Problem

When using native tool protocol, if read_file encountered an error (e.g., file not found), it would create two tool_result blocks with the same tool_call_id:

1. handleError() internally calls pushToolResult() → tool_result Add options to always approve write and execute operations #1
2. Code continues to final pushToolResult() with XML → tool_result Fix vscode compatibility issue #2

This violates the native protocol requirement of exactly one tool_result per tool_use_id, causing 400 errors on subsequent API calls.

Solution

Replace handleError() with task.say("error", ...) in error paths within ReadFileTool.

Agent still receives error details through the XML in the single final pushToolResult() call:

<file><path>nonexistent.txt</path><error>Error reading file: ENOENT</error></file>

Why This Doesn't Break Anything

Native Protocol

• ✅ Only ONE pushToolResult() call = only ONE tool_result per tool_call_id
• ✅ Agent receives error via XML: <error>Error reading file: ...</error>
• ✅ User sees error via task.say()

XML Protocol

• ✅ Only ONE text block with complete XML (cleaner than before with separate error blocks)
• ✅ Agent receives error via XML (same as native)
• ✅ User sees error via task.say() (same as before)

Both protocols: Agent error visibility is preserved through XML in the tool result.

Testing

• ✅ All 44 tests passing in readFileTool.spec.ts
• ✅ All pre-push tests passing (315 test files, 4069 tests)
• ✅ Updated test to verify say() is called with error messages