cookbook-advisor

# cookbook-advisor ## §Purpose The Anthropic Cookbook 2026 catalogues four canonical patterns (Structured output / Chain-of-Verification / Citations API / Message Batches) consumed by framework users at runtime. This skill is the **routing surface** that `check_agent_spawn.py` fires when a spawn prompt or task description matches one of the pattern trigger taxonomies loaded from `.claude/data/cookbook_patterns.json`. Emit is **advisory-only** (UX hint via `audit_emit.emit_cookbook_pattern_advised`) and **never blocks** the agent dispatch path. Rate-cap is enforced via the PLAN-088 global `_plan088_rate_admit("cookbook_pattern_advised")` bucket — a single global cap, not per-pattern-class (R1 P1-7 correction locked into ADR-127 §Migration). The kill-switch env `CEO_COOKBOOK_ADVISOR_ENABLED=0` disables emit entirely. The skill shares advisory-only doctrine with `check_pair_rail.py` (both governed by ADR-127 Tier-B promotion path — see `.claude/adr/ADR-127-pair-rail-advisory-promotion.md` for the canonical governance contract). ## §Pattern catalogue (4 canonical IDs) ### COOK-P1 — Structured output / JSON mode **Intent**: LLM responses destined for downstream automation must be parseable. Free-form text introduces parsing risk (model prefixes a sentence before the JSON, wraps it in Markdown fences, etc.). Structured-output / JSON mode tells the model to emit *only* well-formed JSON conforming to a caller-supplied shape. Anthropic SDK exposes this via the `tool_choice` str...