Mintplex-Labs/anything-llm · warning

[loop-detect] Suspected loop: ${reason}

Error message

[loop-detect] Suspected loop: ${reason}

What it means

The loop-detect heuristic decided the agent is stuck repeating the same behavior (identical tool calls/results). alertLoopingAgent injects a STOP nudge into the proxy interrupt queue (or the pi RPC channel) telling the agent to change approach, and broadcasts an ask_for_help event so the user can intervene; it is rate-limited by pendingHelp and alertedAt so it does not spam.

Source

Thrown at open-computer/services/interface-service/loop-detect/index.js:135

}

/** Call whenever a text message turn arrives (increments no-tool-turn counter). */
function onTextMessage() {
  LOOP_DETECT.turnsSinceToolCall++;
}

// ─── Intervention ─────────────────────────────────────────────────────────

/**
 * Injects a stop nudge into the LLM proxy interrupt queue and, if the proxy
 * is not in use, also writes directly to the pi RPC channel.  Surfaces an
 * ask_for_help event so the user can intervene if the nudge fails.
 */
function alertLoopingAgent(reason) {
  if (!workspace.piProcess) return;
  if (workspace.pendingHelp) return;

  console.warn(`[loop-detect] Suspected loop: ${reason}`);
  LOOP_DETECT.alertedAt = Date.now();

  const nudge =
    `STOP. You are stuck in a loop: ${reason}. ` +
    `The same tool calls are returning identical results. ` +
    `Do NOT repeat the same action. Try a completely different approach: ` +
    `use a different tool, navigate to a different URL, use curl instead of the browser, ` +
    `or move on to the next step of the task.`;

  if (settings.LLM_PROXY_TARGET) {
    setPendingInterrupt(nudge);
  }

  if (workspace.piRpc) {
    workspace.piRpc.write(
      JSON.stringify({
        id: `loop-nudge-${Date.now()}`,
        type: "prompt",

View on GitHub (pinned to 3aec848f28)

Solutions

1. Watch for the ask_for_help event and intervene manually with a hint or by taking over the session.
2. Improve the agent's task prompt to enumerate fallback strategies (different tool, curl instead of browser, skip step).
3. Raise temperature or diversity settings so the model explores alternative actions.
4. Fix the underlying blocker (login, selector, network) that produces identical results each attempt.

When it happens

Trigger: Agent retrying the same failing selector or navigation because the page never changes; identical API/tool responses each turn; a task whose next step keeps erroring so the model re-issues the same call; prompt that strongly biases one tool.

Common situations: Scraping targets that require login the agent cannot complete; flaky selectors; agent temperature too low so it repeats the first plausible action.

AI-assisted analysis of Mintplex-Labs/anything-llm@3aec848f28 (2026-08-18). Data as JSON: /api/errors/0fa770544e46564b. Report an issue: GitHub.