stuck in a loop

[hadronai]

The harness constantly gets stuck in loops searching local files or the internet, repeating the same searches over and over. Is anyone else experiencing this?

[slyyr]

I'd first determine whether the loop is caused by the model repeatedly choosing the same tool call or by the harness failing to preserve the result of a previous tool call.

A useful debugging procedure would be:

1. Record each tool invocation, its arguments, and its returned result.
2. Check whether identical searches are being issued with identical arguments.
3. Verify that successful tool results are added to the conversation history before the next model call.
4. Check whether the agent's stop condition and maximum-iteration limit are working.
5. Repeat the task with a smaller context and a simple, deterministic tool.

If the same tool call is repeated despite a successful result being available in the history, that points toward an agent-planning or prompt issue. If the result is missing from the history, I'd investigate tool-result handling.

As a temporary safeguard, a maximum-step limit and detection of repeated tool-call signatures can prevent runaway loops.

For a reproducible bug report, the most useful evidence would be a short trace showing the model's tool call, the tool result, and the next model response.

[elladeniya-dev]

Yes, this happens when the tool-call output exceeds the truncation threshold or returns an empty/ambiguous match, causing the model to lose track of previous search parameters in its context window and re-issue the exact same file/web search.

How to Stop the Search Loop:

1. Switch or Reset the Session Context: Once a session has 3+ identical tool calls in its message history, the model tends to pattern-match its own prior tool calls. Compact/clear the session history or start a fresh thread.
2. Constrain Search Scope in Your Prompt: Explicitly specify the directory path (e.g., Look only inside src/components) or pass the target file directly with @filename so the agent skips broad recursive grep/glob searches.
3. Avoid Mid-Session Model Switching: Switching between flash/lightweight models mid-session can cause tool-call state mismatches that trigger repeated search retries.

If this helps, feel free to click Mark as answer!