MrSphay/odysseus
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: normalize Gemma 4 thought-channel output (#2224)

Browse Source
This commit is contained in:
RaresKeY
2026-06-04 20:26:58 +03:00
committed by GitHub
parent 7ce6ec7f50
commit c12c2aa233
7 changed files with 249 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
static/js/chat.js
View File

@@ -1120,7 +1120,7 @@ import createResearchSynapse from './researchSynapse.js';
let _measureDiv = null;
function _replyAfterClosedThinking(text) {
const closeRe = /<\/think(?:ing)?>/gi;
const closeRe = /<\/(?:think(?:ing)?|thought)>|<channel\|>/gi;
let match = null;
let last = null;
while ((match = closeRe.exec(text || '')) !== null) last = match;
@@ -1147,7 +1147,7 @@ import createResearchSynapse from './researchSynapse.js';
replyTrimmed = (replyText || '').trim();
} else {
// Non-tag: check for garbled <think> (reasoning\n<think>reply)
const _gm = dt.match(/^[\s\S]+?<think(?:ing)?>\s*([\s\S]*?)(?:<\/think(?:ing)?>)?\s*$/i);
const _gm = dt.match(/^[\s\S]+?<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*([\s\S]*?)(?:<\/(?:think(?:ing)?|thought)>)?\s*$/i);
if (_gm && _gm[1].trim()) {
replyTrimmed = _gm[1].trim();
} else {
@@ -1188,8 +1188,11 @@ import createResearchSynapse from './researchSynapse.js';
const prevLen = contentEl._prevTextLen || 0;
// If thinking is still streaming (unclosed <think>), show indicator instead of raw text
if (markdownModule.hasUnclosedThinkTag && markdownModule.hasUnclosedThinkTag(dt)) {
const thinkStart = dt.search(/<think(?:ing)?>/i);
const thinkContent = dt.substring(thinkStart).replace(/<think(?:ing)?>/i, '').trim();
const thinkStart = dt.search(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>|<\|channel>thought/i);
const thinkContent = dt.substring(Math.max(thinkStart, 0))
.replace(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>|<\|channel>thought\s*\n?/i, '')
.replace(/<channel\|>/gi, '')
.trim();
const lines = thinkContent.split('\n').length;
// Don't show beforeThink text during streaming — it'll appear in the final render
// This prevents the "split into two" duplication
@@ -1449,7 +1452,7 @@ import createResearchSynapse from './researchSynapse.js';
// Detect non-tag thinking patterns: "Thinking:", "Thinking Process:", Gemma-style reasoning
// These patterns don't use <think> tags, so we simulate unclosed thinking during streaming
const _replyPrefixes = ['Hey', 'Hi ', 'Hi!', 'Hello', 'Sure', 'Yes', 'No ', 'No,', 'Yo', 'OK', 'Here', 'Absolutely', 'Of course', 'Great', 'Alright', 'Thanks', 'Welcome', 'Good ', "I'm happy", "I'd be"];
if (!hasUnclosedThink && !roundText.includes('<think')) {
if (!hasUnclosedThink && !/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>|<\|channel>thought/i.test(roundText)) {
const _trimmedRT = roundText.trimStart();
const _isReasoning = markdownModule.startsWithReasoningPrefix(_trimmedRT);
if (_isReasoning) {
@@ -1475,10 +1478,10 @@ import createResearchSynapse from './researchSynapse.js';
}
}
}
if (!hasUnclosedThink && /^<think(?:ing)?>\s*<\/think(?:ing)?>/i.test(roundText)) {
if (!hasUnclosedThink && /^<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*<\/(?:think(?:ing)?|thought)>/i.test(roundText)) {
// Empty <think></think> — the model likely put thinking outside the tags
const afterEmpty = roundText.replace(/^<think(?:ing)?>\s*<\/think(?:ing)?>/i, '').trim();
const closeTags = (afterEmpty.match(/<\/think(?:ing)?>/gi) || []).length;
const afterEmpty = roundText.replace(/^<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*<\/(?:think(?:ing)?|thought)>/i, '').trim();
const closeTags = (afterEmpty.match(/<\/(?:think(?:ing)?|thought)>/gi) || []).length;
if (closeTags === 0 && afterEmpty.length > 0) {
hasUnclosedThink = true; // still waiting for real closing tag
