From the first proof of concept to today — newest first. Built in a nine-day sprint; versions were assigned to capability milestones afterwards.
npm run setup no longer fails on newer npm versions. The product itself is unchanged.tv.inputdevice privilege in the app’s config.xml when it is missing — without it Tizen never hands the remote’s number keys to the page, so the code could not wake the module in a test app that asks for no privileges. Library apps already have it; a build injected earlier gets it on the next inject. The “module did not call home” message now names this cause.launch_app could already do. Leave it empty and the launch is the plain one, as before.running, then the verdict.launch step and the MCP launch_app tool take an optional payload. The box sends it to the TV over the remote channel, and the app reads it from its app control (PAYLOAD) — what it means is the app’s own convention.list_sets takes the same optional project as the other project-aware tools.activate_debug / deactivate_debug.install_package as debugCode.send_keys tool now waits for each key to be confirmed. A key that cannot reach the television — off, or in standby — is reported as an error with the reason, and the keys after it are not pressed.create_upload needs the file's SHA-256, so an upload link that shows up in a chat is useless to anyone else; every eval_js is in the audit log with the code it ran.start_recording / stop_recording tools record work done by hand. A recording ends on stop, after a few minutes without control, or after two hours at the latest.delete_package: removes a build from the package library, so an agent iterating on your app no longer leaves one behind per attempt.create_upload: a one-time link the agent uploads a .wgt to with curl, then installs with install_package — no more squeezing the file into a tool parameter.install_package can also download a build from a public https link, such as a CI artefact. Private and local addresses are refused.launch: false no longer starts the app anyway.install_package, list_packages, get_console, eval_js and capture_screen — the Install and DevTools tabs, for an agent.--fix applies the same remedy.image: tfgsdfg — passes both validators, and rightly: it is a valid name, and the format cannot know what the store holds. The strip said Valid over it, and the run failed hours later with “reference missing”.image: on purpose, the relay refuses that document, and the status strip read Valid · 3 actions over it.similarity, tolerance, offsetTolerance, js length, a camera check inside waitUntil, check.timeout/screenshot/soft, launch with a second field, a carriage return in a note, an unknown repeat field, and the 2 000-action budget.check · image: and nothing more — the reference id is empty until one is captured, so two different conditions rendered identically.image: — a blank nothing you can type will fill, because a reference is a frame taken off the television.PkG1234567.Name from memory, and the editor offered a guess that was almost never the app you meant.screenshot: true step takes a camera frame of the screen and keeps it with the run; the step’s note is its caption. Inside a loop every pass takes a fresh one, up to twenty per run.press now takes times and every, so “press LEFT ten times, 800 ms apart” is a single step. Loops stay for everything with more than one step in the body.repeat with until and max. Only what you put inside is ever done again.