CLI reference

All commands support --help. jevdroid --version prints the package version.

Command	Effect	Calls Jev?
doctor	Find ADB and list connected device states	No
inspect	Print the active accessibility observation locally	No
benchmark --samples 10	Measure sequential accessibility reads	No
demo	Run a deterministic fake navigation task	No
run GOAL --package PACKAGE	Execute a bounded agent task	Yes
scroll --package PACKAGE --count 5	Execute a bounded scrolling task	Yes

inspect, benchmark, run, and scroll accept --serial when more than one device is attached. Diagnostic output may contain your serial number or visible UI text; do not post it publicly without reviewing it.

Run and scroll options

Option	Default	Purpose
--package	Required	Allowed application package
--backend	uiautomator	Persistent UIAutomator2 over ADB, or standard adb
--provider	vercel	vercel or typesafe
--model	Provider default	Jev model ID override
--max-steps	20	Hard ceiling on decisions, including skipped actions
--budget	0.10	Local estimated USD ceiling
--input-price	0.042	Estimated USD per million input tokens
--trace	Off	Path for a new metadata-only JSONL trace
--json	Off	JSON lines for decisions, actions, and final result

Navigation supports --allow-taps, --allow-back, and --allow-checkable (requires taps). All are off by default. Scroll supports --count (default 5) and --pause (default 0.5 seconds between gestures, excluding inference time). The count must fit inside --max-steps; allow extra decisions for launching or waiting.

Repeat --package for a navigation task spanning multiple apps. A scroll task accepts exactly one package. inspect and benchmark also accept --backend.

jevdroid run "Open the settings application" \
  --package com.android.settings --package org.example.yourapp --backend adb

The hidden interactive prompt is preferred for local experiments. Noninteractive jobs use AI_GATEWAY_API_KEY or TYPESAFE_API_KEY. JevDroid does not load .env automatically, accept key arguments, or write credentials to disk.

Result and exit codes

• 0: completed gesture count, model_done, or successful diagnostic/demo.
• 2: incomplete run, model stop, limit, cancellation during a run, or operational error returned by the agent. Inspect the final status.
• 1: setup, credential, configuration, or local I/O error.
• 130: interrupt before/after the agent loop.

Argument parsing errors also use exit code 2. JSON mode's final event contains status and counts but no free-form error body. A nonzero exit requires attention; it is never automatically retried.