feat: configurable aiohttp read_bufsize via AIOHTTP_READ_BUFSIZE env

The shared aiohttp ClientSession was created with the library default
read_bufsize (64 KiB). Some upstream providers emit a single SSE chunk
larger than that ceiling, which fails the stream with HTTP 400 and the
message "Got more than 131072 bytes when reading: b'data: ...'".

Known triggers seen in the wild:
- OpenRouter routing reasoning models (Kimi K2.6, R1-style) through
  providers that batch the full reasoning trace into one chunk
  (Novita, WandB, SiliconFlow)
- Image-generation responses returned as a single large SSE message
  (related: discussion #24543)

Fix: expose AIOHTTP_READ_BUFSIZE (default 1 MiB) and pass it through
to ClientSession. Users hitting unusually large chunks can raise it
further without code changes.

No behavior change at the default; the 1 MiB default is conservative
enough to handle the cases above while still rejecting truly runaway
payloads.

backend/open_webui/env.py

@@ -578,6 +578,20 @@ try:
 except ValueError:
     AIOHTTP_POOL_DNS_TTL = 300
 
+# Max size of a single chunk read from an SSE/HTTP stream, in bytes.
+# aiohttp's default is 64 KiB which can be exceeded when an upstream provider
+# emits a single very large SSE chunk (e.g. OpenRouter sometimes batches a
+# reasoning trace or tool-call payload into one chunk, or an image-generation
+# response is returned in one message). Raising this avoids
+# `Got more than N bytes when reading` 400 errors. Default 1 MiB.
+AIOHTTP_READ_BUFSIZE = os.getenv('AIOHTTP_READ_BUFSIZE', str(1024 * 1024))
+try:
+    AIOHTTP_READ_BUFSIZE = int(AIOHTTP_READ_BUFSIZE)
+    if AIOHTTP_READ_BUFSIZE <= 0:
+        AIOHTTP_READ_BUFSIZE = 1024 * 1024
+except ValueError:
+    AIOHTTP_READ_BUFSIZE = 1024 * 1024
+
 RAG_EMBEDDING_TIMEOUT = os.getenv('RAG_EMBEDDING_TIMEOUT', '')
 
 if RAG_EMBEDDING_TIMEOUT == '':

backend/open_webui/utils/session_pool.py

@@ -9,6 +9,11 @@ All pool parameters are configurable via environment variables:
     - AIOHTTP_POOL_CONNECTIONS (default 100) — max total connections
     - AIOHTTP_POOL_CONNECTIONS_PER_HOST (default 30) — per-host limit
     - AIOHTTP_POOL_DNS_TTL (default 300) — DNS cache TTL in seconds
+    - AIOHTTP_READ_BUFSIZE (default 1048576) — per-chunk read buffer size in
+      bytes. Raise this when an upstream emits very large SSE chunks (e.g.
+      reasoning-trace batching by some OpenRouter providers, or single-message
+      image-generation responses) which otherwise fail with
+      "Got more than N bytes when reading".
 
 Usage:
     from open_webui.utils.session_pool import get_session, cleanup_response
@@ -32,6 +37,7 @@ from open_webui.env import (
     AIOHTTP_POOL_CONNECTIONS,
     AIOHTTP_POOL_CONNECTIONS_PER_HOST,
     AIOHTTP_POOL_DNS_TTL,
+    AIOHTTP_READ_BUFSIZE,
 )
 
 log = logging.getLogger(__name__)
@@ -61,12 +67,14 @@ async def get_session() -> aiohttp.ClientSession:
             connector=connector,
             timeout=timeout,
             trust_env=True,
+            read_bufsize=AIOHTTP_READ_BUFSIZE,
         )
         log.info(
-            'Created shared aiohttp session pool (limit=%s, per_host=%s, dns_ttl=%d)',
+            'Created shared aiohttp session pool (limit=%s, per_host=%s, dns_ttl=%d, read_bufsize=%d)',
             AIOHTTP_POOL_CONNECTIONS or 'unlimited',
             AIOHTTP_POOL_CONNECTIONS_PER_HOST or 'unlimited',
             AIOHTTP_POOL_DNS_TTL,
+            AIOHTTP_READ_BUFSIZE,
         )
     return _session