Dxos Code Style

DXOS Code Style

Authoring conventions for the DXOS monorepo. The always-on subset lives in AGENTS.md; this skill holds the detail. Use @dxos/echo as the reference implementation for the namespace-export pattern.

Casts — fix the type at its source

Do NOT cast to silence a build error; fix the type where it originates.

• "Cast" means as T, as any, as unknown as T, non-null !, or a widened / any signature added to silence a type error. as const is NOT a cast — it narrows a literal rather than bypassing the checker, and is always acceptable (no comment or justification needed).
• Default: fix the type at its source (inference, signature, generic), not the call site that surfaced the error. A red typecheck during a refactor is a finding, not an obstacle to paper over.
• Casts are only acceptable at genuine type-system boundaries (external / untyped data, deliberate coercions), and must carry a concise comment saying why no typed alternative exists.
• Before every commit/PR, audit your diff for new casts: git diff origin/main | grep -nE '\bas (any|unknown|[A-Z])|as unknown as'. Justify or remove each; do not defer to review.
• Casts accumulate fastest during large codemods — treat each as a deliberate decision, never an autopilot stopgap.
• Use @dxos/util trim to create multi-line strings (e.g., prompts).

Unhandled errors — surface, don't suppress

Unhandled errors and rejections are findings; surface them and fix the root cause. Do NOT silence them to make a suite go green.

• Never set dangerouslyIgnoreUnhandledErrors: true in any vitest config (shared vite.base.config.ts or a per-package config). It lets a run exit 0 despite unhandled rejections, hiding real failures — including the teardown races it is usually reached for. It has slipped in before; keep it out.
• A flaky teardown rejection (e.g. vitest worker rpc closing while a console log is pending, browser/birpc close races) is a bug to fix at its source — close the resource, await the pending work, or disable what leaks — not to blanket- ignore. If a specific, understood signature must be tolerated, filter exactly it via onUnhandledError (returning false only for that case) so every other unhandled error stays fatal; never widen that to all errors.
• Same rule for a swallowed unhandledRejection handler, an empty .catch() on real work, and a blanket try/catch that drops the error — handle or rethrow.

Comments — say why, once

A comment earns its place by stating the why the code can't — the constraint, the non-obvious consequence, the reason a reader would otherwise "fix" it wrong. The code already says what it does; a comment that restates that is noise.

• One load-bearing clause. Not a multi-sentence essay. If you're explaining a mechanism, state the constraint in a sentence and stop — resist narrating each step, the alternatives you rejected, or how it used to work. This applies hardest to JSDoc on a new abstraction, where the instinct to over-explain peaks.
• Line-count is itself a signal. If a comment runs past ~2 lines, that length is almost never buying explanatory power proportional to its size — it's restating what a competent reader already infers from the code, the variable names, or the fact that it's test/fixture scaffolding. Compress to the one clause a reader couldn't get any other way, or delete outright. A correct but low-stakes "why" (this is a toolkit stub, this layer is a noop) does not justify three sentences of scene-setting — the reader can see it's a stub.
• Test/fixture code gets a lower comment bar, not a higher one: "this is a minimal/fake X for testing Y" is exactly what the surrounding describe block, filename, and variable names (TestToolkit, *LayerNoop, scripted*) already communicate — restating it in prose adds nothing.
• Never narrate history or the conversation ("previously X, now Y", "as requested", "changed to…"). State the current invariant as if it always was.
• Delete a comment that a competent reader gets from the code itself. Prefer a clearer name or signature over a comment that compensates for an unclear one.
• End with a period. JSDoc public functions.
• Before every commit/PR, audit the comments in your diff: git diff origin/main | grep -nE '^\+\s*(//|\*|/\*)'. Re-read each added line and cut it to its load-bearing clause — or delete it. A verbose comment is the autopilot default; conciseness is the deliberate pass. Do not defer to review. A comment that survives your own audit deserves a second look — try deleting it first, and keep only words that don't come back on their own from re-reading the code.

// ✅ why the code can't be the obvious thing, in one clause:
// Seeded across an identity change so the view refreshes in place, not to empty.

