fix(openai): parse MiniMax thinking tags by shenyankm · Pull Request #3677 · QwenLM/qwen-code

shenyankm · 2026-04-27T10:10:53Z

Add a MiniMax OpenAI-compatible provider that opts into tagged thinking parsing for minimaxi.com endpoints.

Split MiniMax <think> / <thinking> content into thought parts while preserving default OpenAI-compatible behavior, and avoid rendering leading blank stream chunks as empty assistant rows.

Summary

What changed:
- Added a MiniMax OpenAI-compatible provider selected only for *.minimaxi.com endpoints.
- Added provider-gated tagged thinking parsing for MiniMax <think> / <thinking> content.
- Converted tagged MiniMax thinking content into thought: true parts.
- Preserved existing reasoning_content / reasoning thought handling.
- Prevented leading blank streaming chunks from rendering an empty assistant ✦ row in the CLI TUI.
Why it changed:
- MiniMax OpenAI-compatible /v1 responses can place thinking content inside normal content using <think> or <thinking> tags.
- The existing OpenAI-compatible converter only recognized reasoning_content / reasoning, so MiniMax thinking text leaked into the visible assistant response.
- Some MiniMax streams also begin visible output with blank chunks, which could render the assistant icon on a separate empty row.
Reviewer focus:
- Confirm tagged thinking parsing is enabled only for MiniMax *.minimaxi.com OpenAI-compatible endpoints.
- Confirm non-MiniMax OpenAI-compatible providers still preserve normal XML/HTML-like text.
- Confirm streaming parsing handles tags split across chunks.
- Confirm leading blank stream chunks no longer create empty assistant rows.

Validation

Commands run:

cd packages/core && npx vitest run src/core/openaiContentGenerator/converter.test.ts src/core/openaiContentGenerator/provider/minimax.test.ts src/core/openaiContentGenerator/pipeline.test.ts
cd packages/cli && npx vitest run src/ui/hooks/useGeminiStream.test.tsx
cd packages/core && npm run lint && npx tsc --noEmit
cd packages/cli && npm run lint && npx tsc --noEmit
npm run build
npm run typecheck
git diff --check

Prompts / inputs used:
```
原神怎么你了？
```
Expected result:
- MiniMax OpenAI-compatible <think> / <thinking> content is parsed as thought content.
- Raw thinking tags are not shown in the final visible assistant response.
- Non-MiniMax OpenAI-compatible providers keep tagged text unchanged.
- The assistant icon ✦ aligns with the first visible response line instead of appearing on an empty row.
Observed result:
- MiniMax tagged thinking content is split into thought parts.
- Visible assistant output no longer contains raw <think> / <thinking> tags.
- Streaming tags split across chunks, such as <thi + nk>, are handled correctly.
- CLI output no longer renders an empty assistant row before the response.

Quickest reviewer verification path:

cd packages/core && npx vitest run src/core/openaiContentGenerator/converter.test.ts src/core/openaiContentGenerator/provider/minimax.test.ts
cd ../cli && npx vitest run src/ui/hooks/useGeminiStream.test.tsx

Evidence (output, logs, screenshots, video, JSON, before/after, etc.):

Before:

After:

Test results:

packages/core: targeted OpenAI converter/provider/pipeline tests passed.
packages/cli: useGeminiStream.test.tsx passed.
npm run build: passed.
npm run typecheck: passed.
git diff --check: passed.

Scope / Risk

Main risk or tradeoff:
- Tagged thinking parsing is intentionally gated to MiniMax *.minimaxi.com endpoints to avoid changing behavior for other OpenAI-compatible providers.
- The CLI display fix drops leading blank lines before the first visible assistant/thought chunk, which avoids empty assistant rows and matches existing visible rendering expectations.
Not covered / not validated:
- Full automated network E2E against the MiniMax API is not included in CI.
- Manual MiniMax OpenAI-compatible runtime validation was performed locally with npm run dev.
- Docker, Podman, Seatbelt, macOS, and Linux runtime paths were not tested for this change.
Breaking changes / migration notes:
- None.

