jev-desktop

jev-desktop

This skill requires a clone of the agent-desktop repository (the scripts live in scripts/jev/ and ship in no installed package) and a TYPESAFE_API_KEY.

Two entry points over the same screen reading. run.mjs takes a goal and drives until it is met. act.mjs takes one step and hands the plan back to you. Neither puts the tree in your context.

The loop

node scripts/jev/run.mjs --app Finder "open the Applications folder"
node scripts/jev/run.mjs --app TextEdit --cursor --text "A line to write." "write the supplied sentence into the document"

Each turn reads the screen, asks for one operation and a target for that operation in the same request, carries it out, and reads again. It prints one JSON line per turn and one when it stops (shown trimmed).

{"turn":{"step":1,"operation":"CLICK","target":"treeitem \"Applications\"",
         "confidence":1,"ok":true,"delivery":"delivered_verified","changed":true}}
{"stop":"done","confidence":0.96}

The operations are CLICK, TYPE_TEXT, CHECK, UNCHECK, EXPAND, COLLAPSE, SCROLL, DRILL, WIDEN, WAIT, DONE and BLOCKED. Only the ones something on screen can receive are offered: DRILL for a region that reports it holds more, WIDEN only after a DRILL. SCROLL always scrolls down. delivery is waited for WAIT and looked for DRILL and WIDEN.

The target is asked once per operation. click_target is chosen from the elements that advertise Click, type_text_target from those that advertise SetValue, and so on. Only the head matching the chosen operation is read, so a target can never be incompatible with the verb that acts on it.

Looking is an operation. A window is read with --skeleton first, so a dense document costs the same first look as a small panel. A region that was cut off reports how much it holds, and DRILL pins it as the root for later turns. WIDEN gives the whole window back.

CHECK and UNCHECK instead of a toggle. Both are idempotent: a box already in the wanted state stays there.

You supply the text. Nothing here writes a value, so a run never puts a string on screen that you did not choose. Pass --text once per value and they are consumed in order, or pass a function as the text option and it is asked for each field with that field's description. When there is no value left to give, TYPE_TEXT is not offered at all, so the run stops rather than inventing one.

Text goes in by whichever route the application accepts. A direct value write (set-value) is one verified call. Only when the application refuses it and the refusal says a retry is safe does the value go through the clipboard and a cmd+v paste (macOS only). The clipboard is put back when the run ends, or emptied if it held anything but text. One key press per character is never used: it drops characters and loses capitals.

What leaves the machine. Every turn posts a description of the screen to the decision endpoint (TypeSafe's api.typesafe.ai by default, or whatever TYPESAFE_BASE_URL points at): each element's role, its accessible name or description, up to sixty characters of the value it holds, its state, the window title, and the recent actions, excluding supplied text and detailed execution evidence. That is enough to send the contents of a private document or a filled form to the configured remote service. Secure text fields are already withheld by agent-desktop and never reach the request. Pass --no-values to either script to withhold what every other field holds as well; targeting gets harder, because a value is often the only thing that tells two unnamed rows apart. The default is to include values. run.mjs redacts explicit supplied strings from action turns and later recent-action history, including when later target names or execution evidence repeat them. Live screen descriptions remain separate: if the application displays that text, its accessible name or field value may still reach the remote service under the settings above. act.mjs preserves its exact executable argv in stdout, so that field can include a value supplied with --text; stdout may also contain accessible names and element paths. Choose the endpoint and output destination with that exposure in mind.

A step that is hard to undo is not taken quietly. An ordinary step needs 0.70 confidence in its target, one rated destructive (0.50 or more) needs 0.90, and below 0.55 nothing runs. When the bar is not met the run stops and names the candidate it would have acted on, so you decide instead of it. act.mjs gets the rating in its single request. run.mjs rates risk with a second model call (counted toward the 80), only for targets at 0.70 up to 0.90, so by default a destructive step at target confidence 0.90 or higher runs unrated. If that call returns no number, the step counts as destructive. Pass --confirm-destructive to run.mjs or act.mjs to stop for confirmation on every step rated destructive, regardless of confidence.

--cursor makes the run watchable. It starts a session, shows a cursor that travels to each element before the operation lands, and turns it off at the end. If the session cannot start, the run continues without a cursor. The cursor is drawn where the element is, so bring the application in front of your terminal or it arrives behind it.

