🪙 fix: Max Output Tokens Refactor for Responses API

[dustinhealy]

Summary

This pull request fixes an error raised in #8966 stemming from how the maxTokens parameter is mapped for the new models, ensuring that when the useResponsesApi flag is enabled, maxTokens is moved to max_output_tokens instead of max_completion_tokens.

• Updated OpenAIClient.js, memory.ts, and llm.ts to set max_output_tokens (instead of max_completion_tokens) when useResponsesApi is true, and to use max_completion_tokens otherwise. This ensures consistent parameter mapping for GPT-5+ models. [1] [2] [3]

Testing and validation:

• Added and modified tests in client.test.js to verify that maxTokens is correctly moved to either max_output_tokens or max_completion_tokens depending on the model and useResponsesApi flag. Tests also confirm that only GPT-5+ models are affected by this logic. [1] [2]
• Added tests in memory.test.ts to ensure the correct parameter (max_output_tokens or max_completion_tokens) is set in the LLM config passed to the agent runner, based on the value of useResponsesApi.
• Updated llm.spec.ts to expect max_output_tokens instead of max_completion_tokens when useResponsesApi is true.

Change Type

• Bug fix (non-breaking change which fixes an issue)

Testing

Verified absence of error when receiving responses in conversations with gpt-5 as a model and max token params set.
Confirmed no regression with earlier models and non Responses API requests using gpt-4o.

Checklist

• My code adheres to this project's style guidelines
• I have performed a self-review of my own code
• My changes do not introduce new warnings
• I have written tests demonstrating that my changes are effective or that my feature works
• Local unit tests pass with my changes

[danny-avila]

there's no need to update this file as it's no longer used and will be removed soon.