fix: DCP_PAIRED_TAG_REGEX regex bug causes tag fragment leakage into chat

[mengfanbo123]

Bug

After multiple compression rounds, malformed XML tag fragments like </d-message-id> and stale message IDs (e.g. m00097, m00099) leak into the user-visible chat dialog.

Root Cause

lib/messages/utils.ts:14 — DCP_PAIRED_TAG_REGEX:

- const DCP_PAIRED_TAG_REGEX = /]*>[\s\S]*?<\/(?:dcp|acp)[^>]*/gi
+ const DCP_PAIRED_TAG_REGEX = /<(?:dcp|acp)[^>]*>[\s\S]*?<\/(?:dcp|acp)[^>]*/gi

The buggy regex starts with ]*> which matches a literal ] character — missing the < and tag name. As a result, stripHallucinationsFromString only partially deletes paired tags: it removes the closing tag + content but leaves behind opening-tag fragments that then leak into the chat.

Call Path

stripHallucinations (lib/messages/utils.ts:265) → called globally at lib/hooks.ts:246. stripHallucinationsFromString (utils.ts:261) runs the PAIRED regex first (which misses due to the bug), then the UNPAIRED regex /<\/?(?:dcp|acp)[^>]*/gi to clean up residual tag fragments.

Fix

Add <(?:dcp|acp)[^>]*> to correctly match the opening tag. Now the PAIRED regex matches the full open+content+close span and removes it in one pass.

Verification

• node --import tsx --test tests/strip-stale-compress.test.ts tests/inject.test.ts tests/message-utils.test.ts → 32/32 pass
• Reproduced the leak locally before fix; confirmed clean after fix.

Related

• Closes [Bug]: stripHallucinations 正则错误导致 dcp-message-id 标签碎片泄漏到对话框 #123
• Note: lib/messages/utils.ts:11 DCP_BLOCK_ID_TAG_REGEX = /(])[^>]*>)b\d+(...)/g has a similar ] literal issue — suggest review in a follow-up.

[ranxianglei]

缺少devlog和tests因此ci无法通过

[ranxianglei]

Thanks for the fix! Your root cause analysis for DCP_PAIRED_TAG_REGEX (line 14) was spot-on.

While reviewing, I found that lines 11 (DCP_BLOCK_ID_TAG_REGEX) and 13 (DCP_MESSAGE_REF_TAG_REGEX) have the same class of bug. I've created PR #130 which includes your fix plus the two additional regex fixes, with 23 dedicated tests.

Closing this in favor of #130.