deepseek-harness/packages/host/apiproxy/tests/api-proxy-config.spec.ts

688 lines
30 KiB
TypeScript
Raw Normal View History

/**
* Settings/credentials/llm RPC domains and their host-stream frames over
* createApiProxy: layered redacted describe, write-path rejection mapping,
* value-free credential views, the directory/live-route merge, and the three
* invalidation frames (settings/credentials/models changed).
*/
import { describe, expect, it, vi } from 'vitest'
import { Context } from 'cordis'
import z from 'schemastery'
import AgentRegistry from '@deepseek-ai/dsh-agent'
import SessionStore from '@deepseek-ai/dsh-session'
import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
import ToolRegistry from '@deepseek-ai/dsh-tools'
import UserInteractionService from '@deepseek-ai/dsh-user-interaction'
import LlmService, { LlmAdapter } from '@deepseek-ai/dsh-llm'
import type { GenerateOptions, LlmModelInfo, LlmProviderInfo, StreamChunk } from '@deepseek-ai/dsh-llm'
import { Settings, settingsNamespace } from '@deepseek-ai/dsh-settings'
import type { SettingsNamespace } from '@deepseek-ai/dsh-settings'
import { Credentials } from '@deepseek-ai/dsh-credentials'
import type { CredentialInfo, CredentialRef, ResolvedCredential } from '@deepseek-ai/dsh-credentials'
import type { HostFrame } from '../src/api/index.ts'
import type { RpcRequest, RpcResponse } from '../src/api/rpc.ts'
import { RpcId } from '../src/api/rpc.ts'
import { createApiProxy } from '../src/api-proxy.ts'
const DEFAULTS = { provider: 'p', model: 'm', cwd: '/tmp', workspaceRoot: '/tmp' }
let nextRpc = 1
function request<P>(payload: P): RpcRequest<P> {
return { rpcId: RpcId(`req-${String(nextRpc++)}`), payload }
}
function expectOk<T>(response: RpcResponse<T>): T {
expect(response.result.ok).toBe(true)
if (!response.result.ok) throw new Error('unreachable')
return response.result.value
}
function expectErr<T>(response: RpcResponse<T>): { code: string; message: string; details: unknown } {
expect(response.result.ok).toBe(false)
if (response.result.ok) throw new Error('unreachable')
return response.result.error
}
/** In-memory settings provider: the seam base class owns all tested behavior. */
class MemorySettings extends Settings {
doc: Record<string, unknown>
constructor(ctx: ConstructorParameters<typeof Settings>[0], options?: {
doc?: Record<string, unknown>
readOnly?: boolean
documentPath?: string
preparedPath?: string
}) {
super(ctx)
this.doc = structuredClone(options?.doc ?? {})
this.readOnly = options?.readOnly ?? false
this.path = options?.documentPath
this.preparedPath = options?.preparedPath
}
private readonly readOnly: boolean
private readonly path: string | undefined
private readonly preparedPath: string | undefined
get writable(): boolean {
return !this.readOnly
}
override get documentPath(): string | undefined {
return this.path
}
override prepareDocument(): Promise<string | undefined> {
return Promise.resolve(this.preparedPath ?? this.documentPath)
}
protected load(): Promise<Record<string, unknown>> {
return Promise.resolve(structuredClone(this.doc))
}
protected persist(ns: SettingsNamespace, section: Record<string, unknown>): Promise<void> {
this.doc[ns] = structuredClone(section)
return Promise.resolve()
}
}
/** In-memory credential provider with an env-shadow double for the rejection path. */
class MemoryCredentials extends Credentials {
private readonly values = new Map<string, string>()
constructor(ctx: ConstructorParameters<typeof Credentials>[0], options?: { shadowed?: string[] }) {
super(ctx)
this.shadowed = new Set(options?.shadowed ?? [])
}
private readonly shadowed: Set<string>
resolve(ref: CredentialRef): Promise<ResolvedCredential | undefined> {
if (this.shadowed.has(ref)) return Promise.resolve({ value: 'from-env', source: 'env' })
const value = this.values.get(ref)
return Promise.resolve(value === undefined ? undefined : { value, source: 'file' })
}
describe(ref: CredentialRef): Promise<CredentialInfo> {
if (this.shadowed.has(ref)) return Promise.resolve({ configured: true, source: 'env', writable: false })
const configured = this.values.has(ref)
return Promise.resolve({ configured, ...configured ? { source: 'file' } : {}, writable: true })
}
set(ref: CredentialRef, value: string): Promise<void> {
if (this.shadowed.has(ref)) {
return Promise.reject(new Error(`credentials: ${ref} is shadowed by the read-only environment`))
}
this.values.set(ref, value)
this.ctx.emit('credentials/updated', ref)
return Promise.resolve()
}
unset(ref: CredentialRef): Promise<void> {
if (this.shadowed.has(ref)) {
return Promise.reject(new Error(`credentials: ${ref} is shadowed by the read-only environment`))
}
this.values.delete(ref)
this.ctx.emit('credentials/updated', ref)
return Promise.resolve()
}
}
/** Catalog-serving adapter stub for the llm.models path. */
class CatalogAdapter extends LlmAdapter {
constructor(private readonly name: string, private readonly models: readonly string[]) {
super()
}
override providerInfo(provider: string): LlmProviderInfo {
return { id: provider, name: this.name }
}
override listModels(provider: string): Promise<readonly LlmModelInfo[]> {
return Promise.resolve(this.models.map(id => ({ provider, id, name: id })))
}
async * stream(_options: GenerateOptions): AsyncIterable<StreamChunk> {
throw new Error('not exercised')
}
}
class BrokenCatalogAdapter extends CatalogAdapter {
override listModels(): Promise<readonly LlmModelInfo[]> {
return Promise.reject(new Error('catalog backend down'))
}
}
const NS = settingsNamespace('llm-deepseek')
const AdapterConfig = z.object({
apiKey: z.string().role('secret'),
apiKeyEnv: z.string().default('DEEPSEEK_API_KEY'),
baseURL: z.string(),
})
async function harness(options?: {
settings?: false | {
doc?: Record<string, unknown>
readOnly?: boolean
documentPath?: string
preparedPath?: string
}
credentials?: false | { shadowed?: string[] }
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
/** Skip the directory registration to exercise a namespace the proxy does not expose. */
configurableProviders?: false
}): Promise<Context> {
const ctx = new Context()
await ctx.plugin(SessionStore)
await ctx.plugin(SystemPrompt, { persona: '' })
await ctx.plugin(ToolRegistry)
await ctx.plugin(UserInteractionService)
await ctx.plugin(AgentRegistry)
await ctx.plugin(LlmService)
if (options?.settings !== false) await ctx.plugin(MemorySettings, options?.settings)
if (options?.credentials !== false) await ctx.plugin(MemoryCredentials, options?.credentials)
// Model-provider namespaces plus the explicit Web preference and product
// onboarding allowlists are the proxy's complete settings surface.
