From 5ae4aef30e3efd4dd46968a6daf5f0c657966bb6 Mon Sep 17 00:00:00 2001
From: Will Miao <willmiao.dev@gmail.com>
Date: Mon, 7 Sep 2026 09:54:28 +0800
Subject: [PATCH] fix(llm): disable brotli for catalog fetch to prevent native
 crash (#1099, #1101)

models.dev is served by Cloudflare with brotli compression when the
client advertises it, and brotli is a required dependency here, so
aiohttp always negotiates br. A corrupted br stream can crash the
native decoder with a Windows access violation (a Python-level
exception handler cannot catch it), or produce garbage bytes.

Send an explicit "Accept-Encoding: gzip, deflate" header on the model
catalog and Ollama model-list requests so the server never returns
brotli. zlib handles corrupt gzip data by raising ContentEncodingError
(an aiohttp.ClientError subclass), which the existing handlers already
catch and degrade to a warning with an empty-catalog fallback.
---
 py/services/llm_service.py         | 12 ++++++++++--
 tests/services/test_llm_service.py | 30 +++++++++++++++++++++++++++++-
 2 files changed, 39 insertions(+), 3 deletions(-)

diff --git a/py/services/llm_service.py b/py/services/llm_service.py
index 41ef7870..04ccb346 100644
--- a/py/services/llm_service.py
+++ b/py/services/llm_service.py
@@ -34,6 +34,14 @@ _model_output_limits: Dict[str, Dict[str, int]] = {}
 
 _CATALOG_TIMEOUT = aiohttp.ClientTimeout(total=30)
 
+# Cloudflare serves brotli when the client advertises it, and brotli is a
+# required dependency here — a corrupted br stream can crash the native
+# decoder with a Windows access violation (issue #1099).  Request gzip
+# instead; zlib decompression is not affected and corrupt gzip data only
+# raises ContentEncodingError (an aiohttp.ClientError subclass), which the
+# exception handlers below already catch.
+_NO_BROTLI_HEADERS = {"Accept-Encoding": "gzip, deflate"}
+
 
 async def _load_model_catalog() -> Dict[str, List[str]]:
     """Fetch and parse the model catalog.
@@ -53,7 +61,7 @@ async def _load_model_catalog() -> Dict[str, List[str]]:
 
     try:
         async with aiohttp.ClientSession(timeout=_CATALOG_TIMEOUT) as session:
-            async with session.get(_MODEL_CATALOG_URL) as resp:
+            async with session.get(_MODEL_CATALOG_URL, headers=_NO_BROTLI_HEADERS) as resp:
                 if resp.status != 200:
                     logger.warning("Model catalog returned HTTP %s", resp.status)
                     return _catalog_cache or {}
@@ -126,7 +134,7 @@ async def fetch_ollama_models(api_base: str) -> List[str]:
     url = f"{api_base.rstrip('/')}/models"
     try:
         async with aiohttp.ClientSession(timeout=_OLLAMA_API_TIMEOUT) as session:
-            async with session.get(url) as resp:
+            async with session.get(url, headers=_NO_BROTLI_HEADERS) as resp:
                 if resp.status != 200:
                     logger.debug("Ollama API returned HTTP %s from %s", resp.status, api_base)
                     return []
diff --git a/tests/services/test_llm_service.py b/tests/services/test_llm_service.py
index 883c10d3..9c00c5c1 100644
--- a/tests/services/test_llm_service.py
+++ b/tests/services/test_llm_service.py
@@ -322,8 +322,12 @@ class MockGetSession:
 
     def __init__(self, response):
         self._response = response
+        self.last_url = None
+        self.last_headers = None
 
-    def get(self, url):
+    def get(self, url, headers=None):
+        self.last_url = url
+        self.last_headers = headers
         return self._response
 
     async def __aenter__(self):
@@ -373,3 +377,27 @@ class TestModelCatalog:
             models = await fetch_ollama_models("http://localhost:11434/v1")
 
         assert models == []
+
+    @pytest.mark.asyncio
+    async def test_catalog_request_disables_brotli_encoding(self):
+        """The catalog request must not advertise br — a corrupt brotli stream
+        can crash the native decoder (Windows access violation, issue #1099)."""
+        response = MockResponse(200, json_data={})
+        session = MockGetSession(response)
+
+        with mock.patch("aiohttp.ClientSession", return_value=session):
+            await llm_module._load_model_catalog()
+
+        assert session.last_headers == {"Accept-Encoding": "gzip, deflate"}
+
+    @pytest.mark.asyncio
+    async def test_ollama_request_disables_brotli_encoding(self):
+        """The Ollama models request must not advertise br either."""
+        response = MockResponse(200, json_data={"data": [{"id": "llama3"}]})
+        session = MockGetSession(response)
+
+        with mock.patch("aiohttp.ClientSession", return_value=session):
+            models = await fetch_ollama_models("http://localhost:11434/v1")
+
+        assert models == ["llama3"]
+        assert session.last_headers == {"Accept-Encoding": "gzip, deflate"}