Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions src/seclab_taskflow_agent/agent.py
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
from agents.run import DEFAULT_MAX_TURNS
from dotenv import find_dotenv, load_dotenv
from openai import AsyncOpenAI
import httpx

from .capi import get_AI_endpoint, get_AI_token, get_provider

Expand Down Expand Up @@ -182,6 +183,7 @@ def __init__(
base_url=resolved_endpoint,
api_key=resolved_token,
default_headers=provider.extra_headers or None,
timeout=httpx.Timeout(connect=10.0, read=300.0, write=300.0, pool=60.0),
)
set_tracing_disabled(True)
self.run_hooks = run_hooks or TaskRunHooks()
Expand All @@ -198,6 +200,7 @@ def _ToolsToFinalOutputFunction(
else:
model_impl = OpenAIChatCompletionsModel(model=model, openai_client=client)

self._openai_client = client
self.agent = Agent(
name=name,
instructions=instructions,
Expand All @@ -209,6 +212,11 @@ def _ToolsToFinalOutputFunction(
hooks=agent_hooks or TaskAgentHooks(),
)

async def close(self) -> None:
"""Close the underlying AsyncOpenAI client and its httpx connection pool."""
if self._openai_client is not None:
await self._openai_client.close()

async def run(self, prompt: str, max_turns: int = DEFAULT_MAX_TURNS) -> result.RunResult:
"""Run the agent to completion and return the result."""
return await Runner.run(starting_agent=self.agent, input=prompt, max_turns=max_turns, hooks=self.run_hooks)
Expand Down
101 changes: 98 additions & 3 deletions src/seclab_taskflow_agent/runner.py
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,9 @@
import json
import logging
import os
import sys
import threading
import time
import uuid
from typing import Any

Expand Down Expand Up @@ -51,6 +54,49 @@
MAX_API_RETRY = 5 # Maximum number of consecutive API error retries
TASK_RETRY_LIMIT = 3 # Maximum retry attempts for a failed task
TASK_RETRY_BACKOFF = 10 # Initial backoff in seconds between task retries
# Application-level backstop: kill a streaming run if no events yielded for 30 min.
# Complements the TCP-level httpx.Timeout(read=300s) in agent.py which catches
# dead sockets; this catches subtler hangs where the connection stays open but
# the server (or async generator) stops producing events.
STREAM_IDLE_TIMEOUT = 1800

# Watchdog: a non-asyncio thread that force-kills the process if the event
# loop stops making progress. Covers every hang variant (dead connections,
# asyncio cleanup spin, MCP cleanup, etc.) because it runs outside asyncio.
WATCHDOG_IDLE_TIMEOUT = int(os.environ.get("WATCHDOG_IDLE_TIMEOUT", "2100")) # 35 min default

_watchdog_last_activity = time.monotonic()
_watchdog_lock = threading.Lock()


def watchdog_ping() -> None:
"""Call from any coroutine/callback to signal the process is alive."""
global _watchdog_last_activity
with _watchdog_lock:
_watchdog_last_activity = time.monotonic()


def _watchdog_thread(timeout: int) -> None:
"""Background thread: force-exit if no activity for *timeout* seconds."""
check_interval = min(60, max(1, timeout // 5))
while True:
time.sleep(check_interval)
with _watchdog_lock:
idle = time.monotonic() - _watchdog_last_activity
if idle > timeout:
logging.error(
f"Watchdog: no activity for {idle:.0f}s (limit {timeout}s) — "
"force-exiting to prevent hang"
)
sys.stderr.flush()
sys.stdout.flush()
os._exit(2)


def start_watchdog(timeout: int = WATCHDOG_IDLE_TIMEOUT) -> None:
"""Start the watchdog thread (idempotent, daemon thread)."""
t = threading.Thread(target=_watchdog_thread, args=(timeout,), daemon=True)
t.start()
Comment on lines +96 to +99
Copy link

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

start_watchdog is documented as idempotent, but the current implementation always spawns a new daemon thread on every call. Also, _watchdog_last_activity is initialized at module import time and start_watchdog() does not reset it, so if this module is imported and start_watchdog() is invoked later than timeout seconds, the watchdog can force-exit almost immediately. Consider (1) tracking a module-level started flag/thread to make this truly idempotent, and (2) calling watchdog_ping() (or otherwise resetting the timestamp) inside start_watchdog() before starting the thread.

