fix: strip all reasoning blocks in removeReasoningContent, not just the first

web-app/src/utils/__tests__/reasoning.test.ts

@@ -0,0 +1,82 @@
+import { describe, it, expect } from 'vitest'
+import {
+  removeReasoningContent,
+  extractReasoningFromMessage,
+} from '../reasoning'
+
+// Built with escapes so the tag characters can never be mangled.
+const T_OPEN = '\u003Ct\u0068ink\u003E' //  thinking
+const T_CLOSE = '\u003C/t\u0068ink\u003E' //  response
+
+describe('removeReasoningContent', () => {
+  it('strips a single complete think block', () => {
+    const input = `${T_OPEN}reasoning content${T_CLOSE}\nFinal answer is 2.`
+    expect(removeReasoningContent(input)).toBe('Final answer is 2.')
+  })
+
+  it('strips every complete think block, not just the first', () => {
+    const input =
+      `Question\n${T_OPEN}First reasoning pass${T_CLOSE}` +
+      `Second reasoning begins\n${T_OPEN}Second reasoning pass${T_CLOSE}\nFinal answer is 2.`
+    expect(removeReasoningContent(input)).toBe(
+      'Question\nSecond reasoning begins\n\nFinal answer is 2.'
+    )
+  })
+
+  it('strips an unterminated trailing think block', () => {
+    const input = `Question\n${T_OPEN}Reasoning with no closing tag.`
+    expect(removeReasoningContent(input)).toBe('Question')
+  })
+
+  it('strips multiple unterminated think blocks', () => {
+    const input = `${T_OPEN}first${T_CLOSE}\n${T_OPEN}second without close`
+    expect(removeReasoningContent(input)).toBe('')
+  })
+
+  it('returns content unchanged when there is no reasoning', () => {
+    const input = 'Just a normal answer with no think tags.'
+    expect(removeReasoningContent(input)).toBe(input)
+  })
+
+  it('strips a single complete DeepSeek analysis block', () => {
+    const input =
+      '\u003C|channel|>analysis\u003C|message|>reasoning\u003C|start|>assistant\u003C|channel|>final\u003C|message|>\nFinal answer is 2.'
+    expect(removeReasoningContent(input)).toBe('Final answer is 2.')
+  })
+
+  it('strips multiple DeepSeek analysis blocks, not just the first', () => {
+    const input =
+      'Question\n\u003C|channel|>analysis\u003C|message|>first\u003C|start|>assistant\u003C|channel|>final\u003C|message|>' +
+      'Between\n\u003C|channel|>analysis\u003C|message|>second\u003C|start|>assistant\u003C|channel|>final\u003C|message|>\nFinal answer is 2.'
+    expect(removeReasoningContent(input)).toBe(
+      'Question\nBetween\n\nFinal answer is 2.'
+    )
+  })
+
+  it('strips an unterminated DeepSeek analysis block', () => {
+    const input =
+      'Question\n\u003C|channel|>analysis\u003C|message|>no closing tag'
+    expect(removeReasoningContent(input)).toBe('Question')
+  })
+})
+
+describe('extractReasoningFromMessage', () => {
+  it('returns reasoning_content when present', () => {
+    const msg = { role: 'assistant', content: 'hi', reasoning_content: 'think' }
+    expect(extractReasoningFromMessage(msg as never)).toBe('think')
+  })
+
+  it('falls back to reasoning field', () => {
+    const msg = { role: 'assistant', content: 'hi', reasoning: 'think' }
+    expect(extractReasoningFromMessage(msg as never)).toBe('think')
+  })
+
+  it('returns null when no reasoning present', () => {
+    const msg = { role: 'assistant', content: 'hi' }
+    expect(extractReasoningFromMessage(msg as never)).toBeNull()
+  })
+
+  it('returns null for null message', () => {
+    expect(extractReasoningFromMessage(null as never)).toBeNull()
+  })
+})

web-app/src/utils/reasoning.ts

@@ -20,24 +20,27 @@ function getReasoning(
  * @returns
  */
 export function removeReasoningContent(content: string): string {
-  // Reasoning content should not be sent to the model
-  if (content.includes('<think>')) {
-    const match = content.match(/<think>([\s\S]*?)<\/think>/)
-    if (match?.index !== undefined) {
-      const splitIndex = match.index + match[0].length
-      content = content.slice(splitIndex).trim()
-    }
-  }
-  if (content.includes('<|channel|>analysis<|message|>')) {
-    const match = content.match(
-      /<\|channel\|>analysis<\|message\|>([\s\S]*?)<\|start\|>assistant<\|channel\|>final<\|message\|>/
-    )
-    if (match?.index !== undefined) {
-      const splitIndex = match.index + match[0].length
-      content = content.slice(splitIndex).trim()
-    }
-  }
-  return content
+  // Reasoning content should not be sent to the model.
+  let result = content
+
+  // Strip every complete <think>…</think> block. A model can emit several
+  // reasoning spans across one turn, so this has to be global — the previous
+  // code only ever removed the first block.
+  result = result.replace(/<think>([\s\S]*?)<\/think>/g, '')
+
+  // Strip an unterminated <think>… block (streaming cut off before the
+  // closing tag) so partial reasoning still doesn't leak into the prompt.
+  result = result.replace(/<think>[\s\S]*$/, '')
+
+  // Same treatment for the DeepSeek <|channel|>analysis<|message|>…
+  // reasoning format.
+  result = result.replace(
+    /<\|channel\|>analysis<\|message\|>([\s\S]*?)<\|start\|>assistant<\|channel\|>final<\|message\|>/g,
+    ''
+  )
+  result = result.replace(/<\|channel\|>analysis<\|message\|>[\s\S]*$/, '')
+
+  return result.trim()
 }
 
 // Extract reasoning from a message (for completed responses)
@@ -48,4 +51,4 @@ export function extractReasoningFromMessage(
 
   const extendedMessage = message as chatCompletionRequestMessage
   return getReasoning(extendedMessage)
-}
+}