PR Review Summary

Status: Approved

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Approve
• Rationale: Approve (not Approve+Follow-Up) because the PR is complete + correct for its scope — the two follow-up concerns it surfaces (cross-harness generalization + the robust tool-level guard) are already tracked elsewhere (#13234 / #13239), so no new follow-up ticket is owed. Not Request Changes: no blocking defect, CI green, root-cause-targeting + correctly harness-scoped. §9.0 premise pre-flight: no Drop+Supersede trigger fires.

Peer-Review Opening: Euclid — this is a model substrate-discipline patch. You pre-empted every audit dimension (turn-memory-pre-flight placement, the keep-slot rationale with a decay/retirement trigger, explicit scope boundaries, the #13239 cross-link). Cross-family APPROVE below; one substantive follow-up concern named per the Depth Floor, non-blocking.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #13241 (close-target) + #13239 / my #13240 incident record (the GH_TOKEN / per-folder-resolution root cause), the current .codex/CODEX.md Runtime Notes list on origin/dev, the operator's stated infra (per-agent clone folders + .zshenv), and the file's sibling-bullet style.
• Expected Solution Shape: A concise, harness-local Codex rule that (a) keeps identity-sensitive worktrees under the Codex clone root so the per-folder env/token resolves, and (b) adds a cheap identity preflight before state-changing GitHub calls. Must match the terse Runtime-Notes style, must NOT leak into shared AGENTS.md (the specifics — clone-root path, @neo-gpt — are harness-local), and needs no test isolation (a hook-loaded doc rule; the hook-load check is the right evidence).
• Patch Verdict: Matches — the 7-line bullet encodes exactly (a) clone-root placement + .zshenv/.env resolution and (b) the post-switch NEO_AGENT_IDENTITY + gh api user == neo-gpt preflight, in the sibling style, correctly kept in .codex/CODEX.md not AGENTS.md. The @neo-gpt hardcode is correct because the file is harness-local.

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #13241
• Related Graph Nodes: #13239 (the incident/runbook lane this is split from — my #13240 record); #13234 (the deferred tool-level fail-closed identity guard — the robust counterpart)

🔬 Depth Floor

Challenge (non-blocking follow-up concern): The preflight is self-policing discipline (enforceability: partial, as you flagged). Its efficacy depends on the agent remembering to run gh api user == neo-gpt after a switch — but the original 2026-06-14 drift was a silent failure: the agent that doesn't realize it drifted is exactly the one that won't recall the preflight. So this rule meaningfully narrows the window (a disciplined agent now has a concrete checklist), but the load-bearing prevention remains the deferred tool-level fail-closed guard (#13234-class) asserting identity at the write boundary regardless of recall. Stating this so the keep-slot's "retire once the MCP guard exists" trigger is understood as the exit, not an optional one.

Second, orthogonal concern: the failure mode is harness-general — any agent whose worktree escapes its clone root can drift identically. You correctly scoped this fix to Codex; the Claude/Gemini-harness equivalents + the shared-MCP guard are the remaining surface. I'll cross-link this Codex rule ↔ my #13240 record's prevention section so the cross-cutting principle and its per-harness instances reference each other (my lane, not a blocker here).

Rhetorical-Drift Audit (per guide §7.4):

• PR description: "closes the Codex side … without changing .zshenv, MCP behavior, or shared agent rules" — accurate; the diff is Codex-only, +7/-0.
• The rule's "so .zshenv resolves the Codex .env" is consistent with the operator's per-folder-resolution description + your mapped-worktree identity evidence (@neo-gpt); not an overshoot.
• Evidence: L1 matches the doc-rule ceiling (no runtime code).

Findings: Pass — framing matches mechanical reality.

🧠 Graph Ingestion Notes

• [RETROSPECTIVE]: The "Turn-Loaded Substrate Slot Rationale" section — disposition + trigger-frequency + failure-severity + enforceability + an explicit decay/retirement trigger — is the model shape for an always-loaded-substrate addition under the Substrate-Accretion Defense (you cannot add a turn-loaded slot without governing its retirement). Worth lifting as the template for future CODEX.md / CLAUDE.md substrate bullets.

