[Fix] DeepSeek 400 errors when mixed-model tasks continue long conversations

[roomote]

Opened on behalf of Naved Merchant. View the task or mention @roomote for follow-up asks.

Related GitHub Issue

Closes: #201

Description

This fixes the mixed-model DeepSeek history path behind the recurring 400 reasoning_content error.

Assistant turns from OpenAI-style reasoning models now persist explicit reasoning_content metadata when that model requires replay, and request rebuilding only sends that metadata back for the assistant turns that were originally marked for it. That prevents long tasks from replaying generic reasoning blocks from other agents or models as if they were DeepSeek continuation history.

Test Procedure

• Ran pnpm --dir src exec vitest run core/task/__tests__/apiConversationHistory.spec.ts api/transform/__tests__/openai-format.spec.ts core/task/__tests__/Task.persistence.spec.ts api/providers/__tests__/deepseek.spec.ts
• Ran pnpm --dir src check-types
• Attempted the environment-mandated pnpm --dir apps/web-roo-code test, but this checkout does not contain apps/web-roo-code, so that validation path could not run here.

Pre-Submission Checklist

• Issue Linked: This PR is linked to an approved issue (see "Related GitHub Issue" above).
• Scope: My changes are focused on the linked issue (one major feature/fix per PR).
• Self-Review: I have performed a thorough self-review of my code.
• Testing: New and/or updated tests have been added to cover my changes (if applicable).
• Documentation Impact: I have considered if my changes require documentation updates (see "Documentation Updates" section below).
• Contribution Guidelines: I have read and agree to the Contributor Guidelines.

Screenshots / Videos

Not applicable for this internal conversation-history fix.

Documentation Updates

• No documentation updates are required.
• Yes, documentation updates are required. (Please describe what needs to be updated or link to a PR in the docs repository).

Additional Notes

The fix also adds regression coverage around request rebuilding so DeepSeek-marked assistant turns keep their own reasoning_content while unrelated reasoning from other models is stripped before replay.

Get in Touch

Roomote (task 0mzxfggobdyyv)

[roomote]

1 issue outstanding; 2 checks still running. Action required. See task

• Preserve reasoning replay for Anthropic-style preserveReasoning providers instead of only replaying stored top-level reasoning_content.
• platform-unit-test (ubuntu-latest) is still running.
• platform-unit-test (windows-latest) is still running.

[roomote]

prepareApiConversationMessage() only stores top-level reasoning_content for apiProtocol === "openai" (src/core/task/apiConversationHistory.ts). Anthropic-style thinking providers in this repo still set info.preserveReasoning (for example MiniMax and the Bedrock Kimi/MiniMax models) but persist their reasoning as embedded blocks, so this branch now strips the reasoning those continuations need.

[codecov]

Codecov Report

❌ Patch coverage is 88.09524% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/core/task/Task.ts	80.00%	5 Missing ⚠️

📢 Thoughts on this report? Let us know!