mcp-supervisor-squad 0.1.26

MCP server for the supervisor-squad workflow.

mcp-supervisor-squad

Python MCP server that implements a minimal “director + (deterministic) planner + coder” workflow with brief status.

Tools (public MCP API)

• squad_start(cwd, query, options?) → creates a run, executes a two-step planner (light scan → summary, then heavy scan: plan_scan → exec_scan → write_requirements with tool-collected evidence), then dispatches coder jobs. Returns event-driven brief status (summary ≤ 3 lines + forced_wait_ms delay).
• squad_status(run_id, options?) → returns brief status (lines ≤ 3) using cached state (no stdout tail by default).
• squad_apply(run_id, options?) → applies the final patch with git apply (dry_run default true) and returns the same brief status fields plus apply result fields.

Important: for deterministic run lookup, pass options.cwd (any path inside the target repo) to squad_status and squad_apply. Authoritative machine-readable contract is written to <repo_root>/.codex/supervisor-squad/contract.json.

External directors should treat contract.json as the stable integration surface: inject optional kb/inbox/director_report.*, call squad_start, then poll squad_status and read persisted artifacts under .codex/supervisor-squad/ as needed. If you detect the plan is wrong (e.g., during the ~15s review window), cancel with squad_status(run_id, options={"cwd":"...","cancel":true}) and start a new run.

All tool responses include:

{
  "run_id": "ss-...",
  "next_action": "wait|apply",
  "forced_wait_ms": 15000,
  "wait_reason": "agent_running|ready_to_apply|canceled|error",
  "has_update": true|false,
  "event_id": 3,
  "summary": ["...<=3 lines..."] // when has_update=false this is a short forced-wait notice
}

squad_apply additionally includes:

{
  "applied": true,
  "dry_run": true,
  "summary": "files=3",
  "error": ""
}

Storage layout (repo-local)

Under <repo_root>/.codex/supervisor-squad/:

• contract.json – authoritative machine-readable contract (paths/limits/schema/workflow)
• config.json – repo-local defaults for execution (models/reasoning/sandbox)
• kb/inbox/director_report.md|json – optional director “research injection”
• kb/compact.md – compacted shared KB (kept small)
• kb/archive/* – archived injections (pruned)
• runs/<run_id>/state.json – run state and brief cache
• runs/<run_id>/artifacts/requirements*.md – requirements docs written before dispatch
• runs/<run_id>/artifacts/contract.json – dual-project contract (<=4KB)
• runs/<run_id>/artifacts/patch.diff – final patch to apply
• runs/<run_id>/jobs/<job_id>/* – codex exec stdout/stderr/meta
• worktrees/<run_id>/<slug> – per-task git worktrees (best-effort cleanup)
• kb/prune_meta.json – prune rate limit state
• kb/repo_profile.json – best-effort repo scan cache (reused across runs)

Configuration

Create <repo_root>/.codex/supervisor-squad/config.json to set repo-local defaults. Call-time options overrides these values.

Example:

{
  "execution": {
    "sandbox": "workspace-write",
    "approval_policy": "never"
  },
  "planner": {
    "model": "gpt-5.1-codex-max",
    "reasoning_effort": "medium",
    "extra_config": []
  },
  "coder": {
    "model": "gpt-5.1-codex-max",
    "reasoning_effort": "medium",
    "extra_config": []
  },
  "read_limits": {
    "max_files": 30,
    "max_file_bytes": 65536,
    "max_total_bytes": 1000000
  },
  "polling": {
    "max_ms": 60000,
    "jitter_ratio": 0.1
  },
  "kb": {
    "inbox_max_kb": 20,
    "compact_max_kb": 10,
    "archive_keep": 5,
    "archive_max_kb": 200
  }
}

Director Injection (Optional)

External director (outside this MCP server) may write a short report to:

• <repo_root>/.codex/supervisor-squad/kb/inbox/director_report.md (<=20KB), or
• <repo_root>/.codex/supervisor-squad/kb/inbox/director_report.json (<=20KB)

On squad_start, the server reads it, performs a light scan to surface a quick summary, then runs the heavy scan/requirements + shared KB from a deterministic plan + evidence, then archives and clears the inbox.

Development

cd mcp-tools/mcp-supervisor-squad
poetry install

Running

poetry run mcp-supervisor-squad --transport stdio

Testing

python -m compileall src/mcp_supervisor_squad