fix(kosong): add default 120s timeout to create_openai_client

[wintrover]

Problem

create_openai_client() in kosong/chat_provider/openai_common.py does not pass a timeout to AsyncOpenAI(). The OpenAI SDK defaults to 600s, so when an upstream proxy times out earlier (e.g. MiMo API proxy at ~300s), the client silently waits another 5+ minutes before the SDK timeout fires. Users see "infinite loading."

Root Cause

22:43:43  API call starts (450K tokens, openai_legacy → MiMo proxy)
22:48:48  APIConnectionError after 5m5s (proxy-side timeout)
          → connection recovery → retrying once → also hangs
22:49:29  User Ctrl+C after 41s

The AsyncOpenAI client inherits the SDK default of 600s timeout. The proxy times out at ~300s, leaving the client hanging.

Fix

Inject httpx.Timeout(120.0, connect=30.0) as a default when the caller does not specify one. Callers can still override by passing timeout=... in client_kwargs.

def create_openai_client(*, api_key, base_url, client_kwargs):
    kwargs = dict(client_kwargs)
    if "timeout" not in kwargs:
        kwargs["timeout"] = httpx.Timeout(120.0, connect=30.0)
    return AsyncOpenAI(api_key=api_key, base_url=base_url, **kwargs)

Trade-offs

Parameter	Value	Rationale
connect	30s	TCP handshake
read/write/pool	120s	Covers normal large-context requests (30-60s typical)

• DeepSeek, OpenAI: responses typically <30s → no impact
• MiMo large context: 30-60s normal → 120s provides headroom
• Fails 4x faster than current 600s default

Fixes #2384 (for openai_legacy providers)

---

[devin-ai-integration]

Devin Review found 2 potential issues.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🔴 Logging FileHandler created at module level before log directory exists

The logging.basicConfig call at module level (line 79–86) instantiates a logging.FileHandler(LOG_FILE) for ~/.kimi-code/proxy.log. This executes at import time, before main() runs os.makedirs(os.path.dirname(LOG_FILE), exist_ok=True) at line 333. If the ~/.kimi-code/ directory does not yet exist (e.g. on first run), FileHandler.__init__ calls open(...) which raises FileNotFoundError, crashing the script before it ever reaches main().

Prompt for agents

The fix is to ensure the log directory exists before configuring the logging handlers. Move the `os.makedirs(os.path.dirname(LOG_FILE), exist_ok=True)` call from `main()` (line 333) to just before the `logging.basicConfig(...)` call at module level (around line 78). This way the directory is guaranteed to exist when the FileHandler tries to open the log file. Alternatively, move the entire logging configuration into `main()` so it runs after the directory creation.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🟡 Duplicate Content-Length headers sent in _send_upstream_response

In _send_upstream_response, the header-copying loop (lines 152–163) forwards all upstream headers except transfer-encoding, content-encoding, connection, and keep-alive — but content-length is not in the exclusion list. After the loop, line 165 adds another Content-Length header. Since BaseHTTPRequestHandler.send_header appends rather than replaces, this results in duplicate Content-Length headers in the response. If urllib decompressed the body (after stripping content-encoding), the two values will differ, which violates HTTP spec and can confuse clients.

Suggested change

	for key, value in resp.headers.items():
	k = key.lower()
	if k not in (
	"transfer-encoding",
	"content-encoding",
	"connection",
	"keep-alive",
	):
	try:
	self.send_header(key, value)
	except Exception:
	pass
	# Override content-length to match our read body
	self.send_header("Content-Length", str(len(body)))
	for key, value in resp.headers.items():
	k = key.lower()
	if k not in (
	"transfer-encoding",
	"content-encoding",
	"content-length",
	"connection",
	"keep-alive",
	):
	try:
	self.send_header(key, value)
	except Exception:
	pass
	# Override content-length to match our read body
	self.send_header("Content-Length", str(len(body)))

---

Was this helpful? React with 👍 or 👎 to provide feedback.