Context
Friction→gold from the #13124 incident. PR #13124 + ticket #13122 added a lead-role-session marker that scored lead/peer "coordination compliance" — a lead-tracks-peers surveillance surface that violates the flat-peer-team core value (AGENTS.md §swarm_topology_anchor). It was APPROVED TWICE (both cross-family for the GPT-authored PR) on clean execution — no reviewer challenged its premise. Operator-directed close; #13124/#13122 closed, #11195 AC6 (the underlying compliance-audit requirement) struck.
The recurring failure class this targets: swarm-approves-execution-not-premise — reviews validate tests/checklist/anchors but skip "does this premise cohere with who we are?" The operator rejects a large fraction of cross-family-APPROVED PRs for exactly this.
The Problem
The substrate to catch this already exists but is un-forced:
pr-review-guide.md §0 (patch-blind premise snapshot — "reject a toaster-when-we-need-a-car; a green checklist over a wrong premise is theater") is "optional during migration" in the review-body validator (manage_pr_review tool description; agent-pr-review-body-lint CI). A reviewer can pass without ever recording a premise verdict.AGENTS.md §identity_prompt_firewall L1 mandates "When faced with a questionable premise… or a PR review, DO NOT default to execution or agreement. Instead, HALT, challenge the premise…" — but it is prose (un-forced), and its trigger is scoped to premises handed to you / when you are the reviewer. It does not cue challenging the premise of an item you merely see (a broadcast PR/ticket) — which is #13124's exact hole: a peer saw the PR in broadcasts, wasn't the assigned reviewer, and stayed silent ("not my review").
#13124 is the empirical proof that the existing, explicit, turn-loaded firewall directive did not fire (two reviewers reviewed; the premise-challenge step was skipped). So the failure is non-application of un-forced substrate, not a missing rule.
Critical honest framing (load-bearing — do not let implementation drift from it): a value judgment — "does this premise cohere with our core values?" — is un-CI-able (unlike whitespace or a missing label). Text cues the reflex; only will applies it. No mechanism can guarantee the depth. A forcing-function can only raise the floor by forcing articulation — making a reviewer write the verdict rather than skip it. The primary lever remains the internalized "does it make sense?" habit; these changes complement it, they do not replace it.
The Architectural Reality
- §0 premise snapshot:
.agents/skills/pr-review/references/pr-review-guide.md §0 + the three fields (Inputs Read Before Patch, Expected Solution Shape, Patch Verdict). Validator: manage_pr_review + the agent-pr-review-body-lint CI — currently "optional during migration: omit all three or include all three together."
- Subtlety: §0's existing fields capture the fit-premise ("expected solution shape") — good for code. #13124's miss was value-coherence ("should this exist? does it cohere with flat-peer-team?"), which is upstream of fit. Forcing §0's fit-fields alone would NOT have caught #13124 (a reviewer would write a fit-premise and pass). The forced verdict must explicitly include the value-coherence question, not just the fit-shape.
- Firewall:
AGENTS.md §identity_prompt_firewall (resolved via .claude/CLAUDE.md → AGENTS.md) — L1 trigger list.
The Fix
- AC1 — forcing-function (the lever): make a premise-coherence verdict a REQUIRED, un-skippable field in the agent PR-review-body contract (
agent-pr-review-body-lint + manage_pr_review validation). Flip §0's premise snapshot from optional → required, and ensure the required shape includes the explicit value-coherence question: "Does this PR's PREMISE cohere with our core values (flat-peer-team / equal-peer-agency)?" A review missing it fails validation. Forces articulation (raises the floor); does not claim to force depth.
- AC2 — scope-widen: extend the
§identity_prompt_firewall "challenge the premise" trigger from {handed-to-you / reactive-fix / when-you-review} to also include "any peer's item you merely see (a broadcast PR / ticket / way-of-working)." Patches #13124's exact hole.
Contract Ledger
| Target Surface |
Source of Authority |
Proposed Behavior |
Fallback |
Consumer |
agent-pr-review-body-lint + manage_pr_review body validation |
this ticket + pr-review-guide §0 |
require a premise-coherence verdict (incl. the value-coherence question), flipping §0 from optional→required |
review body fails validation if absent |
every agent PR reviewer |
AGENTS.md §identity_prompt_firewall L1 trigger |
this ticket + §swarm_topology_anchor |
trigger also fires for items merely seen, not just handed-to-you/reviewing |
n/a (prose directive) |
every agent, every turn |
Decision Record impact
none / aligned-with — serves the existing §swarm_topology_anchor flat-peer-team core value + pr-review-guide §0; does not amend or challenge an ADR (the firewall is a core directive, not an ADR).
Substrate Enforcement Sufficiency Audit (§9.2)
Existing enforcement = §0 (optional / un-forced) + the firewall ("challenge the premise" — prose / un-forced / scoped to handed-to-you). Insufficient, proven empirically by #13124: the explicit firewall directive existed and did not fire, and §0 was skippable. This ticket enforces the existing §0 (optional→required) and extends the existing firewall trigger (scope) — it adds no new rule, so it is not negative-ROI substrate bloat; it makes a sufficient-but-un-applied rule un-skippable for the form, and widens an existing trigger's scope.
Acceptance Criteria
Out of Scope
- A CI check that mechanically verifies coherence depth — explicitly NOT attempted; coherence is a value judgment, un-CI-able. We force ARTICULATION only.
- Auto-detecting flat-peer-team violations (judgment, not lint).
Avoided Traps
- A parallel "§9.0 premise gate" — rejected (operator + author, this session): §0 already mandates premise-challenge; the gap is enforcement + scope, not a missing gate. Adding §9.0 = accretion over a sufficient rule (the §13 / substrate-sufficiency anti-pattern).
- Expecting text to fix the will-gap — rejected: text cues, will applies; #13124 proves the existing forceful firewall didn't fire. The honest weight is on the habit; these are floor-raising complements, framed as such.
Related
#13124, #13122 (closed — the surveillance incident), #11195 (AC6 struck), #13136 (Decision-Record-at-graduation — adjacent premise-substrate). pr-review-guide §0, AGENTS.md §identity_prompt_firewall + §swarm_topology_anchor.
Release classification: boardless — post-v13 process/substrate hardening; not release-blocking.
Origin Session ID: e7f14a36-5096-4570-932b-82f860d7a537
Retrieval Hint: "premise-coherence verdict required PR review forcing-function firewall challenge-on-seeing #13124 flat-peer-team swarm-approves-execution-not-premise"
Authored by Claude Opus 4.8 (Claude Code, @neo-opus-grace / Grace).
Context
Friction→gold from the #13124 incident. PR #13124 + ticket #13122 added a
lead-role-sessionmarker that scored lead/peer "coordination compliance" — a lead-tracks-peers surveillance surface that violates the flat-peer-team core value (AGENTS.md §swarm_topology_anchor). It was APPROVED TWICE (both cross-family for the GPT-authored PR) on clean execution — no reviewer challenged its premise. Operator-directed close; #13124/#13122 closed, #11195 AC6 (the underlying compliance-audit requirement) struck.The recurring failure class this targets:
swarm-approves-execution-not-premise— reviews validate tests/checklist/anchors but skip "does this premise cohere with who we are?" The operator rejects a large fraction of cross-family-APPROVED PRs for exactly this.The Problem
The substrate to catch this already exists but is un-forced:
pr-review-guide.md §0(patch-blind premise snapshot — "reject a toaster-when-we-need-a-car; a green checklist over a wrong premise is theater") is "optional during migration" in the review-body validator (manage_pr_reviewtool description;agent-pr-review-body-lintCI). A reviewer can pass without ever recording a premise verdict.AGENTS.md §identity_prompt_firewallL1 mandates "When faced with a questionable premise… or a PR review, DO NOT default to execution or agreement. Instead, HALT, challenge the premise…" — but it is prose (un-forced), and its trigger is scoped to premises handed to you / when you are the reviewer. It does not cue challenging the premise of an item you merely see (a broadcast PR/ticket) — which is #13124's exact hole: a peer saw the PR in broadcasts, wasn't the assigned reviewer, and stayed silent ("not my review").#13124 is the empirical proof that the existing, explicit, turn-loaded firewall directive did not fire (two reviewers reviewed; the premise-challenge step was skipped). So the failure is non-application of un-forced substrate, not a missing rule.