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
if (options?.configurableProviders !== false) {
ctx.llm.registerConfigurableProviders([
{ provider: 'deepseek-official', displayName: 'DeepSeek', settingsNs: 'llm-deepseek', settingsPath: [] },
])
}
// Host-stream opener reads the committed-workspace baseline; the stub
// suffices — the real workspace composition is api-proxy-workspace.spec's.
ctx.provide('workspace', { list: () => [] } as never)
return ctx
}
/** Drain `count` host frames matching `types`, then abort the stream. */
async function collectHost(
api: ReturnType<typeof createApiProxy>,
types: string[],
count: number,
run: () => Promise<void>,
): Promise<HostFrame[]> {
const abort = new AbortController()
const frames: HostFrame[] = []
const stream = api.events.host(request({}), abort.signal)
const consume = (async () => {
for await (const frame of stream) {
if (!types.includes(frame.payload.type)) continue
frames.push(frame.payload)
if (frames.length >= count) abort.abort()
}
})()
await run()
await consume
return frames
}
describe('settings domain', () => {
it('reports an actionable error when no settings provider is mounted', async () => {
const ctx = await harness({ settings: false })
const api = createApiProxy(ctx, DEFAULTS)
const error = expectErr(await api.settings.describe(request({})))
expect(error.code).toBe('internal')
expect(error.message).toContain('dsh-settings-local')
})
it('describes layered redacted namespaces with their secret slots', async () => {
const ctx = await harness({ settings: {
doc: { 'llm-deepseek': { apiKey: 'user-secret', baseURL: 'https://user' } },
documentPath: '/tmp/custom-settings.yaml',
} })
ctx.settings.register(NS, AdapterConfig, { base: { baseURL: 'https://base' } })
const api = createApiProxy(ctx, DEFAULTS)
const value = expectOk(await api.settings.describe(request({})))
expect(value.writable).toBe(true)
expect(value.hasDocument).toBe(true)
expect(value.namespaces).toHaveLength(1)
const view = value.namespaces[0]!
expect(view.ns).toBe('llm-deepseek')
expect(view.applies).toBe('live')
expect((view.schema as { refs?: unknown }).refs).toBeDefined()
expect(view.value).toEqual({ apiKeyEnv: 'DEEPSEEK_API_KEY', baseURL: 'https://user' })
expect(view.base).toEqual({ baseURL: 'https://base' })
expect(view.user).toEqual({ baseURL: 'https://user' })
expect(view.secrets).toEqual([{ path: ['apiKey'], set: true }])
expect(JSON.stringify(value)).not.toContain('user-secret')
})
it('opens the provider-resolved document without accepting a browser path', async () => {
const ctx = await harness({ settings: {
documentPath: '/tmp/described-settings.yaml',
preparedPath: '/tmp/custom-settings.yaml',
} })
const opened: string[] = []
const api = createApiProxy(ctx, {
...DEFAULTS,
openTextFile: (path) => {
opened.push(path)
return Promise.resolve()
},
})
expect(expectOk(await api.settings.openDocument(request({}), new AbortController().signal)))
.toEqual({ opened: true })
expect(opened).toEqual(['/tmp/custom-settings.yaml'])
})
it('refuses to open settings when the provider has no local document', async () => {
const ctx = await harness()
const api = createApiProxy(ctx, DEFAULTS)
expect(expectOk(await api.settings.describe(request({}))).hasDocument).toBe(false)
const error = expectErr(await api.settings.openDocument(request({}), new AbortController().signal))
expect(error.code).toBe('internal')
expect(error.message).toContain('no local document')
})
it('does not prepare or open a settings document after cancellation', async () => {
const ctx = await harness({ settings: { documentPath: '/tmp/settings.yaml' } })
const opened: string[] = []
const api = createApiProxy(ctx, {
...DEFAULTS,
openTextFile: (path) => {
opened.push(path)
return Promise.resolve()
},
})
const prepare = vi.spyOn(ctx.settings, 'prepareDocument')
const cancelled = new AbortController()
cancelled.abort()
expect(expectErr(await api.settings.openDocument(request({}), cancelled.signal)).code)
.toBe('cancelled')
expect(prepare).not.toHaveBeenCalled()
const pending = Promise.withResolvers<string | undefined>()
prepare.mockReturnValueOnce(pending.promise)
const duringPrepare = new AbortController()
const opening = api.settings.openDocument(request({}), duringPrepare.signal)
await vi.waitFor(() => { expect(prepare).toHaveBeenCalledOnce() })
duringPrepare.abort()
pending.resolve('/tmp/settings.yaml')
expect(expectErr(await opening).code).toBe('cancelled')
expect(opened).toEqual([])
})
it('serves model-provider and explicitly allowlisted Web namespaces only', async () => {
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
// The settings seam is general: any plugin may register a namespace for
// its own configuration. The Web configuration plane remains opt-in, so a
// future internal plugin cannot become remotely configurable just by
// registering; permission and the product onboarding namespace are the
// non-model namespaces intentionally admitted by this surface.
