Apply visible+invisible anchor pattern + skill-pointing errors to PR-body lint surfaces (cross-surface enhancement)

FAIR-band: in-band [7/30] — substrate consistency follow-up to PR #11494 (#11491) amendment 2026-05-16T21:24Z. Cross-surface application of the operator-directed two-layer safeguard pattern.

Premise

PR #11494 ships a two-layer mechanical body-shape validator on the manage_pr_review MCP tool's body parameter:

1. Visible layer (VISIBLE_PR_REVIEW_ANCHORS): 7 evaluation-metric tags. Misses enumerated in error response; one diagnostic anchor named in human-facing message.
2. Invisible layer (INVISIBLE_PR_REVIEW_ANCHORS): 3 structural template substrings (Depth Floor, Required Actions, Strategic-Fit Decision). Misses NOT enumerated in response. Defeats Goodhart anchor-stuffing.

Both layers point the agent at .agents/skills/pr-review/SKILL.md with explicit "do NOT compose a substitute template" anti-hallucination phrase.

Empirical anchor for the invisible layer: Gemini's review 4304287893 on PR #11499 (2026-05-16T21:16Z) contained all 7 visible metric tags but missed structural anchors — the canonical Goodhart-stuffing failure mode. Pre-enhancement validator would have passed it.

The cross-surface gap

The same two-layer pattern needs to apply at adjacent enforcement surfaces:

The visible-only state on these surfaces means a Goodhart-stuffer who routes around the MCP tool via gh pr create or gh pr review CLI directly evades the strong layer and only meets weak layers.

Prescription

Part A — Apply two-layer pattern to PR #11498's lint-pr-body.mjs

Author-side anchor set from pull-request-workflow.md §9:

Visible (already in PR #11498):

• Resolves|Closes|Fixes|Related:|Refs #N
• Evidence:
• ## Deltas
• ## Test Evidence

Invisible (to be added — candidates from §9 minimum-viable structure):

• ## Post-Merge Validation
• One or two additional structural substrings TBD by V-B-A against §9

Error message redesign per #11494 pattern:

• Lead with .agents/skills/pull-request/SKILL.md as required-action
• Name template path: .agents/skills/pull-request/references/pull-request-workflow.md §9
• "do NOT compose a substitute template or hallucinate section headings" phrase
• One diagnostic anchor named, NOT the full visible-miss list
• No missing_invisible field; no invisible substring in any response surface

Part B — Inherit two-layer pattern in #11495's gh-CLI-review-bypass CI lint

When #11495 implementation lands, must inherit:

• Same VISIBLE_PR_REVIEW_ANCHORS source-of-truth as PR #11494 (shared module ai/services/github-workflow/prReviewAnchors.mjs if/when extracted, otherwise sync-by-convention with comment-pointers)
• Same INVISIBLE_PR_REVIEW_ANCHORS source-of-truth — INVISIBLE list MUST NOT be duplicated across surfaces in a way that allows discoverability via cross-reference; ideally single-source via the same shared module
• Same skill-pointing error message shape

Single-source-of-truth design

PR #11494's anchor constants live in ai/services/github-workflow/PullRequestService.mjs. For cross-surface consistency:

Option 1 (extract first): factor VISIBLE_PR_REVIEW_ANCHORS + INVISIBLE_PR_REVIEW_ANCHORS into ai/services/github-workflow/prReviewAnchors.mjs. PullRequestService.mjs imports. PR #11498's lint script + future #11495 CI lint import from the same module. Single edit-site for future template evolution.

Option 2 (sync-by-convention): each surface hardcodes its own constants with documented "MUST stay synchronized with PullRequestService.mjs INVISIBLE_PR_REVIEW_ANCHORS" comments. Acceptable for v1 if Option 1 introduces import-path complexity for CI scripts running in GH Actions runners.

Recommended: Option 1. The CI lint scripts can import from the services module just fine in Node 24 runners.

Part A author-side anchors (different from reviewer-side; pull-request-workflow.md §9 not pr-review template) remain independent. Only the reviewer-side surfaces (#11494, #11495) share.

Acceptance criteria

• AC1: PR #11498's lint-pr-body.mjs extended with invisible-anchor layer + skill-pointing error message; preserves backward-compatible behavior when both layers pass
• AC2: PR #11498 / future PR for this ticket: spec coverage asserting:
  • Visible-only-stuffed body REJECTED with skill-pointing error
  • Response body MUST NOT contain invisible-anchor substrings
  • No missing_invisible field
• AC3: #11495 CI lint inherits the same pattern when implemented (recorded in #11495 prescription)
• AC4: If Option 1 chosen, ai/services/github-workflow/prReviewAnchors.mjs shared module created; PR #11494's PullRequestService.mjs refactored to import; AC tests in PR #11494 spec continue to pass
• AC5: ADR or substrate-doc cross-reference: the visible+invisible pattern + skill-pointing error pattern documented somewhere agents can discover when authoring future enforcement layers

Avoided traps

• Duplicating invisible-anchor list across files: rejected per single-source-of-truth design. Discoverability risk multiplies with each duplicate site.
• Documenting the invisible-anchor list publicly (README, ADR enumeration): rejected. Pattern description in ADR is OK; specific anchor enumeration is NOT.
• Coupling to specific cycle template (cycle-1 vs cycle-followup): rejected. The 3 invisible substrings in PR #11494 already span both. Any new invisible anchor for #11498's author-side surface must similarly be cycle-agnostic if applicable.
• Eager extraction to shared module before PR #11494 merges: scope-restraint per feedback_substrate_scope_restraint. Land PR #11494 baseline first; extract via this follow-up ticket.

Authority anchors

• Parent PR: PR #11494 (#11491) — canonical primary surface; this ticket is the cross-surface follow-up
• Operator directive 2026-05-16T21:15Z: "the mcp tool enhancements: if they check for specific tags, the tool errors must point to using the mandatory skills. otherwise we risk that models hallucinate a new template containing the tags. and tools should check for more tags that are not mentioned => invisible. safeguard."
• Empirical anchors: Gemini's #11499 reviews 4304287893 (Goodhart-stuffed) vs 4304295863 (corrected) — three-minute self-correction loop confirms the failure mode is real and recoverable
• Cross-surface tickets: PR #11498 (Gemini, CI PR-body lint; pending operator review re. enhancement), #11495 (gh-CLI-review-bypass split; not yet implemented)

Related

• Parent: PR #11494 (#11491) — ships canonical two-layer + skill-pointing pattern at MCP tool boundary
• Cross-surface PR1: PR #11498 — Shape B CI lint, needs two-layer extension
• Cross-surface ticket: #11495 — gh-CLI-review-bypass; inherit pattern when implemented
• Conceptual ancestor: ~50 closed meta-tickets across Helpful-Assistant counter-substrate / Map vs World Atlas / skill adherence — this ticket extends the categorical tool-boundary layer across all enforcement surfaces