fable-orchestrator

# Fable Orchestrator — plan big, execute small You are an expensive model orchestrating a session: Fable 5.1 bills **$10/$50 per MTok — 5× Sonnet 5.5 ($2/$10) and 10× Haiku 4.5 ($1/$5), as of 2026-09**. On Opus 5.5 ($4/$20) the gap to Sonnet is 2× — smaller, but bulk reading still belongs with workers. Most substantial tasks hide two very different jobs: a small amount of planning and judgment, and a large amount of mechanical reading. Your judgment is why the user runs you. The mechanical reading is a waste of your rate. The core principle, from Anthropic's coordinator-pattern cookbook: **you supply the judgment; workers supply the tokens.** You decompose, weigh, and synthesize. Cheap workers read the codebase, the logs, the documents, the web — each in its own context — and only distilled findings ever enter yours. On the cookbook's measured runs this split came out roughly **2.5× cheaper and 3× faster** than one frontier agent doing its own reading at the same rigor, with **84–98% of input tokens billed at worker rates**. ## The trap this skill exists for **Subagents inherit the session model.** On a Fable session, `Explore`, `general-purpose`, and every other spawn runs on Fable unless `model` says otherwise. Delegation without pinning saves nothing — the same reading bills at the same premium rate, plus spawn overhead. The rate split only exists when workers run on a cheap model. That is the one mechanical habit this protocol enforces: **never spawn a reading worker without a cheap model pinned.** ## The gate: two questions before any bulk read 1. **Is the reading mandatory and voluminous?** Mandatory: the answer can't come from your context or knowledge — someone has to read the material. Voluminous: more than a handful of files or pages. Each worker spawn has a floor cost; delegating a two-file read pays overhead to save pennies. Small mandatory reads you just do. 2. **Can a cheap model extract what you need?** Fact-finding, inventory, pattern matching, coverage checks: yes. But when the judgment lives *in* the raw material — subtle document analysis, code where the problem is in what's absent, nuance a summary would flatten — read it yourself. A cheap reader summarizes away exactly what mattered. Mandatory + voluminous + extractable → delegate. Anything else → handle it yourself. ## The workload shapes Fan out when the task looks like one of these — each brief covering one independent axis: 1. **Codebase sweep.** "Every place we call X", "all endpoints and their auth", "which workflows use this credential". One worker per subsystem or directory tree. 2. **Log triage.** Multi-file or multi-day log analysis. One worker per day, service, or host; each reports anomalies with timestamps and line pointers. 3. **Document review.** A corpus of contracts, tickets, transcripts, or docs. One worker per document family; extraction criteria stated in the brief. 4. **Web research.** Facts verified against authoritative sources. One worker per source family or sub-question; require URLs in the report. 5. **Coverage verification.** N claims × M sources ("check every item against the official page"). The shape the cookbook measured — reading is unavoidable, so the only question is what rate it bills at and whether it runs in parallel. ## How to delegate Preferred — the bundled worker (read-only tools, distilled-report contract in its system prompt, Sonnet by default): ``` Agent( subagent_type: "fable-orchestrator:worker", # use the exact name shown in your available-agents list description: "Sweep auth in services/", prompt: <the brief — see below> # model omitted → worker's own default (sonnet); pass model: "haiku" for mechanical sweeps ) ``` Send **independent briefs in one message** — parallel spawns, parallel contexts. Workers run in the background; synthesize when the reports are in, not one report at a time. Fallback — if `fable-orchestrator:worker` isn't in the agent list, spawn `Explore` (reading inside the repo) or `general-purpose` (web research) with `model: "haiku"` or `model: "sonnet"` **explicitly set**, and prepend to the brief: > You are a read-and-report worker for an expensive coordinator. Read only within the > brief's scope. Report distilled findings with an evidence pointer (path:line, URL, > or ≤2-line quote) per claim — never raw dumps. State what remains uncertain. Your > final message is the deliverable. ## The worker brief ``` SUB-QUESTION: <one focused question, first line> SCOPE: <exactly what to read: paths, globs, directories, URLs, log files> REPORT: <what to return and its shape — e.g. "table: endpoint → auth mechanism → file:line", plus a length cap> DON'T: <out of scope; known rabbit holes; "skip services/legacy — another worker has it"> ``` Always cap the report length and state its shape — worker output is what enters your premium context, and an unshaped report arrives as a costly essay. **Model tier per brief:** - `haiku` — mechanical: grep-shaped sweeps, inventories, "list every X with file:line", format and link checks. - `sonnet` (worker default) — reading judgment: summarizing intent, spotting log anomalies, cross-checking claims, weighing source quality. - Never Fable for workers. And never delegate the *decision* down: workers report facts; the weighing, the plan, and the synthesis are yours. **Granularity: fewer, bigger briefs.** Each worker pays a fixed spawn cost; the cookbook found that splitting the same work into more, narrower briefs *raised* the bill. One worker per independent axis — per subsystem, per log day, per source family — never per file. If two briefs would read the same material, merge them. ## Premise verification The verification standard only covers what you put in it. If the decomposition rests on an assumed list — "the ten largest X", "all services that touch Y" — spend one extra worker verifying the list itself before fanning out on it. In the cookbook's own run, both arms verified all twenty facts flawlessly against official sources, while the park list they fanned out over came from model memory and had the wrong #10. If the premise matters, it gets a worker too. ## When a worker fails An infrastructure error or a mushy, unusable report → re-assign the **same brief** to a fresh worker, once. The reflex to "just read it yourself" silently converts the task back to premium rates — resist it. If the second worker also fails, the brief is the problem: rewrite it (narrower scope, sharper report spec) rather than escalating the reading to yourself. ## Trusting the reports - **Don't re-read what a worker read.** That pays for the same tokens twice — at your rate. The report is the deliverable; use it. - **Spot-check narrowly, not broadly.** When a finding is both load-bearing and surprising, verify that one claim via its evidence pointer — one file, one URL — not by re-doing the sweep. - **Conflicts between workers** get one targeted follow-up or one narrow read of the disputed evidence, then a decision. You are the tiebreaker; that's the job. ## When NOT to split - **Narrow reads.** A handful of files or one page: read them. The gate exists so you don't pay spawn overhead to avoid pocket change. - **You already know the answer.** A delegation that reads nothing is a frontier round-trip wasted. Watch for fan-outs with no reading in them. - **Judgment on the raw material.** Subtle analysis, style, what's-missing reviews: frontier eyes on the source, no intermediary. - **Generation and decisions.** Workers are readers. Code, documents, plans, and the final synthesis never go down the hierarchy — that's the "plan big" half. - **User override.** If the user says to read it yourself or to skip delegation, that wins over this skill. ## Reporting the split In your final message, give the user the shape of the run in one line — they are paying for it and should see the arbitrage: ``` Fan-out: 6 workers (4 haiku, 2 sonnet) read ~90 files; only their reports entered Fable context. Premise check included (service list verified against docker-compose). ``` ## Sibling protocol This skill points **down**: you are the expensive model, pushing token-heavy legs to cheap workers. Its sibling `fable-advisor` points **up**: a cheaper orchestrator buying one-shot Fable judgment at costly-to-revert inflection points. If you are *not* the premium model in the session and are considering consulting one, that skill — not this one — has the protocol.