Rooviewer   See task on Roo Cloud

I've reviewed this PR and found no issues. The implementation looks sound:

• The main fix (marking all tool_use blocks as non-partial after finalization) correctly prevents the race condition that caused duplicated tool rendering
• The fix is well-placed after finalizeRawChunks() and catches edge cases where toolUseIndex was undefined or finalizeStreamingToolCall returned null
• The CI/CD workflow changes add reasonable improvements (concurrency, lint/type-check/build steps)
• The Cerebras provider config change adds temperature support

Note: The PR is marked as a draft with "Not yet tested" - consider adding integration testing before merging.

Previous reviews

• d0a45a9: Review #1
• d0a45a9: Review #2
• e621c39: Review #3

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

🔄 Review in Progress - View Job

Summary

✅ No issues found - The PR is ready for approval.

Changes Reviewed

This PR fixes a race condition where the first tool in a parallel batch could be presented twice due to tools retaining partial=true after finalization. It also implements new_task isolation logic to prevent orphaned tools when delegation disposes the parent task.

Key Changes:

1. Task.ts (lines 3286-3295): Explicitly marks ALL tool_use and mcp_tool_use blocks as non-partial after finalization, preventing the double-presentation race condition.
2. Task.ts (lines 3433-3466): Implements new_task isolation enforcement - truncates subsequent tools when new_task is called alongside other tools and injects error tool_result blocks.
3. presentAssistantMessage.ts (line 878): Adds checkpoint save before new_task execution to ensure state is saved before delegation.
4. NewTaskTool.ts: Removed checkpoint logic (moved to presentAssistantMessage.ts).
5. new-task-isolation.spec.ts: New comprehensive test file for new_task isolation enforcement.

TODO

• Verify existing comments against current code
• Review changes and enumerate net new issues
• Submit comments
• Summarize issues flagged and update top-level comment
• Approve the pull request

📋 View Job Details | Roo Code Reviewer

Note: There are still concerns about the task.ask() method design being fragile when multiple asks of the same type occur in quick succession. This catch-all approach should handle the immediate race condition, but will do further testing to verify edge cases.

Tested. LGTM