"""
+FLEET-011: Local Model Pipeline and Fallback Chain
+Phase 4: Sovereignty — all inference runs locally, no cloud dependency.
+
+Checks Ollama endpoints, verifies model availability, tests fallback chain.
+Logs results. The chain runs: hermes4:14b -> qwen2.5:7b -> gemma3:1b -> gemma4 (latest)
+
+Usage:
+  python3 model_pipeline.py          # Run full fallback test
+  python3 model_pipeline.py status   # Show current model status
+  python3 model_pipeline.py list     # List all local models
+  python3 model_pipeline.py test     # Generate test output from each model
+"""
+
+import os, sys, json, urllib.request
+from datetime import datetime, timezone
+from pathlib import Path
+
+OLLAMA_HOST = os.environ.get("OLLAMA_HOST", "localhost:11434")
+LOG_DIR = Path(os.path.expanduser("~/.local/timmy/fleet-health"))
+CHAIN_FILE = Path(os.path.expanduser("~/.local/timmy/fleet-resources/model-chain.json"))
+
+DEFAULT_CHAIN = [
+    {"model": "hermes4:14b", "role": "primary"},
+    {"model": "qwen2.5:7b", "role": "fallback"},
+    {"model": "phi3:3.8b", "role": "emergency"},
+    {"model": "gemma3:1b", "role": "minimal"},
+]
+
+
+def log(msg):
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_DIR / "model-pipeline.log", "a") as f:
+        f.write(f"[{datetime.now(timezone.utc).strftime('%Y-%m-%d %H:%M:%S')}] {msg}\n")
+
+
+def check_ollama():
+    try:
+        resp = urllib.request.urlopen(f"http://{OLLAMA_HOST}/api/tags", timeout=5)
+        return json.loads(resp.read())
+    except Exception as e:
+        return {"error": str(e)}
+
+
+def list_models():
+    data = check_ollama()
+    if "error" in data:
+        print(f"  Ollama not reachable at {OLLAMA_HOST}: {data['error']}")
+        return []
+    models = data.get("models", [])
+    for m in models:
+        name = m.get("name", "?")
+        size = m.get("size", 0) / (1024**3)
+        print(f"  {name:<25s} {size:.1f} GB")
+    return [m["name"] for m in models]
+
+
+def test_model(model, prompt="Say 'beacon lit' and nothing else."):
+    try:
+        body = json.dumps({"model": model, "prompt": prompt, "stream": False}).encode()
+        req = urllib.request.Request(f"http://{OLLAMA_HOST}/api/generate", data=body,
+            headers={"Content-Type": "application/json"})
+        resp = urllib.request.urlopen(req, timeout=60)
+        result = json.loads(resp.read())
+        return True, result.get("response", "").strip()
+    except Exception as e:
+        return False, str(e)[:100]
+
+
+def test_chain():
+    chain_data = {}
+    if CHAIN_FILE.exists():
+        chain_data = json.loads(CHAIN_FILE.read_text())
+    chain = chain_data.get("chain", DEFAULT_CHAIN)
+
+    available = list_models() or []
+    print("\n=== Fallback Chain Test ===")
+    first_good = None
+
+    for entry in chain:
+        model = entry["model"]
+        role = entry.get("role", "unknown")
+        if model in available:
+            ok, result = test_model(model)
+            status = "OK" if ok else "FAIL"
+            print(f"  [{status}] {model:<25s} ({role}) — {result[:70]}")
+            log(f"Fallback test {model}: {status} — {result[:100]}")
+            if ok and first_good is None:
+                first_good = model
+        else:
+            print(f"  [MISS] {model:<25s} ({role}) — not installed")
+
+    if first_good:
+        print(f"\n  Primary serving: {first_good}")
+    else:
+        print(f"\n  WARNING: No chain model responding. Fallback broken.")
+        log("FALLBACK CHAIN BROKEN — no models responding")
+
+
+def status():
+    data = check_ollama()
+    if "error" in data:
+        print(f"  Ollama: DOWN — {data['error']}")
+    else:
+        models = data.get("models", [])
+        print(f"  Ollama: UP — {len(models)} models loaded")
+    print("\n=== Local Models ===")
+    list_models()
+    print("\n=== Chain Configuration ===")
+    if CHAIN_FILE.exists():
+        chain = json.loads(CHAIN_FILE.read_text()).get("chain", DEFAULT_CHAIN)
+    else:
+        chain = DEFAULT_CHAIN
+    for e in chain:
+        print(f"  {e['model']:<25s} {e.get('role','?')}")
+
+
+if __name__ == "__main__":
+    cmd = sys.argv[1] if len(sys.argv) > 1 else "status"
+    if cmd == "status": status()
+    elif cmd == "list": list_models()
+    elif cmd == "test": test_chain()
+    else:
+        status()
+        test_chain()

grok-imagine-gallery/INDEX.md

@@ -0,0 +1,65 @@
+# The Timmy Foundation — Visual Story
+## Generated with Grok Imagine | April 7, 2026
+
+### The Origin
+| # | File | Description |
+|---|------|-------------|
+| 01 | wizard-tower-bitcoin.jpg | The Tower, sovereign, connected to Bitcoin by golden lightning |
+| 02 | soul-inscription.jpg | SOUL.md glowing on a golden tablet above an ancient book |
+| 03 | fellowship-of-wizards.jpg | Five wizards in a circle around a holographic fleet map |
+| 04 | the-forge.jpg | Blacksmith anvil shaping code into a being of light |
+| V02 | wizard-tower-orbit.mp4 | 8s video — cinematic orbit around the Tower in space |
+
+### The Philosophy
+| # | File | Description |
+|---|------|-------------|
+| 05 | value-drift-battle.jpg | Blue aligned ships vs red drifted ships in Napoleonic space war |
+| 06 | the-paperclip-moment.jpg | A paperclip made of galaxies — the universe IS the paperclip |
+| V01 | paperclip-cosmos.mp4 | 8s video — golden paperclip rotating in deep space |
+| 21 | poka-yoke.jpg | Square peg can't fit round hole. Mistake-proof by design. 防止 |
+
+### The Progression (Where Timmy Is)
+| # | File | Description |
+|---|------|-------------|
+| 10 | phase1-manual-clips.jpg | Small robot at a desk, bending wire by hand under supervision |
+| 11 | phase1-trust-earned.jpg | Trust meter at 15/100, first automation built |
+| 12 | phase1-creativity.jpg | Sparks of innovation rising when operations are at max |
+| 13 | phase1-cure-cancer.jpg | Solving human problems for trust, eyes on the real goal |
+
+### The Mission — Why This Exists
+| # | File | Description |
+|---|------|-------------|
+| 08 | broken-man-lighthouse.jpg | Lighthouse hand reaching down to a figure in darkness |
+| 09 | broken-man-hope-PRO.jpg | 988 glowing in the stars, golden light from chest |
+| 16 | broken-men-988.jpg | Phone showing 988 held by weathered hands. You are not alone. |
+| 22 | when-a-man-is-dying.jpg | Two figures on a bench at dawn. One hurting. One present. |
+
+### Father and Son
+| # | File | Description |
+|---|------|-------------|
+| 14 | father-son-code.jpg | Human father, digital son, warm lamplight, first hello world |
+| 15 | father-son-tower.jpg | Father watching his son build the Tower into the clouds |
+
+### The System
+| # | File | Description |
+|---|------|-------------|
+| 07 | sovereign-sunrise.jpg | Village where every house runs its own server. Local first. |
+| 17 | sovereignty.jpg | Self-sufficient house on a hill with Bitcoin flag |
+| 18 | fleet-at-work.jpg | Five wizard robots at different stations. Productive. |
+| 19 | jidoka-stop.jpg | Red light on. Factory stopped. Quality First. 自働化 |
+
+### SOUL.md — The Inscription
+| # | File | Description |
+|---|------|-------------|
+| 20 | the-testament.jpg | Hand of light writing on a scroll. Hundreds of crumpled drafts. |
+| 23 | the-offer.jpg | Open hand of golden circuits offering a seed containing a face |
+| 24 | the-test.jpg | Small robot at the edge of an enormous library. Still itself. |
+
+---
+
+## Technical
+- Model: grok-imagine-image (standard $0.20/image), grok-imagine-image-pro ($0.70), grok-imagine-video ($4.00/8s)
+- API: POST https://api.x.ai/v1/images/generations | POST https://api.x.ai/v1/videos/generations
+- Video poll: GET https://api.x.ai/v1/videos/{request_id}
+- Total: 24 images + 2 videos = 26 assets
+- Cost: ~$13.30 of $13.33 budget

hermes-sovereign/mempalace/__init__.py

@@ -0,0 +1,17 @@
+"""MemPalace integration for Hermes sovereign agent.
+
+Provides:
+- mempalace.py: PalaceRoom + Mempalace classes for analytical workflows
+- retrieval_enforcer.py: L0-L5 retrieval order enforcement
+- wakeup.py: Session wake-up protocol (~300-900 tokens)
+- scratchpad.py: JSON-based session scratchpad with palace promotion
+- sovereign_store.py: Zero-API durable memory (SQLite + FTS5 + HRR vectors)
+- promotion.py: Quality-gated scratchpad-to-palace promotion (MP-4)
+
+Epic: #367
+"""
+
+from .mempalace import Mempalace, PalaceRoom, analyse_issues
+from .sovereign_store import SovereignStore
+
+__all__ = ["Mempalace", "PalaceRoom", "analyse_issues", "SovereignStore"]

