Decision 0042 · Accepted
Outcome-first build
and ship
build and ship are short files whose bodies mostly tell the agent how to run: start every unblocked task at once, one worktree each, prepare that worktree, post a checkpoint, write a failing test first, and review with one worker per axis plus a verifier.
- Status
- Accepted
- Built from:
- da88c48
Status: Accepted (2026-10-28). Amends 0031 (parallel workers and one review worker per axis do not bind build, ship, or the review they run), 0029 (frontier scheduling, and workers testing their own edits while the orchestrator runs the suite), 0028 (ship lands by running build and review as a named pipeline), 0025 (item 8, a failing test before the change), and 0039 (the independent review need not be one worker per axis). Spec: 028. Amended by 0044 (the independent review runs only when the size guard escalates or the user asks; checks once at pull request time). Amended by 0056 (API routes and MCP tools are test-first; no UI tests by default). Amended by 0065 (item 5: other behavior needs evidence, not necessarily a check).
Context
build and ship are short files whose bodies mostly tell the agent how to run: start every unblocked task at once, one worktree each, prepare that worktree, post a checkpoint, write a failing test first, and review with one worker per axis plus a verifier. The outcomes those steps exist to protect are a landed change, a green suite, an independent review, and an honest verification report. Hosts now ship models that can choose a plan. Prescribing the plan spends the instruction on execution, and the test suite checks that the sentences are still present. It does not run an agent.
The catalog stays host-agnostic. A live comparison on Claude Opus 5.5 is how this bet is measured. It is not a requirement that only that model may run the skills.
Decision
- Outcomes, reason, and proof.
buildandshipstate what finished work looks like, why the gate exists, and the evidence that proves it. They do not prescribe workers, scheduling, worktree setup, checkpoint cadence, or test order. - Execution is a choice. Parallel workers, frontier scheduling, one worktree per task, and one review worker per axis are allowed. None is required. Other skills keep 0031 until a later bet changes them.
- The proofs stay mandatory. A green suite or an explicit report that no suite exists; no local merge when no suite exists; an independent review, with requirement verdicts, before spec work or escalated unspecced work lands; a self-check for small unspecced work; the reviewed head, with the shared-file exception in 0033; a red spec head leaves the branch untouched; the acceptance gate, green checks, branch protection, and no self-approval; local
mainreceives the reviewed merge and is pushed only on "ship it"; verification stays 0040. - Landing still includes the work and the review.
shipbrings the change up to those proofs when they are missing. It does not have to invokebuildandreviewas a named pipeline when the proofs are already met. The implementing agent is not the independent review. - Tests prove the behavior, in whatever order. A seam named in a spec or task is covered by a check that fails if the behavior is missing. Writing that check before the change is allowed and not required.
Alternatives
- Shorten the prose and keep the mandates. Rejected: the skills would still tell the agent how to execute, which is the instruction this bet removes.
- Drop the gates with the procedure. Rejected: a mistake on the default branch, a force-push, or a self-approval is costly to reverse, which is why 0029 keeps those rules pinned.
- Apply the same cut to every skill now. Rejected: this is one bet.
explore, install boundaries, and any furtherstarttrim wait.
Consequences
PRODUCT.md,GLOSSARY.md(frontier, checkpoint, area, baseline failure), and the catalog sentences forbuild,review, andshipchange with this spec. Back-pointers on the amended records land when the spec is accepted.- A model may implement a task set in one pass and still pass, as long as the proofs hold.
- Smoke pins that quote the old procedure are rewritten to the proofs. Pins for the safety rules stay.
- One live comparison of no skills, current Flow, and these skills is the measurement. It is not a CI gate, and one run is not an efficiency claim.