feat(reviewing-prs): native review recovery — one codex reviewer for both species

docs/doperpowers/2026-07-10-codex-workers-shakedown.md

@@ -62,12 +62,29 @@ number in a local clone, so a **manual** review shakedown needs no runner.
     correctness **and** the ticket's spec-compliance criteria, posts findings
     + a verdict, and transitions the board (or comments on a PR with no linked
     issue). The review-trail comment names `codex` as the engine.
-- [ ] **SD-4 · review × Claude.** Same PR (or SD-2's), dispatched with
-  `WORKER_ENGINE=claude` or `engine:claude` → `daemon-spawn.sh` Claude
-  reviewer, with the `fallback-claude` block.
-  - **Acceptance:** a Claude reviewer produces findings + verdict over the
-    same diff and criteria; review-trail names `claude`. Confirms both the
-    engine switch and the differing fallback blocks.
+- [ ] ~~**SD-4 · review × Claude.**~~ OBSOLETE 2026-07-12 — the native-review
+  recovery (`specs/2026-07-12-native-review-recovery-design.md`) deleted the
+  Claude reviewer fallback and the per-species fallback blocks this cell was
+  written to exercise; the reviewer is codex-only now. The `engine:claude`
+  worker SHELL still exists (calls the same `review-engine.sh` non-nested)
+  but has no separate engine to shake down. Superseded by SD-5 below.
+
+### SD-5 · review × Codex × native engine (native-review recovery) — PASSED 2026-07-12
+
+- [x] **SD-5 · native `codex exec review` end-to-end.** PASSED 2026-07-12
+  (doperpowers PR #9 — the recovery branch reviewing itself, dogfood).
+  Codex worker (gpt-5.6-sol, xhigh) ran `review-engine.sh` NESTED: 3 native
+  review rounds against `origin/main...HEAD`, 7 findings verified against
+  the code and ALL routed FIX NOW (7 commits pushed to the PR head),
+  round 3 clean → approve. Tier judgment correct on all four failing
+  clauses (size >150 lines, base=default branch, auth risk surface, no CI
+  checks) → HUMAN tier, `confident-ready` label, no board writes (no
+  linked ticket). Zero `sandbox_apply` / code-mode-host errors. The worker
+  ran with the PRE-minimization criteria (commit 0d00934 landed mid-review);
+  its own round-2 [P1] (PR/ticket text elevated into developer
+  instructions) produced the untrusted-carrier hardening that was then
+  reconciled with the human's minimal-content directive (spec Revision
+  Notes 2-4).
 
 ## Recommended path (covers all four, no runner)
 
@@ -285,6 +302,16 @@ command-running work (pure-inference nested calls do work — FU-6's probe).
 - All three SD-3 rounds were protocol-clean failures (honest trails,
   correct parks, no fabricated verdicts). Round 4 runs the in-thread
   design.
+- **Correction (2026-07-12):** "structurally dead" is overturned — the
+  `sandbox_apply` wall is skipped by running the INNER codex with
+  `-c sandbox_mode="danger-full-access"` (it only skips self-profiling;
+  the outer profile still confines every child), plus
+  `CODEX_CODE_MODE_HOST_PATH=$HOME/.local/bin/codex-code-mode-host` for a
+  third, newly-found wall (nested codex misresolves the code-mode host to
+  `/usr/local/bin/`). Nested `codex exec review --base` verified
+  end-to-end (rc=0, both planted findings). The in-thread design this FU
+  produced is superseded by
+  `specs/2026-07-12-native-review-recovery-design.md`.
 
 ### Remaining engine asymmetries — audited, deliberate (not gaps)
 

docs/doperpowers/specs/2026-07-10-codex-workers-design.md

@@ -709,3 +709,25 @@ mechanical rather than lucky.
    `codex-run.*` sets that no live meta's `event_log` references, called at
    spawn/resume, plus purge-time removal of a daemon's own run files. Both
    reviewed and green across the codex, daemon, and review-dispatch suites.
+6. **2026-07-12 (native-review recovery — two conclusions overturned).**
+   Live spikes (codex-cli 0.144.1, gpt-5.6-sol) overturned two of this
+   spec's load-bearing conclusions. (a) Task 2's "no working invocation at
+   all" for `codex exec review --base` + custom criteria (Revision Note 2's
+   promotion of the cookbook form to "the only confirmed path") holds only
+   for criteria as the positional `[PROMPT]` — criteria as a **config
+   value** (`-c developer_instructions="…"`) pass the clap parser and
+   compose correctness + spec-compliance in one native call (rc=0, full
+   multi-commit range, compact `-o` verdict). (b) FU-7's "codex-in-codex is
+   structurally dead on macOS for command-running work" (recorded in the
+   shakedown doc, reflected in the in-thread engine-block design) is
+   overturned by an environment recipe: `SSL_CERT_FILE` (FU-6), inner
+   `-c sandbox_mode="danger-full-access"` (skips the impossible second
+   Seatbelt profile; the outer profile still confines all children), and
+   `CODEX_CODE_MODE_HOST_PATH=$HOME/.local/bin/codex-code-mode-host`
+   (nested codex misresolves the code-mode host) — verified end-to-end:
+   a nested `codex exec review --base` under a seatbelted outer codex
+   returned rc=0 with both planted findings. The species-split engine
+   design (Codex reviews in-thread / Claude calls the cookbook form) is
+   **superseded** by `specs/2026-07-12-native-review-recovery-design.md`:
+   one native `codex exec review` engine for both species, owned by a
+   substrate script.