Mission-enforced Claude agent runner. You write what "done" looks like as shell commands; swarmd keeps a Claude agent working until those commands pass — across crashes, API outages, and context resets. It doesn't let the agent quit early, game its own check, or drift off task.

You give swarmd two things: a mission (natural language goal) and success criteria (shell commands whose exit codes define done). Everything else — planning, code, commits, decisions — is the agent's job, same as running claude directly.

When you swarm launch mission.yaml, this happens:

1. A Temporal workflow starts on the swarm worker daemon. Its job: run enforcement, not write code.
2. The workflow spawns a claude subprocess in your workspace with the mission prose. Claude does the actual work — files, tools, subagents, whatever it picks.
3. In parallel, the workflow runs a verifier loop every run_every_sec: check tamper (are locked files unmodified?) → enforce invariants (no_mock, test_count_floor, etc.) → run every criterion's shell check in parallel → update state.
4. Three child workflows run alongside:
   • PatternDetector — tails events.jsonl, flags loops, oscillation, scope-shrinking
   • LLMCritic — cadence-driven Haiku calls for progress audit + goal-drift; fires the 6-dim anti-cheat panel (scope_reduction, mock_out, tautology, hardcode, off_criterion, coordinated_edit) on every criterion pass-transition
   • ResourceMonitor — zombie processes, memory pressure, disk
5. When every criterion passes, the workflow enters a hold window. If they stay green for hold_window_sec, a completion judge runs six preconditions (no open cheat/fabrication/tamper findings, no critic disagreements, per-criterion anti-cheat verdict pass) before allowing the transition to complete.
6. Transient errors (HTTP 424/429/5xx, timeouts) become Temporal retries; terminal errors (400/401, auth) halt the mission with a clear reason.

What you specify: mission prose, workspace path, success criteria, optional invariants. What you don't specify: any plan, any steps, any agent behavior. Claude figures that out.

• Temporal server — external dependency (brew install temporal). Persistent state.
• swarm worker — long-running daemon that polls Temporal and executes workflows + activities. Run one or more; more workers = more missions in parallel. Restart at will — state survives.
• Per mission at runtime: 1 parent workflow + 3 child workflows + 1 claude subprocess + up to 6 parallel anti-cheat activities on each pass-transition.

pip install swarmd

Requires Python 3.10+, Temporal (brew install temporal), and claude on PATH.

# mission.yaml
mission: "Add full test coverage to auth.py"
workspace: "/abs/path/to/your/project"
success_criteria:
  - id: tests_pass
    check: "pytest auth/ -q"
    timeout_sec: 120
  - id: coverage_floor
    check: "coverage report --include=auth.py --fail-under=90"
    timeout_sec: 30
  - id: no_mocks        # anti-cheat floor
    check: "! grep -rE 'unittest.mock|MagicMock' auth/"
    timeout_sec: 10
verification:
  run_every_sec: 30
  hold_window_sec: 60

temporal server start-dev &
swarm worker &
swarm launch mission.yaml       # → workflow_id=mission-abc123
swarm status mission-abc123
swarm findings mission-abc123 --tail 50
swarm abort mission-abc123 --reason "criteria were wrong"

• Design spec — full architecture
• Mission schema — every field
• Examples — reference missions

MIT