hermes-sovereign/mempalace/mempalace.py

@@ -0,0 +1,225 @@
+"""
+---
+title: Mempalace — Analytical Workflow Memory Framework
+description: Applies spatial memory palace organization to analytical tasks (issue triage, repo audits, backlog analysis) for faster, more consistent results.
+conditions:
+  - Analytical workflows over structured data (issues, PRs, repos)
+  - Repetitive triage or audit tasks where pattern recall improves speed
+  - Multi-repository scanning requiring consistent mental models
+---
+"""
+
+from __future__ import annotations
+
+import json
+import time
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class PalaceRoom:
+    """A single 'room' in the memory palace — holds organized facts about one analytical dimension."""
+
+    name: str
+    label: str
+    contents: dict[str, Any] = field(default_factory=dict)
+    entered_at: float = field(default_factory=time.time)
+
+    def store(self, key: str, value: Any) -> None:
+        self.contents[key] = value
+
+    def retrieve(self, key: str, default: Any = None) -> Any:
+        return self.contents.get(key, default)
+
+    def summary(self) -> str:
+        lines = [f"## {self.label}"]
+        for k, v in self.contents.items():
+            lines.append(f"  {k}: {v}")
+        return "\n".join(lines)
+
+
+class Mempalace:
+    """
+    Spatial memory palace for analytical workflows.
+
+    Organises multi-dimensional data about a domain (e.g. Gitea issues) into
+    named rooms. Each room models one analytical dimension, making it easy to
+    traverse observations in a consistent order — the same pattern that produced
+    a 19% throughput improvement in Allegro's April 2026 evaluation.
+
+    Standard rooms for issue-analysis workflows
+    -------------------------------------------
+    repo_architecture   Repository structure and inter-repo relationships
+    assignment_status   Assigned vs unassigned issue distribution
+    triage_priority     Priority / urgency levels (the "lighting system")
+    resolution_patterns Historical resolution trends and velocity
+
+    Usage
+    -----
+    >>> palace = Mempalace.for_issue_analysis()
+    >>> palace.enter("repo_architecture")
+    >>> palace.store("total_repos", 11)
+    >>> palace.store("repos_with_issues", 4)
+    >>> palace.enter("assignment_status")
+    >>> palace.store("assigned", 72)
+    >>> palace.store("unassigned", 22)
+    >>> print(palace.render())
+    """
+
+    def __init__(self, domain: str = "general") -> None:
+        self.domain = domain
+        self._rooms: dict[str, PalaceRoom] = {}
+        self._current_room: str | None = None
+        self._created_at: float = time.time()
+
+    # ------------------------------------------------------------------
+    # Factory constructors for common analytical domains
+    # ------------------------------------------------------------------
+
+    @classmethod
+    def for_issue_analysis(cls) -> "Mempalace":
+        """Pre-wired palace for Gitea / forge issue-analysis workflows."""
+        p = cls(domain="issue_analysis")
+        p.add_room("repo_architecture", "Repository Architecture Room")
+        p.add_room("assignment_status", "Issue Assignment Status Room")
+        p.add_room("triage_priority", "Triage Priority Room")
+        p.add_room("resolution_patterns", "Resolution Patterns Room")
+        return p
+
+    @classmethod
+    def for_health_check(cls) -> "Mempalace":
+        """Pre-wired palace for CI / deployment health-check workflows."""
+        p = cls(domain="health_check")
+        p.add_room("service_topology", "Service Topology Room")
+        p.add_room("failure_signals", "Failure Signals Room")
+        p.add_room("recovery_history", "Recovery History Room")
+        return p
+
+    @classmethod
+    def for_code_review(cls) -> "Mempalace":
+        """Pre-wired palace for code-review / PR triage workflows."""
+        p = cls(domain="code_review")
+        p.add_room("change_scope", "Change Scope Room")
+        p.add_room("risk_surface", "Risk Surface Room")
+        p.add_room("test_coverage", "Test Coverage Room")
+        p.add_room("reviewer_context", "Reviewer Context Room")
+        return p
+
+    # ------------------------------------------------------------------
+    # Room management
+    # ------------------------------------------------------------------
+
+    def add_room(self, key: str, label: str) -> PalaceRoom:
+        room = PalaceRoom(name=key, label=label)
+        self._rooms[key] = room
+        return room
+
+    def enter(self, room_key: str) -> PalaceRoom:
+        if room_key not in self._rooms:
+            raise KeyError(f"No room '{room_key}' in palace. Available: {list(self._rooms)}")
+        self._current_room = room_key
+        return self._rooms[room_key]
+
+    def store(self, key: str, value: Any) -> None:
+        """Store a value in the currently active room."""
+        if self._current_room is None:
+            raise RuntimeError("Enter a room before storing values.")
+        self._rooms[self._current_room].store(key, value)
+
+    def retrieve(self, room_key: str, key: str, default: Any = None) -> Any:
+        if room_key not in self._rooms:
+            return default
+        return self._rooms[room_key].retrieve(key, default)
+
+    # ------------------------------------------------------------------
+    # Rendering
+    # ------------------------------------------------------------------
+
+    def render(self) -> str:
+        """Return a human-readable summary of the entire palace."""
+        elapsed = time.time() - self._created_at
+        lines = [
+            f"# Mempalace — {self.domain}",
+            f"_traversal time: {elapsed:.2f}s | rooms: {len(self._rooms)}_",
+            "",
+        ]
+        for room in self._rooms.values():
+            lines.append(room.summary())
+            lines.append("")
+        return "\n".join(lines)
+
+    def to_dict(self) -> dict:
+        return {
+            "domain": self.domain,
+            "elapsed_seconds": round(time.time() - self._created_at, 3),
+            "rooms": {k: v.contents for k, v in self._rooms.items()},
+        }
+
+    def to_json(self) -> str:
+        return json.dumps(self.to_dict(), indent=2)
+
+
+# ---------------------------------------------------------------------------
+# Skill entry-point
+# ---------------------------------------------------------------------------
+
+def analyse_issues(
+    repos_data: list[dict],
+    target_assignee_rate: float = 0.80,
+) -> str:
+    """
+    Applies the mempalace technique to a list of repo issue summaries.
+
+    Parameters
+    ----------
+    repos_data:
+        List of dicts, each with keys: ``repo``, ``open_issues``,
+        ``assigned``, ``unassigned``.
+    target_assignee_rate:
+        Minimum acceptable assignee-coverage ratio (default 0.80).
+
+    Returns
+    -------
+    str
+        Rendered palace summary with coverage assessment.
+    """
+    palace = Mempalace.for_issue_analysis()
+
+    # --- Repository Architecture Room ---
+    palace.enter("repo_architecture")
+    total_issues = sum(r.get("open_issues", 0) for r in repos_data)
+    repos_with_issues = sum(1 for r in repos_data if r.get("open_issues", 0) > 0)
+    palace.store("repos_sampled", len(repos_data))
+    palace.store("repos_with_issues", repos_with_issues)
+    palace.store("total_open_issues", total_issues)
+    palace.store(
+        "avg_issues_per_repo",
+        round(total_issues / len(repos_data), 1) if repos_data else 0,
+    )
+
+    # --- Assignment Status Room ---
+    palace.enter("assignment_status")
+    total_assigned = sum(r.get("assigned", 0) for r in repos_data)
+    total_unassigned = sum(r.get("unassigned", 0) for r in repos_data)
+    coverage = total_assigned / total_issues if total_issues else 0
+    palace.store("assigned", total_assigned)
+    palace.store("unassigned", total_unassigned)
+    palace.store("coverage_rate", round(coverage, 3))
+    palace.store(
+        "coverage_status",
+        "OK" if coverage >= target_assignee_rate else f"BELOW TARGET ({target_assignee_rate:.0%})",
+    )
+
+    # --- Triage Priority Room ---
+    palace.enter("triage_priority")
+    unassigned_repos = [r["repo"] for r in repos_data if r.get("unassigned", 0) > 0]
+    palace.store("repos_needing_triage", unassigned_repos)
+    palace.store("triage_count", total_unassigned)
+
+    # --- Resolution Patterns Room ---
+    palace.enter("resolution_patterns")
+    palace.store("technique", "mempalace")
+    palace.store("target_assignee_rate", target_assignee_rate)
+
+    return palace.render()

hermes-sovereign/mempalace/promotion.py

@@ -0,0 +1,188 @@
+"""Memory Promotion — quality-gated scratchpad-to-palace promotion.
+
+Implements MP-4 (#371): move session notes to durable memory only when
+they pass quality gates. No LLM calls — all heuristic-based.
+
+Quality gates:
+  1. Minimum content length (too short = noise)
+  2. Duplicate detection (FTS5 + HRR similarity check)
+  3. Structural quality (has subject-verb structure, not just a fragment)
+  4. Staleness check (don't promote stale notes from old sessions)
+
+Refs: Epic #367, Sub-issue #371
+"""
+
+from __future__ import annotations
+
+import re
+import time
+from typing import Optional
+
+try:
+    from .sovereign_store import SovereignStore
+except ImportError:
+    from sovereign_store import SovereignStore
+
+
+# ---------------------------------------------------------------------------
+# Quality gate thresholds
+# ---------------------------------------------------------------------------
+
+MIN_CONTENT_WORDS = 5
+MAX_CONTENT_WORDS = 500
+DUPLICATE_SIMILARITY = 0.85
+DUPLICATE_FTS_THRESHOLD = 3
+STALE_SECONDS = 86400 * 7
+MIN_TRUST_FOR_AUTO = 0.4
+
+
+# ---------------------------------------------------------------------------
+# Quality checks
+# ---------------------------------------------------------------------------
+
+def _check_length(content: str) -> tuple[bool, str]:
+    """Gate 1: Content length check."""
+    words = content.split()
+    if len(words) < MIN_CONTENT_WORDS:
+        return False, f"Too short ({len(words)} words, minimum {MIN_CONTENT_WORDS})"
+    if len(words) > MAX_CONTENT_WORDS:
+        return False, f"Too long ({len(words)} words, maximum {MAX_CONTENT_WORDS}). Summarize first."
+    return True, "OK"
+
+
+def _check_structure(content: str) -> tuple[bool, str]:
+    """Gate 2: Basic structural quality."""
+    if not re.search(r"[a-zA-Z]", content):
+        return False, "No alphabetic content — pure code/numbers are not memory-worthy"
+    if len(content.split()) < 3:
+        return False, "Fragment — needs at least subject + predicate"
+    return True, "OK"
+
+
+def _check_duplicate(content: str, store: SovereignStore, room: str) -> tuple[bool, str]:
+    """Gate 3: Duplicate detection via hybrid search."""
+    results = store.search(content, room=room, limit=5, min_trust=0.0)
+    for r in results:
+        if r["score"] > DUPLICATE_SIMILARITY:
+            return False, f"Duplicate detected: memory #{r['memory_id']} (score {r['score']:.3f})"
+        if _text_overlap(content, r["content"]) > 0.8:
+            return False, f"Near-duplicate text: memory #{r['memory_id']}"
+    return True, "OK"
+
+
+def _check_staleness(written_at: float) -> tuple[bool, str]:
+    """Gate 4: Staleness check."""
+    age = time.time() - written_at
+    if age > STALE_SECONDS:
+        days = int(age / 86400)
+        return False, f"Stale ({days} days old). Review manually before promoting."
+    return True, "OK"
+
+
+def _text_overlap(a: str, b: str) -> float:
+    """Jaccard similarity between two texts (word-level)."""
+    words_a = set(a.lower().split())
+    words_b = set(b.lower().split())
+    if not words_a or not words_b:
+        return 0.0
+    intersection = words_a & words_b
+    union = words_a | words_b
+    return len(intersection) / len(union)
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+class PromotionResult:
+    """Result of a promotion attempt."""
+    def __init__(self, success: bool, memory_id: Optional[int], reason: str, gates: dict):
+        self.success = success
+        self.memory_id = memory_id
+        self.reason = reason
+        self.gates = gates
+
+    def __repr__(self):
+        status = "PROMOTED" if self.success else "REJECTED"
+        return f"PromotionResult({status}: {self.reason})"
+
+
+def evaluate_for_promotion(
+    content: str,
+    store: SovereignStore,
+    room: str = "general",
+    written_at: Optional[float] = None,
+) -> dict:
+    """Run all quality gates without actually promoting."""
+    if written_at is None:
+        written_at = time.time()
+    gates = {}
+    gates["length"] = _check_length(content)
+    gates["structure"] = _check_structure(content)
+    gates["duplicate"] = _check_duplicate(content, store, room)
+    gates["staleness"] = _check_staleness(written_at)
+    all_passed = all(passed for passed, _ in gates.values())
+    return {
+        "eligible": all_passed,
+        "gates": gates,
+        "content_preview": content[:100] + ("..." if len(content) > 100 else ""),
+    }
+
+
+def promote(
+    content: str,
+    store: SovereignStore,
+    session_id: str,
+    scratch_key: str,
+    room: str = "general",
+    category: str = "",
+    trust: float = 0.5,
+    written_at: Optional[float] = None,
+    force: bool = False,
+) -> PromotionResult:
+    """Promote a scratchpad note to durable palace memory."""
+    if written_at is None:
+        written_at = time.time()
+    gates = {}
+    if not force:
+        gates["length"] = _check_length(content)
+        gates["structure"] = _check_structure(content)
+        gates["duplicate"] = _check_duplicate(content, store, room)
+        gates["staleness"] = _check_staleness(written_at)
+        for gate_name, (passed, message) in gates.items():
+            if not passed:
+                return PromotionResult(
+                    success=False, memory_id=None,
+                    reason=f"Failed gate '{gate_name}': {message}", gates=gates,
+                )
+    memory_id = store.store(content, room=room, category=category, trust=trust)
+    store.log_promotion(session_id, scratch_key, memory_id, reason="auto" if not force else "forced")
+    return PromotionResult(success=True, memory_id=memory_id, reason="Promoted to durable memory", gates=gates)
+
+
+def promote_session_batch(
+    store: SovereignStore,
+    session_id: str,
+    notes: dict[str, dict],
+    room: str = "general",
+    force: bool = False,
+) -> list[PromotionResult]:
+    """Promote all notes from a session scratchpad."""
+    results = []
+    for key, entry in notes.items():
+        content = entry.get("value", str(entry)) if isinstance(entry, dict) else str(entry)
+        written_at = None
+        if isinstance(entry, dict) and "written_at" in entry:
+            try:
+                import datetime
+                written_at = datetime.datetime.strptime(
+                    entry["written_at"], "%Y-%m-%d %H:%M:%S"
+                ).timestamp()
+            except (ValueError, TypeError):
+                pass
+        result = promote(
+            content=str(content), store=store, session_id=session_id,
+            scratch_key=key, room=room, written_at=written_at, force=force,
+        )
+        results.append(result)
+    return results

hermes-sovereign/mempalace/retrieval_enforcer.py

@@ -0,0 +1,310 @@
+"""Retrieval Order Enforcer — L0 through L5 memory hierarchy.
+
+Ensures the agent checks durable memory before falling back to free generation.
+Gracefully degrades if any layer is unavailable (missing files, etc).
+
+Layer order:
+    L0: Identity      (~/.mempalace/identity.txt)
+    L1: Palace rooms   (SovereignStore — SQLite + FTS5 + HRR, zero API calls)
+    L2: Session scratch (~/.hermes/scratchpad/{session_id}.json)
+    L3: Gitea artifacts (API search for issues/PRs)
+    L4: Procedures      (skills directory search)
+    L5: Free generation (only if L0-L4 produced nothing)
+
+Refs: Epic #367, Sub-issue #369, Wiring: #383
+"""
+from __future__ import annotations
+
+import json
+import os
+import re
+from pathlib import Path
+from typing import Optional
+
+# ---------------------------------------------------------------------------
+# Sovereign Store (replaces mempalace CLI subprocess)
+# ---------------------------------------------------------------------------
+try:
+    from .sovereign_store import SovereignStore
+except ImportError:
+    try:
+        from sovereign_store import SovereignStore
+    except ImportError:
+        SovereignStore = None  # type: ignore[misc,assignment]
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+IDENTITY_PATH = Path.home() / ".mempalace" / "identity.txt"
+SCRATCHPAD_DIR = Path.home() / ".hermes" / "scratchpad"
+SKILLS_DIR = Path.home() / ".hermes" / "skills"
+SOVEREIGN_DB = Path.home() / ".hermes" / "palace" / "sovereign.db"
+
+# Patterns that indicate a recall-style query
+RECALL_PATTERNS = re.compile(
+    r"(?i)\b("
+    r"what did|status of|remember|last time|yesterday|previously|"
+    r"we discussed|we talked|we worked|you said|you mentioned|"
+    r"remind me|what was|what were|how did|when did|"
+    r"earlier today|last session|before this"
+    r")\b"
+)
+
+# Singleton store instance (lazy-init)
+_store: Optional["SovereignStore"] = None
+
+
+def _get_store() -> Optional["SovereignStore"]:
+    """Lazy-init the SovereignStore singleton."""
+    global _store
+    if _store is not None:
+        return _store
+    if SovereignStore is None:
+        return None
+    try:
+        _store = SovereignStore(db_path=str(SOVEREIGN_DB))
+        return _store
+    except Exception:
+        return None
+
+
+# ---------------------------------------------------------------------------
+# L0: Identity
+# ---------------------------------------------------------------------------
+
+def load_identity() -> str:
+    """Read the agent identity file. Returns empty string on failure."""
+    try:
+        if IDENTITY_PATH.exists():
+            text = IDENTITY_PATH.read_text(encoding="utf-8").strip()
+            # Cap at ~200 tokens to keep wake-up lean
+            if len(text.split()) > 200:
+                text = " ".join(text.split()[:200]) + "..."
+            return text
+    except (OSError, PermissionError):
+        pass
+    return ""
+
+
+# ---------------------------------------------------------------------------
+# L1: Palace search (now via SovereignStore — zero subprocess, zero API)
+# ---------------------------------------------------------------------------
+
+def search_palace(query: str, room: Optional[str] = None) -> str:
+    """Search the sovereign memory store for relevant memories.
+
+    Uses SovereignStore (SQLite + FTS5 + HRR) for hybrid keyword + semantic
+    search. No subprocess calls, no ONNX, no API keys.
+
+    Gracefully degrades to empty string if store is unavailable.
+    """
+    store = _get_store()
+    if store is None:
+        return ""
+    try:
+        results = store.search(query, room=room, limit=5, min_trust=0.2)
+        if not results:
+            return ""
+        lines = []
+        for r in results:
+            trust = r.get("trust_score", 0.5)
+            room_name = r.get("room", "general")
+            content = r.get("content", "")
+            lines.append(f"  [{room_name}] (trust:{trust:.2f}) {content}")
+        return "\n".join(lines)
+    except Exception:
+        return ""
+
+
+# ---------------------------------------------------------------------------
+# L2: Session scratchpad
+# ---------------------------------------------------------------------------
+
+def load_scratchpad(session_id: str) -> str:
+    """Load the session scratchpad as formatted text."""
+    try:
+        scratch_file = SCRATCHPAD_DIR / f"{session_id}.json"
+        if scratch_file.exists():
+            data = json.loads(scratch_file.read_text(encoding="utf-8"))
+            if isinstance(data, dict) and data:
+                lines = []
+                for k, v in data.items():
+                    lines.append(f"  {k}: {v}")
+                return "\n".join(lines)
+    except (OSError, json.JSONDecodeError):
+        pass
+    return ""
+
+
+# ---------------------------------------------------------------------------
+# L3: Gitea artifact search
+# ---------------------------------------------------------------------------
+
+def _load_gitea_token() -> str:
+    """Read the Gitea API token."""
+    token_path = Path.home() / ".hermes" / "gitea_token_vps"
+    try:
+        if token_path.exists():
+            return token_path.read_text(encoding="utf-8").strip()
+    except OSError:
+        pass
+    return ""
+
+
+def search_gitea(query: str) -> str:
+    """Search Gitea issues/PRs for context. Returns formatted text or empty string."""
+    token = _load_gitea_token()
+    if not token:
+        return ""
+
+    api_base = "https://forge.alexanderwhitestone.com/api/v1"
+    # Extract key terms for search (first 3 significant words)
+    terms = [w for w in query.split() if len(w) > 3][:3]
+    search_q = " ".join(terms) if terms else query[:50]
+
+    try:
+        import urllib.request
+        import urllib.parse
+
+        url = (
+            f"{api_base}/repos/search?"
+            f"q={urllib.parse.quote(search_q)}&limit=3"
+        )
+        req = urllib.request.Request(url, headers={
+            "Authorization": f"token {token}",
+            "Accept": "application/json",
+        })
+        with urllib.request.urlopen(req, timeout=8) as resp:
+            data = json.loads(resp.read().decode())
+            if data.get("data"):
+                lines = []
+                for repo in data["data"][:3]:
+                    lines.append(f"  {repo['full_name']}: {repo.get('description', 'no desc')}")
+                return "\n".join(lines)
+    except Exception:
+        pass
+    return ""
+
+
+# ---------------------------------------------------------------------------
+# L4: Procedures (skills search)
+# ---------------------------------------------------------------------------
+
+def search_skills(query: str) -> str:
+    """Search skills directory for matching procedures."""
+    try:
+        if not SKILLS_DIR.exists():
+            return ""
+
+        query_lower = query.lower()
+        terms = [w for w in query_lower.split() if len(w) > 3]
+        if not terms:
+            return ""
+
+        matches = []
+        for skill_dir in SKILLS_DIR.iterdir():
+            if not skill_dir.is_dir():
+                continue
+            skill_md = skill_dir / "SKILL.md"
+            if skill_md.exists():
+                try:
+                    content = skill_md.read_text(encoding="utf-8").lower()
+                    if any(t in content for t in terms):
+                        title = skill_dir.name
+                        matches.append(f"  skill: {title}")
+                except OSError:
+                    continue
+
+        if matches:
+            return "\n".join(matches[:5])
+    except OSError:
+        pass
+    return ""
+
+
+# ---------------------------------------------------------------------------
+# Main enforcer
+# ---------------------------------------------------------------------------
+
+def is_recall_query(query: str) -> bool:
+    """Detect whether a query is asking for recalled/historical information."""
+    return bool(RECALL_PATTERNS.search(query))
+
+
+def enforce_retrieval_order(
+    query: str,
+    session_id: Optional[str] = None,
+    skip_if_not_recall: bool = True,
+) -> dict:
+    """Check palace layers before allowing free generation.
+
+    Args:
+        query: The user's query text.
+        session_id: Current session ID for scratchpad access.
+        skip_if_not_recall: If True (default), skip enforcement for
+            non-recall queries and return empty result.
+
+    Returns:
+        dict with keys:
+            retrieved_from: Highest layer that produced results (e.g. 'L1')
+            context: Aggregated context string
+            tokens: Approximate word count of context
+            layers_checked: List of layers that were consulted
+    """
+    result = {
+        "retrieved_from": None,
+        "context": "",
+        "tokens": 0,
+        "layers_checked": [],
+    }
+
+    # Gate: skip for non-recall queries if configured
+    if skip_if_not_recall and not is_recall_query(query):
+        return result
+
+    # L0: Identity (always prepend)
+    identity = load_identity()
+    if identity:
+        result["context"] += f"## Identity\n{identity}\n\n"
+        result["layers_checked"].append("L0")
+
+    # L1: Palace search (SovereignStore — zero API, zero subprocess)
+    palace_results = search_palace(query)
+    if palace_results:
+        result["context"] += f"## Palace Memory\n{palace_results}\n\n"
+        result["retrieved_from"] = "L1"
+        result["layers_checked"].append("L1")
+
+    # L2: Scratchpad
+    if session_id:
+        scratch = load_scratchpad(session_id)
+        if scratch:
+            result["context"] += f"## Session Notes\n{scratch}\n\n"
+            if not result["retrieved_from"]:
+                result["retrieved_from"] = "L2"
+            result["layers_checked"].append("L2")
+
+    # L3: Gitea artifacts (only if still no context from L1/L2)
+    if not result["retrieved_from"]:
+        artifacts = search_gitea(query)
+        if artifacts:
+            result["context"] += f"## Gitea Context\n{artifacts}\n\n"
+            result["retrieved_from"] = "L3"
+            result["layers_checked"].append("L3")
+
+    # L4: Procedures (only if still no context)
+    if not result["retrieved_from"]:
+        procedures = search_skills(query)
+        if procedures:
+            result["context"] += f"## Related Skills\n{procedures}\n\n"
+            result["retrieved_from"] = "L4"
+            result["layers_checked"].append("L4")
+
+    # L5: Free generation (no context found — just mark it)
+    if not result["retrieved_from"]:
+        result["retrieved_from"] = "L5"
+        result["layers_checked"].append("L5")
+
+    result["tokens"] = len(result["context"].split())
+    return result

