fix(deepseek): round-trip reasoning_content in thinking mode to prevent 400 errors#775
Merged
Conversation
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
🚧 Files skipped from review as they are similar to previous changes (1)
📝 WalkthroughWalkthroughAdds Changesreasoning_content round-trip
Estimated code review effort: 3 (Moderate) | ~25 minutes Suggested labels: Suggested reviewers: 🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
edelauna
changed the title
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
edelauna
requested review from
JamesRobert20,
hannesrudolph,
navedmerchant and
taltas
as code owners
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/api/transform/openai-format.ts (1)
462-481: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick winMultiple reasoning blocks in one turn will silently drop earlier reasoning text.
extractedReasoningis overwritten on every matching"reasoning"block rather than accumulated, so only the last block's text survives. Since the PR explicitly targets DeepSeek/Z.ai interleaved thinking (reasoning interspersed with tool calls in the same turn), a turn with more than one reasoning segment will lose all but the final one when round-tripped back to the API.♻️ Proposed fix: accumulate instead of overwrite
- let extractedReasoning: string | undefined + const extractedReasoningParts: string[] = [] const { nonToolMessages, toolMessages } = anthropicMessage.content.reduce<{ nonToolMessages: (Anthropic.TextBlockParam | Anthropic.ImageBlockParam)[] toolMessages: Anthropic.ToolUseBlockParam[] }>( (acc, part) => { if (part.type === "tool_use") { acc.toolMessages.push(part) } else if (part.type === "text" || part.type === "image") { acc.nonToolMessages.push(part) } else if ((part as any).type === "reasoning" && (part as any).text) { - // Extract reasoning stored as a content block (DeepSeek / Z.ai interleaved thinking). - // Must be passed back as top-level reasoning_content so providers like DeepSeek - // don't reject the request with "reasoning_content must be passed back to the API". - extractedReasoning = (part as any).text + extractedReasoningParts.push((part as any).text) } // assistant cannot send tool_result messages return acc }, { nonToolMessages: [], toolMessages: [] }, ) + const extractedReasoning = extractedReasoningParts.length > 0 ? extractedReasoningParts.join("\n") : undefined🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/api/transform/openai-format.ts` around lines 462 - 481, In openai-format.ts, the reasoning extraction in the anthropicMessage.content reduce currently overwrites extractedReasoning on each "reasoning" block, so earlier segments are lost. Update the reducer logic in the nonToolMessages/toolMessages extraction path to accumulate all reasoning text blocks (preserving order) instead of replacing the previous value, and ensure the final reasoning_content passed back for DeepSeek/Z.ai interleaved thinking includes every reasoning block found in the turn.
🧹 Nitpick comments (1)
apps/vscode-e2e/src/suite/providers/deepseek-v4.test.ts (1)
382-395: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick winSolid regression assertion for the round-trip fix.
Existence check on
secondRequestbefore asserting onhasReasoningContentInHistoryavoids a brittle false-positive/undefined failure. Consider also assertinghasReasoningContentInHistoryisfalseon the corresponding reasoning-off probe's second request, to guard against the flag becoming a tautology (e.g., always true due to a capture bug).🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@apps/vscode-e2e/src/suite/providers/deepseek-v4.test.ts` around lines 382 - 395, The regression check in deepseek-v4.test.ts should not only verify that the reasoning-enabled flow preserves reasoning_content on the second request via secondRequest.hasReasoningContentInHistory, but also guard against a capture bug by asserting the equivalent second request in the reasoning-off probe keeps that flag false. Use the existing probe/request objects in the DeepSeek test suite to add the negative assertion alongside the current positive one so the flag is validated in both modes.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Outside diff comments:
In `@src/api/transform/openai-format.ts`:
- Around line 462-481: In openai-format.ts, the reasoning extraction in the
anthropicMessage.content reduce currently overwrites extractedReasoning on each
"reasoning" block, so earlier segments are lost. Update the reducer logic in the
nonToolMessages/toolMessages extraction path to accumulate all reasoning text
blocks (preserving order) instead of replacing the previous value, and ensure
the final reasoning_content passed back for DeepSeek/Z.ai interleaved thinking
includes every reasoning block found in the turn.
---
Nitpick comments:
In `@apps/vscode-e2e/src/suite/providers/deepseek-v4.test.ts`:
- Around line 382-395: The regression check in deepseek-v4.test.ts should not
only verify that the reasoning-enabled flow preserves reasoning_content on the
second request via secondRequest.hasReasoningContentInHistory, but also guard
against a capture bug by asserting the equivalent second request in the
reasoning-off probe keeps that flag false. Use the existing probe/request
objects in the DeepSeek test suite to add the negative assertion alongside the
current positive one so the flag is validated in both modes.
