The IDE distribution

An Eclipse Theia distribution with the runtime as its engine, three extensions, and native installers for Linux and Windows. It shipped, it ran, and it was withdrawn. This page records it as research rather than as a product.

Why it was built

An extension is a participant in someone else’s product: its chat, its session list, its approval widgets, its model-selection policy. The argument that carried the work forward was that a harness which could not own the surface would always be a guest in it — and that the runtime was strong enough to deserve a workbench of its own.

What it was

• A distribution, not a plugin. Eclipse Theia with the Contenox workbench, ACP bridge and completion extensions compiled in, packaged as a desktop application for Linux and Windows.
• The runtime as the engine. The same contenox binary the terminal talks to, launched by the application, serving sessions, approvals and model routing over the same protocol.
• One acceptance standard. A capability counted as done only when a clean installed copy of the application completed it without a terminal, a checkout, or maintainer-owned machine state.

What it proved

• The engine can carry a full workbench: sessions, tool cards, approvals, checkpoints and model selection all work through the ACP seam.
• The protocol is enough for a first-party surface. No private editor API was needed for the agent itself.
• A desktop distribution is a company-sized commitment: a framework to track, a release train per platform, signing, and packaging for every dependency.

Why it was withdrawn

The distribution became the definition of done. Once “a clean install of the IDE completes the capability” was the acceptance test, every other surface lost its claim to being finished: the CLI drifted, the terminal client was removed, and the runtime was judged by whether a screen rendered rather than whether an operator could run work.

The runtime was the product. The distribution was a surface — an expensive one, in a framework the project did not own, competing for the same weeks that would have kept the harness operable.

What survived

• ACP. Editors that speak the protocol drive the agent without a fork or a vendor extension API. The distribution was replaceable; the seam was not.
• The acceptance discipline. “A clean install, a real operation” is the right test. The mistake was attaching it to one surface.
• The rule that a surface must not own policy. The gate belongs to the engine; every client, including this one, renders it.

The distribution is archived, not deleted: it is the strongest evidence the project has that the engine can sit under a full application, and the clearest example of what it costs to make a surface the product.