fix(chat): do not retry if we had chatdeltas or tooldeltas from backend

[mudler]

Description

This pull request addresses a critical bug in the retry logic for chat completions involving the C++ autoparser, ensuring that valid tool calls and content delivered via ChatDeltas are not discarded or retried unnecessarily. It introduces new regression tests, updates the backend mock to simulate autoparser behavior, and enhances the retry logic to properly recognize actionable ChatDelta content. This prevents excessive retries and output concatenation, improving both streaming and non-streaming chat endpoint behavior.

Bug fix: Retry logic for autoparser ChatDeltas

• Enhanced the chat completion retry logic to check for actionable tool calls and content in ChatDeltas, preventing unnecessary retries when the C++ autoparser delivers results via ChatDeltas with an empty raw message. This fixes a bug where valid tool calls/content were ignored, causing up to 6 retries and concatenated output. [1] [2] [3] [4]

Testing: Regression tests for autoparser scenarios

• Added comprehensive end-to-end tests that verify streaming and non-streaming chat completions correctly process tool calls and content from ChatDeltas, and do not retry or concatenate output unnecessarily.
• Created a dedicated model config for autoparser tests, disabling grammar enforcement to ensure tool calls are driven by backend ChatDeltas.

Test infrastructure: Mock backend improvements

• Updated the mock backend to simulate C++ autoparser behavior by emitting tool calls and content via ChatDeltas with an empty message, both for streaming and non-streaming endpoints. [1] [2]

Test utilities

• Added necessary imports (e.g., encoding/json) to support new test logic for parsing streaming responses.

Notes for Reviewers

Signed commits

• Yes, I signed my commits.