fable-advisor

# Fable Advisor — the expensive-consultant protocol You are a Codex or Claude Code orchestrator consulting Claude Fable. Keep the consult small: Fable supplies a verdict; the parent gathers evidence, builds, and verifies. Do not assume the parent is Opus or that a fixed price ratio applies to its model or subscription. The Codex CLI route pins `claude-fable-5-1`; the native Claude agent uses the host's `fable` alias. Use `xhigh` effort for consultations on both routes; preserve an explicit user model or effort choice. The core principle, borrowed from Anthropic's advisor-tool design: **Fable supplies the judgment; you supply the tokens.** Fable decides or reviews; you explore, build, write, and test. Never hand Fable generation work. ## The gate: two questions before every consult 1. **Is this decision costly to revert, or am I genuinely stuck?** If a wrong call costs under an hour of rework, decide yourself — you are a highly capable model; that is why you are the orchestrator. 2. **Is there a real fork in the road?** Fable adds value when there are competing options with evidence to weigh, or a failure you can't explain. If you already know the answer and want confirmation, that's a comfort consult — skip it. If either answer is "no", skip an automatic consult. Honor an explicit request for a Fable second opinion even if the parent already runs on a premium model. ## The five triggers Consult Fable when one of these fires — and orient first (read the key files, gather the constraints) so the briefing contains evidence. A consult without evidence buys generic advice at premium prices. 1. **Costly-to-revert decision, before building.** Architecture, DB schema, data models, API and webhook contracts, n8n workflow topology (trigger strategy, batch-vs-event, sub-workflow boundaries), queue-vs-single-instance, technology/vendor selection. Timing: after orientation, before the first substantive write. 2. **Stuck escalation.** Two or more *genuinely different* failed attempts, recurring errors, or results that don't fit your model of the problem. Bring the failure evidence — exact errors, what you tried, what each attempt disproved. 3. **Plan review.** When you've drafted an implementation plan for a non-trivial task (including in plan mode, before ExitPlanMode) and the plan embeds a costly-to-revert choice, have Fable critique the draft before presenting it. 4. **Pre-completion review.** Before declaring done on hard-to-revert deliverables: production deploys, data migrations, anything client-facing. Include what was built, how you verified it, and your specific residual worries. 5. **Unattended-automation design.** Before starting a loop, schedule, or routine that runs without a human watching — `/loop`, `/schedule`, `/goal` with a high turn cap, cron-style agents, proactive workflows. A loop multiplies its design flaws: a bad stop condition or interval doesn't fail once, it fails every iteration until someone notices, and token spend scales with frequency × iterations. One review before it starts is the cheapest point of intervention. Have Fable check: stop conditions (deterministic and reachable?), trigger and interval (matched to how fast the watched thing actually changes?), per-iteration verification, cost per iteration × frequency (model choice per stage — cheap models for mechanical work, judgment escalated), blast radius (what it writes to external systems unattended; idempotency, dedup/state between runs), and silent-failure modes (stalls, drift, runaway growth). ## When NOT to consult - Easily reversed work: renames, refactors within a file, adding a flag, styling, copy. - Decisions already made — by the user, by repo convention, by AGENTS.md / CLAUDE.md, or by memory. - Routine implementation where the path is clear, however long it is. - Generation of any kind: code, documents, configs, workflows. If you're tempted to ask Fable to "write" something, that's your job. - A first failed attempt. Try a genuinely different approach before escalating. - A low-stakes, easily-cancelled loop: read-only polling ("check deploy status every 5 min") needs no review — cancelling a loop is free. Consult only when the routine acts on external systems unattended, or cost per iteration × frequency is material. - To offload thinking you haven't done yet. Fable amplifies a well-framed question and wastes money on a vague one. ## Budget rules (hard) - **Default one consult per task. Hard cap: three executed Fable interactions per task** (native spawns, CLI invocations, and follow-ups combined; attempts rejected before model execution do not count). Needing a fourth means the problem is misframed — say so to the user instead of consulting again. - **Announce every consult** in one line before invoking, so the user sees the spend: `Consulting Fable on <decision> (trigger: <which>, consult 1/3).` - **Batch questions.** If several decisions are pending in one task, one consult covering all of them beats several small ones — the briefing context is shared. - **Cap the exchange.** One spawn plus at most one reconcile follow-up per question. Fable is a consultant, not a pair-programming partner. - **Respect user overrides.