verify-acceptance-items
Verify a GitHub pull request or GitLab merge request (MR) against the acceptance criteria (验收标准/条件, task-list checkboxes) of the issues it links to, then tick the criteria it proves once the user confirms. Use when asked whether a PR or MR is done or meets its issue, when asked to check off an issue's checkboxes, or right after opening or updating a PR or MR that closes an issue.
- License 许可
- MIT
- 所属仓库
- meymchen/skills
- 最近更新
- 2026-10-16
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- verify-acceptance-items
- description
- Verify a GitHub pull request or GitLab merge request (MR) against the acceptance criteria (验收标准/条件, task-list checkboxes) of the issues it links to, then tick the criteria it proves once the user confirms. Use when asked whether a PR or MR is done or meets its issue, when asked to check off an issue's checkboxes, or right after opening or updating a PR or MR that closes an issue.
02 资源文件 5
- 参考文档agents/openai.yaml346 B
- 资产文件assets/judge-brief.md982 B
- 参考文档README.md875 B
- 脚本scripts/verify_acceptance_items.py44.4 KB
- 脚本tests/test_verify_acceptance_items.py53.2 KB
03 SKILL.md 正文
Verify Acceptance Items
Vocabulary
- Request: a GitHub pull request or a GitLab merge request. Report it with
the platform's own term, such as PR #12 or MR !42;
linksreturns it asrequest_label. - Acceptance criterion: a task-list item in an issue body that states
something the work must satisfy. Its section heading varies:
Acceptance criteria,验收标准,验收条件,Definition of done, a bold line, a<details>block. - Item id: the
L<line>idextractgives every task-list item. Ids carry each criterion fromextractthroughbriefand the verdicts toapply. - Verdict:
satisfied,unsatisfied, orundecidable.
The script does every deterministic part: detecting the platform, resolving
links, parsing the body, rendering the judging brief, and ticking. You supply
the two judgements it cannot make: which items are acceptance criteria, and
what the user decides. Resolve the skill directory from the loaded SKILL.md
path and run
uv run --script <skill-dir>/scripts/verify_acceptance_items.py <command>.
The script detects the platform from the current branch's upstream remote, or
origin, and uses gh for GitHub and glab for GitLab. Exit code 2 with
"cannot tell whether ... is GitHub or GitLab" means the host is ambiguous: ask
the user which it is and pass --platform github or --platform gitlab on
every command. Exit code 2 naming auth login means the CLI is missing or
signed out: tell the user to run the command it names. Every command accepts
--repo (a project path such as group/sub/project, or a project URL) and
--host.
1. Resolve the issues
uv run --script <skill-dir>/scripts/verify_acceptance_items.py links [--request <number>]
Without --request it uses the current branch's request. Each candidate
carries a provenance. Use a closing_reference directly. For a
body_mention, related_issue, or branch_name candidate, name it to the user
and get agreement before using it. Mention a rejected number only if the user
expected that issue to be covered.
Process every accepted candidate, each in its own report section, passing its
repo to the later commands.
2. Extract the structure
uv run --script <skill-dir>/scripts/verify_acceptance_items.py extract --repo <project> --issue <number>
Each item comes with id, text (wrapped lines folded in), state
(unchecked, checked, or inapplicable), depth, parent, sub_issue, and
its heading_path. Keep the body_sha256 for step 6.
Report every skipped entry as detected and not handled: table for
checkboxes in a table, comment for task lists in the issue's comments.
3. Choose the acceptance criteria
Pick sections from heading_path and the wording of their items. When more than
one section plausibly holds acceptance criteria, include them all and say why
each was included; over-listing costs the user a glance, under-listing hides a
criterion for good. Leave out sub_issue: true items, which only reference
another issue, and inapplicable items, which the author marked [~]; list
the latter in the report as inapplicable. If the body has checkboxes but none
read like acceptance criteria, say so plainly.
Done when every item is chosen, or excluded with a reason.
4. Judge in one subagent
uv run --script <skill-dir>/scripts/verify_acceptance_items.py brief --repo <request-project> --request <number> --items "#42:L13,L15" "group/sub/spec#3:L4"
One --items selection per issue; the project prefix defaults to --repo. The
output is the complete prompt: the evidence commands for the platform, the
verdict definitions, the answer format, and the chosen criteria. Hand it
verbatim to a single subagent with a fresh context, and add nothing. A session
that wrote the request knows the intent behind the code, and that intent reads
as evidence the diff does not hold; one sentence of your own summary carries it
across.
One subagent covers every issue: the fresh context provides the isolation, and reading the request dominates the cost. Split the selections across subagents only when one brief is too large to hold, and say so in the report.
Report verdicts as returned. Where you disagree because of something outside
the request, carry the criterion as undecidable and let the user settle it.
Re-run brief and the subagent only to fix a faulty selection.
Without subagent support, follow the brief yourself and state that the judging was not isolated.
5. Report, then wait
- A table of every criterion with its id, verdict, and evidence.
- A numbered queue of
undecidablecriteria, so the user can answer "1 and 3 are satisfied". - Warnings for criteria already ticked that this request's evidence does not support.
Then stop. Write nothing until the user responds.
6. Tick
Tick satisfied criteria and undecidable ones the user adjudicated as
satisfied; mark the latter as user-adjudicated in the report. One call per
issue:
uv run --script <skill-dir>/scripts/verify_acceptance_items.py apply --repo <project> --issue <number> --body-sha256 <from extract> --tick L13 L15
apply re-reads the body, aborts with exit code 2 if its hash changed, refuses
any id that is not a tickable item with exit code 64, flips only [ ]
characters, and reads the body back after writing. On exit code 2, re-run from
step 2 for that issue and ask the user again if its criteria changed. Exit code
2 after a permission failure means the account cannot write to that project;
the report still stands. Exit code 1 with a read-back difference means the
platform stored something other than what was sent: show the user the
difference and do not retry.
Post no comment unless the user asks; then comment on the request, not the issue.
Guardrails
- Every write goes through
apply. A body rebuilt any other way loses images, HTML comments, and trailing whitespace the author may not be able to recover. - Ticks only move from
[ ]to[x]. A tick may rest on verification you cannot see, so an unsupported tick earns a warning and stays. unsatisfiedcriteria stay unticked.
04 同仓库更多 Skills
cleanup-merged-branch
Safely clean a merged local branch, reproducible worktree caches, and stale local branches before updating the repository default branch.
update-request-body
Update the title and body of one or more GitHub pull requests or GitLab merge requests (MRs).
commit
Create a Git commit from the current changes.
commit-push-request
Commit current changes, push the branch, and open a GitHub pull request or GitLab merge request (MR).
clean-gone
Remove local Git branches whose upstreams are gone, including associated worktrees.
testing-strategy
Design test strategies and test plans. Trigger with "how should we test", "test strategy for", "write tests for", "test plan", "what tests do we need", or when the user needs help with testing approaches, coverage, or test architecture.