Bug Description

sanitizeUserFacingText() runs on all outgoing messages to external channels (Feishu, Telegram, Slack, etc.) and incorrectly replaces normal AI responses with error messages when the response text contains certain keywords.

Reproduction Steps

1. Send a message to the assistant asking about "context overflow" errors
2. The assistant's response mentioning "context overflow" is replaced with:

   Context overflow: prompt too large for the model. Try again with less input or a larger-context model.
   

Root Cause

In sanitizeUserFacingText(), the function calls isContextOverflowError(trimmed) which matches any text containing keywords like:

• context overflow
• timeout / timed out
• maximum context length
• prompt is too long
• etc.

This is problematic because:

1. WebChat receives messages directly from emitAgentEvent and bypasses sanitizeUserFacingText()
2. External channels (Feishu, Telegram, etc.) use createReplyDispatcher which calls normalizeReplyPayload → sanitizeUserFacingText()
3. Result: Same message appears correctly in WebChat but is replaced with error text in external channels

Affected Keywords (likely to appear in normal conversation)

Suggested Fix

Similar to PR #12226 which removed isBillingErrorMessage() from sanitizeUserFacingText(), the context overflow and other error pattern checks should also be removed from this function.

Real errors are already caught by formatAssistantErrorText() which correctly gates on stopReason === "error" before checking these patterns.

The sanitizeUserFacingText() function should only:

1. Strip <final> tags
2. Collapse duplicate blocks
3. NOT perform error pattern matching on normal assistant responses

Environment

• OpenClaw version: 2026.2.1 (also verified in 2026.2.6-3)
• Channel: Feishu
• WebChat shows correct response, Feishu shows error replacement