fix(openai-agents): use synchronous spans to fix OTel context detachment

[harsh543]

Problem

PR #1197 attempts to fix OpenAI Agents tracing but fails with OTel providers (Langfuse, Logfire) due to context detachment errors:

ValueError: <Token> was created in a different Context

The issue: OTel validates token detachment happens in the exact same contextvars.Context instance where the token was created. Even copy_context() creates a new instance with equivalent values, causing the check to fail.

Root Cause

Activities and child workflows complete in:

• Different machines (distributed execution)
• Different workflow tasks (different context instances)
• Different callback invocations (new context per task)

Callbacks registered with copied context still execute in a new context instance, breaking OTel's validation.

Solution

Follow the pattern used in temporalio/contrib/opentelemetry.py - complete spans synchronously at the orchestration point:

def start_activity(self, input: StartActivityInput) -> ActivityHandle:
    trace = get_trace_provider().get_current_trace()
    if trace:
        with custom_span(name="temporal:startActivity", data={"activity": input.activity}):
            pass  # Completes immediately in same context
    
    set_header_from_context(input, temporalio.workflow.payload_converter())
    return self.next.start_activity(input)

Benefits

1. OTel Compatible: No context detachment errors - span finishes in same context it started
2. Distributed Execution: No callback needed = works across machines
3. Simpler: Removes add_done_callback machinery, cleaner code (-26 lines, +19 lines)
4. Correct Semantics: Zero-duration orchestration span (CLIENT) + real-duration execution span (SERVER)
5. Replay Safe: No long-running spans across workflow tasks

Span Hierarchy

temporal:startActivity (0ns, workflow) ← Orchestration decision
  └─ temporal:executeActivity (2.5s, activity worker) ← Actual work

The zero-duration span is correct - it represents when the orchestration decision was made, not when the distributed work completed.

Testing

• All existing tests should pass
• No ValueError with default provider
• Follows existing pattern from temporalio/contrib/opentelemetry.py

References

• Fixes [Bug] Langfuse Tracing Not Working with Temporal OpenAI Agents Plugin #1136 (Langfuse integration)
• Alternative to OpenAI Agents Tracing: Run span start/finish in the same context #1197
• Pattern matches: temporalio/contrib/opentelemetry.py

---

cc @cretz - This implements the synchronous span approach you suggested in the PR #1197 review. Let me know if you'd like any changes!

🤖 Generated with Claude Code

[tconley1428]

I don't think this is the approach we want to take. I'm working on a PR at #1286 which makes more substantial changes to fix the scenario. The problem statement here

PR #1197 attempts to fix OpenAI Agents tracing but fails with OTel providers (Langfuse, Logfire) due to context detachment errors:

Is also very incomplete given our later investigations. There are a number of additional problems with running otel simultaneous to the agents sdk logging in Temporal beyond the context warning.

[harsh543]

Good catch — I overcomplicated this. Reverted to a single test with clearer structure:

# --- First execution: logs should appear ---
...
# --- Clear logs and continue execution (replay path) ---
# When the new worker starts, it replays the workflow history (signals 1 & 2).
# Replay suppression should prevent those logs from appearing again.
capturer.log_queue.queue.clear()
...
# --- Replay execution: no duplicate logs ---
assert not capturer.find_log("Signal: signal 1")

Also wrapped the global state mutation in try/finally:

original = workflow.logger.full_workflow_info_on_extra
workflow.logger.full_workflow_info_on_extra = True
try:
    ...
finally:
    workflow.logger.full_workflow_info_on_extra = original

No new test, just better visibility into what the existing one validates.

[harsh543]

Changes to test_workflow_logging

Addressed feedback — kept a single test with improved structure:

1. Added try/finally for test hygiene

original_full_workflow_info_on_extra = workflow.logger.full_workflow_info_on_extra
workflow.logger.full_workflow_info_on_extra = True
try:
    ...
finally:
    workflow.logger.full_workflow_info_on_extra = original_full_workflow_info_on_extra

2. Added clear phase markers

• # --- First execution: logs should appear ---
• # --- Clear logs and continue execution (replay path) ---
• # --- Replay execution: no duplicate logs ---

3. Added explanatory note for replay

# When the new worker starts, it replays the workflow history (signals 1 & 2).
# Replay suppression should prevent those logs from appearing again.

No separate replay test — visibility > duplication.