[foureight84]

Ports iOS #140 (coach model catalog, 8833ee0) and adds what the iOS change doesn't have: the coach now reads each provider's own model list instead of relying only on preset model names that go stale.

Live model lists

Once a cloud provider is selected and its key is saved, Settings fetches the provider's model list (CoachModelCatalog):

Provider	Endpoint	What it gives
OpenAI	GET /v1/models	ids; filtered to chat models (no embeddings, TTS, image, realtime, dated snapshots or retired slugs)
Gemini	GET v1beta/models (paged, key in header)	models that support generateContent, with display names
OpenRouter	GET /api/v1/models	tool-capable models only, with each model's reasoning levels and retirement date
MiniMax	GET /v1/models	ids

The picker shows the curated presets first, then "All models from …" with a filter (OpenRouter lists ~390 tool-capable models). Under the picker:

• a status line, with a Refresh/Retry button;
• a warning when the stored model isn't listed any more (probably retired), so every turn doesn't fail with a 404;
• a note when OpenRouter has announced a retirement date.

The listing is advisory: the stored model stays a free string, and the presets are the fallback when the list can't be fetched. It is cached in memory for 6 h, keyed per provider and key. All four endpoints were checked to exist; OpenRouter's fields were read from the live response.

From iOS #140

• Presets:
  • OpenAI: GPT-6 Luna / Sol (default) / Astra.
  • OpenRouter: Claude Sonnet / Opus / Haiku 5.5, GPT-6.1 Sol and GPT-6 Luna, Gemini 4 Argon and 3.8 Flash, DeepSeek V4.1 Flash.
  • MiniMax: M2.5 and M2.7 are dropped. MiniMax closes them to new users Oct 19 and retires them Nov 18. A stored choice keeps working until then, and the warning flags it afterwards.
• Per-model reasoning levels (ReasoningEffort, none to max). The Reasoning picker lists only the levels the selected model accepts. The stored level is dropped when the model rejects it (ReasoningEfforts.resolve, applied in CoachClientResolver.coachSettings and meal analysis), because an unsupported level is an HTTP 400 on every turn. OpenRouter models outside the presets use the live reasoning.supported_efforts. The picker is hidden for MiniMax and local, whose clients send no reasoning field.
• Gemini now sends the level as thinkingConfig.thinkingLevel, only for Gemini 3 and later, since 2.x rejects it. Before this, Android ignored the setting for Gemini.
• Prices for the new models, cross-checked against OpenRouter's live pricing on 2026-10-14. Old entries are kept so stored choices still price.

Android differences

• Gemini presets: Android keeps its rolling -latest aliases (issue Chat Issues Gemini Model Old #22) instead of iOS's versioned slugs.
• Gemini thoughtSignature: Android already echoes it back.
• Upstream gap: iOS could adopt the live lists too.

Review

An independent read-only review found no blocking issues. Fixed in 189fecf:

• a superseded fetch could overwrite a newer result;
• offline/proxy modes dropped the effort;
• the Gemini page token wasn't URL-encoded;
• possible false "unlisted" warnings for MiniMax presets.

Left as is, and noted in the code: for a custom OpenRouter model, the live effort levels apply only after Settings has fetched the list in that session. Until then the stored effort is sent unchanged, as before.

Also: Today Activity tile clipping (4621312)

A user reported that the Activity tile cut off steps and distance and dropped calories. We don't know their build or phone, and I couldn't reproduce it: the tile has sized to its content since 6be5d1c. So the values column now makes clipping impossible. It still reports its full height so the tile grows when it can; if it gets less space anyway, it's drawn scaled down to fit instead of cut off. Not checked on a device.

Tests

./gradlew :app:testDebugUnitTest: 1519 tests, 0 failures. New CoachModelCatalogTest covers:

• the parsers (with an OpenRouter fixture trimmed from the live response);
• the "unlisted model" rule;
• reasoning-level resolution.

GeminiClientTest covers thinkingLevel. Existing fixtures were updated for the new defaults.

Not checked on a device, and no live call was made with a real key.

🤖 Generated with Claude Code