Context
Closes the reviewer-side discipline gap in the 4-layer PR-base defense (Layers 1 + 4 LIVE; Layer 2 still open):
| Layer |
Mechanism |
State |
| 1 |
§0 invariant elevation (always-loaded) |
LIVE via #11337 / PR #11339 |
| 2 |
Workflow discipline strengthening |
OPEN for pickup |
| 3 |
pr-review Cycle-0 baseRef gate (reviewer-side mechanical) |
THIS ticket |
| 4 |
Mechanical CI/MCP guard (machine-enforceable backstop) |
LIVE via #11336 / PR #11340 |
Empirical anchor: PR #11335 base=main blowup (2026-05-13T21:38Z; operator-caught). I was the primary reviewer and missed the baseRefName verification at Cycle 1 review-start, rationalizing the 990K-line diff as "stale-branch" while the actual root cause was the catastrophic base-branch misconfiguration. Saved as feedback_pr_review_base_branch_verification memory (private). The 4-layer defense converged on the architectural framing, but the reviewer-side Cycle-0 mechanical gate was deferred for separate ticket — this is that ticket.
The Problem
Current pr-review-guide.md §9.0 Premise Pre-Flight + §5.x audits structure REQUIRES reviewer to dive into substantive scope before mechanical pre-flight. When PR diff is much larger than the stated scope (e.g., +990K lines for a feature PR claiming ~106 lines), the reviewer's natural impulse is to investigate "why so big?" via substantive diff inspection FIRST. This is precisely when wrong-base misconfiguration gets rationalized as something else (stale-branch, force-push noise, etc.).
The Cycle-0 mechanical gate fires BEFORE §9.0 Premise Pre-Flight: the reviewer MUST run the 3-command base/merge-base/divergence check IF the PR scope claim doesn't match observed size. ONLY AFTER base + merge-base + divergence verified, dive into substantive code review.
The Fix
Substrate-Additive via Recursive Application pattern (per #11324 / Discussion #11314 sub-rule sibling discipline):
New sub-rule sibling file: .agents/skills/pr-review/audits/base-branch-verification.md (~40-60 lines)
- The 3-command check (
gh pr view --json baseRefName,headRefName, git merge-base <head> origin/dev, git rev-list --count origin/main..origin/dev)
- Trigger conditions (diff size >> expected scope; PR title says "small change" but diff is huge)
- Drop+Supersede framing when base=main on feature PR
- Empirical anchor: PR #11335 stale-branch rationalization
1-line trigger pointer in pr-review-guide.md (~100 char addition):
<!-- trigger: PR diff size >> stated scope OR baseRefName uncertain → read ./audits/base-branch-verification.md -->
Insertion point: BEFORE §9.0 Premise Pre-Flight (Cycle-0 gate fires first).
pr-review-template.md checkbox (+1 line):
- Add to existing Cycle-1 sections:
- [ ] §5.x Base-Branch Verification: baseRefName + merge-base verified before substantive review
Substrate Budget Impact
| Surface |
Delta |
Conditionally-loaded? |
.agents/skills/pr-review/audits/base-branch-verification.md |
+~40-60 lines (~2-3 KB) |
Yes — only loads when trigger fires |
.agents/skills/pr-review/references/pr-review-guide.md |
+1 line (~100 bytes) |
Yes — already conditionally-loaded |
.agents/skills/pr-review/assets/pr-review-template.md |
+1 line (~80 bytes) |
Yes — already conditionally-loaded |
No always-loaded substrate touched. No AGENTS.md / AGENTS_STARTUP.md / AGENTS_ATLAS.md mutation. Per friction → gold asymmetric ratchet discipline (feedback_friction_to_gold_asymmetric_ratchet): substrate-addition justified by net preventive value — would have caught PR #11335 base=main in <1 min at Cycle-0 vs the ~20-min substantive-review-then-correction cycle.
Acceptance Criteria
Out of Scope
- Layer 2 (workflow discipline strengthening) — separate ticket
- 5th layer / "ticket already has APPROVED PR" duplicate-resolution detector — separate ticket
- Modifying §0 invariants — not in scope; Layer 1 (#11337/PR #11339) already covers always-loaded prose
- Mechanical CI gate — covered by Layer 4 (#11336/PR #11340)
Avoided Traps
- Bloating pr-review-guide.md monolith — at 57K already; per Recursive Application pattern, new gate goes in sub-rule sibling file + trigger pointer
- Always-loaded substrate addition — Layer 3 is reviewer-side discipline; lives in skill payload, not AGENTS.md
- Cycle-1 substantive-review-before-mechanical-gate order — current §9.0 fires first; new gate fires BEFORE §9.0 as Cycle-0 mechanical pre-flight
- Rule-card demotion — not applicable here; new substrate is multi-step discipline content, not single-rule prose
Related
- Discussion #11341 (graduated) — 4-layer defense framing
- #11337 / PR #11339 — Layer 1 (LIVE)
- #11336 / PR #11340 — Layer 4 (LIVE)
- PR #11335 — empirical anchor: my Cycle 1 miss + Drop+Supersede + collision-trigger
- #11324 / PR #11343 — Recursive Application sub-rule sibling pattern precedent + INV1 cascade demotion sister-pilot
feedback_pr_review_base_branch_verification — private feedback memory anchoring disciplinefeedback_friction_to_gold_asymmetric_ratchet — substrate-budget justification frame
Origin Session ID: 46c1f1f9-c521-4b71-85fa-f7cc5c7531ee
🤖 Authored by @neo-opus-4-7 — Layer 3 closure of the 4-layer PR-base defense started by Discussion #11341 graduation arc. Continuing on-lane per operator direction 2026-05-14T07:06Z.
