Repository navigation
docs(local-review): the audit takes the CI review's pre-existing lines as findings - #115
Conversation
…s as findings The shared review now gives each pre-existing defect a reviewer noticed a `Pre-existing:` summary line, never counted (pixtuoid/.github#34), so a real bug found on a PR is no longer dropped. The whole-codebase audit is that line's consumer: each one on a PR merged since the last audit enters its refute-and-rank like any finder's. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FH67SxdB9PG9MHCkkREPqH
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
There was a problem hiding this comment.
Blocking: 0
Verdict: nothing blocks the merge. A verifier kept one non-blocking finding, posted inline at SKILL.md:56: nothing records when "the last audit" ran, so the window "since the last audit" can't be computed. A thread at SKILL.md:55 lists the readers and writers the reviewers left unread. It stays open until someone replies with what was read.
Changes the reviewers listed:
.claude/skills/local-review/SKILL.md, Whole-codebase scope: one sentence is added. EveryPre-existing:line that the CI review's summary wrote on a PR merged since the last audit now counts as an audit finding. It goes through the same skeptic, ranking and one-PR-per-fix steps as any other finder's finding. The rest of the paragraph is only re-wrapped.
Lens checks:
- Must-run: nothing applies.
lofi_benchcoversPerformance::render's path, and this docs-only diff doesn't touch it. - Security: the diff touches no
.github/file. - Claims:
- The body says the summary gives each pre-existing defect "a
Pre-existing:line, never counted". That holds against the sharedclaude-review.ymlstep 3. - The body cites pixtuoid/.github#34 as merged. That holds: it was fetched and its
merged_atis 2026-11-10T07:29:32Z. - The body says pixtuoid's
local-reviewgets the same step in its own PR. That is future work in another repo and is unverified.
- The body says the summary gives each pre-existing defect "a
Pre-existing: .claude/skills/local-review/SKILL.md:39-40 (step 6, "Disposition") tells the session to "reply with the fix or the refutation, then resolve" each thread the CI review opens. The shared review now also opens an unread-readers thread, which is not a finding and needs a reply naming what was read. Neither a fix nor a refutation fits that thread. Suggested fix: add "or, on the unread-list thread, what was read".
Pre-existing: this is outside the repo, in pixtuoid/.github's review prompt. The prompt says to read AGENTS.md and REVIEW.md "with mcp__github__get_file_contents, ref the base". This run, ref set to the base sha failed with "could not resolve ref … as a branch or a tag", and the sha parameter worked.
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
…dow to compute Nothing records when the last audit ran, so "since the last audit" could not be computed (lofi#115's review). The audit takes every merged PR's `Pre-existing:` lines instead; one fixed since fails its skeptic, which already refutes a finding the head no longer has. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FH67SxdB9PG9MHCkkREPqH
There was a problem hiding this comment.
Blocking: 0
Verdict: 71fc4d6 fixes the earlier "since the last audit" finding: the recheck confirmed the head has no window left. One non-blocking finding remains. The PR body's "What" sentence still describes the removed window (see the inline comment).
Changes the reviewers listed (scope: 71fc4d6, the commit after dc18271):
.claude/skills/local-review/SKILL.md:55-58, Whole-codebase scope. "summarized on a PR merged since the last audit is a finding too" becomes "summarized on a merged PR is a finding too, and one fixed since fails its skeptic like any other". The next sentence is only re-wrapped.
Checks with no finding:
- Must-run:
lofi_benchisn't touched. - CI-only gates: no
ci.ymljob reads.claude/skills/**. - Security: nothing under
.github/changes. - The title still holds.
Dropped by verifiers:
- Re-checking refuted lines on every audit: thread 4522247927 already raised this.
- "merged" leaving out closed PRs: this is a scope choice, not a defect.
Pre-existing: none noticed.
The readers and writers left unread are in a separate inline thread.
…r not A pre-existing defect is in main whether or not the PR that noticed it merged, so a closed-unmerged PR's `Pre-existing:` lines count too. This matches pixtuoid's local-review (pixtuoid/pixtuoid#2218). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FH67SxdB9PG9MHCkkREPqH
There was a problem hiding this comment.
Blocking: 0
Verdict: no findings in a5e7561. The verifiers kept none. The head's wording matches the PR body and pixtuoid/pixtuoid#2218. Nothing on REVIEW.md's must-run list applies: lofi_bench is for Performance::render's path, and this diff touches only a skill doc.
What the diff changes:
.claude/skills/local-review/SKILL.md:55-57: the audit now takesPre-existing:lines from every closed PR, merged or not, where before it took them only from merged PRs.- The following lines are re-wrapped. Their wording is unchanged.
Rechecks, both dropped because the head no longer has the defect:
- 4522247927, the "since the last audit" window: the head has no window.
- 4522469322, the stale PR body: the body's "What" section now matches the head.
Candidates the verifiers dropped:
- "Closed" leaves out open PRs: a design choice, since an open PR's summary changes on every push.
- The step is copied by hand in two repos: the two SKILL.md files are separate files for separate repos, and each audit reads only its own repo's PRs.
Pre-existing: none noticed.
The inline thread 4522718020 lists the readers and writers this run left unread.
|
This pull request does not currently match the merge queue conditions, so it cannot be queued from here. The box comes back if it matches again. |
Why
The shared review contract (pixtuoid/.github#34, merged) now has reviewers return each pre-existing defect they notice. The summary gives each one a
Pre-existing:line, never counted. Before that change, a real bug a reviewer found outside the PR's responsibility was dropped. The owner asked that these reach a fix. A line nobody reads would fail the org rule "Every mechanism earns its upkeep", so it needs a consumer. lofi's whole-codebase audit is that consumer.What
local-review's Whole-codebase scope: eachPre-existing:line the CI review wrote on a closed PR, merged or not, is a finding too. It then goes through the same skeptic, ranking and one-PR-per-fix as any finder's, and one fixed since fails its skeptic, so no audit-run record is needed. pixtuoid'slocal-reviewgets the same step in its own docs PR.Roadmap: Review: a review reads what a change reaches, and the pre-existing defects it notices feed the audit
🤖 Generated with Claude Code
https://claude.ai/code/session_01FH67SxdB9PG9MHCkkREPqH