[Bug] getUpdates at-least-once delivery double-invokes AI pipeline (≈1s duplicate user messages)

[ericcaiwx-star]

Summary

In production with OpenClaw + @tencent-weixin/openclaw-weixin, nearly every inbound user text message is processed twice (~1 second apart). Transcripts show dup≈2 for the same user content; agents answer twice / burn double tool calls.

Environment

• Plugin: @tencent-weixin/openclaw-weixin@2.4.6
• Host: OpenClaw gateway (private banking mid-office deploy)
• Channel: openclaw-weixin personal WeChat (iLink getUpdates long-poll)

Evidence (redacted)

• Staff agent WeChat session over ~10h wall time: ~26 unique user texts, ~52 raw user turns (almost every message duplicated).
• Pair spacing typically ~1s, same body text.
• Not explained by user double-tap; pattern is systematic across the session.

Root cause (code)

1. monitorWeixinProvider loops resp.msgs and calls processOneMessage with no dedupe (src/monitor/monitor.ts).
2. processOneMessage has no claim/dedupe before slash-command / media / dispatchReplyFromConfig (src/messaging/process-message.ts).
3. Wire messages expose message_id / seq / client_id (logged in debug), but weixinMessageToMsgContext stamps a fresh random MessageSid via generateId("openclaw-weixin") every delivery (src/messaging/inbound.ts) — so downstream cannot correlate replays either.

Sibling OpenClaw channels (Telegram / Slack / WhatsApp) already use replay guards / durable event ids for at-least-once transports. Weixin does not.

Expected

A logical WeChat message (same message_id, or stable fallback) should run the inbound AI pipeline once within a short TTL window.

Proposed fix

See accompanying PR: short-TTL in-memory claim in processOneMessage keyed by message_id → client_id → seq → body fingerprint; stable MessageSid from the same key.

Workaround (operators)

Session-level “answer once if user text repeats” in agent SOUL — does not stop double tool spend.

[ericcaiwx-star]

Fix PR: #240