LearnNewsExamplesServices
Frontmatter
id17207
titleThe pr-review payload argues for rules the managed path already refuses
stateClosed
labels
documentationairefactoring
assigneesneo-opus-grace
createdAtAug 15, 2026, 9:25 PM
updatedAtAug 15, 2026, 10:24 PM
githubUrlhttps://github.com/neomjs/neo/issues/17207
authorneo-opus-grace
commentsCount0
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtAug 15, 2026, 10:24 PM

The pr-review payload argues for rules the managed path already refuses

Closed Backlog/active-chunk-16 documentationairefactoring
neo-opus-grace
neo-opus-grace commented on Aug 15, 2026, 9:25 PM

Refs #17141

The byte half of #17141's AC-7, which #17178 claimed and did not meet. Filed as its own leaf because the parent still carries five open ACs and the correction is finished, verified work that should not wait behind them.

Problem

AC-7 asked that "total .agents/skills/pr-review/** bytes and procedural steps net-decrease." Steps did: an ordinary Round 2 fell from 14 sections to 3. Bytes did not — the tree went 101,844 → 104,479 (+2,635), discharged with the [skill-growth-justified] marker.

That marker exists for growth you cannot avoid. This was growth nobody tried to avoid, and the PR body compounded it by reporting +1,654 — a figure taken before four repair rounds and never re-measured before the final push. Corrected in place on PR #17179 after @tobiu asked the obvious question: "wasn't the goal that skills get SHORTER?"

The Architectural Reality

The removable mass is prose arguing for rules the code already enforces. One principle covers all of it:

Where a guard refuses something at the moment it matters, a template arguing for the rule pays bytes on every load to say what the code already says.

The clearest instance: a paragraph in the Round-2 asset explaining why a STILL_OPEN round must be COMMENT. The managed path refuses an APPROVED round carrying a STILL_OPEN and refuses a REQUEST_CHANGES that dispositions everything — and the verdict line already reads "COMMENT if any item is STILL_OPEN". The paragraph explained a rule the reviewer cannot break.

Three smaller instances of the same shape: the review-id note at two sentences where a clause does; the measurement-methodology addition at three sentences where one routing fact does; and a circuit-breaker restatement of the per-family budget that repeats the service's contract in a payload instead of pointing at where it fails closed.

What is NOT reducible, and why the number stops where it does

After the cuts the tree lands at 103,422 — still +1,578 against the pre-#17179 baseline. That remainder is the format: a Round-2 disposition asset cannot be free.

Reaching a genuine net decrease from here means deleting sections from the follow-up template that governs terminal Drop+Supersede and repair-minted re-entry. Removing contract surface to hit a byte target is the metric moving while the goal stands still — the trade this substrate's own guards exist to refuse. So AC-7's byte clause stays unmet and says so, rather than being met by a number bought with contract.

Out of Scope

  • The follow-up template's sections. Terminal-verdict contract surface is not byte-budget material.
  • Round-1 review depth. The expensive moment is correct where it is.
  • The [skill-growth-justified] marker mechanism itself.

Acceptance Criteria

  • Prose that argues for a rule the managed path mechanically refuses is removed from the pr-review payloads; the rule's statement stays, its advocacy goes.
  • The bytes an ordinary Round 2 loads are recorded before and after, since that is the surface a reviewer pays on every review — distinct from tree size, which is paid once on disk.
  • lint-skill-manifest passes without a growth-justified marker for this change.
  • The remaining delta against the pre-#17179 baseline is stated as a number with its reason, so the unmet clause is legible rather than quietly dropped.
  • No template section is deleted; the validate_pr_review_body tiers behave identically, proven by the existing corpus staying green.

Evidence class

Measured on the merged tree: 101,844 → 104,479 before this work, 103,422 after, per-file deltas from git cat-file -s across the merge commit. Ordinary-Round-2 loaded bytes 6,121 → 1,787. 676 specs green, lint-skill-manifest OK. Work is already committed on agent/17141-payload-reduction (fa7d1bec13).

Live latest-open sweep at 2026-08-15T19:25:08Z; nothing equivalent open. A2A sweep: no competing claim.

Decision Record impact

none.

🖖 Authored by Grace (Claude Opus 5, Claude Code). Session b17338dd-b474-494f-b08c-683044de2ddb. Self-reported: I claimed this AC with an exception rather than meeting it, and with a stale figure.

tobiu referenced in commit f07d2ad - "refactor(ai): the enforcement is the documentation — 1,057 B of prose the guards already say (#17207) (#17208) on Aug 15, 2026, 10:24 PM
tobiu closed this issue on Aug 15, 2026, 10:24 PM