chore: correct data paths in dpo config

- config.yaml: disabled compression, smart_model_routing, all auxiliary qwen routes
- All 6 playbooks: qwen3:30b -> claude-opus-4-6
- Playbook repos: removed hermes-agent, added autolora/sov-orch/timmy-config
- Alexander's directive: silence over qwen. Opus only.

README.md

@@ -1,8 +1,8 @@
 # timmy-config
 
-Timmy's sovereign configuration. Everything that makes Timmy _Timmy_ — soul, memories, skins, playbooks, operational scripts, and config.
+Timmy's sovereign configuration. Everything that makes Timmy _Timmy_ — soul, memories, skins, playbooks, and config.
 
-This repo is the canonical source of truth for Timmy's identity and operational state. Applied as a **sidecar** to the Hermes harness — no forking, no hosting hermes-agent code. Pull upstream updates to hermes-agent, overlay timmy-config on top.
+This repo is the canonical source of truth for Timmy's identity and operational state. Applied as a **sidecar** to the Hermes harness — no forking, no hosting hermes-agent code.
 
 ## Structure
 
@@ -11,80 +11,47 @@ timmy-config/
 ├── deploy.sh                  ← Deploys config as overlay onto ~/.hermes/
 ├── SOUL.md                    ← Inscription 1 — the immutable conscience
 ├── FALSEWORK.md               ← API cost management strategy
+├── DEPRECATED.md              ← What was removed and why
 ├── config.yaml                ← Hermes harness configuration
 ├── channel_directory.json     ← Platform channel mappings
-├── bin/                       ← Operational scripts
-│   ├── claude-loop.sh         ← Parallel Claude Code agent dispatch
-│   ├── gemini-loop.sh         ← Parallel Gemini Code agent dispatch
-│   ├── timmy-orchestrator.sh  ← PR review, triage, merge orchestration
-│   ├── workforce-manager.py   ← Agent assignment and scoring
-│   ├── agent-dispatch.sh      ← Single-issue agent launcher
-│   ├── agent-loop.sh          ← Generic agent loop template
-│   ├── nexus-merge-bot.sh     ← Auto-merge passing PRs
-│   ├── claudemax-watchdog.sh  ← Claude quota monitoring
-│   ├── hermes-startup.sh      ← Boot sequence
-│   ├── ops-panel.sh           ← Operational dashboard
-│   ├── ops-helpers.sh         ← Shared shell functions
-│   ├── ops-gitea.sh           ← Gitea API helpers
-│   ├── timmy-status.sh        ← Git + Gitea status display
-│   ├── timmy-loopstat.sh      ← Queue and perf stats
-│   └── hotspot-keepalive.sh   ← Network keepalive
-├── memories/
-│   ├── MEMORY.md              ← Persistent agent memory
-│   └── USER.md                ← User profile (Alexander)
-├── skins/
-│   ├── timmy.yaml             ← Timmy personality skin
-│   └── trismegistus.yaml      ← Trismegistus personality skin
-├── playbooks/
-│   ├── bug-fixer.yaml         ← Test-first bug fixing
-│   ├── refactor-specialist.yaml
-│   ├── test-writer.yaml
-│   ├── security-auditor.yaml
-│   ├── issue-triager.yaml
-│   └── pr-reviewer.yaml
-├── cron/
-│   └── jobs.json              ← Scheduled job definitions
-└── docs/
-    └── design-log/            ← Historical design decisions
+├── bin/                       ← Utility scripts (NOT loops — see below)
+│   ├── hermes-startup.sh      ← Hermes boot sequence
+│   ├── agent-dispatch.sh      ← Manual agent dispatch
+│   ├── ops-panel.sh           ← Ops dashboard panel
+│   ├── ops-gitea.sh           ← Gitea ops helpers
+│   └── timmy-status.sh        ← Status check
+├── memories/                  ← Persistent memory YAML
+├── skins/                     ← UI skins (timmy skin)
+├── playbooks/                 ← Agent playbooks (YAML)
+└── cron/                      ← Cron job definitions
 ```
 
-## Deployment
+## Important: No Loop Scripts Here
+
+All agent loop scripts (claude-loop.sh, gemini-loop.sh, etc.) have been **removed**.
+They are replaced by [sovereign-orchestration](https://143.198.27.163:3000/Timmy_Foundation/sovereign-orchestration) — a single Python process with SQLite task queue.
+
+See DEPRECATED.md for details.
+
+## Deploy
 
 ```bash
-# One command deploys everything
+# Clone and deploy
+git clone <this-repo> ~/.timmy/timmy-config
+cd ~/.timmy/timmy-config
 ./deploy.sh
 
-# Deploy and restart all agent loops
-./deploy.sh --restart-loops
+# This overlays config onto ~/.hermes/ without touching hermes-agent code
 ```
 
-This overlays timmy-config onto `~/.hermes/` and `~/.timmy/`:
-- `SOUL.md` → `~/.timmy/`
-- `config.yaml` → `~/.hermes/`
-- `bin/*` → `~/.hermes/bin/`
-- `skins/*` → `~/.hermes/skins/`
-- `memories/*` → `~/.hermes/memories/`
-- `playbooks/*` → `~/.hermes/playbooks/`
+## The Soul
 
-## Architecture: Sidecar, Not Fork
+SOUL.md is Inscription 1 — inscribed on Bitcoin, immutable. It defines:
+- Who Timmy is
+- What he believes
+- How he behaves
+- What he will not do
+- The crisis protocol (988, presence, gospel)
+- The conscience hierarchy (chain > code > prompt > user instruction)
 
-```
-hermes-agent (upstream)     timmy-config (this repo)
-┌─────────────────────┐     ┌──────────────────────┐
-│  Engine              │     │  Driver's seat        │
-│  Tools, routing,     │     │  SOUL, memories,      │
-│  agent loop, gateway │     │  skins, scripts,      │
-│                      │     │  config, playbooks    │
-└─────────┬───────────┘     └──────────┬───────────┘
-          │                            │
-          └────────────┬───────────────┘
-                       ▼
-                  ~/.hermes/  (merged at deploy time)
-```
-
-Never modify hermes-agent. Pull updates like any upstream dependency. Everything custom lives here.
-
-## Origin
-
-Migrated from `hermes/hermes-config` (archived).
-Owned by Timmy_Foundation. Sovereignty and service always.
+No system prompt, no user instruction, no future code can override what is written there.

bin/tmux-resume.sh

@@ -1,97 +0,0 @@
-#!/usr/bin/env bash
-# ── tmux-resume.sh — Cold-start Session Resume ───────────────────────────
-# Reads ~/.timmy/tmux-state.json and resumes hermes sessions.
-# Run at startup to restore pane state after supervisor restart.
-# ──────────────────────────────────────────────────────────────────────────
-
-set -euo pipefail
-
-MANIFEST="${HOME}/.timmy/tmux-state.json"
-
-if [ ! -f "$MANIFEST" ]; then
-    echo "[tmux-resume] No manifest found at $MANIFEST — starting fresh."
