Skill

fable-check — Fable 5 code review & advisory

Run a structured, read-only code review — or ask an advisory question — using the Claude Fable 5 model via the local claude CLI. The reviewer/advisor gets read-only tools (Read/Grep/Glob plus read-only git commands), so it verifies its claims against the real code instead of guessing.

All commands below run from the repository being reviewed. <skill-dir> is this skill's directory (the folder containing this SKILL.md).

Core constraint

• This skill is read-only. Never fix issues, apply patches, or imply you are about to make changes as part of running it.
• Return the script's output to the user verbatim — do not paraphrase, summarize, or filter the findings or the advisory answer.
• After presenting the output, you may offer to act on it as a separate follow-up if the user wants.

Commands

Standard review of current work (auto-detects: dirty working tree → working-tree diff; clean tree → branch vs default branch):

node "<skill-dir>/scripts/fable-check.mjs" review

Review a branch against a base:

node "<skill-dir>/scripts/fable-check.mjs" review --base main

Adversarial review (skeptical, tries to block the change; accepts steering text):

node "<skill-dir>/scripts/fable-check.mjs" review --adversarial challenge the caching and retry design

Deep review (three parallel lens passes — correctness, security/data-safety, design/failure-modes — merged into one report; the most extensive and most expensive mode):

node "<skill-dir>/scripts/fable-check.mjs" review --deep

Advisory — ask Fable a question instead of requesting a review (architecture choices, design tradeoffs, "is this plan sound?", second opinions on a decision, "how should X work?"). The advisor explores the repo with read-only tools and answers in prose with one clear recommendation:

node "<skill-dir>/scripts/fable-check.mjs" ask "should the job runner move to worker threads, or is the detached-process design right?"

Other flags (both review and ask): --effort low|medium|high|xhigh|max (default xhigh), --model <model> (default claude-fable-5), --background, --json, --quiet (suppress progress stream). review also takes --scope auto|working-tree|branch and --base <ref>.

Progress reporting — IMPORTANT for invoking agents

Runs take minutes, but they are never silent. While running, the script streams continuous progress to stderr (stdout stays clean for the final report):

• a startup banner with the target and an expected duration range,
• a line for every tool call the reviewer makes (tool #14: reading src/foo.js),
• phase changes (phase: review pass running, phase: lens passes: 2/3 complete, phase: merging ...),
• a heartbeat at least every ~20 seconds during thinking stretches (still working — 3m40s elapsed, 14 tool call(s) so far, last: ...).

Interpretation rules:

• Do not kill or abandon a run that is emitting progress lines or heartbeats — it is healthy. Typical durations: standard review 2–8 min, deep review 5–15 min, advisory 1–6 min (longer for big diffs or max effort).
• Silence for more than ~60 seconds is abnormal; only then check on it. The script itself flags a job as possibly stalled in status after 5 minutes without any model events (its own heartbeats don't count as activity).
• If your execution harness has a command timeout shorter than ~15 minutes, run with --background instead of stretching the timeout.

Foreground vs background

• Small change (1–3 files) or a quick advisory question → run in the foreground and wait; relay the streamed progress if your harness shows it.
• Anything larger, unclear size, or --deep → run in the background.

Background options (either works):

1. Use your own background-execution facility to run the foreground command, then report results when it completes.
2. Or use the built-in job runner:

node "<skill-dir>/scripts/fable-check.mjs" review --background     # or: ask --background "..."
node "<skill-dir>/scripts/fable-check.mjs" status                  # live progress: phase, elapsed, tool calls, last activity
node "<skill-dir>/scripts/fable-check.mjs" result                  # final report (latest job)
node "<skill-dir>/scripts/fable-check.mjs" cancel                  # stop an active job

status on a running job shows elapsed time, current phase, tool-call count, the last activity with its age, and an explicit healthy/possibly-stalled verdict — poll it every 30–60 seconds. You can also tail -f the job's log file (path shown at launch and in status).

When you launch a background job, tell the user the job id and that they can ask for status/results at any time.

Setup / troubleshooting

If a run fails because the claude CLI is missing or logged out:

node "<skill-dir>/scripts/fable-check.mjs" setup

It reports what is missing and the exact next step (installing Claude Code or logging in). Runs use the user's existing Claude subscription — no API key is needed.

Notes

• For reviews there must be something to review: uncommitted changes, or commits ahead of the base branch. If the script reports nothing to review, relay that — do not invent a review. (ask has no such requirement.)
• Reports and answers are saved under ~/.fable-check/jobs/<repo>/ and each includes a claude -r <session-id> command to reopen the session interactively.
• Focus text is most effective with --adversarial; the standard review intentionally takes no steering so its judgment stays neutral. For steered questions, prefer ask.