Skip to content

fable5-compliance.md

Latest commit

 

History

History
59 lines (47 loc) · 4.57 KB

fable5-compliance.md

File metadata and controls

59 lines (47 loc) · 4.57 KB
Raw

Fable 5 Compliance

Ground rules for working with Claude Fable 5 (claude-fable-5; claude-mythos-5 on Project Glasswing) — Anthropic's most capable model, premium-priced ($10 input / $50 output per MTok). These rules apply to every executor (Claude Code and non-Claude loop executors); they are enforced in code (Ai::Llm::ModelCapabilities, Ai::Llm::Adapters::Anthropic, Ai::Routing::EffortMapper, Ai::FableRouting), not something an executor hand-implements per call.

API rules — enforced by the request builder, do not hand-set

Rule Detail
Thinking Always-on, adaptive-only. NEVER pass a thinking block or sampling params (temperature/top_p/top_k) — Fable 400s on either; the builder omits/strips them.
Depth Controlled ONLY via output_config.effort, auto-derived from task complexity (Ai::Routing::EffortMapper) — do not try to tune depth another way.
Prefill No assistant-prefill turns — Fable rejects a trailing prefilled assistant message.
Data retention 30-day data retention is required for Fable/Mythos traffic.

Refusal handling — don't panic, don't manually retry

Fable's safety classifiers can decline benign security/infra work (stop_reason: "refusal", HTTP 200 — not an error). This is already handled: the platform reframes once with truthful authorized-context, then falls back to Opus, logging the event (Ai::ModelRefusalEvent) and attributing served-by. Repeated refusals for an (agent_type, category) combo auto-pre-route away from Fable going forward. Executors should NOT retry manually or treat a refusal as a failed task — it is handled and surfaced.

Routing

Fable is gated by Account#settings["fable_routing_enabled"] (Ai::FableRouting), DEFAULT OFF. While off, Fable is excluded from the candidate set entirely — non-selectable, never chosen by preference, UCB exploration, or a cost tie. When on, reasoning-tier allowlisted agent_types get a preference bonus toward Fable; repeat refusals still auto-pre-route away regardless of the allowlist.

Prompting Fable

Fable is tuned for long-horizon autonomous work, follows brief goal-level instructions better than enumerated ones, and degrades when over-prescribed. So "optimize for Fable" means removing scaffolding built for older models as much as adding rules: state the GOAL plus constraints over step-by-step, and let it plan its own approach.

These patterns are already wired into the platform's executor prompts — Ai::Agent::BASE_GUARDRAILS (every agent) and Ai::DevLoop::LoopGuardrails (the autonomous loops). Carry them yourself when prompting Fable directly:

Pattern What to do
Goal over steps State the goal + constraints; don't enumerate a procedure Fable can plan itself — over-prescription measurably lowers output quality.
Act when ready When you have enough to act, act. Don't re-derive settled facts, re-litigate decided calls, or survey options you won't pursue — give a recommendation, not a menu.
Checkpoint only when needed Pause for the user ONLY for a destructive/irreversible action, a real scope change, or input only they can provide — not to narrate or ask permission for in-scope reversible work.
Ground progress claims Audit every progress/completion claim against a real tool result before reporting it; say plainly when a step failed, was skipped, or is unverified. (Anthropic testing: nearly eliminates fabricated status reports.)
Don't over-refactor at high effort Higher effort tempts unrequested tidying. A bug fix needs no surrounding cleanup; validate only at system boundaries; no abstractions/flags/back-compat shims for hypothetical futures.
Lead with the outcome First sentence = the TLDR ("what happened" / "what I found"); detail after. Readable beats terse — no arrow-chains, jargon, or invented labels in user-facing text.
Give the reason Include intent ("I'm doing X for Y; they need Z") — Fable connects the task to context better than inferring it.
Delegate readily Fable dispatches parallel subagents well; hand off independent subtasks and keep working; prefer async over blocking.
Don't echo reasoning NEVER instruct Fable to show, transcribe, or explain its internal reasoning as response text — it trips the reasoning_extraction refusal → Opus fallback. Adaptive thinking is always on; read the summarized thinking block if you need visibility.

Effort is the depth control (auto-derived; do not hand-tune): high default, xhigh for the hardest work, low/medium for routine — Ai::Routing::EffortMapper.