Copilot uses AI. Check for mistakes.


def _resolve_model_config(
Expand Down Expand Up @@ -367,6 +413,7 @@ async def deploy_task_agents(
server_prompts=server_prompts,
important_guidelines=important_guidelines,
)
agent0 = None
agent0 = TaskAgent(
name=primary_name,
instructions=prompt_with_handoff_instructions(system_prompt) if handoffs else system_prompt,
Expand All @@ -391,9 +438,32 @@ async def _run_streamed() -> None:
while rate_limit_backoff:
try:
result = agent0.run_streamed(prompt, max_turns=max_turns)
async for event in result.stream_events():
if event.type == "raw_response_event" and isinstance(event.data, ResponseTextDeltaEvent):
await render_model_output(event.data.delta, async_task=async_task, task_id=task_id)
stream = None
try:
stream = result.stream_events()
async_iter = stream.__aiter__()
while True:
try:
event = await asyncio.wait_for(
async_iter.__anext__(),
timeout=STREAM_IDLE_TIMEOUT,
)
except StopAsyncIteration:
break
except asyncio.TimeoutError:
logging.error(
f"Stream idle for {STREAM_IDLE_TIMEOUT}s — "
"connection likely dead, raising APITimeoutError"
)
raise APITimeoutError("Stream idle timeout exceeded")
if event.type == "raw_response_event" and isinstance(event.data, ResponseTextDeltaEvent):
watchdog_ping()
await render_model_output(event.data.delta, async_task=async_task, task_id=task_id)
finally:
if stream is not None:
aclose = getattr(stream, "aclose", None)
if aclose is not None:
await aclose()
Copy link

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The finally cleanup calls await aclose() directly. Any exception raised during aclose() will override the original streaming error/timeout and can break the retry path. Consider wrapping the aclose() call in a try/except (logging on failure) so cleanup failures don’t mask the root cause.

Suggested change
await aclose()
try:
await aclose()
except Exception:
logging.exception("Failed to close streamed response")

Copilot uses AI. Check for mistakes.
await render_model_output("\n\n", async_task=async_task, task_id=task_id)
return
except APITimeoutError:
Expand Down Expand Up @@ -433,6 +503,11 @@ async def _run_streamed() -> None:
return complete

finally:
# Close the AsyncOpenAI client to release httpx connection pool.
# Dead CLOSE_WAIT sockets in the pool cause kqueue CPU spin if left open.
watchdog_ping()
if agent0 is not None:
await agent0.close()
Comment on lines 505 to +510
Copy link

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This finally only closes the primary TaskAgent’s AsyncOpenAI client. Handoff agents are currently created via TaskAgent(...).agent, which keeps the underlying AsyncOpenAI client alive but makes it impossible to close() later, so their httpx connection pools can still leak and reintroduce the same CLOSE_WAIT/kqueue-spin behavior. Consider retaining the handoff TaskAgent wrappers (or their clients) and closing all of them here; also wrap agent0.close() in a try/except so a close failure doesn’t mask an earlier exception or prevent MCP cleanup.

Copilot uses AI. Check for mistakes.
Comment on lines +506 to +510
Copy link

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Only the primary TaskAgent’s AsyncOpenAI client is being closed here. Any additional handoff agents created earlier via TaskAgent(...).agent will still retain their own AsyncOpenAI/httpx pools (the wrapper instance is discarded, but the Agent keeps a reference to the client through its model). To fully prevent lingering CLOSE_WAIT sockets / event-loop spin, track and close all TaskAgent-owned clients (or share a single client across agents) in this finally block.

