Summary

This PR enables the native tool protocol to execute multiple tools per assistant message turn, with comprehensive failure guardrails to ensure safe execution.

Key Features

1. Multiple Native Tool Calls Per Turn

• Native protocol (OpenAI-style) can now execute multiple tools in a single assistant message
• Removed didAlreadyUseTool restriction for native protocol
• XML protocol maintains single-tool-per-message behavior for backward compatibility
• Protocol detection based on presence of tool call ID (native calls have IDs, XML calls don't)

2. Tool Failure Guardrail System

• Added didToolFailInCurrentTurn flag to prevent attempt_completion after tool failures
• Flag resets at start of each new API request cycle, scoping failures to single assistant messages
• Updated all 19 tools to set the flag when encountering errors
• Prevents scenarios like: read_file: failure! → attempt_completion: blocked!
• Allows user override in subsequent turns: read_file: failure! → user message → attempt_completion: allowed!

3. System Prompt Updates

• Updated tool-use.ts to clarify native protocol supports multiple tools per message
• Removed "ALWAYS wait for confirmation" guideline that only applies to XML protocol
• Adjusted tool-use-guidelines.ts to be protocol-aware

4. Tool Result Handling

• Implemented duplicate tool_result prevention for native protocol
• Each tool call ID only gets one tool_result block
• Maintains separate handling for XML vs native protocols

Technical Details

Protocol Detection:

const toolCallId = (block as any).id
const toolProtocol = toolCallId ? TOOL_PROTOCOL.NATIVE : TOOL_PROTOCOL.XML

Failure Tracking:

• Set on error: task.didToolFailInCurrentTurn = true
• Reset on new turn: In startNewApiRequest()
• Checked before completion: In attempt_completion tool

Tools Updated (19 total):
ReadFileTool, WriteToFileTool, ApplyDiffTool, InsertContentTool, ExecuteCommandTool, BrowserActionTool, GenerateImageTool, UseMcpToolTool, AskFollowupQuestionTool, SwitchModeTool, NewTaskTool, UpdateTodoListTool, CodebaseSearchTool, SearchFilesTool, ListFilesTool, ListCodeDefinitionNamesTool, FetchInstructionsTool, RunSlashCommandTool

Verified Behavior

✅ XML Protocol: Single tool per message with step-by-step confirmation
✅ Native Protocol: Multiple tools per message with intelligent chaining
✅ Failure Blocking: Tool failures block attempt_completion in same message
✅ Cross-Turn Recovery: User can override in subsequent turns

Example Scenarios

Native Protocol - Success:

read_file: success → apply_diff: success → attempt_completion: success ✅

Native Protocol - Failure (Same Turn):

read_file: failure → attempt_completion: BLOCKED ❌
Error: "Cannot execute attempt_completion because a previous tool failed in this turn"

Cross-Turn Recovery:

Turn 1: read_file: failure ❌
Turn 2: user: "just complete the task"
Turn 3: attempt_completion: ALLOWED ✅ (flag reset at turn start)

Testing

• All 4133 tests passing ✅
• Added comprehensive test coverage for attempt_completion guardrail
• Updated test snapshots for prompt formatting changes

Breaking Changes

None - XML protocol behavior unchanged, native protocol gets new capabilities