fix(claude-agent-sdk): test against 0.1.18 after older releases left PyPI

py/noxfile.py

@@ -411,16 +411,17 @@ def test_litellm(session, version):
     _run_tests(session, f"{INTEGRATION_DIR}/litellm/test_litellm.py", version=version)
 
 
-# CLI bundling started in 0.1.10 - older versions require external Claude Code installation
+# CLI bundling started in 0.1.10 - older versions require external Claude Code installation.
+# 0.1.9-0.1.17 were deleted from PyPI, so the oldest matrix entry is 0.1.18.
 CLAUDE_AGENT_SDK_VERSIONS = _get_matrix_versions("claude-agent-sdk")
 
 
 @nox.session()
 @nox.parametrize("version", CLAUDE_AGENT_SDK_VERSIONS, ids=CLAUDE_AGENT_SDK_VERSIONS)
 def test_claude_agent_sdk(session, version):
     _install_test_deps(session)
-    # Claude Agent SDK 0.1.10 calls Server.list_tools(), which MCP 2 removed.
-    constraint_group = "test-mcp-v1" if version == "0.1.10" else None
+    # Claude Agent SDK 0.1.18 calls Server.list_tools(), which MCP 2 removed.
+    constraint_group = "test-mcp-v1" if version == "0.1.18" else None
     _install_matrix_dep(session, "claude-agent-sdk", version, constraint_group)
     _run_tests(session, f"{INTEGRATION_DIR}/claude_agent_sdk/test_claude_agent_sdk.py", version=version)
 

py/pyproject.toml

@@ -515,7 +515,9 @@ latest = "pipecat-ai==1.12.0"
 
 [tool.braintrust.matrix.claude-agent-sdk]
 latest = "claude-agent-sdk==0.2.168"
-"0.1.10" = "claude-agent-sdk==0.1.10"
+# The integration still supports >=0.1.10, but 0.1.9-0.1.17 were deleted from
+# PyPI, so 0.1.18 is the oldest installable version we can test against.
+"0.1.18" = "claude-agent-sdk==0.1.18"
 
 [tool.braintrust.matrix.cursor-sdk]
 # 1.0.25 fixes bridge callback tokens that could be parsed as CLI flags while

py/scripts/session-weights.json

@@ -23,7 +23,7 @@
   "test_btx_openai(1.77.0)": 8,
   "test_btx_openai(1.92.0)": 7,
   "test_btx_openai(latest)": 11,
-  "test_claude_agent_sdk(0.1.10)": 10,
+  "test_claude_agent_sdk(0.1.18)": 10,
   "test_claude_agent_sdk(latest)": 10,
   "test_cli": 11,
   "test_cohere(5.0.0)": 15,

py/src/braintrust/integrations/claude_agent_sdk/_constants.py

@@ -50,6 +50,8 @@ class MCPToolMetadataValues:
 CLAUDE_AGENT_TASK_SPAN_NAME: Final[str] = "Claude Agent"
 CLAUDE_AGENT_RUN_FAILED_ERROR: Final[str] = "Claude Agent run failed"
 ANTHROPIC_MESSAGES_CREATE_SPAN_NAME: Final[str] = "anthropic.messages.create"
+# Bundled Claude Code CLIs before 2.1 (e.g. 2.0.72 in SDK 0.1.18) emit subagent calls as "Task".
+SUBAGENT_TOOL_NAMES: Final[frozenset[str]] = frozenset({"Agent", "Task"})
 
 MCP_TOOL_PREFIX: Final[str] = "mcp__"
 MCP_TOOL_NAME_DELIMITER: Final[str] = "__"

py/src/braintrust/integrations/claude_agent_sdk/_test_transport.py

@@ -5,6 +5,7 @@
 import json
 import os
 import re
+import warnings
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
 from urllib.parse import parse_qsl, urlencode, urlsplit, urlunsplit
@@ -43,6 +44,11 @@ def _resolve_cassettes_dir() -> Path:
 
 CASSETTES_DIR = _resolve_cassettes_dir()
 
+# Timeouts for test synchronization waits. Replay is effectively instant;
+# a live recording waits on real model calls.
+REPLAY_WAIT_TIMEOUT_S = 1.0
+RECORD_WAIT_TIMEOUT_S = 120.0
+
 
 def get_record_mode() -> str:
     """Match VCR-like defaults for subprocess transport cassettes."""
@@ -227,6 +233,10 @@ def _sanitize_url_string(value: str) -> str:
 )
 
 
+def _is_result_message(message: Any) -> bool:
+    return isinstance(message, dict) and message.get("type") == "result"
+
+
 def _is_mcp_tool_call(message: Any) -> bool:
     request = message.get("request") if isinstance(message, dict) else None
     mcp_message = request.get("message") if isinstance(request, dict) else None
@@ -334,6 +344,14 @@ async def close(self) -> None:
         if self._recording:
             assert self._delegate is not None
             await self._delegate.close()
+            if not any(event["op"] == "read" and _is_result_message(event["payload"]) for event in self._events):
+                # A run that fails or hangs mid-recording would otherwise leave a
+                # truncated cassette that "once" mode replays from then on.
+                warnings.warn(
+                    f"Not saving cassette {self._cassette_name}: the recording has no result message",
+                    stacklevel=2,
+                )
+                return
             payload = {
                 "sdk_version": getattr(claude_agent_sdk, "__version__", "unknown"),
                 "cassette_name": self._cassette_name,
@@ -347,6 +365,11 @@ async def close(self) -> None:
     def is_ready(self) -> bool:
         return self._ready
 
+    @property
+    def wait_timeout(self) -> float:
+        """Timeout for waits on this transport's gates, sized for replay or live recording."""
+        return REPLAY_WAIT_TIMEOUT_S if self._should_replay() else RECORD_WAIT_TIMEOUT_S
+
     async def wait_for_mcp_tool_call(self) -> None:
         """Wait until replay delivers a local MCP ``tools/call`` request."""
         await self._mcp_tool_call_read.wait()