feat(openai): surface Gemini thinking + request thought summaries

[philipph-askui]

The problem (conceptually)

A reasoning model keeps its "thinking" separate from its spoken answer. Claude returns that reasoning as a distinct, typed thinking block; the OpenAI-compatible path (how Gemini is served via the AskUI API) is messier and, as a result, dropped the reasoning entirely:

1. We never asked for the thoughts. Gemini (via Vertex's OpenAI-compatible endpoint) only returns a thought summary if you explicitly opt in. We weren't opting in.
2. Even when thoughts came back, nothing extracted them. The summary arrives either in a side field (reasoning_content / reasoning) or — for Gemini — inline in the answer text, wrapped in <think>…</think> tags. The parser ignored the side field and passed the inline tags straight through as part of the answer.

So a Python-SDK user running Gemini through the AskUI API saw no reasoning at all, or raw <think> tags leaking into the answer. (There's also a sharp edge: Vertex 400s if a request carries both reasoning_effort and a custom thinking_config — "found both" — so the opt-in must use thinking_config alone.)

This path previously only carried the Gemini thought_signature (multi-turn tool calling, #290); reasoning was never surfaced.

The fix (conceptually)

Make the OpenAI-compatible path speak the same internal thinking vocabulary as the Anthropic path:

• Ask for the thoughts on the AskUI Gemini path (extra_body.google.thinking_config.include_thoughts, deliberately not reasoning_effort).
• Recognize reasoning in both forms in the response parser — the side field and inline <think>…</think> — and lift either into the same BetaThinkingBlock the Anthropic path already produces, ahead of the spoken text.

Changes

• models/openai/messages_api.py — parse reasoning_content / reasoning; split inline <think>…</think> out of content; emit a BetaThinkingBlock before the text block.
• model_providers/askui_vlm_provider.py — on the Google backend, request thoughts via extra_body.google.thinking_config.include_thoughts (caller provider_options win).
• Unit tests for both parse forms + the request wiring (55 passed on the affected files); ruff + mypy clean.

Related PRs

This is a parity port of the C# SDK change:

• C# SDK (original fix): askui/csharp-sdk#13
• Desktop app (renders the result): askui/integrated-task-platform#177 — where the fix was validated against a live Gemini run.

The desktop app runs on the C# SDK, so it does not depend on this PR; this is a standalone-product parity fix for Python-SDK users.

🤖 Generated with Claude Code