fix: address ds-review-bot v8 findings

- llm-deepseek: the uncatalogued resolveModel fallback declares text-only
  modalities — the wire route is text-only regardless of catalog
  membership, so "unknown" must not let the host persist-then-fail images.
- session.selectModel also consults the pending-inbox mirror: a queued
  image prompt enters the log only when claimed, after a switch would land.
- attachment store: ensureDurableDirectory syncs every ancestor entry up to
  a caller-vouched boundary regardless of what mkdir reports — a raced
  "already existed" is not "already durable".
- One image walker (imageBlockIn/imageInEvent) now serves both attachment
  authorization and the selection gate; referencedImage therefore also
  authorizes references inside wrapped message content.
- InputHub: the scope disposer resolves the conversation service optionally
  (teardown/HMR must reach quiescence), and a send failing after its scope
  died releases the in-flight drafts instead of restoring them onto a
  disposed shell.
- http-bridge destroys declared-oversize requests with connection: close
  instead of draining a body the client can trickle indefinitely.
- LlmService validates AND detaches modality arrays identically on the
  advisory and exact routes; READMEs record the fourth INVALID_MODEL_INFO
  rejection reason.
- CLI provider docs (JSDoc, README pair, Agent Note pair) describe the
  reuse behavior; llm-route.spec now parses the SHIPPED cordis.yml through
  the production extraction, pinning the row coupling.
- image-display lane pins gallery/rail shape in inline snapshots and the
  object-URL scheme this environment must take; stale host.schema comment
  dropped.
This commit is contained in:
creatixchu
2026-07-29 19:40:17 +08:00
parent adce3b833d
commit f73bf425ec
27 changed files with 337 additions and 100 deletions

View File

@@ -79,21 +79,28 @@ async function syncDirectory(path: string): Promise<void> {
}
/**
* Create one private directory tree and persist every newly published ancestor.
* Create one private directory tree and persist every ancestor entry up to a
* caller-vouched durable boundary. The walk deliberately ignores what mkdir
* reports as newly created: a concurrent first save can create a level this
* process then merely observes, so "already existed" is not "already durable"
* — the entry may still be unsynced in the creator, and a crash would drop a
* directory the session checkpoint already references. Re-syncing a durable
* entry is harmless; skipping an unsynced one is not.
* @param path - absolute directory to create.
* @param boundary - absolute ancestor the caller vouches is already durable.
*/
async function ensureDurableDirectory(path: string): Promise<void> {
async function ensureDurableDirectory(path: string, boundary: string): Promise<void> {
const target = resolve(path)
const firstCreated = await mkdir(target, { recursive: true, mode: 0o700 })
const stop = resolve(boundary)
await mkdir(target, { recursive: true, mode: 0o700 })
await chmod(target, 0o700)
if (firstCreated === undefined) return
const highestCreated = resolve(firstCreated)
let created = target
while (true) {
await syncDirectory(dirname(created))
if (created === highestCreated) return
created = dirname(created)
let level = target
while (level !== stop) {
const parent = dirname(level)
await syncDirectory(parent)
/* v8 ignore next -- filesystem-root guard: callers pass a boundary that is an ancestor of path, so the walk reaches it first. */
if (parent === level) return
level = parent
}
}
@@ -110,8 +117,12 @@ export async function saveImageFile(root: string, input: SaveImageAttachment, li
const sha256 = digest(input.data)
const bucket = join(root, 'objects', sha256.slice(0, 2))
const staging = join(root, 'tmp')
await ensureDurableDirectory(bucket)
await ensureDurableDirectory(staging)
// The durable boundary is the root's grandparent (DSH_HOME for the
// documented `DSH_HOME/attachments/v1` layout): `attachments`/`v1` may be
// first-created by a concurrent save, so their entries sync on every path.
const boundary = dirname(dirname(resolve(root)))
await ensureDurableDirectory(bucket, boundary)
await ensureDurableDirectory(staging, boundary)
const temporary = join(staging, randomUUID())
const target = objectPath(root, sha256)
let handle