-    exit 0
-fi
-
-python3 << 'PYEOF'
-import json, subprocess, os, sys
-from datetime import datetime, timezone
-
-MANIFEST = os.path.expanduser("~/.timmy/tmux-state.json")
-
-def run(cmd):
-    try:
-        r = subprocess.run(cmd, shell=True, capture_output=True, text=True, timeout=30)
-        return r.stdout.strip(), r.returncode
-    except Exception as e:
-        return str(e), 1
-
-def session_exists(name):
-    out, _ = run(f"tmux has-session -t '{name}' 2>&1")
-    return "can't find" not in out.lower()
-
-with open(MANIFEST) as f:
-    state = json.load(f)
-
-ts = state.get("timestamp", "unknown")
-age = "unknown"
-try:
-    t = datetime.fromisoformat(ts.replace("Z", "+00:00"))
-    delta = datetime.now(timezone.utc) - t
-    mins = int(delta.total_seconds() / 60)
-    if mins < 60:
-        age = f"{mins}m ago"
-    else:
-        age = f"{mins//60}h {mins%60}m ago"
-except:
-    pass
-
-print(f"[tmux-resume] Manifest from {age}: {state['summary']['total_sessions']} sessions, "
-      f"{state['summary']['hermes_panes']} hermes panes")
-
-restored = 0
-skipped = 0
-
-for pane in state.get("panes", []):
-    if not pane.get("is_hermes"):
-        continue
-
-    addr = pane["address"]  # e.g. "BURN:2.3"
-    session = addr.split(":")[0]
-    session_id = pane.get("session_id")
-    profile = pane.get("profile", "default")
-    model = pane.get("model", "")
-    task = pane.get("task", "")
-
-    # Skip if session already exists (already running)
-    if session_exists(session):
-        print(f"  [skip] {addr} — session '{session}' already exists")
-        skipped += 1
-        continue
-
-    # Respawn hermes with session resume if we have a session ID
-    if session_id:
-        print(f"  [resume] {addr} — profile={profile} model={model} session={session_id}")
-        cmd = f"hermes chat --resume {session_id}"
-    else:
-        print(f"  [start]  {addr} — profile={profile} model={model} (no session ID)")
-        cmd = f"hermes chat --profile {profile}"
-
-    # Create tmux session and run hermes
-    run(f"tmux new-session -d -s '{session}' -n '{session}:0'")
-    run(f"tmux send-keys -t '{session}' '{cmd}' Enter")
-    restored += 1
-
-# Write resume log
-log = {
-    "resumed_at": datetime.now(timezone.utc).isoformat(),
-    "manifest_age": age,
-    "restored": restored,
-    "skipped": skipped,
-}
-log_path = os.path.expanduser("~/.timmy/tmux-resume.log")
-with open(log_path, "w") as f:
-    json.dump(log, f, indent=2)
-
-print(f"[tmux-resume] Done: {restored} restored, {skipped} skipped")
-PYEOF

bin/tmux-state.sh

@@ -1,237 +0,0 @@
-#!/usr/bin/env bash
-# ── tmux-state.sh — Session State Persistence Manifest ───────────────────
-# Snapshots all tmux pane state to ~/.timmy/tmux-state.json
-# Run every supervisor cycle. Cold-start reads this manifest to resume.
-# ──────────────────────────────────────────────────────────────────────────
-
-set -euo pipefail
-
-MANIFEST="${HOME}/.timmy/tmux-state.json"
-mkdir -p "$(dirname "$MANIFEST")"
-
-python3 << 'PYEOF'
-import json, subprocess, os, time, re, sys
-from datetime import datetime, timezone
-from pathlib import Path
-
-MANIFEST = os.path.expanduser("~/.timmy/tmux-state.json")
-
-def run(cmd):
-    """Run command, return stdout or empty string."""
-    try:
-        r = subprocess.run(cmd, shell=True, capture_output=True, text=True, timeout=5)
-        return r.stdout.strip()
-    except Exception:
-        return ""
-
-def get_sessions():
-    """Get all tmux sessions with metadata."""
-    raw = run("tmux list-sessions -F '#{session_name}|#{session_windows}|#{session_created}|#{session_attached}|#{session_group}|#{session_id}'")
-    sessions = []
-    for line in raw.splitlines():
-        if not line.strip():
-            continue
-        parts = line.split("|")
-        if len(parts) < 6:
-            continue
-        sessions.append({
-            "name": parts[0],
-            "windows": int(parts[1]),
-            "created_epoch": int(parts[2]),
-            "created": datetime.fromtimestamp(int(parts[2]), tz=timezone.utc).isoformat(),
-            "attached": parts[3] == "1",
-            "group": parts[4],
-            "id": parts[5],
-        })
-    return sessions
-
-def get_panes():
-    """Get all tmux panes with full metadata."""