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
const ctx = await harness()
ctx.settings.register(NS, AdapterConfig)
ctx.settings.register(settingsNamespace('some-other-plugin'), z.object({ secretPath: z.string() }))
ctx.settings.register(settingsNamespace('permission'), z.object({
defaultPreset: z.union(['read-only', 'workspace-write']).required(),
}), {
base: { defaultPreset: 'read-only' },
})
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
const api = createApiProxy(ctx, DEFAULTS)
const value = expectOk(await api.settings.describe(request({})))
expect(value.namespaces.map(view => view.ns)).toEqual(['llm-deepseek', 'permission'])
const permission = expectOk(await api.settings.mutate(request({
ns: 'permission',
ops: [{ op: 'set', path: ['defaultPreset'], value: 'workspace-write' }],
})))
expect(permission.value).toEqual({ defaultPreset: 'workspace-write' })
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
for (const response of [
await api.settings.update(request({ ns: 'some-other-plugin', patch: { secretPath: '/etc/shadow' } })),
await api.settings.replace(request({ ns: 'some-other-plugin', section: {} })),
]) {
const error = expectErr(response)
expect(error.code).toBe('settings-not-exposed')
expect(error.details).toEqual({ ns: 'some-other-plugin' })
}
// The write never reached the seam.
expect(ctx.settings.describe().find(d => String(d.ns) === 'some-other-plugin')?.value).toEqual({})
})
it('serves the product onboarding namespace without invalidating the model catalog', async () => {
const ctx = await harness()
ctx.settings.register(settingsNamespace('ui-onboarding'), z.object({ welcomeNoticeVersion: z.string() }))
const api = createApiProxy(ctx, DEFAULTS)
expect(expectOk(await api.settings.describe(request({}))).namespaces.map(view => view.ns))
.toEqual(['ui-onboarding'])
const frames = await collectHost(api, ['host/settings-changed'], 1, async () => {
expectOk(await api.settings.mutate(request({
ns: 'ui-onboarding',
ops: [{ op: 'set', path: ['welcomeNoticeVersion'], value: 'v1' }],
})))
})
expect(frames).toEqual([{ type: 'host/settings-changed', ns: 'ui-onboarding' }])
})
it('serves the agent-preset namespace, so a browser preset picker can persist its choice', async () => {
const ctx = await harness()
ctx.settings.register(settingsNamespace('agent-presets'), z.object({ default: z.string() }))
const api = createApiProxy(ctx, DEFAULTS)
expectOk(await api.settings.update(request({ ns: 'agent-presets', patch: { default: 'minimal' } })))
// Both browser surfaces that offer the choice — the General row and the
// management section — write the default through `settings.update`, so a
// namespace outside this boundary makes the picker move and then silently
// forget, which is worse than refusing the control.
expect(ctx.settings.describe().find(view => String(view.ns) === 'agent-presets')?.value)
.toEqual({ default: 'minimal' })
})
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
it('refuses even a model-provider namespace once its directory entry is gone', async () => {
const ctx = await harness({ configurableProviders: false })
ctx.settings.register(NS, AdapterConfig)
const api = createApiProxy(ctx, DEFAULTS)
expect(expectOk(await api.settings.describe(request({}))).namespaces).toEqual([])
expect(expectErr(await api.settings.update(request({ ns: 'llm-deepseek', patch: { baseURL: 'https://x' } }))).code)
.toBe('settings-not-exposed')
})
feat(settings): detect stale writers with a revision, and announce raw changes The remaining P1 from the #939 review, plus the P2 it shares a mechanism with. Nothing carried a version, so two tabs editing one namespace silently overwrote each other — reproduced as tab B's `reasoning` lost to tab A's older draft. The seam's per-namespace write queue orders writes; it cannot tell a fresh writer from one replaying a snapshot a predecessor superseded. Each namespace now carries a monotonic `revision` over its RAW section. A write may send `expectedRevision`, checked at the FRONT of the queue (not at call time, which would race the very predecessor it guards against); a mismatch rejects with `SettingsConflictError` → `settings-conflict` on the wire, carrying both revisions. The editor captures the revision it opened at and, on conflict, asks the user to reopen rather than replaying its snapshot. The same counter fixes the missing broadcast. `settings/updated` is gated on the resolved value — correct for consumers, wrong for configuration surfaces: storing an override equal to the composition base leaves the resolved value alone while changing what the document says (the field is now overridden, not inherited) and moving every open editor's revision. `settings/document-updated (ns, revision)` fires on any raw-section change, in-process or external, and `host/settings-changed` now rides it. That event also closes the stale model picker: editing a provider's `models` changes no route, so `llm/adapters-updated` never fired and an open picker kept serving the old catalog. A change to an exposed provider namespace now emits `host/models-changed` too — that namespace holds the catalog. Docs: both sides of the five touched README pairs, a type-equiv block for `SettingsPathOp`, and an Agent Note recording what the plane exposes and who may overwrite what. The deferred wire-redaction gaps (secrets behind union/intersection/transform, `.default(...)` in the served envelope, schema text in rejection messages, `new Function` rehydration, pi-ai's `headers`) are recorded as TODO(settings-wire-redaction) and in Known Limitations rather than half-fixed.
