Summary

In Telegram group chats, when a tool call fails (e.g. exec), OpenClaw may send the warning line (e.g. ⚠️ 🛠️ Exec: ... failed) and the user can miss the assistant’s normal final reply, even though the run completes successfully.

I observed this in a real conversation where:

• tool call failed first (gh not installed),
• the agent recovered and completed the task,
• the final assistant message exists in session transcript,
• but the user reported only seeing the tool-failure warning message in Telegram.

Environment

• OpenClaw: 2026.2.15 (3fe22ea)
• Channel: Telegram (group chat)
• Runtime: gateway local systemd
• Model during repro: openai-codex/gpt-5.3-codex

Reproduction (practical)

1. In Telegram group chat, ask the assistant to do a multi-step task that triggers exec.
2. Ensure first command fails (example: gh --version on host without gh).
3. Let the assistant continue with fallback logic and complete task.
4. Observe Telegram output: warning line appears (⚠️ 🛠️ Exec: ... failed).
5. In some cases, user reports not seeing the final assistant reply (though transcript contains it).

Evidence from local run

• Tool failure logged:
  • error [tools] exec failed: /bin/sh: 1: gh: not found ...
• Same run completed normally:
  • embedded run done ... aborted=false
• Session transcript contains a normal assistant final message for that turn.

Suspected code paths

1. Tool warning payload generation:

   • src/agents/pi-embedded-runner/run/payloads.ts
   • buildEmbeddedRunPayloads() adds warning text:
     • ⚠️ ${formatToolAggregate(...)} failed...

2. Potential blocking before final payload send:

   • src/auto-reply/reply/agent-runner.ts
   • tool result delivery tasks are tracked and awaited (Promise.allSettled(pendingToolTasks)) before final reply payload dispatch.
   • If tool-result delivery path stalls/fails in-channel, final reply timing/delivery may be impacted.

Expected behavior

• Tool-failure diagnostics should not replace/block the final assistant response.
• If warning is shown, final answer should still reliably arrive.

Note

No fix proposal in this report — sharing user-visible impact + likely code areas for investigation.