}
@@ -1487,13 +1490,13 @@ import createResearchSynapse from './researchSynapse.js';
// Only applies when there's a second </think> later (model leaked thinking outside tags)
// Do NOT trigger if the text after </think> contains tool calls (that's real content)
if (!hasUnclosedThink && isThinking) {
const _thinkMatch = roundText.match(/<think(?:ing)?>([\s\S]*?)<\/think(?:ing)?>/i);
const _thinkMatch = roundText.match(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>([\s\S]*?)<\/(?:think(?:ing)?|thought)>/i);
const _thinkLen = _thinkMatch ? _thinkMatch[1].trim().length : 0;
if (_thinkLen < 20) {
const _afterClose = roundText.replace(/<think(?:ing)?>([\s\S]*?)<\/think(?:ing)?>/i, '').trim();
const _afterClose = roundText.replace(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>([\s\S]*?)<\/(?:think(?:ing)?|thought)>/i, '').trim();
// Only keep waiting if there's trailing text that looks like thinking (not tool calls)
const _hasToolCall = /```(?:bash|python|web_search|read_file|write_file|create_document|edit_document|manage_|generate_image)/i.test(_afterClose);
const _hasOrphanClose = /<\/think(?:ing)?>/i.test(_afterClose);
const _hasOrphanClose = /<\/(?:think(?:ing)?|thought)>/i.test(_afterClose);
if (!_hasToolCall && (_hasOrphanClose || (Date.now() - thinkingStartTime) < 500)) {
hasUnclosedThink = true; // keep waiting for real </think>
}
@@ -1550,8 +1553,12 @@ import createResearchSynapse from './researchSynapse.js';
}
} else if (hasUnclosedThink && isThinking) {
if (_liveThinkInner) {
// Extract raw thinking text (strip all <think>/<thinking> open/close tags and prefixes)
var thinkText = roundText.replace(/<\/?think(?:ing)?>/gi, '');
// Extract raw thinking text (strip known thinking wrappers and prefixes)
var thinkText = roundText
.replace(/<\/?(?:think(?:ing)?|thought)(?:\s+[^>]*)?>/gi, '')
.replace(/<\|channel>thought\s*\n?/gi, '')
.replace(/<\|channel>response\s*\n?/gi, '')
.replace(/<channel\|>/gi, '');
thinkText = thinkText.replace(/^\s*Thinking(?:\s+Process)?:\s*/i, '');
_liveThinkInner.innerHTML = markdownModule.mdToHtml(thinkText);
// Keep thinking box scrolled to bottom
@@ -2402,8 +2409,8 @@ import createResearchSynapse from './researchSynapse.js';
_finalReply = (_extracted.content || '').trim();
} else {
// Non-tag thinking: extract reply from raw text
// Handle garbled <think> tag: "Thinking: reasoning\n<think>reply"
const _garbledMatch = finalDisplay.match(/^[\s\S]+?<think(?:ing)?>\s*([\s\S]*?)(?:<\/think(?:ing)?>)?\s*$/i);
// Handle garbled thinking tag: "Thinking: reasoning\n<think>reply"
const _garbledMatch = finalDisplay.match(/^[\s\S]+?<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*([\s\S]*?)(?:<\/(?:think(?:ing)?|thought)>)?\s*$/i);
if (_garbledMatch && _garbledMatch[1].trim()) {
_finalReply = _garbledMatch[1].trim();
} else {
@@ -2452,8 +2459,8 @@ import createResearchSynapse from './researchSynapse.js';
_body4b.innerHTML = _sourcesData ? _buildSourcesBox(_sourcesData, _sourcesType, _wasExpanded2) : _sourcesHtml;
} else if (roundHolder !== holder) {
// Check if there's thinking content worth showing
const _thinkMatch = roundText.match(/<think(?:ing)?>([\s\S]*?)<\/think(?:ing)?>/i);
if (_thinkMatch && _thinkMatch[1].trim()) {
const _thinkingOnly = markdownModule.extractThinkingBlocks(roundText);
if (_thinkingOnly.thinkingBlocks?.length && !_thinkingOnly.content) {
// Show thinking in a collapsed section even if no visible reply text
const _body4c = roundHolder.querySelector('.body');
if (_body4c) _body4c.innerHTML = markdownModule.processWithThinking(roundText);
@@ -4534,9 +4541,10 @@ import createResearchSynapse from './researchSynapse.js';
// never closes (so it would otherwise hide the whole answer). Peel all of
// those off so what's left is just the rewritten text.
