fix: extract LangChain content-block text and apply model parameters after eval

[andrewklatzke]

Summary

• Extract visible text from LangChain AIMessage.content when it is a list of thinking/text blocks so callers no longer get an empty output while tokens are still billed.
• Accept llm as a factory (config) => model (in addition to a pre-built instance) and spread flag model.parameters unchanged into default ChatAnthropic/ChatOpenAI constructors after evaluation.
• Add langchain-thinking example (tool-free prompt) plus unit tests for mixed content blocks, factory vs instance, and default constructor kwargs.

Test plan

• packages/langchain-messages and packages/langchain-agents unit tests
• python main.py langchain-thinking launch-darkly-documentation-summarizer-messages-claude "Reason it out yourself without any tools: what is 17 times 23?" — non-empty response
• Same scenario with doesnt-exist — exit 1

Made with Cursor

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 7beaa6f. Configure here.}

[cursor]

Example factory kwargs collide with parameters

Medium Severity

The thinking example spreads flag model.parameters into ChatAnthropic and then passes overlapping keywords (timeout, stop, model_name, thinking, max_tokens_to_sample). A second value for any of those keys is a TypeError, and max_tokens aliases max_tokens_to_sample, so a typical Claude flag never constructs the model.

 

^{Reviewed by Cursor Bugbot for commit 7beaa6f. Configure here.}

[tracisiebel]

just curious - why the change from string -> any?

[andrewklatzke]

This PR is specifically targeting langchain's "content blocks" pattern it uses for some providers. So no longer a string. Any type we provided here given the breadth of what it can return would be incorrect for someone depending on provider, etc.