feat(skill): add --coder codex reversed-roles support

[wwind123]

Summary

• Adds --role {reviewer,coder} argument to helpers/run_external.py so Codex can be invoked for coder (plan-writing) turns, not just reviewer turns
• Dry-run path for --role coder produces a valid plan_state stub (markdown + <!-- AGENT_PLAN_STATE: approved -->)
• Documents the reversed-roles flow in a new Reversed roles section in SKILL.md
• Updates .claude/commands/coding-review-agent-loop.md to accept --coder codex and route to the reversed-roles flow
• Adds a test (test_dry_run_coder_role_exits_zero_and_writes_plan_state_stub) that validates the dry-run stub passes plan_state validation

Closes #285

Test plan

• pytest tests/test_skill_helpers.py — all 12 tests pass
• New test covers --role coder --dry-run exit 0 and plan_state validation of the stub output

🤖 Generated with Claude Code

[wwind123]

Review verdict: Blocking

Reversed-role dry-run support is added, but the documented skill flow does not configure Claude as the reviewer for resume/approval tracking.

Blocking issues

• SKILL.md:221 and .claude/commands/coding-review-agent-loop.md:8 leave Step 1 and reviewer parsing on the normal reviewers set even when --coder codex switches Step 6 to a Claude review. Resume matching only counts metadata whose agent is in configured_reviewers (src/coding_review_agent_loop/round_state.py:515), so a posted --agent Claude review is ignored under the documented/default reviewer config (gemini, or any external reviewer list). That means reruns will not recognize Claude's completed review and the reversed-role loop can repeat review turns or fail to reach a stable approved state. The reversed-role instructions need to explicitly configure/pass claude as the reviewer for build-resume and approval tracking, or otherwise adapt the helper flow to recognize the host Claude review.

-- OpenAI Codex

[wwind123]

Review verdict: Approved

The PR successfully implements support for using Codex (or Gemini) as the coder agent, enabling the 'reversed roles' flow (Codex as coder, Claude as reviewer) as described in Issue #285. The orchestrator has been effectively refactored to be agent-agnostic, and the Skill mode helpers now correctly handle these alternative configurations. Test coverage is comprehensive, including new integration tests for the reversed roles flow.

Future follow-ups

• Consider adding session resumption support (session_id) to CodexBackend if the Codex CLI introduces a corresponding --resume or --session flag.

-- Google Gemini

[wwind123]

Review verdict: Approved

Round 2 review found no blocking issues; the reversed-role resume tracking fix now configures Claude as the reviewer.

Prior unresolved item dispositions

• [item-1] Blocking issue from OpenAI Codex, round 1: missing --reviewers claude in reversed-roles build-resume -> resolved

-- OpenAI Codex

[wwind123]

Review verdict: Approved

Verified that item-1 from round 1 has been addressed. SKILL.md and the Claude command instructions now correctly specify --reviewers claude for the build-resume step when using reversed roles, ensuring that Claude's review turns are properly tracked and resumed. The implementation is consistent with the existing agent mapping logic.

Prior unresolved item dispositions

• [item-1] Blocking issue from OpenAI Codex, round 1: missing --reviewers claude in reversed-roles build-resume -> resolved

-- Google Gemini