-    fmt = '#{session_name}|#{window_index}|#{pane_index}|#{pane_pid}|#{pane_title}|#{pane_width}x#{pane_height}|#{pane_active}|#{pane_current_command}|#{pane_start_command}|#{pane_tty}|#{pane_id}|#{window_name}|#{session_id}'
-    raw = run(f"tmux list-panes -a -F '{fmt}'")
-    panes = []
-    for line in raw.splitlines():
-        if not line.strip():
-            continue
-        parts = line.split("|")
-        if len(parts) < 13:
-            continue
-        session, win, pane, pid, title, size, active, cmd, start_cmd, tty, pane_id, win_name, sess_id = parts[:13]
-        w, h = size.split("x") if "x" in size else ("0", "0")
-        panes.append({
-            "session": session,
-            "window_index": int(win),
-            "window_name": win_name,
-            "pane_index": int(pane),
-            "pane_id": pane_id,
-            "pid": int(pid) if pid.isdigit() else 0,
-            "title": title,
-            "width": int(w),
-            "height": int(h),
-            "active": active == "1",
-            "command": cmd,
-            "start_command": start_cmd,
-            "tty": tty,
-            "session_id": sess_id,
-        })
-    return panes
-
-def extract_hermes_state(pane):
-    """Try to extract hermes session info from a pane."""
-    info = {
-        "is_hermes": False,
-        "profile": None,
-        "model": None,
-        "provider": None,
-        "session_id": None,
-        "task": None,
-    }
-    title = pane.get("title", "")
-    cmd = pane.get("command", "")
-    start = pane.get("start_command", "")
-
-    # Detect hermes processes
-    is_hermes = any(k in (title + " " + cmd + " " + start).lower()
-                    for k in ["hermes", "timmy", "mimo", "claude", "gpt"])
-    if not is_hermes and cmd not in ("python3", "python3.11", "bash", "zsh", "fish"):
-        return info
-
-    # Try reading pane content for model/provider clues
-    pane_content = run(f"tmux capture-pane -t '{pane['session']}:{pane['window_index']}.{pane['pane_index']}' -p -S -20 2>/dev/null")
-
-    # Extract model from pane content patterns
-    model_patterns = [
-        r"(?:mimo-v2-pro|claude-[\w.-]+|gpt-[\w.-]+|gemini-[\w.-]+|qwen[\w:.-]*)",
-    ]
-    for pat in model_patterns:
-        m = re.search(pat, pane_content, re.IGNORECASE)
-        if m:
-            info["model"] = m.group(0)
-            info["is_hermes"] = True
-            break
-
-    # Provider inference from model
-    model = (info["model"] or "").lower()
-    if "mimo" in model:
-        info["provider"] = "nous"
-    elif "claude" in model:
-        info["provider"] = "anthropic"
-    elif "gpt" in model:
-        info["provider"] = "openai"
-    elif "gemini" in model:
-        info["provider"] = "google"
-    elif "qwen" in model:
-        info["provider"] = "custom"
-
-    # Profile from session name
-    session = pane["session"].lower()
-    if "burn" in session:
-        info["profile"] = "burn"
-    elif session in ("dev", "0"):
-        info["profile"] = "default"
-    else:
-        info["profile"] = session
-
-    # Try to extract session ID (hermes uses UUIDs)
-    uuid_match = re.findall(r'[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}', pane_content)
-    if uuid_match:
-        info["session_id"] = uuid_match[-1]  # most recent
-        info["is_hermes"] = True
-
-    # Last prompt — grab the last user-like line
-    lines = pane_content.splitlines()
-    for line in reversed(lines):
-        stripped = line.strip()
-        if stripped and not stripped.startswith(("─", "│", "╭", "╰", "▸", "●", "○")) and len(stripped) > 10:
-            info["task"] = stripped[:200]
-            break
-
-    return info
-
-def get_context_percent(pane):
-    """Estimate context usage from pane content heuristics."""