2026-07-30 19:24:21 +08:00
it('invalidates the model catalog when a provider namespace changes, and broadcasts a raw-only change', async () => {
// Editing `models` changes no route, so llm/adapters-updated never fires
// and an open model picker kept serving the old catalog. And storing an
// override equal to the resolved value emits nothing on settings/updated,
// so another tab never learned the field became overridden.
const ctx = await harness()
ctx.settings.register(NS, AdapterConfig, { base: { baseURL: 'https://base' } })
const api = createApiProxy(ctx, DEFAULTS)
const frames = await collectHost(api, ['host/settings-changed', 'host/models-changed'], 2, async () => {
await api.settings.update(request({ ns: 'llm-deepseek', patch: { baseURL: 'https://base' } }))
})
expect(frames).toEqual([
{ type: 'host/settings-changed', ns: 'llm-deepseek' },
{ type: 'host/models-changed' },
])
// The resolved value never moved: base already said https://base.
expect(expectOk(await api.settings.describe(request({}))).namespaces[0]!.value)
.toEqual({ apiKeyEnv: 'DEEPSEEK_API_KEY', baseURL: 'https://base' })
})
it('broadcasts a permission change without invalidating the model catalog', async () => {
const ctx = await harness()
const permission = ctx.settings.register(settingsNamespace('permission'), z.object({
defaultPreset: z.union(['read-only', 'workspace-write']).required(),
}), {
base: { defaultPreset: 'read-only' },
})
const api = createApiProxy(ctx, DEFAULTS)
const frames = await collectHost(api, ['host/settings-changed', 'host/models-changed'], 1, async () => {
await permission.update({ defaultPreset: 'workspace-write' })
})
expect(frames).toEqual([{ type: 'host/settings-changed', ns: 'permission' }])
})
feat(settings): detect stale writers with a revision, and announce raw changes The remaining P1 from the #939 review, plus the P2 it shares a mechanism with. Nothing carried a version, so two tabs editing one namespace silently overwrote each other — reproduced as tab B's `reasoning` lost to tab A's older draft. The seam's per-namespace write queue orders writes; it cannot tell a fresh writer from one replaying a snapshot a predecessor superseded. Each namespace now carries a monotonic `revision` over its RAW section. A write may send `expectedRevision`, checked at the FRONT of the queue (not at call time, which would race the very predecessor it guards against); a mismatch rejects with `SettingsConflictError` → `settings-conflict` on the wire, carrying both revisions. The editor captures the revision it opened at and, on conflict, asks the user to reopen rather than replaying its snapshot. The same counter fixes the missing broadcast. `settings/updated` is gated on the resolved value — correct for consumers, wrong for configuration surfaces: storing an override equal to the composition base leaves the resolved value alone while changing what the document says (the field is now overridden, not inherited) and moving every open editor's revision. `settings/document-updated (ns, revision)` fires on any raw-section change, in-process or external, and `host/settings-changed` now rides it. That event also closes the stale model picker: editing a provider's `models` changes no route, so `llm/adapters-updated` never fired and an open picker kept serving the old catalog. A change to an exposed provider namespace now emits `host/models-changed` too — that namespace holds the catalog. Docs: both sides of the five touched README pairs, a type-equiv block for `SettingsPathOp`, and an Agent Note recording what the plane exposes and who may overwrite what. The deferred wire-redaction gaps (secrets behind union/intersection/transform, `.default(...)` in the served envelope, schema text in rejection messages, `new Function` rehydration, pi-ai's `headers`) are recorded as TODO(settings-wire-redaction) and in Known Limitations rather than half-fixed.
2026-07-30 19:24:21 +08:00
it('maps a stale expectedRevision to settings-conflict carrying both revisions', async () => {
const ctx = await harness()
ctx.settings.register(NS, AdapterConfig)
const api = createApiProxy(ctx, DEFAULTS)
const opened = expectOk(await api.settings.describe(request({}))).namespaces[0]!.revision
expect(expectOk(await api.settings.update(request({ ns: 'llm-deepseek', patch: { baseURL: 'https://first' }, expectedRevision: opened })))
.revision).toBe(opened + 1)
const error = expectErr(await api.settings.update(request({ ns: 'llm-deepseek', patch: { baseURL: 'https://second' }, expectedRevision: opened })))
expect(error.code).toBe('settings-conflict')
expect(error.details).toEqual({ ns: 'llm-deepseek', expected: opened, actual: opened + 1 })
// The refused write changed nothing.
expect(expectOk(await api.settings.describe(request({}))).namespaces[0]!.user).toEqual({ baseURL: 'https://first' })
})
it('updates the user layer, answers with the new redacted view, and broadcasts the frame', async () => {
const ctx = await harness()
ctx.settings.register(NS, AdapterConfig, { base: { baseURL: 'https://base' } })
const api = createApiProxy(ctx, DEFAULTS)
const frames = await collectHost(api, ['host/settings-changed'], 1, async () => {
const view = expectOk(await api.settings.update(request({ ns: 'llm-deepseek', patch: { apiKey: 'sk-new', baseURL: 'https://next' } })))
expect(view.value).toEqual({ apiKeyEnv: 'DEEPSEEK_API_KEY', baseURL: 'https://next' })
expect(view.user).toEqual({ baseURL: 'https://next' })
expect(view.secrets).toEqual([{ path: ['apiKey'], set: true }])
expect(JSON.stringify(view)).not.toContain('sk-new')
})
expect(frames).toEqual([{ type: 'host/settings-changed', ns: 'llm-deepseek' }])
})
it('replace resets the user layer wholesale', async () => {
const ctx = await harness({ settings: { doc: { 'llm-deepseek': { baseURL: 'https://user' } } } })
ctx.settings.register(NS, AdapterConfig)
const api = createApiProxy(ctx, DEFAULTS)
const view = expectOk(await api.settings.replace(request({ ns: 'llm-deepseek', section: {} })))
expect(view.value).toEqual({ apiKeyEnv: 'DEEPSEEK_API_KEY' })
expect(view.user).toEqual({})
})
it.each([
['an invalid namespace name', 'Not A Namespace', {}],
['a schema-invalid patch', 'llm-deepseek', { baseURL: 42 }],
])('rejects %s as settings-rejected', async (_case, ns, patch) => {
const ctx = await harness()
ctx.settings.register(NS, AdapterConfig)
const api = createApiProxy(ctx, DEFAULTS)
const error = expectErr(await api.settings.update(request({ ns, patch })))
expect(error.code).toBe('settings-rejected')
expect(error.details).toEqual({ ns })
})
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
it('answers an unregistered namespace exactly like an unexposed one', async () => {
// Deliberately indistinguishable: separating "does not exist" from
// "exists but is not yours to configure" would let a caller enumerate the
// registered namespaces one probe at a time.
