game-controller

Controller mode — a gamepad drives the dashboard through the action registry, and voice is its text input. PENDING.

Status: pending — this node is the contract; no code exists yet. When it lands, the code: list names the dashboard's pad module and the backend's ASR endpoint.

what it is

Controller mode: sit back with a game controller and run the whole manager loop — walk the board, open reviews, drive sessions — without touching a keyboard. A pad has no keys, so the mode is two halves that only together make a complete input surface: the pad for navigation and verbs, and voice for every place the dashboard wants text (a session prompt, search, a new-session instruction). Dictation is not a companion feature; it is the mode's text input method.

the contract

• In the browser, on the action registry. The dashboard reads the pad with the Gamepad API (standard mapping — the browser has already normalized the common pads) and resolves buttons and sticks to the same stable action ids keyboard-nav's registry names. Not synthetic keystrokes: nothing is synthesized, so nothing is untrusted — the pad is simply the registry's second dispatcher, beside the keyboard handler. One registry means pad and keys cannot drift, and a user's key rebind changes which key fires an action, never the action the pad is bound to.
• Zero install, every screen. Because the mode lives in the page, it works wherever the dashboard is open — any machine, a TV or tablet with a Bluetooth pad, the public gateway — nothing installed per-device, no OS permissions.
• Voice: the browser captures, the backend recognizes. A reserved control is push-to-talk: the page records the mic (a secure-context surface) and posts audio to the backend, whose pluggable ASR adapter turns it to text — a local engine or a cloud key, the operator's choice — landing in whatever input has focus. The browser's own recognition (Web Speech) is not the path: it is vendor-bound and unreliable where we run. Owning the voice path means owning the adapter, not the recognizer.
• The backend knows the board. Recognition is biased with the project's own vocabulary — node ids, session names, spec titles — so mixed-language dev speech resolves to the right identifiers. A generic recognizer can never do this, and it is the seam a hosted offering can charge for: self-hosters bring an ASR key; hosted supplies one, ready out of the box.

why not an OS-level remapper

An earlier shape of this node was a standalone OS remapper in its own repo (an extensions satellite) emitting real keystrokes, because a page could neither fake trusted keys nor reach the OS voice hotkey. Both reasons fell: dispatching action ids needs no key events at all, and voice stops needing any OS facility once the backend owns ASR. What survives from that survey: no third-party mapper (AntiMicroX / JoyToKey / Steam Input / ClaudeGamepad) is a dependency — the mapping stays our artifact, in the dashboard, beside the registry it reads.