fable-agent/CORE/README.md

Core Framework

The markdown spec in this directory maps to the TypeScript engine in src/. This file tells you what each module does, which phase it maps to, and how to use it.

---

Mapping: Markdown Spec → TypeScript Engine

Phase	Markdown (PHASES/)	TypeScript (src/)
1	phase-01-session-engine.md	tier1-foundation/session-engine.ts
2	phase-02-context-manager.md	tier1-foundation/context-manager.ts
3	phase-03-tool-orchestrator.md	tier1-foundation/tool-orchestrator.ts
4	phase-04-feedback-loop.md	tier2-primitives/loops/feedback-loop.ts
5	phase-05-state-accumulator.md	tier2-primitives/loops/state-accumulator.ts
6	phase-06-convergence-check.md	tier2-primitives/loops/convergence-check.ts
7	phase-07-workflow-graph.md	tier2-primitives/workflows/workflow-graph.ts
8	phase-08-adaptive-router.md	tier2-primitives/workflows/adaptive-router.ts
9	phase-09-recovery-handler.md	tier2-primitives/workflows/recovery-handler.ts
10	phase-10-skill-registry.md	tier2-primitives/routines/skill-registry.ts
11	phase-11-execution-engine.md	tier2-primitives/routines/execution-engine.ts
12	phase-12-routine-evolution.md	tier2-primitives/routines/routine-evolution.ts
13	phase-13-state-repository.md	tier3-compounding/state-repository.ts
14	phase-14-skill-sharpener.md	tier3-compounding/skill-sharpener.ts

Upgrades

Module	Markdown	TypeScript (src/upgrades/)
Dreaming System	Dream cycle (every 3 goals)	dreaming-system.ts
Self-Validation	Phase output validation	self-validator.ts
Multi-Agent	Orchestrator → specialists	multi-agent.ts
Rubric Engine	N-criteria evaluation	rubric-engine.ts
Persistent Memory	Episodic/semantic/procedural	persistent-memory.ts
Enhanced Meta-Agent	Dream → Act → Validate → Codify	enhanced-meta-agent.ts
Vision Self-Check	Visual output verification	vision-self-check.ts
Safety Boundary	Model routing + blocked detection	safety-boundary.ts
Routine Scheduler	Cron-like scheduled runs	routine-scheduler.ts
Goal Queue	Persistent priority queue	goal-queue.ts
Daemon Engine	24/7 autonomous operation	daemon-engine.ts
PAI Adapter	PAIMM / TELOS / LifeOS	pai-adapter.ts

Executors

Executor	TypeScript (src/examples/executors/)	Used For
Sonnet	sonnet-executor.ts	Direct Anthropic API
OpenAI	openai-executor.ts	GPT-4.1 / o3
Local	local-executor.ts	Ollama, vLLM, open models
Composite	composite-executor.ts	Route phases to different models
Weak-to-Strong	weak-to-strong.ts	Bootstrap skills from weak models

Usage

The system runs two ways:

1. TypeScript CLI (structured execution)

fable-agent run "<goal>" --loop 10
fable-agent daemon start
fable-agent skills list

2. Agent-driven (agent reads markdown files)

1. Load AGENT.md — understands the system
2. Read CONFIG.md — knows what models are available
3. Read STATE.md — loads current state
4. Read relevant PHASES/ — understands the current phase
5. Read relevant SKILLS/ — applies accumulated knowledge
6. Execute using the TS CLI or direct work
7. Write back to STATE.md and sharpen SKILLS/

Both approaches converge on the same architecture. The engine and the spec are the same system.