Canopy in May 2026 — Highlights
A two-minute summary of a month of work. For the day-by-day version, see the Japanese original — the source for this one reads more like a raw work journal than a polished log, and that carries over here a little.
Canopy is a structure editor written in MoonBit: it treats source code as structure (IR) rather than strings, with text kept as the source of truth. Concurrent editing is built on a CRDT implementing the eg-walker paper. Around it sit a few sibling projects: Loom (incremental parsing), incr (incremental computation), moondsp (a MoonBit DSP/music DSL experiment), and js_engine (a JS interpreter validated against test262).
1. Rabbita and CodeMirror finally get along
The glue code connecting Rabbita (Canopy’s UI layer) to CodeMirror had been fragile enough that the editor could just stop responding mid-session. That class of bug got fixed early in the month, which mattered more than any single feature — nothing else is worth building on top of an editor that randomly locks up.
2. Hidden buttons become explicit boundaries
Interactions that had been routed through invisible DOM controls moved to explicit event subscriptions and boundaries instead — a small change that made the system’s behavior something you could reason about rather than something you had to reverse-engineer.
3. Inspector and op-log tooling
Internal editor state became something you could actually watch happen: an Inspector panel and an operation log turned “what is the editor doing right now” from a debugging guess into something visible on screen.
4. incr’s API gets rethought
Working through Build Systems à la Carte, the incr library’s evaluation model and public API got a deliberate design pass — groundwork that quietly pays off two months later, in July’s 0.13.0 and 0.14.0 API-boundary cleanups.
5. Cognition: the first AI-context scaffolding
The first pieces of “Cognition” appeared: a workspace concept, context packing, and a provider boundary that tracks what goes to an AI provider, what comes back, and how cancellation/retry are handled. The goal — letting the editor’s own state become context an AI can use — doesn’t pay off yet this month, but the scaffolding is where it starts.
6. Lambda gets real navigation
A scope graph and go-to-definition landed for Lambda, with Ideal’s own scope annotations converging on the same resolution results — the first sign of the name-resolution consolidation that would fully land in July. Shared infrastructure work (an ephemeral-presence store, a byte codec) and the first performance/demo surfaces (incr’s typed-spreadsheet demo, js_engine’s bytecode benchmark) rounded out the month.
May is the quiet month where the AI-context (Cognition) and performance-measurement threads both start — neither pays off yet, but both become central later.
- GitHub: dowdiness/canopy · dowdiness/incr · mooncakes.io/user/dowdiness
- Full log: May 2026 devlog (Japanese)