Fix Gemini thinking-only STOP handling

[davideast]

The Problem

Gemini requests were treating omitted or disabled reasoning effort as a request for visible thought summaries. That made ordinary calls more likely to end with thinking output but no visible assistant text or tool call, especially on long ReAct turns.

The failure was also hard for callers to classify. A thinking-only STOP looked like a plain message string, and the agent layer dropped provider metadata as the event moved through strategy and session surfaces.

A typical failure looked like this: a host asks the agent to make a change, Gemini successfully calls a write tool, the tool result is dispatched, and the next model pass ends after internal thinking with finishReason=STOP but no visible output. Before this change, that final provider failure could make the whole turn look failed even though useful tool progress had already completed.

What Changed

Gemini now only requests visible thought summaries when the caller explicitly asks for reasoning with low, medium, or high effort. Omitted reasoning effort and reasoningEffort: off leave includeThoughts out of the request body.

When reasoning is requested, Gemini request construction uses deterministic model-family controls: Gemini 3.5 and Gemini 3 Flash receive thinkingLevel, Gemini 2.5 receives the configured thinkingBudget, and unknown Gemini model families only receive includeThoughts.

Gemini no-output streams now emit structured model errors with stable codes, retryability, and stream diagnostics. Thinking-only STOP, truncated no-output, malformed function call, and deterministic no-output failures are distinguishable without parsing prose. Retry-exhausted errors also carry final attempt metadata.

The Agent Behavior

The shared model error event now keeps message while allowing code, retryable, and details. Agent strategy, retrieval, relay, and session surfaces preserve that metadata instead of flattening it away.

ReAct stays strict by default. The new toolProgressErrorPolicy option is opt-in, and complete-with-warning only applies after successful tool progress in the current run and only for the known retryable gemini.thinking_only_stop error. In that narrow case, the strategy emits react_loop_model_warning and returns normally without synthesizing another turn completion.

Why This Is Safer

The default library behavior remains fatal on provider errors. Hosts that need conservative side-effect completion can opt into the progress-aware policy, while docs-site and ordinary ReAct callers keep the existing strict semantics.

The fix also avoids changing Gemini tool-call accumulation, schema sanitization, max output token behavior, and thought-signature replay, so the existing Gemini tool path keeps its current shape while the no-output failure path becomes explicit and actionable.