Skip to content

Use Gemini's default temperature and preserve empty-response errors #1211

Description

@ebhills

Gemini URL retrieval currently sends temperature: 0.1. A development run with a minimal prompt returned HTTP 200 with finishReason: TOO_MANY_TOOL_CALLS and content: {}. Other requests timed out during prefill. Google recommends its default temperature of 1.0 for Gemini 3 because lower settings may cause looping; causation for these failures is not yet confirmed.

The SDK parses content: {} as a truthy content object with parts=None. The client then tries to iterate the absent parts, masking the provider's finish reason with an unexpected error.

Scope:

  • Remove the packaged Gemini temperature override, preserving explicit configuration overrides.
  • Treat missing or empty content as failed generation and retain its finish reason/message, even if URL retrieval succeeded.
  • Add focused offline regressions and update the configuration documentation.

No recipe API, retry, timeout, or concurrency changes. After merge and package promotion to the development recipe Lambda, repeat the same URL and prompt to test the temperature hypothesis. Eric will handle the Lambda dev deployment.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

No labels
No labels

Type

No type

Fields

Stage

None yet

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions