17 KiB
Plugin SDK Namespaces Plan
TL;DR
OpenClaw should introduce a few clear SDK namespaces like plugin, channel,
and provider, instead of keeping so much of the public surface flat.
The safe way to do that is:
- add thin ESM facade entrypoints, not TypeScript
namespace - keep the root
openclaw/plugin-sdksurface small - replace flat registration methods on
OpenClawPluginApiwith namespace groups - ship the cutover in one coordinated release instead of dragging old flat APIs along
- forbid leaf modules from importing back through namespace facades
That gives plugin authors a cleaner SDK that feels closer to VS Code, without turning the SDK into a giant barrel or creating circular import problems.
Goal
Introduce public namespaces to the OpenClaw Plugin SDK so the surface feels closer to the VS Code extension API, while keeping the implementation tight, isolated, and resistant to circular imports.
This plan is about the public SDK shape. It is not a proposal to merge everything into one giant barrel.
Why This Is Worth Doing
Today the Plugin SDK has three visible problems:
- The public package export surface is large and mostly flat.
src/plugin-sdk/core.tsandsrc/plugin-sdk/index.tscarry too many unrelated meanings.OpenClawPluginApiis still a flat registration API even thoughapi.runtimealready proves grouped namespaces work well.
The result is harder docs, harder discovery, and too many helper names that look equally important even when they are not.
Current Facts In The Repo
- Package exports are generated from a flat entrypoint list in
src/plugin-sdk/entrypoints.tsandscripts/lib/plugin-sdk-entrypoints.json. - The root
openclaw/plugin-sdkentry is intentionally tiny insrc/plugin-sdk/index.ts. api.runtimeis already a successful namespace model. It groups behavior asagent,subagent,media,imageGeneration,webSearch,tools,channel,events,logging,state,tts,mediaUnderstanding, andmodelAuthinsrc/plugins/runtime/index.ts.- The main plugin registration API is still flat in
OpenClawPluginApiinsrc/plugins/types.ts. - The concrete API object is assembled in
src/plugins/registry.ts, and a second partial copy exists insrc/plugins/captured-registration.ts.
Those facts suggest a path that is low-risk:
- keep leaf modules as the source of truth
- add namespace facades on top
- cut docs, examples, and templates over in the same release as the namespace model
Design Principles
1. Do Not Use TypeScript namespace
Use normal ESM modules and package exports.
The SDK already ships as package export subpaths. The namespace model should be
implemented as public facade modules, not TypeScript namespace syntax.
2. Keep The Root Tiny
Do not turn openclaw/plugin-sdk into a giant VS Code-style monolith.
The closest safe equivalent is:
- a tiny root for shared types and a few universal values
- a small number of explicit namespace entrypoints
- optional ergonomic aggregation only after the namespace surfaces settle
3. Namespace Facades Must Be Thin
Namespace entrypoints should contain no real business logic.
They should only:
- re-export stable leaves
- assemble small namespace objects
That keeps cycles and accidental coupling down.
4. Types Stay Direct And Easy To Import
Like VS Code, namespaces should mostly group behavior. Common types should stay directly importable from the root or the owning domain surface.
Examples:
ChannelPluginProviderPluginOpenClawPluginApiPluginRuntime
5. Do Not Namespace Everything At Once
Only namespace areas that already have a clear public identity.
Phase 1 should focus on:
pluginchannelprovider
runtime already has a good public namespace shape on api.runtime and should
not be reopened as a giant package-export family in the first pass.
Proposed Public Model
Namespace Entry Points
Canonical public entrypoints:
openclaw/plugin-sdk/pluginopenclaw/plugin-sdk/channelopenclaw/plugin-sdk/provideropenclaw/plugin-sdk/runtimeopenclaw/plugin-sdk/testing
What each should mean:
plugin- plugin entry helpers
- shared plugin definition helpers
- plugin-facing config schema helpers that are truly universal
channel- channel entry helpers
- chat-channel builders
- stable channel-facing contracts and helpers
provider- provider entry helpers
- auth, catalog, models, onboard, stream, usage, and provider registration helpers
runtime- the existing
api.runtimestory and runtime-related public helpers that are truly stable
- the existing
testing- plugin author testing helpers
Nested Leaves
Under those namespaces, the long-term canonical leaves should become nested:
-
openclaw/plugin-sdk/channel/setup -
openclaw/plugin-sdk/channel/pairing -
openclaw/plugin-sdk/channel/reply-pipeline -
openclaw/plugin-sdk/channel/contract -
openclaw/plugin-sdk/channel/targets -
openclaw/plugin-sdk/channel/actions -
openclaw/plugin-sdk/channel/inbound -
openclaw/plugin-sdk/channel/lifecycle -
openclaw/plugin-sdk/channel/policy -
openclaw/plugin-sdk/channel/feedback -
openclaw/plugin-sdk/channel/config-schema -
openclaw/plugin-sdk/channel/config-helpers -
openclaw/plugin-sdk/provider/auth -
openclaw/plugin-sdk/provider/catalog -
openclaw/plugin-sdk/provider/models -
openclaw/plugin-sdk/provider/onboard -
openclaw/plugin-sdk/provider/stream -
openclaw/plugin-sdk/provider/usage -
openclaw/plugin-sdk/provider/web-search
Not every current flat subpath needs a namespaced replacement. The goal is to promote the stable public domains, not to preserve every current export forever.
