Build

Turn a child's locked failing tests green without touching them, prove the change on the running app, and hand back a pull request a human can check in minutes.

Origin: TDFlow (human-owned tests as the oracle, 94.3% vs 68.0%), ImpossibleBench (read-only tests plus "stop and say so" took cheating from 54% to 9%) and Anthropic's long-running harness (fresh sessions carried by commits; end-to-end checks before "done").

This session is the Delegator: it holds the spec, the unit list and the results, never a diff, so it stays small however big the child is. Every unit is a fresh subagent, one at a time; one writer per checkout. The briefs are in references/briefs.md.

Steps

1. Read the ticket, whose body is the spec, and the test: lock commit on feat/<capability>. Bring the branch up to date with the base — merge it in once the branch is pushed, never rebase — then run the lock check and the locked tests: every one red for the reason tester reported. Anything green, or red for another reason, stops the build.
2. Fit the loop from the repo, not memory: typecheck, lint, the locked command, the full suite and the spec's ## Verify start command. Run each once at the base. A missing command is recorded as absent; a missing service or credential is a stop (step 6).
3. Plan units. Group the locked tests into units a fresh child can turn green, ordered so each starts from the last one's green tree. A child that needs more than a few units goes back to /cut.
4. Build each unit with a fresh builder. It commits once, on a real green, and hands back the commit and the command's output. Check that output yourself before the next unit.
5. Vet and verify, both in fresh subagents and neither the builder:
   • /crucible over <lock-sha>...HEAD with the ticket. Delete the scratch tests on its prune list in one commit, test: prune scratch tests.
   • The verifier with the spec's ## Verify steps only, never the diff or the tests.