model-routing

name	model-routing
description	Canonical model-tier reference and the Fable-5 manual-only enforcement protocol. Consult before writing or changing any model field (agent frontmatter, skill frontmatter, Task dispatch parameter, settings.json availableModels, env vars). Auto-invocable — any agent or skill authoring a new artifact should load this first.
model	haiku

model-routing — consult before touching any model field

This skill is a pointer to the single source of truth, ai_docs/model-routing-and-fable-policy.md — it summarizes rather than duplicates that document at length. If anything here and that file conflict, the ai_docs file wins; re-read it directly for the full text.

If you are about to write model: fable anywhere except .claude/skills/plan-deep/SKILL.md, STOP — that is a CONSTITUTION.md N2 violation. No exceptions, no "just for this one special agent."

Tier table (Opus / Sonnet / Haiku — freely auto-routed)

Tier	Used for
Haiku	Cheap classifiers/routers/triage — triage, profile-loader, taxonomy-mapper, judge-router, missability-inspector, visual-regression-runner.
Sonnet	Default workhorse — most generation/implementation/review agents (spec-author, engineer, api-designer, data-modeler, test-strategist, release-planner, ops-author, ai-controls-author, retirement-planner, discovery-researcher, designer, design-system-curator, reflexion-coach, oracle-evaluator, master-plan-patcher, run-finalizer, agents-md-author, governance-author, browser-validator, docs-author).
Opus	Highest-stakes judging/architecture/security/synthesis — judge-cross-vendor, verifier, meta-agent, strategy-author, architect, security-reviewer.
Fable	Never auto-routed. Reachable only through /plan-deep (disable-model-invocation: true), and only after either (a) direct user invocation of that command, or (b) an agent-suggested path that has gone through the AskUserQuestion-then-token flow below.

The six surfaces Fable-5 could leak through (why frontmatter discipline alone is not enough)

1. Subagent frontmatter model: field.
2. A per-invocation model parameter when dispatching a Task sub-agent.
3. The CLAUDE_CODE_SUBAGENT_MODEL environment variable.
4. The interactive /model command.
5. /advisor / advisorModel.
6. teammateDefaultModel (Agent Teams).
7. (Structural, addresses all of the above) the project/user availableModels setting in .claude/settings.json.

The full enforcement protocol

1. .claude/settings.json sets availableModels to exclude Fable-5 from the pool available to the main interactive session and to all subagent/teammate dispatch. This is the primary structural gate — if Fable isn't in the available set, none of surfaces 1-6 can silently select it.
2. No agent file (.claude/agents/*.md) ever sets model: fable.
3. Fable is reachable only through /plan-deep, gated by disable-model-invocation: true.
4. The suggestion-and-approval flow lives entirely in the interactive main-loop/skill layer, never inside a Workflow() script (workflows have no mid-run human input). An agent that judges a task Fable-worthy uses AskUserQuestion in the normal conversational turn. If the user says yes, the harness writes a single-use .fable/fable-approval.token; only then does a separate, freshly-invoked /plan-deep call proceed with Fable. If the user says no, the token is never written.
5. A UserPromptSubmit hook denies any literal /model fable or /advisor-to-Fable switch unless .fable/fable-approval.token exists and is unconsumed.
6. This document + ai_docs/model-routing-and-fable-policy.md are the canonical reference every future agent/skill author (including meta-agent) must consult before ever touching the model field.

Constitution citation

Cite N2 when refusing any attempt to write model: fable outside plan-deep, or when denying an unapproved /model//advisor switch to Fable: Refused per N2: Fable-5 is manual-only, never auto-routed.