Surface model thinking controls in Chat and Agent Mode

[AnthonyRonning]

Context

Once OpenSecret exposes authoritative per-model thinking metadata and accepts a normalized thinking level on initial and follow-up requests, Maple should make that control available in both of its chat surfaces. Maple should consume OpenSecret's capability metadata rather than maintaining its own hard-coded list of reasoning models or provider-specific parameters.

Dependency

• Blocked on: Expose model thinking capabilities and per-response controls opensecret#278

Scope

• Add a compact thinking control (button, segmented control, or dropdown as appropriate to the existing composer UI) to:
  • the standard Maple chat composer;
  • the Maple Agent chat composer backed by Goose.
• Populate the available choices and default from OpenSecret model metadata for the selected model.
• Hide or disable the control for models that do not support configurable thinking.
• Support the normalized options advertised by the backend, expected to include combinations such as off, low, high, and max depending on the model.
• Send the selected value on initial requests and every user-initiated follow-up request.
• Wire Agent Mode through Maple's custom Goose provider so it sends the OpenSecret-supported setting without relying on Goose model-name heuristics or duplicating provider-specific translation in Maple.
• Keep the normal chat and Agent Mode behavior consistent.

Preference persistence

• Persist the user's selection similarly to Maple's current model preferences.
• Key the preference by model so a level selected for one model does not leak into an unsupported or differently configured model.
• Use the model's advertised default when no user preference exists.
• Reconcile stale stored values safely if backend metadata changes and a previously selected level is no longer supported.
• Use the same model-specific preference across both chat surfaces unless there is a concrete product reason to keep them separate.

UX considerations

• Make the current setting visible without overcrowding the composer.
• Clearly distinguish non-thinking/off from active thinking levels.
• Avoid implying that every reasoning-capable model supports every level.
• Preserve the selected setting as the user switches conversations, models, and between standard Chat and Agent Mode, subject to each model's advertised capabilities.

Acceptance criteria

• Both composers render controls from backend metadata rather than hard-coded model IDs.
• Unsupported models do not show an actionable thinking selector.
• Supported selections are included on initial and follow-up requests in standard Chat and Agent Mode.
• Model-specific selections persist across application restarts and do not leak between models.
• A removed or changed backend capability falls back to the advertised default without breaking message submission.
• Tests cover capability loading, model switching, preference persistence, follow-up requests, and both provider paths.
• Existing conversations and models without configurable thinking remain backward compatible.