const ctx = await harness()
ctx.settings.register(NS, AdapterConfig)
ctx.settings.register(settingsNamespace('some-other-plugin'), z.object({ secretPath: z.string() }))
const api = createApiProxy(ctx, DEFAULTS)
const unknown = expectErr(await api.settings.update(request({ ns: 'unknown-ns', patch: {} })))
const unexposed = expectErr(await api.settings.update(request({ ns: 'some-other-plugin', patch: {} })))
expect(unknown.code).toBe('settings-not-exposed')
expect(unexposed.code).toBe(unknown.code)
expect(unexposed.message.replace('some-other-plugin', 'unknown-ns')).toBe(unknown.message)
})
it('maps a read-only provider refusal onto the same rejection', async () => {
const ctx = await harness({ settings: { readOnly: true } })
ctx.settings.register(NS, AdapterConfig)
const api = createApiProxy(ctx, DEFAULTS)
const value = expectOk(await api.settings.describe(request({})))
expect(value.writable).toBe(false)
const error = expectErr(await api.settings.update(request({ ns: 'llm-deepseek', patch: {} })))
expect(error.code).toBe('settings-rejected')
expect(error.message).toContain('read-only')
})
})
describe('credentials domain', () => {
it('reports an actionable error when no credential provider is mounted', async () => {
const ctx = await harness({ credentials: false })
const api = createApiProxy(ctx, DEFAULTS)
const error = expectErr(await api.credentials.describe(request({ refs: ['A'] })))
expect(error.code).toBe('internal')
expect(error.message).toContain('dsh-credentials-local')
})
it('describes value-free views and flips state through set/unset with frames', async () => {
const ctx = await harness()
const api = createApiProxy(ctx, DEFAULTS)
const before = expectOk(await api.credentials.describe(request({ refs: ['OPENAI_API_KEY'] })))
expect(before.credentials).toEqual({ OPENAI_API_KEY: { configured: false, writable: true } })
const frames = await collectHost(api, ['host/credentials-changed'], 2, async () => {
expectOk(await api.credentials.set(request({ ref: 'OPENAI_API_KEY', value: 'sk-secret' })))
const after = expectOk(await api.credentials.describe(request({ refs: ['OPENAI_API_KEY'] })))
expect(after.credentials).toEqual({ OPENAI_API_KEY: { configured: true, source: 'file', writable: true } })
expect(JSON.stringify(after)).not.toContain('sk-secret')
expectOk(await api.credentials.unset(request({ ref: 'OPENAI_API_KEY' })))
})
expect(frames).toEqual([
{ type: 'host/credentials-changed', ref: 'OPENAI_API_KEY' },
{ type: 'host/credentials-changed', ref: 'OPENAI_API_KEY' },
])
})
it('maps a shadowed write onto credential-rejected for set and unset alike', async () => {
const ctx = await harness({ credentials: { shadowed: ['DEEPSEEK_API_KEY'] } })
const api = createApiProxy(ctx, DEFAULTS)
const described = expectOk(await api.credentials.describe(request({ refs: ['DEEPSEEK_API_KEY'] })))
expect(described.credentials['DEEPSEEK_API_KEY']).toEqual({ configured: true, source: 'env', writable: false })
const setError = expectErr(await api.credentials.set(request({ ref: 'DEEPSEEK_API_KEY', value: 'x' })))
expect(setError.code).toBe('credential-rejected')
expect(setError.details).toEqual({ ref: 'DEEPSEEK_API_KEY' })
const unsetError = expectErr(await api.credentials.unset(request({ ref: 'DEEPSEEK_API_KEY' })))
expect(unsetError.code).toBe('credential-rejected')
})
})
describe('llm domain', () => {
it('merges the configurable directory with live routes and appends undeclared ones', async () => {
fix(web-config): close the wire boundary, the redacted-replace data loss, and three P2s Five findings from the #939 review, each reproduced before being fixed. **Configuration reads are as privileged as writes.** `settings.describe` returns every exposed namespace's configuration and `credentials.describe` reports whether an arbitrary environment-variable name is configured and from where — reconnaissance no anonymous caller should have. Both join PRIVILEGED_METHODS, so the whole configuration plane is loopback-only until real authentication exists; `trustedHosts` was never authentication. The model catalog stays reachable: it carries no endpoints or key state, and a LAN client's model picker legitimately needs it. Asserted over a real HTTP server, because the Host header a browser actually sends is what decides this. **The proxy serves only namespaces a registered model provider addresses.** The settings seam is general — any plugin may register one — but the Web configuration plane is the model-provider surface. Without the gate, every future `settings.register()` would silently become remotely readable and writable configuration. An unregistered namespace and an unexposed one answer identically, so no caller can enumerate the registry one probe at a time. **Path-addressed writes replace the redacted-document rebuild.** The editor reads the REDACTED descriptor, so rebuilding a section from it and replacing wholesale deleted every literal secret the wire never returned — reproduced as `{baseURL, reasoning}` in, stored `apiKey` gone out. `settings.mutate` applies set/unset ops to the section as it stands at the front of the seam's write queue, and the client names only fields it can see, so an unseen secret is untouched by construction rather than by care. P2s in the same pass: `llm/adapters-updated` now contains async listener rejections (an uncontained one escaped as unhandledRejection, contradicting the documented "observer failures are contained"); llm-deepseek's retry-policy swap uses the atomic `registration.replace` instead of dispose-then-register, which published `[]` then `["deepseek-official"]` so an observer saw the provider disappear and come back; and a transport rejection no longer strands the page in `loading` or a card in `busy`, with removal failures surfaced on the page banner instead of swallowed.
