feat(agent): support per-task model selection for subagents (expose user role profiles with cost class)

[7jrxt42BxFZo4iAnN4CX]

Summary

Codewhale cannot run a sub-agent on a different model per task. The model-facing agent tool advertises only the 8 built-in roles and, absent a profile, the child inherits the parent's model. There is no supported way for the main agent to say "run this sub-task on the free model" or "run the critic on the strong model", and user-authored role profiles are invisible to the model.

In 0.9.9 (#5324) the agent tool schema was intentionally slimmed from 33 to 12 advertised fields; the routing overrides model, model_strength, and thinking were removed from the advertised surface. They are still parsed and honored when a caller sends them (parse_spawn_request reads model/model_strength/thinking; parse_optional_subagent_model feeds resolve_spawn_model_selection), but because they are absent from the schema and description, the model never learns they exist — so per-task model selection is effectively unusable.

Current behavior

• agent(action=roster) returns only the 8 built-in roles and states: "There are no saved members: every spawn resolves a role only."
• User agent profiles (~/.codewhale/agents/*.toml, workspace .codewhale/agents/, saved Fleet members) are not listed to the model. Passing profile=<user id> fails with "Unknown Fleet role/profile … Fleet profiles are roles" (resolve_spawn_role requires a built-in FleetRole).
• A sub-agent inherits the parent's model, or uses a global [subagents] <role>_model default. A per-call model is validated against the parent's provider (normalize_requested_subagent_model(…, runtime.client.api_provider())), so requesting a model from a different provider is rejected.
• The only "cheaper sibling" today is model_strength: "faster" (same provider family), not an arbitrary cheaper model.
• Modes (Plan / Work / Operate) are permission postures, not model selectors, so they offer no per-mode model either.

Desired behavior

1. Per-task role/model choice visible to the model. Either re-advertise model/model_strength/thinking on the agent surface, or expose user profiles as selectable rows, so the main agent can pick the appropriate model/role per spawn instead of being limited to one global default.
2. Cost-aware selection. When profiles are listed, include enough for a deliberate choice: provider, model, and a cost class (free / paid / unknown), derived from route pricing metadata or an optional profile field. Example roster row: bug-hunter · openrouter · minimax/minimax-m3:free · class=free. The model could then use the free profile for wide bug-scanning and a paid/strong profile when a hard critic is needed.
3. Unchanged safety: a child never widens the parent's authority; containment clamps stay as-is.

Proposed (maintainer's choice)

• Minimal: re-add model/model_strength/thinking to the advertised agent schema and description (the fields already parse and are honored), so the parent model can request a per-spawn model on its own provider.
• Fuller: action=roster also lists user profiles (id, base role, provider, model, cost class); profile=<name> resolves user profiles as well as built-in roles; cost class comes from pricing metadata or an explicit (new) profile field.
• Cross-provider support is a separate question; if not added, requests for a foreign-provider model should keep failing with a clear message rather than silently inheriting.

Acceptance criteria

• A user with two profiles (e.g. bug-hunter on a free model, critic on a paid model) can tell the main agent "spawn bug-hunter for this scan, critic for the review", and the agent selects the intended profile/model per spawn.
• action=roster (or equivalent) shows the model which roles/profiles exist and which are free vs paid.
• Per-spawn model/thinking overrides on the parent's provider work when the model sends them; a different-provider model is rejected with a clear error.
• Authority/containment behavior is unchanged.

Environment

• codewhale 0.9.12 (dev), Linux, tmux.

[github-actions]

Thanks @7jrxt42BxFZo4iAnN4CX for the report.

This issue is staying open for maintainer triage. CodeWhale gets better because people bring us real edge cases from real machines, providers, regions, and workflows.

If you can add a reproduction, logs, version output, screenshots, or the provider/model involved, that makes it much easier for us to verify and harvest the fix. Maintainers may comment /lgtmi to mark recurring issue reporters as approved so this intake note is skipped next time.

[7jrxt42BxFZo4iAnN4CX]

Closing as subsumed by #5915 (opened today, founder direction): the model shortlist flow (provider → model → shortlist → role, with price-class/strength/context tags on each fleet model and model constrained to the list) covers the core ask here — letting the operator pick a per-task model deliberately (e.g. a free model for bug-hunting, a stronger one as critic). The one part of this request #5915 does not currently mention is listed as a suggestion there.

Thanks — tracking #5915.