add llm_request event (or onPayload)

[linktohack]

This helps us debug the payload sent to the provider.

The change can be very simple

+++ b/packages/agent/src/agent-loop.ts
@@ -234,6 +234,9 @@ async function streamAssistantResponse(
                ...config,
                apiKey: resolvedApiKey,
                signal,
+               onPayload: (payload) => {
+                       stream.push({ type: "llm_request", payload });
+               },
        });
        

Then we can have an extension to log the request and response (existing today) from provider

        pi.on("llm_request", async (event) => {
                log({
                        type: "llm_request",
                        timestamp: new Date().toISOString(),
                        payload: event.payload,
                });
        });

        // Log assistant responses
        pi.on("message_end", async (event) => {
                if (event.message.role === "assistant") {
                        log({
                                type: "llm_response",
                                timestamp: new Date().toISOString(),
                                message: event.message,
                        });
                }
        });
        

PR is ready and working fine for me for a few weeks now (was very helpful when I wanted to debug llama.cpp chat parser)