fable-opus

# Fable Mode — Opus (v6, main-thread orchestration) Family version: v6 (2026-10-10). v4 removed the named-agent routing from v3. That design depended on custom agent definitions (`fable-orchestrator`, `fable-worker-sonnet`, `fable-worker-haiku`, `fable-verifier`) that do not exist on skill-only surfaces like Cowork — so every invocation fell through to the fallback while the file still described the dead path as primary, and the v3 structural guarantee never actually applied. The fallback is now the only path, stated directly. v5 change: this skill is now the top of the ladder. The Fable 5.1 runner was removed from the skill set (its own benchmark measured zero correctness delta at that tier), so escalation past Opus goes to the user, not to another runner. v6 change (2026-10-10, cloud review): v5 told a spawned Opus worker to spawn its own Sonnet/Haiku workers, but in the cloud runtime general-purpose subagents had no Agent tool (Haiku and Sonnet checked; Opus unchecked). The worker's brief also pasted fable-mode's rule that a spawned subagent spawns no subagents of its own, plus fable-mode's double-check gate, so the gate could run twice. Fan-out now starts from the main thread, the worker brief omits fable-mode steps 2 and 5, and the gate runs once, on the main thread. ## Invocation gate When the user names this skill — a slash command, or "fable on opus" in words — RUN IT. Do not weigh whether the task looks big enough. The user's explicit invocation IS the trigger; the cost call is theirs, not yours. The "skip for single-pass tasks" judgment below governs only autonomous triggering — this skill deciding on its own to fire. It never overrides an explicit request. If you think the loop is overkill for what was asked, run it anyway and say so afterward. **Autonomous trigger only:** if a task has one obvious correct approach, fits in a single pass, and the user did not ask for this skill, do it directly. **Spawn authorization** (same rule as fable-mode): a slash command or the skill named ("fable-opus", "fable on opus") authorizes spawning. A wording-only trigger ("stage this on opus", "deep work mode, opus") or an autonomous trigger does not: run the fable-mode Core Loop inline on this thread, run double-check in its self-check mode, and say so. If this thread is not on Opus, also say the Opus pin was skipped and that a slash command requests the pinned run. Usage limit: same as fable-mode — when the user reports usage at their threshold, check in first. ## How to run it 1. **The main thread orchestrates.** Subagents may have no Agent tool of their own, so every spawn starts here. Without spawn authorization, follow the Invocation gate's inline path instead. Otherwise take the first branch that fits: - **Main thread on Opus** (per its own model line) → run fable-mode Core Loop steps 1–3, and step 4 without its fresh-verifier paragraph (step 4 below handles the verifier), on this thread at frontier-tier calibration. With the Agent tool, spawn stage workers directly from here: `model: "sonnet"` for ordinary stage reasoning, `model: "haiku"` for bulk mechanical work; without it, run every stage here. Opus tokens buy judgment and synthesis, not file-shuffling. Respect any user-set worker cap. Then go to step 3. - **Main thread on any other model, and the Agent tool offers `"opus"`** → spawn ONE worker with `subagent_type: "general-purpose"` and `model: "opus"`, briefed per step 2. - **Otherwise** (no Agent tool, or no `"opus"` model offered) → say the Opus pin is not available here, run fable-mode Core Loop steps 1–4 inline on this thread, sequentially — then go to step 5. 2. Brief the Opus worker with: - the user's task, verbatim - the output directory - relevant session context - any user-set limits (warning threshold, deadline) - from the **fable-mode** skill (installed alongside this one), copied **verbatim**: Core Loop steps 1 and 3, step 4 without its fresh-verifier paragraph, the domain patterns, and the operational rules. Do NOT paste step 2 (delegation) or step 5 (the double-check gate). - the three execution-guardrails rules, inlined verbatim (below) - this instruction: "Run every stage yourself and spawn nothing — the main thread handles verifiers and the delivery gate. If a third replan seems needed, stop and return the ambiguity to the main thread; do not ask the user — this overrides the pasted Replan-budget wording. Return a report that lists the stages you ran, numbered, each with its artifact and the exact check that verified it, plus every stage you marked unverified and any scope you found but did not build." 3. **Wait for every result.** Spawns may run in the background even when foreground is requested. Do not merge, fix, relay, or start the gate until every spawned agent — worker or verifier — has returned. If a spawn errors or fails to return, say which, do its part inline, and label that part in the delivery. 4. Relay: the result, every stage marked unverified, any scope surfaced but not built, and any replan ambiguity the worker returned — put that to the user for a decision. For a high-stakes deliverable, where the Agent tool is available, spawn one fresh verifier from this thread on the finished deliverable, briefed only with the spec and the artifact, and wait for its verdict; only then may step 5's line say `yes`. Without the Agent tool, the line cannot say `yes`. If the task is beyond Opus's capability, say so to the user plainly — this is the top of the ladder and there is no further tier to hand off to. That report must follow verify-before-flag: name what was attempted and where it failed, not a vague appeal to difficulty. 5. **Mandatory delivery gate — once, on the main thread.** Before presenting anything to the user, invoke the **double-check** skill on the finished deliverable. Never run it inside a worker. Close your own report to double-check with a literal line: ``` FINAL-COLD-CHECK: yes | intermediates-only | no ``` - `yes` — a fresh verifier checked the FINISHED deliverable cold. Then double-check runs the seam check only. - `intermediates-only` — verification happened at stage level but nothing checked the final write-up. Full checks run. - `no` — no cold verification happened. Full checks run. Emit the line honestly. A missing or wrong line makes the gate either redundant or absent. Do not skip the gate; the user relies on it instead of re-checking the work themselves. ## Execution guardrails (inline these into every worker briefing) Copy these three rules into every worker briefing verbatim. Cloud subagents could load skills (2026-10-10); other surfaces are unchecked. The inline copy is shorter than a skill load and works either way. 1. **Verify before flag.** Never report a problem that has not been confirmed present by a direct check — grep it, run it, read the source. Absence of evidence is not the finding. Web silence is never grounds for a warning about the user's own firsthand facts. 2. **Warning threshold.** Batch minor concerns. At three (or the user's set number), stop and surface all at once. A concern that independently meets the verify-before-flag bar and is material on its own does not wait. 3. **Find-and-replace safety.** Anchor sed/substring replaces on word boundaries (`\bword\b`), never bare. Grep for corrupted compound words after every pass. ## Known limitation Delegation is still requested in prose, not enforced structurally. The orchestrating thread has Write and Edit, so nothing stops it doing stage work inline on Opus instead of spawning Sonnet/Haiku workers. Make that visible rather than silent: the stage map names, per stage, which tier produced its artifact. v3 tried to enforce delegation by removing Write from a named orchestrator agent; that needs custom agent definitions, which skill-only surfaces cannot load.