autopilot

/autopilot: autonomous subagent orchestrator

Drives a single subagent (pi / opencode / codex / cursor) toward a goal without user-driven polling. Wraps the existing dispatchers at ~/.claude/skills/<name>/bin/<name>-agent and adds two wake sources, a checkpoint protocol, and a final review gate.

This skill is the autonomous counterpart to /pi, /opencode, /codex, /cursor (which forbid auto-polling). When operating inside /autopilot, you are explicitly authorized to auto-poll those same dispatchers.

Invocation

/autopilot <pi|opencode|codex|cursor> <task description>

• First token after /autopilot selects the subagent: one of pi, opencode, codex, cursor. Reject anything else.
• Everything after the subagent token is the task.

Session naming and files

Session: cc-<subagent>-autopilot (e.g. cc-pi-autopilot). Distinct from the user-driven cc-<subagent> sessions so the two don't collide. Pass it explicitly: <dispatcher> --session cc-<subagent>-autopilot <verb>.

Fresh session per task. If a session with this name already exists when /autopilot is invoked fresh, stop it first.

Three files per run:

File	Purpose
/tmp/cc-autopilot-<subagent>-task.md	the full task, so the subagent can re-read it
/tmp/cc-autopilot-<subagent>-decisions.md	the decision log; the single artifact the user reads when they return
/tmp/cc-autopilot-<subagent>.done	the done-sentinel; its presence means the loop has ended

The control flow

Every turn begins by checking the sentinel, then routing on the wake source. The full loop:

EVERY WAKE:
  sentinel present?  ──yes──▶ stale wake; one-line note; END turn (never restart)
        │ no
        ▼
  session absent?    ──yes──▶ PHASE 1 (start), then END turn
        │ no (a wake-up)
        ▼
  route on wake source:
    wait ping = idle/completion ──────────────▶ PHASE 3 (review)
    wait ping = timeout (still busy) ──────────▶ re-arm wait; END turn
    120s report tick ─────────────────────────▶ EVALUATE
    60s permission re-check / 10-min tick ─────▶ CHECKPOINT TICK

CHECKPOINT TICK (capture pane, classify one state):
    idle        ──▶ PHASE 3 (review)
    permission  ──▶ answer prompt; schedule 60s re-check; wait stays armed; END turn
    crashed     ──▶ BAIL (summary, stop, set sentinel, surface)
    busy        ──▶ kill wait; Escape + confirm; request 3-para report;
                    schedule 120s tick; END turn

EVALUATE (read the report):
    ground-truth with git diff; harvest DECISION lines
    direction OK?  ──yes──▶ send bare "continue" ──▶ RE-ARM
                   ──no───▶ inspect myself (Read/Bash/Grep)
                            same blocker 2nd round?  ──yes──▶ BAIL
                                                     ──no───▶ send "continue" + concrete fix;
                                                              log intervention ──▶ RE-ARM

PHASE 3 REVIEW (agent idle):
    git diff + run tests myself; harvest DECISION lines
    good?  ──yes──▶ summary to log; stop session; set sentinel; END loop (no re-arm)
           ──no───▶ send numbered fixes; log verdict ──▶ RE-ARM

RE-ARM: start fresh bg wait; schedule 10-min timer (subject to the one-pending-timer invariant)