openclaw - 💡(How to fix) Fix [Bug]: Assistant can reply to stale user intent after reasoning-only/tool retry [2 comments, 3 participants]

Summary

In Telegram direct chat, the assistant can produce a final visible reply that answers an older nearby user message instead of the latest inbound message after a tool-use / reasoning-only retry path.

This appears to be an OpenClaw runtime/session continuation bug rather than Telegram message ordering. The latest user message is present in the session history before the assistant reply, but the final answer can still target a previous intent.

Observed behavior

Recent direct-chat sequence:

1. User: asks to treat ffb as shorthand for "future feature backlog list".
2. Assistant saves/acknowledges that preference.
3. User: Add to ffb as part of item 1 on list: ingest tab on memory dashboard...
4. Assistant performs a tool call to read the backlog file for the new request.
5. Assistant final reply incorrectly says: Yes, Sir — ffb now means future feature backlog list. I’ve saved it.

So the assistant answered the previous shorthand-confirmation message instead of the latest actionable backlog-edit message.

A similar stale-response incident happened immediately before this, where a final answer repeated the previous backlog-update completion instead of answering the current shorthand-confirmation message.

Evidence / diagnostics

Session history showed the messages in the correct order:

• latest Telegram user message was already present in transcript before assistant final response;
• the wrong answer occurred after a tool call;
• gateway logs shortly before the first symptom included:

reasoning-only assistant turn detected: ... retrying ... with visible-answer continuation

This suggests the issue is in the assistant retry/continuation/session state path rather than Telegram delivery order.

Relevant source area found locally:

• src/agents/pi-embedded-runner/run.ts
  • reasoning-only / empty-response retry loop
  • appends visible-answer retry instruction and continues
• src/agents/pi-embedded-runner/run/incomplete-turn.ts
  • REASONING_ONLY_RETRY_INSTRUCTION
  • EMPTY_RESPONSE_RETRY_INSTRUCTION

The retry instruction currently asks the model to continue from the partial turn and produce visible output, but does not strongly re-anchor the final answer to the latest inbound user message / current run intent after tool use.

Expected behavior

After every tool-use continuation, reasoning-only retry, empty-response retry, or final visible answer generation, the assistant must answer the latest inbound user message for that run/session.

If the runtime cannot verify that the final answer is anchored to the latest user message, it should prefer a safe blocked/retry state rather than delivering a stale answer.

Impact

High trust impact for direct chat: the assistant appears to ignore the newest user message and answer an old one. This is especially damaging in rapid back-and-forth Telegram workflows and when the user has issued an actionable edit/request.

Possible fix direction

Add a runtime-level guard around retry/continuation/final-delivery paths:

• include/latest user message fingerprint or sequence ID in retry prompt construction;
• ensure final assistant answer is associated with the current run/latest user message;
• avoid reusing stale assistant-visible-answer state across tool/retry continuations;
• add regression tests for: user A -> assistant/tool/retry -> user B arrives/current run -> assistant must not answer user A.

Environment

• OpenClaw: 2026.5.4 (325df3e)
• Runtime: main Telegram direct session
• Model: openai-codex/gpt-5.5
• Channel: Telegram direct