fix(a2a): a2a-sdk 1.x compat + input/output capture

[SuhaniNagpal7]

Summary

A2AInstrumentor silently no-op'd on a2a-sdk 1.x — instrument() returned cleanly with a single WARNING log, no methods were wrapped, every A2A call produced zero spans. The SDK restructured its client surface in 1.0: the class was renamed A2AClient → Client, the two send methods collapsed into a single async-streaming Client.send_message, and request/response types moved to protobuf. The wrapper's _get_client_class couldn't find A2AClient anywhere, hit its "SDK structure may have changed" warning path, and bailed.

This PR (a) makes the instrumentor work on both v0 and v1, and (b) fills the FI dashboard's Input/Output panels which were never populated even on v0. Verified end-to-end against a2a-sdk 1.0.2.

What was broken on a2a-sdk 1.x

Concern	a2a-sdk 0.x	a2a-sdk 1.x	Effect on the unfixed wrapper
Client class	A2AClient	Client	_get_client_class returns None → warning, no patching
Send methods	send_task + send_task_streaming	unified send_message (always async, always streaming)	Even if class found, neither attribute exists → silent skip
Request type	dict payload	protobuf SendMessageRequest	_extract_payload's dict.get(\"payload\") finds nothing
Stream events	direct event.artifact / event.status	oneof of task / message / status_update / artifact_update	_extract_event_state / artifact access all return None
Part.type	discriminator string	gone — type encoded by populated field	get_artifact_type always returns None
Header injection	kwargs[\"headers\"] passes through to httpx	Client.send_message has no headers kwarg	Would TypeError if we forced it
Sync/async	sync	async def returning AsyncIterator	Existing sync __call__ would return the coroutine without awaiting; the dead __call_async__ was never reached by wrapt

What this patch changes

All three files in traceai_a2a/:

__init__.py

• _resolve_client_class returns (class, api_version) and tries a2a.client.Client first, falls back to a2a.client.A2AClient, then to getattr(a2a_module, \"A2AClient\", None).
• _methods_for_api(api_version) returns [(\"send_message\", True)] on v1 and [(\"send_task\", False), (\"send_task_streaming\", True)] on v0.
• _instrument / _uninstrument iterate that list instead of hard-coding the two v0 methods.
• The _a2a_streaming marker is still attached to each wrapped function for symmetry — always True on v1.

_a2a_client.py

• A2AClientWrapper.__init__ takes an api_version (default v0 for backward compat with tests that don't pass it).
• __call__ dispatches to _sync_call or _async_call based on inspect.iscoroutinefunction(wrapped). Removed the dead __call_async__ method.
• _async_call awaits the wrapped coroutine, wraps the returned AsyncIterator with _AsyncStreamingSpanWrapper, handles the streaming-only v1 path.
• W3C header injection is skipped on v1 (no headers kwarg). The W3C trace-id is still recorded locally as gen_ai.a2a.propagated_trace_id.
• Span name is Client.send_message on v1, unchanged on v0.
• New _extract_event_state walks the v1 oneof (event.status_update.status → event.task.status → fallback to v0's event.status).
• New _extract_event_artifact walks the v1 oneof (event.artifact_update.artifact → event.task.artifacts[-1] → fallback to v0's event.artifact).
• Both streaming wrappers now accumulate text from each artifact into _output_fragments, joined and emitted as output.value in _finalize_span.

_attributes.py

• New text helpers: _part_text, extract_parts_text, extract_message_text, extract_artifact_text (handle both v0 and v1 part shapes).
• New get_send_message_request_attributes extracts task_id / role / parts.count / input text from a protobuf SendMessageRequest.
• get_send_task_payload_attributes (v0) now also emits input text from payload[\"message\"] parts.
• get_task_attributes (v0 non-streaming) now also emits output text from task.artifacts[*].parts.
• get_artifact_type falls back to inferring \"text\" / \"file\" / \"data\" from the populated v1 Part field when the v0 .type discriminator is absent.

What's NOT changed (out of scope)

• Cross-wire W3C propagation on v1. The trace-id is recorded locally, but the outbound request doesn't carry traceparent because there's no headers kwarg. The proper hook is ClientCallInterceptor — separate PR.
• Server-side _a2a_server.py. Not audited or changed. If the inbound A2A server shape also moved in 1.x, it's a separate fix.
• Protobuf enum readability. gen_ai.a2a.task.state = 3 and gen_ai.a2a.message.role = 1 surface as integers; readable names would need a pb.TaskState.Name(value) translation in _enum_or_str. Cosmetic.
• pyproject — no version pin tightening. The [tool.poetry.extras] a2a = [\"a2a-sdk\"] is unpinned and the package now works on both lines, so left alone.

Test plan

• python -m py_compile on all three changed files — syntax clean.
• Fresh Python 3.12 venv with a2a-sdk 1.0.2, fi-instrumentation-otel, and the local traceai-a2a installed editable.
• Before fix: A2AInstrumentor().instrument() logs WARNING traceai_a2a: Could not locate A2AClient class in a2a module; Client.send_message.__wrapped__ is False; zero spans on any call.
• After fix: instrument() runs cleanly; Client.send_message.__wrapped__ is True; _a2a_streaming = True.
• Smoke test: FakeClient(Client) subclass returning a canned AsyncIterator of StreamResponse (status_update WORKING → artifact_update with a text Part → status_update COMPLETED). All three events consumed, span flushed.
• Span in Future AGI Observe (traceai-a2a-smoke project) shows:
  • status OK
  • gen_ai.span.kind = \"A2A_CLIENT\"
  • gen_ai.a2a.agent.url = \"http://weather-agent.local\"
  • gen_ai.a2a.streaming = true
  • gen_ai.a2a.task.id = \"task-smoke-1\" (extracted from protobuf SendMessageRequest.message.task_id)
  • gen_ai.a2a.message.role = 1 (ROLE_USER)
  • gen_ai.a2a.message.parts.count = 1
  • gen_ai.a2a.task.state = 3 (TASK_STATE_COMPLETED)
  • gen_ai.a2a.artifact.type = \"text\" (inferred from populated Part.text field)
  • gen_ai.a2a.propagated_trace_id matches the trace's own W3C id
  • Input panel populated with the user prompt text (from input.value)
  • Output panel populated with the accumulated artifact text (from output.value)

Files changed

• python/frameworks/a2a/traceai_a2a/__init__.py — 78 / 57 ins/del
• python/frameworks/a2a/traceai_a2a/_a2a_client.py — 196 / 34 ins/del
• python/frameworks/a2a/traceai_a2a/_attributes.py — 137 / 8 ins/del