fable-fable

# Fable Mode — Fable 5.1 (v3, agent-routed) Same structural delegation as fable-opus: the orchestrator is a real agent definition (`agents/fable-orchestrator.md`) with no Write/Edit tool, so it cannot do the work inline. The only difference from fable-opus is the model the orchestrator runs on. Fable 5.1 is the top of the ladder (Haiku → Sonnet → Opus → Fable → user); gaps it cannot close go to the user. Fable 5.1's characteristic gaps under this discipline: none measured (benchmarked 2026-09-02, 16 paired runs, n=2 per cell: 100/100 in every cell, both arms; skill cost ~1.1× tokens, ~2× wall). This is a provenance runner — worklog, reproducible checks, flagged judgment calls — not an accuracy runner. The orchestrator definition keeps the frontier-tier guards (external-artifact verification, scope rule, replan budget) because those produce the trail. If a task has one obvious correct approach and fits in a single pass, skip this loop and do it directly. On a Fable-class model the stage map is ceremony for anything short of multi-session or many-file work. ## How to run it 1. Confirm the fable agents are installed (`fable-orchestrator`, `fable-worker-sonnet`, `fable-worker-haiku`, `fable-verifier` appear in the available agent types) AND the Agent/Task tool accepts `model: "fable"`. - Agents missing → fall back to the inline method: spawn a general-purpose agent with `model: "fable"` and pass it the Core loop and operational rules verbatim from `agents/fable-orchestrator.md`. - `"fable"` not accepted as a model → say so and run fable-opus instead. - Per-invocation `model` override is documented for Claude Code (code.claude.com/docs/en/sub-agents: "the per-invocation `model` parameter" outranks the definition's frontmatter; `fable` is a listed alias) and confirmed on the Cowork Agent tool. On a runtime that lacks it, copy `agents/fable-orchestrator.md` to `agents/fable-orchestrator-fable.md`, set `name: fable-orchestrator-fable` and `model: fable`, and route to that instead. 2. Spawn **@fable-orchestrator** via the Task tool (`subagent_type: "fable-orchestrator"`, `model: "fable"`). Brief it with: the user's task, the output directory, relevant session context, and any user-set limits (warning threshold, worker cap, deadline). 3. Do not restate the Core Loop or operational rules in the briefing — the orchestrator's agent definition carries them. Brief the task, not the method. 4. When it returns, relay the result, every stage it marked unverified, and its recommendations (surfaced scope it did not build). 5. **Mandatory delivery gate:** before presenting the result to the user, invoke the **double-check** skill on the finished deliverable. If the orchestrator's fable-verifier already cold-checked the *final* document (not just intermediate outputs), double-check will detect that and run only the synthesis-seam check instead of a full panel — its own rules handle this. Do not skip the gate; the user relies on it instead of re-checking the work themselves. ## Known limitation Same side door as fable-opus: the orchestrator keeps Bash for verification commands, and Bash can technically create files. Its definition forbids that use.