# Cross-review: Codex (`openai/codex`) Rev.2 plan + v4.2.0 implementation (re-review)

- **Reviewer**: `reviewer-creator-grok-01`
- **Job**: `607ee97f` (re-review after `5bfd80b8` NOT PASS)
- **Prior finding**: F1 — OpenCode Rev.3 유실 at `.agents/reports/implementation_plan.md` (`plan-de667f0f.md` was a 16-line stub).
- **Diff**: 39 files +1223/−277, plus untracked `codex.py` and `plan-e30e3123.md`.

---

## 1. Prior F1 — fixed

Independent check (not taken from the creator's word):

| Path | Now |
|---|---|
| `.agents/reports/planner-reviewer-claude-01/plan-de667f0f.md` | 253 lines. 4-line NOTE names it the durable OpenCode Rev.3 archive (job `78f332a9`) and notes the mirror now holds Codex. Body after the NOTE is **byte-identical** to `git show HEAD:.agents/reports/implementation_plan.md` (249 lines). |
| `.agents/reports/implementation_plan.md` | Codex Rev.2. `cmp` identical to `plan-e30e3123.md`. |
| `.agents/reports/planner-reviewer-claude-01/plan-e30e3123.md` | Full Codex Rev.2 (454 lines). |
| Root `implementation_plan.md` | NATS *MAM 메시징 백플레인 전환 실행 로드맵* intact. |
| Codex **Status** | `Implemented (v4.2.0) — full code, tests, and version lockstep present in repo.` (was “Plan only — no code changed”). |

Required fix from `5bfd80b8` items 1–4 all landed. OpenCode Rev.3 is no longer only in git history.

The archived OpenCode body's own “Durable Path” line still lists `.agents/reports/implementation_plan.md` as a mirror — that is the original Rev.3 sentence, preserved because the body is identical. The NOTE header supersedes it. Residual nit, not 유실.

---

## 2. Codex implementation (unchanged, still correct)

Spot-checked after the archive edit; wiring was not regressed.

| Item | Result |
|---|---|
| Adapter | Untracked `codex.py` (158 lines). Class A spawn; `resume_spec` with positional UUID after `--dangerously-bypass-approvals-and-sandbox`; `_state_db()` `state_*.sqlite` glob by mtime; DB-first / glob fallback; `purge_artifacts` `codex delete <uuid> --force` then file+SQL fallback. `py_compile` OK. |
| Challenge 1 | `orc_onboard.sh:133-134` flag-tolerant regex; `codex) env_var=""`; `test_orc_onboard_codex_flag_tolerant_regex` present. |
| Challenge 2 | No hardcoded `state_5` in the adapter; dedicated `# === drift C (codex) ===` with PRAGMA-tolerant order column. |
| Class A resume | `resume_session.sh:65` `agy\|hermes\|opencode\|codex) FRESH_SPAWN=1`. |
| Spawn | `create_session.sh` `agy\|hermes\|grok\|codex)` — not folded into OpenCode's `OPENCODE_PERMISSION` branch. Correct. |
| Shell | `bash -n` on create/resume/stop/orc_onboard/reconcile/status — OK. |
| SemVer | `lib.sh:32` `MAM_VERSION="4.2.0"`; `VERSIONS.md` header/date/8-row matrix/changelog; 8× SKILL.md `version: 4.2.0`. MINOR class matches OpenCode v4.1.0. |

Live CLI facts from the previous pass still hold on this machine (`codex-cli 0.153.4`; `--force` exists; `updated_at` is unix seconds).

---

## 3. Tests

```
pytest tests/test_a4_adapter_contract.py tests/test_c1_tui_readiness.py \
       tests/test_tier1_unit.py tests/test_version_consistency.py \
       tests/test_tier2_component.py -q --tb=line
→ 160 passed in 316.03s
```

---

## 4. Residuals (non-blocking)

1. **TUI empirical gate still open.** `ready_tokens='Codex|Ask Codex to do anything'` remains a candidate. Do not tick plan §7 from this review.
2. **`auth_ok` uses `HOME_DIR`/`HOME`**, not `ctx` — same as grok/agy/opencode; tests patch `HOME_DIR`.
3. **`test_codex_purge_artifacts` exercises fallback only** (`u-del` is not a UUID, so `codex delete --force` fails). CLI-success path untested.
4. **Resume `SKILL.md` Workflow** still interpolates `$UUID` into `CMD_FULL` after the Class A empty-UUID branch. Real `resume_session.sh` uses spawn-spec. Pre-existing nit (`901bfbe8`).
5. Untracked `.agents/reports/reviewer-opencode-01/report-9cc33e22.md` is a peer review written against the pre-archive tree (that report's NOT PASS is now stale). Not a Codex code defect.

No `[ESCALATE: PLANNER]`.

---

[VERDICT: PASS]