Suggested change
# Close the AsyncOpenAI client to release httpx connection pool.
# Dead CLOSE_WAIT sockets in the pool cause kqueue CPU spin if left open.
if agent0 is not None:
await agent0.close()
# Close all AsyncOpenAI clients reachable from the task agent graph to
# release every httpx connection pool created for handoff agents too.
# Dead CLOSE_WAIT sockets in the pool cause kqueue CPU spin if left open.
if agent0 is not None:
root_agent = getattr(agent0, "agent", None)
agents_to_visit: list[Agent[Any]] = []
if root_agent is not None:
agents_to_visit.append(root_agent)
seen_agent_ids: set[int] = set()
extra_clients: list[Any] = []
seen_client_ids: set[int] = set()
primary_client = getattr(getattr(root_agent, "model", None), "client", None)
if primary_client is not None:
seen_client_ids.add(id(primary_client))
while agents_to_visit:
current_agent = agents_to_visit.pop()
current_agent_id = id(current_agent)
if current_agent_id in seen_agent_ids:
continue
seen_agent_ids.add(current_agent_id)
current_client = getattr(getattr(current_agent, "model", None), "client", None)
if current_client is not None and id(current_client) not in seen_client_ids:
seen_client_ids.add(id(current_client))
extra_clients.append(current_client)
for handoff in getattr(current_agent, "handoffs", ()) or ():
handoff_agent = getattr(handoff, "agent", handoff)
if handoff_agent is not None:
agents_to_visit.append(handoff_agent)
await agent0.close()
for client in extra_clients:
close_client = getattr(client, "close", None)
if close_client is None:
continue
try:
await close_client()
except Exception:
logging.exception("Exception while closing handoff agent client")

Copilot uses AI. Check for mistakes.
start_cleanup.set()
cleanup_attempts_left = len(entries)
while cleanup_attempts_left and entries:
Expand All @@ -443,6 +518,14 @@ async def _run_streamed() -> None:
continue
except Exception:
logging.exception("Exception in mcp server cleanup task")
# Cancel the MCP session task if it's still running to prevent
# the asyncio event loop from spinning on a dangling task.
if not mcp_sessions.done():
mcp_sessions.cancel()
try:
await mcp_sessions
except (asyncio.CancelledError, Exception):
pass
Comment on lines +526 to +528
Copy link

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

After mcp_sessions.cancel(), the code does await mcp_sessions with no timeout. This can hang indefinitely if the session task suppresses cancellation during cleanup (it catches CancelledError inside its cleanup loop). Use a bounded wait (e.g., asyncio.wait_for(..., timeout=MCP_CLEANUP_TIMEOUT)) and log if cancellation doesn’t complete, so shutdown can’t re-hang in the cleanup path.

Suggested change
await mcp_sessions
except (asyncio.CancelledError, Exception):
pass
await asyncio.wait_for(mcp_sessions, timeout=MCP_CLEANUP_TIMEOUT)
except asyncio.TimeoutError:
logging.warning(
"Timed out waiting for MCP session task cancellation after %s seconds",
MCP_CLEANUP_TIMEOUT,
)
except asyncio.CancelledError:
pass
except Exception:
logging.exception("Exception while waiting for mcp session task cancellation")

Copilot uses AI. Check for mistakes.


async def run_main(
Expand All @@ -465,12 +548,18 @@ async def run_main(
"""
from .session import TaskflowSession

# Start the watchdog thread — if the process hangs for any reason
# (asyncio spin, dead connections, MCP cleanup), this kills it.
start_watchdog()

last_mcp_tool_results: list[str] = []

async def on_tool_end_hook(context: RunContextWrapper[TContext], agent: Agent[TContext], tool: Tool, result: str) -> None:
watchdog_ping()
last_mcp_tool_results.append(result)

async def on_tool_start_hook(context: RunContextWrapper[TContext], agent: Agent[TContext], tool: Tool) -> None:
watchdog_ping()
await render_model_output(f"\n** 🤖🛠️ Tool Call: {tool.name}\n")

async def on_handoff_hook(context: RunContextWrapper[TContext], agent: Agent[TContext], source: Agent[TContext]) -> None:
Expand Down Expand Up @@ -731,3 +820,9 @@ async def _deploy(ra: dict, pp: str) -> bool:
if session is not None and not session.error:
session.mark_finished()
await render_model_output(f"** 🤖✅ Session {session.session_id} completed\n")

# Force-exit to prevent asyncio event loop spin on dangling
# tasks/connections from the responses API path. Flush first.
sys.stdout.flush()
sys.stderr.flush()
os._exit(0 if (session is None or session.finished) else 1)
Loading