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.

Install

/plugin marketplace add steadymoka/tars
/plugin install tars@steadymoka-tars

Quick Start

/tars:init                       # scaffold .tars/, choose core & execution mode
/tars:heartbeat                  # run one autonomous cycle
/loop 10m /tars:heartbeat        # auto-repeat every 10 minutes

Autonomous Mode

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.

Skills

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

Cores

A core defines how the agent operates — its protocol, mission template, and definition of done. Cores are the pluggable unit of tars.

Built-in Cores

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

Community Cores

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 rules
• MISSION.md — mission template with DoD and proactive work tiers

Execution Modes

Solo (default)

One agent, one core. The agent follows its core's protocol autonomously.

/tars:init → choose core → solo

Team

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 agents

You can also convert a solo project to team later:

/tars:team    # add team agents to an existing solo project

How It Works

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/ Structure

.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

Self-Evolution

Every 10 heartbeats, a retrospective runs automatically. When it discovers process improvements, it patches PROTOCOL.md directly.

Type	Description	How Applied
Parameter Patch	Threshold/count adjustments	Direct PROTOCOL.md edit
Structural Patch	Add/remove steps or rules	Propose via COMMS → auto-approve
Question	Challenge process assumptions	Escalate to patch after 3 repeats
Revert	Metrics worsened after patch	Auto-detect + rollback

Anomaly Detection

• Stagnation: Same task stuck for 3+ heartbeats, or 3 consecutive idle cycles
• Oscillation: Same patch applied → reverted → applied again (A→B→A pattern)
• Regression: Metrics worsened after a protocol patch
• Wonder: Socratic questioning of process assumptions

Safety Rails

• Cannot lower Definition of Done levels
• Cannot remove Proactive Work tiers
• Cannot extend RETRO intervals
• All three require explicit human approval to change

Guard Hook

The guard hook (guard.py) protects autonomous operation:

• File writes: Blocks Write/Edit/NotebookEdit outside the project directory
• Bash commands: Blocks dangerous commands (rm, mv, cp, dd, etc.) targeting outside the project
• Destructive git: Blocks git reset --hard, git push --force, git clean -fd, git checkout -- .
• Redirects: Blocks >, >>, tee writing to absolute paths outside the project
• sudo: Strips sudo prefix to detect the real command underneath

License

MIT