View File

@@ -47,7 +47,7 @@ afterEach(async () => {
})
describe('local attachment store', () => {
it.skipIf(process.platform === 'win32')('syncs every newly created object ancestor before returning', async () => {
it.skipIf(process.platform === 'win32')('syncs every object ancestor up to the durable boundary before returning', async () => {
const storageRoot = await root()
const base = join(storageRoot, '..', '..')
const sha256 = createHash('sha256').update(PNG).digest('hex')
@@ -57,12 +57,20 @@ describe('local attachment store', () => {
await saveImageFile(storageRoot, { data: PNG, mediaType: 'image/png' }, LIMITS)
// Every level between each created directory and the vouched boundary
// syncs unconditionally — "already existed" is not "already durable"
// when a concurrent first save may have created but not yet synced it.
expect(fsControl.syncedDirectories).toEqual([
// bucket chain: every parent entry between the bucket and the boundary.
objects,
storageRoot,
join(storageRoot, '..'),
base,
// staging chain re-walks the shared ancestors after creating tmp.
storageRoot,
join(storageRoot, '..'),
base,
// publication: the settled object's bucket and its parent for the rename.
bucket,
objects,
])

View File

@@ -30,9 +30,13 @@ export async function bridge(
})
const declaredLength = req.headers['content-length']
if (declaredLength !== undefined && Number(declaredLength) > maxRequestBodyBytes) {
res.writeHead(413)
// Same discipline as the chunked-overrun path below: destroy, never
// drain. resume() would keep the socket open while the client trickles
// its declared length — an already-rejected request holding a server
// socket for as long as it likes.
res.writeHead(413, { connection: 'close' })
res.end()
req.resume()
req.destroy()
return
}
const chunks: Buffer[] = []

View File

@@ -5,6 +5,34 @@ import { describe, expect, it } from 'vitest'
import { bridge } from '../src/http-bridge.ts'
describe('HTTP bridge abort', () => {
it('destroys a declared-oversize request instead of draining it', async () => {
const destroyed: true[] = []
const request = Readable.from([]) as unknown as IncomingMessage
Object.assign(request, {
url: '/api/session.prompt',
method: 'POST',
headers: { 'content-type': 'application/json', 'content-length': '999999' },
destroy: () => { destroyed.push(true) },
})
let status: number | undefined
let headers: unknown
const response = Object.assign(new EventEmitter(), {
writableEnded: false,
writeHead(code: number, values?: unknown) { status = code; headers = values; return this },
write() { return true },
end(this: { writableEnded: boolean }) { this.writableEnded = true; return this },
}) as unknown as ServerResponse
await bridge(request, response, {
fetch: () => { throw new Error('a rejected request must never reach the handler') },
}, 1000)
// The socket must not stay parked draining a body the client can trickle
// at will after the rejection — same discipline as the chunked overrun.
expect(status).toBe(413)
expect(headers).toMatchObject({ connection: 'close' })
expect(destroyed).toHaveLength(1)
})
it('aborts a pending native picker request when the browser disconnects', async () => {
const body = JSON.stringify({
type: 'client-request', rpcId: 'picker-1', method: 'host.pickDirectory', payload: {},

View File

@@ -87,11 +87,17 @@ export class InputHub implements InputService {
for (const off of offs) off()
// Draft attachments die with the scope: the shell only holds ids, so
// the service-owned File objects and object URLs must be released
// here or they leak for the page lifetime.
// here or they leak for the page lifetime. The lookup is optional —
// during application teardown or HMR of this plugin the root
// `conversation` service can already be unregistered while session
// scopes are still alive; a throwing disposer would abort teardown
// quiescence, and the service's own disposal effect revokes every
// remaining URL in that case anyway.
const drafts = shell.snapshot.imageIds
shell.dispose()
this.shells.delete(id)
for (const imageId of drafts) this.conversation().releaseDraftImage(imageId)
const conversation = this.rootCtx.get('conversation') as ConversationAttachmentFace | undefined
for (const imageId of drafts) conversation?.releaseDraftImage(imageId)
}
}, 'conversation.input: session shell')
return shell
@@ -139,8 +145,18 @@ export class InputHub implements InputService {
// Commit, not an editable clear: undo must not resurrect sent content.
shell?.commitSend(imageIds)
void this.conversation().sendSession(session, text, mode, imageIds).catch(() => {
shell?.restoreImages(imageIds)
if (shell?.snapshot.draft === '') shell.setDraft(text)
// Restore only into the shell that still owns the session: if the scope
// died while the send was in flight, `commitSend` already removed the
// ids from the (now disposed) shell, so the teardown release could not
// see them — release the drafts here instead of resurrecting them onto
// a dead instance where they would leak for the page lifetime.
if (this.shells.get(session.sessionId) === shell) {
shell?.restoreImages(imageIds)
if (shell?.snapshot.draft === '') shell.setDraft(text)
return
}
const conversation = this.rootCtx.get('conversation') as ConversationAttachmentFace | undefined
for (const id of imageIds) conversation?.releaseDraftImage(id)
})
}

