codex - 💡(How to fix) Fix Agent ignores explicit tool restriction and loops on exec_command no-ops [1 comments, 2 participants]

TL;DR

The agent's tool-selection loop can be mitigated by ensuring explicit user corrections are properly handled and prioritized over default or cached tool selections.

Guidance

• Review the agent's logic for handling user corrections, particularly for tool selection, to identify potential flaws in updating the intended tool usage.
• Verify that the exec_command calls are correctly overridden or skipped when a user explicitly instructs the agent to use a specific tool, such as Chrome MCP.
• Investigate if there's a caching or state management issue that causes the agent to revert to default or previously used tools despite user corrections.
• Consider implementing a more robust feedback mechanism for the agent to acknowledge and adapt to user corrections, ensuring it either complies with the instruction or reports its inability to do so.

Example

No specific code example can be provided without more context on the agent's implementation, but ensuring that user corrections are handled as high-priority events that update the agent's state regarding tool selection is crucial.

Notes

The exact fix may depend on the internal implementation of the agent and its tool selection mechanism, which is not detailed in the provided information. This guidance assumes a potential issue with how user corrections are integrated into the agent's decision-making process.

Recommendation

Apply a workaround by enhancing the agent's logic to prioritize and correctly handle user corrections regarding tool selection, ensuring that it either uses the specified tool or stops and reports its inability to comply. This is recommended because it directly addresses the observed mismatch between the agent's declared intent and its actual actions.