Shadow workspace and LSP

Cursor-class IDEs often validate AI edits in a shadow editor: a second extension host, duplicate language servers, and heavy incremental parsers. That pattern correlates with UI lag. The harness refuses it.

Target loop

  1. Model proposes a patch (byte/character offsets).

  2. The harness applies it to an in-memory copy-on-write layer over the real workspace.

  3. An LSP multiplexer sends textDocument/didOpen / didChange with synthetic URIs to the one language server already serving the project.

  4. Diagnostics feed a self-correction prompt when needed.

  5. Disk commit happens only on user confirmation.

No second rust-analyzer / tsserver process per speculative edit.

Forbidden live engines

  • Tree-sitter (and similar) as the live editor/DOM incremental engine — reflow and update tax.

  • Duplicate extension hosts or full IDE clones for validation.

  • Python/JIT stacks on the speculative hot path.

AST awareness comes from LSP diagnostics and offset patches, not a second parser driving the UI tree.

LSP policy

Prefer language servers that stay cheap under virtual documents. When a language’s LSP is chronically expensive under muxed virtual docs, plan a D reimplementation (serve-d quality bar for D first; others as measured).

GC

The harness defaults to tgc (thread-local GC) for actor swarms — collections do not stop sibling threads. Use @nogc on hot paths: patch apply, IPC framing, LSP JSON-RPC framing, Mixr feature extract. Do not chase whole-program @nogc.

Status

This page is the design contract. Harness v0.5 ships Mixr + provider HTTP seams; CoW buffers and the LSP mux are roadmap (see project-plans seeds).

Related