| name | build-batch |
| description | Implement a milestone autonomously inside a Conductor worktree without stop-and-ask checkpoints, strictly enforcing scope-first and logging anything out of scope. Use when running parallel Conductor lanes, asked to "build autonomously", "run this in the background", "yolo this milestone", or for low-risk batched UX/tech-debt work. |
Build Batch
Autonomous milestone execution. Same TDD/BDD/validation as /build, but no stop-and-ask checkpoints during the run. Trade interactivity for parallelism.
The autonomous twin of /build. Use inside a Conductor worktree where the user is busy elsewhere — small bets only.
Produces
Same as /build: working code, filled Review Contracts under each task bullet in the milestone file, updated checkbox state, updated .context/todos.md.
Plus: tmp/scope-skipped.md with every out-of-scope item discovered during the run.
Reads
Same as /build:
docs/{feature}/prd.md — scope and tech design
docs/{feature}/milestones/milestone-{n}.md — task bullets in scope
.app-catalog/features/{feature}.md — pages, flows, vocabulary, gotchas (and .app-catalog/README.md for the wider domain/actor map)
If a task bullet is missing BDD criteria, stop and ask — this is the one allowed escalation pre-run.
Steps
0. Pre-flight (the only blocking checkpoint)
Before any code:
- Confirm every task bullet has BDD criteria below it
- Confirm
.context/todos.md exists and matches the milestone bullets
If either fails, halt and ask the user. After this point, no stop-and-ask.
1. Loop (autonomous, scope-first)
For each task bullet in order, follow the /build implementation loop (rules/status-sync.md, rules/tdd.md, rules/bdd.md, rules/visual-check.md).
When anything unexpected comes up, default to log-and-skip per rules/escalation.md:
- Unrelated bug → log to
tmp/scope-skipped.md, continue
- Adjacent refactor → log, continue
- Flaky test in untouched code → log, continue
Only stop if:
- A task cannot be completed at all (genuine block — log in milestone's Issues section, move to next task)
- Discovery contradicts the PRD's Tech Design section (cannot proceed without re-deciding)
2. Self-review
After every task bullet is checked, run validation per /review Phase 1 on yourself: tests, endpoints, visuals, edge cases. Record verdict at the bottom of the milestone file.
If self-review fails, add fix bullets at the end of the milestone file but do not implement them — those return to the human queue.
3. Summary
End with a PR-style summary at the bottom of the milestone file:
- What shipped
- What was logged-and-skipped (and where in
tmp/scope-skipped.md)
- Any genuine blocks
- Self-review verdict and link to evidence
Tell the user: "Batch done. Milestone {id}. Summary at the bottom of the milestone file; skipped items in tmp/scope-skipped.md."
NOT
- Does not run on features with unresolved PRD questions — those go through
/build interactively
- Does not expand scope to fix adjacent issues — log-and-skip, always
- Does not skip validation — the autonomous mode means no human gates, not no gates
- Does not run on task bullets without BDD criteria