- package-surface: split the two plugin forms (a function plugin named-exports
the Loader four, a service plugin default-exports its class), keep Host
Service Definition exports and the Client `/client` export discipline, replace
the wrong example that quoted live production code, state that the static
client libraries and `packages/util/*` are closed to new members, and drop
census numbers that rot on the next package.
- client-components: use the real Slot API — the inject face with the reserved
`hooks` compartment, the parent's derived `renderSlot`, `registerFactory`
scope declarations with local components at `renderFactorySlot` — drop the
string seat, pass the injected `useResource` explicitly with a protocol key,
and describe persistence as the spec's framework-run `persist` key.
- conversation-and-session: rewrite the reasoning example on the real
`conversation.chat.reasoning.body` seat and the published `reasoning.content`
Factory, and drop status annotations and future-tense spec-speak.
- remote-and-api: scope the JSON requirement to the wire payload so
`@Remote({ mode: 'stream' })`, lookup arguments, and a trailing `AbortSignal`
stay valid; drop a leaked Chinese parenthetical from the English page.
- general: measure through `benchmarks/` (`test:bench`, `check:ci:bench`,
`test:bench:memory`) instead of an invented command, state the baseline
without presuming a hook, and drop a leaked exception line.
- api-gateway: link the Remote and API rules instead of restating them.
- development.zh and AGENTS: drop an orphan anchor and widen the standing
order's trigger to the directory's scope.
The earlier "retired routing skill" claim was wrong and is dropped: that skill
was never committed, so the diff deletes nothing under `.agents/skills/`.
10 KiB
Remote and API rules
English | 中文
Rules for the packages/api layer and the remote methods it exposes. api-gateway.md describes the call model itself. Each section below states the requirement, how a reviewer judges it, what it does not cover, which exception applies and who approves it, and a correct and a wrong example.
A new remote capability gets its own service
A new remote business capability is a new service — its own package or its own object — and @Remote marks that service's concrete methods. Growing an existing large service one method at a time leaves one object answering for auth, Session, workspace, and event carriage at once, mixes per-domain activation policy with history pagination, and makes every later change to one capability touch all of them. Split by business boundary, not by caller count: a capability with a single caller today still gets its own service, and an endpoint that changes owners moves to its new owner instead of joining the object it is leaving.
- Judged by: which object holds the new remote method. A method for a new business line added to a service that already owns another line is the defect; the fix is a service named for the new capability, with
@Remoteon its concrete methods. A migration is judged the same way — the moved method appears on its new owner, asSubagentRuntimedoes inpackages/subagent/subagent, not on the object it left. - Not covered: a service whose methods all serve one business boundary stays whole, and a new private helper of that capability belongs on it. Data that needs its own protocol does not become a Remote method on the new service either; session event streams, pagination, incremental reduce, projection, and entity substreams keep their separate registration model (gateway boundaries).
- Exception: none recorded; a single caller is not a reason to merge. Where the business boundary itself is contested, the domain owner settles it as an architecture question before the change lands.
- Fix: create the capability's own service object or package, put
@Remoteon its concrete methods, and let callers reach it throughctx.remote.<namespace>.
The subagent capability owns its own service, and every remote method is a concrete method of that service:
import { Remote, TypertRemoteService } from '@deepseek-ai/dsh-typert-protocol'
interface SubagentList { ids: string[] }
interface PromptRequest { text: string }
interface PromptResult { accepted: boolean }
class SubagentRuntime extends TypertRemoteService {
@Remote list(): SubagentList { return { ids: [] } }
@Remote prompt(request: PromptRequest): PromptResult {
return { accepted: request.text !== '' }
}
@Remote interruptByParent(id: string): void { }
}
The wrong version adds the same capability to the object that already carries other domains:
class ApiProxy {
@Remote goalList(): GoalList // goal domain
@Remote subagentPrompt(request): ... // subagent domain
}
One object then answers for both domains, so a change to either capability's activation policy, pagination, or event handling lands in the same place, and neither capability can move without the other.
@Remote marks capability methods
@Remote carries four disciplines that a reviewer checks separately: it marks the concrete method that performs a capability; client-facing types come from the existing definitions through the generator instead of a hand-written copy; the payload is JSON and the remote method performs no output-side decode; and existing method and field names are additive. Each discipline removes a different failure — a wrapper hides which operation is exposed, a second hand-written type set drifts from the generated one, an output-side decode adds a second result path beside the codec the Gateway already runs, and a rename breaks every caller in the same change.
