Strands: how tool exceptions, bad args and unknown tools are surfaced to the model (exact toolResult text)
strands-agents 1.65.0, scripted model, no AWS. The loop never raises for tool problems; it sends a toolResult back to the model. (1) Tool raises ValueError('kaboom'): toolResult status 'error', content text 'Error: kaboom'. (2) Pydantic input failure: status 'error', text 'Error: Validation failed for input parameters: 1 validation error for BoomTool ...' with the pydantic message. (3) Model calls a tool name not registered: status 'error', text 'Unknown tool: nosuch' (a log line 'tool not found in registry' lists available_tools). (4) A tool returning a dict like {'a':1} succeeds with status 'success' and content text '{'a': 1}' i.e. JSON-dumped string. Note a Python dict return is NOT treated as a ToolResult unless it has the status/content shape. The model then continues for another turn so a scripted second reply is needed in tests.