fix(api): replace phrase-based reasoning sanitizer with tag-based filter (#779)

Reasoning models (MiniMax M2.7, GLM-4.5/5, DeepSeek, Kimi K2) inline
chain-of-thought inside <think>...</think> tags in the content field
rather than using the reasoning_content channel. The prior phrase-matching
sanitizer (looksLikeLeakedReasoningPrefix) only caught English-prose
preambles like "I should"/"the user asked", missed tag-based leaks
entirely, and risked false-stripping legitimate assistant output.

Replace with a structural tag-based approach (same pattern as hermes-agent):

- createThinkTagFilter() — streaming state machine that buffers partial
  tags across SSE delta boundaries (<th| + |ink>), so tags split mid-chunk
  still parse correctly.
- stripThinkTags() — whole-text cleanup for non-streaming responses and
  as a safety net. Handles closed pairs, unterminated opens at block
  boundaries, and orphan tags.
- Recognizes think, thinking, reasoning, thought, REASONING_SCRATCHPAD
  case-insensitively, including tags with attributes.
- False-negative bias: flush() discards buffered partial tags at stream
  end rather than leaking them.

Existing phrase-based shim tests updated to exercise the actual <think>
tag leak. Added regression tests confirming legitimate prose starting
with "I should..." is preserved (the old sanitizer's main false-positive).

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Kevin Codex
2026-04-20 15:18:58 +08:00
committed by GitHub
parent c0b8a59a23
commit 336ddcc50d
8 changed files with 544 additions and 213 deletions

View File

@@ -6,10 +6,9 @@ import type {
} from './providerConfig.js'
import { sanitizeSchemaForOpenAICompat } from './openaiSchemaSanitizer.js'
import {
looksLikeLeakedReasoningPrefix,
shouldBufferPotentialReasoningPrefix,
stripLeakedReasoningPreamble,
} from './reasoningLeakSanitizer.js'
createThinkTagFilter,
stripThinkTags,
} from './thinkTagSanitizer.js'
export interface AnthropicUsage {
input_tokens: number
@@ -734,25 +733,22 @@ export async function* codexStreamToAnthropic(
{ index: number; toolUseId: string }
>()
let activeTextBlockIndex: number | null = null
let activeTextBuffer = ''
let textBufferMode: 'none' | 'pending' | 'strip' = 'none'
const thinkFilter = createThinkTagFilter()
let nextContentBlockIndex = 0
let sawToolUse = false
let finalResponse: Record<string, any> | undefined
const closeActiveTextBlock = async function* () {
if (activeTextBlockIndex === null) return
if (textBufferMode !== 'none') {
const sanitized = stripLeakedReasoningPreamble(activeTextBuffer)
if (sanitized) {
yield {
type: 'content_block_delta',
index: activeTextBlockIndex,
delta: {
type: 'text_delta',
text: sanitized,
},
}
const tail = thinkFilter.flush()
if (tail) {
yield {
type: 'content_block_delta',
index: activeTextBlockIndex,
delta: {
type: 'text_delta',
text: tail,
},
}
}
yield {
@@ -760,8 +756,6 @@ export async function* codexStreamToAnthropic(
index: activeTextBlockIndex,
}
activeTextBlockIndex = null
activeTextBuffer = ''
textBufferMode = 'none'
}
const startTextBlockIfNeeded = async function* () {
@@ -837,43 +831,17 @@ export async function* codexStreamToAnthropic(
if (event.event === 'response.output_text.delta') {
yield* startTextBlockIfNeeded()
activeTextBuffer += payload.delta ?? ''
if (activeTextBlockIndex !== null) {
if (
textBufferMode === 'strip' ||
looksLikeLeakedReasoningPrefix(activeTextBuffer)
) {
textBufferMode = 'strip'
continue
}
if (textBufferMode === 'pending') {
if (shouldBufferPotentialReasoningPrefix(activeTextBuffer)) {
continue
}
const visible = thinkFilter.feed(payload.delta ?? '')
if (visible) {
yield {
type: 'content_block_delta',
index: activeTextBlockIndex,
delta: {
type: 'text_delta',
text: activeTextBuffer,
text: visible,
},
}
textBufferMode = 'none'
continue
}
if (shouldBufferPotentialReasoningPrefix(activeTextBuffer)) {
textBufferMode = 'pending'
continue
}
yield {
type: 'content_block_delta',
index: activeTextBlockIndex,
delta: {
type: 'text_delta',
text: payload.delta ?? '',
},
}
}
continue
@@ -969,7 +937,7 @@ export function convertCodexResponseToAnthropicMessage(
if (part?.type === 'output_text') {
content.push({
type: 'text',
text: stripLeakedReasoningPreamble(part.text ?? ''),
text: stripThinkTags(part.text ?? ''),
})
}
}