Deep Inspection: the agent that asks its own questions

What is different about it

The recording is handed to the helper once, and from then on the agent reads it through a set of tools rather than being sent a fixed excerpt. It can list events, filter them, open one in full, search inside response bodies, pull the video frame at a given moment, and ask what happened around a request. It follows what it finds.

Two consequences worth knowing before you start. It is slower and it costs more, because it makes many calls rather than one. And it runs on your machine, so whether anything reaches a cloud model is decided by the agent you configured, not by webQsee.

Choose the agent, and where the conversation happens

Selecting the Deep Inspection card opens one row inside it: Which agent, and where does the conversation happen? It is a single question because the answer is a single setting. The list is grouped by vendor, one group per CLI you have, each collapsed to a line that says how many profiles are under it, so six agents with three profiles each stay a list you can read rather than eighteen radio buttons. Open the one you want and pick a profile. The three below are what those profiles are, and all three run the same agent on the same recording.

Run one

1. Press Analyze, pick Deep Inspection, pick "in this window"

   If Deep Inspection is greyed out, the card says why: the helper is not reachable, or, in a browser tab, Deep Inspection is not set up yet, or your browser has not allowed (or is blocking) the page reaching the helper. It says so on the card itself rather than letting you press it and wait.

   Sanity check: the sheet's status line reads "The webQsee helper is running" with a version number.

2. Ask the first question

   Nothing is transferred until you do. Opening the panel costs nothing: the recording is handed over when the first question is asked, which is also when you will see the progress line move through preparing, sending and reading.

   Sanity check: the panel shows the agent's steps appearing one by one, not just a spinner.

3. Close the tab if you want to

   The analysis keeps running. The agent is a detached process on your machine, not something owned by the browser tab, and when you come back to the recording the panel offers to pick the conversation up rather than starting a second one.

   Sanity check: reopen the recording and press Analyze. You are asked whether to carry on the earlier analysis, and choosing that shows the answer that was produced while nothing was watching.

4. Push on the answer

   The conversation stays open and the agent still has the recording. Follow-ups are where Deep Inspection earns its cost:

   "Was that 500 the first one, or had it been failing before the recording started?"

   Sanity check: the second answer cites different rows than the first. If it cites the same ones, the question was probably a rephrasing rather than a new one.

5. Or carry on in your own terminal

   Under the conversation, Continue in a terminal has a line to copy, and the sessions list in the AI Analysis settings has the same one per conversation:

   wqs-helper continue <session>

   It opens the same conversation with the same agent, in the folder it ran in, and switches the webQsee tools on for it, with a token for that conversation that is revoked when you leave it. Cursor and Antigravity are the exception: their CLIs take MCP servers only from their own global configuration, so there the tools are available only if you registered the helper with them once.

   Sanity check: ask the continued conversation to list its tools. The webQsee ones, such as get_overview, are there (with Cursor and Antigravity, once you have registered the helper with them). Resuming with the agent's own command instead (claude --resume and the like) starts it without them, and an agent without them tends to report that the recording does not exist.

Let it read your code

This is where most of the value is. The recording says what happened; the source says why. Map the application's origin to a checkout on your machine and the agent can open the file behind the request it is looking at.

1. Map an origin to a directory

   In the extension's analysis settings, add a project mapping: the origin of the site you record, and the local directory of its source. Wildcards are allowed in the host part, for example https://*.example.com.

   Sanity check: start an analysis of a recording of that site. The destination line above the question box now names the project instead of saying no project source is in scope.

2. Decide whether it may change code

   Reading is the default. Writing is a separate, explicit choice per analysis, and it is only ever possible when a project is mapped in the first place.

   Sanity check: with fix mode off, the agent proposes a change in prose. With it on, it can edit the file, and the panel says so before you start. Either way nothing is committed, branched or backed up: review the edits in your editor or with git diff, as you would after any coding agent.

A whole collection at a time

Select several items, or open a collection, and analyze them together. The agent can then answer the question that is actually being asked most of the time: what is different between the run that works and the run that does not. See comparing two recordings.

Everything in the collection goes to the agent, not only the recordings. It looks at the screenshots, reads the Excel events overview as a table, and reads an HTTP archive as the event snapshot it would become on import, so it can compare that one too. The collection's description and notes go along as context. The analysis is then saved on the collection itself, together with a record of what the collection contained at that moment; if items are added or removed later, the saved analysis says so.

Your own files, with a question

The agent reads the recording by itself, but some things only you have: the spec, the server log, the screenshot a customer mailed. Attach them to a question with the paperclip, drop them onto the question box, or paste a long log, which becomes an attached text file the way it does in a chat app. They go to the webQsee helper on your machine for that question. webQsee keeps their names and sizes with the analysis and never their content.

Troubleshooting

Next