RunHooksBase on_tool_end result type error

[sandsc]

Traceback (most recent call last):
  File "/app/.venv/lib/python3.12/site-packages/agents/run_internal/tool_execution.py", line 1598, in _run_single_tool
    result = await self._execute_single_tool_body(
  File "/app/.venv/lib/python3.12/site-packages/agents/run_internal/tool_execution.py", line 1743, in _execute_single_tool_body
    return await self._await_invoke_task(outer_task=outer_task, invoke_task=invoke_task)
  File "/app/.venv/lib/python3.12/site-packages/agents/run_internal/tool_execution.py", line 1812, in _await_invoke_task
    return await asyncio.shield(invoke_task)
  File "/app/.venv/lib/python3.12/site-packages/agents/run_internal/tool_execution.py", line 1796, in _invoke_tool_and_run_post_invoke
    self.hooks.on_tool_end(tool_context, self.public_agent, func_tool, final_result),
TypeError: Argument 'result' has incorrect type (expected str, got dict)

Debug information

• Agents SDK version: 0.17.4

[scosemicolon]

This looks like a concrete hook contract mismatch rather than just a user callback issue.

The stack gets all the way to _execute_single_tool_body, then calls self.hooks.on_tool_end(..., final_result) where final_result is still a dict. If RunHooksBase.on_tool_end currently validates or narrows result to str, then any function/MCP tool that returns structured output can trip this path before the SDK has normalized it for model-visible output.

Two checks that would make the regression easier to pin down:

• minimal tool returning {"ok": true} with a custom RunHooksBase.on_tool_end implementation
• the same tool returning json.dumps({"ok": true}) as a control case

Expected behavior should probably be one of these, but not a runtime TypeError:

1. hook receives the raw tool result as Any, matching what tools can return, or
2. SDK normalizes structured tool results before invoking on_tool_end, and the hook type stays str.

Until this is fixed, a local workaround is to stringify dict/list tool returns before they reach the hook, but that loses the distinction between structured tool output and display text.