2026-07-30 18:30:15 +08:00
const ctx = await harness({ configurableProviders: false })
ctx.llm.registerConfigurableProviders([
{ provider: 'deepseek-official', displayName: 'DeepSeek', settingsNs: 'llm-deepseek', settingsPath: [] },
{ provider: 'openai', displayName: 'openai', settingsNs: 'llm-pi-ai', settingsPath: ['providers', 'openai'] },
])
ctx.llm.registerAdapter(['deepseek-official'], new CatalogAdapter('DeepSeek', ['deepseek-v4-flash']))
ctx.llm.registerAdapter(['undeclared'], new CatalogAdapter('Undeclared', ['u-1']))
fix(llm): answer a catalog route's models from pi-ai's own registry Clicking "fetch available models" on a built-in provider went to the network. That is the wrong source: pi-ai's registry is the authoritative list for its own providers, and it carries the context windows and output caps a `GET /models` listing does not disclose. Asking api.deepseek.com what DeepSeek serves is both slower and worse, and against an endpoint that answers a different shape it failed outright. Interrogation is still keyed by settings namespace — the provider being added has no route — but the request may now name the route it is editing. An adapter that already describes that route answers from what it knows, needs no endpoint at all, and never touches the network; only a route the catalog does not describe reaches the wire, and one naming no endpoint is told to set one or enter its models by hand. `ConfigurableProviderView` gained `supportsDiscovery` so a surface offers the action where a namespace can answer instead of hardcoding an adapter family. Three narrower corrections ride along. Discovery no longer claims Azure or Codex: Azure authenticates with an `api-key` header and an `api-version` query despite its OpenAI lineage, and Codex uses OAuth, so both reported an authentication failure as a provider with no models. Cancellation during the body read escaped as the raw abort reason rather than a coded ABORTED. And the schema comment claiming the probe key is never logged overstated it: the host neither stores nor returns it, but it rides the client's outgoing envelope like every other secret-bearing payload, and redacting that tap is a configuration-plane-wide change.
2026-08-04 12:30:41 +08:00
// Only one namespace can answer an interrogation, so the flag follows the
// entry's namespace rather than being assumed for every row.
ctx.llm.registerModelDiscovery('llm-pi-ai', () => Promise.resolve([]))
const api = createApiProxy(ctx, DEFAULTS)
const value = expectOk(await api.llm.providers(request({})))
expect(value.providers).toEqual([
fix(host): pin model discovery to loopback and drop its unread wire field llm.discoverModels was reachable from any declared trusted host. The method takes a caller-supplied baseURL and makes the host issue a GET to it, then reports the status or the parsed body — so on a LAN deployment an anonymous caller had a probe for whatever the host can reach and the browser cannot, plus a path that carries a draft credential. The PRIVILEGED_METHODS doc already states the rule this broke: trustedHosts is a DNS-rebinding fence, not authentication, so the configuration plane stays loopback-same-origin. It is in that set now, asserted both against the hand-built fence and over real HTTP beside the catalog reads that deliberately stay reachable. supportsDiscovery and listModelDiscoveryNamespaces are gone. The field was required on the wire and read by nobody: its own contract said a surface should offer the action "instead of naming an adapter family it would have to hardcode", while the surface hardcodes llm-pi-ai in two places and gates the button on whether there is anything to probe. Its shape did not fit the second caller either — the create card has no row to read a per-row field from. Keeping a required field alive for a consumer that may never arrive costs every producer and fixture a value nobody consults, which is exactly how the fixtures drifted. The registry that fed it had no other production consumer, so registration and disposal are now observed through the offer itself. The Agent Note claimed the key is never logged, which the wire schema beside it already contradicts, and predated both the provider field and the catalog-answer path. The two new public types pointed at core.md without a type-equiv block or manifest entry, so the generated service catalog named documentation that did not exist.
2026-08-04 16:08:07 +08:00
{ provider: 'deepseek-official', displayName: 'DeepSeek', settingsNs: 'llm-deepseek', settingsPath: [], active: true },
{ provider: 'openai', displayName: 'openai', settingsNs: 'llm-pi-ai', settingsPath: ['providers', 'openai'], active: false },
fix(llm): answer a catalog route's models from pi-ai's own registry Clicking "fetch available models" on a built-in provider went to the network. That is the wrong source: pi-ai's registry is the authoritative list for its own providers, and it carries the context windows and output caps a `GET /models` listing does not disclose. Asking api.deepseek.com what DeepSeek serves is both slower and worse, and against an endpoint that answers a different shape it failed outright. Interrogation is still keyed by settings namespace — the provider being added has no route — but the request may now name the route it is editing. An adapter that already describes that route answers from what it knows, needs no endpoint at all, and never touches the network; only a route the catalog does not describe reaches the wire, and one naming no endpoint is told to set one or enter its models by hand. `ConfigurableProviderView` gained `supportsDiscovery` so a surface offers the action where a namespace can answer instead of hardcoding an adapter family. Three narrower corrections ride along. Discovery no longer claims Azure or Codex: Azure authenticates with an `api-key` header and an `api-version` query despite its OpenAI lineage, and Codex uses OAuth, so both reported an authentication failure as a provider with no models. Cancellation during the body read escaped as the raw abort reason rather than a coded ABORTED. And the schema comment claiming the probe key is never logged overstated it: the host neither stores nor returns it, but it rides the client's outgoing envelope like every other secret-bearing payload, and redacting that tap is a configuration-plane-wide change.
