lac-cli /watch: Your Browser Session, Sent Straight to the Agent

The debugging loop I got tired of

You click through a form, something breaks, and then you spend the next ten minutes reconstructing what happened. Open DevTools Network tab, scroll through a wall of requests, copy the failing XHR body into a message, paste the error from the console, try to remember which button you actually clicked. Then you explain all of that to an AI and hope the context is enough.

It usually isn't. You leave something out, the AI guesses wrong, and you're back to square one.

/watch in lac-cli fixes the reconstruction problem. It runs a browser monitoring session that captures the whole thing automatically — every click, input, navigation, network request, and JS error — then puts a floating button in your browser that sends the entire timeline to lac agent in one shot.

Starting a session

Once you have lac-cli installed (pip install lac-cli or the curl installer at lacai.io/lac-cli), starting a watch session is one command:

Then from inside the agent, type /watch. That's it. A browser panel opens and a small floating "Send to AI" button appears in the corner of the page. From that point on, everything you do in that browser window is being recorded.

What it actually captures

This is the part that surprised me the first time I used it. It's not just "you clicked here." The session timeline includes:

• Clicks — which element, what its text or label was
• Inputs — what you typed into form fields
• Navigation — every URL change, including redirects
• XHR and fetch requests — method, URL, request headers, request body, response status, response body. Yes, full bodies.
• JavaScript errors — message, stack trace, and which line threw it

That last two are the ones that matter most on a typical debugging session. When a POST to /api/checkout comes back 422 and your UI just shows a generic error toast, you want to know what was in the request body and what the server actually said. /watch has both. You don't have to touch DevTools at all.

The "Send to AI" button

After you've reproduced whatever you were looking at, you hit the floating Send to AI button. The full session timeline — formatted, structured, everything from the first click to the last network response — gets piped directly into lac agent.

At that point you can just ask it something. "Why did that checkout request fail?" or "What triggered the TypeError on line 84?" The agent has the complete picture. It's not working from your paraphrasing, it has the raw data.

I used this last week on a bug where a file upload was silently failing. I'd been staring at it for 30 minutes. Ran /watch, reproduced the upload, hit Send to AI, asked "what went wrong with that multipart request?" — the agent spotted a missing Content-Type boundary in about ten seconds. That was it.

Using it alongside lac agent's other tools

Because /watch runs inside lac agent, you get the rest of the agent's toolset right after you send the session. The agent can read the relevant source files, write a fix, show you a diff, and wait for you to confirm before touching anything. The undo stack is there if the fix makes things worse.

The flow ends up being: reproduce the bug in the browser, send the session, let the agent find the cause in the timeline, let it trace that back to the source file, review the proposed change, confirm. Five steps, no copy-pasting, no switching between five different windows.

If the task is complex enough that it would benefit from parallel work, you can break out of the single-agent flow into /multi mode right from the same session — something we covered in the /multi post last week.

When /watch is most useful

Not every problem needs it. If you already know exactly where the bug is, just fix it. But /watch earns its place in a few specific situations:

• A multi-step user flow is broken and you're not sure which step is the actual failure point
• An API call is returning something unexpected and you want both the request and response captured without opening DevTools
• A frontend JS error is happening in a context that's hard to reliably reproduce and describe
• You want to hand a full reproduction to someone else (or to the AI) without writing a long bug report

It's also genuinely useful for auditing what a third-party script on your page is actually doing. Drop it in, click around, send the session, ask the agent to summarize all outbound network requests. You'll see things you didn't expect.

Getting set up

If you haven't installed lac-cli yet:

Or use the shell installer:

Run lac agent, pick your provider (Claude, GPT, or a local Ollama model), and type /watch. The first session takes about a minute to set up, then it's instant after that.

One practical tip: keep the session focused. Don't record 20 minutes of random clicking and then send it — the timeline gets noisy and the agent has to sift through a lot of irrelevant events. Reproduce the specific thing that's broken, hit send, ask a targeted question. That's the workflow that actually moves fast.