conductor-max

Madmax conductor for pi — split a big goal into many tiny slices (up to ~50), run them as parallel pi-subagents workers in worktrees, and let TypeSafe Jev make every routing, review, QA and supervisio

SKILL.md

GitHub

Conductor Max (pi)

Large model plans. Jev decides. Workers execute. Verification checks.

You (the main pi session) plan once and handle escalations. jevo-worker* subagents write code. The jevo tool runs the shared decision CLI; its thresholds and vetoes live in code and fail closed. Inside every jevo-worker* child, this package supervises it:

  • Finish gate: when the worker ends a turn without tool calls, the package reruns the slice gate and asks Jev for QA and a staged review. If the slice isn't done, the reason is queued as a follow-up and the worker keeps going in the same run. After JEVO_MAX_BLOCKS (default 2) rounds, or on any error, the slice is marked escalate.
  • Health: every JEVO_HEALTH_EVERY (default 8) tool calls, Jev checks for stuck or off-track work and steers the worker.

The workers are background children (async: true), because only background children load installed extensions. Jev key: TYPESAFE_API_KEY, else .env in the project, else ~/.pi/agent/secrets/typesafe_api_key.

Tool: jevo

jevo(["slice","new","--id","S3","--goal","...","--acceptance","...","--allow","src/cart.py,tests/test_cart.py","--gate","pytest -q tests/test_cart.py"])
jevo(["triage","--slice","S3"])        # tier -> agent; exit 3 = human; wave serial|parallel
jevo(["slice","list"])                 # status: new / fixing / approved / escalate
jevo(["check","--slice","S3"])         # manual gate -> QA -> review (the finish gate runs it for you)
jevo(["report","--html",".jev-orchestrator/scoreboard.html"])

Exit codes in the JSON: 0 proceed/approve · 1 fix · 3 escalate · 2 error (treat as escalate). /jevo in the TUI shows slice list (or any args you give it).

Loop

  1. Scope + authority. Restate the goal. Quote what the human authorizes. By default workers commit in their worktree only; no push, PR, deploy or ticket write without quoted authority.

  2. Plan once. Build a slice DAG. A slice is one acceptance criterion, ≤ ~3 files, ≤ ~150 changed lines, and has its own fast gate command. slice new each. Commit your plan first: worktree isolation needs a clean checkout.

  3. Triage all slices. Map tiers to agents: fast → jevo-worker-fast, balanced → jevo-worker, reasoning → jevo-worker-reasoning. Exit 3 goes to the human list. wave: serial slices run after the wave.

  4. Dispatch a wave in one subagent call with worktree isolation (cap 12 in flight; raise toward 50 only when slices touch disjoint files):

    subagent({ isolation: "worktree", workflowScript: `
      const r = await runs.all([
        { key: "S1", agent: "jevo-worker-fast", task: "JEV-SLICE: S1\nGoal: ...\nAllowed: a.py, tests/test_a.py\nGate: pytest -q tests/test_a.py" },
        { key: "S2", agent: "jevo-worker", task: "JEV-SLICE: S2\n..." }
      ]);
      return r.map(x => x.output);
    ` })
    

    Every task must start with JEV-SLICE: <id>. That's how the supervisor finds the slice. Put the packet inline, because the slice file isn't visible inside the worktree.

  5. Collect. jevo(["slice","list"]). approved → merge that worker's worktree branch in dependency order. escalate → run jevo-reviewer with the slice id, the ledger reason and the worktree path. VERDICT: fix → a new slice id and a fresh worker; human → the human list.

  6. Integration verify after each wave merges: full test/lint/typecheck. Red → find the culprit merge, revert it, requeue it as a fix slice.

  7. Scoreboard: jevo(["report", ...]). Report only numbers the ledger measured.

Rules

  • A worker's "done" is not evidence. The finish gate and your integration tests are.
  • Never two workers in one checkout. Never let a worker delete files it didn't create.
  • Jev reads its input literally and can be argued with. Code vetoes (allowlist, secrets, failure counts, exit codes) run first, and a Jev approval never pushes, merges to main or deploys.
  • A high escalation rate means the slices are too big. Re-slice instead of loosening thresholds.
  • Agent models are defaults. Override them with pi-subagents subagents.agentOverrides.<name>.model.