// ❌ restates the code / multi-sentence narration / history:
// Set displayItems to initialItems if it has items, otherwise the empty array.
// We used to reset here but that flashed empty, so now we hold the previous page
// and only replace it once the new query delivers its own results, which means…

// ❌ correct but over-explained test scaffolding — the name/context already says this:
// A minimal echo tool: the deterministic developer code the loop invokes when the
// (scripted) model emits a tool call. Its handler runs for real, so a genuine
// tool-call → result → continue cycle is exercised without any live model.
const TestToolkit = Toolkit.make(Tool.make('Echo', { ... }));

// ✅ same fact, one clause, or just delete it and let the name carry it:
// Real handler, so tool-call → result → continue is a genuine cycle, not a mock.
const TestToolkit = Toolkit.make(Tool.make('Echo', { ... }));

Namespace-export packages

Packages are increasingly organized as namespace exports. Modules have capital-case names and are re-exported as namespaces:

src/
  Foo.ts
  Bar.ts
  errors.ts
  index.ts
  testing/
    index.ts
  internal/
    foo.ts
    bar.ts
    baz.ts

// index.ts
export * as Foo from './Foo';
export * as Bar from './Bar';
export * from './errors';

// Foo.ts

// @import-as-namespace
export const
export const two = 2;
export const func: {
  (a: string): number;
  (a: number): string;
} = (a) => {
  return a;
};

• The @import-as-namespace linter directive marks a file as a namespace export.
• Internal code is hidden in internal/, which is not exported.
• testing/ and errors.ts are the exceptions (exported directly).
• For a namespace file, avoid prefixing top-level types with the namespace name — inside Foo.ts prefer Manager, Service, Options over FooManager, FooService, FooOptions (callers see Foo.Manager either way).

Internal module imports

For @dxos/echo-style entrypoints importing src/internal/<Module>/: import the capitalized internal barrel as a lowercase *Internal namespace — import * as objInternal from './internal/Obj', import * as queryInternal from './internal/Query'. Do not deep-import submodules (./internal/Obj/atoms, ./internal/Ref/ref, etc.); re-export needed symbols from the module's index.ts instead. The top-level ./internal barrel is for cross-cutting re-exports only — prefer the per-module barrel when a single entrypoint owns the dependency. Atom factories inside internal modules use the makeAtom name (not make) to avoid clashing with public make APIs.

Types and signatures

• Common suffix for constructor option-bag types is Options (e.g. SpawnOptions, ManagerImplOptions) — pick this over Opts / Props / Config.
• Keep React component Props types immediately before the component function.
• Take an options object when a constructor or function has more than a few readonly props, especially when several are optional or share a logical group.
• For exported functions with multiple overloads, declare them as const with the overload signatures inline in the type annotation rather than export function with repeated declarations:

  export const myFn: {
    <T extends Foo>(a: T): Bar<T>;
    (a: string): Bar<any>;
  } = (a): Bar<unknown> => { ... };
  

Class member ordering

Consider: static fields → public readonly → public mutable → private readonly (incl. constructor-injected) → private mutable → constructor → public methods → private methods. Within each group, rank properties roughly most-important to least — "further up the stack" (closer to public API), required over optional, readonly over mutable.

Testing

• Place tests near modules as module.test.ts. Use vitest with describe / test (not it); prefer test('foo', ({ expect }) => ...).
• Prefer extending existing test suites over creating new ones. Look for a suite that already covers the area before adding a file. A small number of cohesive suites beats many fragmented ones.
• Test at the level that is naturally the public API. Exercise the seam consumers actually use (exported surface, a service/manager's public methods), not private internals. This keeps tests resilient to refactors and documents real usage.
• Prefer a unified TestLayer for all tests rather than one per test. TestLayer(opts?) can be parametrized so tests configure it.
• Place test layer, configuration, and main definitions at the top of the suite; helpers at the bottom.
• Never wrap an official API in a trivial local helper. A one-liner like const makeBody = (text: string) => Obj.make(Body, { text }) renames the API rather than removing duplication: the reader has to jump to the definition to see what is under test, and several of them turn a suite into an ad-hoc DSL. Inline the real call — Obj.make(Body, { text: 'x' }) — so the API being exercised stays visible next to the assertion. A helper earns its place only when it composes several calls or encodes a non-obvious setup sequence.
• Avoid sleep and polling. Use events and TestClock instead.