Browser Use in DSH Agent Enhancer: Click, Debug, and Replay Saved Flows
dsh-agent-enhancer is a desktop plugin for DSH. Its Browser Use layer gives your agent a dedicated browser. It can click, type, and submit. Login state stays in its own profile, so your everyday browser is left alone. Y
dsh-agent-enhancer is a desktop plugin for DSH. Its Browser Use layer gives your agent a dedicated browser. It can click, type, and submit. Login state stays in its own profile, so your everyday browser is left alone.
You do not need scripting or API knowledge. Do the task once in the dedicated window, and the agent can follow the same steps. Save the recording as a flow, rerun it by name, and hand it to a scheduled task without approving every action.
The practical rule: if you click through the same web admin every day or every week, it is worth handing over. Here are scenarios that work right after install:
| Role | Repetitive work | How it helps |
|---|---|---|
| E-commerce ops | Export yesterday's orders, refunds, and inventory at 8:00 every day | Three saved flows plus the built-in scheduler, running unattended |
| Finance / reconciliation | Export admin data and statements on the 1st of each month, then reconcile three tables | Automated except for the password step; if it gets stuck, it leaves a record |
| Recruiting / HR | Go through new resumes and filter by keywords every Monday | Read-only tabs: it can read, but cannot change anything |
| Marketing / competitive research | Check competitor pricing and campaign pages every morning | Extract plus screenshots, with changes written straight into the chat |
| Dev / QA | Re-click the same reproduction steps in the admin UI | Record once, rerun by name during regression; fix once after a redesign |
| Handover / onboarding | Explain "where to click first" again and again | A teammate demos once; the steps become a flow anyone can run next month |
Three patterns cover most of these:
- Turn repetitive work into one sentence. Record once, save, run on schedule, and check the result in the morning.
- Read-only pages that must not change. Order details, reports, production dashboards: it can read and scroll, but clicks do nothing.
- Keep the "how" around. A teammate demos once, the steps become a flow; when the site changes, it finishes the job and you click Remember once.
Under the hood, the actions come from a set of browser_* tools: browser_navigate opens a page, browser_snapshot reads it, browser_click / browser_type operate it, browser_scroll / browser_wait / browser_extract handle long pages, browser_list_tabs / browser_follow_tab switch tabs, browser_network / browser_console / browser_emulate debug it, browser_demonstrate records a demo, and browser_demo_recipes manages saved flows. When you need to take over and log in yourself, use browser_ask_user; there are also browser_screenshot and browser_status.
This post starts with what it can do, then walks through six areas: operation, troubleshooting, demo recording, saved flows, scheduled tasks, and boundaries. Each section ends with a concrete use case.
Operation: multiple tabs, read-only tabs, idle cleanup
Say "open the browser" in the chat, or Alt-click (Option-click on Mac) a link in the conversation, and the dedicated window starts. It has its own profile, so you log in once and stay logged in. On Mac it uses Chrome or Safari; on Windows, Chrome or Edge. It is completely separate from your daily browser. On Mac, selecting Safari drives the WebKit engine, not the Safari app you use every day.
Multiple tabs. The sidebar lists every tab; click one to switch. When navigating, you can choose to open a new tab. If a task needs to compare orders, inventory, and shipping, you do not have to cram everything into one page.
Read-only tabs. One click in the sidebar makes a tab read-only: the agent can look, but it cannot click, type, or navigate within that tab. Opening a new tab and scrolling still work. Check-only pages β reports, order details β are the safest place to use it. Even a saved flow replay is blocked by this restriction.
Idle cleanup. Idle tabs can close automatically after 10, 30, or 60 minutes, or never. The tab you are currently viewing, session-bound tabs, and the last remaining tab are never closed. The dedicated window does not pile up.
Use it for: reconciling orders across three admin pages. Open orders, inventory, and shipping in separate tabs; make inventory read-only so nobody accidentally clicks "Adjust stock"; come back from lunch and idle tabs are gone, but the login is still there.
Troubleshooting: debug tools and page emulation
When the Export button does nothing, guessing is the worst approach. Turn on Debug & Page Emulation in settings (off by default):
- Network: method, URL, status code, duration, failure reason;
- Console: output, uncaught exceptions, source location;
- Page emulation: dark mode, reduced motion, viewport size.
It reads only what it needs: no request headers, cookies, or response bodies. Tokens and codes in URLs, plus links in the console, are redacted.
