PR #17545: fix(weixin): route send_message media uploads through a fresh aiohttp session

• Repository: NousResearch/hermes-agent
• Author: alexzhu0
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/17545

Description (problem / solution / changelog)

Closes #17347.

Symptom

send_message tool call to a Weixin chat with a .docx / image attachment fails 5/5 times with:

WARNING [Weixin] send chunk failed to=…: Timeout context manager should be used inside a task
ERROR   [Weixin] send failed to=…: Timeout context manager should be used inside a task

Text replies through the same platform continue to work — only tool-initiated media sends break.

Root cause

send_weixin_direct (gateway/platforms/weixin.py:2005) optimistically reuses the live long-poll adapter's _send_session when that adapter is running:

live_adapter = _LIVE_ADAPTERS.get(resolved_token)
send_session = getattr(live_adapter, '_send_session', None)
if live_adapter is not None and send_session is not None and not send_session.closed:
    # reuse for text AND media

That _send_session was created inside the long-poll reader task. Reusing it from the tool-call task works for short request timeouts (the text-send path uses API_TIMEOUT_MS ≈ 8 s), but the 120-second upload-CDN timeout in _upload_ciphertext:

timeout = aiohttp.ClientTimeout(total=120)
async with session.post(upload_url, data=ciphertext, …, timeout=timeout) as response:

trips aiohttp's asyncio.current_task() check inside the session's timeout handler when the session crosses task boundaries.

Text sends work because their timer registers and tears down within the same tool-call task turn; the 120 s upload timer spans enough turns that the task-binding check fails.

Fix

Reuse the live adapter only when media_files is empty. Any media send falls through to the existing fresh-session path (the same aiohttp.ClientSession that cron delivery and adapter-unavailable tool calls already use), so the upload runs in the tool-call task's own session and the timeout-context check sees a valid current task.

-if live_adapter is not None and send_session is not None and not send_session.closed:
+has_media = bool(media_files)
+if (
+    live_adapter is not None
+    and send_session is not None
+    and not send_session.closed
+    and not has_media
+):
     # text-only reuse, preserved to keep the common-case perf

Text-only behavior is preserved byte-for-byte. The adapter-unavailable branch (cron / standalone) already fell through to the fresh-session path, so its behavior also unchanged.

Not in scope

• Doesn't restructure the live adapter's internal task model. That may be the right long-term fix (own task for each send op), but it's a bigger change and this one-line gate unblocks the user-visible bug first.
• Doesn't add a regression test — the bug requires a live weixin account + CDN flight to reproduce. Adding a mock-aiohttp test that hits the same asyncio.current_task() branch is possible but significantly more scaffolding than the fix; happy to add in a follow-up if preferred.

File

• gateway/platforms/weixin.py: +20 / -10

Changed files

• gateway/platforms/weixin.py (modified, +20/-10)