From 2833bd34e58db803310c14fac9133a9047dd2eb5 Mon Sep 17 00:00:00 2001
From: Felix <0xfelixli@gmail.com>
Date: Mon, 29 Jun 2026 17:59:25 +0800
Subject: [PATCH] fix: prevent long scans from hanging or dying on fd
 exhaustion

Two independent failure modes surfaced on long multi-agent scans:

1. Too many open files (OSError 24). A long scan accumulates file
   descriptors (httpx pools, docker socket polling, per-subagent
   resources). The process never raised its soft RLIMIT_NOFILE, so on
   macOS (default soft limit 256) a long scan blew past the ceiling and
   crashed at whatever fd-opening call lost the race. Raise the soft
   limit toward the hard limit at startup.

2. Agents stuck at "Starting agent...". LLM_TIMEOUT was only applied to
   the startup warm-up call, never to agent-loop calls. Models on the
   openai/ prefix go through the SDK's OpenAI provider whose default
   timeout is 600s; as an httpx read timeout that parks an agent for ten
   minutes on a stalled stream. Install a default AsyncOpenAI client
   whose read timeout is bounded by LLM_TIMEOUT so a stalled stream
   raises and the existing retry policy recovers the agent.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 strix/config/models.py  | 33 ++++++++++++++++++++++++++++++++-
 strix/interface/main.py | 27 +++++++++++++++++++++++++++
 2 files changed, 59 insertions(+), 1 deletion(-)

diff --git a/strix/config/models.py b/strix/config/models.py
index 7f6227f3..396701ae 100644
--- a/strix/config/models.py
+++ b/strix/config/models.py
@@ -3,7 +3,7 @@
 from __future__ import annotations
 
 import os
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Any
 
 from agents import set_default_openai_api, set_default_openai_key, set_tracing_disabled
 from agents.models.multi_provider import MultiProvider
@@ -75,6 +75,37 @@ def configure_sdk_model_defaults(settings: Settings) -> None:
         set_default_openai_api("chat_completions")
     else:
         set_default_openai_api("responses")
+    _configure_openai_client_timeout(llm)
+
+
+def _configure_openai_client_timeout(llm: Any) -> None:
+    """Install a default AsyncOpenAI client whose read timeout is bounded.
+
+    Models on the ``openai/`` prefix (e.g. ``openai/gpt-5.5`` against a
+    custom ``api_base``) go through the SDK's OpenAI provider, NOT
+    litellm — so ``litellm.request_timeout`` does not apply to them. The
+    OpenAI SDK's default timeout is 600s, and as a *read* timeout that
+    means a stalled stream (server stops sending bytes without closing
+    the socket) parks the agent for 10 minutes — surfaced in the TUI as
+    an agent stuck at "Starting agent...".
+
+    Passing a float ``timeout`` makes the underlying httpx read timeout an
+    *idle* timeout: it resets on every received chunk, so a healthy long
+    generation is never cut off, but a truly stalled stream raises after
+    ``LLM_TIMEOUT`` seconds and the SDK retry policy recovers the agent.
+
+    The client carries the key/base_url because ``set_default_openai_client``
+    takes precedence over ``set_default_openai_key``.
+    """
+    if not llm.api_key or llm.timeout <= 0:
+        return
+    from agents import set_default_openai_client
+    from openai import AsyncOpenAI
+
+    client_kwargs: dict[str, Any] = {"api_key": llm.api_key, "timeout": float(llm.timeout)}
+    if llm.api_base:
+        client_kwargs["base_url"] = llm.api_base
+    set_default_openai_client(AsyncOpenAI(**client_kwargs), use_for_tracing=False)
 
 
 def _mirror_api_key_to_provider_env(model_name: str | None, api_key: str) -> None:
diff --git a/strix/interface/main.py b/strix/interface/main.py
index 4eae0527..49e0d2d9 100644
--- a/strix/interface/main.py
+++ b/strix/interface/main.py
@@ -737,8 +737,35 @@ def pull_docker_image() -> None:
     console.print()
 
 
+def _raise_open_files_limit() -> None:
+    """Raise the soft RLIMIT_NOFILE toward the hard limit.
+
+    A long multi-agent scan accumulates file descriptors (litellm httpx
+    pools, docker socket polling, per-subagent resources). On macOS the
+    default soft limit is just 256, which a long scan blows past and then
+    dies with ``OSError: [Errno 24] Too many open files`` — surfacing at
+    whatever random fd-opening call happens to lose the race. Bumping the
+    soft limit to the hard limit removes that ceiling. Not available on
+    Windows, where ``resource`` is absent.
+    """
+    try:
+        import resource
+    except ImportError:
+        return
+
+    soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
+    target = hard if hard != resource.RLIM_INFINITY else 1_048_576
+    if soft >= target:
+        return
+    try:
+        resource.setrlimit(resource.RLIMIT_NOFILE, (target, hard))
+    except (ValueError, OSError):
+        logger.debug("Could not raise RLIMIT_NOFILE from %s to %s", soft, target, exc_info=True)
+
+
 def main() -> None:
     configure_dependency_logging()
+    _raise_open_files_limit()
 
     if sys.platform == "win32":
         asyncio.set_event_loop_policy(asyncio.WindowsSelectorEventLoopPolicy())