Expose tool call arguments in RunHooks on_tool_start and on_tool_end

[skyfallsin]

Background

Currently, the RunHooks API only exposes the Tool instance (e.g. FunctionTool or ComputerTool) to on_tool_start and on_tool_end, but does not include the raw arguments passed in the tool call. The ToolContext object does not carry these call arguments, so hook implementers cannot inspect or log the payload that the model requested.

Proposal

1. Include arguments on ToolContext
   • Add an arguments: Any field to ToolContext, populated with the JSON-decoded payload of the tool call.
2. Populate ToolContext with arguments
   • Update runner code to call ToolContext.from_agent_context(context_wrapper, call_id, tool_call.arguments) when building the context.
3. Expose context.arguments in hooks
   • Users can then read context.arguments in on_tool_start and on_tool_end.

Benefits

• Hooks can log or audit exactly what inputs were provided to each tool.
• Enables richer lifecycle monitoring and debugging.

Example Hook Usage

class LoggingHooks(RunHooks):
    async def on_tool_start(self, context: ToolContext, agent: Agent, tool: Tool):
        print(f"Tool {tool.name} started with args: {context.arguments!r}")