orcs-oss/orcs-code
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

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

Browse Source 
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
Download Patch File Download Diff File Expand all files Collapse all files

89
src/services/api/codexShim.test.ts
View File

@@ -547,7 +547,7 @@ describe('Codex request translation', () => {
])
})
test('strips leaked reasoning preamble from completed Codex text responses', () => {
test('strips <think> tag block from completed Codex text responses', () => {
const message = convertCodexResponseToAnthropicMessage(
{
id: 'resp_1',
@@ -560,7 +560,7 @@ describe('Codex request translation', () => {
{
type: 'output_text',
text:
'The user just said "hey" - a simple greeting. I should respond briefly and friendly.\n\nHey! How can I help you today?',
'<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?',
},
],
},
@@ -578,6 +578,37 @@ describe('Codex request translation', () => {
])
})
test('strips unterminated <think> tag at block boundary in Codex completed response', () => {
const message = convertCodexResponseToAnthropicMessage(
{
id: 'resp_1',
model: 'gpt-5.4',
output: [
{
type: 'message',
role: 'assistant',
content: [
{
type: 'output_text',
text:
'Here is the answer.\n<think>wait, let me reconsider the user request',
},
],
},
],
usage: { input_tokens: 12, output_tokens: 4 },
},
'gpt-5.4',
)
expect(message.content).toEqual([
{
type: 'text',
text: 'Here is the answer.',
},
])
})
test('translates Codex SSE text stream into Anthropic events', async () => {
const responseText = [
'event: response.output_item.added',
@@ -609,7 +640,7 @@ describe('Codex request translation', () => {
])
})
test('strips leaked reasoning preamble from Codex SSE text stream', async () => {
test('strips <think> tag block from Codex SSE text stream', async () => {
const responseText = [
'event: response.output_item.added',
'data: {"type":"response.output_item.added","item":{"id":"msg_1","type":"message","status":"in_progress","content":[],"role":"assistant"},"output_index":0,"sequence_number":0}',
@@ -618,13 +649,13 @@ describe('Codex request translation', () => {
'data: {"type":"response.content_part.added","content_index":0,"item_id":"msg_1","output_index":0,"part":{"type":"output_text","text":""},"sequence_number":1}',
'',
'event: response.output_text.delta',
'data: {"type":"response.output_text.delta","content_index":0,"delta":"The user just said \\"hey\\" - a simple greeting. I should respond briefly and friendly.\\n\\nHey! How can I help you today?","item_id":"msg_1","output_index":0,"sequence_number":2}',
'data: {"type":"response.output_text.delta","content_index":0,"delta":"<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?","item_id":"msg_1","output_index":0,"sequence_number":2}',
'',
'event: response.output_item.done',
'data: {"type":"response.output_item.done","item":{"id":"msg_1","type":"message","status":"completed","content":[{"type":"output_text","text":"The user just said \\"hey\\" - a simple greeting. I should respond briefly and friendly.\\n\\nHey! How can I help you today?"}],"role":"assistant"},"output_index":0,"sequence_number":3}',
'data: {"type":"response.output_item.done","item":{"id":"msg_1","type":"message","status":"completed","content":[{"type":"output_text","text":"<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?"}],"role":"assistant"},"output_index":0,"sequence_number":3}',
'',
'event: response.completed',
'data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","model":"gpt-5.4","output":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"The user just said \\"hey\\" - a simple greeting. I should respond briefly and friendly.\\n\\nHey! How can I help you today?"}]}],"usage":{"input_tokens":2,"output_tokens":1}},"sequence_number":4}',
'data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","model":"gpt-5.4","output":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"<think>user wants a greeting, respond briefly</think>Hey! How can I help you today?"}]}],"usage":{"input_tokens":2,"output_tokens":1}},"sequence_number":4}',
'',
].join('\n')
@@ -646,6 +677,50 @@ describe('Codex request translation', () => {
}
}
expect(textDeltas).toEqual(['Hey! How can I help you today?'])
expect(textDeltas.join('')).toBe('Hey! How can I help you today?')
})
test('preserves prose without tags (no phrase-based false positive)', async () => {
// Regression test: older phrase-based sanitizer would incorrectly strip text
// starting with "I should" or "The user". The tag-based approach leaves it alone.
const responseText = [
'event: response.output_item.added',
'data: {"type":"response.output_item.added","item":{"id":"msg_1","type":"message","status":"in_progress","content":[],"role":"assistant"},"output_index":0,"sequence_number":0}',
'',
'event: response.content_part.added',
'data: {"type":"response.content_part.added","content_index":0,"item_id":"msg_1","output_index":0,"part":{"type":"output_text","text":""},"sequence_number":1}',
'',
'event: response.output_text.delta',
'data: {"type":"response.output_text.delta","content_index":0,"delta":"I should note that the user role requires a briefly concise friendly response format.","item_id":"msg_1","output_index":0,"sequence_number":2}',
'',
'event: response.output_item.done',
'data: {"type":"response.output_item.done","item":{"id":"msg_1","type":"message","status":"completed","content":[{"type":"output_text","text":"I should note that the user role requires a briefly concise friendly response format."}],"role":"assistant"},"output_index":0,"sequence_number":3}',
'',
'event: response.completed',
'data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","model":"gpt-5.4","output":[{"type":"message","role":"assistant","content":[{"type":"output_text","text":"I should note that the user role requires a briefly concise friendly response format."}]}],"usage":{"input_tokens":2,"output_tokens":1}},"sequence_number":4}',
'',
].join('\n')
const stream = new ReadableStream({
start(controller) {
controller.enqueue(new TextEncoder().encode(responseText))
controller.close()
},
})
const textDeltas: string[] = []
for await (const event of codexStreamToAnthropic(
new Response(stream),
'gpt-5.4',
)) {
const delta = (event as { delta?: { type?: string; text?: string } }).delta
if (delta?.type === 'text_delta' && typeof delta.text === 'string') {
textDeltas.push(delta.text)
}
}
expect(textDeltas.join('')).toBe(
'I should note that the user role requires a briefly concise friendly response format.',
)
})
})