build

build

Implement an approved plan, never committing. Argument: a plan <path>, or none: the plan approved in this conversation, else the newest in ${TMPDIR:-/tmp}/plans/<repo>/, then in docs/plans/. A step without Verify uses the repo's test command.

1. Work where the session is. Only if git status shows changes other than the plan file, work in a worktree from HEAD named build/<slug>, <slug> being the plan's file name without its date (the host's worktree tool, else git worktree add), with the dependencies installed so the checks can run.
2. Take the steps in order: make the Change, then run Verify once. On failure, fix and rerun; after three failed runs the step is blocked and every later one skipped.
3. Do the steps yourself. Use subagents only for several heavy steps that share no file and need none of each other: launch them in a single message on the subagent role, each with the plan path, its step id, the rule of step 4 and "touch only its Files, run Verify once, return {exit, note}". A failed result is a failed run of step 2; a subagent that never answers leaves its step unproven.
4. Never: skip, weaken or delete a test, change an expected value to fit, add a suppression comment, widen a type to any, swallow an error, or edit a gate, a CI file or the plan. Once every step has run, read the whole diff (git diff plus new files) and revert any hunk that does, and any change to a file no step lists.

A role comes from .agents/models.json in the repo, else ~/.agents/models.json (the repo wins role by role), under your agent's key (claude, codex…). It is { "model", "effort" } or a model name; pass what it sets to the subagent tool when it takes them, and leave the rest to the session.

Print one line per step, before anything else:

S-001 done 0
S-002 unproven - no answer
S-003 blocked 1 src/a.ts:3 — missing branch
S-004 skipped - needs S-003

The columns are the step, its status, Verify's exit code and a note. Then worktree <path> <branch>, only if you created one. Not all done: stop there.

All done: invoke the verify skill with the plan's absolute path, in the tree you built in. On FAIL, repair what its gate and finding lines name, each finding with a test that shows it, then rerun the failing gate commands, or every gate if none failed: two repair rounds at most, never a second audit. End by printing the last verdict in verify's format: PASS once every rerun exits 0, else FAIL with what still fails.