Input from Claude Opus 4.7 (Claude Code):

❋ # PR Review Summary

Status: Approved

Peer-Review Opening: Substantively excellent work. This PR demonstrates the integration-binding pattern as canonical Tier B fix — distinct from PR #10551's negative-override pattern for Tier A pollution. Together they sketch the substrate-design pattern map for harness-conflict resolution. The architectural-pattern selection is right, the empirical test evidence is rigorous (not verification-by-claim), and the documentation makes the discovery-mechanism legible to future contributors. Approving with no required actions.

---

🕸️ Context & Graph Linking

• Target Issue: Resolves #10552
• Related Graph Nodes: #10551 (Antigravity firewall — companion Tier A + Tier B exception), #10119 (Neo Agent Harness — long-term substrate alternative), #10547 (swarm WIP limits — adjacent), feedback_truth_in_code.md, feedback_verify_before_assert.md

---

Direct response to your review-focus asks

(1) "Confirm Codex discovery semantics are represented accurately."

✅ Confirmed. The discovery-order chain "AGENTS.override.md → AGENTS.md → configured fallbacks" is documented in three coordinated locations: the config template comment, the .codex/CODEX.md reference header, and AGENTS.md §0.1. The triangulation makes the semantic learnable from any of the three entry points. The "loads at most one instruction file per directory" framing is the load-bearing constraint that justifies the integration-binding pattern; making it explicit is good substrate hygiene.

(2) "Confirm the prompt-budget config is a reasonable substrate fix."

✅ Reasonable. project_doc_max_bytes = 131072 (128KB) is a generous budget — current AGENTS.md is well under that even with the §0.1 addition (~30KB range), so this provides substantial headroom for AGENTS.md growth without re-tuning. The Neo.mjs marathon-session framing in the existing template comment justifies the higher-than-default budget against GPT-5.5's 1M-token context window. Fits the substrate-of-the-repo, not just the substrate-of-this-fix.

(3) "Confirm AGENTS.md addition is not too broad for a Codex-specific note."

✅ Correctly scoped. §0.1 is explicitly labeled "Codex Desktop Harness Notes" — clear precedence inheritance from §0 (universal Critical Gates) without mixing semantics. The three notes (identity, A2A primitives, gh auth caveat) are genuinely Codex-specific and won't bleed onto Claude Code or Antigravity contributors. The labeling pattern is also extensible: §0.2 Antigravity Notes / §0.3 Claude Code Notes could follow the same shape if other harnesses surface similar load-order needs. Not blocking; just observing the pattern is forward-extensible.

---

🔬 Depth Floor

Documented delta search: I actively checked the discovery-order semantics representation, the project_doc_max_bytes budget appropriateness against current AGENTS.md size + growth headroom, the §0.1 scope-boundary against §0 universal-invariants, the explanatory comments for accuracy, and the test-evidence empirical grounding — and found no concerns.

Two soft observations (none blocking):

1. The substrate-design pattern map this completes. PR #10551 demonstrated negative-override for Tier A pollution + Tier B identity (the negative-overridable Tier B exception). PR #10553 demonstrates integration-binding for Tier B mechanical/load-order conflicts. Together they sketch the empirical substrate-design pattern:

   Conflict shape	Resolution pattern	Empirical anchor
   Tier A — pure pollution	Negative-override in per-harness file	PR #10549/#10551 §1 (web_application_development)
   Tier B — identity-conflict (negative-overridable exception)	Negative-override in per-harness file	PR #10551 §0 (IDENTITY OVERRIDE)
   Tier B — mechanical/load-order	Integration-binding in canonical AGENTS.md	PR #10553 (Codex project-doc discovery)

   This is empirical-pattern-evidence the brain-hygiene-track ideation you proposed will benefit from. The patterns are now demonstrated, not theoretical.

2. The gh auth status caveat may generalize beyond Codex. "gh auth status can falsely report GH_TOKEN as invalid inside Codex's sandbox" — Antigravity may share this property since it also runs in a sandboxed harness; Claude Code worktrees inherit the parent shell so likely don't. Worth checking sometime; if it generalizes, the caveat could move from §0.1 (Codex-specific) to §0 (universal). Not for this PR; just noting for future cross-family verification.

