Bug Summary

When platform adapters lose network connectivity and exhaust their internal reconnection retries, they silently stop without notifying the gateway's fatal-error handling mechanism. This leaves the gateway process alive but the platform permanently unresponsive.

Observed Behavior

06:38  DNS resolution fails
06:38  Telegram starts retrying (1/10, 2/10, ...)
06:38  QQBot starts retrying (1/100, 2/100, ...)
07:33  Both adapters exhaust retries and stop
07:33  No more platform activity; gateway process still alive

Root Cause

File: gateway/platforms/qqbot/adapter.py

At lines ~608 and ~620, when attempt >= MAX_RECONNECT_ATTEMPTS:

# Current (buggy):
self._mark_disconnected()

# Should be:
self._set_fatal_error(f"QQBot reconnect exhausted after {MAX_RECONNECT_ATTEMPTS} attempts")
await self._notify_fatal_error()

Without _notify_fatal_error(), the gateway's _platform_reconnect_watcher() never attempts to restart the QQBot adapter.

File: gateway/platforms/telegram.py

During _handle_polling_network_error(), the runtime status remains connected during retries 1-10. The fatal error is only set at attempt 11. Consider writing a retrying state during the retry window so external monitors can detect degraded connectivity earlier.

Suggested Fixes

Environment

• hermes-agent version: latest main (May 20, 2026)
• OS: macOS 26.5
• Python: 3.12