It stops on DONE, on BLOCKED, on a confidence too low for the risk, after 40 operations (WAIT, DRILL and WIDEN count), after 80 model calls (risk ratings count), or after three turns in a row that changed nothing (WAIT turns do not count). It also stops with action_failed when an operation fails, and with nothing on this screen can be acted on when no element advertises an action.

A screen with more actionable elements than a choice can carry says so, in the request and in the turn it reports (truncated), and the policy is told to look inside a region rather than call the goal impossible.

--root @ref starts inside a region when you already know which one.

One step at a time

You keep the goal, the plan and the memory. You send one sentence. You get back one small object.

node scripts/jev/act.mjs --app TextEdit --execute \
  --text "Morning over the dock." "type this text into the main writing area"

{
  "ok": true,
  "intent": "type this text into the main writing area",
  "app": "TextEdit", "window": "Untitled 3", "window_id": "w-4521", "surface": "sheet",
  "element": { "ref": "@s1:e139", "role": "textfield", "name": "Save As:",
               "where": "sheet \"save\" > group" },
  "command": "set-value",
  "argv": ["set-value", "@s1:e139", "poem.txt"],
  "decision": "act",
  "why": null,
  "confidence": { "target": 0.99, "command": 0.88 },
  "gates": { "present": 0.94, "destructive": 0.31, "needs_text": 0.87 },
  "runner_up": [{ "ref": "@s1:e137", "what": "textfield", "p": 0.02 }],
  "notes": [],
  "executed": { "ok": true, "delivery": "delivered_verified" }
}

Calling it

Flag	Meaning
--app <name>	Required.
--window-id <id>	Read one window of the app, by the id agent-desktop list-windows --app <name> reports. Needed only when the app has more than one visible window and none of them is focused, which otherwise fails with AMBIGUOUS_TARGET. The run stays on that window: a window the run opens elsewhere is not read. If the window closes after an action, run.mjs reports that turn and stops with window_closed. A --root or an open surface outside the window gives root_outside_window or surface_outside_window (a stop in run.mjs, the error in act.mjs). Every run.mjs stop reports the window it read in screen.window_id. run.mjs takes it too.
--no-values	Withhold accessible field values from the decision request. Supported by both scripts; the default includes values.
--execute	act.mjs only. Run the command when decision is act. Without it, nothing runs.
--text "…"	Text the intent needs. Jev returns choices, never strings. run.mjs takes it repeatedly; act.mjs takes one value.
--root @ref	Resolve inside one container instead of the whole window.
--bin <path>	act.mjs only. agent-desktop binary. Defaults to the release build, then PATH. run.mjs reads AGENT_DESKTOP_BIN instead, with the same default.

Environment: TYPESAFE_API_KEY is required (both scripts fail with TYPESAFE_API_KEY unset). TYPESAFE_BASE_URL and TYPESAFE_MODEL are covered below. JEV_VERIFY_MS sets the paste read-back deadline in ms (default 2000, 0 to 60000).

argv is the exact agent-desktop command. double-click and hover carry --headed; scroll always goes down.

Phrase the intent as an action, not as an element. "type this text into the main writing area" resolves to type. "the main writing area" names no operation and can resolve to focus. Describe the target the way a person would, "the field holding the name the file will be saved under", not the way the tree names it.

Reading decision

decision	What it means	What you do
act	One element clearly matches and the confidence clears the bar for this action's risk.	Nothing. With --execute it already ran.
confirm	The match is plausible but under the bar.	Ask the user, or re-phrase the intent and call again.
abstain	Jev answered none, the element is probably not on this screen, or two elements fit equally.	The screen is not where you think. Open the surface you need, then call again.
needs_text	The command takes text and none was supplied.	Call again with --text.

why carries the reason in one sentence for every decision except act (where it is null). runner_up shows what else it considered, which is usually enough to tell a wrong screen from a vague intent.

What the script decides, not Jev

• Risk sets the bar. The answer says what; confidence says whether to act. An ordinary action needs 0.70. One that Jev rates destructive at 0.50 or more needs 0.90: writing a file, deleting, sending, confirming a warning. Below 0.55 nothing acts at all.
• The command is reconciled against the element. Both questions are answered in parallel and neither sees the other, so code checks the chosen verb against the element's advertised actions. A readonly combobox gets click, never set-value. A --text payload with a non-text verb corrects the verb, because the caller supplying text is evidence Jev does not have.
• An open surface wins. When a sheet, alert, menu or popover is up, the script reads the window, then that surface, and offers the surface's elements instead of the window's. In the window tree those elements carry offscreen and would all be dropped.
• Disabled, hidden and action-less elements are never offered, nor a cell inside a treeitem (the row beside it acts). Everything else is, including unnamed rows: a row is told apart by the value it holds, and a Choice does better with the full list than with a shortlist.

