Add Codex reviewer backend to structured review

README.md

@@ -73,7 +73,7 @@ Please read ~/.agent-protocols/closeout/SKILL.md
 Scope: current implementation task
 ```
 
-For Claude Code structured-review reviewer passes, use the bundled runner:
+For structured-review reviewer passes, use the bundled runner:
 
 ```bash
 python structured-review/scripts/claude_structured_review.py \
@@ -89,6 +89,11 @@ python structured-review/scripts/claude_structured_review.py \
 The runner loads the shared skill from this repo and target-repo overlays from
 the explicit `--worktree`.
 
+The runner drives either Claude Code or Codex as the reviewer.
+`--reviewer-backend` defaults to `auto`, which picks the cross-vendor
+reviewer for the detected driver (Claude Code driver -> Codex reviewer, Codex
+driver -> Claude reviewer); pass `--reviewer-backend claude|codex` to pin it.
+
 `--protocol-dir` is optional and mainly for tests or intentional alternate
 checkouts; normal usage relies on the script's own `structured-review`
 directory.

closeout/SKILL.md

@@ -44,10 +44,9 @@ This trigger list is the canonical Closeout Review trigger list. Maintain it
 through structured-review skill self-evolution when protocol rules change.
 
 When Closeout Review is triggered, use the `structured-review` protocol with
-`Type: closeout-review`. For repo-backed Closeout Review, use the bundled
-Claude runner when available. Default to `print-review` for durable closeout
-reports or reference artifacts; use `write-commit-to-plan` only when the driver
-explicitly provides a thread file with `## Review Threads`.
+`Type: closeout-review`. The structured-review skill's default runner rule and
+runner-mode guidance decide how the reviewer pass runs; this protocol does not
+duplicate that policy.
 
 Closeout Review checks whether the closeout evidence or report is accurate and
 complete. It returns to closeout with `ready to resume closeout` or named

docs/CURRENT.md

@@ -1,7 +1,7 @@
 # Current Protocol Map
 
 Status: active source-of-truth map
-Last updated: 2026-06-06
+Last updated: 2026-06-10
 
 This file points humans and agents to the active protocol entrypoints. It does
 not replace root `AGENTS.md`; agents should start there, then use this map to
@@ -19,8 +19,8 @@ avoid treating historical plans as current instructions.
 
 - `structured-review/SKILL.md` - structured review for plans,
   implementations, optional closeout evidence reviews, and review-response
-  passes. Repo-backed review gates default to the bundled Claude runner when
-  available.
+  passes. Repo-backed review gates default to the bundled runner when
+  available, with cross-vendor Claude and Codex reviewer backends.
 - `closeout/SKILL.md` - final hygiene checklist and delivery handoff after
   implementation, documentation, review, or phase work.
 - `planning/SKILL.md` - discussion-to-plan workflow before implementation.

docs/agent_plans/README.md

@@ -28,3 +28,5 @@ responses, and closeout evidence for changes to this repo's protocols.
   provenance hardening plan.
 - `2026-06-10_scout_validation_provenance_closeout.md` - closeout report for
   the Scout validation provenance work (PR #13).
+- `2026-06-10_codex_reviewer_backend_plan.md` - Codex reviewer backend and
+  cross-vendor default runner plan.

structured-review/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: structured-review
-description: Use when a written reviewable artifact needs structured human-agent review, including plans, implementation plans, implementation reviews, optional closeout evidence reviews, and review-response passes. Defaults repo-backed review gates to the bundled Claude runner when available, with explicit reviewer/driver roles and driver-owned post-review decisions.
+description: Use when a written reviewable artifact needs structured human-agent review, including plans, implementation plans, implementation reviews, optional closeout evidence reviews, and review-response passes. Defaults repo-backed review gates to the bundled runner when available, with Claude and Codex reviewer backends, explicit reviewer/driver roles, and driver-owned post-review decisions.
 ---
 
 # Structured Review
@@ -48,26 +48,40 @@ skill manually, read these references before judging readiness.
 For high-touch product UI, dashboards, charts, data review tools, and operator
 consoles, also apply `references/ui-review.md`.
 
-## Default Claude Runner Rule
+## Default Runner Rule
 
 For repo-backed Plan Review, Plan Re-review, Implementation Review, and
 Implementation Re-review gates, the driver defaults to invoking the bundled
-Claude runner for the reviewer pass when it is available.
+runner for the reviewer pass when it is available.
 
 `Closeout Review` is not part of this default gate list. If closeout explicitly
 triggers a repo-backed Closeout Review, use the bundled runner when available;
 the trigger comes from closeout, not from structured-review by default.
 
+The runner supports two reviewer backends: `claude` (Claude Code) and `codex`
+(Codex CLI). `--reviewer-backend` defaults to `auto`, which selects the
+cross-vendor reviewer from the driver's environment: a Claude Code driver gets
+the `codex` reviewer and a Codex driver gets the `claude` reviewer. When both
+driver markers are present the runner fails and requires an explicit
+`--reviewer-backend`; when neither is present it uses `claude`. Review threads
+record the reviewer backend in each pass heading.
+
+If an auto-selected backend binary is unavailable, the runner fails with an
+actionable error. Substituting the same-vendor reviewer is an explicit
+human/driver decision via `--reviewer-backend`, never a silent fallback.
+
 `Default` means use the runner unless:
-- the human explicitly says not to use Claude;
+- the human explicitly says not to use the runner or a specific reviewer
+  backend;
 - the review is not repo-backed;
 - the runner is unavailable and the blocker is reported.
 
 The runner is available when:
 - the runner script exists in this protocol checkout or submodule;
 - the target worktree is a git repository;
 - Python can run the script;
-- `claude` is available through the configured `--claude-bin`.
+- the selected reviewer backend binary is available through the configured
+  `--claude-bin` or `--codex-bin`.
 
 Do not silently substitute driver self-review, chat-only commentary, or a
 manually constructed reviewer prompt for a required repo-backed review gate. If
@@ -89,6 +103,11 @@ python structured-review/scripts/claude_structured_review.py \
   --topic "example plan"
 ```
 
+In both modes the reviewer runs read-only and returns the review text. In
+`write-commit-to-plan` the runner itself appends that text verbatim under the
+thread file's `Review Threads` section and creates the `structured-review:`
+commit, so the append-only thread contract holds by construction.
+
 Use `print-review` when appending review threads would pollute a durable
 reference artifact, such as a published doc, skill, protocol overlay, or other
 reference file.
@@ -100,6 +119,9 @@ provides a thread file that already has a `## Review Threads` section.
 The explicit runner `--mode` is the write-back authorization for that run. The
 driver still owns the decision after receiving reviewer output.
 
+Pass `--reviewer-backend claude|codex` to pin the reviewer backend; the
+default `auto` selects the cross-vendor reviewer for the detected driver.
+
 Use `--protocol-dir` only when testing or intentionally running against a
 different checkout of this protocol. By default the runner uses the
 `structured-review` directory that contains the script.
@@ -199,7 +221,7 @@ Pause and discuss with the human before editing when a reviewer finding:
 
 After applying or rejecting reviewer feedback, the driver summary must
 distinguish:
-- Claude reviewer suggestions;
+- reviewer suggestions, named by reviewer backend;
 - driver decisions;
 - human decisions;
 - remaining risks;