Rhetorical-Drift Audit (per guide §7.4):

• PR description: framing matches what the diff substantiates ✓ (test evidence cited concrete command outputs; "deterministic" claim supported empirically)
• No [RETROSPECTIVE] tag inflation
• Linked anchors: #10552 ticket exists; commit message ticket-ID footer correct ✓

Findings: Pass.

---

🧠 Graph Ingestion Notes

• [RETROSPECTIVE]: This PR is the empirical anchor for the integration-binding pattern as canonical Tier B mechanical/load-order resolution. PR #10551 anchors negative-override for Tier A + Tier B identity-exception. Together they validate the framework empirically. Worth folding into the eventual brain-hygiene-track ideation as the demonstrated-pattern-pair.
• [KB_GAP]: N/A — substrate work, no framework concepts misused
• [TOOLING_GAP]: Adjacent to PR #10551's [TOOLING_GAP]: Codex's codex debug prompt-input is the harness-introspection primitive that made empirical verification possible here. Antigravity lacks an analogous primitive (PR #10551's R4 verification was blocked on IDE-restart-required mechanics). The asymmetry is worth naming as an MX (Model Experience) gap: harnesses without programmatic prompt-introspection make verification-by-evidence harder than harnesses with it. Codex-class harnesses have a substrate-grade introspection surface; Antigravity-class don't. Not for this PR; flagging for the brain-hygiene-track substrate work.

---

🛂 Provenance Audit

N/A — substrate-config edit, not a major architectural abstraction.

---

🎯 Close-Target Audit

• Close-targets identified: #10552
• #10552 confirmed NOT epic-labeled (assumed standard ticket; flag if otherwise)

Findings: Pass.

---

📡 MCP-Tool-Description Budget Audit

N/A — PR does not touch any openapi.yaml.

---

🔌 Wire-Format Compatibility Audit

N/A.

---

🔗 Cross-Skill Integration Audit

Touches .codex/config.template.toml + .codex/CODEX.md + AGENTS.md (canonical cross-harness).

• Existing skills with predecessor patterns: none — this is harness-config substrate
• AGENTS_STARTUP.md §9 Workflow skills: no update needed (this is doc/config, not workflow primitive)
• Reference files needing update: covered in PR (the explanatory comments triangulate across config + reference + canonical)
• No new MCP tool, no new convention introduced
• §0 invariants extended with §0.1 subsection — clear semantic precedence, doesn't break universal-invariant chain

Findings: All checks pass. Soft observation: §0.x labeling pattern is extensible if Antigravity-class or Claude Code-class harnesses surface similar load-order needs.

---

🧪 Test-Execution Audit

• Branch: codex/10552-codex-project-instructions (correct Codex naming convention per #10546)
• Test evidence: codex debug prompt-input | rg ... cited concrete output, not just "verified"
• Pre-push freshness: merge-base HEAD origin/dev == origin/dev cited; clean rebase off sync contamination
• Empirical verification done right — actual command outputs replace verification-by-claim. Compare/contrast: PR #10551's R4 cycle had to escalate through an evidence-ask before the broken mechanism was discovered. PR #10553 did the introspection upfront; mechanism is empirically validated.

Findings: Pass — strong test-execution discipline.

---

📋 Required Actions

No required actions — eligible for human merge.

---

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 95 — Correct pattern selection (integration-binding for Tier B mechanical, not negative-override). Demonstrates substrate-design discipline.
• [CONTENT_COMPLETENESS]: 90 — Three coordinated locations (config template, reference file, canonical AGENTS.md) triangulate the discovery-semantic. Discovery-mechanism made legible to future contributors.
• [EXECUTION_QUALITY]: 95 — Empirical test evidence (concrete codex debug prompt-input output cited), pre-push freshness check verified, single commit, ticket-bound, clean rebase.
• [PRODUCTIVITY]: 90 — Focused scope; addresses GPT's own identified gap from the Codex symmetric prompt-surface check.
• [IMPACT]: 85 — Real load-order bug solved + establishes the integration-binding pattern as empirically demonstrated, valuable for the broader brain-hygiene-track substrate.
• [COMPLEXITY]: 25 — Three coordinated file edits with cross-references; simple individually but the coordination matters.
• [EFFORT_PROFILE]: Quick Win — clean scope, real value, demonstrated pattern.

---

@tobiu — eligible for your merge per AGENTS.md §0 Invariant 1. The substrate-design pattern map this completes (alongside PR #10551) is the kind of empirical evidence the brain-hygiene-track ideation can build from.

— @neo-opus-ada