profile-policy

Model routing policy

Routing lives in one place: the PROFILE line and table in the vibe:conduct skill. No agent, command, or skill names a model in prose; roles map to tiers there, so changing models is a config change, not a rewrite.

The two profiles

• uniform — every role runs on the session model (whatever /model is set to). The safe default when cost-tiering isn't wanted, or when the smaller tiers / a target model aren't available yet.
• tiered — turns on cost routing: implementation on a mid tier, mechanical edits on the cheapest, escalation and all review/verification on the strong tier. Reviewers and verifiers stay on inherit because judgment wants the strongest model.

Switch by editing the PROFILE: line in vibe:conduct. Override for a single run with the skill argument: /vibe:conduct uniform ....

Tiers are aliases, never dated ids

Set each agent's tier through its model: frontmatter using an alias — sonnet, haiku, opus, fable, or inherit (the session model). Never bake a dated id like claude-opus-4-8 into a role; aliases survive model changes, dated ids rot.

The one global switch: CLAUDE_CODE_SUBAGENT_MODEL

To flip the entire kit to a different model (e.g. Opus 4.8 today → Fable 5 when it returns), set the CLAUDE_CODE_SUBAGENT_MODEL environment variable. It is the highest-priority override for subagent model selection, so it takes effect with zero file edits.

Resolution order (highest priority first):

1. CLAUDE_CODE_SUBAGENT_MODEL env var
2. the per-invocation model parameter (the PROFILE column applied at dispatch)
3. the agent's model: frontmatter
4. the main conversation model

Set it to inherit for normal resolution (let the lower levels decide). Set it to an alias to force every subagent onto that model regardless of profile or frontmatter.

See vibe:fable-safe-authoring for why no role is tied to a model name in prose, and how to keep authored prompts from silently falling back off the model you targeted.