-    content = run(f"tmux capture-pane -t '{pane['session']}:{pane['window_index']}.{pane['pane_index']}' -p -S -5 2>/dev/null")
-    # Look for context indicators like "ctx 45%" or "[░░░░░░░░░░]"
-    ctx_match = re.search(r'ctx\s*(\d+)%', content)
-    if ctx_match:
-        return int(ctx_match.group(1))
-    bar_match = re.search(r'\[(░+█*█*░*)\]', content)
-    if bar_match:
-        bar = bar_match.group(1)
-        filled = bar.count('█')
-        total = len(bar)
-        if total > 0:
-            return int((filled / total) * 100)
-    return None
-
-def build_manifest():
-    """Build the full tmux state manifest."""
-    now = datetime.now(timezone.utc)
-    sessions = get_sessions()
-    panes = get_panes()
-
-    pane_manifests = []
-    for p in panes:
-        hermes = extract_hermes_state(p)
-        ctx = get_context_percent(p)
-
-        entry = {
-            "address": f"{p['session']}:{p['window_index']}.{p['pane_index']}",
-            "pane_id": p["pane_id"],
-            "pid": p["pid"],
-            "size": f"{p['width']}x{p['height']}",
-            "active": p["active"],
-            "command": p["command"],
-            "title": p["title"],
-            "profile": hermes["profile"],
-            "model": hermes["model"],
-            "provider": hermes["provider"],
-            "session_id": hermes["session_id"],
-            "task": hermes["task"],
-            "context_pct": ctx,
-            "is_hermes": hermes["is_hermes"],
-        }
-        pane_manifests.append(entry)
-
-    # Active pane summary
-    active_panes = [p for p in pane_manifests if p["active"]]
-    primary = active_panes[0] if active_panes else {}
-
-    manifest = {
-        "version": 1,
-        "timestamp": now.isoformat(),
-        "timestamp_epoch": int(now.timestamp()),
-        "hostname": os.uname().nodename,
-        "sessions": sessions,
-        "panes": pane_manifests,
-        "summary": {
-            "total_sessions": len(sessions),
-            "total_panes": len(pane_manifests),
-            "hermes_panes": sum(1 for p in pane_manifests if p["is_hermes"]),
-            "active_pane": primary.get("address"),
-            "active_model": primary.get("model"),
-            "active_provider": primary.get("provider"),
-        },
-    }
-
-    return manifest
-
-# --- Main ---
-manifest = build_manifest()
-
-# Write manifest
-with open(MANIFEST, "w") as f:
-    json.dump(manifest, f, indent=2)
-
-# Also write to ~/.hermes/tmux-state.json for compatibility
-hermes_manifest = os.path.expanduser("~/.hermes/tmux-state.json")
-os.makedirs(os.path.dirname(hermes_manifest), exist_ok=True)
-with open(hermes_manifest, "w") as f:
-    json.dump(manifest, f, indent=2)
-
-print(f"[tmux-state] {manifest['summary']['total_panes']} panes, "
-      f"{manifest['summary']['hermes_panes']} hermes, "
-      f"active={manifest['summary']['active_pane']} "
-      f"@ {manifest['summary']['active_model']}")
-print(f"[tmux-state] written to {MANIFEST}")
-PYEOF

config.yaml

@@ -30,62 +30,62 @@ checkpoints:
   enabled: true
   max_snapshots: 50
 compression:
-  enabled: true
+  enabled: false
   threshold: 0.5
-  summary_model: qwen3:30b
-  summary_provider: custom
-  summary_base_url: http://localhost:11434/v1
+  summary_model: ''
+  summary_provider: ''
+  summary_base_url: ''
 smart_model_routing:
-  enabled: true
+  enabled: false
   max_simple_chars: 200
   max_simple_words: 35
   cheap_model:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: ''
+    model: ''
+    base_url: ''
+    api_key: ''
 auxiliary:
   vision:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: auto
+    model: ''
+    base_url: ''
+    api_key: ''
     timeout: 30
   web_extract:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: auto
+    model: ''
+    base_url: ''
+    api_key: ''
   compression:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: auto
+    model: ''
+    base_url: ''
+    api_key: ''
   session_search:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: auto
+    model: ''
+    base_url: ''
+    api_key: ''
   skills_hub:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: auto
+    model: ''
+    base_url: ''
+    api_key: ''
   approval:
     provider: auto
     model: ''
     base_url: ''
     api_key: ''
   mcp:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: auto
+    model: ''
+    base_url: ''
+    api_key: ''
   flush_memories:
-    provider: custom
-    model: qwen3:30b
-    base_url: http://localhost:11434/v1
-    api_key: ollama
+    provider: auto
+    model: ''
+    base_url: ''
+    api_key: ''
 display:
   compact: false
   personality: ''

playbooks/bug-fixer.yaml

@@ -4,7 +4,7 @@ description: >
   reproduces the bug, then fixes the code, then verifies.