Testing Matrix

	🍏	🪟	🐧
npm run	⚠️	✅	⚠️
npx	⚠️	✅	⚠️
Docker	⚠️	⚠️	⚠️
Podman	⚠️	N/A	N/A
Seatbelt	⚠️	N/A	N/A

Testing matrix notes:

Validated on Windows with targeted npx vitest run commands, package-level npm run lint, package-level npx tsc --noEmit, root npm run build, and root npm run typecheck.
macOS and Linux were not tested locally.
Docker, Podman, and Seatbelt were not tested because this change is limited to TypeScript response conversion and CLI rendering logic.

Linked Issues / Bugs

Fixes #3669
Fixes #3228
Fixes #3387

Add a MiniMax OpenAI-compatible provider that opts into tagged thinking parsing for minimaxi.com endpoints. Split MiniMax <think>/<thinking> content into thought parts while preserving default OpenAI-compatible behavior, and avoid rendering leading blank stream chunks as empty assistant rows.

Merge upstream/main into the MiniMax thinking-tags branch. Resolve the RequestContext conflict by preserving both MiniMax response parsing fields and upstream splitToolMedia support. Recognize both official MiniMax OpenAI-compatible API hosts: api.minimaxi.com and api.minimax.io.

shenyankm · 2026-04-28T02:00:31Z

I merged the latest upstream/main into this branch to resolve the merge conflict.

The conflict was in RequestContext, and the resolution keeps both sides:

MiniMax thinking-tags parsing fields from this PR
splitToolMedia support from upstream main

I also added support for the MiniMax international OpenAI-compatible endpoint:

China: https://api.minimaxi.com/v1
International: https://api.minimax.io/v1

Validation completed:

provider/minimax.test.ts and converter.test.ts passed
npm run lint passed
npm run build passed
npm run typecheck passed

wenshao

⚠️ Downgraded from Approve to Comment: CI has 1 failing check (Qwen Code CI/Test on macos-latest, 24.x).

…thinking parser - perf(taggedThinkingParser): pre-compute lowercase buffer once per parse() call, use startsWith(tag, offset) to avoid O(N²) per-char slice+toLowerCase allocations - feat(minimax): expand host matching from exact Set to *.minimaxi.com / *.minimax.io wildcard suffix, covering gateway / custom subdomains - docs(converter): add comment clarifying finish_reason flush of buffered tagged-thinking content on stream end

QwenLM/qwen-code#3677 (minimax thinking tag parser + blank-chunk UI fix), aaif-goose/goose#8984 (officeqa/minimax blog). INDEX.md backfills drip-333 (which was missed) plus drip-334.

wenshao

Follow-up review on 814ea0d: previous review findings (O(N²) tag matching, finish_reason flush comment) have been addressed. All CI checks pass.

Remaining suggestion:

TaggedThinkingParser 缺少独立单元测试 — 107 行的状态机类仅通过 converter.test.ts 间接测试。建议按项目约定创建 taggedThinkingParser.test.ts，覆盖：空标签内容、text 模式下遇 </think>（应为普通文本）、纯 partial-tag-prefix chunk 等边界场景。

— deepseek-v4-pro via Qwen Code /review

Add taggedThinkingParser.test.ts with 23 test cases covering the boundary scenarios requested in PR review #4219047370: empty tag content, close tags in text mode, pure partial-tag-prefix chunks, multi-chunk splitting, final flag flush behavior, and case insensitivity.

shenyankm · 2026-05-04T13:40:57Z

TaggedThinkingParser 缺少独立单元测试

@wenshao Thanks for the review! The suggested standalone unit test has been added in 15520c4:

Created \packages/core/src/core/openaiContentGenerator/taggedThinkingParser.test.ts\ with 23 test cases directly testing the 108-line state machine class
Covered all three requested boundary scenarios:
- Empty tag content: <think>\ / <thinking>\
- Close tags in text mode: </think>\ treated as normal text when no opening tag precedes it
- Pure partial-tag-prefix chunk: <thi\ buffered across chunks, full-chunk prefix returning []\
Extra coverage: case insensitivity, 3+ chunk tag splitting, final flag flush behavior, static convenience method
All 23 new tests + 86 existing tests pass; TypeScript compilation clean

Ready for re-review.

wenshao

[Critical] TaggedThinkingParser 完全缺少可观测性 — 解析器在没有任何 debug 日志的情况下静默运行。当它消费 <think> / <thinking> 标签内容并将其转换为对用户不可见的 thought: true parts 时，没有任何日志记录；在流截断时刷新未关闭的 thought 缓冲区也没有警告。建议至少添加 debugLogger 调用来记录标签检测、parts 计数和未关闭缓冲区的刷新。

