Repeated identical local tool calls should be blocked to prevent infinite reasoning loops

[andrew-ellis-engineering]

Summary

Nanobot currently guards repeated external lookups (for example web_fetch and web_search) but does not guard repeated identical local tool calls in the same turn.

In practice, this can produce loops where the model reissues the same local call with unchanged arguments until cancellation or max_iterations.

Reproduction Pattern

Examples of repeated identical local calls:

• read_file(path=...)
• list_dir(path=...)
• glob(pattern=...)
• grep(pattern=...)

The same (tool_name, args) signature can repeat many times with no new signal.

Expected Behavior

Nanobot should short-circuit repeated identical local calls in a single turn after a small retry budget.

Proposed Scope

Initial guarded tools:

• read_file
• list_dir
• glob
• grep

Behavior:

1. Normalize a signature from tool_name + normalized args
2. Count repeats within one turn
3. Allow a small retry budget (for example 2-3)
4. After threshold, return synthetic tool error:
   Error: repeated identical tool call blocked. Use the prior result or choose a different tool.
5. Emit warning log with tool, abbreviated signature, attempt count

Suggested Implementation Area

• nanobot/agent/runner.py
• optionally shared helper in nanobot/utils/runtime.py

Test Expectations

• Repeated identical read_file is blocked after threshold
• Repeated identical glob is blocked after threshold
• Repeated identical grep is blocked after threshold
• Different arguments remain allowed
• Existing external lookup guard behavior remains unchanged

Notes

This issue is intended as a general core safety fix, not workflow-specific behavior.

[chengyongru]

Closing — the existing max_iterations cap and repeated_external_lookup_error safeguard already address runaway tool-call loops. There are no user reports of infinite local-tool-call loops being a practical problem. If this becomes a real pain point with concrete reproduction steps, happy to revisit.

[andrew-ellis-engineering]

This is an issue I personally am facing, although it's a limitation to the model I am using more so than nanobot.

I figured since this is meant to be a super lightweight agent that others may struggle with the same issues as me on locally hosted models, so I thought I would open this issue.

Completely understand if it is not relevant or a priority, just wanted to share! Thanks for reviewing!