Adversarial Review

Adversarial code review

Two reviewers from different model families — Claude and Codex/GPT — review the same diff independently, then each tries to refute the other's findings. A finding's confidence comes from whether it survives that cross-examination. This kills the two failure modes of solo LLM review: self-ratification (a model won't critique its own work) and confident false positives.

You are the orchestrator — and one of the two reviewers

This skill runs from either Claude Code or Codex. First, identify which model family you are (Claude or Codex/GPT). Then:

• You are reviewer #1. You review natively, in this session, using your own tools.
• The other family is reviewer #2. You invoke it as a subprocess CLI for an independent pass: a fresh process, no shared context — that independence is the point.

The CLI for "the other model":

Everything else in the flow is symmetric. Resolve the prompts/ and schemas/ paths below relative to this skill's own directory (where this SKILL.md lives).

Inputs

Two independent, optional inputs:

• BASE — the ref to diff against. Default main.
• SCOPE — a pathspec to narrow the review (e.g. src/basic_memory). Default: none (whole diff).

These are separate: a ref and a pathspec are not interchangeable. Build the canonical diff command once in preflight and reuse it everywhere below — never re-spell the diff inline (the scattered, inconsistent spelling is what broke earlier). Build it as an argv array, not a string, so a $SCOPE containing spaces or glob characters survives intact:

BASE="${BASE:-main}"
DIFF=(git diff "$BASE...HEAD")          # argv array — never a scalar string
[ -n "$SCOPE" ] && DIFF+=(-- "$SCOPE")  # pathspec stays one argument even with spaces
DIFF_STR=$(printf '%q ' "${DIFF[@]}")   # shell-quoted rendering, for embedding in a prompt

To run it, use "${DIFF[@]}" (quoted, no word-splitting). To embed it as text inside a subprocess prompt, use $DIFF_STR.

Preflight

0. Set SKILL_DIR to the directory this SKILL.md lives in. Canonical location is .agents/skills/adversarial-review (the shared agent-skills store); Claude Code reaches it via the .claude/skills/adversarial-review symlink, Codex via its own skills path. The prompts/ and schemas/ subdirs are siblings of this file in every case.
1. Confirm the other model's CLI is on PATH (codex if you're Claude, claude if you're Codex). If it's missing, tell the user the panel falls back to single-model (which loses the cross-vendor benefit) and ask whether to proceed or stop.
2. Run "${DIFF[@]}". If it prints nothing, report "nothing to review against $BASE" (mention $SCOPE if set) and stop.
3. RUN=$(mktemp -d) — scratch dir for the other model's output. Transient, never committed. No persisted artifacts, no state file.

Phase 0 — Deterministic gates (before the models)

Models are statistically blind to negation ("never do X"). Enforce mechanical house rules with tools, not prompts, and treat hits as high-confidence facts (reported separately from model findings):

• just lint and just typecheck if the diff touches src/.
• Grep the diff for catchable house-rule violations: getattr(.*,.*, defaults, bare except: / except Exception: pass, function-scope imports.

Phase 1 — Independent review (you + the other model, concurrently)

Both reviewers get the same brief: prompts/review.md + the repo's CLAUDE.md house rules, reviewing the diff from "${DIFF[@]}". Both emit findings matching schemas/findings.schema.json.

Your native pass: review as yourself, following prompts/review.md. Hold your findings as that JSON shape.

The other model's pass — run, from the repo root, the row that matches you:

Always redirect codex stdin from /dev/null — if stdin is a pipe (e.g. the call gets backgrounded), codex exec blocks "Reading additional input from stdin..." and fails.

# You are Claude → run Codex:
codex exec -s read-only \
  --output-schema "$SKILL_DIR/schemas/findings.schema.json" \
  -o "$RUN/other_findings.json" \
  "$(cat "$SKILL_DIR/prompts/review.md")

Review the diff: $DIFF_STR" </dev/null

# You are Codex → run Claude (read-only via plan mode; parse the JSON block it returns):
claude -p --permission-mode plan --output-format json \
  "$(cat "$SKILL_DIR/prompts/review.md")

Review the diff: $DIFF_STR
Return ONLY a JSON object matching this schema:
$(cat "$SKILL_DIR/schemas/findings.schema.json")" </dev/null > "$RUN/other_raw.json"
# claude --output-format json output shape varies by CLI version: it may be a JSON ARRAY
# of event objects, OR a single result object. Normalize before reading: if it's an array,
# take the element with type=='result'; otherwise use the object as-is. Then read its
# .result string, strip the ```json fence if present, and parse that.
# (Verified empirically: the CLI in this environment emits the array form.)

Runtime note for Codex orchestrating: claude -p needs network access, which Codex's default sandbox blocks. Run it from a Codex session whose project is trusted with network allowed (or approve the claude call when prompted). Keep Codex's own sandbox on — do not bypass it just to reach the network.

Tag each finding with its origin (claude / codex).

Phase 2 — Cross-refute

Each model tries to refute the other's findings, per prompts/refute.md (verdicts match schemas/verdicts.schema.json).

• You refute the other model's findings natively.
• The other model refutes your findings — invoke it again the same way (swap prompts/review.md for prompts/refute.md, append your findings JSON and $DIFF_STR so it judges against the right base and scope, and for Codex use --output-schema "$SKILL_DIR/schemas/verdicts.schema.json").

Match verdicts to findings by id.

Phase 3 — Synthesize and report (no auto-fix)

Merge, dedupe (same file + overlapping lines + same root cause = one finding), assign confidence from provenance:

• High — both models raised it independently, OR one raised it and the other upheld it.
• Medium — one raised it; the other could not refute it but did not independently find it.
• Low / contested — one raised it and the other refuted it. Keep it, show both sides, let the human judge. Never silently drop a contested finding.
• Deterministic-gate hits are reported as facts, separate from the model panel.

Rank by severity × confidence. Present a compact table: severity | confidence | file:line | claim | found-by / upheld-or-refuted-by. Expand the high-confidence ones with why and any suggested fix.

End by asking which findings, if any, to fix. Do not edit code until the user picks. Convergence between the models is not correctness — your job is to surface a ranked, cross-examined list, not to declare the branch clean.

Deliberately NOT done

• No loop-until-both-agree (models converge by going silent, not by being right).
• No persisted artifacts / state machine — the scratch dir is thrown away.
• No auto-applying fixes.