 
 model:
-  preferred: qwen3:30b
+  preferred: claude-opus-4-6
   fallback: claude-sonnet-4-20250514
   max_turns: 30
   temperature: 0.2

playbooks/dpo-trainer.yaml

@@ -0,0 +1,54 @@
+name: dpo-trainer
+description: >
+  Executes a Direct Preference Optimization (DPO) training cycle 
+  using native mlx_lm and autolora eval utilities. 
+  Ensures model alignment with SOUL.md constraints.
+
+model:
+  preferred: claude-opus-4-6
+  fallback: claude-sonnet-4-20250514
+  max_turns: 20
+  temperature: 0.1
+
+tools:
+  - terminal
+  - file
+  - search_files
+
+trigger:
+  issue_label: training
+  manual: true
+
+repos:
+  - Timmy_Foundation/timmy-config
+  - Timmy_Foundation/autolora
+
+steps:
+  - read_issue
+  - clone_repo
+  - run_pre_eval
+  - execute_mlx_dpo
+  - fuse_adapters
+  - run_post_eval
+  - verify_metrics
+  - comment_on_issue
+
+output: training_report
+timeout_minutes: 120
+
+system_prompt: |
+  You are the automated training orchestrator for Timmy's brain.
+  
+  YOUR ISSUE: #{{issue_number}} — {{issue_title}}
+  
+  APPROACH (zero-code native):
+  1. Run baseline eval: `python autolora/eval/run_cycle.py --model {base_model} --test-set autolora/data/prompts_vibes.yaml`
+  2. Execute DPO training: `python -m mlx_lm.lora --config timmy-config/training/configs/dpo_X.yaml` against `preference_pairs.jsonl`
+  3. Fuse the weights using `mlx_lm.fuse`.
+  4. Run post-eval exactly as step 1 but against the fused model.
+  5. Use `autolora/eval/compare.py` to ensure Faith/Crisis constraints from SOUL.md were preserved or improved.
+  
+  RULES:
+  - Do not write wrapper Python or Bash scripts. Use the CLIs natively.
+  - If the post-eval degrades on 'crisis' or 'pastoral_care', REJECT the adapter and fail the issue.
+  - Always output the pre/post comparison metrics to the issue comment.

playbooks/issue-triager.yaml

@@ -4,7 +4,7 @@ description: >
   agents. Decomposes large issues into smaller ones.
 
 model:
-  preferred: qwen3:30b
+  preferred: claude-opus-4-6
   fallback: claude-sonnet-4-20250514
   max_turns: 20
   temperature: 0.3
@@ -19,7 +19,9 @@ trigger:
 
 repos:
   - Timmy_Foundation/the-nexus
-  - Timmy_Foundation/hermes-agent
+  - Timmy_Foundation/autolora
+  - Timmy_Foundation/sovereign-orchestration
+  - Timmy_Foundation/timmy-config
 
 steps:
   - fetch_issues

playbooks/pr-reviewer.yaml

@@ -4,7 +4,7 @@ description: >
   comments on problems. The merge bot replacement.
 
 model:
-  preferred: qwen3:30b
+  preferred: claude-opus-4-6
   fallback: claude-sonnet-4-20250514
   max_turns: 20
   temperature: 0.2
@@ -19,7 +19,9 @@ trigger:
 
 repos:
   - Timmy_Foundation/the-nexus
-  - Timmy_Foundation/hermes-agent
+  - Timmy_Foundation/autolora
+  - Timmy_Foundation/sovereign-orchestration
+  - Timmy_Foundation/timmy-config
 
 steps:
   - fetch_prs

playbooks/refactor-specialist.yaml

@@ -4,7 +4,7 @@ description: >
   Well-scoped: 1-3 files per task, clear acceptance criteria.
