Reliability Hardening: Fixed tool_use conversion, streaming bridge, and upstream retries

[swe-squad-alpha]

Overview

This issue documents the root cause analysis and resolution of several critical reliability and performance issues discovered during development.

Issues Resolved

1. Tool Use Conversion Gate (ISSUE-004)

• Problem: Many OpenAI-compatible models (qwen, deepseek, minimax) were returning tool calls as plain text or were being suppressed by a model-prefix gate.
• Fix: Removed the model-prefix gate in convert_litellm_to_anthropic and added a parser fallback for manual tool calls in content_text.
• Result: Models like deepseek-v3.1 now correctly produce structured tool_use blocks.

2. Streaming Response RuntimeError

• Problem: Agent would 'stall' or connections would drop during streaming.
• Root Cause: HTTPException was being raised after the first SSE blocks were yielded.
• Fix: Implemented a response_started flag. Any errors occurring after headers are sent are now yielded as standard Anthropic SSE error events.

3. Upstream Stalls (502/504 Errors)

• Problem: Transient 502/504 errors from upstream providers were not being handled gracefully.
• Fix: Implemented retry_with_backoff for both non-streaming and the initial phase of streaming requests.

4. Streaming Lifecycle Fix

• Problem: AsyncClient created outside the generator lifecycle was being closed prematurely.
• Fix: Moved AsyncClient lifecycle inside the stream generator.

Status

All fixes have been implemented and verified with complex tool-use tests. We have prepared a PR with these changes.