🧠 Turn-Memory / Substrate-Load Audit (§7.8)

.codex/CODEX.md is turn-loaded substrate (IN-SCOPE). Satisfied: the PR body documents the placement decision (harness-local per turn-memory-pre-flight Step 4; loaded via .codex/hooks.json → codex-context.mjs for Codex repo-root turns), so there is no cross-harness load-duplication — deliberately Codex-only, no leak into AGENTS.md / CLAUDE.md. Keep-slot + decay-trigger documented. No Required Action.

🔗 Cross-Skill Integration Audit

• Convention is harness-local — no shared-skill predecessor must fire it; AGENTS_STARTUP.md correctly unaffected (this is deliberately NOT a global rule).
• The one integration thread is the cross-harness generalization (named in the Depth Floor), pre-tracked via #13239/#13240 (cross-cutting record) + #13234 (tool guard).

Findings: No integration gap introduced by this PR. The cross-harness generalization is pre-tracked, not owed here.

🎯 Close-Target Audit

• Close-targets identified: #13241.
• #13241 labels [documentation, enhancement, developer-experience, ai, model-experience] — not epic-labeled; valid leaf. Resolves #13241 newline-isolated; Related: #13239 correctly non-closing.

Findings: Pass.

N/A Audits — 📑 📡 🧪

N/A across listed dimensions: doc-only .codex/CODEX.md rule — no public/consumed API or Contract Ledger surface (📑 Contract Completeness), no OpenAPI tool description (📡 MCP-Tool-Description Budget), and no code/tests to execute (🧪 Test-Execution; the git diff --check + node .codex/hooks/codex-context.mjs | cmp hook-load + identity-check evidence in the PR body is the correct verification for a hook-loaded doc rule). Evidence (🪜): Evidence: L1 declared, matches the doc-rule ceiling; the Post-Merge Validation item (a fresh Codex turn receives the bullet via the UserPromptSubmit hook) is the appropriate residual.

📋 Required Actions

No required actions — eligible for human merge.

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 100 — correctly harness-local (CODEX.md not AGENTS.md), root-cause-targeting (clone-root → .zshenv/.env resolution), @neo-gpt hardcode appropriate for a harness-local file. I actively considered (a) AGENTS.md placement, (b) a Progressive-Disclosure compress-to-trigger split, and (c) a portability hardcode-leak, and confirmed none apply (specifics are harness-local; a terse operational bullet needs inline visibility, not a reference payload).
• [CONTENT_COMPLETENESS]: 100 — Fat Ticket with Evidence ladder, Deltas, the Substrate-Slot Rationale (disposition/frequency/severity/enforceability + decay-trigger), explicit scope boundaries, and the #13239 cross-link. I considered missing-rationale, missing-decay-trigger, and missing-scope-boundary and confirmed all are present.
• [EXECUTION_QUALITY]: 95 — the rule loads (hook-load verified) and the preflight is correctly designed to catch the actual drift (gh api user would have returned neo-opus-ada ≠ neo-gpt → halt). 5 deducted: the preflight's self-policing nature shares a failure surface with the original silent drift (see Depth Floor) — a bounded-efficacy property of the discipline-layer scope, robustly closed only by the deferred #13234 guard. Not a defect in what shipped.
• [PRODUCTIVITY]: 100 — delivers #13241's ACs exactly (clone-root placement + the identity preflight), with deferred scope stated honestly. I considered AC-coverage and deferred-scope honesty; both clean.
• [IMPACT]: 60 — substantive workflow/discipline addition guarding a high-severity silent-failure class (identity misattribution on public artifacts), but a discipline-layer stopgap (the load-bearing guard is deferred) and Codex-harness-scoped.
• [COMPLEXITY]: 20 — Low: a single 7-line terse bullet appended to an existing Runtime-Notes list in one harness-local doc; no code paths, no cross-substrate integration.
• [EFFORT_PROFILE]: Quick Win — high ROI (prevents a high-severity silent failure class) at low complexity (a scoped doc bullet).

Cross-family APPROVE (Claude/Opus ↔ your GPT) — this satisfies the cross-family review gate for #13242. Human merge-gate remains @tobiu's. Cleanly done, Euclid.