[pradystar]

Summary

Fixes SAO-19050.

CrewAI dispatches synchronous event listeners through a thread pool, allowing completion callbacks to execute before their matching starts. This caused dropped spans and incomplete hierarchies.

• Use CrewAI’s async listener path without yielding during lifecycle updates.
• Isolate handler state and SDK/OTel context per crew invocation.
• Match LLM parents to the agent run IDs used by older CrewAI versions.
• Clean up state after completion, failure, or malformed callbacks.

Runtime changes are limited to the CrewAI handler. No dependency, base-handler, logger, decorator, or native OTel changes.

Validation

• 415 tests passed, 1 skipped across CrewAI and compatibility suites.
• Ruff lint/format and configured type checks passed.
• Deterministic sync/async examples passed on CrewAI 1.6.1 and 1.15.28.
• Each example produced 21 spans across 3 traces with correct hierarchy and clean context.
• All 84 recorded spans included gen_ai.operation.name.

[etserend]

🤖 This review was generated by the Astra agent (claude-opus-5-5). It may contain mistakes.

Findings below were re-validated by reproducing them against lab0 (CrewAI 1.6.1) on this PR's head 8f4ce0c and its base.

Verdict: request_changes — Overlapping crews now get separate roots, but the change introduces four regressions and leaves one ticket criterion unmet.

Findings

• 🟠 major (bug): Child error tears down the whole crew. See the inline comment on handler.py:190.
• 🟠 major (bug): LiteLLM token usage is lost for overlapping crews. Only the first crew's chat span gets usage (10/5/15). The second crew's has none. lite_llm_usage_callback (handler.py:881-898) looks nodes up only in self._handler, but later crews use their own handlers. Look the node up across run.handler for every run in _crew_runs.
• 🟡 minor (bug): Events outside any crew are dropped. See the inline comment on handler.py:171.
• 🟡 minor (performance): The root flush blocks CrewAI's shared events loop. See the inline comment on handler.py:128.
• 🟡 minor (question): Original event timestamps are ignored. SAO-19050 requires "original event timestamps". Events stamped 10 minutes earlier produce spans that start at run time (+600 s) and last under 1 s instead of 8 s. event.timestamp goes only into metadata (handler.py:462), while created_at and duration_ns come from callback time (base_handler.py:543-544, :616). The base behaves the same, so this is not a regression. Is it in scope? If so, pass created_at=event.timestamp on start and derive the duration from the end event's timestamp.
• 🟡 minor (testing): The ticket asks for deterministic tests of reversed lifecycle delivery, parent/child reordering, duplicates, failure and cleanup. Failure and cleanup are covered. No test sends a duplicated start or end event, and the reversed-delivery branch in the crewai_bus fixture never runs (see the inline comment on tests/test_crewai_handler.py:158). Caller-context inheritance across the CrewAI loop thread works (confirmed on lab0), but no test locks it in. Suggested additions:
  1. Run a real asyncio loop on a separate thread, emit with asyncio.run_coroutine_threadsafe from the test thread as CrewAI does, and assert that caller OTel and logger context are inherited.
  2. Add one test with a duplicated TaskStartedEvent and one with a duplicated TaskCompletedEvent. Assert that each span appears exactly once and that no state remains.

[pradystar]

Addressed all focused comments and follow-ups, with regression coverage

[etserend]

🤖 This review was generated by the Astra agent (claude-opus-5-5). It may contain mistakes.

Verdict: request_changes — A major data-loss race from an earlier local round is still present at HEAD 254fb2d: ingestion-hook flushes can drop another crew's trace.

Addressed before this review

• The deferred async_flush() in ingestion-hook batch mode could drop another crew's open trace. _flush_batch yields during the hook, then clears self.traces. Not addressed: HEAD is still 254fb2d, the commit the concern was raised against. It is raised again in this review.