AI-LongTerm Coordination — A reusable framework for long-horizon AI collaboration.
Turn multi-session AI work from chaotic handoffs into a stable, verifiable operating system.
Most AI coding workflows burn expensive models continuously, lose context between sessions, and have no recovery path when things go wrong. AI-LTC solves this with:
- Staged model分工: GPT designs architecture, Qwen executes day-to-day, GPT returns only for audits
- File-based state: Context lives in
.ai/state.json, not in conversation history - Kernel-verified transitions: Every phase change validated against formal state machine rules
- Error recovery: 6 error types with defined detection and recovery strategies
- Multi-session orchestration: Parallel independent sessions via OpenCode
task()
# Clone AI-LTC into your project
git clone https://github.com/Hope2333/AI-LTC.git .ai/AI-LTC
# Copy the runtime template
cp -r .ai/AI-LTC/.ai-template/* .ai/
# Configure
edit .ai/system/ai-ltc-config.jsonApply qwen-init-routing.prompt.md to your AI operator. It will:
- Classify your project state (greenfield / midstream / chaotic)
- Write resolver config
- Recommend the next model and prompt
- Normal execution:
qwen-generalist-autopilot.prompt.md - Architecture bootstrap:
gpt-bootstrap-architect.prompt.md - Review/checkpoint:
qwen-supervisory-generalist.prompt.md
cd examples/demo-cli
python main.py greet --name Alice
python main.py wordcount hello world from AI-LTC
python -m pytest tests/test_main.py -v # 8 tests, all passing| File | Purpose |
|---|---|
kernel/state_schema.json |
SSOT — the only valid state structure |
kernel/control.yaml |
Authority chain: who can write what |
kernel/state_machine.yaml |
Legal phase transitions |
kernel/error_model.yaml |
6 error types + recovery strategies |
kernel/arbitration.yaml |
Conflict resolution when agents disagree |
| File | Purpose |
|---|---|
.ai/state.json |
Current runtime state (SSOT) |
.ai/system/ai-ltc-config.json |
Resolver config, model routing, language policy |
.ai/logs/ |
Decision, state, and error logs |
.ai/history/snapshots/ |
State snapshots for rollback |
| Role | Prompt | When to Use |
|---|---|---|
| Architect | gpt-bootstrap-architect.prompt.md |
Initial design, skeleton setup |
| Generalist | qwen-generalist-autopilot.prompt.md |
Day-to-day execution (default) |
| Supervisor | qwen-supervisory-generalist.prompt.md |
Checkpoints, sequencing |
| Strategist | gpt-corrective-strategist.prompt.md |
Architecture drift, long-range replanning |
| Optimizer | gpt-optimizer-auditor.prompt.md |
Narrow audits, hard blockers |
INIT → HANDOFF_READY → EXECUTION → REVIEW → OPTIMIZER → EXECUTION
│ │
└──────────────────────────────────────────────┘
↓
CHECKPOINT → (new batch or close)
Each transition is validated against kernel/state_machine.yaml. Illegal transitions are rejected.
AI-LTC integrates with oh-my-litecode (OML) via a thin adapter bridge:
- AI-LTC = Brain: State machine, memory, error recovery, cross-repo sync
- OML = Body: Plugin loading, MCP gateway, session management, worker pool, hooks engine
Architecture: docs/OML-BRIDGE-ARCHITECTURE.md
Integration plan: docs/OML-INTEGRATION-PLAN.md
Platform adapters: docs/OML-PLUGIN-ADAPTER.md
Design principles: docs/BRAIN-BODY-SEPARATION.md
| Tag | What Changed |
|---|---|
v1.5.3 |
Kernel v0.1 + Runtime v0.1 + Demo CLI + public README rewrite |
v1.5.4 |
Branch governance + benchmark framework + multi-session config |
v1.5.5 |
Context overflow, circuit breakers, transition hooks, memory system, cross-repo management |
v1.5.6 |
Code quality gate: thresholds, regression rules, execution integration |
v1.5.7 |
enve-derived templates: multi-distro CI, packaging workflows, version adaptation, execution prompts, dependency ledger, manual testing |
v1.5.11 |
OML bridge integration: bridge layer, platform adapters (OpenCode, Claude Code, Aider), memory/context bridge, deployment scripts |
v1.5.12 |
Enhanced task system (priority, tags, QA, evidence, timestamps), unified security model (hash chain, audit trail, secret detection, tamper detection, atomic writes) |
v1.5.13 |
enve-derived templates: cross-CLI adapter architecture, OML Core spec, AI-LTC integration plan |
v1.5.14 |
Upstream throttle retry (OpenCode Zen Alibaba routing), line rename to Qwen3.6-Plus-WITH-OMO, AI-LTC Todo Tasks document |
v1.5.15 |
Reasoning efficiency kernel: Caveman Compression, Chain-of-Draft (Zoom), Think Deep Not Just Long (Google), Headroom. Prompts moved to prompts/. Intuition file system. |
AI-LTC/
├── kernel/ # Formal kernel (rules, schemas, state machine)
├── adapters/ # Model-specific adapters (preview only)
│ ├── qwen36/ # Qwen 3.6 Plus Preview adapter
│ ├── opencode/ # OpenCode plugin adapter
│ ├── claude-code/ # Claude Code adapter
│ ├── aider/ # Aider adapter
│ ├── registry.ts # Platform adapter registry
│ └── types.ts # Shared adapter types
├── bridge/ # OML integration bridge layer
│ ├── index.ts # Bridge entry point
│ ├── oml-bridge.ts # Core bridge logic
│ ├── event-map.yaml # Event mapping table
│ ├── capability-registry.ts # Plugin capability registry
│ ├── memory-adapter.ts # Memory bridge
│ ├── context-compact.ts # Context compaction
│ ├── cross-session.ts # Cross-session sharing
│ └── protocol.md # Task/result protocol
├── .ai-template/ # Runtime template (copy to .ai/ in target projects)
├── examples/
│ ├── demo-cli/ # Minimum runnable demo (8 tests passing)
│ ├── collaboration-system/ # Copyable collaboration template
│ └── benchmark/ # Cross-model comparison tasks
├── scripts/ # Validators and tools
├── prompts/ # All agent role prompts
│ ├── qwen-*.prompt.md # Qwen role prompts
│ └── gpt-*.prompt.md # GPT role prompts
├── kernel/reasoning-policy.yaml # Reasoning efficiency rules (Caveman, CoD, DTR, Headroom)
├── PROMPTS.md # Root prompt guide (minimal)
├── BRANCH-GOVERNANCE.md # Dual-branch responsibilities and merge rule
MIT. See LICENSE.
Small, focused improvements preferred. See CONTRIBUTING.md.
- Issues: GitHub Issues
- Discussions: Recommended for questions and ideas
- Discord / WeChat: Not configured yet