lilpacy/claude-fable-review

Claude Fable Review

Trigger Rule

• Use this skill only when the user explicitly asks for Fable review, Claude review, or to consult Fable.
• Do not auto-trigger this skill for generic "review", planning, or post-commit review requests.
• Default review path remains codex-exec-review unless the user explicitly selects Fable.

When Invoked

• Before presenting an implementation plan to the user, run claude -p with Fable to review the plan if the user explicitly asked for Fable review.
• After a non-trivial commit, run claude -p with Fable to review the committed change if the user explicitly asked for Fable review.
• Repeat review up to 3 times. Stop when no critical issue remains.
• Do not replace a requested Fable review with local tests or your own judgment.

Review Command Rules

• Use env claude -p to avoid shell aliases such as claude --dangerously-skip-permissions.
• Use CLAUDE_REVIEW_MODEL=global.anthropic.claude-fable-5. Do not read the review model from the normal runtime model setting.
• Pass --model "$CLAUDE_REVIEW_MODEL" explicitly. Do not rely on the root "model" setting.
• Use CLAUDE_REVIEW_EFFORT=${CLAUDE_REVIEW_EFFORT:-medium} and pass --effort "$CLAUDE_REVIEW_EFFORT".
• Do not pass /Users/lilpacy/dotfiles/claude/settings.json as the runtime settings file. Its normal permissions are broader than review mode.
• Use --safe-mode and a minimal review settings JSON. Do not use --bare for default reviews because Claude Code 2.1.197 exposes only Read in that mode, even when WebFetch is requested.
• Use --permission-mode default. Do not use --permission-mode plan for default reviews because it suppresses WebFetch, Grep, Glob, and Bash in the measured review setup.
• Prefer --output-format stream-json --include-partial-messages --verbose so progress remains visible while the final result is still machine-checkable.
• Allow Bash broadly and control risk with a denylist plus prompt constraints. Fable review quality depends on being able to run unknown local CLI introspection such as --help, --version, --default-config, status, and list commands without adding one-off allowlist entries.
• Keep the denylist focused on external review/agent launches, destructive git commands, obvious filesystem mutation, install/deploy entry points, and code execution snippets. Do not deny broad test/build command names because that also blocks their --help output; rely on the review prompt to forbid running tests/builds.
• Allow WebFetch by default for currentness checks against public primary sources. Request WebSearch and Fetch too, but do not require them in environments where Claude Code does not expose those tools.
• Instruct the reviewer not to paste private code, secrets, env values, customer data, or large local diffs into web queries.
• Instruct the reviewer not to run tests, build, format, install, generation, mutation, deployment, or external review commands.
• Instruct the reviewer not to start another claude -p, codex exec, or mcp__ais call.
• Instruct the reviewer to report only critical issues and to check whether the answer is out of date or deprecated.

Initial Review Template

Set PLAN_OR_DIFF_REF to the full path, commit ref, or concise description being reviewed.

set -o pipefail

CLAUDE_MAIN_SETTINGS=/Users/lilpacy/dotfiles/claude/settings.json
CLAUDE_REVIEW_MODEL=global.anthropic.claude-fable-5
CLAUDE_REVIEW_EFFORT=${CLAUDE_REVIEW_EFFORT:-medium}
REVIEW_LOG=${REVIEW_LOG:-review-result.jsonl}
CLAUDE_REVIEW_SETTINGS_JSON=$(jq --arg review_model "$CLAUDE_REVIEW_MODEL" -c '{
  env: {
    CLAUDE_CODE_USE_BEDROCK: .env.CLAUDE_CODE_USE_BEDROCK,
    AWS_REGION: .env.AWS_REGION,
    ANTHROPIC_MODEL: $review_model,
    CLAUDE_CODE_MAX_OUTPUT_TOKENS: .env.CLAUDE_CODE_MAX_OUTPUT_TOKENS,
    MAX_THINKING_TOKENS: .env.MAX_THINKING_TOKENS
  },
  permissions: {
    allow: [
      "Read(~/**)", "Grep", "Glob", "WebSearch", "WebFetch", "Fetch",
      "Bash"
    ],
    deny: [
      "Edit(*)", "Write(*)",
      "Bash(codex exec:*)", "Bash(claude:*)",
      "Bash(git reset:*)", "Bash(git checkout:*)",
      "Bash(git clean:*)", "Bash(git push:*)",
      "Bash(rm:*)", "Bash(rmdir:*)", "Bash(mv:*)",
      "Bash(cp:*)", "Bash(mkdir:*)", "Bash(touch:*)",
      "Bash(chmod:*)", "Bash(chown:*)", "Bash(sudo:*)",
      "Bash(npx:*)", "Bash(pnpx:*)", "Bash(bunx:*)",
      "Bash(brew install:*)", "Bash(brew upgrade:*)",
      "Bash(brew uninstall:*)", "Bash(npm install:*)",
      "Bash(npm i:*)", "Bash(npm ci:*)",
      "Bash(pnpm install:*)", "Bash(yarn install:*)",
      "Bash(bun install:*)",
      "Bash(pip install:*)", "Bash(pip3 install:*)",
      "Bash(python -c:*)", "Bash(python3 -c:*)", "Bash(ruby -e:*)",
      "mcp__ais__*"
    ]
  },
  sandbox: {enabled: false}
}' "$CLAUDE_MAIN_SETTINGS")