How the request is shaped

act.mjs sends one call with five questions. They are answered in parallel, so the speculative ones cost tokens and no latency.

Question	Type	Asks
target	choice	Which element the intent refers to, plus none.
command	choice	Which of the 16 interaction verbs it asks for.
present	noul	Is the thing on this screen at all?
destructive	noul	Would this be hard to undo?
needs_text	noul	Does this need text from the caller?

A Choice accepts 255 options, so up to 254 elements go in one pass (both scripts). When the first act.mjs pass lands under 0.70, the top five are re-asked with richer descriptions. A screen with more than 254 elements says so in notes; use --root @ref.

Deciding through OpenRouter

The decision endpoint is TYPESAFE_BASE_URL, defaulting to TypeSafe's native https://api.typesafe.ai/v1/systemone. Configured endpoints must use HTTPS; HTTP is allowed only for localhost or a literal loopback IP, which keeps local test routers usable. URL credentials are rejected, and redirects fail closed so the bearer key cannot be forwarded to another endpoint. Point it at OpenRouter's Decisions router to use an OpenRouter key as TYPESAFE_API_KEY:

TYPESAFE_BASE_URL=https://openrouter.ai/api/alpha/decisions \
TYPESAFE_MODEL=typesafe/jev-1.13 \
TYPESAFE_API_KEY=sk-or-... \
node scripts/jev/run.mjs --app Finder "open the Applications folder"

TYPESAFE_MODEL may stay at its jev-latest default: against OpenRouter a bare name is sent as the registered ~typesafe/<name> alias. Or pin a routed id such as typesafe/jev-1.13. Do not set typesafe/jev-latest: OpenRouter answers 400 Model does not exist. OpenRouter's Decisions endpoint is alpha and may change shape.

Known limits

• act.mjs reads the screen once per call. run.mjs reads it before and after each action, and skips the read after DRILL or WIDEN. Without a pinned region, each read takes a second snapshot when a sheet, alert, menu or popover is open. A read takes about 3 s on a dense app; --root is far cheaper.
• act.mjs does not retry: a 429 ends the call. run.mjs retries 429, 503 and 529 twice, after 500 ms then 1000 ms, before it fails.
• Held input is unavailable, so no sustained key and no drag with hold. See crates/core/src/commands/input_hold_policy.rs.
• press is not resolved here. It needs no element, so send it directly.

Files

File	Holds
scripts/jev/policy.mjs	The operations, how a screen is described, and when to stop. No application is touched here.
scripts/jev/screen.mjs	Which elements a command can reach, which surface is open, how an element is described. Shared by both entry points.
scripts/jev/desktop.mjs	The only code that speaks to agent-desktop.
scripts/jev/run.mjs	The loop and its command line.
scripts/jev/act.mjs	The single-step resolver.

Checks

WAIT in run.mjs pauses 250 ms before the screen is read again. A failed operation stops the run at once as action_failed, keeping the error code and delivery state, and is never repeated. A failed focus, clipboard write or paste ends the run the same way. A binary whose failed set-value reports no error disposition fails with BINARY_TOO_OLD.

run.mjs exits 0 only when the run stops at done; every other stop, and any error ({"ok":false,"error":…}), exits 1, and a usage error exits 2. act.mjs exits 0 unless the executed command failed or was not executed (executed.code is NOT_EXECUTED when decision is not act); an error exits 1 and a usage error exits 2.

After a paste the field is read back. If it does not hold the requested text the run stops with TEXT_VERIFICATION_FAILED and does not type again. The read-back compares the exact string, so a field that normalizes text (trims, changes case or line endings) reports a failure even when the text landed. The read is retried until the JEV_VERIFY_MS deadline before it fails.

node scripts/jev/act.test.mjs
node scripts/jev/run.test.mjs
node scripts/jev/run.integration.test.mjs
node scripts/jev/evidence.test.mjs

Turns keep native steps, post_state and details for the caller; they are excluded from model history. A verified AlreadyInState step with outcome: "skipped" and delivery: "not_delivered" means no input was needed. A failed postcondition keeps ok: false and any nested action evidence.