Problem

When messages.suppressToolErrors: true is configured, non-mutating tool errors are correctly suppressed from being forwarded to the user. However, mutating tool errors (exec, write, edit, etc.) always bypass this setting and emit ⚠️ warnings to the chat channel.

This is problematic for assistants that handle errors gracefully in-context — the user sees raw ⚠️ 🛠️ Exec: ... failed: ... messages even though the assistant already retried or handled the failure in the next turn.

Root Cause

In shouldShowToolErrorWarning():

function shouldShowToolErrorWarning(params) {
    // Mutating tools always show, regardless of config
    if (params.lastToolError.mutatingAction ?? 
        isLikelyMutatingToolName(params.lastToolError.toolName)) return true;
    if (params.suppressToolErrors) return false;
    return !params.hasUserFacingReply && !isRecoverableToolError(params.lastToolError.error);
}

The mutating check runs first and returns true unconditionally, ignoring suppressToolErrors.

Proposal

Extend suppressToolErrors to accept a string mode:

{
  "messages": {
    "suppressToolErrors": "all"  // true | false | "all"
  }
}

• false (default): current behavior
• true: suppress non-mutating tool errors (current behavior)
• "all": suppress ALL tool errors including mutating

This is backward-compatible — existing boolean configs work unchanged.

Use Case

Running an always-on personal assistant via Telegram. The assistant frequently runs shell commands that may fail on first attempt (SSH timeouts, git conflicts, permission issues) but handles errors by retrying or adjusting. The ⚠️ warnings clutter the Telegram chat with raw error output that the assistant has already addressed.

Current config:

{
  "tools": {
    "exec": {
      "notifyOnExit": false,
      "notifyOnExitEmptySuccess": false
    }
  },
  "messages": {
    "suppressToolErrors": true
  }
}

Despite all three settings, exec errors still leak through to the user.

Environment

• OpenClaw 2026.2.15
• Channel: Telegram (DM)