Summary

When Telegram streaming is enabled together with visible reasoning (reasoning=true / thinking on), intermediate toolUse turns are emitted to user chat as message_end block replies.

This leaks internal progress text like "让我检查状态…" / "让我等待完成…" into the visible conversation, causing the user-visible final answer flow to be interrupted/overwritten in practice.

Why this is a regression-like behavior

This reproduces on openclaw 2026.2.22-2, whose npm gitHead is:

• 45febecf2a2d91fd1a378bb2cae38ec21e71857e

And this gitHead already contains the prior Telegram fixes:

• #17973 (dddb1bc...) ✅ ancestor
• #20774 (ab256b8...) ✅ ancestor
• #22613 (8b1fe0d...) ✅ ancestor
• #23202 (63b4c50...) ✅ ancestor

So this is not an “old version missing fix” report.

Steps to reproduce

1. Configure Telegram channel with streaming enabled (block/partial behavior is enough to hit this path).
2. Enable reasoning/thinking visibility.
3. Trigger a run where assistant performs tool polling/execution before final answer (e.g. exec + repeated process:poll).
4. Observe visible Telegram output sequence.

Expected behavior

• Intermediate toolUse turns should stay internal.
• User should only see intended answer/reasoning messages, without tool-step status chatter being emitted as final chat messages.

Actual behavior

• Intermediate toolUse text is emitted to channel as message_end block replies.
• Final answer appears interleaved with these status lines, creating "final output disappears / replaced by repeated progress lines / then reappears" user perception.

OpenClaw version

• 2026.2.22-2
• npm metadata: { "version": "2026.2.22-2", "gitHead": "45febecf2a2d91fd1a378bb2cae38ec21e71857e" }

Operating system

• macOS (arm64)

Install method

• npm global package (~/.local/lib/node_modules/openclaw)

Logs, screenshots, and evidence

1) Session evidence showing intermediate toolUse status chatter

From local session transcript (~/.openclaw/agents/main/sessions/0e9e7c38-d6d5-4993-95c3-04e08d35a897.jsonl):

• stopReason:"toolUse" with text "让我检查状态：1"
• Later merged text: "让我检查状态：2 ... 让我检查状态：3 ... 最终结论..."

From another long-running transcript (~/.openclaw/agents/coder/sessions/b75fc08b-0210-45bf-b67c-5303bc48a230.jsonl):

• repeated assistant stopReason:"toolUse" texts:
  • "任务仍在生成。让我检查进程状态："
  • "任务正在读取所有文档。让我继续等待："
  • "让我直接检查文件："
  • "任务仍在运行。让我检查最终状态："

2) Code-path evidence

src/agents/pi-embedded-subscribe.handlers.messages.ts (handleMessageEnd) currently emits onBlockReply for message_end text without excluding intermediate toolUse assistant turns that include tool-call blocks.

That allows toolUse progress text to be delivered as user-visible message blocks.

3) Reproducible failing test before fix

Added targeted test (locally) to encode expected behavior:

• src/agents/pi-embedded-subscribe.subscribe-embedded-pi-session.suppresses-message-end-block-replies-message-tool.test.ts

Before fix, it fails with:

• expected onBlockReply not called
• actual called with text "让我检查状态："

Impact and severity

• Affected: Telegram users with stream + reasoning flows involving tool calls.
• Severity: Medium-High (user-visible response corruption/noise and confusing output order).
• Frequency: High for workflows that poll tools before final answer.
• Consequence: perceived output overwrite/regression and reduced trust/readability.

Additional information

Related historical and active items:

• #17935, #17973, #20774, #22613, #23202
• #19180, #19193, #20568

This report is specifically about intermediate toolUse message_end emission into Telegram-visible replies.