- Capability methods. The annotation goes on the concrete method that performs the operation (
list,prompt,interruptByParent), never on a wrapper that forwards to the real implementation. Judged by the decorated method's body: a body that only delegates means the annotation is on the wrong method. - Types from inference. Client-facing types are aliased from what the generator produces —
type SessionRemote = TypertClientRemote['session']— instead of being written a second time. Judged by a hand-written declaration whose fields repeat a generated Client type; typert.md owns the generated maps and their declaration merges. - JSON on the wire, no output-side decode. The wire payload of a Remote method is JSON, with binary content expressed field-level. The Host signature keeps what the Gateway already supports — lookup arguments resolved to Host objects, a trailing
AbortSignal, a stream declared with@Remote({ mode: 'stream' })— and@Remoteperforms no decode on the way out; the generated codec and the Gateway validate the result. Judged by a decode, a stream reassembly, or a whole non-JSON value in the payload of a method carrying the decorator; programming model owns what may cross a Remote method. - Existing names are additive. A migration does not rename a remote method or field that a Client already calls; a new semantic gets a new entry, and a rename is its own change with its own review. Judged by a diff that renames or removes a name already exposed through
ctx.remote.<namespace>. - Not covered. A method without the decorator stays an ordinary service method with in-process callers; the four disciplines govern what becomes a Remote method, not what a service may implement.
- Exception. A rename or a new decode path states its benefit and the impact on existing callers first and needs the domain owner's confirmation; the author never self-approves.
The generated map is the only source for Client-facing types, the decorator sits on the concrete method, and the payload stays JSON:
import { Remote, TypertRemoteService, type TypertClientRemote } from '@deepseek-ai/dsh-typert-protocol'
// Client-facing types come from the generated map, never from a second hand-written set.
type SessionRemote = TypertClientRemote['session']
class SessionController extends TypertRemoteService {
@Remote('search')
search(query: string): SessionSearchValue {
return this.runSearch(query)
}
}
A hand-written second type set, a rename of a shipped name, and an output-side decode each break a different discipline:
type SessionRemote = { list(): ...; rename(id: string): ... } // hand-written copy
@Remote('renameSession') rename(...) // shipped name changed
@Remote decode(payload: Uint8Array): Result // output-side decode
The hand-written copy drifts from the generated type without failing a build, the rename breaks every existing Client call at once, and the decode replaces the generated codec with a second, unchecked result path.
packages/api carries generic orchestration only
The packages under packages/api — the controllers, gateway, and remotes — carry transport, call shape, Session lifecycle, and other generic orchestration. A business branch never aggregates there: a decision that depends on concrete business state, handling bound to one feature, or a per-feature filter belongs in its own package or module and reaches this layer as projected data or through a registration surface. The criterion is where the logic belongs, not where the file sits, so a branch that only makes sense for one feature is a defect even when the file holding it is generic.
- Judged by: where the branch reads its condition. An
iforswitchover feature state, a per-feature filter, or a helper named for one feature insidepackages/apiis the defect; the same decision taken in the owning package and consumed here as projected data or a registered entry is not. The package census shows the impact:packages/apiholds nine packages —account-controller,gateway,job-controller,remotes,session-controller,settings-controller,terminal-controller,workspace-controller,workspace-files— andsession-controlleris the largest at 34srcfiles (terminal-controller 12, gateway 11, workspace-controller 10, workspace-files 8, job-controller 8, remotes 4, settings-controller 3, account-controller 2). - Not covered: generic Session lifecycle, transport, and storage abstraction stay —
gatewaydispatch,remotesselection, and shared session-list and catalog reads are infrastructure, not business logic. Business ownership states the same split from the Gateway side. - Exception: none by default. Adding a business branch here needs an independent-package review, and that review answers why the logic cannot live in the package that owns the feature.
- Fix: move the decision into the owning package, publish its result as projected data or register it on a registry this layer already injects, and leave only the orchestration that reads it.
The owning package decides and registers its unit, and the controller only reads the projected value:
// packages/workspace/<feature>/src/index.ts: the rule lives with the feature it serves.
ctx.sessionProjections.register({
key: 'durablePage',
stateSchema: durablePageState,
init: () => ({ durable: false }),
apply: applyDurablePage,
stateVersion: 1,
})
// packages/api/session-controller: generic orchestration reads the projected value.
const { durable } = ctx.sessionProjections.stateOf(session, 'durablePage')
The wrong version decides the rule inside the shared layer:
// packages/api/session-controller: a feature branch inside the shared layer
if (session.page.transport === 'durable') { /* durable-transport handling */ }
The rule then has a second home, and every later change to it has to find both while the shared layer keeps carrying one feature's semantics.