View File

@@ -69,6 +69,34 @@ describe('ConversationService', () => {
await b.runtime.dispose()
})
it('releases in-flight send images when the scope dies before the failure lands', async () => {
const b = await bench()
const created = vi.spyOn(URL, 'createObjectURL').mockReturnValue('blob:inflight-1')
const revoked = vi.spyOn(URL, 'revokeObjectURL').mockReturnValue(undefined)
try {
const [attachment] = b.root.createDraftImages([new File([new Uint8Array(4)], 'b.png', { type: 'image/png' })])
if (attachment === undefined) throw new Error('draft attachment missing')
const shell = b.hub.shell(b.runtime.sessions.behavior('s1').sessionId)
shell.addImages([attachment.id])
let reject!: (error: Error) => void
b.prompt.mockReturnValueOnce(new Promise((_resolve, rej) => { reject = rej }) as never)
shell.setDraft('x')
shell.submit('queue')
// commitSend already removed the ids from the shell; kill the scope
// while the RPC is still pending, then land the failure.
await b.runtime.sessions.remove('s1')
reject(new Error('transport died'))
await vi.waitFor(() => {
expect(revoked).toHaveBeenCalledWith('blob:inflight-1')
})
expect(b.root.draftImages([attachment.id])).toEqual([])
} finally {
created.mockRestore()
revoked.mockRestore()
}
await b.runtime.dispose()
})
it('fails loudly from the root scope, on an unbound session, or without SessionsService', async () => {
const b = await bench()
await expect(b.root.send('x', 'queue')).rejects.toThrow(/requires a session scope/)

