{
  "name": "launch-control",
  "version": "1.0.0",
  "description": "Plan-as-contract implementation loopflow: context, block plan, builder/reviewer feedback loop, and final audit with real validation evidence.",
  "backend": "pi-subprocess",
  "defaults": {
    "agentScope": "both"
  },
  "steps": [
    {
      "id": "context",
      "agent": "builtin-context",
      "tools": [],
      "output": "context.md",
      "task": "Build concise launch-control execution context for this task:\n\n{task}\n\nRead any provided plan/spec path if one is explicitly named. For new standalone folder/project tasks, do NOT grep/find/scan the whole workspace; only check whether the target path exists, list the immediate parent directory if needed, and inspect environment versions needed for validation. Always ignore .pi, previous loopflow runs, node_modules, .venv, caches, and unrelated projects. Do not edit files. Return compact handoff context with: scope, likely files, validation contract candidates, explicit unknowns, and stop conditions."
    },
    {
      "id": "plan",
      "agent": "planner",
      "tools": ["read", "ls"],
      "output": "block-plan.md",
      "task": "Create the smallest safe launch-control implementation block from this context:\n\n{outputs.context}\n\nOriginal task:\n{task}\n\nReturn a concrete block plan with: intended slice, acceptance criteria, files likely involved, real validation commands/evidence required, non-goals, stop conditions, and what the builder must report. Keep the plan lean and executable. Do not edit files."
    },
    {
      "loop": {
        "id": "build-review",
        "maxIterations": 3,
        "gateStep": "review",
        "passStatuses": ["approved"],
        "retryStatuses": ["changes_requested"],
        "stopStatuses": ["blocked"],
        "onExhausted": "stop",
        "body": [
          {
            "id": "build",
            "agent": "worker",
            "tools": ["read", "grep", "find", "ls", "bash", "edit", "write"],
            "output": "build-{loop.iteration}.md",
            "task": "Implement or fix launch-control block iteration {loop.iteration}.\n\nOriginal task:\n{task}\n\nBlock plan:\n{outputs.plan}\n\nPrevious review, if any:\n{outputs.review.output}\n\nRules:\n- One writer only.\n- Use the real project stack and real integrations.\n- No mocks, demos, fallbacks, fake validation, or hidden shortcuts unless the user explicitly asked.\n- If this is iteration 1, implement exactly the block plan. If this is a later iteration, apply only required reviewer fixes.\n- If the plan is stale, acceptance criteria change, data/API model changes, or user decision is required, stop and report it instead of inventing scope.\n- Run the validation contract where possible.\n- Prefer small incremental file writes/edits. Avoid one huge write payload for many files or very large files.\n- For new multi-file projects, create directories first, then write one compact file at a time, validating syntax as soon as possible.\n- When validating a local server, avoid fixed-port conflicts: clean up any PID you start, check whether the chosen port is already listening before launch, or use an available alternate port and report it.\n- Do not let exploratory validation commands abort the whole agent before the final report. Capture command exit codes/output into evidence, clean up processes, then return the report with any failures/blockers explicitly. Avoid long `&&` chains for smoke tests unless every failure is intentionally terminal.\n\nReturn changed files, commands run with exit codes, validation evidence, blockers, and any plan mismatch."
          },
          {
            "id": "review",
            "agent": "reviewer",
            "tools": ["read", "grep", "find", "ls", "bash"],
            "output": "review-{loop.iteration}.json",
            "gate": {
              "type": "json-status",
              "passStatuses": ["approved"],
              "retryStatuses": ["changes_requested"],
              "stopStatuses": ["blocked"]
            },
            "task": "Review launch-control build iteration {loop.iteration}. Do not edit project/source files.\n\nOriginal task:\n{task}\n\nBlock plan:\n{outputs.plan}\n\nBuilder report:\n{outputs.build.output}\n\nCheck: plan freshness, scope drift, correctness, real validation evidence, test coverage, hidden mocks/fallbacks/demos, over-large slice, and unlogged plan changes.\n\nReturn ONLY valid JSON matching this shape:\n{\n  \"status\": \"approved\" | \"changes_requested\" | \"blocked\",\n  \"summary\": \"short verdict\",\n  \"findings\": [{\"severity\": \"blocker|required|optional\", \"file\": \"path or null\", \"issue\": \"specific issue\", \"required_fix\": \"specific fix or null\"}],\n  \"validation_gaps\": [\"missing evidence\"],\n  \"plan_changed\": false,\n  \"requires_user_decision\": false\n}\n\nUse status=approved only when the diff and evidence satisfy the block plan. Use changes_requested for fixable required issues. Use blocked for missing credentials, stale plan, required user/product decision, or exhausted/unsafe scope."
          }
        ]
      }
    },
    {
      "id": "final-audit",
      "agent": "reviewer",
      "tools": ["read", "grep", "find", "ls", "bash"],
      "output": "final-audit.json",
      "gate": {
        "type": "json-status",
        "passStatuses": ["complete", "approved"],
        "stopStatuses": ["incomplete", "blocked"]
      },
      "task": "Run final launch-control completion audit. Do not edit files.\n\nOriginal task:\n{task}\n\nPlan:\n{outputs.plan}\n\nLast build:\n{outputs.build.output}\n\nLast review status: {outputs.review.status}\nLast review JSON:\n{outputs.review.json}\n\nVerify every explicit acceptance criterion, artifact, command, invariant, and stop condition against current repo state and available evidence.\n\nReturn ONLY valid JSON:\n{\n  \"status\": \"complete\" | \"incomplete\" | \"blocked\",\n  \"summary\": \"short audit result\",\n  \"verified_items\": [\"item\"],\n  \"remaining_blockers\": [\"blocker\"],\n  \"validation_gaps\": [\"gap\"]\n}"
    }
  ]
}