fix: restore _current_span in scheduler to fix task failures (issue #…

[NexionisJake]

Commit 1cc2a61 changed _TaskOperation to carry trace_id/span_id strings but the in-memory path never restored the
span in the worker, silently dropping trace context and causing every task to fail when the worker tried to access
_current_span.

Changes:

• base.py: revert _TaskOperation fields from trace_id/span_id back to _current_span: Span; import Span from
  opentelemetry.trace; update docstring to clarify Redis divergence
• memory_scheduler.py: update _send_operation() to pass _current_span=get_current_span() instead of
  trace_id/span_id; remove unused math import; replace math.inf buffer with 100 to restore backpressure; drop
  get_trace_context import

workers/base.py already handles the None-span case via nullcontext() so no change is needed there. The Redis
scheduler correctly serialises trace_id/span_id strings for cross-process transport (CASE 2).

Summary

• Problem: 1cc2a61 refactored _TaskOperation to use serialisable trace_id/span_id strings, but never updated
  InMemoryScheduler to pass a live span — so _current_span was never set in the task dict, the worker received None,
  and every task was silently marked failed
• Why it matters: 100% of tasks submitted via the A2A protocol failed immediately with no artifacts produced; the
  broad except block in the worker swallowed the error with no visible signal
• What changed: _TaskOperation carries _current_span: Span again; _send_operation() in InMemoryScheduler now
  calls get_current_span() directly; math.inf buffer replaced with 100 to prevent unbounded memory growth under
  backpressure
• What did NOT change (scope boundary): RedisScheduler is untouched — it legitimately needs trace_id/span_id
  strings because live Span objects cannot be JSON-serialised across processes; workers/base.py is untouched — it
  already guards against a missing span with nullcontext()

Change Type (select all that apply)

• Bug fix
• Feature
• Refactor
• Documentation
• Security hardening
• Tests
• Chore/infra

Scope (select all touched areas)

• Server / API endpoints
• Extensions (DID, x402, etc.)
• Storage backends
• Scheduler backends
• Observability / monitoring
• Authentication / authorization
• CLI / utilities
• Tests
• Documentation
• CI/CD / infra

Linked Issue/PR

• Closes [Bug]: Document Processing Failures #353

User-Visible / Behavior Changes

Tasks submitted via the A2A message/send endpoint now complete successfully and return artifacts. Previously every task
immediately transitioned to "failed" state with no artifacts.

Security Impact (required)

• New permissions/capabilities? No
• Secrets/credentials handling changed? No
• New/changed network calls? No
• Database schema/migration changes? No
• Authentication/authorization changes? No
• If any Yes, explain risk + mitigation: N/A

Verification

Environment

• OS: Linux 6.17.0-19-generic
• Python version: 3.12.3
• Storage backend: InMemoryStorage
• Scheduler backend: InMemoryScheduler

Steps to Test

1. Start the server: python examples/beginner/echo_simple_agent.py
2. Submit a task: curl -X POST http://localhost:3773/ -H 'Content-Type: application/json' -d '{"jsonrpc":"2.0","id":"<uu id>","method":"message/send","params":{"configuration":{"acceptedOutputModes":["text"]},"message":{"messageId":"<uuid>"," contextId":"<uuid>","taskId":"<uuid>","kind":"message","role":"user","parts":[{"kind":"text","text":"Hello"}]}}}'
3. Poll the task: curl -X POST http://localhost:3773/ -H 'Content-Type: application/json' -d '{"jsonrpc":"2.0","id":"<uuid>","method":"tasks/get","params":{"taskId":"<task-uuid>"}}'

Expected Behavior

• Task transitions: submitted → working → completed
• Response includes a non-empty artifacts array

Actual Behavior

• Task transitions immediately to failed
• No artifacts field in response
• Worker logs a silent KeyError on _current_span swallowed by the broad except block

Evidence (attach at least one)

• Failing test before + passing after
• Test output / logs
• Screenshot / recording
• Performance metrics (if relevant)

Full suite after fix: 666 passed, 18 skipped, 0 failures