refactor(weave): migrate Claude Agent SDK tracing to GenAI SDK

AGENTS.md

@@ -439,6 +439,13 @@ pnpm exec tsx examples/claudeAgents.ts
 - `Turn.messages` stores input messages, while `Turn.output_messages` stores
   the terminal agent response. `Turn.record(messages=..., output_messages=...)`
   replaces the two lists independently.
+- Mypy requires explicit `return None` paths in functions annotated with
+  `T | None`; bare `return` and implicit fallthrough trigger return-value
+  errors.
+- Python `Turn` spans are same-process `invoke_agent` operations and use OTel's
+  default `SpanKind.INTERNAL`. Do not set `gen_ai.provider.name` on a `Turn`;
+  set it on child `LLM`/`chat` spans, because one turn can use multiple
+  providers.
 - Keep streaming and batch paths aligned: `Turn._build_attrs()` must apply
   content gating and PII redaction to both lists before passing them to
   `invoke_agent_attributes()`, and `log_turn()` must accept both fields.
@@ -457,6 +464,15 @@ pnpm exec tsx examples/claudeAgents.ts
   rollups.
 - Regression coverage must exercise both the calls-based and OTel integrations
   with nonzero cache-read and cache-creation counts.
+- The OTel-selected Claude Agent SDK path composes the Python GenAI
+  `Conversation`, `Turn`, `LLM`, and `Tool` handles; it must not create raw
+  OTel spans or call the low-level GenAI attribute builders itself. Use
+  `set_attributes()` only for semantic fields the typed handles do not expose.
+  The legacy calls-based path remains separate.
+- Stream adapters can create child handles when work starts, then enter them
+  with a normal `with` when the completion message arrives. Preserve logical
+  timing with `LLM.started_at` and an explicit `Tool.started_at` instead of
+  keeping contexts open with `ExitStack`.
 
 ### Credential-shaped fields in call inputs and attributes
 

tests/integrations/claude_agent_sdk/claude_agent_sdk_otel_test.py

@@ -19,13 +19,14 @@
 from opentelemetry.sdk.trace import TracerProvider as SDKTracerProvider
 from opentelemetry.sdk.trace.export import SimpleSpanProcessor
 from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
-from opentelemetry.trace import StatusCode
+from opentelemetry.trace import SpanKind, StatusCode
 
 from tests.integrations.claude_agent_sdk.conftest import ReplayTransport, load_cassette
 from weave.conversation import agent_name_override
 from weave.integrations.claude_agent_sdk.otel_integration import (
     get_claude_agent_sdk_otel_patcher,
 )
+from weave.trace.settings import override_settings
 
 
 @pytest.fixture
@@ -55,6 +56,16 @@ def patch_claude_agent_sdk_otel() -> Generator[None]:
     patcher.undo_patch()
 
 
+@pytest.fixture(autouse=True)
+def disable_capture_info() -> Generator[None]:
+    """Keep exact span payload assertions independent of host metadata."""
+    with override_settings(
+        capture_client_info=False,
+        capture_system_info=False,
+    ):
+        yield
+
+
 # --- helpers ----------------------------------------------------------------
 
 
@@ -65,7 +76,7 @@ def get_attrs(span: Any) -> dict[str, Any]:
 def check_integration_and_strip(attrs: dict[str, Any]) -> dict[str, Any]:
     """Assert + remove the flattened integration.* provenance keys.
 
-    The agent OTel processor stamps integration provenance on every span; pop it
+    Conversation attributes stamp integration provenance on every span; pop it
     here so the exact-shape assertions below stay focused on the GenAI semconv keys.
     """
     assert attrs["integration.name"] == "claude_agent_sdk"
@@ -116,6 +127,7 @@ async def run_query(cassette: str, prompt: str) -> None:
 async def test_simple_text_query_otel(otel_spans: InMemorySpanExporter) -> None:
     await run_query("simple_text_response", "What is 2+2?")
     spans = otel_spans.get_finished_spans()
+    assert {span.instrumentation_scope.name for span in spans} == {"weave.conversation"}
 
     agent_spans = get_spans_by_op(spans, "invoke_agent")
     chat_spans = get_spans_by_op(spans, "chat")
@@ -126,10 +138,10 @@ async def test_simple_text_query_otel(otel_spans: InMemorySpanExporter) -> None:
     # glance and any spec drift (added/removed/renamed keys) fails the test.
     agent_span = agent_spans[0]
     assert agent_span.name == "invoke_agent claude_agent_sdk"
+    assert agent_span.kind == SpanKind.INTERNAL
     assert check_integration_and_strip(get_attrs(agent_span)) == {
         "gen_ai.operation.name": "invoke_agent",
         "gen_ai.agent.name": "claude_agent_sdk",
-        "gen_ai.provider.name": "anthropic",
         "gen_ai.conversation.id": "s-abc123",
         "gen_ai.request.model": "claude-sonnet-4-6",
         "gen_ai.input.messages": (
@@ -358,8 +370,7 @@ async def test_custom_agent_name_query_otel(otel_spans: InMemorySpanExporter) ->
     assert agent_span.name == "invoke_agent research_agent"
     attrs = check_integration_and_strip(get_attrs(agent_span))
     assert attrs["gen_ai.agent.name"] == "research_agent"
-    # Only the agent name is overridden — the rest of the GenAI shape is intact.
-    assert attrs["gen_ai.provider.name"] == "anthropic"
+    # Only the agent name is overridden — the model remains intact.
     assert attrs["gen_ai.request.model"] == "claude-sonnet-4-6"