Summary

• When adding a Custom provider whose URL looks like Ollama (port 11434 or hostname contains ollama), prefer native GET /api/tags so tools / thinking / vision / context window prefills are richer than sparse /v1/models.
• Soft-fail the probe (network error, non-Ollama JSON, empty catalog) and fall back to /v1/models so non-Ollama custom endpoints keep working.
• Persist advertised capabilities into the model row (toolUsage, startWithReasoning, contextWindow) on create and edit, show them on model detail, and add a per-conversation Thinking chip in the composer for models with startWithReasoning === 1.
• Chip off does not mutate the stored model (so the chip stays visible); it overrides the send path (startWithReasoning view + reasoningEffort: 'none' for OpenAI-compat / Ollama, and Pi thinkingLevel: 'off' in the built-in adapter).

Durable contract is the model row + session override — not “we are permanently bound to Ollama.” Autodetection is opportunistic.

Heuristic / failover notes

• Probe gate: isLikelyOllamaBaseUrl — intentional incomplete match so llama.cpp / corporate OpenAI-compat URLs are not hit with an extra failed /api/tags.
• Ollama on an uncommon port without ollama in the host still works via /v1/models (no rich caps).
• /api/tags body is zod-validated (ollamaTagsResponseSchema envelope + per-tag ollamaTagModelSchema); malformed tags are skipped; empty/invalid → soft-fail.

Test plan

• Unit: ollama-catalog origin/heuristic/mapping/schema
• Unit: fetchModelsForProvider prefers /api/tags; fails over on throw and non-Ollama body
• Unit: add-model form persists Ollama-advertised flags
• Unit: thinking-session helpers + openaiCompatThinkingProviderOptions
• Unit: withThinkingSessionOverride (ACP / Pi path)
• Unit: ThinkingChip a11y pressed/toggle
• Manual: Custom URL http://localhost:11434/v1 → catalog shows tools/thinking/context; create model; chip appears in chat
• Manual: toggle Thinking off → send does not stream reasoning (Ollama thinking model)
• Manual: Custom URL non-Ollama (e.g. llama.cpp other port) → no /api/tags probe failure in UX; /v1/models path works
• Manual: Edit model → change Thinking / Tools / Context window → detail + chip reflect updates
• Manual: switch to a non-thinking model → chip hidden; prior chat’s thinkingEnabled does not leak (reset on model switch)

Reviewer notes

Addressed prior review:

1. Dropped ALL_CAPS REASONING_OFF (inlined 0)
2. Removed dead contextLengthFromModelInfo; wired ollamaTagsResponseSchema into the fetch path (envelope stays unknown[] for per-tag skip)
3. Removed as never casts on stubJsonResponse mocks
4. Spread ResolvedPiModel consistently in both withThinkingSessionOverride branches

Made with Cursor