2026-08-04 12:30:41 +08:00
// An undeclared live route has no settings address, so nothing can be
// interrogated on its behalf either.
fix(host): pin model discovery to loopback and drop its unread wire field llm.discoverModels was reachable from any declared trusted host. The method takes a caller-supplied baseURL and makes the host issue a GET to it, then reports the status or the parsed body — so on a LAN deployment an anonymous caller had a probe for whatever the host can reach and the browser cannot, plus a path that carries a draft credential. The PRIVILEGED_METHODS doc already states the rule this broke: trustedHosts is a DNS-rebinding fence, not authentication, so the configuration plane stays loopback-same-origin. It is in that set now, asserted both against the hand-built fence and over real HTTP beside the catalog reads that deliberately stay reachable. supportsDiscovery and listModelDiscoveryNamespaces are gone. The field was required on the wire and read by nobody: its own contract said a surface should offer the action "instead of naming an adapter family it would have to hardcode", while the surface hardcodes llm-pi-ai in two places and gates the button on whether there is anything to probe. Its shape did not fit the second caller either — the create card has no row to read a per-row field from. Keeping a required field alive for a consumer that may never arrive costs every producer and fixture a value nobody consults, which is exactly how the fixtures drifted. The registry that fed it had no other production consumer, so registration and disposal are now observed through the offer itself. The Agent Note claimed the key is never logged, which the wire schema beside it already contradicts, and predated both the provider field and the catalog-answer path. The two new public types pointed at core.md without a type-equiv block or manifest entry, so the generated service catalog named documentation that did not exist.
2026-08-04 16:08:07 +08:00
{ provider: 'undeclared', displayName: 'Undeclared', settingsNs: '', settingsPath: [], active: true },
])
})
it('serves the host-scoped catalog with per-provider failures contained', async () => {
const ctx = await harness()
ctx.llm.registerAdapter(['deepseek-official'], new CatalogAdapter('DeepSeek', ['deepseek-v4-flash', 'deepseek-v4-pro']))
ctx.llm.registerAdapter(['broken'], new BrokenCatalogAdapter('Broken', []))
const api = createApiProxy(ctx, DEFAULTS)
const value = expectOk(await api.llm.models(request({})))
expect(value.groups).toEqual([{
id: 'deepseek-official',
name: 'DeepSeek',
models: [
{ id: 'deepseek-v4-flash', name: 'deepseek-v4-flash' },
{ id: 'deepseek-v4-pro', name: 'deepseek-v4-pro' },
],
}])
expect(value.failures).toEqual([{ id: 'broken', name: 'Broken', message: 'catalog backend down' }])
})
it('broadcasts host/models-changed at every topology commit point', async () => {
const ctx = await harness()
const api = createApiProxy(ctx, DEFAULTS)
const frames = await collectHost(api, ['host/models-changed'], 2, async () => {
const dispose = ctx.llm.registerAdapter(['deepseek-official'], new CatalogAdapter('DeepSeek', []))
dispose()
return Promise.resolve()
})
expect(frames).toEqual([{ type: 'host/models-changed' }, { type: 'host/models-changed' }])
})
})
feat(llm): interrogate a draft provider endpoint for its models Once a pi-ai route became a declaration rather than a catalog lookup, adding an OpenAI-compatible gateway meant knowing its model ids up front. Most such endpoints publish that list at `GET /models`, but no seam operation could ask: every one is keyed by a registered provider route, and the provider being added has no route, no stored profile, and no stored credential — the endpoint and key are values in a form. Interrogation is therefore keyed by settings namespace, which a configuration surface already holds from the configurable-provider directory. `registerModelDiscovery` offers it per namespace, `discoverModels` asks, and the request carries the draft itself. The reply is candidates, not a catalog: every field but the id is optional because most listings disclose nothing else, and adopting one is a settings write like any other. Nothing here reads or writes settings or credentials, so `settings.yaml` still decides what a route serves. `llm.discoverModels` carries the same draft over the wire. Its apiKey is the third and last payload a secret may ride, and it is never stored, logged, or echoed; every refusal folds into `model-discovery-failed`, naming the endpoint asked but never the credential offered. The pi-ai side is a plain GET for OpenAI-compatible protocols only — their listing shape is the one gateways, self-hosted servers, and the official endpoints agree on. Others say so, sending the user to hand-entry rather than reporting a guessed shape as an empty provider. The reply is read under a four-megabyte ceiling held on the bytes actually received, because the endpoint is a URL the user typed.
2026-08-04 10:14:46 +08:00
describe('llm.discoverModels', () => {
it('carries a draft to its namespace and returns candidates without storing anything', async () => {
const ctx = await harness()
const seen: unknown[] = []
ctx.llm.registerModelDiscovery('llm-pi-ai', (probe) => {
seen.push({ baseURL: probe.baseURL, api: probe.api, apiKey: probe.apiKey })
return Promise.resolve([
{ id: 'acme-large', name: 'Acme Large', contextWindow: 65_536, maxTokens: 4096 },
{ id: 'acme-small' },
])
})
const api = createApiProxy(ctx, DEFAULTS)
const value = expectOk(await api.llm.discoverModels(request({
settingsNs: 'llm-pi-ai',
baseURL: 'https://gateway.acme.example/v1',
api: 'openai-completions',
apiKey: 'probe-key',
})))
expect(value.models).toEqual([
{ id: 'acme-large', name: 'Acme Large', contextWindow: 65_536, maxTokens: 4096 },
{ id: 'acme-small' },
])
expect(seen).toEqual([{
baseURL: 'https://gateway.acme.example/v1',
api: 'openai-completions',
apiKey: 'probe-key',
}])
// Interrogating a draft is a read: no namespace gained a section, and no
// credential reference was written.
