Skip to content

meta-audit: fix doc-audit scope rot, close openspec/README.md gap#708

Open
stphung wants to merge 3 commits into
mainfrom
claude/meta-audit-doc-audit-v2
Open

meta-audit: fix doc-audit scope rot, close openspec/README.md gap#708
stphung wants to merge 3 commits into
mainfrom
claude/meta-audit-doc-audit-v2

Conversation

@stphung

@stphung stphung commented Jul 8, 2026

Copy link
Copy Markdown
Owner

Summary

meta-audit for doc-audit, triggered by the 5-run threshold (runs since the last eval marker: PR #666, #693, #697, #702, #706).

Phase 2 — adversarial re-verification: Sampled the 20 most recent findings across those 5 runs (of 38 total; 18 skipped — the 11 from PR #666 and 6 from PR #693, the two oldest runs, plus 1 low-value repeat finding from PR #697), each independently re-derived from source at its own commit by a fresh agent blind to the logged answer. Result: 20/20 confirmed — zero inaccuracies found. doc-audit's findings across this window are reliable.

Phase 3 — scope-staleness check, using the scope-universe comment at the top of doc-audit/SKILL.md:

  • Scope rot (fixed): Agent 1's Scope: line still named 6 docs/*.md files — system-design.md, core-systems.md, secondary-systems.md, balancing.md, challenge-minigames.md, infrastructure.md — that were deleted on 2026-07-04 (Adopt OpenSpec as source of truth; archive the old design docs #675, "Adopt OpenSpec as source of truth"). That deletion predates 4 of the 5 evaluated runs (only PR docs: fix stale constants and missing entries found by doc-audit #666's audit commit predates it). The line now reflects what's actually under docs/*.md today (decisions.md, README.md) with a pointer to why the others are gone.
  • Scope gap (fixed): openspec/README.md carries its own copy of the OpenSpec capability/requirement/scenario counts, independent of CLAUDE.md's copy of the same numbers — and PR doc-audit: fix stale OpenSpec counts, facade note, file attribution #702's run already caught it drifting on its own (197/500+ → actual 208/498), without it being named anywhere in the scope. Added it to both the scope-universe comment and Agent 1's Scope: line so coverage isn't incidental.

No confirmed-inaccuracy pattern emerged from Phase 2, so there's nothing to add to the anti-patterns table this round.

Test plan

  • Confirmed via git log/timestamps that the OpenSpec-adoption deletion (974bdbb, 2026-07-04 17:04) predates 4 of 5 evaluated runs' audit-base commits (PR693 2026-07-05, PR697 2026-07-06, PR702 2026-07-07, PR706 2026-07-08), while PR666's audit base (2026-07-04 12:43) predates it
  • Confirmed openspec/README.md's count claims are not covered by any existing Scope: glob (CLAUDE.md, docs/*.md, docs/dossiers/*.md, src/*/CLAUDE.md, README.md)
  • Doc-only change (skill instructions); no code/build impact

Generated by Claude Code

claude added 3 commits July 8, 2026 17:28
Records this run's findings (PR #706) in the meta-audit history log.
Agent 1's Scope: line still named 6 docs/*.md files (system-design,
core-systems, secondary-systems, balancing, challenge-minigames,
infrastructure) that were deleted on 2026-07-04 when OpenSpec became
the source of truth (#675) -- predating 4 of the last 5 evaluated
doc-audit runs. Updated it to reflect the current docs/*.md contents.

Also closes a scope gap: openspec/README.md carries its own copy of
the capability/requirement/scenario counts (independent of CLAUDE.md's
copy), and PR #702 already caught it drifting on its own -- formalized
it into the scope-universe and Agent 1's Scope: line so coverage isn't
incidental.

Re-verified all 20 sampled findings from the 5 runs since the last
eval marker (PR666, PR693, PR697, PR702, PR706) against source as of
each run's commit -- zero inaccuracies found across the sample.
Marks PR666/693/697/702/706 as evaluated -- 20/20 sampled findings
confirmed accurate, scope rot and a scope gap fixed.
@stphung stphung enabled auto-merge (squash) July 8, 2026 17:35
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants