Omnis
/

ADR-0017 — The TUI is dropped as a surface

  • Status: Accepted · Date: 2026-09-06
  • Supersedes: the two-backend model in ADR-0001
  • Closes: D15

Context

ADR-0001 established two renderer backends over one scene tree: the GPU compositor, and omnis-tui rendering the same scene tree as ANSI escape sequences in a real terminal. The second backend brought a parity contract (§9.4), a degradation ladder per primitive, terminal capability detection, kitty and SGR input decoding, a terminal capability floor to decide (D15), and an epic to build it (E22).

Two things changed since. Remote access is now a web surface over Tailscale (ADR-0015), which covers the case the TUI was mostly wanted for — reaching the workspace from elsewhere. And the compositor targets WebGPU, so the browser already hosts the real interface rather than a reduced one.

That leaves the TUI carrying a large, permanent tax for a shrinking benefit.

Decision

There is one renderer backend: the GPU compositor. It targets a desktop window natively and a browser canvas over WebGPU. omnis-tui is not built.

What this does not remove:

  • cell-grid presentation mode stays. The terminal aesthetic — fixed advance, ANSI palettes, pane-grid keyboard navigation, command-palette-first interaction — is a layout mode the GPU compositor draws, and was always separate from running in a terminal. The Zed-style look survives intact.
  • The CLI stays a surface. omnis with --json on every command is unaffected; it is not a renderer.
  • The scene tree stays. It remains the single input to the renderer and the abstraction feature code targets. One backend does not make it optional — it is what keeps sources from reaching into rendering.

What this removes: the omnis-tui crate, the parity contract, per-primitive terminal fallbacks (P7), terminal capability detection, D15, and E22.

Alternatives rejected

  • Keep both backends (ADR-0001 as written). Honest, and expensive: every primitive needs a declared terminal fallback forever, the material layer can never render, and every new source pays a degradation tax. Justified when the TUI was the remote story; not justified once the web surface is.
  • Keep a reduced TUI — a status view rather than a full surface. Sounds cheap and is not: it still needs a renderer, capability detection, and input decoding, and it sets an expectation of parity it does not meet. A CLI with good output serves that need without pretending to be a UI.
  • Drop the scene tree too, since there is one backend. That would let sources reach into rendering, which is the coupling ADR-0001 existed to prevent. The abstraction earns its place with one backend.

Consequences

  • A machine with no display and no browser gets the CLI only. This is the real loss and it should be named rather than discovered: SSH into a headless box and there is no interactive Omnis, only omnis commands. If that becomes intolerable, this ADR is what gets superseded.
  • Accessibility gets more important, not less. The TUI was one path to text-addressable output; with it gone, the compositor's published accessibility tree (UIA/AX/AT-SPI) is the only path. That work moves from important to non-negotiable.
  • The roadmap loses E22 and shrinks E3's acceptance criteria. The parity clauses come out of the architecture.
  • ADR-0001's core claim — sources are not renderers, and everything meets in one scene tree — is unaffected. Only its second backend is.
Omnis