[Suggestion] reasoning_content / reasoning 与 taggedThinkingTags 的隐式耦合：convertOpenAIChunkToGemini 中现有的 reasoning_content 提取（约第 1019 行）无条件添加 thought parts，而新增的 convertOpenAITextToParts 独立从标签中解析 thought parts。如果未来有 provider 同时启用 taggedThinkingTags 并返回 reasoning_content，thought 内容会静默重复。建议在提取 reasoning_content 前检查 !requestContext.responseParsingOptions?.taggedThinkingTags。

— deepseek-v4-pro via Qwen Code /review

…feguard for MiniMax tagged thinking - Add debugLogger.warn on unclosed thought flush (observability) - Add debugLogger.debug for tag detection and parts count - Add cross-matching tag comment explaining binary mode toggle - Add MINIMAX_KNOWN_HOSTS exact match layer with suffix fallback - Add reasoning_content guard to avoid duplicating thought parts - Add cross-matching and unclosed flush tests (+4 cases) - Add known host exact match and custom proxy tests (+4 cases)

shenyankm · 2026-05-04T18:19:12Z

@wenshao All review comments from #4221337642 have been addressed in commit 30d5d66:

Observability (Critical): Added debugLogger.warn on unclosed thought flush + debugLogger.debug for tag detection and parts count in taggedThinkingParser.ts.

Cross-matching comment: Added comment explaining the binary mode toggle design at L12-14.

MiniMax host detection: Added MINIMAX_KNOWN_HOSTS exact match layer (api.minimaxi.com, api.minimax.io) before suffix fallback in minimax.ts.

reasoning_content safeguard: Added guard in both convertOpenAIResponseToGemini and convertOpenAIChunkToGemini to skip reasoning_content extraction when taggedThinkingTags is enabled, avoiding duplicate thought parts.

Test coverage: +8 tests (cross-matching x2, unclosed flush x2, known hosts x2, custom proxy x2). All 116 tests pass, typecheck clean.

Ready for re-review.

wenshao

No issues found in the PR diff. LGTM! ✅

Note: tsc --noEmit reports 2 pre-existing type errors (ServerGeminiEventType.Text → should be Content) in useGeminiStream.test.tsx:3067,3373, but these are inherited from the merge-base (commit 431a87c) and not introduced by this PR.

— deepseek-v4-pro via Qwen Code /review

* fix(openai): parse MiniMax thinking tags Add a MiniMax OpenAI-compatible provider that opts into tagged thinking parsing for minimaxi.com endpoints. Split MiniMax <think>/<thinking> content into thought parts while preserving default OpenAI-compatible behavior, and avoid rendering leading blank stream chunks as empty assistant rows. * fix(openai): address PR #3677 review — optimize MiniMax tagged thinking parser - perf(taggedThinkingParser): pre-compute lowercase buffer once per parse() call, use startsWith(tag, offset) to avoid O(N²) per-char slice+toLowerCase allocations - feat(minimax): expand host matching from exact Set to *.minimaxi.com / *.minimax.io wildcard suffix, covering gateway / custom subdomains - docs(converter): add comment clarifying finish_reason flush of buffered tagged-thinking content on stream end * test(openai): add standalone unit tests for TaggedThinkingParser Add taggedThinkingParser.test.ts with 23 test cases covering the boundary scenarios requested in PR review #4219047370: empty tag content, close tags in text mode, pure partial-tag-prefix chunks, multi-chunk splitting, final flag flush behavior, and case insensitivity. * fix(openai): add observability, host guards, and reasoning_content safeguard for MiniMax tagged thinking - Add debugLogger.warn on unclosed thought flush (observability) - Add debugLogger.debug for tag detection and parts count - Add cross-matching tag comment explaining binary mode toggle - Add MINIMAX_KNOWN_HOSTS exact match layer with suffix fallback - Add reasoning_content guard to avoid duplicating thought parts - Add cross-matching and unclosed flush tests (+4 cases) - Add known host exact match and custom proxy tests (+4 cases)

