model-selection

name	model-selection
description	Guidance for picking the right model per role ("droid whispering"). Activate when configuring agents, starting a new mission, or diagnosing why a role is misbehaving. Model strings age fast — verify current names via search per P-2 before relying on the defaults below.

Model selection (droid whispering)

No single model is best at planning, implementing, and validating. Each agent in .claude/agents/ sets its own model: field. Change them per project and per role.

P-2 applies to model strings

Per the version-freshness skill, the model identifiers below age faster than almost anything else in the system. Anthropic ships new models frequently; deprecated model IDs may still resolve for a transition window but get worse responses than the current generation.

Before relying on the defaults below, search to verify which model IDs are current. Useful queries: Anthropic latest models, claude opus current version, claude sonnet current version. Update the agent files' model: field with the verified-current ID and annotate with a source URL comment if you want auditability.

Default mapping (as of scaffold creation; verify before use)

Role	Capability needed	Default model
Orchestrator (main session)	Slow careful reasoning, conversational scoping, judgement at milestone boundaries, web-search execution	claude-opus-4-8
Worker	Fast code fluency, willingness to iterate, tight instruction-following inside a fresh context	claude-sonnet-4-6
Scrutiny validator	Precise instruction-following, adversarial bias, willingness to reject	claude-opus-4-8 (ideally a different model family entirely if you have access — see below)
UX validator	Patience with flaky UI, reliable tool-use, ability to design real user flows	claude-opus-4-8

Search-verify these strings before assuming they're still current. If a search reveals a newer generation, update the agent files.

When to deviate

• Throwaway prototype mission, cost matters more than iterations — Sonnet or Haiku everywhere. Expect more validator-driven follow-ups.
• Production-critical mission (migration, refactor, anything customer-visible) — Opus everywhere. The serial bottleneck dominates wall-clock time; model cost is rounding error compared to a human review.
• Scrutiny validator suspiciously lenient — if you have access to a non-Anthropic frontier model via a proxy, put it here. Shared training data means an Anthropic model reviewing Anthropic-written code is more likely to share the same blind spots. This is the single highest-leverage swap.
• Worker keeps producing vague handoffs — first tighten the structured-handoffs skill template, then upgrade the worker model. Not the other way around.
• Orchestrator keeps drifting from the plan — upgrade the orchestrator model. Planning is the role most sensitive to model quality.

Per-mission overrides

Drop missions/<id>/model-overrides.yaml:

worker: claude-sonnet-4-6
scrutiny-validator: claude-opus-4-8
ux-validator: claude-opus-4-8

The orchestrator reads this file (if present) before spawning each subagent and passes the model via the Agent tool call, overriding the agent file's default.

Watching for failure modes

After every milestone, glance at:

• Average worker iterations before COMPLETE — rising means worker model is below the task's difficulty
• Scrutiny rejection rate — falling toward zero across milestones is suspicious (could mean coverage is dropping, not quality rising)
• UX validator INCONCLUSIVE rate — rising means the app is getting harder to drive, often a sign of accidental coupling

If any of these trend wrong over 2 milestones, change models before they compound across a multi-day run.

When model IDs in agent files are out of date

If you spin up a mission and a worker subagent fails to spawn with "unknown model" or returns degraded responses, the model ID is likely deprecated. Web-search for the current name (per P-2), update the relevant .claude/agents/<role>.md file's model: field, and re-run the failed phase command.