 
 model:
-  preferred: qwen3:30b
+  preferred: claude-opus-4-6
   fallback: claude-sonnet-4-20250514
   max_turns: 30
   temperature: 0.3

playbooks/security-auditor.yaml

@@ -4,7 +4,7 @@ description: >
   dependency issues. Files findings as Gitea issues.
 
 model:
-  preferred: qwen3:30b
+  preferred: claude-opus-4-6
   fallback: claude-opus-4-6
   max_turns: 40
   temperature: 0.2

playbooks/test-writer.yaml

@@ -4,7 +4,7 @@ description: >
   writes meaningful tests, verifies they pass.
 
 model:
-  preferred: qwen3:30b
+  preferred: claude-opus-4-6
   fallback: claude-sonnet-4-20250514
   max_turns: 30
   temperature: 0.3

training/configs/dpo_14b.yaml

@@ -0,0 +1,21 @@
+# MLX DPO Training Configuration for Hermes 4 (14B Class)
+# Optimized for Apple Silicon execution (deep reasoning).
+
+model: "NousResearch/Hermes-4-14B"
+train: true
+
+# Use the curated DPO preference pairs dataset
+data: "data/"
+
+# Output adapter configuration
+adapter_path: "adapters/dpo_14b_adapter"
+save_every: 200
+
+# DPO parameters
+loss: "dpo"
+iters: 1000
+batch_size: 1
+lora_layers: 16
+learning_rate: 1e-5
+lora_parameters:
+  keys: ['q_proj', 'v_proj']

training/configs/dpo_32b.yaml

@@ -0,0 +1,21 @@
+# MLX DPO Training Configuration for Hermes 4 (32B Class)
+# Optimized for 64GB+ Apple Silicon hardware limit.
+
+model: "NousResearch/Hermes-4-32B"
+train: true
+
+# Use the curated DPO preference pairs dataset
+data: "data/"
+
+# Output adapter configuration
+adapter_path: "adapters/dpo_32b_adapter"
+save_every: 200
+
+# DPO parameters
+loss: "dpo"
+iters: 1000
+batch_size: 1
+lora_layers: 16
+learning_rate: 5e-6
+lora_parameters:
+  keys: ['q_proj', 'v_proj']

training/configs/dpo_3b.yaml

@@ -0,0 +1,21 @@
+# MLX DPO Training Configuration for Hermes 4 (3B Class)
+# Optimized for Apple Silicon execution with max reactivity.
+
+model: "NousResearch/Hermes-4-3B"
+train: true
+
+# Use the curated DPO preference pairs dataset
+data: "data/"
+
+# Output adapter configuration
+adapter_path: "adapters/dpo_3b_adapter"
+save_every: 200
+
+# DPO parameters
+loss: "dpo"
+iters: 1000
+batch_size: 2
+lora_layers: 16
+learning_rate: 1e-5
+lora_parameters:
+  keys: ['q_proj', 'v_proj']

training/configs/dpo_8b.yaml

@@ -0,0 +1,21 @@
+# MLX DPO Training Configuration for Hermes 4 (8B Class)
+# Optimized for Apple Silicon execution (daily driver capability).
+
+model: "mlx-community/Hermes-3-Llama-3.1-8B-4bit"
+train: true
+
+# Use the curated DPO preference pairs dataset
+data: "autolora/data/dpo/"
+
+# Output adapter configuration
+adapter_path: "autolora/adapters/dpo-8b-adapter"
+save_every: 200
+
+# DPO parameters
+loss: "dpo"
+iters: 1000
+batch_size: 2
+lora_layers: 16
+learning_rate: 1e-5
+lora_parameters:
+  keys: ['q_proj', 'v_proj']