* fix(openai): parse MiniMax thinking tags Add a MiniMax OpenAI-compatible provider that opts into tagged thinking parsing for minimaxi.com endpoints. Split MiniMax <think>/<thinking> content into thought parts while preserving default OpenAI-compatible behavior, and avoid rendering leading blank stream chunks as empty assistant rows. * fix(openai): address PR QwenLM#3677 review — optimize MiniMax tagged thinking parser - perf(taggedThinkingParser): pre-compute lowercase buffer once per parse() call, use startsWith(tag, offset) to avoid O(N²) per-char slice+toLowerCase allocations - feat(minimax): expand host matching from exact Set to *.minimaxi.com / *.minimax.io wildcard suffix, covering gateway / custom subdomains - docs(converter): add comment clarifying finish_reason flush of buffered tagged-thinking content on stream end * test(openai): add standalone unit tests for TaggedThinkingParser Add taggedThinkingParser.test.ts with 23 test cases covering the boundary scenarios requested in PR review #4219047370: empty tag content, close tags in text mode, pure partial-tag-prefix chunks, multi-chunk splitting, final flag flush behavior, and case insensitivity. * fix(openai): add observability, host guards, and reasoning_content safeguard for MiniMax tagged thinking - Add debugLogger.warn on unclosed thought flush (observability) - Add debugLogger.debug for tag detection and parts count - Add cross-matching tag comment explaining binary mode toggle - Add MINIMAX_KNOWN_HOSTS exact match layer with suffix fallback - Add reasoning_content guard to avoid duplicating thought parts - Add cross-matching and unclosed flush tests (+4 cases) - Add known host exact match and custom proxy tests (+4 cases)

tanzhenxin added the type/bug Something isn't working as expected label Apr 28, 2026

github-actions Bot mentioned this pull request Apr 28, 2026

📊 AI CLI 工具社区动态日报 2026-04-28 gsscsd/big_model_radar#256

Open

Merge branch 'QwenLM:main' into sheny/fix-minimax-thinking-tags

8e9a7b7

github-actions Bot mentioned this pull request May 2, 2026

📊 AI CLI 工具社区动态日报 2026-05-02 gsscsd/big_model_radar#280

Open

wenshao reviewed May 3, 2026

View reviewed changes

Comment thread packages/core/src/core/openaiContentGenerator/taggedThinkingParser.ts

Comment thread packages/core/src/core/openaiContentGenerator/converter.ts

shenyankm requested a review from wenshao May 4, 2026 06:51

wenshao reviewed May 4, 2026

View reviewed changes

shenyankm requested a review from wenshao May 4, 2026 13:55

wenshao requested changes May 4, 2026

View reviewed changes

Comment thread packages/core/src/core/openaiContentGenerator/taggedThinkingParser.ts

Comment thread packages/core/src/core/openaiContentGenerator/taggedThinkingParser.ts

Comment thread packages/core/src/core/openaiContentGenerator/provider/minimax.ts

shenyankm requested a review from wenshao May 4, 2026 18:30

wenshao approved these changes May 4, 2026

View reviewed changes

wenshao merged commit 5984540 into QwenLM:main May 5, 2026
13 checks passed

github-actions Bot mentioned this pull request May 5, 2026

📊 AI CLI 工具社区动态日报 2026-05-05 gsscsd/big_model_radar#295

Open

shenyankm deleted the sheny/fix-minimax-thinking-tags branch May 6, 2026 03:38

tanzhenxin mentioned this pull request Jun 11, 2026

feat(minimax): add MiniMax provider with M3 as default #3165

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(openai): parse MiniMax thinking tags#3677

fix(openai): parse MiniMax thinking tags#3677
wenshao merged 6 commits into
QwenLM:mainfrom
shenyankm:sheny/fix-minimax-thinking-tags

shenyankm commented Apr 27, 2026

Uh oh!

shenyankm commented Apr 28, 2026

Uh oh!

wenshao left a comment

Uh oh!

Uh oh!

Uh oh!

wenshao left a comment

Uh oh!

shenyankm commented May 4, 2026

Uh oh!

wenshao left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

shenyankm commented May 4, 2026

Uh oh!

wenshao left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

shenyankm commented Apr 27, 2026

Summary

Validation

Scope / Risk

Testing Matrix

Linked Issues / Bugs

Uh oh!

shenyankm commented Apr 28, 2026

Uh oh!

wenshao left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

wenshao left a comment

Choose a reason for hiding this comment

Uh oh!

shenyankm commented May 4, 2026

Uh oh!

wenshao left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

shenyankm commented May 4, 2026

Uh oh!

wenshao left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants