Fix run cancellation and live activity stability

src/openai_compat.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import os
 from typing import Any, Iterator
 from urllib import error, request
 
@@ -31,6 +32,7 @@ ANTHROPIC_MESSAGES_MODEL_PREFIXES = (
 
 ANTHROPIC_VERSION = '2023-06-01'
 ANTHROPIC_MAX_TOKENS = 4096
+DEFAULT_MODEL_IDLE_TIMEOUT_SECONDS = 60.0
 
 
 def _join_url(base_url: str, suffix: str) -> str:
@@ -274,6 +276,23 @@ class OpenAICompatClient:
     def __init__(self, config: ModelConfig) -> None:
         self.config = config
 
+    def _request_timeout_seconds(self) -> float:
+        """Return per-socket model timeout.
+
+        `ModelConfig.timeout_seconds` is the whole request budget used by the
+        product, but urllib applies it as an idle socket timeout. Keeping it at
+        one hour makes a bad upstream stream hold a session lock for an hour.
+        Use a shorter idle timeout by default; any normally streaming response
+        keeps extending this naturally because bytes continue to arrive.
+        """
+        raw = os.environ.get('CLAW_MODEL_IDLE_TIMEOUT_SECONDS', '').strip()
+        try:
+            configured = float(raw) if raw else DEFAULT_MODEL_IDLE_TIMEOUT_SECONDS
+        except ValueError:
+            configured = DEFAULT_MODEL_IDLE_TIMEOUT_SECONDS
+        configured = max(5.0, configured)
+        return min(float(self.config.timeout_seconds), configured)
+
     def complete(
         self,
         messages: list[dict[str, Any]],
@@ -351,7 +370,7 @@ class OpenAICompatClient:
             method='POST',
         )
         try:
-            with request.urlopen(req, timeout=self.config.timeout_seconds) as response:
+            with request.urlopen(req, timeout=self._request_timeout_seconds()) as response:
                 yield StreamEvent(type='message_start')
                 for event_payload in self._iter_sse_payloads(response):
                     yield from self._parse_stream_payload(event_payload)
@@ -380,7 +399,7 @@ class OpenAICompatClient:
             method='POST',
         )
         try:
-            with request.urlopen(req, timeout=self.config.timeout_seconds) as response:
+            with request.urlopen(req, timeout=self._request_timeout_seconds()) as response:
                 raw = response.read()
         except error.HTTPError as exc:
             detail = exc.read().decode('utf-8', errors='replace')
@@ -521,7 +540,7 @@ class OpenAICompatClient:
             method='POST',
         )
         try:
-            with request.urlopen(req, timeout=self.config.timeout_seconds) as response:
+            with request.urlopen(req, timeout=self._request_timeout_seconds()) as response:
                 raw = response.read()
         except error.HTTPError as exc:
             detail = exc.read().decode('utf-8', errors='replace')
@@ -561,7 +580,7 @@ class OpenAICompatClient:
             method='POST',
         )
         try:
-            with request.urlopen(req, timeout=self.config.timeout_seconds) as response:
+            with request.urlopen(req, timeout=self._request_timeout_seconds()) as response:
                 yield StreamEvent(type='message_start')
                 tool_block_indexes: dict[int, int] = {}
                 next_tool_index = 0