View File

@@ -110,33 +110,48 @@ async function durablePromptContent(ctx: Context, content: readonly PromptConten
}))
}
function imageInContent(content: unknown, attachmentId: string): ImageAttachmentRef | undefined {
/**
* The ONE recursive block walk shared by attachment authorization and the
* model-selection gate (nested tool-result content included). Both consumers
* must agree on what counts as replayed image content — a route added to one
* walker but not the other would silently skip authorization or stranding
* protection — so there is exactly one walker, parameterized by match.
*/
function imageBlockIn(content: unknown, match: (ref: ImageAttachmentRef) => boolean): ImageAttachmentRef | undefined {
if (!Array.isArray(content)) return undefined
for (const value of content) {
if (typeof value !== 'object' || value === null || Array.isArray(value)) continue
const block = value as { type?: unknown; attachment?: unknown; content?: unknown }
if (block.type === 'image' && typeof block.attachment === 'object' && block.attachment !== null) {
const ref = block.attachment as ImageAttachmentRef
if (String(ref.attachmentId) === attachmentId) return ref
if (match(ref)) return ref
}
if (block.type === 'tool-result') {
const nested = imageInContent(block.content, attachmentId)
const nested = imageBlockIn(block.content, match)
if (nested !== undefined) return nested
}
}
return undefined
}
/** Every replayed content route of one event: direct content, wrapped message content, streamed block-end. */
function imageInEvent(event: SessionEvent, match: (ref: ImageAttachmentRef) => boolean): ImageAttachmentRef | undefined {
const data = event.data as { content?: unknown; message?: { content?: unknown }; chunk?: { type?: unknown; block?: unknown } }
const direct = imageBlockIn(data.content, match)
if (direct !== undefined) return direct
if (data.message !== undefined) {
const wrapped = imageBlockIn(data.message.content, match)
if (wrapped !== undefined) return wrapped
}
if (event.type === 'assistant/chunk' && data.chunk?.type === 'block-end') {
return imageBlockIn([data.chunk.block], match)
}
return undefined
}
/** True when any block (nested tool-result content included) is an image block. */
function contentHasImage(content: unknown): boolean {
if (!Array.isArray(content)) return false
for (const value of content) {
if (typeof value !== 'object' || value === null || Array.isArray(value)) continue
const block = value as { type?: unknown; content?: unknown }
if (block.type === 'image') return true
if (block.type === 'tool-result' && contentHasImage(block.content)) return true
}
return false
return imageBlockIn(content, () => true) !== undefined
}
/**
@@ -145,23 +160,13 @@ function contentHasImage(content: unknown): boolean {
* The log is immutable, so a true here is permanent for the session's life.
*/
function sessionHasImage(events: readonly SessionEvent[]): boolean {
return events.some((event) => {
const data = event.data as { content?: unknown; message?: { content?: unknown }; chunk?: { type?: unknown; block?: unknown } }
if (contentHasImage(data.content)) return true
if (data.message !== undefined && contentHasImage(data.message.content)) return true
return event.type === 'assistant/chunk' && data.chunk?.type === 'block-end' && contentHasImage([data.chunk.block])
})
return events.some(event => imageInEvent(event, () => true) !== undefined)
}
function referencedImage(events: readonly SessionEvent[], attachmentId: string): ImageAttachmentRef | undefined {
for (const event of events) {
const data = event.data as { content?: unknown; chunk?: { type?: unknown; block?: unknown } }
const direct = imageInContent(data.content, attachmentId)
if (direct !== undefined) return direct
if (event.type === 'assistant/chunk' && data.chunk?.type === 'block-end') {
const streamed = imageInContent([data.chunk.block], attachmentId)
if (streamed !== undefined) return streamed
}
const found = imageInEvent(event, ref => String(ref.attachmentId) === attachmentId)
if (found !== undefined) return found
}
return undefined
}
@@ -1141,7 +1146,11 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
// wire routes reject image content on text-only models — accepting
// this selection would strand the session (every turn fails, no
// in-product recovery). Refuse at the selection boundary instead.
if (sessionHasImage(found.agent.session.events)) {
// The pending inbox counts too: a queued image prompt enters the log
// only when claimed, which would happen AFTER this switch landed.
const queuedImage = (queuedMirror.get(sessionId) ?? [])
.some(entry => contentHasImage(entry.message.content))
if (queuedImage || sessionHasImage(found.agent.session.events)) {
const info = await ctx.llm.resolveModelInfo(resolved.provider, resolved.model)
if (info.inputModalities !== undefined && !info.inputModalities.includes('image')) {
return err(request, {

View File

@@ -37,8 +37,6 @@ export const hostDescribeValueSchema = z.object({
mediaTypes: z.array(imageMediaTypeSchema),
}).optional(),
attachedSessions: z.number().int().nonnegative(),
// Open string, not a literal union: unknown kinds must survive the wire so
// a merge-added capability can advertise (the client hides the affordance).
}) satisfies z.ZodType<Wire<ResponseValue<'host.describe'>>>
/** host.pickDirectory request payload (empty object literal). */