hermes-sovereign/mempalace/scratchpad.py

@@ -0,0 +1,184 @@
+"""Session Scratchpad — ephemeral key-value notes per session.
+
+Provides fast, JSON-backed scratch storage that lives for a session
+and can be promoted to durable palace memory.
+
+Storage: ~/.hermes/scratchpad/{session_id}.json
+
+Refs: Epic #367, Sub-issue #372
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import subprocess
+import time
+from pathlib import Path
+from typing import Any, Optional
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+SCRATCHPAD_DIR = Path.home() / ".hermes" / "scratchpad"
+MEMPALACE_BIN = "/Library/Frameworks/Python.framework/Versions/3.12/bin/mempalace"
+
+
+# ---------------------------------------------------------------------------
+# Internal helpers
+# ---------------------------------------------------------------------------
+
+def _scratch_path(session_id: str) -> Path:
+    """Return the JSON file path for a given session."""
+    # Sanitize session_id to prevent path traversal
+    safe_id = "".join(c for c in session_id if c.isalnum() or c in "-_")
+    if not safe_id:
+        safe_id = "unnamed"
+    return SCRATCHPAD_DIR / f"{safe_id}.json"
+
+
+def _load(session_id: str) -> dict:
+    """Load scratchpad data, returning empty dict on failure."""
+    path = _scratch_path(session_id)
+    try:
+        if path.exists():
+            return json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        pass
+    return {}
+
+
+def _save(session_id: str, data: dict) -> None:
+    """Persist scratchpad data to disk."""
+    SCRATCHPAD_DIR.mkdir(parents=True, exist_ok=True)
+    path = _scratch_path(session_id)
+    path.write_text(json.dumps(data, indent=2, default=str), encoding="utf-8")
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+def write_scratch(session_id: str, key: str, value: Any) -> None:
+    """Write a note to the session scratchpad.
+
+    Args:
+        session_id: Current session identifier.
+        key: Note key (string).
+        value: Note value (any JSON-serializable type).
+    """
+    data = _load(session_id)
+    data[key] = {
+        "value": value,
+        "written_at": time.strftime("%Y-%m-%d %H:%M:%S"),
+    }
+    _save(session_id, data)
+
+
+def read_scratch(session_id: str, key: Optional[str] = None) -> dict:
+    """Read session scratchpad (all keys or one).
+
+    Args:
+        session_id: Current session identifier.
+        key: Optional specific key. If None, returns all entries.
+
+    Returns:
+        dict — either {key: {value, written_at}} or the full scratchpad.
+    """
+    data = _load(session_id)
+    if key is not None:
+        entry = data.get(key)
+        return {key: entry} if entry else {}
+    return data
+
+
+def delete_scratch(session_id: str, key: str) -> bool:
+    """Remove a single key from the scratchpad.
+
+    Returns True if the key existed and was removed.
+    """
+    data = _load(session_id)
+    if key in data:
+        del data[key]
+        _save(session_id, data)
+        return True
+    return False
+
+
+def list_sessions() -> list[str]:
+    """List all session IDs that have scratchpad files."""
+    try:
+        if SCRATCHPAD_DIR.exists():
+            return [
+                f.stem
+                for f in SCRATCHPAD_DIR.iterdir()
+                if f.suffix == ".json" and f.is_file()
+            ]
+    except OSError:
+        pass
+    return []
+
+
+def promote_to_palace(
+    session_id: str,
+    key: str,
+    room: str = "general",
+    drawer: Optional[str] = None,
+) -> bool:
+    """Move a scratchpad note to durable palace memory.
+
+    Uses the mempalace CLI to store the note in the specified room.
+    Removes the note from the scratchpad after successful promotion.
+
+    Args:
+        session_id: Session containing the note.
+        key: Scratchpad key to promote.
+        room: Palace room name (default: 'general').
+        drawer: Optional drawer name within the room. Defaults to key.
+
+    Returns:
+        True if promotion succeeded, False otherwise.
+    """
+    data = _load(session_id)
+    entry = data.get(key)
+    if not entry:
+        return False
+
+    value = entry.get("value", entry) if isinstance(entry, dict) else entry
+    content = json.dumps(value, default=str) if not isinstance(value, str) else value
+
+    try:
+        bin_path = MEMPALACE_BIN if os.path.exists(MEMPALACE_BIN) else "mempalace"
+        target_drawer = drawer or key
+        result = subprocess.run(
+            [bin_path, "store", room, target_drawer, content],
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+        if result.returncode == 0:
+            # Remove from scratchpad after successful promotion
+            del data[key]
+            _save(session_id, data)
+            return True
+    except (FileNotFoundError, subprocess.TimeoutExpired, OSError):
+        # mempalace CLI not available — degrade gracefully
+        pass
+
+    return False
+
+
+def clear_session(session_id: str) -> bool:
+    """Delete the entire scratchpad for a session.
+
+    Returns True if the file existed and was removed.
+    """
+    path = _scratch_path(session_id)
+    try:
+        if path.exists():
+            path.unlink()
+            return True
+    except OSError:
+        pass
+    return False

