diff --git a/CHANGELOG.md b/CHANGELOG.md index 876b76b..c257f92 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -15,6 +15,10 @@ - **Topbar border invisible in light mode** — added `:root:not(.dark)` border override. (PR #627) - **User message bubble text contrast** — accent-colored bubbles now use skin-aware text colors meeting WCAG AA (Poseidon dark improved from 2.8 → 6.5 ratio). (PR #627) - **Settings skin persistence race condition** — save now waits for server confirmation before applying. (PR #627) +## [v0.50.78] — 2026-04-17 + +### Fixed +- **Gemma 4 thinking tokens no longer shown raw in chat** — added `<|turn|>thinking\n...` to the streaming think-token parser in `static/messages.js` and `_strip_thinking_markup()` in `api/streaming.py`. Previously Gemma 4's reasoning output appeared as raw text prepended to the answer. (Closes #607) ## [v0.50.76] — 2026-04-17 diff --git a/api/streaming.py b/api/streaming.py index fb0eba7..63b6472 100644 --- a/api/streaming.py +++ b/api/streaming.py @@ -68,6 +68,7 @@ def _strip_thinking_markup(text: str) -> str: s = str(text) s = re.sub(r'.*?', ' ', s, flags=re.IGNORECASE | re.DOTALL) s = re.sub(r'<\|channel\|>thought.*?', ' ', s, flags=re.IGNORECASE | re.DOTALL) + s = re.sub(r'<\|turn\|>thinking\n.*?', ' ', s, flags=re.IGNORECASE | re.DOTALL) # Gemma 4 s = re.sub(r'^\s*(the|ther)\s+user\s+is\s+asking.*$', ' ', s, flags=re.IGNORECASE | re.MULTILINE) s = re.sub(r'\s+', ' ', s).strip() return s @@ -96,7 +97,7 @@ def _looks_invalid_generated_title(text: str) -> bool: if not s.strip(): return True return bool( - re.search(r'|<\|channel\|>thought', s, flags=re.IGNORECASE) + re.search(r'|<\|channel\|>thought|<\|turn\|>thinking', s, flags=re.IGNORECASE) or re.search(r'^\s*(the|ther)\s+user\s+', s, flags=re.IGNORECASE) or re.search(r'^\s*user\s+\w+\s+', s, flags=re.IGNORECASE) or re.search(r'\b(they|user)\s+want(s)?\s+me\s+to\b', s, flags=re.IGNORECASE) diff --git a/static/index.html b/static/index.html index d427e53..1d556cc 100644 --- a/static/index.html +++ b/static/index.html @@ -591,7 +591,7 @@
System
Instance version and access controls.
- v0.50.77 + v0.50.78
diff --git a/static/messages.js b/static/messages.js index f2e0022..b96ba5e 100644 --- a/static/messages.js +++ b/static/messages.js @@ -144,7 +144,8 @@ function attachLiveStream(activeSid, streamId, uploaded=[], options={}){ // Thinking tag patterns for streaming display const _thinkPairs=[ {open:'',close:''}, - {open:'<|channel>thought\n',close:''} + {open:'<|channel>thought\n',close:''}, + {open:'<|turn|>thinking\n',close:''} // Gemma 4 ]; function _isActiveSession(){ diff --git a/static/ui.js b/static/ui.js index a110bb3..391c43d 100644 --- a/static/ui.js +++ b/static/ui.js @@ -1095,7 +1095,7 @@ function _messageHasReasoningPayload(m){ if(!m||m.role!=='assistant') return false; if(m.reasoning) return true; if(Array.isArray(m.content)) return m.content.some(p=>p&&(p.type==='thinking'||p.type==='reasoning')); - return /[\s\S]*?<\/think>|<\|channel>thought\n[\s\S]*?/.test(String(m.content||'')); + return /[\s\S]*?<\/think>|<\|channel>thought\n[\s\S]*?|<\|turn\|>thinking\n[\s\S]*?/.test(String(m.content||'')); } function _assistantRoleHtml(tsTitle=''){ const _bn=window._botName||'Hermes'; @@ -1167,12 +1167,21 @@ function renderMessages(){ content=content.replace(/[\s\S]*?<\/think>\s*/,'').trimStart(); } if(!thinkingText){ + // Historical name "gemmaMatch" refers to MiniMax <|channel>thought format. const gemmaMatch=content.match(/<\|channel>thought\n([\s\S]*?)/); if(gemmaMatch){ thinkingText=gemmaMatch[1].trim(); content=content.replace(/<\|channel>thought\n[\s\S]*?\s*/,'').trimStart(); } } + if(!thinkingText){ + // Gemma 4 uses asymmetric <|turn|>thinking\n... delimiters. + const gemmaTurnMatch=content.match(/<\|turn\|>thinking\n([\s\S]*?)/); + if(gemmaTurnMatch){ + thinkingText=gemmaTurnMatch[1].trim(); + content=content.replace(/<\|turn\|>thinking\n[\s\S]*?\s*/,'').trimStart(); + } + } } const isUser=m.role==='user'; const isLastAssistant=!isUser&&vi===visWithIdx.length-1; diff --git a/tests/test_issue607.py b/tests/test_issue607.py new file mode 100644 index 0000000..15d0c4e --- /dev/null +++ b/tests/test_issue607.py @@ -0,0 +1,98 @@ +"""Tests for PR #648 — Gemma 4 thinking token stripping (closes #607).""" +import re +import pathlib +import pytest + + +# --------------------------------------------------------------------------- +# _strip_thinking_markup tests +# --------------------------------------------------------------------------- + +from api.streaming import _strip_thinking_markup, _looks_invalid_generated_title + + +class TestGemma4ThinkingTokenStrip: + """Verify that <|turn|>thinking\n...