fix: prevent time-travel bug in parallel tool calling

[daniel-lxs]

Problem

When multiple tools are called in parallel (e.g., update_todo_list + new_task), tool_results could appear in API history BEFORE their corresponding tool_use blocks, causing Anthropic API 400 errors:

"unexpected `tool_use_id` found in `tool_result` blocks. Each `tool_result` block must have a corresponding `tool_use` block in the previous message."

Example of malformed history:

[user msg] → [user msg with tool_results] → [assistant msg with tool_uses] ❌

Should be:

[user msg] → [assistant msg with tool_uses] → [user msg with tool_results] ✓

Root Cause

Tools executed during streaming via presentAssistantMessage() BEFORE the assistant message was saved to history. The sequence was:

1. Stream starts, tool_use chunks arrive
2. Each tool_use is finalized (marked non-partial)
3. presentAssistantMessage immediately executes completed tools
4. update_todo_list executes → pushes tool_result to userMessageContent
5. new_task executes → calls flushPendingToolResultsToHistory()
6. User message with tool_results is saved BEFORE assistant message exists in history
7. Stream completes, assistant message finally saved at Task.ts:3475

Solution

Added assistantMessageSavedToHistory coordination flag to ensure correct message ordering:

1. Flag is reset at the start of each API request (Task.ts:2724)
2. Flag is set to true after assistant message is saved to history (Task.ts:3506)
3. flushPendingToolResultsToHistory() waits for the flag before flushing (Task.ts:1074-1110)
4. This ensures: [assistant with tool_uses] → [user with tool_results]

Also removed the "CRITICAL: This tool MUST be called alone" restriction from new_task tool description to enable parallel tool calling.

Changes

• src/core/task/Task.ts: Added assistantMessageSavedToHistory flag and waiting logic in flushPendingToolResultsToHistory()
• src/core/prompts/tools/native-tools/new_task.ts: Removed restriction preventing parallel tool calling
• src/core/task/tests/flushPendingToolResultsToHistory.spec.ts: Added tests for waiting behavior

Testing

All 8 tests pass in the new test file, including tests verifying:

• ✓ Skips waiting when flag is already true
• ✓ Waits via pWaitFor when flag is false
• ✓ Aborts without flushing when task is aborted during wait

All 5228 tests pass across the entire codebase.

Related

This fix enables parallel tool calling (when the multipleNativeToolCalls experiment is enabled) by ensuring the assistant message is always saved to history before tool_results are flushed, preventing the time-travel bug.

---

Important

Fixes a bug in Task.ts by adding a coordination flag to ensure tool results are saved after assistant messages, preventing API errors.

• Behavior:
  • Added assistantMessageSavedToHistory flag in Task.ts to ensure tool results are saved after assistant messages.
  • flushPendingToolResultsToHistory() now waits for the flag before proceeding.
  • Removed restriction on new_task tool for parallel execution.
• Testing:
  • Added tests in flushPendingToolResultsToHistory.spec.ts to verify waiting behavior and flag handling.
  • Tests cover scenarios where the flag is true, false, and task is aborted.
• Misc:
  • Updated logic in Task.ts to reset and set the flag at appropriate points in the task lifecycle.

^{This description was created by for bf7a662. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task on Roo Cloud

Review complete. No issues found.

The implementation correctly addresses the time-travel bug in parallel tool calling by:

1. Adding an assistantMessageSavedToHistory coordination flag to ensure proper message ordering
2. Making flushPendingToolResultsToHistory() wait for the assistant message before flushing tool results
3. Properly handling abort scenarios during the wait
4. Adequate test coverage for the new waiting behavior

Previous reviews

• b90fe8d: Review #1

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