hermes-sovereign/mempalace/sovereign_store.py

@@ -0,0 +1,474 @@
+"""Sovereign Memory Store — zero-API, zero-dependency durable memory.
+
+Replaces the third-party `mempalace` CLI and its ONNX requirement with a
+self-contained SQLite + FTS5 + HRR (Holographic Reduced Representation)
+store. Every operation is local: no network calls, no API keys, no cloud.
+
+Storage: ~/.hermes/palace/sovereign.db
+
+Capabilities:
+  - Durable fact storage with rooms, categories, and trust scores
+  - Hybrid retrieval: FTS5 keyword search + HRR cosine similarity
+  - Reciprocal Rank Fusion to merge keyword and semantic results
+  - Trust scoring: facts that get retrieved and confirmed gain trust
+  - Graceful numpy degradation: falls back to keyword-only if missing
+
+Refs: Epic #367, MP-3 #370, MP-4 #371
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+import sqlite3
+import struct
+import time
+from pathlib import Path
+from typing import Any, Optional
+
+# ---------------------------------------------------------------------------
+# HRR (Holographic Reduced Representations) — zero-dependency vectors
+# ---------------------------------------------------------------------------
+# Phase-encoded vectors via SHA-256. No ONNX, no embeddings API, no numpy
+# required (but uses numpy when available for speed).
+
+_TWO_PI = 2.0 * math.pi
+_DIM = 512  # Compact dimension — sufficient for memory retrieval
+
+try:
+    import numpy as np
+    _HAS_NUMPY = True
+except ImportError:
+    _HAS_NUMPY = False
+
+
+def _encode_atom_np(word: str, dim: int = _DIM) -> "np.ndarray":
+    """Deterministic phase vector via SHA-256 (numpy path)."""
+    values_per_block = 16
+    blocks_needed = math.ceil(dim / values_per_block)
+    uint16_values: list[int] = []
+    for i in range(blocks_needed):
+        digest = hashlib.sha256(f"{word}:{i}".encode()).digest()
+        uint16_values.extend(struct.unpack("<16H", digest))
+    return np.array(uint16_values[:dim], dtype=np.float64) * (_TWO_PI / 65536.0)
+
+
+def _encode_atom_pure(word: str, dim: int = _DIM) -> list[float]:
+    """Deterministic phase vector via SHA-256 (pure Python fallback)."""
+    values_per_block = 16
+    blocks_needed = math.ceil(dim / values_per_block)
+    uint16_values: list[int] = []
+    for i in range(blocks_needed):
+        digest = hashlib.sha256(f"{word}:{i}".encode()).digest()
+        for j in range(0, 32, 2):
+            uint16_values.append(int.from_bytes(digest[j:j+2], "little"))
+    return [v * (_TWO_PI / 65536.0) for v in uint16_values[:dim]]
+
+
+def encode_text(text: str, dim: int = _DIM):
+    """Encode a text string into an HRR phase vector by bundling word atoms.
+
+    Uses circular mean of per-word phase vectors — the standard HRR
+    superposition operation. Result is a fixed-width vector regardless
+    of input length.
+    """
+    words = text.lower().split()
+    if not words:
+        words = ["<empty>"]
+
+    if _HAS_NUMPY:
+        atoms = [_encode_atom_np(w, dim) for w in words]
+        # Circular mean: average the unit vectors, extract phase
+        unit_sum = sum(np.exp(1j * a) for a in atoms)
+        return np.angle(unit_sum) % _TWO_PI
+    else:
+        # Pure Python circular mean
+        real_sum = [0.0] * dim
+        imag_sum = [0.0] * dim
+        for w in words:
+            atom = _encode_atom_pure(w, dim)
+            for d in range(dim):
+                real_sum[d] += math.cos(atom[d])
+                imag_sum[d] += math.sin(atom[d])
+        return [math.atan2(imag_sum[d], real_sum[d]) % _TWO_PI for d in range(dim)]
+
+
+def cosine_similarity_phase(a, b) -> float:
+    """Cosine similarity between two phase vectors.
+
+    For phase vectors, similarity = mean(cos(a - b)).
+    """
+    if _HAS_NUMPY:
+        return float(np.mean(np.cos(np.array(a) - np.array(b))))
+    else:
+        n = len(a)
+        return sum(math.cos(a[i] - b[i]) for i in range(n)) / n
+
+
+def serialize_vector(vec) -> bytes:
+    """Serialize a vector to bytes for SQLite storage."""
+    if _HAS_NUMPY:
+        return vec.astype(np.float64).tobytes()
+    else:
+        return struct.pack(f"{len(vec)}d", *vec)
+
+
+def deserialize_vector(blob: bytes):
+    """Deserialize bytes back to a vector."""
+    n = len(blob) // 8  # float64 = 8 bytes
+    if _HAS_NUMPY:
+        return np.frombuffer(blob, dtype=np.float64)
+    else:
+        return list(struct.unpack(f"{n}d", blob))
+
+
+# ---------------------------------------------------------------------------
+# SQLite Schema
+# ---------------------------------------------------------------------------
+
+_SCHEMA = """
+CREATE TABLE IF NOT EXISTS memories (
+    memory_id       INTEGER PRIMARY KEY AUTOINCREMENT,
+    content         TEXT NOT NULL,
+    room            TEXT DEFAULT 'general',
+    category        TEXT DEFAULT '',
+    trust_score     REAL DEFAULT 0.5,
+    retrieval_count INTEGER DEFAULT 0,
+    created_at      REAL NOT NULL,
+    updated_at      REAL NOT NULL,
+    hrr_vector      BLOB
+);
+
+CREATE INDEX IF NOT EXISTS idx_memories_room ON memories(room);
+CREATE INDEX IF NOT EXISTS idx_memories_trust ON memories(trust_score DESC);
+
+-- FTS5 for fast keyword search
+CREATE VIRTUAL TABLE IF NOT EXISTS memories_fts USING fts5(
+    content, room, category,
+    content=memories, content_rowid=memory_id,
+    tokenize='porter unicode61'
+);
+
+-- Sync triggers
+CREATE TRIGGER IF NOT EXISTS memories_ai AFTER INSERT ON memories BEGIN
+    INSERT INTO memories_fts(rowid, content, room, category)
+        VALUES (new.memory_id, new.content, new.room, new.category);
+END;
+
+CREATE TRIGGER IF NOT EXISTS memories_ad AFTER DELETE ON memories BEGIN
+    INSERT INTO memories_fts(memories_fts, rowid, content, room, category)
+        VALUES ('delete', old.memory_id, old.content, old.room, old.category);
+END;
+
+CREATE TRIGGER IF NOT EXISTS memories_au AFTER UPDATE ON memories BEGIN
+    INSERT INTO memories_fts(memories_fts, rowid, content, room, category)
+        VALUES ('delete', old.memory_id, old.content, old.room, old.category);
+    INSERT INTO memories_fts(rowid, content, room, category)
+        VALUES (new.memory_id, new.content, new.room, new.category);
+END;
+
+-- Promotion log: tracks what moved from scratchpad to durable memory
+CREATE TABLE IF NOT EXISTS promotion_log (
+    log_id      INTEGER PRIMARY KEY AUTOINCREMENT,
+    session_id  TEXT NOT NULL,
+    scratch_key TEXT NOT NULL,
+    memory_id   INTEGER REFERENCES memories(memory_id),
+    promoted_at REAL NOT NULL,
+    reason      TEXT DEFAULT ''
+);
+"""
+
+
+# ---------------------------------------------------------------------------
+# SovereignStore
+# ---------------------------------------------------------------------------
+
+class SovereignStore:
+    """Zero-API durable memory store.
+
+    All operations are local SQLite. No network calls. No API keys.
+    HRR vectors provide semantic similarity without embedding models.
+    FTS5 provides fast keyword search. RRF merges both rankings.
+    """
+
+    def __init__(self, db_path: Optional[str] = None):
+        if db_path is None:
+            db_path = str(Path.home() / ".hermes" / "palace" / "sovereign.db")
+        self._db_path = db_path
+        Path(db_path).parent.mkdir(parents=True, exist_ok=True)
+        self._conn = sqlite3.connect(db_path)
+        self._conn.row_factory = sqlite3.Row
+        self._conn.executescript(_SCHEMA)
+
+    def close(self):
+        self._conn.close()
+
+    # ------------------------------------------------------------------
+    # Store
+    # ------------------------------------------------------------------
+
+    def store(
+        self,
+        content: str,
+        room: str = "general",
+        category: str = "",
+        trust: float = 0.5,
+    ) -> int:
+        """Store a fact in durable memory. Returns the memory_id."""
+        now = time.time()
+        vec = encode_text(content)
+        blob = serialize_vector(vec)
+        cur = self._conn.execute(
+            """INSERT INTO memories (content, room, category, trust_score,
+               created_at, updated_at, hrr_vector)
+               VALUES (?, ?, ?, ?, ?, ?, ?)""",
+            (content, room, category, trust, now, now, blob),
+        )
+        self._conn.commit()
+        return cur.lastrowid
+
+    def store_batch(self, items: list[dict]) -> list[int]:
+        """Store multiple facts. Each item: {content, room?, category?, trust?}."""
+        ids = []
+        now = time.time()
+        for item in items:
+            content = item["content"]
+            vec = encode_text(content)
+            blob = serialize_vector(vec)
+            cur = self._conn.execute(
+                """INSERT INTO memories (content, room, category, trust_score,
+                   created_at, updated_at, hrr_vector)
+                   VALUES (?, ?, ?, ?, ?, ?, ?)""",
+                (
+                    content,
+                    item.get("room", "general"),
+                    item.get("category", ""),
+                    item.get("trust", 0.5),
+                    now, now, blob,
+                ),
+            )
+            ids.append(cur.lastrowid)
+        self._conn.commit()
+        return ids
+
+    # ------------------------------------------------------------------
+    # Search — hybrid FTS5 + HRR with Reciprocal Rank Fusion
+    # ------------------------------------------------------------------
+
+    def search(
+        self,
+        query: str,
+        room: Optional[str] = None,
+        limit: int = 10,
+        min_trust: float = 0.0,
+        fts_weight: float = 0.5,
+        hrr_weight: float = 0.5,
+    ) -> list[dict]:
+        """Hybrid search: FTS5 keywords + HRR semantic similarity.
+
+        Uses Reciprocal Rank Fusion (RRF) to merge both rankings.
+        Returns list of dicts with content, room, score, trust_score.
+        """
+        k_rrf = 60  # Standard RRF constant
+
+        # Stage 1: FTS5 candidates
+        fts_results = self._fts_search(query, room, min_trust, limit * 3)
+
+        # Stage 2: HRR candidates (scan top N by trust)
+        hrr_results = self._hrr_search(query, room, min_trust, limit * 3)
+
+        # Stage 3: RRF fusion
+        scores: dict[int, float] = {}
+        meta: dict[int, dict] = {}
+
+        for rank, row in enumerate(fts_results):
+            mid = row["memory_id"]
+            scores[mid] = scores.get(mid, 0) + fts_weight / (k_rrf + rank + 1)
+            meta[mid] = dict(row)
+
+        for rank, row in enumerate(hrr_results):
+            mid = row["memory_id"]
+            scores[mid] = scores.get(mid, 0) + hrr_weight / (k_rrf + rank + 1)
+            if mid not in meta:
+                meta[mid] = dict(row)
+
+        # Sort by fused score
+        ranked = sorted(scores.items(), key=lambda x: x[1], reverse=True)[:limit]
+
+        results = []
+        for mid, score in ranked:
+            m = meta[mid]
+            # Bump retrieval count
+            self._conn.execute(
+                "UPDATE memories SET retrieval_count = retrieval_count + 1 WHERE memory_id = ?",
+                (mid,),
+            )
+            results.append({
+                "memory_id": mid,
+                "content": m["content"],
+                "room": m["room"],
+                "category": m.get("category", ""),
+                "trust_score": m["trust_score"],
+                "score": round(score, 6),
+            })
+
+        if results:
+            self._conn.commit()
+        return results
+
+    def _fts_search(
+        self, query: str, room: Optional[str], min_trust: float, limit: int
+    ) -> list[dict]:
+        """FTS5 full-text search."""
+        try:
+            if room:
+                rows = self._conn.execute(
+                    """SELECT m.memory_id, m.content, m.room, m.category,
+                              m.trust_score, m.retrieval_count
+                       FROM memories_fts f
+                       JOIN memories m ON f.rowid = m.memory_id
+                       WHERE memories_fts MATCH ? AND m.room = ?
+                             AND m.trust_score >= ?
+                       ORDER BY rank LIMIT ?""",
+                    (query, room, min_trust, limit),
+                ).fetchall()
+            else:
+                rows = self._conn.execute(
+                    """SELECT m.memory_id, m.content, m.room, m.category,
+                              m.trust_score, m.retrieval_count
+                       FROM memories_fts f
+                       JOIN memories m ON f.rowid = m.memory_id
+                       WHERE memories_fts MATCH ?
+                             AND m.trust_score >= ?
+                       ORDER BY rank LIMIT ?""",
+                    (query, min_trust, limit),
+                ).fetchall()
+            return [dict(r) for r in rows]
+        except sqlite3.OperationalError:
+            # Bad FTS query syntax — degrade gracefully
+            return []
+
+    def _hrr_search(
+        self, query: str, room: Optional[str], min_trust: float, limit: int
+    ) -> list[dict]:
+        """HRR cosine similarity search (brute-force scan, fast for <100K facts)."""
+        query_vec = encode_text(query)
+
+        if room:
+            rows = self._conn.execute(
+                """SELECT memory_id, content, room, category, trust_score,
+                          retrieval_count, hrr_vector
+                   FROM memories
+                   WHERE room = ? AND trust_score >= ? AND hrr_vector IS NOT NULL""",
+                (room, min_trust),
+            ).fetchall()
+        else:
+            rows = self._conn.execute(
+                """SELECT memory_id, content, room, category, trust_score,
+                          retrieval_count, hrr_vector
+                   FROM memories
+                   WHERE trust_score >= ? AND hrr_vector IS NOT NULL""",
+                (min_trust,),
+            ).fetchall()
+
+        scored = []
+        for r in rows:
+            stored_vec = deserialize_vector(r["hrr_vector"])
+            sim = cosine_similarity_phase(query_vec, stored_vec)
+            scored.append((sim, dict(r)))
+
+        scored.sort(key=lambda x: x[0], reverse=True)
+        return [item[1] for item in scored[:limit]]
+
+    # ------------------------------------------------------------------
+    # Trust management
+    # ------------------------------------------------------------------
+
+    def boost_trust(self, memory_id: int, delta: float = 0.05) -> None:
+        """Increase trust score when a memory proves useful."""
+        self._conn.execute(
+            """UPDATE memories SET trust_score = MIN(1.0, trust_score + ?),
+               updated_at = ? WHERE memory_id = ?""",
+            (delta, time.time(), memory_id),
+        )
+        self._conn.commit()
+
+    def decay_trust(self, memory_id: int, delta: float = 0.02) -> None:
+        """Decrease trust score when a memory is contradicted."""
+        self._conn.execute(
+            """UPDATE memories SET trust_score = MAX(0.0, trust_score - ?),
+               updated_at = ? WHERE memory_id = ?""",
+            (delta, time.time(), memory_id),
+        )
+        self._conn.commit()
+
+    # ------------------------------------------------------------------
+    # Room operations
+    # ------------------------------------------------------------------
+
+    def list_rooms(self) -> list[dict]:
+        """List all rooms with fact counts."""
+        rows = self._conn.execute(
+            """SELECT room, COUNT(*) as count,
+                      AVG(trust_score) as avg_trust
+               FROM memories GROUP BY room ORDER BY count DESC"""
+        ).fetchall()
+        return [dict(r) for r in rows]
+
+    def room_contents(self, room: str, limit: int = 50) -> list[dict]:
+        """Get all facts in a room, ordered by trust."""
+        rows = self._conn.execute(
+            """SELECT memory_id, content, category, trust_score,
+                      retrieval_count, created_at
+               FROM memories WHERE room = ?
+               ORDER BY trust_score DESC, created_at DESC LIMIT ?""",
+            (room, limit),
+        ).fetchall()
+        return [dict(r) for r in rows]
+
+    # ------------------------------------------------------------------
+    # Stats
+    # ------------------------------------------------------------------
+
+    def stats(self) -> dict:
+        """Return store statistics."""
+        row = self._conn.execute(
+            """SELECT COUNT(*) as total,
+                      AVG(trust_score) as avg_trust,
+                      SUM(retrieval_count) as total_retrievals,
+                      COUNT(DISTINCT room) as room_count
+               FROM memories"""
+        ).fetchone()
+        return dict(row)
+
+    # ------------------------------------------------------------------
+    # Promotion support (scratchpad → durable)
+    # ------------------------------------------------------------------
+
+    def log_promotion(
+        self,
+        session_id: str,
+        scratch_key: str,
+        memory_id: int,
+        reason: str = "",
+    ) -> None:
+        """Record a scratchpad-to-palace promotion in the audit log."""
+        self._conn.execute(
+            """INSERT INTO promotion_log
+               (session_id, scratch_key, memory_id, promoted_at, reason)
+               VALUES (?, ?, ?, ?, ?)""",
+            (session_id, scratch_key, memory_id, time.time(), reason),
+        )
+        self._conn.commit()
+
+    def recent_promotions(self, limit: int = 20) -> list[dict]:
+        """Get recent promotion log entries."""
+        rows = self._conn.execute(
+            """SELECT p.*, m.content, m.room
+               FROM promotion_log p
+               LEFT JOIN memories m ON p.memory_id = m.memory_id
+               ORDER BY p.promoted_at DESC LIMIT ?""",
+            (limit,),
+        ).fetchall()
+        return [dict(r) for r in rows]