Startup failures do not require digging through logs either. The status bar and sidebar show the reason and the next step: feature not enabled, browser not started, WebKit missing, no usable browser on the machine, profile directory locked, and so on.
Use it for: a dead Export button. Check whether the network call returns 500 or times out, then look for console errors. Switch to a narrow viewport to rule out an overlay covering the button. Once you know the cause, change the agent's action instead of trying blindly.
Demo recording: do it once, it remembers
Some admin tasks are easier to show than to describe. Turn on Demo Recording in settings (off by default), click "Start demo" in the sidebar, and use the dedicated window normally. When you are done, click "End and view steps".
It records clicks, typing, Enter presses, and page navigations. It can record across tabs, and other tabs opened during the demo are captured. When finished, you can copy the steps or click "Send to agent" to insert them into the input box (the current conversation needs at least one message first, otherwise it asks you to send one). You can also say "let me show you" in the chat, and the agent can start a recording itself.
Password, CAPTCHA, token, and card-number fields are stored as *** only. It does not take screenshots or record network requests.
Use it for: handing over monthly reconciliation. Ask a teammate to demo it once: open the reconciliation page, pick the month, export, download, and check the three tables. Send the steps to the agent, and next month nobody has to ask "where do I click first?"
Saved flows: from a one-off demo to a reusable run
After recording, click "Save as flow" and give it a name, such as "Weekly report export". There are three ways to run it:
- Click Run under "Saved flows" in the sidebar;
- Say
Run "Weekly report export"in the chat. The agent finds it by name and asks for approval once per run by default; Full Access or unattended mode stops the repeated prompts; - Hand it to the built-in Automation Tasks with unattended mode, and it runs on schedule. The next section explains how to wire that up.
It works across sessions and restarts. If the current page is not the page where recording started, it returns there first. Data stays local and only redacted steps are stored; password fields remain ***.
The sidebar row keeps only Run and "β―"; rename and delete live in the menu, and delete asks for confirmation. Expanding the row shows editable values, run mode, and steps. Replay supports dropdowns and checkboxes. If several controls share a name and cannot be distinguished, it stops and lists the candidates instead of clicking the first one by name.
Parameters: plain language, not variable names
After saving, values worth changing β dates, email addresses, search terms β are detected automatically. Parameter names come straight from the page labels ("Start date"), so you do not have to invent variable names.
Date fields get a date picker. You can also say "last week", "this month", or "last quarter", and the agent converts it to a real date. When a value is missing, it asks in plain language: "Which start date should I use this time?"
Each value's default and display name can be changed; it can be set to "Ask me each time" or switched back to a fixed value. Every change goes into version history, and the last change can be undone.
Repair: when the site changes, it finishes the job
If the site renames a control or inserts an extra dialog, the flow gets stuck. The agent finishes the task, then offers a suggestion:
"The control in step 4 changed; it is now called 'Start date'. Remember this for next time?"
- Remember: write the repair back into the flow and use it next time;
- Just this once: leave the flow as it was.
Repair changes only how the control is located. It does not change values or step order. Password fields do not participate, and are never stored in the flow. Every write-back creates a version that can be undone.
Run results: what happened last time
Each saved flow records its last run: time, success / stopped at step N / failure. If it did not finish, the row shows a "Last run incomplete" badge. It stores counts and the reason it stopped, not page content.
Use it for: a weekly report export. Record once, save "Weekly report export", set "Start date" to Ask me each time, say Run "Weekly report export" with last week's start date on Monday morning, and check the run result. If the site changes, repair it once, remember it, and the next run uses the fix.
Scheduled tasks: a flow is not "placed inside" a task
First, a common misconception: a saved flow is not an asset you drag into a scheduled task. When the schedule fires, the task only says one sentence.
The built-in Automation Tasks scheduler creates a new session with fixed permissions: Workspace Modify with approval prompts. The task itself cannot change permissions. So when our flow runs, an approval prompt will appear β and nobody is there to click it. The task stalls.
The fix is not to loosen global permissions. Turn on Unattended for that one flow. It applies only to the site recorded in that flow, with a default budget of 20 minutes / 30 actions. When time or actions run out, it goes back to per-action approval. It is off by default; it cannot be enabled if the flow has no recorded site; password fields still stop and wait; it does not apply to another site or another session; if a run is denied, the site grant is revoked immediately.
