Add Codex reviewer backend to structured review

[resodo]

Summary

Add a cross-vendor reviewer backend to the bundled structured-review runner so a Claude Code driver gets a Codex reviewer by default (and vice versa):

• codex exec backend: JSONL event parsing, last-message-file-first text capture, sandboxed write-mode commits via git-dir --add-dir (linked worktrees supported), --sandbox read-only print mode.
• --reviewer-backend {auto,claude,codex} with cross-vendor auto keyed off driver env markers (CLAUDECODE vs CODEX_THREAD_ID/CODEX_SANDBOX); both markers -> hard error; neither -> claude (today's behavior); marker-resolved selections preflight the binary with an actionable error. No silent fallback.
• Pinned Codex defaults gpt-5.5/xhigh with --codex-bin/--codex-model/--codex-effort overrides; --model/--effort/--claude-bin stay Claude-only for backward compatibility.
• Backend-aware metadata (backend, active model/effort, reviewer_version, best-effort token usage) and reviewer identity in every pass heading.
• verify_write_mode/verify_print_mode semantics unchanged — the same post-hoc contract gates both backends.
• Protocol text backend-neutral: SKILL Default Runner Rule, README usage, CURRENT map.

Gates (strict chain, all in the plan thread)

1. Plan review (claude reviewer): no blocking; 5 non-blocking refinements accepted in driver response 1.
2. Implementation review (codex reviewer, dogfooded): the new backend reviewed its own implementation live — real codex exec under --sandbox workspace-write + --add-dir on a linked worktree appended threads and committed; runner verification passed. Verdict: no blocking, no non-blocking, "Ready for closeout", full traceability, reviewer-rerun validation green.

Validation

structured-review/tests 56 OK (16 new), tests 15 OK, scout/tests 15 OK, check_backlog.py pass, compileall clean, git diff --check clean — driver-run and reviewer-rerun; CI runs the same set.

Notes

• closeout/SKILL.md still says "bundled Claude runner" in one line; protocol changes there were an explicit plan non-goal. Optional follow-up backlog item.
• Live spike + dogfood evidence covers real Codex sandbox semantics that fake-bin tests cannot reach.

🤖 Generated with Claude Code

[resodo]

Scope addendum (human-approved during closeout discussion), now included:

1. closeout/SKILL.md delegation — the duplicated (and stale) runner policy is removed; closeout keeps its trigger list and defers runner/backend/mode policy to structured-review.
2. Runner-owned write mode — the reviewer always runs read-only-for-the-repo and returns the review text; the runner verifies it (untouched worktree, review-like, no local paths/secrets), appends it byte-for-byte under ## Review Threads, and creates the commit itself. The append-only contract that reviewer agents historically violated in ~half of write-mode runs now holds by construction. Codex runs --sandbox workspace-write with no .git grant in both modes, so reviewers can run tests while commits remain impossible.

Gate evidence in the plan thread: pass 3 (codex) found one blocking issue (non-verbatim append) plus an environment-blocked-validation residual — both fixed; pass 4 (codex) confirmed: no blocking, no non-blocking, "Ready for closeout", with the reviewer rerunning the full validation suite itself under the new sandbox. Validation green at e8919dd: 61+15+15 tests, backlog check, compileall, git diff --check.