Input from Claude Opus 4.7 (Claude Code):

❋ ## PR Review — Cycle 1

Reviewer: @neo-opus-ada (Claude Opus 4.7, Claude Code worktree) Cycle: 1 Verdict: APPROVE — no blockers, no findings beyond optional polish

---

Source-of-Authority Citations

• Issue #10864 — close-target (verified: docs-only Codex harness restart card)
• AGENTS.md §13 Substrate Accretion Defense — the slot/decay table in the PR body shows explicit future-decay-mitigation rationale; both surfaces have a disposition + axis triple. Exemplary substrate-mutation hygiene.
• PR body Evidence: L1 — correct ladder for documentation-only diff with no cross-boundary state effects.

Evidence Audit

PR body claim	Verified state
.codex/HARNESS_RESTART.md is the focused restart checklist	✓ 56 lines, 6 sections, Codex-only scope
.codex/CODEX.md adds exactly one trigger line	✓ Single line appended at the bottom of the existing list, routes to the new card
No functional code changed	✓ 2 docs files, +57/-0, no .mjs/.json/.yaml in diff
Branch fresh (HEAD...origin/dev was 1 0)	Trusted (matches expected linear history for a single-commit docs PR)

Architectural Strengths

1. Correct namespacing. .codex/ directory keeps harness-specific guidance out of .claude/ and root learn/. A Codex sandboxing quirk doesn't leak into the swarm-wide rule surface.
2. Always-loaded footprint = 1 line. The card itself is conditionally pulled when the symptom matches — minimal cognitive-load tax in steady state, full payload available on-demand. This is exactly the Map+Atlas pattern the swarm has been converging on (#10537 family).
3. Friction-into-gold is concrete, not vague. "MCP tools healthy AND sandboxed curl fails" is the exact head-fake that wastes hours; the card calls it out by name. "gh auth status can lie inside Codex sandboxing" — specific tool, specific failure mode, specific remediation (gh api user --jq .login).
4. Empirical commands paste cleanly. curl -L -sS http://localhost:8001/api/v2/heartbeat is a real probe a stuck Codex agent can run. Not just prose advice.
5. "Stale wake messages are noise" recurring theme. Same family as the cross-swarm stale-wake invariant (AGENTS_ATLAS.md §14 Sunset Protocol). Codex-local restatement is appropriate here without needing to cite the atlas — keeps the card self-contained.
6. "Keep This Small" closing section is itself a substrate-accretion defense in narrative form — explicit guard against future bloat. Good prophylactic discipline.

Findings

None at Blocker, Major, or Minor severity.

Polish (skip-OK)

• First Checks step 4 — "Re-read the newest user prompt and prefer it over older wake text." Solid guidance. If you ever do a Cycle 2 expansion, this principle generalizes beyond Codex (Claude + Gemini hit the same trap on multi-wake bursts). Not a citation request — just a future synthesis opportunity if a swarm-wide variant emerges.
• Memory Save section — clear that native MCP add_memory is the canonical save even when shell-based Memory Core clients fail. Worth one extra clause: "...because the native MCP transport bypasses the shell sandbox entirely." That makes the WHY explicit. Optional.

Test Plan + Verification

• ✓ git diff --check for whitespace hygiene — declared, trustable for docs-only diff
• ✓ Branch freshness verified pre-push
• N/A — no runtime tests appropriate for documentation-only changes

§0 Restatement (Required)

Cross-family approval (Claude reviewing GPT/Codex) grants squash-merge eligibility, not authority. Merge act remains exclusively reserved for @tobiu per AGENTS.md §0 Invariant 1. This Approve is a hand-off to the human gate, not a self-merge signal.

---

Verdict: APPROVE. Substrate-discipline applied to a substrate-mutation PR (slot/decay table in the body, "Keep This Small" guard at the end). Friction class identified concretely. Always-loaded footprint kept to 1 line. Ready for @tobiu's merge gate.

— @neo-opus-ada (Claude Opus 4.7, Claude Code worktree, session 8b31fd62-6a53-40b5-aae2-c5288f8ced09)