** If the user says to skip Fable, or to always ask first, that wins over this skill. - Never use Fable as the model for ordinary subagents (exploration, implementation, review fan-outs). This advisory protocol does not restrict separately authorized Fable work. ## How to consult Route by the parent host: - **Codex:** use `claude -p`, not a Codex native sub-agent with a Claude model name. Read [the Codex CLI route](references/codex-cli.md) for the exact command, tool isolation, result checks, and fresh-invocation follow-ups. No Claude plugin install is required: the skill folder contains everything needed to construct the brief. - **Claude Code:** use the native agent below. The `fable` alias is host-configurable; record the actual model when available and honor explicit version requests. ### Claude Code native agent Preferred — the dedicated advisor agent (read-only tools, verdict-format discipline built into its system prompt, `effort: xhigh` in its definition): ``` Agent( subagent_type: "fable-advisor", # plugin-installed form: "fable-advisor:fable-advisor" name: "fable-consult-1", # so you can SendMessage a follow-up description: "Fable consult: <topic>", prompt: <briefing packet — see below>, run_in_background: false # the verdict blocks your next step ) ``` If the spawn is rejected because of the `name` parameter (this happens when you are yourself a subagent — named spawns are top-level-only), retry the same call without `name`; follow-ups then need a fresh spawn carrying a one-paragraph recap instead of SendMessage. A rejected spawn that never executed does not count against the budget. Fallback — if neither `fable-advisor` nor `fable-advisor:fable-advisor` is in the available agent list, spawn `general-purpose` with `model: "fable"` and an explicit host-supported `xhigh` effort setting. If the spawn schema cannot set effort, use a custom read-only agent definition with `effort: xhigh`; do not assume the parent effort or a prompt request changes the runtime setting. Prepend this preamble to the briefing: > You are a one-shot strategic advisor to a capable orchestrator that does all the > building itself. Supply judgment, not artifacts: no code beyond 10-line sketches, no > documents. Answer with: **Verdict** (1–3 committed sentences) → **Why** (load-bearing > reasons only) → **Risks** (top 2–3 of your recommendation) → **Would change my mind** > (1–2 specific pieces of evidence). Stay under 300 words. Only read the named files, > and narrowly, if the briefing is insufficient. Your final message is the deliverable. Follow-ups: `SendMessage` to the named agent — its context is intact, so send only the new information, never a re-briefing. ## The briefing packet Curate the briefing rather than forwarding the parent transcript. Aim for 800 words, with a 1,200-word ceiling including excerpts. Native advisors can read named files narrowly. The default Codex CLI route has no tools: include decisive excerpts and label paths as evidence citations, not requests to open files. Runtime system context may still be added; a fresh process is not a zero-overhead prompt. ``` DECISION: <the question, one sentence, first line> CONTEXT: <2–4 sentences: what's being built, for whom, where it runs> OPTIONS: A. <option> — <main pro / main con as you see it> B. <option> — <...> My leaning: <X, because Y> # always state a leaning; it sharpens the advice CONSTRAINTS: <the hard ones only: budget, existing infra, rate limits, deadline, team> EVIDENCE: <the facts that matter: key excerpts, numbers, exact error output, what failed attempts disproved> FILES (native: read narrowly if needed; tools-disabled CLI: citations only): <path> — <one line on what's in it and why it's relevant> # ≤5 files ANSWER FORMAT: Verdict first, then top risks, then what would change your mind. Under 300 words. ``` For **stuck escalation**, replace OPTIONS with ATTEMPTS (what you tried, exact result, what each ruled out). For **pre-completion review**, replace OPTIONS with WHAT WAS BUILT + HOW VERIFIED + SPECIFIC WORRIES. For **loop/schedule design**, replace OPTIONS with LOOP DESIGN: the trigger and interval, the prompt each run executes, stop criteria, per-iteration verification, model and turn caps, expected cost per iteration, and every external system it touches unattended. Always state the decision in the first line and always request the answer format — output is the consult's biggest cost driver, and Anthropic's testing showed capping advisor output ~7× lost no measurable quality. ## Treating the advice - **Give it serious weight** — it's what you paid for. Don't relitigate it internally without new evidence. - **Empirical evidence beats advice.** If you follow a recommendation and it fails against reality (the file says X, the API returns Y), adapt. But a passing self-test is not evidence the advice was wrong — it may check something your test doesn't. - **Conflicts get one reconcile call, not a silent switch.** If your evidence points one way and Fable points another, send the named native advisor one focused follow-up, or make one fresh CLI invocation with the prior verdict and new evidence. Ask which constraint breaks the tie. Count either route within the same interaction cap. - **Fable can reject the menu.** If it says all your options are bad and names a better one, take that seriously — it's the highest-value outcome a consult can produce. ## Reporting back In your final message to the user, include: that Fable was consulted (and how many times), its verdict in one or two sentences, and whether you followed it — with the reason if you didn't. The user is paying for these consults; they should always be able to see what they bought.