Repository navigation
skills(checklist-test): one strict qa-run title shape, and a parser that holds it - #10941
Merged
Merged
Conversation
…hat holds it Maintainer ruling 2026-08-20 (option 2 on the drift card): tighten the run-record title convention to exactly ONE shape, new records only. QA run · <selector> (<judged>/<total>) · <sha8> · <YYYY-MM-DD> · <counts> SKILL.md now documents that as the contract rather than a shape no record wrote. `(<judged>/<total>)` is mandatory and bare-digits-only, which retires all five phrasings the corpus grew -- `(FULL area)`, `(77 items)`, `(5 of 10 items)`, `(6/14 items consulted)`, and omitting the parenthetical. `(FULL area)` is why the ruling exists: it declares no total, and a record wearing it judged 13 of its area's 33 items while reading as full coverage. NOT-RUN keeps one spelling -- a `<n> NOT-RUN` counts segment; the trailing `(11 not-run)` parenthetical is retired. qa-rollup.mjs is strict, not tolerant: five fields, mandatory bare judged/total with judged <= total, exactly 8 lowercase hex, a real calendar date, and counts as ` / `-separated `<n> <VERDICT>` segments, uppercase, in report order, each bucket at most once. The multi-phrasing branches are gone, and so are judgedCell's three inferred denominators -- they existed only because the retired phrasings declared no total, and inferring a denominator the record never claimed is the confident nonsense the contract removes. An omitted bucket still means "not declared", never zero. Existing records are not migrated, so the render splits the unparsed set by `CANONICAL_FROM` using the API's created_at -- never anything read out of a title. Nothing is extracted from a legacy title and no legacy shape is recognised; only the explanation differs. Measured over the live corpus (34 records, 2026-08-21): 3 parse, 31 predate the contract, 0 deviate. Without the split the enforcement surface would be 34 rows deep and the first genuine deviation invisible inside it. --self-test carries both halves: 5 canonical shapes that must parse and 16 retired ones that must be rejected, each asserting the reason names the field that failed. The rejections are the half that holds the ruling -- a parser re-widened to accept one of them passes every happy-path assertion. Fixes #9617 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ApyDuQY2fkunMCqXiqvBhR
…sentials The ratchet ceiling for this file is 232 with zero headroom, so the first draft's 21 added lines were 21 over. Compressed to the binding elements only: the shape, the five field rules, the retired phrasings as a bare list, and the enforcement surface. Provenance is one line and the (FULL area) story is one clause, per the ratchet's own remedy text. Still 241 — 9 over. The remaining 9 cannot be reclaimed without either editing the ceiling ledger (scripts/pm/check-skill-line-ratchet.mjs, outside this card's declared file surface) or deleting unrelated content from the file. Reported for a maintainer decision rather than resolved by either. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ApyDuQY2fkunMCqXiqvBhR
…n SKILL.md PM ruling on the ratchet question (option C): .claude/skills/checklist-test/SKILL.md was at its line ceiling of 232 with zero headroom, so documenting the contract in it overflowed by 9 lines. Rather than raise the ceiling (needs a maintainer ruling the 2026-08-20 title ruling is not) or delete maintainer-authored content, the field-by-field grammar moves to the operational doc this skill already cites throughout. docs/qa/platform-checklist/RUNNER.md gains 'The record title — one canonical shape' under its existing '## Run records' section: the shape, a field-by-field table, the declared-vs-zero rule, the retired phrasings with why (FULL area) died, and the new-records-only note. SKILL.md keeps the binding rule — one shape, strict parsing, judged/total mandatory — plus a pointer, and is back to exactly 232. This is the ratchet's own remedy: principles in SKILL.md, on-demand detail elsewhere. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ApyDuQY2fkunMCqXiqvBhR
Contributor
⛔ merge queue 构建失败 — 先分诊，再决定要不要重排队列构建 32543326482 红了。队列跑的是全量套件（PR 侧 CI 只跑 affected 子集）， 失败的 job（日志抽取，best effort）：
跨 PR 相同签名（24h，按失败测试文件聚合）：
历史信号：
分诊清单：
Generated by Claude Code · merge-queue-triage workflow (#4859) |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #9617
Implements the maintainer ruling recorded on the card 2026-08-20 — option 2, tighten the
qa-runrecord title to ONE strict shape, new records only. Provenance quoted verbatim anduntranslated: 「其他接受你的建议。」
Files
.claude/skills/checklist-test/SKILL.md— the binding rule and a pointer (governed surface)docs/qa/platform-checklist/RUNNER.md— the field-by-field grammarscripts/qa/qa-rollup.mjs— the strict parser and its embedded--self-testThe contract
Exactly five middle-dot fields.
