Codex adversary

/codex-adversary — Cross-Model Adversarial Review + Triage

Run Codex as adversarial reviewer (different model family → different blind spots than Claude), then auto-triage findings and drive Claude implementation of approved ones.

When to Use This vs /swarm-review

Skill	Reviewer	When
/swarm-review	Claude sub-agents (multi-perspective: security, perf, architecture, SOTA)	Fast intra-family multi-angle review. Same training data, same blind spots.
/codex-adversary	Codex CLI (GPT-5.x family) via plugin runtime	Cross-model second opinion. Slower. Catches things Claude family miss.

Use both on big work — complement each other. /swarm-review first (cheaper, faster), then /codex-adversary for independent cross-model challenge before merge.

Model selection (tiered)

Codex has three GPT-5.6 tiers (luna≈Haiku, terra≈Sonnet, sol≈Opus); match model weight to review weight. Callers pass --model <alias> per tier — full alias→slug table + resolution rules: references/model-selection.md. Default when --model omitted → gpt-5.6-terra.

Context for Codex

Codex reads AGENTS.md at repo root every invocation. Grimoire AGENTS.md already contains:

• Tech stack and golden-path choices
• Path → subsystem rule map (so Codex consults .claude/rules/subsystem-*.md before flagging)
• Language-level quality rule pointers (.claude/rules/quality-*.md)
• Hard-rule conventions (commits, git workflow, code patterns)
• Security attack surfaces
• Review guidance framing adversarial pass

Do not re-inject this context in review prompt — Codex already loads it. Pass extra focus text only when user wants to steer review to specific concern.

Workflow

1. Scope detection

Two scope families, caller picks:

• code-diff (default; scopes working-tree / branch / --base) — reviews git diff. Used by /swarm-execute and direct user invocation.
• plan-artifact (scope plan-artifact, --target-file <path>) — reviews plan / ADR / design markdown file. Invoked by /swarm-plan after Claude review panel converges. See references/plan-artifact-scope.md.

For code-diff scopes, run in parallel:

• git status --porcelain=v1
• git diff --shortstat and git diff --shortstat --cached
• git log -5 --oneline
• git rev-parse --abbrev-ref HEAD

Pick review scope:

• working-tree (default): unstaged + staged + untracked changes
• branch: main..HEAD — everything since branch diverged from main
• --base <ref>: user-supplied base
• plan-artifact: file path via --target-file is target; skip git probes.

If user passed focus text as argument, treat as intent (e.g., "security review of the new archive extractor") and pass through to Codex unchanged. No rewrite.

2. Run the review via the plugin runtime

Use Codex companion script directly — same entry point plugin /codex:adversarial-review command uses. Keeps Codex context loading, auth, runtime identical to first-party command.

Resolve the --model alias to its slug (Model selection; default gpt-5.6-terra), pass as --model <slug>:

node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-companion.mjs" adversarial-review --model "<slug>" "$ARGUMENTS"

CLAUDE_PLUGIN_ROOT env var set by Claude Code harness when Codex plugin loaded. If unset (plugin missing), stop and tell user to install/enable Codex plugin.

Execution mode:

• --wait in args → foreground
• --background in args → Claude background Bash task
• Neither → estimate size. Small (1–2 files, no directory-wide change): ask via AskUserQuestion, two options (Wait for results (Recommended) vs Run in background). Bigger: default background, tell user check /codex:status.

3. Auto-triage (Claude's job)

When review returns, Claude processes Codex free-text output and makes triaged report. No passthrough-prompt user for every finding — filter smart.

See references/triage.md for full triage class table (Auto-fix / Needs confirmation / Discuss / Filtered-trivia / Filtered-stated-convention / Filtered-false-positive), heuristics for "safe to auto-fix", and triaged report format.

4. Implementation (auto-fix items)

For each auto-fix item:

1. Read affected file(s) to confirm Codex description matches reality
2. Apply fix with Edit (not Write)
3. Run targeted checks — cargo check -p <crate> for Rust, relevant task lint:* for markdown/config, etc.

After all auto-fixes:

4. Run task verify once (not per file)
5. Update report with outcome of each auto-fix (✅ applied / ⚠️ reverted-and-promoted / ❌ failed)

Do not commit. Commit always user explicit call — follow /commit conventions when invoked.

5. Needs-confirmation items

After auto-fix batch done, walk through needs confirmation items one at a time using AskUserQuestion with Apply / Skip / Modify options. Apply approved fixes same way as auto-fix items (Edit + targeted check + final task verify).

6. Discuss items

Present verbatim at session end. Claude may add own read per item, but no forceful advocate — these user decisions.

Flags

• --wait — foreground execution, show review as soon as returns
• --background — run review as Claude background Bash task, return immediately; user checks /codex:status then asks for triage
• --base <ref> — review ref..HEAD instead of working tree
• --scope auto|working-tree|branch|plan-artifact — explicit scope
• --target-file <path> — required when --scope plan-artifact; plan/ADR markdown file to hand to Codex
• --model <luna|terra|sol|slug> — Codex model tier (see Model selection). Alias or full slug; default terra (gpt-5.6-terra)
• --no-auto-fix — skip auto-fix batch; treat every real finding as needs confirmation. Use on sensitive code or when want see everything before change lands.
• --review-only — run Codex review, dump raw output, skip triage entirely. Equivalent to running /codex:adversarial-review directly.

Plan-artifact scope (invoked by /swarm-plan)

When called with --scope plan-artifact --target-file .claude/state/plans/plan_<feature>.md, contract changes — triage simplified (no auto-fix batch; plans edited, not compiled) and pass one-shot. See references/plan-artifact-scope.md for full contract.

Failure modes

See references/failure-modes.md — Codex unavailable, no changes to review, empty/garbled Codex output, task verify failures after auto-fix.

References

• AGENTS.md — project context Codex loads every invocation
• /codex:adversarial-review — first-party plugin command; this skill wraps with Grimoire-specific triage
• /swarm-review — Claude-intra-family multi-perspective review
• .claude/rules/quality-core.md — anti-pattern severity definitions (Block / Warn / Suggest) used in triaged report
• .claude/skills/builder/SKILL.md — implementation pattern for applying approved fixes
• references/triage.md — classes, heuristics, report template
• references/plan-artifact-scope.md — plan-artifact contract
• references/failure-modes.md — error paths