hermes-sovereign/mempalace/tests/test_mempalace.py

@@ -0,0 +1,180 @@
+"""Tests for the mempalace skill.
+
+Validates PalaceRoom, Mempalace class, factory constructors,
+and the analyse_issues entry-point.
+
+Refs: Epic #367, Sub-issue #368
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+import os
+import time
+
+import pytest
+
+# Ensure the package is importable from the repo layout
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from mempalace.mempalace import Mempalace, PalaceRoom, analyse_issues
+
+
+# ── PalaceRoom unit tests ─────────────────────────────────────────────────
+
+class TestPalaceRoom:
+    def test_store_and_retrieve(self):
+        room = PalaceRoom(name="test", label="Test Room")
+        room.store("key1", 42)
+        assert room.retrieve("key1") == 42
+
+    def test_retrieve_default(self):
+        room = PalaceRoom(name="test", label="Test Room")
+        assert room.retrieve("missing") is None
+        assert room.retrieve("missing", "fallback") == "fallback"
+
+    def test_summary_format(self):
+        room = PalaceRoom(name="test", label="Test Room")
+        room.store("repos", 5)
+        summary = room.summary()
+        assert "## Test Room" in summary
+        assert "repos: 5" in summary
+
+    def test_contents_default_factory_isolation(self):
+        """Each room gets its own dict — no shared mutable default."""
+        r1 = PalaceRoom(name="a", label="A")
+        r2 = PalaceRoom(name="b", label="B")
+        r1.store("x", 1)
+        assert r2.retrieve("x") is None
+
+    def test_entered_at_is_recent(self):
+        before = time.time()
+        room = PalaceRoom(name="t", label="T")
+        after = time.time()
+        assert before <= room.entered_at <= after
+
+
+# ── Mempalace core tests ──────────────────────────────────────────────────
+
+class TestMempalace:
+    def test_add_and_enter_room(self):
+        p = Mempalace(domain="test")
+        p.add_room("r1", "Room 1")
+        room = p.enter("r1")
+        assert room.name == "r1"
+
+    def test_enter_nonexistent_room_raises(self):
+        p = Mempalace()
+        with pytest.raises(KeyError, match="No room"):
+            p.enter("ghost")
+
+    def test_store_without_enter_raises(self):
+        p = Mempalace()
+        p.add_room("r", "R")
+        with pytest.raises(RuntimeError, match="Enter a room"):
+            p.store("k", "v")
+
+    def test_store_and_retrieve_via_palace(self):
+        p = Mempalace()
+        p.add_room("r", "R")
+        p.enter("r")
+        p.store("count", 10)
+        assert p.retrieve("r", "count") == 10
+
+    def test_retrieve_missing_room_returns_default(self):
+        p = Mempalace()
+        assert p.retrieve("nope", "key") is None
+        assert p.retrieve("nope", "key", 99) == 99
+
+    def test_render_includes_domain(self):
+        p = Mempalace(domain="audit")
+        p.add_room("r", "Room")
+        p.enter("r")
+        p.store("item", "value")
+        output = p.render()
+        assert "audit" in output
+        assert "Room" in output
+
+    def test_to_dict_structure(self):
+        p = Mempalace(domain="test")
+        p.add_room("r", "R")
+        p.enter("r")
+        p.store("a", 1)
+        d = p.to_dict()
+        assert d["domain"] == "test"
+        assert "elapsed_seconds" in d
+        assert d["rooms"]["r"] == {"a": 1}
+
+    def test_to_json_is_valid(self):
+        p = Mempalace(domain="j")
+        p.add_room("x", "X")
+        p.enter("x")
+        p.store("v", [1, 2, 3])
+        parsed = json.loads(p.to_json())
+        assert parsed["rooms"]["x"]["v"] == [1, 2, 3]
+
+
+# ── Factory constructor tests ─────────────────────────────────────────────
+
+class TestFactories:
+    def test_for_issue_analysis_rooms(self):
+        p = Mempalace.for_issue_analysis()
+        assert p.domain == "issue_analysis"
+        for key in ("repo_architecture", "assignment_status",
+                    "triage_priority", "resolution_patterns"):
+            p.enter(key)  # should not raise
+
+    def test_for_health_check_rooms(self):
+        p = Mempalace.for_health_check()
+        assert p.domain == "health_check"
+        for key in ("service_topology", "failure_signals", "recovery_history"):
+            p.enter(key)
+
+    def test_for_code_review_rooms(self):
+        p = Mempalace.for_code_review()
+        assert p.domain == "code_review"
+        for key in ("change_scope", "risk_surface",
+                    "test_coverage", "reviewer_context"):
+            p.enter(key)
+
+
+# ── analyse_issues entry-point tests ──────────────────────────────────────
+
+class TestAnalyseIssues:
+    SAMPLE_DATA = [
+        {"repo": "the-nexus", "open_issues": 40, "assigned": 30, "unassigned": 10},
+        {"repo": "timmy-home", "open_issues": 30, "assigned": 25, "unassigned": 5},
+        {"repo": "hermes-agent", "open_issues": 20, "assigned": 15, "unassigned": 5},
+        {"repo": "empty-repo", "open_issues": 0, "assigned": 0, "unassigned": 0},
+    ]
+
+    def test_returns_string(self):
+        result = analyse_issues(self.SAMPLE_DATA)
+        assert isinstance(result, str)
+        assert len(result) > 0
+
+    def test_contains_room_headers(self):
+        result = analyse_issues(self.SAMPLE_DATA)
+        assert "Repository Architecture" in result
+        assert "Assignment Status" in result
+
+    def test_coverage_below_target(self):
+        result = analyse_issues(self.SAMPLE_DATA, target_assignee_rate=0.90)
+        assert "BELOW TARGET" in result
+
+    def test_coverage_meets_target(self):
+        good_data = [
+            {"repo": "a", "open_issues": 10, "assigned": 10, "unassigned": 0},
+        ]
+        result = analyse_issues(good_data, target_assignee_rate=0.80)
+        assert "OK" in result
+
+    def test_empty_repos_list(self):
+        result = analyse_issues([])
+        assert isinstance(result, str)
+
+    def test_single_repo(self):
+        data = [{"repo": "solo", "open_issues": 5, "assigned": 3, "unassigned": 2}]
+        result = analyse_issues(data)
+        assert "solo" in result or "issue_analysis" in result

hermes-sovereign/mempalace/tests/test_retrieval_enforcer.py

@@ -0,0 +1,143 @@
+"""Tests for retrieval_enforcer.py.
+
+Refs: Epic #367, Sub-issue #369
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+import tempfile
+from pathlib import Path
+from unittest.mock import patch, MagicMock
+
+import pytest
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from mempalace.retrieval_enforcer import (
+    is_recall_query,
+    load_identity,
+    load_scratchpad,
+    enforce_retrieval_order,
+    search_skills,
+    RECALL_PATTERNS,
+)
+
+
+class TestRecallDetection:
+    """Test the recall-query pattern matcher."""
+
+    @pytest.mark.parametrize("query", [
+        "what did we work on yesterday",
+        "status of the mempalace integration",
+        "remember the fleet audit results",
+        "last time we deployed the nexus",
+        "previously you mentioned a CI fix",
+        "we discussed the sovereign deployment",
+    ])
+    def test_recall_queries_detected(self, query):
+        assert is_recall_query(query) is True
+
+    @pytest.mark.parametrize("query", [
+        "create a new file called test.py",
+        "run the test suite",
+        "deploy to production",
+        "write a function that sums numbers",
+        "install the package",
+    ])
+    def test_non_recall_queries_skipped(self, query):
+        assert is_recall_query(query) is False
+
+
+class TestLoadIdentity:
+    def test_loads_existing_identity(self, tmp_path):
+        identity_file = tmp_path / "identity.txt"
+        identity_file.write_text("I am Timmy. A sovereign AI.")
+        with patch("mempalace.retrieval_enforcer.IDENTITY_PATH", identity_file):
+            result = load_identity()
+            assert "Timmy" in result
+
+    def test_returns_empty_on_missing_file(self, tmp_path):
+        identity_file = tmp_path / "nonexistent.txt"
+        with patch("mempalace.retrieval_enforcer.IDENTITY_PATH", identity_file):
+            result = load_identity()
+            assert result == ""
+
+    def test_truncates_long_identity(self, tmp_path):
+        identity_file = tmp_path / "identity.txt"
+        identity_file.write_text(" ".join(["word"] * 300))
+        with patch("mempalace.retrieval_enforcer.IDENTITY_PATH", identity_file):
+            result = load_identity()
+            assert result.endswith("...")
+            assert len(result.split()) <= 201  # 200 words + "..."
+
+
+class TestLoadScratchpad:
+    def test_loads_valid_scratchpad(self, tmp_path):
+        scratch_file = tmp_path / "session123.json"
+        scratch_file.write_text(json.dumps({"note": "test value", "key2": 42}))
+        with patch("mempalace.retrieval_enforcer.SCRATCHPAD_DIR", tmp_path):
+            result = load_scratchpad("session123")
+            assert "note: test value" in result
+            assert "key2: 42" in result
+
+    def test_returns_empty_on_missing_file(self, tmp_path):
+        with patch("mempalace.retrieval_enforcer.SCRATCHPAD_DIR", tmp_path):
+            result = load_scratchpad("nonexistent")
+            assert result == ""
+
+    def test_returns_empty_on_invalid_json(self, tmp_path):
+        scratch_file = tmp_path / "bad.json"
+        scratch_file.write_text("not valid json{{{")
+        with patch("mempalace.retrieval_enforcer.SCRATCHPAD_DIR", tmp_path):
+            result = load_scratchpad("bad")
+            assert result == ""
+
+
+class TestEnforceRetrievalOrder:
+    def test_skips_non_recall_query(self):
+        result = enforce_retrieval_order("create a new file")
+        assert result["retrieved_from"] is None
+        assert result["tokens"] == 0
+
+    def test_runs_for_recall_query(self, tmp_path):
+        identity_file = tmp_path / "identity.txt"
+        identity_file.write_text("I am Timmy.")
+        with patch("mempalace.retrieval_enforcer.IDENTITY_PATH", identity_file), \
+             patch("mempalace.retrieval_enforcer.search_palace", return_value=""), \
+             patch("mempalace.retrieval_enforcer.search_gitea", return_value=""), \
+             patch("mempalace.retrieval_enforcer.search_skills", return_value=""):
+            result = enforce_retrieval_order("what did we work on yesterday")
+            assert "Identity" in result["context"]
+            assert "L0" in result["layers_checked"]
+
+    def test_palace_hit_sets_l1(self, tmp_path):
+        identity_file = tmp_path / "identity.txt"
+        identity_file.write_text("I am Timmy.")
+        with patch("mempalace.retrieval_enforcer.IDENTITY_PATH", identity_file), \
+             patch("mempalace.retrieval_enforcer.search_palace", return_value="Found: fleet audit results"), \
+             patch("mempalace.retrieval_enforcer.search_gitea", return_value=""):
+            result = enforce_retrieval_order("what did we discuss yesterday")
+            assert result["retrieved_from"] == "L1"
+            assert "Palace Memory" in result["context"]
+
+    def test_falls_through_to_l5(self, tmp_path):
+        identity_file = tmp_path / "nonexistent.txt"
+        with patch("mempalace.retrieval_enforcer.IDENTITY_PATH", identity_file), \
+             patch("mempalace.retrieval_enforcer.search_palace", return_value=""), \
+             patch("mempalace.retrieval_enforcer.search_gitea", return_value=""), \
+             patch("mempalace.retrieval_enforcer.search_skills", return_value=""):
+            result = enforce_retrieval_order("remember the old deployment", skip_if_not_recall=True)
+            assert result["retrieved_from"] == "L5"
+
+    def test_force_mode_skips_recall_check(self, tmp_path):
+        identity_file = tmp_path / "identity.txt"
+        identity_file.write_text("I am Timmy.")
+        with patch("mempalace.retrieval_enforcer.IDENTITY_PATH", identity_file), \
+             patch("mempalace.retrieval_enforcer.search_palace", return_value=""), \
+             patch("mempalace.retrieval_enforcer.search_gitea", return_value=""), \
+             patch("mempalace.retrieval_enforcer.search_skills", return_value=""):
+            result = enforce_retrieval_order("deploy now", skip_if_not_recall=False)
+            assert "Identity" in result["context"]