View File

@@ -274,6 +274,52 @@ describe('Web session model selection', () => {
await ctx.fiber.dispose()
})
it('refuses a text-only selection while an image prompt is still queued (not yet logged)', async () => {
const { ctx, sessionId, agent } = await harness()
ctx.llm.registerAdapter(['text-only'], new class extends CatalogAdapter {
override resolveModel(provider: string, model: string): Promise<LlmResolvedModelInfo> {
return Promise.resolve({ provider, id: model, name: model, inputModalities: ['text'] })
}
}('Text Only', []))
const api = createApiProxy(ctx, { provider: 'deepseek', model: 'deepseek-chat', cwd: '/tmp', workspaceRoot: '/tmp' })
// The queued message enters the session log only when claimed — after a
// model switch would already have landed. The pending-inbox mirror must
// therefore gate the switch too.
ctx.emit('agent/inbox/enqueue', agent, {
id: 'q-1', role: 'user', source: { kind: 'user' },
content: [{ type: 'image', attachment: { attachmentId: 'att-q', mediaType: 'image/png', bytes: 8, width: 1, height: 1 } }],
} as never, 'queued')
const stranded = await api.sessions.selectModel(request({ sessionId, provider: 'text-only', model: 'plain' }))
expect(stranded.result.ok).toBe(false)
// Claiming the message drains the mirror; the log now owns the decision.
ctx.emit('agent/inbox/dequeue', agent, { id: 'q-1' } as never, 'queued')
expect(expectValue(await api.sessions.selectModel(request({
sessionId, provider: 'text-only', model: 'plain',
}))).selected).toEqual({ provider: 'text-only', model: 'plain' })
await ctx.fiber.dispose()
})
it('authorizes an attachment read referenced only from wrapped message content', async () => {
const { ctx, sessionId, agent } = await harness()
const ref = { attachmentId: 'att-w', mediaType: 'image/png' as const, bytes: 4, width: 1, height: 1 }
ctx.provide('attachments', {
readImage: () => Promise.resolve({ ref, data: new Uint8Array([1, 2, 3, 4]) }),
} as never)
const api = createApiProxy(ctx, { provider: 'deepseek', model: 'deepseek-chat', cwd: '/tmp', workspaceRoot: '/tmp' })
// The only reference lives inside an assistant/message wrapper — the same
// walk that gates model selection must authorize the read, or a real host
// denies galleries the fixture (with its own authorization mirror) serves.
agent.session.append('assistant/message', {
turn: 1, step: 0,
message: { id: 'a-1', role: 'assistant', source: { kind: 'model', provider: 'p', model: 'm' }, content: [{ type: 'image', attachment: ref }] },
} as never, { surfaceOp: 'append' })
const got = await api.sessions.attachment(request({ sessionId, attachmentId: 'att-w' as never }))
expect(got.result).toMatchObject({ ok: true, value: { attachment: ref } })
const denied = await api.sessions.attachment(request({ sessionId, attachmentId: 'att-other' as never }))
expect(denied.result).toMatchObject({ ok: false, error: { details: { reason: 'ATTACHMENT_NOT_REFERENCED' } } })
await ctx.fiber.dispose()
})
it('detects images on every replayed route: wrapped messages, streamed blocks, nested tool results', async () => {
const image = { type: 'image', attachment: { attachmentId: 'att-x', mediaType: 'image/png', bytes: 8, width: 1, height: 1 } }
const cases: { label: string; append: (agent: Agent) => void }[] = [

View File

@@ -166,8 +166,12 @@ export class DeepSeekAdapter extends LlmAdapter {
const contextWindow = configured?.contextWindow
?? this.options.defaultContextWindow
return Promise.resolve({
// The chat-completions wire route is text-only regardless of catalog
// membership, so the uncatalogued fallback declares the same negative
// capability — "unknown" here would let the host accept and persist
// images the serializer must then reject.
...configured === undefined
? { provider, id: model, name: model }
? { provider, id: model, name: model, inputModalities: ['text' as const], outputModalities: ['text' as const] }
: modelInfo(provider, configured),
...contextWindow === undefined ? {} : { context: { contextWindow } },
...this.options.defaults?.thinking === 'disabled'

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write packages/llm/llm/README.md
README.md: d343449d1530bf70a3a8c57f883894e29c42d18f
README.zh.md: 6ac57b1e6010b58c45b516f13ec6361d47ca8d12
README.md: ecd6da304a894a687153608f5b468f867ad32e6b
README.zh.md: 2812f2d74ec3dbab521f8e3148e9efe2640c6929