env claude -p \
  --safe-mode \
  --settings "$CLAUDE_REVIEW_SETTINGS_JSON" \
  --model "$CLAUDE_REVIEW_MODEL" \
  --effort "$CLAUDE_REVIEW_EFFORT" \
  --permission-mode default \
  --tools "Read,Grep,Glob,WebSearch,WebFetch,Fetch,Bash" \
  --allowedTools "Read,Grep,Glob,WebSearch,WebFetch,Fetch,Bash" \
  --output-format stream-json \
  --include-partial-messages \
  --verbose \
  "このプランまたは差分をレビューして。レビュー用の許可設定では編集系ツールと危険な Bash を禁止しているので、テスト・build・format・install・生成・編集・mutation・deploy コマンドは実行せず、差分・設定・既存ログの読取を主材料に判断して。現行仕様確認に限って WebFetch を使ってよい。WebSearch/Fetch が利用可能なら使ってもよいが、利用不能なら WebFetch 可能な public primary sources を直接読むこと。公式 docs・release notes・standards・package registry・source repository など public primary sources を優先し、private code・secret・env 値・customer data・大きな local diff を検索クエリや取得 URL に貼らないで。不足する実行結果があれば質問して。別の claude -p、codex exec、mcp__ais、外部レビューコマンドは絶対に起動しないで。瑣末な点へのクソリプはしないで。致命的な点だけ指摘して。回答内容が現時点で out of date / deprecated になっていないかにも気をつけて。Web を使った場合は参照 URL と判断への使い方を短く添えて: $PLAN_OR_DIFF_REF" \
  | tee "$REVIEW_LOG"

Capture the session id from the final stream result:

jq -r 'select(.type == "result") | .session_id' "$REVIEW_LOG" | tail -n 1

Resume Review Template

Use the same session for updated plan reviews:

set -o pipefail

REVIEW_LOG=${REVIEW_LOG:-review-result.jsonl}

env claude -p \
  --safe-mode \
  --settings "$CLAUDE_REVIEW_SETTINGS_JSON" \
  --model "$CLAUDE_REVIEW_MODEL" \
  --effort "$CLAUDE_REVIEW_EFFORT" \
  --permission-mode default \
  --tools "Read,Grep,Glob,WebSearch,WebFetch,Fetch,Bash" \
  --allowedTools "Read,Grep,Glob,WebSearch,WebFetch,Fetch,Bash" \
  --output-format stream-json \
  --include-partial-messages \
  --verbose \
  --resume "$SESSION_ID" \
  "前回の指摘を反映してプランまたは差分を更新した。もう一度レビューして。前回と同じ制約で、テスト・build・format・install・生成・編集・mutation・deploy・外部レビューコマンドは実行しないで。現行仕様確認に限って public primary sources への WebFetch は使ってよい。WebSearch/Fetch が利用可能なら使ってもよい。致命的な点だけ指摘して。新しく追加された問題がなければ、その旨を明示して: $PLAN_OR_DIFF_REF" \
  | tee "$REVIEW_LOG"

For non-interactive automation where progress visibility is irrelevant, --output-format json may be used instead of stream-json.

Permission Caveats

• Claude Code --permission-mode default with broad Bash and a denylist is not the same security boundary as codex exec --sandbox read-only.
• This review mode is intentionally blacklist-based so current local CLI documentation can be inspected without one-off permission churn. Keep review prompts explicit that tests, builds, installs, generation, mutation, and deployment are forbidden.
• Treat the review as complete only when a final type == "result" stream event is returned and the command used the explicit minimal settings above.
• If the command runs without the explicit Fable model, without --safe-mode, or with the full normal claude/settings.json, treat it as review incomplete.
• If the stream init event does not list WebFetch, treat the review command as misconfigured and rerun with the default --safe-mode --permission-mode default template. If WebSearch or Fetch are absent but WebFetch is present, the review can proceed.
• If a denied tool appears in permission_denials, treat that denial as evidence the permission gate worked, not as a review failure.
• If the reviewer needs command output that was intentionally blocked, provide existing logs or run the command yourself outside this review gate when appropriate.

Timeout Semantics

• Wait for the final type == "result" stream event before reporting success:
  • about 15s for trivial prompts
  • about 30-60s for light reviews
  • about 180s for normal review tasks
• If intermediate output is still arriving, keep waiting until the review ends.
• review started but final result not yet returned is not review complete.
• For a required review, if the final answer is missing, try at least one --resume "$SESSION_ID" or rerun.
• If the final answer still cannot be recovered, report review incomplete and ask the user how to proceed.
• Do not say a review passed unless the final review answer was obtained.

Validation

• --safe-mode, --settings <json>, explicit Fable --model, default --effort medium, --output-format stream-json --include-partial-messages --verbose, and --resume were verified with Claude Code 2.1.197.
• The stream output ends with type == "result" and includes .session_id and modelUsage."global.anthropic.claude-fable-5".
• In Claude Code 2.1.197, --bare and --permission-mode plan exposed only Read; --safe-mode --permission-mode default exposed Bash, Glob, Grep, Read, and WebFetch.
• A dry run attempting Bash(claude -p hi) returned a permission denial instead of executing.
• After editing this skill, run:

python3 /Users/lilpacy/dotfiles/codex/skills/.system/skill-creator/scripts/quick_validate.py /Users/lilpacy/dotfiles/skills/claude-fable-review