\n blocks are stripped.""" + + def test_strip_gemma4_basic(self): + """Basic Gemma 4 thinking block stripped, answer kept.""" + raw = "<|turn|>thinking\nSome internal reasoning\nFinal answer" + result = _strip_thinking_markup(raw) + assert result == "Final answer" + + def test_strip_gemma4_multiline_reasoning(self): + """Multi-line reasoning block stripped cleanly.""" + raw = "<|turn|>thinking\nLine 1\nLine 2\nLine 3\nAnswer here" + result = _strip_thinking_markup(raw) + assert result == "Answer here" + + def test_strip_gemma4_no_thinking_passthrough(self): + """Normal response without thinking tokens passes through unchanged.""" + raw = "Normal response without thinking tokens" + result = _strip_thinking_markup(raw) + assert result == raw + + def test_strip_gemma4_with_leading_whitespace(self): + """Leading whitespace before the thinking block is handled.""" + raw = "\n\n<|turn|>thinking\nReasoning\nAnswer" + result = _strip_thinking_markup(raw) + assert result == "Answer" + + def test_strip_gemma4_empty_reasoning(self): + """Empty reasoning block (just delimiters) is stripped.""" + raw = "<|turn|>thinking\nResponse" + result = _strip_thinking_markup(raw) + assert result == "Response" + + def test_strip_gemma4_case_insensitive(self): + """Pattern is case-insensitive (though Gemma 4 uses fixed case).""" + raw = "<|TURN|>THINKING\nreasoning\nanswer" + result = _strip_thinking_markup(raw) + # The regex uses re.IGNORECASE — should strip uppercase variant too + assert "THINKING" not in result + assert "reasoning" not in result + + def test_existing_think_tag_still_works(self): + """Ensure ... still stripped (no regression).""" + raw = "inner reasoningFinal" + result = _strip_thinking_markup(raw) + assert result == "Final" + + def test_existing_channel_tag_still_works(self): + """Ensure <|channel|>thought... still stripped.""" + raw = "<|channel|>thoughtSome reasoningAnswer" + result = _strip_thinking_markup(raw) + assert result == "Answer" + + +class TestGemma4TitleLeakDetection: + """Verify _looks_invalid_generated_title catches Gemma 4 leak.""" + + def test_detects_gemma4_leak_in_title(self): + raw = "<|turn|>thinking\nUser asked about X\nSession Title" + assert _looks_invalid_generated_title(raw) is True + + def test_clean_title_not_flagged(self): + assert _looks_invalid_generated_title("Python debugging session") is False + + +class TestGemma4MessagesJsThinkPairs: + """Verify static/messages.js contains the correct Gemma 4 pair.""" + + def test_messages_js_has_correct_gemma4_open(self): + js = pathlib.Path("static/messages.js").read_text() + # Must have double-pipe format: <|turn|>thinking + assert "<|turn|>thinking" in js, ( + "messages.js is missing correct Gemma 4 open delimiter '<|turn|>thinking'" + ) + + def test_messages_js_no_wrong_gemma4_open(self): + js = pathlib.Path("static/messages.js").read_text() + # Must NOT have single-pipe wrong format: <|turn>thinking + assert "<|turn>thinking" not in js, ( + "messages.js still contains wrong Gemma 4 delimiter '<|turn>thinking' (missing |)" + ) + + def test_messages_js_has_gemma4_close(self): + js = pathlib.Path("static/messages.js").read_text() + assert "" in js, "messages.js missing Gemma 4 close delimiter ''" diff --git a/tests/test_sprint38.py b/tests/test_sprint38.py index 21e9a5d..020574f 100644 --- a/tests/test_sprint38.py +++ b/tests/test_sprint38.py @@ -57,6 +57,35 @@ def test_gemma_content_removal_uses_replace_not_slice(): "ui.js must call .trimStart() on content after removing the Gemma channel block" +def test_gemma_turn_regex_in_ui_js(): + """The Gemma 4 <|turn|>thinking\\n... pattern must be extracted from persisted content.""" + # Detection in _messageHasReasoningPayload (correct double-pipe format) + assert "<\\|turn\\|>thinking" in UI_JS, ( + "ui.js _messageHasReasoningPayload must detect Gemma 4 <|turn|>thinking\\n... pattern" + " (note: double-pipe: <|turn|> not <|turn>)" + ) + # Extraction block + match = re.search(r'const gemmaTurnMatch=content\.match\((/[^/]+/)\)', UI_JS) + assert match, "gemmaTurnMatch line not found in ui.js" + pattern = match.group(1) + assert not pattern.startswith('/^'), ( + f"gemmaTurnMatch regex must not use ^ anchor — got {pattern}" + ) + + +def test_gemma_turn_content_removal_uses_replace_not_slice(): + """Gemma 4 turn token removal must use .replace() not .slice().""" + idx = UI_JS.find("if(gemmaTurnMatch){") + assert idx >= 0, "gemmaTurnMatch handler block not found in ui.js" + block = UI_JS[idx:idx+240] + assert "content.replace(" in block, ( + "ui.js must use content.replace() to remove Gemma 4 turn block (not .slice())" + ) + assert ".trimStart()" in block, ( + "ui.js must call .trimStart() on content after removing the Gemma 4 turn block" + ) + + # ── messages.js: streaming render path ─────────────────────────────────────── def test_stream_display_trims_before_startswith():