Autonomous agent orchestration plugin for Claude Code
npx claudepluginhub steadymoka/tarsAutonomous agent orchestration plugin — self-evolving .tars/ protocol with heartbeat cycles, retrospectives, and protocol evolution
English | 한국어
tars
Autonomous agent orchestration plugin for Claude Code.
The agent self-governs through a .tars/ protocol and evolves its own process over time.
/plugin marketplace add steadymoka/tars
/plugin install tars@steadymoka-tars
/tars:init # scaffold .tars/, choose core & execution mode
/tars:heartbeat # run one autonomous cycle
/loop 10m /tars:heartbeat # auto-repeat every 10 minutes
For fully autonomous operation, start Claude Code with --dangerously-skip-permissions:
claude --dangerously-skip-permissions
The guard.py hook (installed by /tars:init) blocks file modifications and dangerous Bash commands outside the project directory, so autonomous operation remains safe within your project scope.
| Skill | Description |
|---|---|
/tars:init | Initialize .tars/ — choose core (dev/design/...) and execution mode (solo/team) |
/tars:heartbeat | Read PROTOCOL.md and execute one autonomous cycle |
/tars:status | Dashboard — heartbeat count, backlog depth, anomaly flags |
/tars:retro | Immediate retrospective — metrics, anomaly detection, protocol improvement proposals |
/tars:evolve | Protocol evolution analysis — patch effectiveness, lateral thinking options |
/tars:upgrade | Upgrade existing .tars/ to match current plugin version |
/tars:team | Add team agent configuration to an existing solo project |
A core defines how the agent operates — its protocol, mission template, and definition of done. Cores are the pluggable unit of tars.
| Core | Description |
|---|---|
dev | Build verification, code patrol, testing strategy. 3-level DoD (build → functional → integration) |
design | Design iteration, artifact management, feedback loops. Output structure with specs/wireframes/research |
Adding a community core is as simple as copying a directory:
cores/
├── dev/ ← built-in
├── design/ ← built-in
├── research/ ← copy a community core here
│ ├── core.json
│ ├── PROTOCOL.md
│ └── MISSION.md
└── shared/ ← reserved (common templates)
A core directory needs 3 files:
core.json — metadata (name, label, description, placeholders)PROTOCOL.md — heartbeat cycle and operational rulesMISSION.md — mission template with DoD and proactive work tiersOne agent, one core. The agent follows its core's protocol autonomously.
/tars:init → choose core → solo
An orchestrator coordinates multiple specialized agents, each powered by a different core.
/tars:init → choose cores → team
Team mode creates:
.claude/agents/orchestrator.md — coordinates task distribution.claude/agents/{core}-agent.md — specialized agent per core.tars/_workspace/ — shared artifact exchange between agentsYou can also convert a solo project to team later:
/tars:team # add team agents to an existing solo project
Heartbeat Cycle (each cycle):
Read State (no cache) → Process COMMS → CLARIFY if needed → Execute BACKLOG → Write LOG → RETRO Check
Self-Evolution (improvements found in RETRO):
Stagnation Detection → Oscillation Detection → Regression Detection
→ Wonder Gap Analysis → Protocol Patch apply/propose
.tars/
├── COMMS.md # Human <> Agent communication
├── BACKLOG.md # Task list (P0~P3 priority)
├── PROGRESS.md # Current heartbeat count, task state
├── MILESTONES.md # Mid-term goal checkpoints
├── protocol/
│ ├── MISSION.md # Mission definition, DoD, Proactive Work
│ ├── PROTOCOL.md # Autonomous operation protocol (self-evolution target)
│ └── EVOLUTION.md # Protocol change history
├── logs/
│ ├── LOG.md # Work history
│ ├── RETRO.md # Retrospective records
│ ├── archive/
│ └── backlog-archive/
├── _workspace/ # Team mode: agent artifact exchange
└── output/ # Design core: specs, wireframes, research
Every 10 heartbeats, a retrospective runs automatically. When it discovers process improvements, it patches PROTOCOL.md directly.