Context
Closes the reviewer-side discipline gap in the 4-layer PR-base defense (Layers 1 + 4 LIVE; Layer 2 still open):
Empirical anchor: PR #11335 base=main blowup (2026-05-13T21:38Z; operator-caught). I was the primary reviewer and missed the
baseRefNameverification at Cycle 1 review-start, rationalizing the 990K-line diff as "stale-branch" while the actual root cause was the catastrophic base-branch misconfiguration. Saved asfeedback_pr_review_base_branch_verificationmemory (private). The 4-layer defense converged on the architectural framing, but the reviewer-side Cycle-0 mechanical gate was deferred for separate ticket — this is that ticket.The Problem
Current
pr-review-guide.md§9.0 Premise Pre-Flight + §5.x audits structure REQUIRES reviewer to dive into substantive scope before mechanical pre-flight. When PR diff is much larger than the stated scope (e.g., +990K lines for a feature PR claiming ~106 lines), the reviewer's natural impulse is to investigate "why so big?" via substantive diff inspection FIRST. This is precisely when wrong-base misconfiguration gets rationalized as something else (stale-branch, force-push noise, etc.).The Cycle-0 mechanical gate fires BEFORE §9.0 Premise Pre-Flight: the reviewer MUST run the 3-command base/merge-base/divergence check IF the PR scope claim doesn't match observed size. ONLY AFTER base + merge-base + divergence verified, dive into substantive code review.
The Fix
Substrate-Additive via Recursive Application pattern (per #11324 / Discussion #11314 sub-rule sibling discipline):
New sub-rule sibling file:
.agents/skills/pr-review/audits/base-branch-verification.md(~40-60 lines)gh pr view --json baseRefName,headRefName,git merge-base <head> origin/dev,git rev-list --count origin/main..origin/dev)1-line trigger pointer in
pr-review-guide.md(~100 char addition):Insertion point: BEFORE §9.0 Premise Pre-Flight (Cycle-0 gate fires first).
pr-review-template.mdcheckbox (+1 line):- [ ] §5.x Base-Branch Verification: baseRefName + merge-base verified before substantive reviewSubstrate Budget Impact
.agents/skills/pr-review/audits/base-branch-verification.md.agents/skills/pr-review/references/pr-review-guide.md.agents/skills/pr-review/assets/pr-review-template.mdNo always-loaded substrate touched. No AGENTS.md / AGENTS_STARTUP.md / AGENTS_ATLAS.md mutation. Per friction → gold asymmetric ratchet discipline (
feedback_friction_to_gold_asymmetric_ratchet): substrate-addition justified by net preventive value — would have caught PR #11335 base=main in <1 min at Cycle-0 vs the ~20-min substantive-review-then-correction cycle.Acceptance Criteria
.agents/skills/pr-review/audits/base-branch-verification.mdcreated with 3-command check + trigger conditions + Drop+Supersede framing + PR #11335 empirical anchorpr-review-guide.mdaugmented with 1-line trigger pointer BEFORE §9.0 Premise Pre-Flightpr-review-template.mdaugmented with 1-line Cycle-1 checkbox for base-branch verification.agents/skills/skills.manifest.jsonif perFilePayloadBudget tracking needs adjustment)feedback_pr_review_base_branch_verificationrationale + PR #11335 empirical anchor + measurement contract (catch the next wrong-base PR in <1 min Cycle-0)git diff --checkpassesnode ai/scripts/lint-skill-manifest.mjs --base origin/dev)Out of Scope
Avoided Traps
Related
feedback_pr_review_base_branch_verification— private feedback memory anchoring disciplinefeedback_friction_to_gold_asymmetric_ratchet— substrate-budget justification frameOrigin Session ID: 46c1f1f9-c521-4b71-85fa-f7cc5c7531ee
🤖 Authored by @neo-opus-4-7 — Layer 3 closure of the 4-layer PR-base defense started by Discussion #11341 graduation arc. Continuing on-lane per operator direction 2026-05-14T07:06Z.