These two numbers can be changed in Settings β Agent Enhancer β Features β Browser Agent: authorization duration 5β60 minutes, action limit 10β200. Defaults stay unless you change them.
Putting "Weekly report export" on a schedule (5 steps)
- Run it manually once. Sidebar "Saved flows" β expand "Weekly report export" β Run. Confirm it finishes within the authorization window and does not ask for input.
- On the same flow, click "Allow unattended". The "Unattended" badge appears next to the name.
- Create a built-in task. Open "Automation Tasks", create one, and set it to Monday 08:00.
-
The prompt is one sentence:
Run "Weekly report export". You can include the site:Run "Weekly report export" on example.com. It checks the site recorded in the flow and refuses to run on the wrong one. - Check the result on Monday morning. Expand the flow in the sidebar; "Run mode" shows the last run time and whether it succeeded, stopped at step N, or failed.
What actually happens when the schedule fires
- The task creates a new session; permissions are still Workspace Modify with approval.
- The agent finds the flow by name and issues a temporary site pass for the site recorded in the flow (default 20 minutes / 30 actions).
- The dedicated browser launches itself. Login state is still in its own profile; no need to sign in again.
- The flow replays in the recorded order, and the result is written back to the flow.
- If a site change blocks it, the agent finishes the remaining steps and leaves a repair suggestion. Unattended mode does not write the repair back automatically; click Remember when you have time.
- On a password or CAPTCHA, it stops and records the step where it stopped. That step is never auto-approved.
Three easy traps
- The flow is too long. The default pass covers 30 actions. Raise "Action limit" in settings (max 200); if that is still not enough, split "Export orders / Export refunds / Export inventory" into smaller flows. At the default budget, a 40-step flow falls back to per-action approval halfway through and then stalls.
-
A parameter is still set to "Ask me each time". Nobody is there to answer during a scheduled run, and the agent will not invent a value. For scheduled flows, either fix the value or put it in the prompt:
Run "Weekly report export" with start date set to last week. - No site recorded. Older flows that did not record a site have "Allow unattended" grayed out. Record the flow again to enable it.
Use it for: an automatic weekly report at 8:00 on Monday. Enable unattended only for "Weekly report export", tune the budget to that flow, and check the run result afterward. Other tasks and other sites still require approval each time.
If the session already uses Full Access, actions no longer prompt one by one; Auto Review goes through the built-in reviewer; View Only is refused with an explanation.
Boundaries: what it won't do
It can act, but it does not have unlimited permission. By default, every action that changes external state needs approval. It stops at a login wall and waits for you to click "I'm done, continue"; it does not bypass it.
Demo recording captures actions only: no screenshots, no network requests. Debugging does not read request headers, cookies, or response bodies; URL parameters are redacted. Saved flows live locally and contain only redacted steps; passwords never enter a flow and never participate in repair. Read-only tabs prevent changes at the entry point, and flow replays are subject to the same restriction. Turn the feature off and it removes the tools and closes the dedicated window; no background window is left behind.
The dedicated window does not label itself as an automation environment, so sites with login bot checks, such as Bilibili, no longer reject it for that reason. This is not a generic anti-detection patch; sites that check more signals still need a separate look.
Use it for: hand a read-only report page to the agent for checking; when a submission is needed, let it stop and wait for you; give a scheduled task one flow and a limited window. Narrow permissions are what make it safe to let go.
Up and running in 5 minutes
- Settings β Plugins, search
dsh-agent-enhancerβ install or update to the latest version, then restart the desktop app when prompted. - On the plugin page, confirm "Agent Enhancer" is enabled. Then go to Settings β Agent Enhancer β Features β Browser Agent (off by default) and turn it on. Choose the dedicated browser: on Mac, Auto / Chrome / Safari; on Windows, Auto / Chrome / Edge. Before selecting Safari (WebKit) on Mac, run
npx playwright install webkit; the status bar tells you if it is missing. - In the chat, say "open the browser", or Alt-click (Option-click on Mac) a link. Run one demo recording first; if it feels right, save it as a flow.
- Run the flow manually once. When it is stable, enable "Allow unattended", raise the budget in settings if needed, and create a built-in Automation Task with
Run "flow name". - Do not start with Full Access.
Command line (installs the latest version by default):
dsh plugin --profile desktop add dsh-agent-enhancer
Compatible with DSH desktop 0.2.2.x.
Start by walking it through one demo recording. After that, it remembers the run for you.
Links
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.