Bug Description

User and assistant messages containing certain phrases (like the two words "context" + "overflow" together) are incorrectly passed through sanitizeUserFacingText(), which rewrites them to error messages.

Steps to Reproduce

1. Start a fresh session with /new
2. Type a message containing "context" followed by "overflow" (as one phrase)
3. Observe that the message is replaced with: "Context overflow: prompt too large for the model. Try again with less input or a larger-context model."

This happens even with minimal context usage (e.g., 8-15% of context window).

Expected Behavior

User messages should pass through unchanged. The sanitizeUserFacingText() function should only be applied to actual API error responses, not user/assistant chat messages.

Actual Behavior

Any message containing error-pattern phrases gets rewritten. This includes:

• User input messages
• Assistant responses (if they mention the phrase)

The bug affects both directions of chat.

Root Cause

In dist/agents/pi-embedded-helpers/errors.js:

export function sanitizeUserFacingText(text) {
    // ...
    if (isContextOverflowError(trimmed)) {
        return ("Context overflow: prompt too large for the model. " +
            "Try again with less input or a larger-context model.");
    }
    // ...
}

And isContextOverflowError() matches on lower.includes("context overflow").

This function is being called somewhere in the message flow where it processes regular chat messages, not just error payloads.

Suggested Fix

Either:

1. Add a flag/context to sanitizeUserFacingText() to distinguish error payloads from regular messages
2. Or ensure this function is only called on actual API error responses, not on user/assistant message content

Environment

• Clawdbot version: npm latest 2026.1.24-3
• Model: claude-opus-4-5
• Channel: Telegram

Workaround

Avoid typing the exact phrase in messages until fixed.