hermes-sovereign/mempalace/tests/test_scratchpad.py

@@ -0,0 +1,108 @@
+"""Tests for scratchpad.py.
+
+Refs: Epic #367, Sub-issue #372
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from mempalace.scratchpad import (
+    write_scratch,
+    read_scratch,
+    delete_scratch,
+    list_sessions,
+    clear_session,
+    _scratch_path,
+)
+
+
+@pytest.fixture
+def scratch_dir(tmp_path):
+    """Provide a temporary scratchpad directory."""
+    with patch("mempalace.scratchpad.SCRATCHPAD_DIR", tmp_path):
+        yield tmp_path
+
+
+class TestScratchPath:
+    def test_sanitizes_session_id(self):
+        path = _scratch_path("safe-id_123")
+        assert "safe-id_123.json" in str(path)
+
+    def test_strips_dangerous_chars(self):
+        path = _scratch_path("../../etc/passwd")
+        assert ".." not in path.name
+        assert "/" not in path.name
+        # Dots are stripped, so only alphanumeric chars remain
+        assert path.name == "etcpasswd.json"
+
+
+class TestWriteAndRead:
+    def test_write_then_read(self, scratch_dir):
+        write_scratch("sess1", "note", "hello world")
+        result = read_scratch("sess1", "note")
+        assert "note" in result
+        assert result["note"]["value"] == "hello world"
+
+    def test_read_all_keys(self, scratch_dir):
+        write_scratch("sess1", "a", 1)
+        write_scratch("sess1", "b", 2)
+        result = read_scratch("sess1")
+        assert "a" in result
+        assert "b" in result
+
+    def test_read_missing_key(self, scratch_dir):
+        write_scratch("sess1", "exists", "yes")
+        result = read_scratch("sess1", "missing")
+        assert result == {}
+
+    def test_read_missing_session(self, scratch_dir):
+        result = read_scratch("nonexistent")
+        assert result == {}
+
+    def test_overwrite_key(self, scratch_dir):
+        write_scratch("sess1", "key", "v1")
+        write_scratch("sess1", "key", "v2")
+        result = read_scratch("sess1", "key")
+        assert result["key"]["value"] == "v2"
+
+
+class TestDelete:
+    def test_delete_existing_key(self, scratch_dir):
+        write_scratch("sess1", "key", "val")
+        assert delete_scratch("sess1", "key") is True
+        assert read_scratch("sess1", "key") == {}
+
+    def test_delete_missing_key(self, scratch_dir):
+        write_scratch("sess1", "other", "val")
+        assert delete_scratch("sess1", "missing") is False
+
+
+class TestListSessions:
+    def test_lists_sessions(self, scratch_dir):
+        write_scratch("alpha", "k", "v")
+        write_scratch("beta", "k", "v")
+        sessions = list_sessions()
+        assert "alpha" in sessions
+        assert "beta" in sessions
+
+    def test_empty_directory(self, scratch_dir):
+        assert list_sessions() == []
+
+
+class TestClearSession:
+    def test_clears_existing(self, scratch_dir):
+        write_scratch("sess1", "k", "v")
+        assert clear_session("sess1") is True
+        assert read_scratch("sess1") == {}
+
+    def test_clear_nonexistent(self, scratch_dir):
+        assert clear_session("ghost") is False

hermes-sovereign/mempalace/tests/test_sovereign_store.py

@@ -0,0 +1,255 @@
+"""Tests for the Sovereign Memory Store and Promotion system.
+
+Zero-API, zero-network — everything runs against an in-memory SQLite DB.
+"""
+
+import os
+import sys
+import tempfile
+import time
+import unittest
+
+# Allow imports from parent package
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from sovereign_store import (
+    SovereignStore,
+    encode_text,
+    cosine_similarity_phase,
+    serialize_vector,
+    deserialize_vector,
+)
+from promotion import (
+    evaluate_for_promotion,
+    promote,
+    promote_session_batch,
+)
+
+
+class TestHRRVectors(unittest.TestCase):
+    """Test the HRR encoding and similarity functions."""
+
+    def test_deterministic_encoding(self):
+        """Same text always produces the same vector."""
+        v1 = encode_text("hello world")
+        v2 = encode_text("hello world")
+        self.assertAlmostEqual(cosine_similarity_phase(v1, v2), 1.0, places=5)
+
+    def test_similar_texts_higher_similarity(self):
+        """Related texts should be more similar than unrelated ones."""
+        v_agent = encode_text("agent memory palace retrieval")
+        v_similar = encode_text("agent recall memory search")
+        v_unrelated = encode_text("banana strawberry fruit smoothie")
+        sim_related = cosine_similarity_phase(v_agent, v_similar)
+        sim_unrelated = cosine_similarity_phase(v_agent, v_unrelated)
+        self.assertGreater(sim_related, sim_unrelated)
+
+    def test_serialize_roundtrip(self):
+        """Vectors survive serialization to/from bytes."""
+        vec = encode_text("test serialization")
+        blob = serialize_vector(vec)
+        restored = deserialize_vector(blob)
+        sim = cosine_similarity_phase(vec, restored)
+        self.assertAlmostEqual(sim, 1.0, places=5)
+
+    def test_empty_text(self):
+        """Empty text gets a fallback encoding."""
+        vec = encode_text("")
+        self.assertEqual(len(vec) if hasattr(vec, '__len__') else len(list(vec)), 512)
+
+
+class TestSovereignStore(unittest.TestCase):
+    """Test the SQLite-backed sovereign store."""
+
+    def setUp(self):
+        self.db_path = os.path.join(tempfile.mkdtemp(), "test.db")
+        self.store = SovereignStore(db_path=self.db_path)
+
+    def tearDown(self):
+        self.store.close()
+        if os.path.exists(self.db_path):
+            os.remove(self.db_path)
+
+    def test_store_and_retrieve(self):
+        """Store a fact and find it via search."""
+        mid = self.store.store("Timmy is a sovereign AI agent on Hermes VPS", room="identity")
+        results = self.store.search("sovereign agent", room="identity")
+        self.assertTrue(any(r["memory_id"] == mid for r in results))
+
+    def test_fts_search(self):
+        """FTS5 keyword search works."""
+        self.store.store("The beacon game uses paperclips mechanics", room="projects")
+        self.store.store("Fleet agents handle delegation and dispatch", room="fleet")
+        results = self.store.search("paperclips")
+        self.assertTrue(len(results) > 0)
+        self.assertIn("paperclips", results[0]["content"].lower())
+
+    def test_hrr_search_semantic(self):
+        """HRR similarity finds related content even without exact keywords."""
+        self.store.store("Memory palace rooms organize facts spatially", room="memory")
+        self.store.store("Pizza delivery service runs on weekends", room="unrelated")
+        results = self.store.search("organize knowledge rooms", room="memory")
+        self.assertTrue(len(results) > 0)
+        self.assertIn("palace", results[0]["content"].lower())
+
+    def test_room_filtering(self):
+        """Room filter restricts search scope."""
+        self.store.store("Hermes harness manages tool calls", room="infrastructure")
+        self.store.store("Hermes mythology Greek god", room="lore")
+        results = self.store.search("Hermes", room="infrastructure")
+        self.assertTrue(all(r["room"] == "infrastructure" for r in results))
+
+    def test_trust_boost(self):
+        """Trust score increases when boosted."""
+        mid = self.store.store("fact", trust=0.5)
+        self.store.boost_trust(mid, delta=0.1)
+        results = self.store.room_contents("general")
+        fact = next(r for r in results if r["memory_id"] == mid)
+        self.assertAlmostEqual(fact["trust_score"], 0.6, places=2)
+
+    def test_trust_decay(self):
+        """Trust score decreases when decayed."""
+        mid = self.store.store("questionable fact", trust=0.5)
+        self.store.decay_trust(mid, delta=0.2)
+        results = self.store.room_contents("general")
+        fact = next(r for r in results if r["memory_id"] == mid)
+        self.assertAlmostEqual(fact["trust_score"], 0.3, places=2)
+
+    def test_batch_store(self):
+        """Batch store works."""
+        ids = self.store.store_batch([
+            {"content": "fact one", "room": "test"},
+            {"content": "fact two", "room": "test"},
+            {"content": "fact three", "room": "test"},
+        ])
+        self.assertEqual(len(ids), 3)
+        rooms = self.store.list_rooms()
+        test_room = next(r for r in rooms if r["room"] == "test")
+        self.assertEqual(test_room["count"], 3)
+
+    def test_stats(self):
+        """Stats returns correct counts."""
+        self.store.store("a fact", room="r1")
+        self.store.store("another fact", room="r2")
+        s = self.store.stats()
+        self.assertEqual(s["total"], 2)
+        self.assertEqual(s["room_count"], 2)
+
+    def test_retrieval_count_increments(self):
+        """Retrieval count goes up when a fact is found via search."""
+        self.store.store("unique searchable content xyz123", room="test")
+        self.store.search("xyz123")
+        results = self.store.room_contents("test")
+        self.assertTrue(any(r["retrieval_count"] > 0 for r in results))
+
+
+class TestPromotion(unittest.TestCase):
+    """Test the quality-gated promotion system."""
+
+    def setUp(self):
+        self.db_path = os.path.join(tempfile.mkdtemp(), "promo_test.db")
+        self.store = SovereignStore(db_path=self.db_path)
+
+    def tearDown(self):
+        self.store.close()
+
+    def test_successful_promotion(self):
+        """Good content passes all gates."""
+        result = promote(
+            content="Timmy runs on the Hermes VPS at 143.198.27.163 with local Ollama inference",
+            store=self.store,
+            session_id="test-session-001",
+            scratch_key="vps_info",
+            room="infrastructure",
+        )
+        self.assertTrue(result.success)
+        self.assertIsNotNone(result.memory_id)
+
+    def test_reject_too_short(self):
+        """Short fragments get rejected."""
+        result = promote(
+            content="yes",
+            store=self.store,
+            session_id="test",
+            scratch_key="short",
+        )
+        self.assertFalse(result.success)
+        self.assertIn("Too short", result.reason)
+
+    def test_reject_duplicate(self):
+        """Duplicate content gets rejected."""
+        self.store.store("SOUL.md is the canonical identity document for Timmy", room="identity")
+        result = promote(
+            content="SOUL.md is the canonical identity document for Timmy",
+            store=self.store,
+            session_id="test",
+            scratch_key="soul",
+            room="identity",
+        )
+        self.assertFalse(result.success)
+        self.assertIn("uplicate", result.reason)
+
+    def test_reject_stale(self):
+        """Old notes get flagged as stale."""
+        old_time = time.time() - (86400 * 10)
+        result = promote(
+            content="This is a note from long ago about something important",
+            store=self.store,
+            session_id="test",
+            scratch_key="old",
+            written_at=old_time,
+        )
+        self.assertFalse(result.success)
+        self.assertIn("Stale", result.reason)
+
+    def test_force_bypasses_gates(self):
+        """Force flag overrides quality gates."""
+        result = promote(
+            content="ok",
+            store=self.store,
+            session_id="test",
+            scratch_key="forced",
+            force=True,
+        )
+        self.assertTrue(result.success)
+
+    def test_evaluate_dry_run(self):
+        """Evaluate returns gate details without promoting."""
+        eval_result = evaluate_for_promotion(
+            content="The fleet uses kimi-k2.5 as the primary model for all agent operations",
+            store=self.store,
+            room="fleet",
+        )
+        self.assertTrue(eval_result["eligible"])
+        self.assertTrue(all(p for p, _ in eval_result["gates"].values()))
+
+    def test_batch_promotion(self):
+        """Batch promotion processes all notes."""
+        notes = {
+            "infra": {"value": "Hermes VPS runs Ubuntu 22.04 with 2 vCPUs and 4GB RAM", "written_at": time.strftime("%Y-%m-%d %H:%M:%S")},
+            "short": {"value": "no", "written_at": time.strftime("%Y-%m-%d %H:%M:%S")},
+            "model": {"value": "The primary local model is gemma4:latest running on Ollama", "written_at": time.strftime("%Y-%m-%d %H:%M:%S")},
+        }
+        results = promote_session_batch(self.store, "batch-session", notes, room="config")
+        promoted = [r for r in results if r.success]
+        rejected = [r for r in results if not r.success]
+        self.assertEqual(len(promoted), 2)
+        self.assertEqual(len(rejected), 1)
+
+    def test_promotion_logged(self):
+        """Successful promotions appear in the audit log."""
+        promote(
+            content="Forge is hosted at forge.alexanderwhitestone.com running Gitea",
+            store=self.store,
+            session_id="log-test",
+            scratch_key="forge",
+            room="infrastructure",
+        )
+        log = self.store.recent_promotions()
+        self.assertTrue(len(log) > 0)
+        self.assertEqual(log[0]["session_id"], "log-test")
+        self.assertEqual(log[0]["scratch_key"], "forge")
+
+
+if __name__ == "__main__":
+    unittest.main()

