// Handle tool_use/tool_result mismatch errors (transcript corruption) if (/tool_use_id|tool_result.*corresponding.*tool_use|unexpected.*tool.*block/i.test(errorText)) { return { payloads: [ { text: "Conversation history corruption detected. " + "Please use /new to start a fresh session.", isError: true, }, ], meta: { durationMs: Date.now() - started, agentMeta: { sessionId: sessionIdUsed, provider, model: model.id, }, systemPromptReport: attempt.systemPromptReport, error: { kind: "tool_result_mismatch", message: errorText }, }, }; }

Fix Plan

To address the tool_use/tool_result mismatch error and prevent an infinite loop of error messages, we will implement the following steps:

• Modify the formatAssistantErrorText() function in errors.js to catch tool_use/tool_result mismatch errors and return a user-facing message suggesting the use of /new to start a fresh session.
• Update the sanitizeUserFacingText() function in errors.js to include the same pattern for consistency.
• Modify the promptError handler in run.js to handle tool_use/tool_result mismatch errors and return a payload with an actionable message.

Code Changes

// In errors.js - formatAssistantErrorText()
if (/tool_use_id|tool_result.*corresponding.*tool_use|unexpected.*tool.*block/i.test(raw)) {
    return ("Conversation history corruption detected. " +
        "Please use /new to start a fresh session.");
}

// In errors.js - sanitizeUserFacingText()
// Add the same pattern for consistency

// In run.js - promptError handler
if (/tool_use_id|tool_result.*corresponding.*tool_use|unexpected.*tool.*block/i.test(errorText)) {
    return {
        payloads: [
            {
                text: "Conversation history corruption detected. " +
                    "Please use /new to start a fresh session.",
                isError: true,
            },
        ],
        meta: {
            durationMs: Date.now() - started,
            agentMeta: {
                sessionId: sessionIdUsed,
                provider,
                model: model.id,
            },
            systemPromptReport: attempt.systemPromptReport,
            error: { kind: "tool_result_mismatch", message: errorText },
        },
    };
}

Verification

To verify that the fix worked, test the following scenarios:

• Trigger a tool_use/tool_result mismatch error and verify that the user-facing message suggests using /new to start a fresh session.
• Verify that the infinite loop of error messages is broken and the user can start a new session using /new.

Extra Tips

Consider implementing a more robust fix that attempts automatic transcript repair when this error is detected, such as calling repairToolUseResultPairing() and retrying. This would provide a better user experience and prevent the need for manual intervention.