Closes #96681

What Problem This Solves

After an agent calls the message tool (action=send) to deliver its main reply, models often append a short trailing meta-acknowledgement -- patterns like 已发 #22141, Sent above, 核心回答如下, OK, Roger, Got it, and compound forms like Sent. Replied in thread., 已发, 不再追加总结.

These acks carry no information beyond the message-tool send that already went out, but OpenClaw currently delivers them as a second visible message in the channel.

The existing filterMessagingToolDuplicates only catches full-content duplicates via substring match with a 10-char minimum, so short meta-acks (e.g. 已发 #22141, 9 chars) escape the floor and reach the channel as a duplicate message.

Why This Change Was Made

This patch adds a complementary filter -- filterMessagingToolMetaCommentary -- that recognizes post-tool-send meta commentary by pattern (Chinese + English) and suppresses it from the final reply payload.

Key design decisions:

1. Compound ack coverage -- Two-pass detection: first try full-text anchored patterns, then split on sentence boundaries and verify every segment independently matches an ack pattern. Catches Sent. Replied in thread., 已发, 不再追加总结, OK. Done. while preserving mixed forms like OK, that's the fix.

2. Only fires after message-tool sends (sentTexts.length > 0). Agents with no message-tool sends are unaffected.

3. Length-capped at 200 chars and anchored patterns only -- prefix matches like Oklahoma weather, 已发现问题 are not suppressed.

4. Media payloads always preserved, even if caption is a meta-ack.

5. Wired into both paths: agent-runner-payloads.ts (final reply) and followup-delivery.ts (queued follow-up).

Evidence

Real behavior proof (tsx runtime)

============================================================
Real Behavior Proof: Post-Message-Tool Meta-Ack Filter
============================================================

[1] Meta-ack pattern detection:
  ✓ "已发 #22141" → true
  ✓ "Sent above" → true
  ✓ "OK" → true
  ✓ "核心回答如下" → true
  ✓ "Sent. Replied in thread." → true
  ✓ "已发, 不再追加总结" → true
  ✓ "OK. Done." → true
  ✓ "Got it. Will do." → true
  ✓ "了解, 收到" → true
  ✓ "OK, that's the fix." → false
  ✓ "Done, the migration is complete." → false
  ✓ "Got it, let me take a look at the code." → false
  ✓ "Oklahoma weather" → false
  ✓ "已发现问题" → false
  Accuracy: 14/14

[2] Pipeline: 3 meta-acks + 1 real reply → filters to 1 payload
  Output: 1 payloads, kept: "Here is the real follow-up analysis..."

[3] Media payloads always preserved:
  1 media payload(s) preserved after filter

✓ All behavior checks passed.

Test coverage (reply-payloads.test.ts)

All 134 tests pass (includes 109 new test cases for meta-ack detection):

 Test Files  1 passed (1)
      Tests  134 passed (134)
   Duration  5.10s

Standalone meta-acks covered (46 positive cases): 已发, 已发 #22141, 已发送, 已发完毕, 主回复已发, 消息已发出, 回复已发, 好了, 收到, 完毕, 完成, 了解, 知道了, 明白, 核心回答如下, 总结如下, 以下为核心回答, 以下为总结, 不再追加总结, 不再追加, 以下为回复, 答案如下, 如上, 回复如上, Sent, Sent above, Done., Replied above, Replied in thread, See above, As above, Posted., Acknowledged., OK, Okay, Roger, Got it, Gotcha, Copy that, Copied, Ack, Will do, On it, Noted, Understood, Thanks

Compound meta-acks covered (13 cases): Sent. Replied in thread., 已发, 不再追加总结, 已发. 不再追加., OK. Done., Roger, copy., Done. Sent., 收到. 已发., Sent above. Replying in thread., Got it. Will do., 已发, 完成, OK, noted, 了解, 收到, Copied, thanks

False positives excluded (18+ cases): Oklahoma weather, sentence fixed, 已发现问题, 已发现, 已收到消息, sentry deployed, okay let's go, OK, that's the fix., Got it, let me take a look at the code., Done, the migration is complete., Sent, but I noticed the logs have an error., 收到, 但是这个方案有一个问题, real reply content, and more.

Build verification

The full project builds successfully with these changes:

[build-all] phase timings: total 506.3s

No type errors, no build failures.

Merge Risk

Low. The new filter is additive and only activates when message-tool sends happened this run (sentTexts.length > 0). It does not modify the existing dedupe logic. Media payloads are explicitly preserved. The worst-case failure mode is a false positive suppressing a real reply, which the anchored patterns + 200-char limit + compound segment validation guard against.

🤖 Generated with Claude Code