What Happens To core
core is overloaded today. In a namespace model it should shrink, not grow.
Target split:
- plugin-wide entry helpers move toward
plugin - channel builders and channel-oriented shared helpers move toward
channel corestops being a first-class public destination and shrinks to the smallest possible remaining shared surface
Rule: no new public API should be added to core once namespace entrypoints
exist.
Proposed OpenClawPluginApi Shape
Keep context fields flat:
idnameversiondescriptionsourcerootDirregistrationModeconfigpluginConfigruntimeloggerresolvePath
Move registration behavior behind namespaces:
| Current flat method | Proposed namespace location |
|---|---|
registerTool |
api.tool.register |
registerHook |
api.hook.register |
on |
api.hook.on |
registerHttpRoute |
api.http.registerRoute |
registerChannel |
api.channel.register |
registerProvider |
api.provider.register |
registerSpeechProvider |
api.provider.registerSpeech |
registerMediaUnderstandingProvider |
api.provider.registerMediaUnderstanding |
registerImageGenerationProvider |
api.provider.registerImageGeneration |
registerWebSearchProvider |
api.provider.registerWebSearch |
registerGatewayMethod |
api.gateway.registerMethod |
registerCli |
api.cli.register |
registerService |
api.service.register |
registerInteractiveHandler |
api.interactive.register |
registerCommand |
api.command.register |
registerContextEngine |
api.contextEngine.register |
registerMemoryPromptSection |
api.memory.registerPromptSection |
The cutover should replace the flat methods in one coordinated change.
That gives plugin authors a clearer public shape and avoids carrying two public registration models at the same time.
Example Public Usage
Proposed style:
import { definePluginEntry } from "openclaw/plugin-sdk/plugin";
import { channel } from "openclaw/plugin-sdk/channel";
import { provider } from "openclaw/plugin-sdk/provider";
import type { ChannelPlugin, OpenClawPluginApi } from "openclaw/plugin-sdk";
const chatPlugin: ChannelPlugin = channel.createChatPlugin({
id: "demo",
/* ... */
});
export default definePluginEntry({
id: "demo",
register(api: OpenClawPluginApi) {
api.channel.register(chatPlugin);
api.command.register({
name: "status",
description: "Show plugin status",
run: async () => ({ text: "ok" }),
});
},
});
This is close to the VS Code mental model:
- grouped behavior
- direct types
- obvious public areas
without requiring a single monolithic root import.
Optional Ergonomic Surface
If the project later wants the closest possible VS Code feel, add a dedicated
opt-in facade such as openclaw/plugin-sdk/sdk.
That facade can assemble:
pluginchannelproviderruntimetesting
It should not be phase 1.
Why:
- it is the highest-risk barrel from a cycle and weight perspective
- it is easier to add once the namespace surfaces already exist
- it preserves the root
openclaw/plugin-sdkentry as a small type-oriented surface
Internal Implementation Rules
These rules are the important part. Without them, namespaces will rot into barrels and cycles.
Rule 1: Namespace Facades Are One-Way
Namespace entrypoints may import leaf modules.
Leaf modules may not import their namespace entrypoint.
Examples:
- allowed:
src/plugin-sdk/channel.tsimporting./channel-setup.ts - forbidden:
src/plugin-sdk/channel-setup.tsimporting./channel.ts
Rule 1A: Allowed Dependency Directions Must Be Explicit
The allowed directions should be:
- namespace facade -> leaves in the same namespace
- leaf -> local implementation helpers
- leaf -> dedicated shared internal leaf
- leaf -> another leaf in the same namespace only by direct relative import, never through the namespace facade
The forbidden directions should be:
- leaf -> its own namespace facade
- leaf -> another namespace facade
- namespace facade -> another namespace facade
- channel leaf -> provider leaf, or provider leaf -> channel leaf, unless the dependency is first extracted into a shared internal leaf
Short version:
- facades point downward
- leaves never point back upward
- cross-namespace sharing must go sideways through a shared internal leaf, not directly through another public namespace
Rule 1B: If Two Namespaces Need Each Other, Extract A Shared Leaf
If channel and provider start needing each other directly, that is the sign
that the seam is wrong.
Do not allow:
src/plugin-sdk/channel/*importing fromsrc/plugin-sdk/provider/*src/plugin-sdk/provider/*importing fromsrc/plugin-sdk/channel/*
Instead:
- extract the shared logic into a dedicated internal leaf
- let both sides depend on that leaf
- keep the public namespaces separate
This is the main cycle-prevention rule. Shared logic moves to a lower layer before it creates a back-edge.
Rule 2: No Public-Specifier Self-Imports Inside The SDK
Files inside src/plugin-sdk/** should never import from
openclaw/plugin-sdk/....