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 9edf7d50-fd91-4a2f-9e82-45ae5020c01e
📒 Files selected for processing (5)
apps/vscode-e2e/fixtures/deepseek-v4.jsonapps/vscode-e2e/src/suite/providers/deepseek-v4.test.tssrc/api/providers/__tests__/openai.spec.tssrc/api/transform/__tests__/openai-format.spec.tssrc/api/transform/openai-format.ts
github-actions
Bot
added
the
awaiting-review
github-actions
Bot
added
awaiting-review
taltas
previously approved these changes
Jul 1, 2026
taltas
left a comment
taltas
left a comment
Contributor
There was a problem hiding this comment.
Clean PR, except one question with the typing.
| } | ||
| } | ||
| } else if (anthropicMessage.role === "assistant") { | ||
| const messageWithDetails = anthropicMessage as any |
Contributor
There was a problem hiding this comment.
I am not a big fan of this any here... maybe we should better define the type here? It's not really a blocker, but something to consider.
github-actions
Bot
removed
the
awaiting-review
Add a regression test that passes a non-object content part (a stray string) through convertToOpenAiMessages, exercising the defensive early-return in getReasoningBlockText (line 276). This was the only uncovered line in the PR patch, causing the codecov/patch check to fail.
navedmerchant
approved these changes
Jul 3, 2026
github-actions
Bot
added
the
awaiting-review
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
3 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.
Related GitHub Issue
Closes: #201
Description
When DeepSeek thinking mode is active, the API requires
reasoning_contentto be passed back on every assistant message in conversation history. Without it, the API returns HTTP 400:The reasoning_content in the thinking mode must be passed back to the API.Root cause — two bugs, one symptom:
convertToOpenAiMessages(used by the generic OpenAI-compatible provider path) silently droppedreasoningcontent blocks and never setreasoning_contenton outgoing messages. This is the primary fix.aimock fixture format — the e2e fixtures had
reasoning+toolCallsbut nocontent, so aimock classified them asToolCallResponse(which ignoresreasoning) instead ofContentWithToolCallsResponse. Fixed by adding"content": "".What changed:
src/api/transform/openai-format.ts—convertToOpenAiMessagesnow extractsreasoning_contentfrom both the top-level field (already round-tripped) and{ type: "reasoning" }content blocks (stored by Task.ts during streaming), and sets it on the outgoing assistant message.src/api/transform/__tests__/openai-format.spec.ts— 5 new unit tests covering the round-trip: string-content path, array-content path, array-content with tool calls (the exact failure case), preference ordering, and absent-reasoning no-op.src/api/providers/__tests__/openai.spec.ts— provider-level integration test:OpenAiHandler.createMessagewithpreserveReasoning: trueand an assistant message containing areasoningblock + tool call; assertsreasoning_contentappears in the body sent to the OpenAI client. This directly exercises the fixedOpenAiHandler → convertToOpenAiMessagespath.apps/vscode-e2e/fixtures/deepseek-v4.json— added"content": ""to reasoning-on fixtures so aimock routes them throughbuildContentWithToolCallsChunksand streamsreasoning_contentin SSE deltas.apps/vscode-e2e/src/suite/providers/deepseek-v4.test.ts— e2e regression guard for the native DeepSeek provider path: assertshasReasoningContentInHistoryon the second request (after the tool call), provingreasoning_contentfrom turn 1 survives into turn 2's history.Coverage note: The native
DeepSeekHandler → convertToR1Formatpath was already correct before this PR. The e2e test guards that path. The fix toconvertToOpenAiMessagesis guarded by unit tests + the new provider-level integration test inopenai.spec.ts.Test Procedure
Unit tests:
All 408 test files pass (6,584 tests).
E2e tests (mock):
All 4 DeepSeek V4 provider tests pass, including the
reasoning-onprobes that asserthasReasoningContentInHistoryon the second request.Live API validation (performed during development):
content: ""+reasoning_content: "..."+tool_calls: [...]on turn 1 tool-call responses.reasoning_contentround-tripped returns HTTP 200.Pre-Submission Checklist
Summary by CodeRabbit
New Features
reasoning_contentin OpenAI-compatible message formatting, including DeepSeek/Z.ai thinking mode and mixed tool+reasoning turns.Bug Fixes
Tests / Chores