Using the AI Assistant (Skills · My Features · Credits)

1. First, the idea: the product ships skills, not features

The traditional approach: the vendor thinks of a scenario and builds a page for it; a site with a new question has to wait for the next release. But questions from the field are a long tail — "list the machines that didn't finish downloading today, with a one-click re-download", "show uptime per room", "which machines are still on the old image". Adding one page per new question never catches up with the rate at which questions arrive.

vDisk5 turns this around: every capability of the platform is exposed as a "skill" the assistant can call (query data, capture a screen, read logs, push commands, query a switch, run a command line…). The administrator states the goal in one sentence; the assistant picks the skills and works through them, stopping for your confirmation wherever something has to be changed. "Does the product have this feature?" becomes "can it be composed from existing skills?" — and the answer to the latter is usually yes.

So the AI Assistant is not a chat box bolted onto the product. It is a second way to operate this console:

	Traditional way	Skill-based way
Finding a feature	Hunt through menus; if it isn't there, it doesn't exist	Say it in one sentence; the assistant picks skills
A page that doesn't exist	File a request, wait for a release	Say it in one sentence, generated on the spot (see §6)
Troubleshooting	Check machines one by one, read logs one by one	The assistant gathers the evidence and shows it to you
Security	Permission checks	Permission checks unchanged, plus "high-risk requires a human click"

Three entry points

• Top-level "AI Assistant" menu in the console (placed directly under "Desktop"): Chat, Credits & Billing, Assistant Settings.
• "My Features" in the left menu: the pages you created yourself, plus "Feature Management".
• WeChat mini program: the same brain, the same skills and the same conversation history as the console.

2. Ready out of the box — no on-site configuration

Which model to use, the output cap for a single answer, and the credit price per thousand tokens are all pushed from the cloud. Any on-site server that can reach the cloud picks them up automatically; change it once in the cloud and every server follows within 5 minutes.

The "Assistant Settings" page has only two switches left:

1. Master switch — turn the assistant off entirely for this site (new installations ship with it enabled).

2. Periodic inspection — let the assistant run a health check on a schedule. ⚠ Model calls made by inspections also consume credits.

The same page shows Cloud connection status: connected or not, the billing entity, the current model, the remaining credits, the gateway address and the output cap — plus a Refetch configuration button.

Upgrading from an older version

• If this server previously had an endpoint URL, model name or API key entered by hand, those settings no longer take effect.
• There is exactly one outbound path (the cloud gateway). No silent fallback, no local direct connection. If the cloud has not yet enabled this server, the assistant plainly says "not connected to the cloud" instead of timing out. Confirm the server is enabled in the cloud before upgrading.

3. Asking questions

• Enter sends, Shift+Enter inserts a line break. Answers stream token by token.
• You can attach images — upload a screenshot or paste with Ctrl+V. Blue screens, error dialogs and "stuck on this screen" are faster shown than described.
• You can narrow the scope — name a target terminal, or restrict the assistant to specific groups.
• The process is visible. "Reasoning" and "Evidence gathering" are each collapsible (collapsed by default); expand them to see which machine it is querying and which log it is reading.
• Long tasks report as they run instead of showing a spinner.
• Stop mid-run, start a new conversation; history can be reviewed, continued and deleted; every answer can be rated up or down (down-votes are kept for regression comparison before a model change).
• The empty state offers suggested questions, refreshed each time you open it.

4. Which skills exist, and which need your approval

Skill category	Examples	Execution
Read-only platform queries	Online counts per group, download progress, terminal status, software inventory, SSD health, system logs	Assistant runs it on its own
Client screen capture	Grab the current screen of a given client and analyse it	Assistant runs it on its own
Client file read	List a directory, read a file	Assistant runs it on its own
Client NIC health check	Whether the link negotiated down to 100 Mb, whether error counts are rising (Windows and Linux)	Assistant runs it on its own
Switch port query	Per-port link state, speed, duplex, spanning tree, errors / collisions / drops	Assistant runs it on its own (read-only)
Client write operations	Write a file, delete a file, run a command on the client	Confirmation card
Console operations	Restart / shut down / wake, push a desktop, set the default desktop, change access mode, create computers · groups · scheduled tasks, and more	High-risk items: confirmation card
Server command line	Run a command on the server	Confirmation card; anything that would destroy the OS or the cloud-desktop software is refused outright
Remote assistance	Open a remote-assistance session and operate that machine	Confirmation card
Debug mode	Offer to enable debug mode to collect client logs	Confirmation card

The client side adds another layer: files are backed up before being written, deletion is really a rename, and only empty directories may be removed.

5. The confirmation card — the only gate for taking action

When the assistant is about to change something, it does not act. It raises a card stating which machine, what action, with which parameters (and, for desktop operations, exactly which desktop).

1. After you click confirm, it executes directly — the model is no longer involved. What runs is the snapshot taken when the card appeared: what you saw on the card is what is executed.

2. The approval token is valid for 15 minutes, usable exactly once, and cannot be used by anyone else; permission is still checked before execution.

3. A multi-step procedure (enter update mode → restart → install software → submit the update → check the merge) is packaged into one plan card listing every step. One confirmation, executed in order, and it stops immediately if any step fails.

4. When something is ambiguous it stops and asks rather than choosing for you. If the site has Win10, Win10-A and ATA-Win10H2-S, saying just "Win10" makes the assistant list them and ask you to pick.

6. My Features: say one sentence, get your own console page

This is where "skills, not features" becomes concrete — a page the console doesn't have is one you can simply ask for.

How to use it

1. In Chat, state what you want: "Build me a page listing the machines that didn't finish downloading today, with a one-click re-download."

2. The assistant generates it and test-runs it on the spot (visible progress: validating the definition → test-running → 12 rows returned → added to the menu). It is only saved if the test run passes.

3. Refresh the console and it appears under "My Features". From then on you just click it — no need to ask the assistant again, and therefore no further credit cost.

What such a page can do

Summary figures at the top · filters (text / number / dropdown / machine / group) · a table with configurable columns and formats (percentage, capacity, time, tag) · an optional bar, line or pie chart · inline action buttons. When there is a lot of data it states plainly "showing the first N of M rows".

Why it is safe

1. What gets created is a feature definition, not code — what to query, how to display it, which existing operations may be invoked. A generic renderer draws it at runtime. No new code is produced; deleting that one record retracts it completely, leaving nothing behind.

2. Data may only come from read-only queries. Using a state-changing operation as a data source is rejected on the spot (in testing, "restart computer" as a data source was refused).

3. Buttons may only reference operations the console already has, and clicking them still goes through the existing risk tiers and administrator confirmation.

4. Permissions are re-evaluated every time. What data you can see is recomputed against the current user on each open — sharing a feature with a colleague does not lend them your permissions.

5. The page runs no scripts; the user can only fill in values. Which operation is called with which parameters is fixed by the definition, and editing the address bar cannot change it.

Feature Management

"My Features → Feature Management" lists what you created and what others shared with you: change the sharing scope, disable, delete, or inspect the definition. Up to 60 per person, so the menu cannot be flooded.

7. Teach it your local vocabulary (Memory)

Every site has its own shorthand — "Room One" is really "Group 1", "the exam machines" means a particular group. Teach the assistant once and it understands from then on. ⚠ Memory is maintained by each site itself and is not pushed from the cloud.

8. Network and switch troubleshooting

The assistant can read port data from registered switches and settle slow links (speed/duplex negotiation), loops (spanning tree blocking), collisions and errors / drops in one pass, with per-port remediation advice and commands in Huawei / H3C / Cisco syntax.

1. It samples twice by default and judges on the delta. Error counters are cumulative since boot — several thousand on a device that has run for six months is normal, and judging by absolute values is the classic mistake here. Without sampling, the assistant states explicitly that the result must not be used as evidence.

2. Only problem ports are returned by default (a full dump of a 48-port switch would swamp the conversation); ask explicitly for everything when you need one port in full.

3. ⚠ Read-only, and only read-only: SNMP read access, no write capability whatsoever, and only devices registered by an administrator may be queried — an unregistered address is refused.

4. The switch must have read-only queries enabled and the port allowed. Where the site's switches are unmanaged, use the client-side NIC health check instead.

9. Credits and billing

• Usage is billed in credits, issued centrally by the cloud; the balance sits in the top right of the chat page and warns when it runs low.
• Self-service top-up via WeChat Pay or Alipay; credits arrive immediately after payment. The top-up history shows order and transaction numbers for reconciliation.
• Creating a "My Feature" costs one question's worth of credits; opening it every day afterwards costs nothing.
• Invoices: once CNY 200 has accumulated you can request an invoice from the console. One request can cover several top-ups and cannot double-invoice; the invoicing details are snapshotted at submission. Download it yourself once the back office has issued it.

10. On a phone (WeChat mini program)

The mini program uses the same brain, the same skills and the same conversation history as the console, and investigates just as deeply — depth is decided by the question, not by the device you asked from. You can continue a console conversation from your phone.

The phone supports Markdown-formatted answers, collapsible reasoning and evidence, confirmation cards, image attachments, scoping to terminals and groups, new conversations and stopping mid-run, history browsing and rating, a balance bar and top-up inside the mini program.

Not available on the phone

Skill-library management, memory maintenance and action-catalogue export are desk work and were not moved to the phone — use the console for those.

FAQ

Symptom	Cause and remedy
"Not connected to the cloud"	This server is not enabled in the cloud yet. Once it is, click "Refetch configuration" in Assistant Settings.
"This machine does not have that skill"	The client is out of date. New skills ship with the client, so rebuild and redeploy it.
Screen capture fails	Capture requires a logged-in user on the terminal. Linux terminals need a capture tool present, otherwise the assistant reports "not supported".
No switch data	The device is not registered, or read-only queries / the port are not enabled on the switch. The assistant explains why and falls back to terminal-side evidence.
Vague answer to "can it do X?"	The assistant carries a capability overview of the product and can look up details. When it genuinely finds nothing it says "the product most likely does not have this capability" rather than inventing a page name.
Confirmed a card after 15 minutes	The approval token expired. Ask again and confirm the new card.