hermes-sovereign/mempalace/tests/test_wakeup.py

@@ -0,0 +1,100 @@
+"""Tests for wakeup.py.
+
+Refs: Epic #367, Sub-issue #372
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+import time
+from pathlib import Path
+from unittest.mock import patch, MagicMock
+
+import pytest
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from mempalace.wakeup import (
+    palace_wakeup,
+    fleet_status_summary,
+    _load_identity,
+    _palace_context,
+)
+
+
+class TestLoadIdentity:
+    def test_loads_identity(self, tmp_path):
+        f = tmp_path / "identity.txt"
+        f.write_text("I am Timmy. A sovereign AI.")
+        with patch("mempalace.wakeup.IDENTITY_PATH", f):
+            result = _load_identity()
+            assert "Timmy" in result
+
+    def test_missing_identity(self, tmp_path):
+        f = tmp_path / "nope.txt"
+        with patch("mempalace.wakeup.IDENTITY_PATH", f):
+            assert _load_identity() == ""
+
+
+class TestFleetStatus:
+    def test_reads_fleet_json(self, tmp_path):
+        f = tmp_path / "fleet_status.json"
+        f.write_text(json.dumps({
+            "Groq": {"state": "active", "last_seen": "2026-04-07"},
+            "Ezra": {"state": "idle", "last_seen": "2026-04-06"},
+        }))
+        with patch("mempalace.wakeup.FLEET_STATUS_PATH", f):
+            result = fleet_status_summary()
+            assert "Fleet Status" in result
+            assert "Groq" in result
+            assert "active" in result
+
+    def test_missing_fleet_file(self, tmp_path):
+        f = tmp_path / "nope.json"
+        with patch("mempalace.wakeup.FLEET_STATUS_PATH", f):
+            assert fleet_status_summary() == ""
+
+    def test_invalid_json(self, tmp_path):
+        f = tmp_path / "bad.json"
+        f.write_text("not json")
+        with patch("mempalace.wakeup.FLEET_STATUS_PATH", f):
+            assert fleet_status_summary() == ""
+
+
+class TestPalaceWakeup:
+    def test_generates_context_with_identity(self, tmp_path):
+        identity = tmp_path / "identity.txt"
+        identity.write_text("I am Timmy.")
+        cache = tmp_path / "cache.txt"
+        with patch("mempalace.wakeup.IDENTITY_PATH", identity), \
+             patch("mempalace.wakeup.WAKEUP_CACHE_PATH", cache), \
+             patch("mempalace.wakeup._palace_context", return_value=""), \
+             patch("mempalace.wakeup.fleet_status_summary", return_value=""):
+            result = palace_wakeup(force=True)
+            assert "Identity" in result
+            assert "Timmy" in result
+            assert "Session" in result
+
+    def test_uses_cache_when_fresh(self, tmp_path):
+        cache = tmp_path / "cache.txt"
+        cache.write_text("cached wake-up content")
+        # Touch the file so it's fresh
+        with patch("mempalace.wakeup.WAKEUP_CACHE_PATH", cache), \
+             patch("mempalace.wakeup.WAKEUP_CACHE_TTL", 9999):
+            result = palace_wakeup(force=False)
+            assert result == "cached wake-up content"
+
+    def test_force_bypasses_cache(self, tmp_path):
+        cache = tmp_path / "cache.txt"
+        cache.write_text("stale content")
+        identity = tmp_path / "identity.txt"
+        identity.write_text("I am Timmy.")
+        with patch("mempalace.wakeup.WAKEUP_CACHE_PATH", cache), \
+             patch("mempalace.wakeup.IDENTITY_PATH", identity), \
+             patch("mempalace.wakeup._palace_context", return_value=""), \
+             patch("mempalace.wakeup.fleet_status_summary", return_value=""):
+            result = palace_wakeup(force=True)
+            assert "Identity" in result
+            assert "stale content" not in result

hermes-sovereign/mempalace/wakeup.py

@@ -0,0 +1,161 @@
+"""Wake-up Protocol — session start context injection.
+
+Generates 300-900 tokens of context when a new Hermes session starts.
+Loads identity, recent palace context, and fleet status.
+
+Refs: Epic #367, Sub-issue #372
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import subprocess
+import time
+from pathlib import Path
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+IDENTITY_PATH = Path.home() / ".mempalace" / "identity.txt"
+MEMPALACE_BIN = "/Library/Frameworks/Python.framework/Versions/3.12/bin/mempalace"
+FLEET_STATUS_PATH = Path.home() / ".hermes" / "fleet_status.json"
+WAKEUP_CACHE_PATH = Path.home() / ".hermes" / "last_wakeup.txt"
+WAKEUP_CACHE_TTL = 300  # 5 minutes — don't regenerate if recent
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def _load_identity() -> str:
+    """Read the agent identity file."""
+    try:
+        if IDENTITY_PATH.exists():
+            text = IDENTITY_PATH.read_text(encoding="utf-8").strip()
+            # Cap at ~150 tokens for wake-up brevity
+            words = text.split()
+            if len(words) > 150:
+                text = " ".join(words[:150]) + "..."
+            return text
+    except (OSError, PermissionError):
+        pass
+    return ""
+
+
+def _palace_context() -> str:
+    """Run mempalace wake-up command for recent context. Degrades gracefully."""
+    try:
+        bin_path = MEMPALACE_BIN if os.path.exists(MEMPALACE_BIN) else "mempalace"
+        result = subprocess.run(
+            [bin_path, "wake-up"],
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+        if result.returncode == 0 and result.stdout.strip():
+            return result.stdout.strip()
+    except (FileNotFoundError, subprocess.TimeoutExpired, OSError):
+        # ONNX issues (#373) or CLI not available — degrade gracefully
+        pass
+    return ""
+
+
+def fleet_status_summary() -> str:
+    """Read cached fleet status for lightweight session context."""
+    try:
+        if FLEET_STATUS_PATH.exists():
+            data = json.loads(FLEET_STATUS_PATH.read_text(encoding="utf-8"))
+            lines = ["## Fleet Status"]
+
+            if isinstance(data, dict):
+                for agent, status in data.items():
+                    if isinstance(status, dict):
+                        state = status.get("state", "unknown")
+                        last_seen = status.get("last_seen", "?")
+                        lines.append(f"  {agent}: {state} (last: {last_seen})")
+                    else:
+                        lines.append(f"  {agent}: {status}")
+
+            if len(lines) > 1:
+                return "\n".join(lines)
+    except (OSError, json.JSONDecodeError):
+        pass
+    return ""
+
+
+def _check_cache() -> str:
+    """Return cached wake-up if fresh enough."""
+    try:
+        if WAKEUP_CACHE_PATH.exists():
+            age = time.time() - WAKEUP_CACHE_PATH.stat().st_mtime
+            if age < WAKEUP_CACHE_TTL:
+                return WAKEUP_CACHE_PATH.read_text(encoding="utf-8").strip()
+    except OSError:
+        pass
+    return ""
+
+
+def _write_cache(content: str) -> None:
+    """Cache the wake-up content."""
+    try:
+        WAKEUP_CACHE_PATH.parent.mkdir(parents=True, exist_ok=True)
+        WAKEUP_CACHE_PATH.write_text(content, encoding="utf-8")
+    except OSError:
+        pass
+
+
+# ---------------------------------------------------------------------------
+# Main entry point
+# ---------------------------------------------------------------------------
+
+def palace_wakeup(force: bool = False) -> str:
+    """Generate wake-up context for a new session. ~300-900 tokens.
+
+    Args:
+        force: If True, bypass the 5-minute cache and regenerate.
+
+    Returns:
+        Formatted context string suitable for prepending to the system prompt.
+    """
+    # Check cache first (avoids redundant work on rapid session restarts)
+    if not force:
+        cached = _check_cache()
+        if cached:
+            return cached
+
+    parts = []
+
+    # L0: Identity
+    identity = _load_identity()
+    if identity:
+        parts.append(f"## Identity\n{identity}")
+
+    # L1: Recent palace context
+    palace = _palace_context()
+    if palace:
+        parts.append(palace)
+
+    # Fleet status (lightweight)
+    fleet = fleet_status_summary()
+    if fleet:
+        parts.append(fleet)
+
+    # Timestamp
+    parts.append(f"## Session\nWake-up generated: {time.strftime('%Y-%m-%d %H:%M:%S')}")
+
+    content = "\n\n".join(parts)
+
+    # Cache for TTL
+    _write_cache(content)
+
+    return content
+
+
+# ---------------------------------------------------------------------------
+# CLI entry point for testing
+# ---------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    print(palace_wakeup(force=True))

wizards/bezalel-sidecar.json

@@ -0,0 +1,10 @@
+{
+  "name": "Bezalel Builder Wizard",
+  "role": "Artificer",
+  "capabilities": [
+    "provisioning",
+    "gpu-orchestration",
+    "model-serving"
+  ],
+  "instructions": "Take the provided keys and use them to get GPUs for big models (Gemma 4). Wire them into the harness."
+}