A fleet-wide sweep for sync LLM .create() calls running directly inside an
async def on the uvicorn event loop found three more wedge sites (same class
as the claude-sdk-python fix in this PR):
- integrations/ag2/src/agents/beautiful_chat.py
- integrations/llamaindex/src/agents/a2ui_dynamic.py
- integrations/llamaindex/src/agents/agent.py (missed by the original report)
Each extracts the blocking secondary-LLM round-trip into a sync _generate_a2ui
helper and offloads it via await asyncio.to_thread(...) from the async
generate_a2ui wrapper (lowest blast radius; sync body unchanged). ag2's other
agents already use AsyncOpenAI; all other sync .create sites are inside plain
def framework tools dispatched off-loop by their frameworks, so they do not
wedge. entrypoint.sh alert-scoping left untouched (claude-sdk-python-specific).
Adds a dev-only OpenAI-SDK repro harness (slow_openai.py, prod_server_openai.py,
run_prod_openai.sh) that drives the REAL production _generate_a2ui via a slow
local OpenAI-compatible endpoint, with a tool_dispatch_fired>=1 anti-false-green
guard. RED (sync-on-loop) -> GREEN (to_thread) verified for all three sites.
The claude-sdk-python agent :8000 wedges under D6/LLM load: two synchronous
anthropic.Anthropic().messages.create() calls run directly on the uvicorn
asyncio event loop, freezing it for the full LLM round-trip so /health stops
responding. The watchdog counts 3 consecutive failures (~90s) and kill-restarts
the container, dropping active sessions.
Root cause (sync-in-async), all in integrations/claude-sdk-python/:
- src/agents/agent.py: _execute_tool's generate_a2ui branch builds a sync
anthropic.Anthropic() and calls messages.create() synchronously; invoked on
the loop from run_agent's agentic loop AND from the Claude-Agent-SDK MCP tool
handler in claude_agent_sdk_adapter.py.
- src/agents/a2ui_dynamic.py: _generate_a2ui, same sync pattern, invoked on the
loop from the run_a2ui_dynamic_agent generator.
Fix: wrap every async call site in `await asyncio.to_thread(...)` (lowest blast
radius — the sync functions and the shared ExecuteTool callback type are
unchanged, and the whole tool-dispatch path is fixed uniformly, not just
generate_a2ui):
- agent.py run_agent call site
- claude_agent_sdk_adapter.py MCP tool handler
- a2ui_dynamic.py secondary call site
Blast radius: claude-sdk-python only. a2ui_dynamic.py is per-integration (each
framework has its own copy); every other claude-sdk-python agent already uses
AsyncAnthropic. The `tools` symlink to shared/python was not touched.
entrypoint.sh: drop the Slack alert from the :8000 agent watchdog branch (keep
the kill-restart — it self-heals silently now that the root cause is fixed);
keep the LOUD #oss-alerts page on the public $PORT /api/health branch.
Adds showcase/tests/repro/async-wedge/ — a faithful RED/GREEN harness driving
the real anthropic sync client against a controllable slow endpoint, plus a
mutation guard on the real _generate_a2ui.
Faithful node:22-slim repro: fd1 through the same awk process-substitution as entrypoint.sh, a Railway-capped drain reader, a uvicorn+CVDIAG-shaped flood, and the static no-log /api/health as victim. RED wedges (200->502, CPU->0, heartbeat frozen); the FIXED lane stays 200 throughout. run.sh asserts the outcome (exit 3/4/5 on a false result, proven). watchdog.sh runs the entrypoint public-guard loop verbatim and needle-anchors it against entrypoint.sh.
github-actions[bot]