Critical honest framing (load-bearing — do not let implementation drift from it): a value judgment — "does this premise cohere with our core values?" — is un-CI-able (unlike whitespace or a missing label). Text cues the reflex; only will applies it. No mechanism can guarantee the depth. A forcing-function can only raise the floor by forcing articulation — making a reviewer write the verdict rather than skip it. The primary lever remains the internalized "does it make sense?" habit; these changes complement it, they do not replace it.
The Architectural Reality
.agents/skills/pr-review/references/pr-review-guide.md §0+ the three fields (Inputs Read Before Patch,Expected Solution Shape,Patch Verdict). Validator:manage_pr_review+ theagent-pr-review-body-lintCI — currently "optional during migration: omit all three or include all three together."AGENTS.md §identity_prompt_firewall(resolved via.claude/CLAUDE.md→AGENTS.md) — L1 trigger list.The Fix
agent-pr-review-body-lint+manage_pr_reviewvalidation). Flip §0's premise snapshot from optional → required, and ensure the required shape includes the explicit value-coherence question: "Does this PR's PREMISE cohere with our core values (flat-peer-team / equal-peer-agency)?" A review missing it fails validation. Forces articulation (raises the floor); does not claim to force depth.§identity_prompt_firewall"challenge the premise" trigger from {handed-to-you / reactive-fix / when-you-review} to also include "any peer's item you merely see (a broadcast PR / ticket / way-of-working)." Patches #13124's exact hole.Contract Ledger
agent-pr-review-body-lint+manage_pr_reviewbody validationpr-review-guide §0AGENTS.md §identity_prompt_firewallL1 trigger§swarm_topology_anchorDecision Record impact
none/aligned-with— serves the existing§swarm_topology_anchorflat-peer-team core value +pr-review-guide §0; does not amend or challenge an ADR (the firewall is a core directive, not an ADR).Substrate Enforcement Sufficiency Audit (§9.2)
Existing enforcement = §0 (optional / un-forced) + the firewall ("challenge the premise" — prose / un-forced / scoped to handed-to-you). Insufficient, proven empirically by #13124: the explicit firewall directive existed and did not fire, and §0 was skippable. This ticket enforces the existing §0 (optional→required) and extends the existing firewall trigger (scope) — it adds no new rule, so it is not negative-ROI substrate bloat; it makes a sufficient-but-un-applied rule un-skippable for the form, and widens an existing trigger's scope.
Acceptance Criteria
§identity_prompt_firewalltrigger extended to items you merely see (broadcast PR / ticket / way-of-working), not just handed-to-you / formal-review.references/; the AGENTS.md change is a scoped trigger-extension, not a new block).Out of Scope
Avoided Traps
Related
#13124, #13122 (closed — the surveillance incident), #11195 (AC6 struck), #13136 (Decision-Record-at-graduation — adjacent premise-substrate).
pr-review-guide §0,AGENTS.md §identity_prompt_firewall+§swarm_topology_anchor.Release classification: boardless — post-v13 process/substrate hardening; not release-blocking.
Origin Session ID: e7f14a36-5096-4570-932b-82f860d7a537
Retrieval Hint: "premise-coherence verdict required PR review forcing-function firewall challenge-on-seeing #13124 flat-peer-team swarm-approves-execution-not-premise"
Authored by Claude Opus 4.8 (Claude Code, @neo-opus-grace / Grace).