ship-loop-master

name	ship-loop-master
description	24/7 ~1.5-day autonomous shipping orchestrator for Synthex. Drives 7 child loops (build, test, sanity, watch, smoke, pr, merge) on differentiated cadences. Reads state from .claude/scratchpad/ship-loop-state.json on every tick, dispatches the next eligible child loop, and exits cleanly after 36 hours OR when state.completion === 'merged'. Use when the CEO says "/loop ship-loop-master" to start an autonomous shipping run.
type	orchestrator
context	persistent

ship-loop-master — autonomous shipping orchestrator

Activation

CEO invokes /loop ship-loop-master (default) or /loop ship-loop-master --enable-auto-merge (auto-merge mode).

The literal phrase --enable-auto-merge in the activation prompt is the only way to set auto_merge: true. Without it the loop forces auto_merge: false regardless of any state file content. This is the audit trail.

Pre-flight (runs once at first tick)

1. Verify git status is on a feature branch (refuse to run on main)
2. Verify gh auth status shows authenticated session
3. Verify a Linear ticket reference exists in env SHIP_LOOP_TICKET=SYN-XXX OR in .claude/scratchpad/ship-loop-ticket.txt (refuse to start without one — CLAUDE.md hard rule "all work traces to Linear")
4. Initialise .claude/scratchpad/ship-loop-state.json if absent:

   {
     "started_at": "<iso>",
     "deadline_at": "<iso + 36h>",
     "ticket": "SYN-XXX",
     "branch": "<current branch>",
     "auto_merge": false,
     "completion": "in_progress",
     "layers": {
       "build": { "state": "unknown", "last_run": null, "retries": 0 },
       "test": { "state": "unknown", "last_run": null, "retries": 0 },
       "sanity": { "state": "unknown", "last_run": null, "retries": 0 },
       "watch": { "state": "unknown", "last_run": null, "retries": 0 },
       "smoke": { "state": "unknown", "last_run": null, "retries": 0 },
       "pr": { "state": "not_opened", "url": null, "ci_state": null },
       "merge": { "state": "not_attempted" }
     }
   }
   

5. Initialise ship-loop-events.jsonl (touch) and ship-loop-escalations.md (touch with header)

Tick loop (every dynamic-paced wakeup)

Read ship-loop-state.json. Then:

Hard stops (highest priority, check first)

• Deadline reached: now() >= deadline_at → set completion = 'timeout', write summary to escalations.md, exit
• Completion state: completion === 'merged' → exit cleanly with summary
• Unresolved P0 escalation: parse escalations.md for ^## P0 headings; if any remain unchecked, refuse to invoke any child loop, write current status, sleep 1800s and re-check

Eligibility decisions (in priority order)

The master picks ONE child to invoke per tick (not parallel — sequential discipline keeps debugging tractable):

If state	Then invoke
layers.build.state !== 'green' OR last_run > 30 min ago	ship-loop-build
layers.build.state === 'green' AND (layers.test.state !== 'green' OR last_run > 30 min ago)	ship-loop-test
build+test green AND layers.sanity.last_run > 60 min ago OR sanity.state !== 'green'	ship-loop-sanity
any layer green AND layers.watch.last_run > 15 min ago	ship-loop-watch
build+test+sanity green AND layers.smoke.last_run > 60 min ago OR smoke.state !== 'green'	ship-loop-smoke
all 5 inner layers green AND git status --short shows staged/unstaged changes AND pr.state === 'not_opened'	ship-loop-pr
pr.state === 'opened' AND pr.ci_state === 'green' AND merge.state === 'not_attempted'	ship-loop-merge
nothing eligible	sleep (longer cache-miss interval)

After invoking the child

• Read child's exit state from ship-loop-state.json (children update their layer's section atomically)
• Append a tick entry to ship-loop-events.jsonl:

  {"ts":"<iso>","tick":N,"invoked":"ship-loop-build","result":"green","duration_ms":12345}
  

• Decide next sleep duration via ScheduleWakeup:
  • Active work (something just changed state, more eligible loops queued): 60–270s (in-cache)
  • Idle (everything green, just monitoring): 1200–1800s (cache miss but acceptable)
  • Approaching deadline: shorter intervals to maximise final-hour work

Hard rules compliance

• Linear ticket required — refuses to start without SHIP_LOOP_TICKET
• Branch protection — refuses to run on main
• Auto-merge gating — defaults false; only flips on if literal --enable-auto-merge was in activation prompt; recorded in state for audit
• No env modifications — never reads/writes .env*
• Max 1 retry per child loop per tick — child loops handle their own recovery; master does not retry beyond what child reports
• Phase 8 stays human-gated — even with auto-merge enabled, master never invokes Vercel CLI or pushes to main directly

Exit conditions

Condition	Action
completion === 'merged'	Write summary to escalations.md (positive), exit clean
Deadline reached	Mark completion = 'timeout', summary, exit
P0 escalation written by any child	Halt; require human to clear escalation before resume
Activation prompt missing --enable-auto-merge AND merge layer reaches ready_for_human	Halt with "ready for human review" message; do not exit (loop sleeps so CEO can resume by clearing the escalation)

Files this skill writes

• .claude/scratchpad/ship-loop-state.json (read + write atomically)
• .claude/scratchpad/ship-loop-events.jsonl (append-only)
• .claude/scratchpad/ship-loop-escalations.md (append-only)

Verification

/loop ship-loop-master on a branch with one trivially failing test should:

1. Tick 1: invoke ship-loop-build (state was unknown)
2. Tick 2: invoke ship-loop-test (build went green)
3. Tick 3: child reports red with one fail; master logs and re-queues
4. Eventually: escalations.md has the test failure escalated after retry exhaustion

Out of scope

• Production deploy invocation (Vercel auto-deploys on merge to main; loop does not call Vercel CLI)
• Cross-repo coordination
• Slack/email notifications (escalations.md is the surface)