v0.20.4

v0.20.4

AgentCrew v0.20.4 — this release focuses on the A2A protocol server: cross-agent conversation continuity with tenant-scoped sessions, per-turn streaming artifacts, recovery of orphaned tasks after crashes/restarts, and a significant performance boost for long-running streaming tasks. It also introduces truly async agent activation so MCP discovery runs natively on the caller's event loop.

What's Changed

🚀 Features

• Agent-agnostic sessions keyed by tenant (7dea77e4) — Conversation history is now shared across agents and keyed by {owner}:{context_id}, so a conversation can continue seamlessly when switching between agents. Task and pending-tool state stays isolated per agent namespace + owner.

• Truly async agent activation (a865214b) — LocalAgent.activate_async now replicates the full sync activation path and runs MCP discovery as a non-blocking background task on the caller's event loop instead of spawning a thread with asyncio.run(). Also fixes cross-event-loop OAuth lock binding by scoping asyncio.Locks per (server, event loop).

🐛 Bug Fixes

• Orphaned A2A tasks no longer hang on resubscribe (82568a6f) — Resubscribing to a persisted non-terminal task with no live producer (server restart or an execution that died mid-stream) previously hung on an idle SDK producer. The new AgentCrewRequestHandlerV2 terminates such resubscribes by marking the orphaned task FAILED, and startup reconciliation marks stale tasks FAILED after a restart.

• Performance fix for long-running streaming tasks (c8793cc7) — Token chunks are now buffered and flushed every 100ms as one batched artifact event, so SDK task-store persistence runs once per batch instead of once per token (the cause of A2A streaming being slower than the interactive UI). The file-backed task store also switched to JSONL appends with atomic compaction — steady-state saves no longer reread the whole file, and torn trailing lines from crashes are tolerated.

🧹 Refactors & Cleanup

• Per-turn artifacts for every LLM round-trip (bf896114) — Each LLM/tool round now streams into its own per-turn artifact (answer_{task_id}_turn_{n}, plus per-turn thinking/tool artifacts), giving clients a clean view of multi-turn tool workflows.

• Separated concerns in the A2A executor (3237a55d) — The executor was split into dedicated modules: artifact_stream.py (streaming state + chunk buffer), attachment_processor.py (file/URI attachment preprocessing), and tool_executor.py (tool-call orchestration).

• Cleanup and formatting (9c757943)

Thanks for the continued contributions, bug reports, and feedback! 🎉