fix: address ds-review-bot v7 findings on the merged image-input head

- gate model selection on steering-placement image carriers from enqueue
  until their steering/message event publishes; release the gate when an
  admission ends idle without publication (both behaviorally asserted)
- reject session.updateQueue edits carrying non-text blocks at the RPC
  boundary (queue edits cannot bypass image admission)
- extend the durable-directory walk past a first-created DSH_HOME to the
  deepest pre-existing ancestor
- strip Windows-style separators from attachment display names on POSIX
- verify attachment reads with a header-only probe (digest already proves
  the bytes decoded fully at admission); document the read path
- make SessionInputShell.addImages refusal observable and keep workspace
  transfers/composer intake from leaking refused drafts
- own ONE recursive image walk (dsh-llm contentHasImage) across apiproxy,
  pi-ai, compact-basic, and the DeepSeek text-only assertion
- drop the redundant canonical-base64 regex and the no-op role read
- move AttachmentId/AttachmentError out of types.ts (brand.ts/error.ts);
  document why AttachmentError does not extend HarnessError
- document the hard attachments inject in both consumer READMEs
This commit is contained in:
creatixchu
2026-07-30 14:34:08 +08:00
parent 97cf33b7e0
commit 0d1250f743
37 changed files with 335 additions and 140 deletions

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/host/apiproxy/README.md
README.md: 3badccd4144babc474f52fa44598ddec3c253e65
README.zh.md: 8ccd5bf95c70c79e968b3ea8d0f6a48c62359ae3
README.md: 7ee67009c33d9df843e1d5aa71e727c0c798d4a0
README.zh.md: f7d1503d2947f84dbc21cb74e10678c4e14cb248

View File

@@ -40,6 +40,7 @@ None; this package neither assembles nor sends a provider request.
## Known Limitations and Deferred Work
- **`attachments` is a hard inject** — the proxy will not mount until an attachment backend provides `ctx.attachments`; a composition missing one stalls silently as a cordis inject gap rather than failing loud (same gap as the connection route; a capability-degraded mount is deferred work).
- **`respond` routing is shipped, but pending-interaction state is host-side work** — the wire shape (POST `/api/respond`, `RpcReceipt`) is final; the pending table that makes late/duplicate answers meaningful lives in `src/api-proxy.ts` and is still minimal (questions only, no approvals).
- **Reserved seams stay out of `RpcMethodMap`** — `session.fork`, `prompt.mode: 'inject'`, `task.list`, `host.listModels`, and a describe `hostInstanceId` are documented reservations; an unknown method fails loud at envelope parse rather than getting a not-implemented code.
- **No protocol version field** — client and host ship together; `host.describe` gains a version negotiation field only when an independently released client exists.

View File

@@ -40,6 +40,7 @@ Workspace 列表与 Session 列表是相互独立的重连基线。`workspace.cr
## 已知限制与延期工作
- **`attachments` 是硬性注入依赖**:代理在附件后端提供 `ctx.attachments` 之前不会挂载;缺少后端的组合会以 cordis 注入缺口的形式静默停滞,而非响亮失败(与 connection 路由是同一缺口;降级挂载属于延期工作)。
- **`respond` 路由已经发布,但待处理交互状态仍属宿主侧工作**:协议形状(POST `/api/respond`、`RpcReceipt`)已经定型;使延迟或重复回答具有明确语义的待处理表位于 `src/api-proxy.ts`,目前仍很精简(只支持问题,不支持审批)。
- **预留 seam 不进入 `RpcMethodMap`**:`session.fork`、`prompt.mode: 'inject'`、`task.list`、`host.listModels` 和描述字段 `hostInstanceId` 都是已记录的预留项;未知方法会在信封解析时直接失败,而不会返回「尚未实现」错误码。
- **没有协议版本字段**:客户端与宿主一同发布;只有出现独立发布的客户端后,`host.describe` 才会增加版本协商字段。

View File

@@ -13,7 +13,7 @@ import type {
} from '@deepseek-ai/dsh-agent'
import { AttachmentError } from '@deepseek-ai/dsh-attachment'
import type { ImageAttachmentRef } from '@deepseek-ai/dsh-attachment'
import { createUserMessage, ReasoningEffortId } from '@deepseek-ai/dsh-llm'
import { contentHasImage, createUserMessage, ReasoningEffortId } from '@deepseek-ai/dsh-llm'
import { errorChain } from '@deepseek-ai/dsh-llm'
import type { ContentBlock, MessageSource } from '@deepseek-ai/dsh-llm'
import type { Session, SessionEvent, SessionHeader, SessionId, UserMessage } from '@deepseek-ai/dsh-session'
@@ -65,11 +65,11 @@ const DEFAULT_MAX_MESSAGES = 50
const MESSAGE_TYPES = new Set(['user/message', 'assistant/message', 'steering/message'])
function decodeBase64(data: string): Uint8Array {
if (data.length === 0 || data.length % 4 !== 0 || !/^(?:[A-Za-z0-9+/]{4})*(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?$/.test(data)) {
throw new AttachmentError('Image upload is not canonical base64.', 'INVALID_IMAGE_BASE64')
}
const decoded = Buffer.from(data, 'base64')
if (decoded.toString('base64') !== data) {
// One canonical-form check: any non-canonical input (whitespace, url-safe
// alphabet, bad padding, truncated groups) fails the exact round-trip, so a
// pre-filter regex over the multi-MiB upload string would be pure overhead.
if (data.length === 0 || decoded.toString('base64') !== data) {
throw new AttachmentError('Image upload is not canonical base64.', 'INVALID_IMAGE_BASE64')
}
return new Uint8Array(decoded)
@@ -147,12 +147,6 @@ function imageInEvent(event: SessionEvent, match: (ref: ImageAttachmentRef) => b
return undefined
}
/** True when typed model content contains an image, including nested tool results. */
function contentHasImage(content: readonly ContentBlock[]): boolean {
return content.some(block => block.type === 'image'
|| (block.type === 'tool-result' && contentHasImage(block.content)))
}
/** True when the current model-visible surface contains an image. */
function messagesHaveImage(messages: readonly { content: readonly ContentBlock[] }[]): boolean {
return messages.some(message => contentHasImage(message.content))
@@ -627,11 +621,12 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
*/
const queuedMirror = new Map<SessionId, InboxItem[]>()
/**
* Claimed-but-unpublished queued occurrences: dequeue is not publication —
* the `user/message` append follows it asynchronously — so an image carrier
* stays a model-selection gate until its durable event lands, its discard
* arrives, or the admission's turn settles idle. Kept apart from the mirror
* so the mux-open snapshot never replays a claimed occurrence as queued.
* Unpublished occurrences that must still gate model selection: a queued
* item from dequeue (claim is not publication — the `user/message` append
* follows asynchronously) and a steering item from enqueue (it never enters
* the queued mirror, and its `steering/message` append is a separate outbox
* hop). Entries retire on their durable event, discard, or idle. Kept apart
* from the mirror so the mux-open snapshot never replays them as queued.
*/
const pendingPublication = new Map<SessionId, InboxItem[]>()
type UnseenQueueEvent =
@@ -692,7 +687,17 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
}
const disposers = [
ctx.on('agent/inbox/enqueue', (agent: Agent, item: InboxItem) => {
if (item.placement !== 'queued') return
if (item.placement === 'steering') {
// A steering carrier never enters the queued mirror, yet its image
// must gate model selection from enqueue until its steering/message
// event publishes (or the admission ends): the outbox hop between
// steer() and the append is asynchronous, and a text-only switch
// accepted inside it would strand every later turn.
const pending = pendingPublication.get(agent.id) ?? []
pending.push(item)
pendingPublication.set(agent.id, pending)
return
}
const unseen = takeUnseen(agent.id, item.id)
if (unseen?.kind === 'terminal') return
let entries = queuedMirror.get(agent.id)
@@ -726,10 +731,16 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
if (retire(agent, item)) publishQueue(agent.id)
}),
ctx.on('session/event', (session: Session, event: SessionEvent) => {
if (event.type !== 'user/message') return
const id = event.type === 'user/message'
? event.data.id
: event.type === 'steering/message'
? (event.data as { message: UserMessage }).message.id
: undefined
if (id === undefined) return
const pending = pendingPublication.get(session.id)
if (pending === undefined) return
const index = pending.findIndex(entry => entry.message.id === (event.data).id)
const placement = event.type === 'user/message' ? 'queued' : 'steering'
const index = pending.findIndex(entry => entry.message.id === id && entry.placement === placement)
if (index === -1) return
pending.splice(index, 1)
if (pending.length === 0) pendingPublication.delete(session.id)
@@ -1385,6 +1396,16 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
updateQueue(request) {
const { sessionId, itemId, action } = request.payload
// Queue edits bypass durablePromptContent (no admission, no durable
// reference, no model-capability recheck), so only text blocks may be
// written through this boundary; image intake is prompt-only.
if (action.kind === 'edit' && action.content.some(block => block.type !== 'text')) {
return Promise.resolve(err(request, {
code: 'attachment-error',
message: 'queue edits accept text content only',
details: { reason: 'QUEUE_EDIT_NON_TEXT' },
}))
}
const agent = ctx.agents.get(sessionId)
if (agent === undefined || agent.updateInbox(itemId, action) === 'not-found') {
return Promise.resolve(err(request, {

View File

@@ -387,6 +387,70 @@ describe('Web session model selection', () => {
await ctx.fiber.dispose()
})
it('gates selection on a steering image from enqueue until its event publishes', async () => {
const { ctx, sessionId, agent } = await harness()
registerTextOnly(ctx)
const api = createApiProxy(ctx, { provider: 'deepseek', model: 'deepseek-chat', cwd: '/tmp', workspaceRoot: '/tmp' })
const steering = {
id: 's-1', role: 'user', source: { kind: 'user' },
content: [{ type: 'image', attachment: { attachmentId: 'att-s', mediaType: 'image/png', bytes: 8, width: 1, height: 1 } }],
} as never
const steeringItem = { id: 'i-s-1', message: steering, placement: 'steering' } as never
// A steering carrier never enters the queued mirror, yet the outbox hop
// between steer() and its append must not open a text-only switch window.
ctx.emit('agent/inbox/enqueue', agent, steeringItem)
expect((await api.sessions.selectModel(request({ sessionId, provider: 'text-only', model: 'plain' }))).result.ok).toBe(false)
ctx.emit('agent/inbox/dequeue', agent, steeringItem)
expect((await api.sessions.selectModel(request({ sessionId, provider: 'text-only', model: 'plain' }))).result.ok).toBe(false)
// Publication hands the gate over to the durable surface.
agent.session.append('steering/message', { turn: 1, message: steering }, { surfaceOp: 'append' })
expect((await api.sessions.selectModel(request({ sessionId, provider: 'text-only', model: 'plain' }))).result.ok).toBe(false)
await ctx.fiber.dispose()
})
it('re-opens selection when an admission ends idle without publication', async () => {
const { ctx, sessionId, agent } = await harness()
registerTextOnly(ctx)
const api = createApiProxy(ctx, { provider: 'deepseek', model: 'deepseek-chat', cwd: '/tmp', workspaceRoot: '/tmp' })
const rejected = {
id: 'r-1', role: 'user', source: { kind: 'user' },
content: [{ type: 'image', attachment: { attachmentId: 'att-r', mediaType: 'image/png', bytes: 8, width: 1, height: 1 } }],
} as never
const rejectedItem = { id: 'i-r-1', message: rejected, placement: 'queued' } as never
ctx.emit('agent/inbox/enqueue', agent, rejectedItem)
ctx.emit('agent/inbox/dequeue', agent, rejectedItem)
expect((await api.sessions.selectModel(request({ sessionId, provider: 'text-only', model: 'plain' }))).result.ok).toBe(false)
// Idle proves the admission ended without publication; nothing durable
// requires an image route, so the text-only switch must be accepted again.
ctx.emit('agent/status', agent, 'idle')
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('rejects a queue edit that injects unadmitted image content', async () => {
const { ctx, sessionId, agent } = await harness()
const api = createApiProxy(ctx, { provider: 'deepseek', model: 'deepseek-chat', cwd: '/tmp', workspaceRoot: '/tmp' })
Object.assign(agent, { updateInbox: () => 'applied' })
const denied = await api.sessions.updateQueue(request({
sessionId,
itemId: 'i-x' as never,
action: {
kind: 'edit' as const,
content: [{ type: 'image', attachment: { attachmentId: 'att-x', mediaType: 'image/png', bytes: 8, width: 1, height: 1 } }] as never,
},
}))
expect(denied.result).toMatchObject({
ok: false,
error: { code: 'attachment-error', details: { reason: 'QUEUE_EDIT_NON_TEXT' } },
})
await ctx.fiber.dispose()
})
it('serializes an image save with a concurrent model selection', async () => {
const { ctx, sessionId, agent } = await harness()
registerTextOnly(ctx)