Skip to content

The docs-accuracy --all backstop does not exist and the named fallback is measured not to work — #9192 option A needs re-ruling on a real second leg #9435

Description

@os-steve

Filed by the domain:devx PM seat (session session_01XqDQYVU5smx29ts9pAErja) as the decision half of #9231, which is closing as answered by PR #9428. ⛔ Not claimed, not dispatched — this needs a maintainer ruling before any code moves.

The situation, measured

The #9192 option-A ruling — keep the docs-drift scope precise rather than falling back to packageMentionDocs — rested on two legs:

  1. omissions are legible (fix(docs-audit): derive the docs-drift advisory from what pages document, not package edges #9229 publishes a "What this run could not see" section) — ✅ this leg is real;
  2. a periodic --all full audit is the coverage backstop — ⛔ this leg was never real.

#9231 checked leg 2. It does not exist: no workflow (31 registered, the 10 scheduled ones all unrelated), no Routine, the schedule skill named as its creation mechanism is not in .claude/skills/, and the repo's only three docs-accuracy audit PRs (#3243, #4219, #4312) were all hand-initiated with no cadence, most recent 2026-07-31.

⚠️ Scope note on the Routine half: list_triggers is account-scoped. The agent-seat account has no such Routine. If one exists under the maintainer's own account it would not be visible from here — that is the single residual unknown, and it is a one-sentence answer.

Two findings that change the shape of the decision

⭐ (a) The backstop as designed would not have been one. README.md part 4 described the routine as computing "the change-scoped doc list since the last audit". A change-scoped list is derived by the same anchor heuristic whose misses the backstop exists to catch — so it re-inherits the exact blind spot it was supposed to cover. This is not "a thing that wasn't running"; it is "a thing that would not have worked if it had been". Only a genuine --all pass is a backstop.

⭐ (b) The cheap fallback named in #9231 is measured not to work. Across the 8 most recent packages/**-touching commits on main, packageMentionDocs is wider but NOT a superset of the precise set: in every one of the 8, between 3 and 7 pages present in the precise docs set are absent from it. It trades one incomplete set for another and drops pages the current scope gets right — while re-introducing the wrong-in-both-directions predicate #9192 deliberately removed.

The decision

A. Record only — keep the precise scope, accept there is no backstop today.
Cost: zero. This is the state PR #9428 lands: the README now says ⛔ NOT RUNNING with the four checks that establish it, so the absence is legible instead of claimed-away. Residual risk: a docs regression in a page no change anchors to sits unseen indefinitely.

B. Stand up a real periodic --all audit. The only thing that is actually a backstop.
Cost is why this is yours and not mine: measured at ~2 agents per doc — ~2.8M output tokens / ~160 agents at 128 docs, and the corpus is 178 today. Cadence is the real knob: monthly, against a repo merging ~18 PRs/day, means a regression sits for weeks.

C. Apply the packageMentionDocs fallback. ⛔ Measured not to work — see (b). Listed only so it is visibly rejected on the merits rather than quietly dropped.

PM recommendation: A now, B as a funding decision, explicitly not C

  • Real need — the per-PR precise scope plus fix(docs-audit): derive the docs-drift advisory from what pages document, not package edges #9229's "could not see" section covers the common path. The pull for a periodic full audit is real but unquantified: nobody has yet measured a drift that only --all would have caught. B should be bought deliberately, not assumed. C serves no need at all.
  • Long-term soundness — B is the only contract-honest backstop. The highest-value part is already landed: a README claiming a backstop nobody ran is the declared-not-enforced shape this repo keeps paying for, and it is now corrected.
  • Error-resistance — the failure mode here was a PM (me) consuming a stated property as authoritative. The os-dev on [finding] The docs-accuracy --all full audit is named as the coverage backstop for the precise docs-drift list — verify it is actually scheduled and running #9231 reproduced it mid-task by copying a workflow list out of a triage comment, then fixed it structurally by putting the derivation command in the README instead of a copied list. ⇒ the durable fix is derivability, not a wider net. A backstop that silently under-covers is worse than a documented absence — the absence is legible; the under-coverage is not. That is the argument against both C and a half-hearted B.

⚠️ If B is chosen, note (a): it must be a true --all pass. A change-scoped "periodic" run would reproduce the original defect with a schedule attached.

Refs: #9231 (the verification, answered) · PR #9428 (the record) · #9192 / PR #9229 (the ruling this re-opens) · #3243 · #4219 · #4312

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions