v3.2.0a0

Pre-release: migrate to llm-rosetta 0.7.x ConversionPipeline.

Install: pip install argo-proxy==3.2.0a0 or pip install --pre argo-proxy

Changed

• Adopted llm-rosetta ConversionPipeline (#135): Replaced manual converter instantiation and IR round-trips with the declarative ConversionPipeline API from llm-rosetta 0.7.x. Pipeline construction is centralized in _build_pipeline() and reused across streaming, non-streaming, and retry paths
• Bumped llm-rosetta to >=0.7.0a0: Breaking change from 0.6.x — the new ConversionPipeline API replaces the previous manual converter + shim workflow
• Threaded provider types explicitly: Pipeline internals are no longer accessed directly; provider types are passed through the dispatch layer

Fixed

• Reuse pipeline in retry: Retry logic now reuses the existing ConversionPipeline instance instead of rebuilding it from scratch, avoiding redundant setup

Added

• Multi-entry E2E tests: New test_multi_entry.py covering Anthropic Messages, Google GenAI, and OpenAI Responses as input formats, with cross-format and same-format upstream routing (9 test paths)
• agentabi E2E tests: test_agentabi.py covering all 29 upstream models via OpenAI Chat entry with codex CLI

Removed

• Dead _sanitize_tool_schemas: Removed unused _sanitize_tool_schemas function and stale sanitize_schema import

Full Changelog: v3.1.2...v3.2.0a0