record-evidence

Record acceptance-criterion evidence

Evidence is the bridge between "I claim it works" and "it is demonstrably done". Record one true, specific row per acceptance criterion, plus the smallest-change note, then stop. You do not submit, push or open a PR: the runner runs the gates, records the delivery and submits for review; it pushes and opens a PR only when the operator has enabled that.

Steps (delivery agent)

1. Commit first. Evidence describes committed work: git status --porcelain must be empty apart from runner-excluded files, and git log -1 --oneline shows your deliver #N: … commit. Note the short SHA for your summaries.
2. List the ACs. get_ticket → acceptance_criteria: each id, its text, and whether it carries a check_command.
3. For each AC, find the proof that exercises THAT AC's own behaviour. Name the test file and test name, and say what it asserts. A green suite is not evidence for an AC unless a test in it would fail if the AC were not met. If no such test exists, write it now (and re-run) — do not paper over the gap; reviewers reject an AC without one. For an AC with a check_command, run that exact command yourself first: the runner runs it after you stop, a failure sends the delivery back, and only the runner's own check result can satisfy that AC — your row for it is supporting evidence only.
4. Record one row per AC with record_ac_evidence: ticket_id, ac_id, evidence_type, and a summary under 5,000 characters. Pick the type that matches the proof: test_output (command + pass summary + the named test), coverage_report, static_analysis (lint/typecheck), diff_summary, commit, screenshot, log, ci_run (a CI run you actually observed), manual_note (a verified observation no command could capture). No claim_token is needed; the runner injects it. Recording with an ac_id marks that AC satisfied (an AC with a check_command is still not done until the runner's own check passes), so only record what is true. A good summary: "pnpm test src/lock.test.ts → 14 passed. two concurrent saves both persist fires 2 writers via Promise.all and asserts both records exist (AC2). abc123f".
5. Record the smallest-change note — one manual_note row with NO ac_id, whose summary starts smallest-change check: and names the files you changed by file name, what you cut or refused, and why this size is the floor. The runner scans evidence for the phrase "smallest-change" (or "smallest change"), uses the MOST RECENT such row (its own needs_human_review flags excluded), and flags one that names none of the changed files (by file name or a stem of four or more characters) as boilerplate. Record it once, after your final edit. This is the same line the self-review skill asks for: if you already recorded it in this run, do not repeat it. On a rework or resume, record a fresh note for the diff as it now stands — the newest note is the one the runner reads (on a resume the call is refused; see below). See the minimalism skill.
6. Record the digest delta with the prepare-digest-delta skill.
7. Stop. Once every AC has a row, the note and the delta are recorded, you are done. Do not call submit_ticket_for_review, push, open a PR, or keep polishing.