Description

The Gateway is injecting the same inbound message into a session multiple times, causing the agent to respond to each duplicate independently. This results in a flood of duplicate replies being delivered to the channel.

Reproduction

This has been observed on two separate channels within hours:

Case 1: Telegram (2026-03-31 ~17:36 JST)

• A single Telegram message (message_id 6627) was injected dozens of times into the main session
• SIGUSR restart did not resolve it; required /new session reset (3 times)

Case 2: Slack (2026-03-31 ~23:05–00:16 JST)

• A single Slack message from a user in channel CXXXXXXXXXX was injected 4 times into the session
• Each injection triggered a full agent response, resulting in 40+ delivered reply log entries over ~1 hour
• Gateway restart and session reset did not prevent recurrence

Evidence from gateway.log

Slack duplicate deliveries (burst pattern, ~250ms apart):

23:26:35.196 [slack] delivered reply to channel:CXXXXXXXXXX
23:26:35.563 [slack] delivered reply to channel:CXXXXXXXXXX
23:26:35.814 [slack] delivered reply to channel:CXXXXXXXXXX
23:26:36.068 [slack] delivered reply to channel:CXXXXXXXXXX
23:26:36.717 [slack] delivered reply to channel:CXXXXXXXXXX

Session transcript shows same message injected 4 times:

{"ts":"2026-03-31T15:13:24.312Z","content":"System: [2026-04-01 00:04:48 GMT+9] Slack message in #channel from user: ..."}
{"ts":"2026-03-31T15:14:13.018Z","content":"System: [2026-04-01 00:04:48 GMT+9] Slack message in #channel from user: ..."}
{"ts":"2026-03-31T15:15:02.082Z","content":"System: [2026-04-01 00:04:48 GMT+9] Slack message in #channel from user: ..."}
{"ts":"2026-03-31T15:16:28.457Z","content":"System: [2026-04-01 00:04:48 GMT+9] Slack message in #channel from user: ..."}

Additional symptoms during the period:

• WebSocket connections disconnecting/reconnecting every 30-60 seconds
• Repeated live session model switch detected logs for multiple models cycling every 20-30s
• [xai-auth] bootstrap config fallback: no config-backed key found logged every 20-30s (xAI not configured)
• Discord health-monitor restart due to stale-socket

Environment

• OpenClaw version: 2026.3.28 (f9b1079)
• OS: macOS (Darwin arm64)
• Channels affected: Telegram (polling), Slack (socket mode)
• Agent: embedded

Expected Behavior

Each inbound message should be injected into the session exactly once, with deduplication preventing re-injection of already-processed messages.

Workaround

Currently the only mitigation is resetting the session via /new, but the issue can recur.