fable-agent/CONFIG.md

80 lines
2.9 KiB
Markdown

# Configuration
> Global config for the fable-agent system.
> Read by AGENT.md on startup. Edit to change behavior across all sessions.
---
## Model Routing
| Tier | Model | Use Case | Status |
|------|-------|----------|--------|
| Mythos | ~~Fable 5~~ | Orchestration, planning | **BLOCKED** (June 12, 2026) |
| Mythos | ~~Mythos 5~~ | Glasswing-only | **BLOCKED** (June 12, 2026) |
| Opus | Opus 4.8 | Orchestration, vision, planning, routing | Available |
| Sonnet | Sonnet 4.6 | Bounded subtasks, code review, vision | Available |
| Haiku | Haiku | Quick grading, routine edits, file ops | Available |
| Frontier | GPT-4.1 | Alternative orchestrator | Available |
| Frontier | GPT-4o | Vision checks | Available |
| Reasoning | o3-mini | Verification, analysis | Available |
| Local | Local (Ollama/vLLM) | Offline, cheap tasks | Check status |
| Compound | OpenRouter Fusion | Fable-tier compound (draft→critique→fuse) | Available (needs key) |
Routing priority:
1. Never route to blocked models (Fable 5, Mythos 5)
2. Orchestration/planning → Opus 4.8 (or GPT-4.1 fallback)
3. Vision tasks → Opus 4.8 or GPT-4o (vision-capable)
4. Grading/verification → Haiku (cheapest, fast)
5. Code review → Sonnet 4.6
6. Fusion/compound → OpenRouter Fusion (or local panel dispatch)
7. Offline fallback → local model
---
## Data Paths
| Path | Purpose |
|------|---------|
| `~/.fable-agent/` | Persistent state root |
| `~/.fable-agent/sessions/` | Session checkpoints |
| `~/.fable-agent/skills/` | Skill registry (auto-synced with SKILLS/) |
| `~/.fable-agent/knowledge/` | Knowledge base entries |
| `~/.fable-agent/memory/` | Episodic, semantic, procedural memory |
| `~/.fable-agent/safety/` | Model overrides + incident log |
| `~/.fable-agent/scheduler/` | Scheduled task definitions |
| `~/.fable-agent/reviews/` | Skill sharpener reviews |
| `./LOGS/runs/` | Session logs (gitignored) |
---
## Loop Defaults
| Param | Default | Description |
|-------|---------|-------------|
| `maxIterations` | 20 | Max feedback loop iterations per goal |
| `convergenceThreshold` | 0.05 | Min improvement delta to avoid plateau |
| `contextTokenBudget` | 128000 | Max context tokens |
| `dreamInterval` | 3 | Dream every N goals |
| `heartbeatIntervalMs` | 30000 | Session heartbeat frequency |
| `loopDelayMs` | 1000 | Delay between loop iterations |
---
## Self-Validation
| Param | Default | Description |
|-------|---------|-------------|
| `maxRetries` | 2 | Max phase re-executions on validation failure |
| `enforceMinLength` | true | Reject outputs below phase-specific minimums |
| `forbidUncertainty` | true | Reject "maybe", "I don't know", "perhaps" |
---
## Safety
| Param | Default | Description |
|-------|---------|-------------|
| `autoBlockOnAuthError` | true | Block model on 401/403/404 |
| `maxRetriesPerModel` | 3 | Attempts before failing to fallback |
| `preferCheapGrader` | true | Route grading to cheapest available model |