summary

This PR implements ConditionalModelSettingsMiddleware, a middleware for dynamically configuring model bind settings based on runtime conditions.

What it does

Provides a fluent API (.when(condition).use(settings)) to apply different model_settings (e.g., parallel_tool_calls, strict) based on runtime conditions like message count or state flags. Multiple matching conditions apply cumulatively, with later settings overriding earlier ones for the same keys.

Use cases

• Disable parallel tool calls for long conversations
• Enable strict mode based on state flags
• Dynamically adjust tool calling behavior based on context
• Fine-grained control over model behavior in different scenarios

Why this PR

ref #34031 and #34010. Adding model_settings directly to create_agent would cause problems:

• User confusion: Unclear which settings belong in model initialization vs. model_settings
• Limited scope: Only agent-level configuration; users with scenario-specific needs would still need custom middleware
• Unnecessary complexity: Having both create_agent and middleware modify model_settings simultaneously

Solution advantages

• Flexible configuration: Supports both agent-level and condition-based local configurations
• Clear timing: model_settings application happens explicitly in wrap_model_call hooks
• Easy migration: Existing middleware users can easily adopt this built-in solution
• Separation of concerns: Keeps create_agent simple while encapsulating dynamic logic in dedicated middleware

Issue: Fixes #34031