Agent-Skills.md

Agent Skills: dual-review

Run the ca dual-model review (Codex second opinion + blind Claude review + Claude synthesis) standalone on any PR — no implement loop required. Use to get a Claude-and-Codex joint review of a pull request, to re-review a PR after the ca loop finished, or when the user says "dual review this PR", "review with both Claude and Codex", or invokes /ca:dual-review. Emits the same ca_claude_review.v1 verdict as the loop's final review.

UncategorizedID: ken2403/.claude-paralell-dev-plugin/dual-review
3

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ken2403/claude-parallel-dev-plugin/tree/HEAD/ca/claude/skills/dual-review

Skill Files

Browse the full folder contents for dual-review.

Download Skill

Loading file tree…

ca/claude/skills/dual-review/SKILL.md

Skill Metadata

Name
dual-review
Description
Review a pull request in ca. Runs the dual-model review — a blind Claude review and an offline Codex second opinion in parallel, then a Claude synthesis pass that adjudicates them into one ca_claude_review.v1 verdict. This is the entry point for reviewing any PR by hand, inside or outside the implement loop. Use when the user says review this PR, dual review this PR, review with both Claude and Codex, re-review after the ca loop, or invokes /ca:dual-review. Pass --claude-only for a single-model review.

dual-review

The way to review a PR in ca. Run the loop's dual-model final review as a standalone command: an offline Codex second opinion and a blind Claude review run in parallel, then a separate Claude synthesis call adjudicates the Codex findings into one ca_claude_review.v1 verdict. Works on any PR — mid-loop, after the loop, or on PRs the loop never touched. You are the ORCHESTRATOR: the judgment happens in the subprocesses; you set up inputs, run the script, and report the result.

Preconditions (check first, fail loudly)

  • claude on PATH (or CLAUDE_BIN) and the ca plugin resolvable by claude -p (installed, or export CA_CLAUDE_PLUGIN_DIR=<repo>/ca/claude).
  • Network + authenticated gh (gh auth status).
  • codex (or CODEX_BIN) is OPTIONAL: absent/failing degrades visibly to a Claude-only review with the reason in the meta sidecar — never a hard failure.
  • The launcher materializes its exact bundled explicit-only $ca-second-opinion into an isolated temporary root. A standalone review therefore needs the Codex binary, not an ambient ca Codex plugin install. An isolated CODEX_HOME excludes global instructions/skills/plugins/config; apps, hooks, plugins, remote plugins, plugin sharing, Web search, browser, computer, image-generation, and multi-agent are disabled; shell commands inherit only the core environment and filter secret-shaped variables; approvals are never; reviewed paths are JSON encoded and their files are untrusted data.
  • The reviewed worktree must be a clean checkout whose HEAD equals the PR's headRefOid. The launcher rechecks that the PR head did not move while fetching the diff; mismatches degrade visibly as input_fetch_failed instead of reviewing stale callers/tests. Codex reads an immutable archive of that commit, and both legs must return the same pr and head_sha before clean-skip or synthesis.

Step 1 — Resolve inputs

PR="<first argument, or auto-detect>"
[ -n "$PR" ] || PR="$(gh pr view --json number --jq .number 2>/dev/null)"
[ -n "$PR" ] || { echo "no PR number given and none found for the current branch" >&2; exit 1; }
ROOT="$(git rev-parse --show-toplevel)"

Plan/intent file — the reviewers judge the PR against it:

  • If the user gave a plan path (e.g. docs/ca/plans/<id>.md), use it. For a re-review after a loop, the loop's staged copy also works: .ca/runs/<plan-id>/plan.md.

  • If no plan is given, derive an intent file from the PR itself — the review then judges the PR against its stated intent:

    INTENT="$ROOT/.ca/reviews/pr-$PR/intent.md"
mkdir -p "$(dirname "$INTENT")"
gh pr view "$PR" --json title,body --jq '"# PR intent: \(.title)\n\n\(.body)"' > "$INTENT"

Output dir and round — keep every artifact (gitignored under .ca/):

OUTDIR="$ROOT/.ca/reviews/pr-$PR"
ROUND=1; while [ -f "$OUTDIR/review-round-$ROUND.json" ]; do ROUND=$((ROUND+1)); done

Step 2 — Run the orchestrator

CLAUDE_SKILL_CA_DIR="${CLAUDE_SKILL_DIR}"
bash "$CLAUDE_SKILL_CA_DIR/scripts/dual-review.sh" \
  --pr "$PR" --plan "$INTENT_OR_PLAN" --worktree "$ROOT" \
  --round "$ROUND" --out-dir "$OUTDIR"

Pass --claude-only when the user asks for a single-model review (or codex is deliberately out of the picture). It runs the same blind Claude leg and emits the same verdict contract; no Codex process is started and the meta sidecar records dual_review:false. Do NOT reach for /ca:review-pr for this — that skill is this command's internal blind leg, not a second entry point.

It runs both legs in parallel and keeps the current-round Codex output outside the worktree until the blind review finishes, synthesizes when Codex found anything, skips synthesis on a clean full-coverage Codex pass, and degrades to Claude-only with a machine-readable reason when the Codex leg fails. A timeout is reported distinctly as codex_timeout; the entire child process group is terminated and a bounded head+tail diagnostic is retained in the leg log. If blind Claude fails first, the Codex launcher process group is cancelled immediately. If synthesis fails, the script removes any partial final JSON, writes synthesis.status: failed, and exits non-zero. If it exits non-zero, report the failure verbatim — do not improvise a verdict.

Step 3 — Report

Read review-round-N.json and the meta sidecar; report to the human:

  • Verdict (approve / request_changes / blocked) and the summary paragraph.
  • Findings as a table: id, blocking, severity, title, file:line.
  • When synthesis ran: the second_opinion.ledger adjudications (confirmed/refuted/not_applicable/unresolved) and any resolved_blind_findings (blind blockers that synthesis downgraded, with evidence).
  • The leg statuses from the meta sidecar (e.g. Codex used coverage full, unavailable: codex_timeout, unavailable: invalid_configuration, or unavailable: review_input_oversized, or unavailable: unsupported_codex_cli — say plainly when the second opinion did NOT happen).

If --comment was passed, also post the summary to the PR:

gh pr review "$PR" --comment --body "<summary>"

Notes

  • This is the same orchestrator the implement loop runs for its final review (dual by default) — same parallel isolation, scripts, contract, and fail-closed semantics.
  • Verdicts from this command do NOT promote a draft PR; only the loop's final-mode approve does (the loop owns gh pr ready).
  • The code-review standards skill (canonical risky-surface list, four dimensions) auto-activates in the claude -p review sessions — the criteria are the same as sa/ha; only the execution differs.

References

  • references/review-contract.md — ca_claude_review.v1 / ca_codex_review.v1 shapes and gate semantics (byte-identical copy; the canonical set lives with each skill that consumes it).
dual-review Skill | Agent Skills