/cursor-review

name: cursor-review
description: "Run a Cursor Composer review as an advisory second opinion while keeping the primary review verdict on the host brain. Use when user invokes cursor:review, asks Cursor to review, or wants composer to sanity-check a diff. Cursor never owns APPROVE/REQUEST_CHANGES."
description-en: "Run a Cursor Composer review as an advisory second opinion while keeping the primary review verdict on the host brain. Use when user invokes cursor:review, asks Cursor to review, or wants composer to sanity-check a diff. Cursor never owns APPROVE/REQUEST_CHANGES."
description-ja: "Cursor Composer に advisory second opinion としてレビューさせるスキル。cursor:review、Cursor にレビューして、composer で差分を sanity check して、という時に使う。primary verdict は常に host brain が持ち、Cursor は APPROVE/REQUEST_CHANGES を所有しない。"
allowed-tools: ["Read", "Bash", "Grep"]
argument-hint: "[--base <ref>] [review-scope]"
user-invocable: true

cursor:review - Advisory Cursor Review

Cursor Composer (`composer-2.5-fast`) を read-only second opinion として使う review skill。primary verdict は host brain (Opus / Claude role) が出す。

Quick Reference

cursor:review --base origin/main
cursor:review "Phase 88.5 command namespace diff"

Rules

• Cursor review is advisory. The final verdict must come from the host reviewer.
• Do not pass `--write` to `cursor-companion.sh`.
• Use resolver/model routing, not direct env checks:

  bash scripts/resolve-impl-backend.sh --backend cursor --role reviewer
  bash scripts/model-routing.sh --host cursor --role worker --field model

• Prefer the existing `harness-review` contract when a full review is requested. This command exists for users who explicitly ask for the Cursor lane.

Flow

1. Resolve the Harness helper root:

   HARNESS_PLUGIN_ROOT="${HARNESS_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-}}"
   if [ -z "$HARNESS_PLUGIN_ROOT" ] && [ -n "${CLAUDE_SKILL_DIR:-}" ]; then
     probe="$(cd "${CLAUDE_SKILL_DIR}" && pwd)"
     while [ "$probe" != "/" ] && [ ! -d "$probe/scripts" ]; do
       probe="$(cd "$probe/.." && pwd)"
     done
     [ -d "$probe/scripts" ] && HARNESS_PLUGIN_ROOT="$probe"
   fi
   if [ -z "$HARNESS_PLUGIN_ROOT" ]; then
     echo "ERROR: HARNESS_PLUGIN_ROOT is not set and could not be derived from CLAUDE_PLUGIN_ROOT or CLAUDE_SKILL_DIR" >&2
     exit 2
   fi

2. Determine review scope:

   BASE_REF="${TASK_BASE_REF:-origin/main}"
   REVIEW_SCOPE="${ARGUMENTS:-}"
   if [ -n "${REVIEW_SCOPE}" ]; then
     set -- ${REVIEW_SCOPE}
     while [ "$#" -gt 0 ]; do
       case "$1" in
         --base)
           BASE_REF="${2:?--base requires a ref}"
           shift 2
           ;;
         --base=*)
           BASE_REF="${1#--base=}"
           shift
           ;;
         *)
           shift
           ;;
       esac
     done
   fi
   DIFF_STAT="$(git diff --stat "${BASE_REF}..HEAD")"
   DIFF_TEXT="$(git diff "${BASE_REF}..HEAD")"

3. Ask Cursor in read-only mode:

   PROMPT="$(cat <<EOF
Review this diff as an advisory second opinion.
Base ref: ${BASE_REF}
Requested scope: ${ARGUMENTS:-full diff}
Requirement, purpose and DoD: <original request and completion criteria>
Spec and validation evidence: <relevant paths and observed results>

Focus on bugs, regressions, missing tests, and unsafe assumptions. Do not propose broad refactors.
Use the supplied requirement, DoD, spec and validation evidence to assess the requested scope; inspect missing context read-only.
For each finding, state its location, trigger, impact and checkable evidence. Separate hypotheses from observed defects; do not invent findings to fill a quota.
Return an assessment only. Do not implement changes or treat the implementation worker's self-report as verification.

Diff stat:
${DIFF_STAT}

Diff:
${DIFF_TEXT}
EOF
)"
   bash "${HARNESS_PLUGIN_ROOT}/scripts/cursor-companion.sh" task "${PROMPT}"

4. Host reads Cursor's output and performs the primary review:

• Findings first, ordered by severity.
• Mark Cursor-originated points as advisory when used.
• `APPROVE` / `REQUEST_CHANGES` must be the host decision, not a copied Cursor verdict.

Output

Use the standard review shape:

• findings
• open questions
• validation run
• final host verdict