feat: apply reasoning effort and surface thinking blocks for Google chat models - #28273

…king level

applyReasoningEffort had no case for the google provider, so a chat's
selected reasoning effort was silently dropped for Gemini models. With
only include_thoughts configured, Gemini 3.7 Flash never thinks
(reasoning_tokens is 0 on every step), so no reasoning parts stream,
persist, or render.

Map the global effort scale onto thinking_level (clamped to
MINIMAL..HIGH), preserving any explicitly configured thinking_budget
since fantasy rejects requests carrying both.

…dels

Gemini 2.5 and older reject requests carrying thinking_level, so
applying it for every Google model would turn a configured reasoning
effort into a hard generation failure on those models. Parse the major
version from the model ID and keep dropping the effort for pre-3 Gemini,
non-Gemini, and unrecognized model IDs.

…Gemini-3 models

The config conversion path copied a pinned thinking_level for every
Google model, so an admin pinning a level on a Gemini 2.5 config would
fail every generation. Gate conversion on the same model capability
check as the reasoning-effort path. Gating at call time rather than
config save time also covers updates that switch a config's model
without resubmitting options.

… model's supported set

Gemini models accept different thinking_level subsets: 3 Pro launched
with LOW and HIGH only, 3.1 Pro added MEDIUM, the Flash family accepts
all four, and image models accept only HIGH (plus MINIMAL for flash).
Replace the boolean Gemini 3+ gate with a per-model supported set and
round-up clamping so an effort or pinned level maps to the nearest
level with at least the requested depth instead of a rejected request.
Versionless -latest aliases are recognized as current-generation
models; unknown Gemini 3+ variants fall back to LOW and HIGH, the
intersection of every documented non-image set.

…e openai-compat path

Google AI providers route through aibridge as openai-compat clients, so
the native thinking_level path never runs for chat generation. Google's
OpenAI-compatible endpoint translates reasoning_effort into the model's
thinking configuration but validates instead of clamping, so efforts
outside the model's supported set (none/minimal/xhigh/max on Gemini 3.1
Pro, verified live) fail the whole request with HTTP 400. Clamp the
effort into each Gemini model's supported set before dispatch, reusing
the thinking-level tables from the native path.

…pat patch overview

…rite on model capability

Pre-2.5 and unrecognized Gemini variants have no thinking support, so
sending thinking_config could reject previously working requests.
Restrict the rewrite to models with a known thinking_level set or the
Gemini 2.5 thinking_budget family and leave other Gemini requests
untouched.

… the 2.5 chat families

Specialized Gemini 2.5 variants such as image and TTS models reject
thinking_config outright, so a version-only predicate would break
previously working requests for them. Recognize only the Pro, Flash,
and Flash-Lite family name shapes and fail closed on unknown tokens.

…ig on the AI Bridge route

Google models backed by an AI Provider route through the
OpenAI-compatible client, which ignores the fantasygoogle options key,
so an administrator's pinned thinking_level or thinking_budget was
silently dropped. Translate the pinned config into the compat request's
extra_body and merge it in the transport patch with native-path
precedence: a pinned budget wins over the per-turn effort, the effort
overrides a pinned level, and an explicitly disabled include_thoughts
is preserved.

… Gemini's thought metadata

An answer that legitimately begins with the <thought> marker text was
reclassified as reasoning and truncated. Real thought output always
carries message.extra_content.google.thought, which the streaming path
already requires, so demand it before splitting non-streaming content.

…sponses from the raw upstream body

The typed openai.ChatCompletion round trip drops provider-specific
fields, so a blocking Gemini response with thought output lost its
extra_content thought metadata while keeping the <thought> markers
inline, leaking them to clients as answer text. Serialize the final
blocking response from the raw body for Google upstreams, mirroring
marshalChunk on the streaming path, with the same ID and usage
overrides.

… Gemini 3.7+ Flash

gemini-3.7-flash rejects thinking_level MINIMAL with HTTP 400, so the
default reasoning effort (none) failed every generation. Versions 3
through 3.6 of the flash and flash-lite families still accept it
(live-verified against the Google API). The versionless latest aliases
and unknown future flash versions also stay off MINIMAL because
gemini-flash-latest already tracks 3.7 and returns the same 400, while
clamping up to LOW is accepted by every flash release.

…3 variants

Codex found that the thinking-level map classified every Gemini 3+ model
containing the flash token as thinking-capable, so specialized variants
like gemini-3.1-flash-live-preview and gemini-3.1-flash-tts-preview got
thinking_config injected on the compat path and rejected the request.
Live-verified: the tts variant returns 400 (thinking level not
supported), the live variant does not serve generateContent, and
gemini-omni-flash-preview only supports the Interactions API. Gate the
Gemini 3+ set to the Pro and Flash chat families with the same
unknown-token fail-closed rule as googleSupportsThinkingBudget.

Also address human review feedback: assert the rewritten completion
keeps the bridge-side usage override, and join read and close errors
with errors.Join instead of a hand-rolled firstError helper.

…and budget mutual exclusion

Codex asked for interaction coverage of the new Google thinking_level
selector's conflicts_with wiring. The story selects a level and asserts
the budget input is disabled, clears it, sets a budget, and asserts the
level select is disabled, covering both directions of the exclusion.