Skip to content

Telegram surface leaks assistant pre-tool scratch/commentary into user-visible chat #66161

@Abdo-codes

Description

@Abdo-codes

Summary

On Telegram, assistant pre-tool scratch/commentary text is sometimes being delivered to the user as normal chat messages instead of remaining internal.

Expected

Only user-facing final responses should appear in the Telegram chat.
Internal planning/scratch/tool-chatter should never be surfaced to the user.

Actual

Short internal-looking lines sometimes appear in chat, for example scratch notes like:

  • "Need add hasLoadedChildren maybe."
  • "Need maybe built-in skill? no. use apprank apps list/status."
  • similar pre-tool implementation/thought fragments

These are not intended as user-facing replies.

Why this matters

  • leaks low-value internal scratch text into the user conversation
  • degrades trust / polish
  • can expose implementation chatter that should stay hidden
  • repeated even after changing assistant behavior, which suggests a surface/runtime handling issue rather than only prompting style

Observed context

  • Surface: Telegram direct chat
  • Runtime: OpenClaw main session
  • Tools involved: normal tool use (exec, read, etc.)
  • The leak appears around turns where tool work is happening and the assistant would normally keep intermediate commentary internal

Working hypothesis

A class of assistant intermediate text/commentary is being treated as a normal outbound chat message on Telegram instead of remaining internal/non-user-visible.

This does not appear to come from GitHub/Codex/AppRank output itself; it looks like assistant-side pre-tool chatter/scratch text is being surfaced by the runtime/channel pipeline.

Suggested investigation areas

  • Telegram outbound message path for assistant intermediate/commentary content
  • any distinction between internal commentary vs final user-facing assistant text
  • tool-call narration handling on Telegram compared with other surfaces
  • whether certain assistant messages before tool calls are being flushed as normal channel replies

Repro sketch

  1. Use OpenClaw in a Telegram DM
  2. Ask for a multi-step repo/tool task
  3. During tool-heavy turns, observe whether short internal scratch lines appear before the actual final response

Notes

The safest behavioral workaround is to avoid emitting any pre-tool chatter at all, but that only masks the issue. The platform should not surface this class of internal assistant text to end users.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions