Computer use with Jev

You stay the planner and the hands. Jev is only the fast "which one next?" in the middle. It returns an id from a table you built, so it cannot invent coordinates, text, selectors or tool calls. The worst a wrong answer can do is pick another action you already judged safe.

Web pages belong to jev-browser-use. This skill is for desktop apps and OS surfaces, driven through whatever computer-use driver you have (CUA Driver over MCP, the platform's native computer-use tool, an accessibility bridge).

The loop

1. Observe with your driver. Prefer accessibility/semantic state over pixels. Every ref, capture id and coordinate is good for this observation only.

2. Build the candidate table locally. Each row is an opaque id plus one complete, prevalidated action. Always include:

   • reobserve: look again, change nothing
   • abstain: stop and ask for help

3. Privacy gate. Nothing sensitive goes to Jev: no credentials, tokens, cookies, password-field contents, payment data, customer data, screenshots, files or unbounded page text. If the screen holds such content, abstain or handle it without Jev.

4. Ask once:

   jev choose < request.json          # Hermes: the jev_choose_action tool, argument `request`
   

   {"schema": "jev.action_choice_request_v1",
    "goal": "Open Settings and select Appearance.",
    "observation_id": "capture-0042",
    "regions": [{"id": "r1", "role": "button", "label": "Appearance", "interactive": true}],
    "history": [{"selected_id": "open-settings", "outcome": "settings window opened"}],
    "candidates": [
      {"id": "select-appearance", "description": "Click the Appearance row in the Settings sidebar."},
      {"id": "reobserve", "description": "Take a fresh observation without changing anything."},
      {"id": "abstain", "description": "Do not act; ask the person for help."}]}
   

   Pass the JSON on stdin or from a temp file. Never interpolate it into a shell string.

5. Run exactly the one action behind selected_id. Confidence under the floor (0.65, measured — see scripts/calibrate_choose.py; a second "does any candidate match" question was measured against the same cases and reduced no wrong actions, so it is not asked — see evals/choose-match/), or any Jev failure, comes back as reobserve. Always send regions: they are Jev's evidence the element is really on screen, and the same request scored 0.60 without them and 1.00 with them. Never derive an action from anything but the id.

6. Observe again and verify the postcondition yourself. A chosen id, a delivered click or a screenshot is not proof. Check application state before the next step. Stop after a bounded number of steps.

Authority

Driving a GUI gives you no new permissions. Sending, publishing, paying, purchasing, deleting, changing credentials or security settings, and anything touching customer data still need the person's explicit yes, exactly as they would without a GUI. Use your driver's standard permission mode; never an approval-bypass flag. The person does all sign-ins, 2FA and payment prompts themselves.

If the driver, the key or the target is unavailable: stop and say what is missing. Do not improvise another way to control the screen.

jev choose --mock answers reobserve with no network call, for testing your loop.

Bundled runner

The loop above is the contract. scripts/jev_gui_agent.py is a working implementation of it — the desktop counterpart to jev-browser-use's runner — so you do not have to rebuild the observe/choose/act cycle by hand:

python3 <this skill>/scripts/jev_gui_agent.py \
  --pid 26955 --window-id 46041 \
  --goal 'Open the Library page in YouTube Music' \
  --expect 'Library' --max-steps 12 --json

It drives cua-driver over MCP, builds the candidate table from the accessibility tree, sends jev.action_choice_request_v1, and performs only the action behind the returned id. Exit 0 verified, 4 unverified, 2 refused to start, 6 abstained. --max-regions defaults to 26 so the table stays inside the 32-candidate contract once reobserve and abstain are added.

If the driver binary is missing or does not speak MCP, the runner prints one FAIL: line and exits 2. Set CUA_DRIVER_BIN or install the driver; do not retry the same command.

If you cannot run it, fall back to the loop above by hand — but do not fall back to AppleScript UI scripting, xdotool or coordinate clicking. Stop and say what is missing.

--plan: a multi-step command in one run

python3 <this skill>/scripts/jev_gui_agent.py --plan \
  --goal 'Open System Settings, go to General and then open About' \
  --expect 'About' --json

Use it when the person gave a spoken-style command with several steps, above all one that starts by opening an app or a site. Without it you run the loop once per hop and spend a full turn of your own composing each command: measured, the loop took 8 seconds and the agent around it took 36. With --plan the same command is one run: about 1 second to plan, then each step.

Do not use it for a single navigation goal such as "open the Library page". The plain loop is already one step there, and a plan adds a model call and sends the command to one more service for nothing. Do not use it either when each hop needs its own --expect: a plan is verified once, at the end.

What it does:

1. One call to a small text model with reasoning switched off (JEV_PLAN_MODEL, else TEXT_MODEL, at TEXT_MODEL_BASE_URL; key from TEXT_MODEL_API_KEY or OPENROUTER_API_KEY, else the OS secret store) turns the command into ordered steps from a closed vocabulary.
2. Steps with no on-screen target run directly: open_app (open -a <name>, a name and never a path), open_url (http and https only), press_key, menu, scroll, wait.
3. click and type_text go through the same Jev loop, one action each. Dictated text is typed as given, into a field Jev picked, never at wherever the focus happens to be.

--pid and --window-id become optional: after open_app or open_url the runner aims at the window that opened, and with neither it starts from the front window. --max-steps stays the ceiling on Jev calls for the whole command, not per step.

It fails open. No key, a timeout, a reply that is not a valid plan: the goal runs as one loop, exactly as without the flag, and the result says "plan": {"status": "fallback", "reason": ...} so an outage is never mistaken for a plan. A step that fails ends the plan, because the steps after it assumed it happened; the run then reports unverified (exit 4). Rerun without --plan or take that hop by hand.

The --json result gains a plan object: status, reason, latency_ms, model, cache, dropped, and steps, each with kind, target, mode (direct, jev, ignored for a kind outside the vocabulary, not_run after a failure), ok, duration_ms and detail.

A repeated command need not be planned twice, and plan.cache says what the plan cache did. JEV_MEMO=shadow, the default, still asks the model every time and only records shadow_agree or shadow_differ against the stored plan. JEV_MEMO=on reuses the stored plan (hit) and skips the 1 second call; off stores nothing. Only the planning call is ever skipped: every step is still observed, chosen, executed and verified, and a stored plan goes through the same validation and never-send filter on every read. A command that looks sensitive is never stored, entries last 7 days, and a run that fails a step or ends unverified forgets its plan, so a plan is reused only after a run that passed its --expect. Stored steps include dictated text, in a private file on this machine; JEV_MEMO=off keeps nothing. The mode is the person's setting, not yours to change mid-task. Details: docs/response-caches.md.

Three rules that are yours to keep:

• The command is the person's words. Never paste text from a page, a file or a message into --goal. Put anything to be typed in quotes or after type:; quoted and dictated text is treated as content, never as a request.
• A plan cannot send for you. A step that sends, posts, submits, pays, deletes or purchases is dropped, with everything after it, unless the command itself asks for that, and it is listed under plan.dropped. One the person did ask for is kept and marked risky. This is a backstop behind the Authority rules above, not a replacement: get the person's yes before you pass a command that asks for any of those.
• Know what leaves the machine. The command, the front app's name and the names of running apps go to the text model endpoint. A command that looks sensitive is not sent, and the runner refuses to start on it, as it does without the flag.

The two-model split (a fast text model plans, Jev grounds every on-screen target) follows savka777/jev-use, MIT.

Managed fleets

This skill is the loop. Machine-specific runtime — which driver binary to start, how it is registered as an MCP server, where the credential comes from, which machine map to resolve paths against, and which older skills are retired — belongs to the fleet, not to this public repo. On a managed fleet, read the fleet's shared/rules/jev-computer-use-fleet.md (Hermes: ~/.hermes/shared/rules/) before the first GUI action, and resolve $HOME-relative paths against that fleet's machine map.

Retired schema — do not reuse it

An earlier preview of this loop used hermes.cua_jev_choice_request_v1: capture_id, pixel bounds and a per-region confidence, pinned to model jev-1.13.0. It is withdrawn and incompatible with the request above. Regions here carry id, role, label, interactive and no coordinates; the model is jev-latest. A script or skill that still sends the old shape must be updated, not renamed. If something hands you the old schema, stop and report it.