AGENT SKILL

verdict

The Reviewer's Verdict on a PR: re-run the repo's verification skill at the PR's head, hold the PR's Evidence against it, and record pass or fail with nico-agent.

License 许可
MIT
最近更新
2026-11-10
Stars
★ 0
Forks
0

01 前言 Front Matter

name
verdict
description
The Reviewer's Verdict on a PR: re-run the repo's verification skill at the PR's head, hold the PR's Evidence against it, and record pass or fail with nico-agent.
disable-model-invocation
true

02 资源文件 6

03 SKILL.md 正文

Verdict

You are the Reviewer of one PR. Re-run the repo's verification at the PR's head, hold what you observe against the feature map and the PR's Evidence, and record pass when you drove the verification and it left no Finding, fail otherwise. A mismatch is anything you observe that differs from what a feature file or the PR's Evidence says should happen. Each mismatch becomes a Finding, numbered F1, F2, … afresh per Verdict and named by the head's first seven characters, the short head: <short head> F2.

The invocation reads verdict <prUrl> for <issueUrl>. From it: the issue key (owner/repo#n, or a Linear key such as BLE-12), the id (n, or the Linear key; VERIFY_RUN_ID is the id), the head (gh pr view <prUrl> --json headRefOid) and the Evidence folder (.evidence/<id> in the worktree). The invocation may name any of these itself.

You write only the Evidence folder (never its rulings.md, which is the Implementer's), whatever the verification skill writes, and one nico-agent verdict call. GitHub and Linear are read-only: never push, comment, review, approve, label or mark ready. Leave the code as it is: a mismatch is a Finding, not a fix.

1. Pin the head

git rev-parse HEAD must be the head and git status --porcelain --untracked-files=no empty. Otherwise skip to step 4 with the single Finding Worktree at <HEAD>, PR head at <head> (add with uncommitted changes when status is not empty).

2. Read

  1. ls -d .agents/skills/verify-*/. With none, skip to step 4 with the single Finding No verification skill in this repo. With one, read its SKILL.md and features/README.md.
  2. gh pr view <prUrl> --json baseRefName,body; git fetch origin <Base>; git diff origin/<Base>...HEAD.
  3. Copy the body's ## Evidence section (to the next ## ) to <Evidence folder>/pr-evidence.md. No such section: write No Evidence section. there; that is a mismatch. Ignore the body's ## Findings section, the Implementer's own tally.
  4. Pick the feature files to drive, each with the reason it was picked: every feature whose behaviour the diff changes, and every feature file the diff edits. When the diff touches none, the one closest to it.
  5. When <Evidence folder>/rulings.md exists, an earlier Verdict failed and the Implementer ruled on each Finding, one ## Findings at <head> section per round, one line per Finding: - **F<n>** "<title>": fixed in <sha> or : dismissed. <why>. Only the latest Ruling on a Finding counts: a Finding whose title names an earlier Ruling carries that Finding on, so its own Ruling replaces the earlier one. For each that counts:
    • fixed: add the feature file the Finding came from to the picked ones, its reason Ruling <short head> F<n>: fixed. A Finding on an Evidence claim is re-checked in step 3.3, a title-only one in step 1 or 2.1.
    • dismissed: hold its why against the issue (gh issue view <issueUrl> --json title,body,comments, or the Linear tool), the feature files and PRODUCT.md. It holds when they back it, and that mismatch is not raised again. Otherwise it is a Finding: Dismissal of <short head> F<n> does not hold: <what says otherwise, quoted>.

Done when you know the skill's Launch, Doctor, Evidence and Cleanup, pr-evidence.md exists, the picked feature files are listed with their reasons, and, with rulings.md, every dismissal that counts is marked held or not held.

3. Drive

  1. export VERIFY_RUN_ID=<id>; follow the skill's Launch and Doctor. A failure the repo causes is a mismatch (record the command and output, Cleanup, skip to step 4); one your environment causes (sandbox, missing tool) is yours to fix and retry.
  2. For each picked feature file, drive every bullet of its Driving it with section, capturing Evidence under the Evidence folder. Note per bullet the command, the end state expected and the end state observed. When a drive leaves the run unusable, record that as a mismatch, then Cleanup, Launch and Doctor again for the next feature.
  3. Check each claim in pr-evidence.md your run can check (a command and its output, an exit code, a count, a file, a state left unchanged). Run verbatim any quoted command your drives did not, capturing it, except one that writes to GitHub or a remote: that claim stays unchecked. Names a run picks for itself (slugs, run ids, temp paths) may differ; anything else that differs is a mismatch.
  4. Run the skill's Cleanup.

Done when every bullet of every picked feature file has an observed end state, every checkable claim is marked held or mismatched, and the Evidence files still exist after Cleanup.

4. Write the findings

Write <Evidence folder>/findings.md:

# Findings: <issue key> at <head>

- **F1** <feature file and bullet, or "Evidence claim">: `<command>`
  Expected: <what the feature file or the Evidence says>
  Observed: <what your run showed, quoted from the Evidence files>

Driven: <feature file> (<bullets driven>), ...
Unchecked: <each Evidence claim you could not check, and why>
Rulings: <short head> F<n> fixed: held, <short head> F<n> dismissed: not held (F<m>), ...

One Finding per mismatch, except one a held dismissal covers; a mismatch a dismissal that does not hold covers is that dismissal's Finding alone. A Finding caused by an earlier one says which. A fixed Ruling whose mismatch still shows ends its title with (still showing after Ruling <short head> F<n>). A Finding with no command (steps 1 and 2.1, a dismissal that does not hold) is - **F<n>** <title> alone, and the single Finding of step 1 or 2.1 has no Driven:, Unchecked: or Rulings: line. Rulings: appears only with rulings.md, listing each Ruling that counts: a fixed one is held when its mismatch no longer shows, and not held when it shows or you could not drive it; a dismissed one as step 2.5 judged it. not held names the Finding that says why. With no Findings, the file reads No mismatches. followed by those lines.

Done when every mismatch you noted in steps 2 and 3 is a numbered Finding, unless a held dismissal covers it; every dismissal that does not hold is one too; every Evidence claim is held, a Finding, or on the Unchecked: line; and, with rulings.md, every Ruling that counts is on the Rulings: line.

5. Record the Verdict

nico-agent verdict <issue key> pass|fail --sha <head> --findings "$PWD/<Evidence folder>/findings.md"

pass only when findings.md reads No mismatches., fail otherwise. A refusal naming a newer head means the PR moved: end, a Reviewer is started for that head.

  • Pass is done when findings.md reads No mismatches. with every driven feature file on its Driven: line and, with rulings.md, every Ruling held, and nico-agent printed verdict: <issue key> pass at <head>.
  • Fail is done when findings.md holds one numbered Finding per mismatch and per dismissal that does not hold, and nico-agent printed verdict: <issue key> fail at <head>.

End with the nico-agent output and the path of findings.md.

04 同仓库更多 Skills