expect(expectOk(await api.settings.describe(request({}))).namespaces.map(view => view.ns))
.not.toContain('llm-pi-ai')
})
fix(llm): answer a catalog route's models from pi-ai's own registry Clicking "fetch available models" on a built-in provider went to the network. That is the wrong source: pi-ai's registry is the authoritative list for its own providers, and it carries the context windows and output caps a `GET /models` listing does not disclose. Asking api.deepseek.com what DeepSeek serves is both slower and worse, and against an endpoint that answers a different shape it failed outright. Interrogation is still keyed by settings namespace — the provider being added has no route — but the request may now name the route it is editing. An adapter that already describes that route answers from what it knows, needs no endpoint at all, and never touches the network; only a route the catalog does not describe reaches the wire, and one naming no endpoint is told to set one or enter its models by hand. `ConfigurableProviderView` gained `supportsDiscovery` so a surface offers the action where a namespace can answer instead of hardcoding an adapter family. Three narrower corrections ride along. Discovery no longer claims Azure or Codex: Azure authenticates with an `api-key` header and an `api-version` query despite its OpenAI lineage, and Codex uses OAuth, so both reported an authentication failure as a provider with no models. Cancellation during the body read escaped as the raw abort reason rather than a coded ABORTED. And the schema comment claiming the probe key is never logged overstated it: the host neither stores nor returns it, but it rides the client's outgoing envelope like every other secret-bearing payload, and redacting that tap is a configuration-plane-wide change.
2026-08-04 12:30:41 +08:00
it('carries the route being edited so an adapter can answer from its own registry', async () => {
const ctx = await harness()
let probe: unknown
ctx.llm.registerModelDiscovery('llm-pi-ai', (request_) => {
probe = request_
return Promise.resolve([{ id: 'from-registry', contextWindow: 65_536, maxTokens: 4096 }])
})
const api = createApiProxy(ctx, DEFAULTS)
const value = expectOk(await api.llm.discoverModels(request({
settingsNs: 'llm-pi-ai',
provider: 'deepseek',
})))
// No endpoint at all: a route the adapter already describes needs none.
expect(probe).toEqual({ provider: 'deepseek' })
expect(value.models).toEqual([{ id: 'from-registry', contextWindow: 65_536, maxTokens: 4096 }])
})
feat(llm): interrogate a draft provider endpoint for its models Once a pi-ai route became a declaration rather than a catalog lookup, adding an OpenAI-compatible gateway meant knowing its model ids up front. Most such endpoints publish that list at `GET /models`, but no seam operation could ask: every one is keyed by a registered provider route, and the provider being added has no route, no stored profile, and no stored credential — the endpoint and key are values in a form. Interrogation is therefore keyed by settings namespace, which a configuration surface already holds from the configurable-provider directory. `registerModelDiscovery` offers it per namespace, `discoverModels` asks, and the request carries the draft itself. The reply is candidates, not a catalog: every field but the id is optional because most listings disclose nothing else, and adopting one is a settings write like any other. Nothing here reads or writes settings or credentials, so `settings.yaml` still decides what a route serves. `llm.discoverModels` carries the same draft over the wire. Its apiKey is the third and last payload a secret may ride, and it is never stored, logged, or echoed; every refusal folds into `model-discovery-failed`, naming the endpoint asked but never the credential offered. The pi-ai side is a plain GET for OpenAI-compatible protocols only — their listing shape is the one gateways, self-hosted servers, and the official endpoints agree on. Others say so, sending the user to hand-entry rather than reporting a guessed shape as an empty provider. The reply is read under a four-megabyte ceiling held on the bytes actually received, because the endpoint is a URL the user typed.
2026-08-04 10:14:46 +08:00
it('omits a credential and protocol the draft does not name', async () => {
const ctx = await harness()
let probe: unknown
ctx.llm.registerModelDiscovery('llm-pi-ai', (request_) => {
probe = request_
return Promise.resolve([])
})
const api = createApiProxy(ctx, DEFAULTS)
expectOk(await api.llm.discoverModels(request({
settingsNs: 'llm-pi-ai',
baseURL: 'https://gateway.acme.example/v1',
})))
// Absent fields stay absent rather than crossing as explicit undefined:
// the adapter distinguishes "no protocol named" from "protocol undefined".
expect(probe).toEqual({ baseURL: 'https://gateway.acme.example/v1' })
})
it('reports a failed interrogation as the form\'s next move, naming no credential', async () => {
const ctx = await harness()
ctx.llm.registerModelDiscovery('llm-pi-ai', () =>
Promise.reject(new Error('https://gateway.acme.example/v1/models answered 401; check the API key')))
const api = createApiProxy(ctx, DEFAULTS)
const error = expectErr(await api.llm.discoverModels(request({
settingsNs: 'llm-pi-ai',
baseURL: 'https://gateway.acme.example/v1',
apiKey: 'wrong',
})))
expect(error.code).toBe('model-discovery-failed')
expect(error.message).toContain('answered 401; check the API key')
expect(error.details).toEqual({ settingsNs: 'llm-pi-ai', baseURL: 'https://gateway.acme.example/v1' })
expect(JSON.stringify(error)).not.toContain('wrong')
})
it('reports a namespace no adapter family serves', async () => {
const ctx = await harness()
const api = createApiProxy(ctx, DEFAULTS)
const error = expectErr(await api.llm.discoverModels(request({
settingsNs: 'llm-deepseek',
baseURL: 'https://api.deepseek.com',
})))
expect(error.code).toBe('model-discovery-failed')
expect(error.message).toContain('no model discovery is registered')
})
})