fix(core): allow thought-only responses in GeminiChat stream validation#3251
Merged
Conversation
Contributor
📋 Review SummaryThis PR fixes a critical bug where reasoning/thinking models (e.g., 🔍 General Feedback
🎯 Specific Feedback🟢 Medium
🔵 Low
✅ Highlights
|
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
yiliang114
reviewed
Apr 14, 2026
Models using thinking/reasoning modes may emit only thought content without explicit text output. The stream validation previously rejected these as 'empty' responses. Now accepts responses that contain either text content or thought content when a finish reason is present. (cherry picked from commit a0b1391)
365f0ef to
6026843
Compare
wenshao
commented
Apr 14, 2026
wenshao
left a comment
Collaborator
Author
There was a problem hiding this comment.
No issues found. LGTM! ✅ — gpt-5.4 via Qwen Code /review
yiliang114
approved these changes
Apr 15, 2026
yiliang114
left a comment
Collaborator
There was a problem hiding this comment.
Main fix looks right to me. The current head also already addressed the earlier stale-comment nit in the test file.
I also did a quick local pass on the GeminiChat path and the related thought-only subagent behavior, and the targeted tests, typecheck, and lint looked good. No further comments from me.
|
This PR doesn't fix the error for me. I keep getting > /status
╭──────────────────────────────────────────────────────────────────────────────────────────────────╮
│ │
│ Status │
│ │
│ Qwen Code 0.15.6-nightly.20260504.e617f20d1 (d4baaf9e0) │
│ Runtime Node.js v22.22.2 / npm 10.9.7 │
│ OS linux x64 (6.8.0-110-generic) │
│ │
│ Auth API Key - openai │
│ Base URL http://127.0.0.1:11434/v1 │
│ Model qwen3.6:27b Ollama service for self hosted qwen3.6: services:
ollama:
image: ollama/ollama:latest
container_name: ollama
restart: unless-stopped
ports:
- "127.0.0.1:11434:11434"
volumes:
- ollama:/root/.ollama
environment:
- OLLAMA_HOST=0.0.0.0:11434
- OLLAMA_KEEP_ALIVE=24h
- OLLAMA_FLASH_ATTENTION=1
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]Here are my relevant qwen settings: ...
"env": {
"OLLAMA_API_KEY": "***"
},
"modelProviders": {
"openai": [
{
"id": "qwen3.6:27b",
"name": "Local Qwen3.6 27B",
"envKey": "OLLAMA_API_KEY",
"baseUrl": "http://127.0.0.1:11434/v1",
"generationConfig": {
"timeout": 300000,
"maxRetries": 1,
"contextWindowSize": 262144,
"modalities": {
"image": true
},
"samplingParams": {
"temperature": 0.2
}
}
}
]
},
"security": {
"auth": {
"selectedType": "openai"
}
},
"model": {
"name": "qwen3.6:27b"
},
"$version": 3,
....
Also qwen just ocasionally stops executing without any error at all, but I suspect its related to the error above as well. |
xaelistic
pushed a commit
to xaelistic/qwen-code
that referenced
this pull request
Jun 7, 2026
…on (QwenLM#3251) Models using thinking/reasoning modes may emit only thought content without explicit text output. The stream validation previously rejected these as 'empty' responses. Now accepts responses that contain either text content or thought content when a finish reason is present. (cherry picked from commit e8db8679305370de91e466475e8e672b2e782499) Co-authored-by: mingholy.lmh <mingholy.lmh@alibaba-inc.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
Subagents running against thinking/reasoning models (qwen3-thinking, qwen-plus-thinking, etc.) hit
Failed to run subagent: Model stream ended with empty response text.on code-review-style tasks like/review's undirected audit:The existing transient-stream retry budget (
INVALID_STREAM_RETRY_CONFIG.maxRetries = 2) retries twice, but the same prompt + same model consistently produces the same thought-only response, so all three attempts fail.Root cause
packages/core/src/core/geminiChat.ts:939validates that a successful stream has either a tool call or non-emptycontentText.contentTextis built fromconsolidatedHistoryParts, which filters out parts wherepart.thought === true:So when a reasoning model emits only
thought: trueparts and finishes with a validfinishReason: STOP,contentTextis empty and the validator throwsInvalidStreamError('NO_RESPONSE_TEXT'). This is a legitimate response shape for thinking-mode providers — they emit internal reasoning and sometimes end without producing a distinct visible-text part.Fix
Accept responses that contain thought text when no visible-text content is present. The validator now asks "is there any content?" instead of "is there visible text?":
Truly empty responses (no text AND no thought AND no tool call) still throw — only the "thought-only is empty" misclassification is fixed.
Attribution
This is cherry-picked from the previously abandoned PR #2636 by @Mingholy (commit
a0b13911f), which was closed on 2026-03-26 without human review despite being a correct fix for two user-reported issues (#2530 and #1700). The authorship on the commit is preserved viagit cherry-pick -x. Reopening here so the fix is not lost — it matches a production failure hit today on a/reviewsubagent.Test plan
npx vitest run packages/core/src/core/geminiChat.test.ts— 46 tests pass (43 pre-existing + 3 new thought-only coverage cases)The three new test cases (from mingholy's original PR) cover:
The existing "empty response" test is retained but restructured to assert that truly empty responses (no text, no thoughts, no tool call) still throw — regression-proofing the negative case.
Related issues