(JUDGED/TOTAL)is mandatory and bare-digits-only, withjudged not exceeding total. The sha is exactly 8 lowercase hex. The date is a real calendar
date. COUNTS is
/-separatedN VERDICTsegments — uppercase, in the report orderPASS / PARTIAL / FAIL / BLOCKED / NOT-RUN, each bucket at most once. An omitted bucket still
means "not declared", never zero, so an author who means zero writes
0 FAIL.Retired, all of them:
(FULL area),(N items),(N of M items),(N/M items consulted),omitting the parenthetical entirely, and the trailing
(11 not-run)spelling of NOT-RUN.(FULL area)is why the ruling exists — it declares no total, and a record wearing it judged13 of its area's 33 items while reading as complete coverage.
Where the contract is written
The grammar lives in
docs/qa/platform-checklist/RUNNER.md, as a new"The record title — one canonical shape" subsection under its existing "## Run records"
section: the shape, a field-by-field table, the declared-vs-zero rule, the retired phrasings
with the reason each died, and the new-records-only note.
SKILL.mdkeeps the binding rule — one shape, strict parsing, judged/total mandatory — plus apointer. That split is the skill line ratchet's own stated remedy ("principles in SKILL.md,
on-demand detail elsewhere"): the file was at its ceiling of 232 with zero headroom, so
documenting the contract inside it overflowed by 9 lines. It is back to exactly 232, with no
ceiling exception and nothing maintainer-authored deleted.
The parser is strict, not tolerant
scripts/qa/qa-rollup.mjs(landed by PR #9616) matches that shape and nothing else. Themulti-phrasing branches are gone, and so are
judgedCell's three inferred denominators — theyexisted only because the retired phrasings declared no total, and inferring a denominator the
record never claimed is the confident nonsense the contract removes.
Every rejection names the field that failed, because that reason is what the "Not parsed"
section prints and that section is the enforcement surface.
Existing records are not migrated, and the enforcement surface stays legible
The ruling is new-records-only. Measured against the live corpus on 2026-08-21 (34 records,
open and closed) the strict parser gives 3 parsed / 31 predating the contract / 0 deviating.
An undifferentiated "Not parsed" list would therefore be 31 rows deep on day one, and the first
genuine deviation would be invisible inside it — defeating the surface the ruling relies on.
So the render partitions the unparsed set by a
CANONICAL_FROMconstant, using the GitHub API'screated_at. This is not tolerant multi-shape parsing: nothing is extracted from a legacytitle, no legacy shape is recognised, and every record is equally unparsed. Only the explanation
differs — records predating the contract are labelled expected and not defects; records written
after it are the signal. Every record is still printed, never dropped.
Verification
node scripts/qa/qa-rollup.mjs --self-test— green, 78 assertions over 5 canonical shapes and16 retired ones. Both halves are load-bearing: a parser re-widened to accept a retired phrasing
passes every happy-path assertion, so the rejections are the half that actually holds the ruling.
[^)]*) to re-accept(6/14 items consulted)turns the self-test red with exactly 1 failure, naming that fixtureas having parsed. Mutation confirmed on disk before the run (injected form present 1, strict form
present 0, plus the diff of the line); the restore leg ran under an
EXIT INT TERMtrap and wasconfirmed absent afterwards, tree clean, self-test green again.
node scripts/pm/dispatch-gates.mjs(no path args) on a clean treeat
e249bff62— it read the 3-path change set itself. 10 named, all 10 green, pluscheck-platform-checklistrun as an adjacent extra, also green.check:pm-skill-ratchetnowreports: "✓ check-skill-line-ratchet: .claude/skills/checklist-test/SKILL.md is 232 lines
(ceiling 232; headroom 0)."
Governed surface
.claude/skills/**— draft-only, human merge (Prime Directive #14). Not queued, not armed, neverflipped ready.
RUNNER.mdis not a governed face; that edit is confined to the grammar section.No changeset: this PR publishes nothing, matching 12 of 12 recent
.claude/skills/**merges, so itcarries
skip-changeset.Generated by Claude Code