They should import local source files directly.
Rule 3: Shared Code Lives In Shared Leaves
If channel and provider need the same implementation detail, move that code
to a shared leaf instead of importing one namespace from the other.
Good shared homes:
- a dedicated internal shared leaf
- a very small shared core leaf only if it has a precise, stable reason to exist
- existing domain-neutral helpers
Bad pattern:
provider/*importing fromchannel/indexchannel/*importing fromprovider/index
Rule 4: Assemble The API Surface Once
OpenClawPluginApi should be built by one canonical factory.
src/plugins/registry.ts and src/plugins/captured-registration.ts should stop
hand-building separate versions of the API object.
That factory can expose:
- the namespaced shape only
from the same underlying implementation.
Rule 5: Namespace Entry Files Stay Small
Namespace facades should stay close to pure exports. If a namespace file grows real orchestration logic, split that logic back into leaf modules.
Dependency Shape
The intended import graph is:
public facade
-> same-namespace leaves
-> local helpers
-> shared internal leaves
Not this:
channel facade -> provider facade
channel leaf -> channel facade
provider leaf -> channel leaf
Concrete examples:
- allowed:
src/plugin-sdk/channel.ts->./channel/setup.ts - allowed:
src/plugin-sdk/channel/setup.ts->./_internal/channel-shared.ts - allowed:
src/plugin-sdk/provider/auth.ts->../_internal/provider-shared.ts - forbidden:
src/plugin-sdk/channel/setup.ts->./channel.ts - forbidden:
src/plugin-sdk/channel/setup.ts->../provider/index.ts - forbidden:
src/plugin-sdk/channel.ts->./provider.ts
Migration Strategy
This should be a cutover, not a long overlap period.
That means:
- one coordinated release
- one migration guide
- one docs/templates/test update
- one public SDK shape after the release
Phase 1: Extract The Canonical API Builder
Do this first, before changing the public surface.
Why:
- it removes duplicated API assembly
- it gives one place to switch the public shape
- it reduces cutover risk
Implementation:
- extract one canonical API builder from
src/plugins/registry.tsandsrc/plugins/captured-registration.ts - make that builder assemble the new namespaced registration API
Phase 2: Add Canonical Namespace Entrypoints
Add:
pluginchannelprovider
as thin public facades over existing flat leaves.
Implementation detail:
- the first pass can re-export current flat files
- do not move source layout and package exports in the same commit if it can be avoided
Examples:
src/plugin-sdk/channel/setup.tscan initially re-export from../channel-setup.jssrc/plugin-sdk/provider/auth.tscan initially re-export from../provider-auth.js
This lets the public namespace story land before the internal source move, without forcing all implementation files to move in the same commit.
Phase 3: Cut Public API, Docs, And Templates Together
In the same release:
- docs prefer namespaced entrypoints
- templates prefer namespaced imports
- tests and examples switch to the namespaced shape
OpenClawPluginApichanges to the namespaced registration model- flat registration methods are removed instead of carried as aliases
Phase 4: Remove The Old Public Story
After the cutover release lands:
- stop documenting superseded flat leaves as public API
- keep only the namespace model in author-facing docs
- remove any leftover flat registration surface that survived only as transitional scaffolding during implementation
What Should Not Be Namespaced In Phase 1
To keep the refactor tight, do not force these into the first milestone:
- every
*-runtimehelper subpath - extension-branded public subpaths
- one-off utilities that do not yet have a stable domain home
- the root
openclaw/plugin-sdkbarrel
If a subpath is only public because history leaked it, namespace work should not promote it.
Guardrails And Validation
The namespace rollout should ship with explicit checks.
Existing Checks To Reuse
src/plugin-sdk/subpaths.test.tssrc/plugin-sdk/runtime-api-guardrails.test.tspnpm buildfor[CIRCULAR_REEXPORT]warningspnpm plugin-sdk:api:check
New Checks To Add
- namespace facade files may only re-export or compose approved leaves
- leaf files under a namespace may not import their parent
indexfacade - leaf files under one namespace may not import another namespace facade
- cross-namespace leaf imports should fail unless the target is an approved shared internal leaf
- namespace facades may not import other namespace facades
- no new API should be added to
coreonce namespace facades exist OpenClawPluginApimust not expose both flat and namespaced registration methods after cutover
Recommended End State
The elegant end state is:
- a tiny root
- a few first-class namespaces
- direct types
- a grouped
apiregistration surface - stable leaves under each namespace
- no reverse imports from leaves back into namespace facades
That gives OpenClaw a VS Code-like feel where the public SDK has clear domains, but still respects the repo's existing build, lazy-loading, and package-boundary constraints.
Short Recommendation
If this work starts soon, the first implementation step should be:
- extract one canonical
OpenClawPluginApibuilder - switch that builder to the namespaced registration shape
- add
plugin,channel, andproviderfacade entrypoints - cut docs, templates, and examples over in the same release
- remove the old flat registration story instead of maintaining dual public APIs
That sequence keeps the refactor elegant and minimizes the chance that namespaces become another layer of accidental coupling.