const _stripThink = (t) => {
t = t.replace(/<think>[\s\S]*?<\/think>/gi, ''); // complete blocks
if (/<\/think>/i.test(t)) t = t.replace(/^[\s\S]*?<\/think>/i, ''); // reasoning w/o opener
return t.replace(/<\/?think>/gi, '').trim(); // any orphan tag
t = markdownModule.normalizeThinkingMarkup(t || '');
t = t.replace(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>[\s\S]*?<\/(?:think(?:ing)?|thought)>/gi, ''); // complete blocks
if (/<\/(?:think(?:ing)?|thought)>/i.test(t)) t = t.replace(/^[\s\S]*?<\/(?:think(?:ing)?|thought)>/i, ''); // reasoning w/o opener
return t.replace(/<\/?(?:think(?:ing)?|thought)(?:\s+[^>]*)?>/gi, '').trim(); // any orphan tag
};
newText = _stripThink(newText);

47
static/js/markdown.js
View File

@@ -116,8 +116,13 @@ function sanitizeAllowedHtml(html) {
* Check if text has unclosed think tag
*/
export function hasUnclosedThinkTag(text) {
const openCount = (text.match(/<think(?:ing)?>/gi) || []).length;
const closeCount = (text.match(/<\/think(?:ing)?>/gi) || []).length;
text = text || '';
const openCount =
(text.match(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>/gi) || []).length
+ (text.match(/<\|channel>thought/gi) || []).length;
const closeCount =
(text.match(/<\/(?:think(?:ing)?|thought)>/gi) || []).length
+ (text.match(/<channel\|>/gi) || []).length;
return openCount > closeCount;
}
@@ -125,8 +130,25 @@ export function startsWithReasoningPrefix(text) {
return /^\s*(?:thinking(?:\s+process)?\s*:|the user |i need |i should |i will |they are |the question |i can )/i.test(text || '');
}
export function normalizeThinkingMarkup(text) {
if (!text) return text;
let normalized = text;
normalized = normalized.replace(/<thought(\s+[^>]*)?>/gi, (_m, attrs = '') => `<think${attrs || ''}>`);
normalized = normalized.replace(/<\/thought>/gi, '</think>');
normalized = normalized.replace(/<\|channel>thought\s*\n?([\s\S]*?)<channel\|>\s*/gi, (_m, content = '') => {
const thought = String(content || '').trim();
return thought ? `<think>${thought}</think>\n` : '';
});
normalized = normalized.replace(/<\|channel>response\s*\n?([\s\S]*?)<channel\|>/gi, (_m, content = '') => content || '');
normalized = normalized.replace(/<\|channel>response\s*\n?/gi, '');
normalized = normalized.replace(/<channel\|>/gi, '');
return normalized;
}
function normalizePlainThinking(text) {
if (!text || /<think/i.test(text)) return text;
if (!text) return text;
text = normalizeThinkingMarkup(text);
if (/<think/i.test(text)) return text;
const trimmed = text.trimStart();
if (!startsWithReasoningPrefix(trimmed)) return text;
@@ -220,11 +242,21 @@ export function extractThinkingBlocks(text) {
// (b) Cut-off mid-generation — there's already real reply text before the
// opener. Drop from the tag onward as before (it's truncated thinking).
if (hasUnclosedThinkTag(normalized)) {
const strayOpener = cleanContent.match(/^\s*<think(?:ing)?(?:\s+[^>]*)?>([\s\S]*)$/i);
if (strayOpener) {
cleanContent = strayOpener[1];
const gemmaThoughtStart = cleanContent.search(/<\|channel>thought/i);
if (gemmaThoughtStart >= 0) {
const leakedThought = cleanContent
.slice(gemmaThoughtStart)
.replace(/^<\|channel>thought\s*\n?/i, '')
.trim();
if (gemmaThoughtStart === 0 && leakedThought) thinkingBlocks.push(leakedThought);
cleanContent = cleanContent.slice(0, gemmaThoughtStart);
} else {
cleanContent = cleanContent.replace(/<think(?:ing)?(?:\s+[^>]*)?>[\s\S]*$/gi, '');
const strayOpener = cleanContent.match(/^\s*<think(?:ing)?(?:\s+[^>]*)?>([\s\S]*)$/i);
if (strayOpener) {
cleanContent = strayOpener[1];
} else {
cleanContent = cleanContent.replace(/<think(?:ing)?(?:\s+[^>]*)?>[\s\S]*$/gi, '');
}
}
}
@@ -686,6 +718,7 @@ const markdownModule = {
createCollapsible,
hasUnclosedThinkTag,
extractThinkingBlocks,
normalizeThinkingMarkup,
startsWithReasoningPrefix,
renderMermaid
};