View File

@@ -23,7 +23,7 @@ An adapter registry plus a single streaming call surface, interceptable via a wa
Provider and model metadata is a discovery surface, not a routing whitelist. `registerAdapter()` still owns provider exclusivity and captures the adapter's retry policy for each route, while an adapter may accept model ids absent from `listModels()`; consumers must not reject a request because its model is unlisted. Returned selector metadata is detached and invalid or duplicate adapter entries fail with `INVALID_ADAPTER` or `INVALID_CATALOG`.
Exact-model metadata is a separate correctness query, not a catalog decoration or global LLM setting. `resolveModelInfo()` asks the adapter that owns the exact provider/model route once; an adapter can describe an unlisted dynamic model, and absent `context` or `reasoning` fields mean only that those capabilities are unavailable. Invalid identity, context, or reasoning metadata fails with `INVALID_MODEL_INFO`, `INVALID_MODEL_CONTEXT`, or `INVALID_MODEL_REASONING`.
Exact-model metadata is a separate correctness query, not a catalog decoration or global LLM setting. `resolveModelInfo()` asks the adapter that owns the exact provider/model route once; an adapter can describe an unlisted dynamic model, and absent `context` or `reasoning` fields mean only that those capabilities are unavailable. Invalid identity or modality metadata fails with `INVALID_MODEL_INFO`, and invalid context or reasoning metadata with `INVALID_MODEL_CONTEXT` or `INVALID_MODEL_REASONING`.
Reasoning identifiers are opaque adapter-owned strings rather than a core enum. An adapter publishes its ordered selectable list, including an `off` id when that model's capability API exposes one. `resolveCallConfig()` accepts only an exact advertised identifier, materializes `defaultEffort` when present, and otherwise preserves the provider default. Asynchronous model resolvers receive the caller's signal and must settle promptly after cancellation. `prepareCall()` additionally retains the exact adapter registration through header logging and terminal dispatch, so HMR cannot combine one adapter's capability result with another adapter's request; reusing its one-shot handle or changing its call-config fields fails with `INVALID_PREPARED_CALL`. An unsupported explicit or configured effort fails with `UNSUPPORTED_REASONING_EFFORT` before provider I/O.

View File

@@ -23,7 +23,7 @@
提供方与模型元数据是发现表层,不是路由白名单。`registerAdapter()` 仍拥有提供方排他性,并为每条路由捕获适配器的重试策略;适配器则可以接受 `listModels()` 中不存在的模型 id消费方禁止因模型未列出而拒绝请求。返回的 selector 元数据与输入脱离,无效或重复适配器配置项会以 `INVALID_ADAPTER``INVALID_CATALOG` 失败。
确切模型元数据是独立的正确性查询,不是 catalog 装饰或全局 LLM 设置。`resolveModelInfo()` 会向拥有精确提供方/模型路由的适配器查询一次;适配器可以描述未列出的动态模型,缺少 `context``reasoning` 字段只表示相应能力不可用。无效的身份、上下文或推理元数据会以 `INVALID_MODEL_INFO``INVALID_MODEL_CONTEXT``INVALID_MODEL_REASONING` 失败。
确切模型元数据是独立的正确性查询,不是 catalog 装饰或全局 LLM 设置。`resolveModelInfo()` 会向拥有精确提供方/模型路由的适配器查询一次;适配器可以描述未列出的动态模型,缺少 `context``reasoning` 字段只表示相应能力不可用。无效的身份或模态元数据会以 `INVALID_MODEL_INFO` 失败,无效的上下文或推理元数据则以 `INVALID_MODEL_CONTEXT``INVALID_MODEL_REASONING` 失败。
推理标识符是由适配器持有的不透明字符串,而非核心枚举。适配器会公布有序可选列表;模型能力 API 提供 `off` id 时,列表也会包含它。`resolveCallConfig()` 只接受与已公布标识符完全一致的值,在存在 `defaultEffort` 时填入它,否则保留提供方默认值。异步模型解析器会接收调用方的 signal并且必须在取消后迅速完成结算。`prepareCall()` 还会让精确适配器注册跨越请求头记录和最终分派,因此 HMR热模块替换不会将一个适配器的能力结果与另一个适配器的请求混用复用其一次性句柄或更改调用配置字段会以 `INVALID_PREPARED_CALL` 失败。不支持的显式或配置推理强度会在提供方 I/O 前以 `UNSUPPORTED_REASONING_EFFORT` 失败。

