Talk to your Mac.
Jev decides. The app verifies.

One request routes the clause

Every clause the local parser can't settle becomes a single systemOne call with the clause, the full transcript, the frontmost app, installed apps and aliases as state — and seven questions.

One Choice picks the next step

For UI work the app reads the target window's Accessibility tree, ranks the buttons, links, rows, tabs and text fields by overlap with the goal words, and asks Jev one next_action Choice over those real candidates plus done and stuck. Alongside it: goal_reached, wrong_surface, needs_text as probabilities.

candidates:
  [1] AXButton  "New session"
  [2] AXTextArea "Ask Devin…"     ← dictation goals see only text inputs
  [3] AXRow     "Check RAM and CPU consuming resources"
  [4] OCR       "Send"  (x=1412,y=884)  ← when the tree is thin
  done · stuck

next_action → 1   (p=0.91)
goal_reached → 0.08

Arguments — URLs, queries, the words to type — are extracted deterministically by SlotExtractor: "type in the prompt box to check CPU" types check CPU. Words that are content never influence which control gets picked.

A complete, fresh tree

The in-process walker spends up to 1.5 s (4 s on a full re-observe) and marks snapshots partial when it hits the deadline. Electron apps expose a thin tree until touched, so a wake-up pass focuses the window and re-walks. A partial tree is never evidence that the UI is stuck.

Text goes into text fields

When the goal contains text to type, only AXTextField, AXTextArea, AXSearchField and AXComboBox are candidates. Create/send controls fire at most once per task; if the screen already changed, the step is done regardless of Jev's confidence.

Real events, grounded geometry

Clicks use AXPress; if the element refuses and its frame is known, one real CGEvent click at the frame centre. Typing is in-process Unicode key events into the focused field. Nothing is ever aimed at a coordinate a model made up.

Postconditions, not vibes

Each action has its own check: the field contains the text, the element count changed, the prompt box emptied, a window appeared. Only a verified change yields "Done". If the step budget runs out but the goal is verifiably met, that's success — not a false failure.

Push-to-talk, no guessing

Key-down starts the on-device engine, key-up finalizes. Toggle mode keeps a silence timer with a Jev complete? judgment on top (≥0.75 finalize now, ≤0.35 keep listening). Only one analyzer may own the audio stream; restarts are serialized so each utterance finalizes exactly once.

Fail clearly, fast

Every run has a 90 s wall clock; fallbacks have 6 steps / 30 s. When the ladder is exhausted the app says why — "I can't see a text box in Devin", "Allow Jev Voice to control System Events" — instead of guessing or pretending.