View File

@@ -13,6 +13,7 @@ import type {
LlmModelInfo,
LlmResolvedModelInfo,
LlmProviderInfo,
ModelModality,
StreamChunk,
} from './types.ts'
import { freezeMessage, type Message } from './message.ts'
@@ -253,6 +254,28 @@ export class LlmService extends Service {
return this.registration(provider).retryPolicy
}
/**
* Validate adapter-owned modality arrays and detach them. One rule for the
* advisory catalog and exact resolution: both validate, both copy — two
* readings of the same adapter field with different trust or detachment
* would be an unexplained asymmetry.
* @param provider - provider route (diagnostic context).
* @param code - error code matching the calling surface.
* @param modalities - adapter-owned array, or undefined for unknown.
* @returns a detached copy, or undefined when absent.
*/
private detachedModalities(
provider: string,
code: 'INVALID_CATALOG' | 'INVALID_MODEL_INFO',
modalities: readonly unknown[] | undefined,
): ModelModality[] | undefined {
if (modalities === undefined) return undefined
if (!Array.isArray(modalities) || modalities.some(entry => typeof entry !== 'string')) {
throw new LlmError(`adapter returned invalid modality metadata for provider "${provider}"`, code)
}
return [...(modalities as readonly ModelModality[])]
}
/**
* Discover models advertised by one registered provider. Catalog membership
* is advisory and never changes routing or request validation.
@@ -277,13 +300,15 @@ export class LlmService extends Service {
throw new LlmError(`adapter returned invalid or duplicate model metadata for provider "${provider}"`, 'INVALID_CATALOG')
}
seen.add(model.id)
const inputModalities = this.detachedModalities(provider, 'INVALID_CATALOG', model.inputModalities)
const outputModalities = this.detachedModalities(provider, 'INVALID_CATALOG', model.outputModalities)
return {
provider: model.provider,
id: model.id,
name: model.name,
...model.description === undefined ? {} : { description: model.description },
...model.inputModalities === undefined ? {} : { inputModalities: [...model.inputModalities] },
...model.outputModalities === undefined ? {} : { outputModalities: [...model.outputModalities] },
...inputModalities === undefined ? {} : { inputModalities },
...outputModalities === undefined ? {} : { outputModalities },
}
})
}
@@ -333,23 +358,17 @@ export class LlmService extends Service {
'INVALID_MODEL_CONTEXT',
)
}
for (const modalities of [resolved.inputModalities, resolved.outputModalities]) {
if (modalities !== undefined && (!Array.isArray(modalities) || modalities.some(m => typeof m !== 'string'))) {
throw new LlmError(
`adapter returned invalid modality metadata for provider "${provider}" model "${model}"`,
'INVALID_MODEL_INFO',
)
}
}
// Capability metadata rides through: an explicit modality omission is
// negative capability downstream preflights act on (image admission).
const inputModalities = this.detachedModalities(provider, 'INVALID_MODEL_INFO', resolved.inputModalities)
const outputModalities = this.detachedModalities(provider, 'INVALID_MODEL_INFO', resolved.outputModalities)
const info: LlmResolvedModelInfo = {
provider,
id: model,
name: resolved.name,
...resolved.description === undefined ? {} : { description: resolved.description },
// Capability metadata rides through: an explicit modality omission is
// negative capability downstream preflights act on (image admission).
...resolved.inputModalities === undefined ? {} : { inputModalities: resolved.inputModalities },
...resolved.outputModalities === undefined ? {} : { outputModalities: resolved.outputModalities },
...inputModalities === undefined ? {} : { inputModalities },
...outputModalities === undefined ? {} : { outputModalities },
...context === undefined ? {} : { context: { contextWindow: context.contextWindow } },
}
const reasoning = resolved.reasoning

View File

@@ -1177,6 +1177,8 @@ describe('LlmService', () => {
[{ provider: 'route', id: 'm', name: 1 }, 'non-string name'],
[{ provider: 'route', id: 'm', name: '' }, 'empty name'],
[{ provider: 'route', id: 'm', name: 'M', description: 1 }, 'non-string description'],
[{ provider: 'route', id: 'm', name: 'M', inputModalities: 'text' }, 'non-array input modalities'],
[{ provider: 'route', id: 'm', name: 'M', outputModalities: [1] }, 'non-string output modality'],
] as const)('rejects invalid model metadata (%s: %s)', async (metadata, _label) => {
const ctx = new Context()
await ctx.plugin(LlmService)