fix(sdk-client): address ds-review-bot findings

- api: resolve a relative workspace cwd to absolute before the handshake —
  the child spawns relative to the parent cwd, but the wire cwd is resolved
  again inside the child, so a relative value double-resolved
  (worker -> worker/worker).
- api: make the documented handshake retry real — HarnessClient.close() is
  permanent, so a failed initialize now reaps the runtime and swaps in a
  fresh client; DeepSeekHarness.close() is terminal and stops the respawns.
- api: validate session.event envelopes, assistant/message content, and
  session.finished reasons at the wire boundary — a malformed runtime
  surfaces as SdkProtocolError instead of type-invalid TurnResult data or a
  TypeError out of finalResponse.
- client: a throwing subscribe() filter fails and detaches only its own
  subscription (normalized to Error); sibling fan-out and the transport read
  loop are undisturbed.
- client: NotificationSubscription.close() drops its queued notifications,
  matching its documented contract; runtime-death fail() still leaves
  already-delivered items drainable.
- client: subscribe() after close()/runtime death returns a born-failed
  subscription so next() rejects instead of parking forever.
- client/transport: bounded requests abandon via AbortSignal — the transport
  drops the pending entry at timeout, so repeated bounded calls against a
  hung method retain no per-call state.

One test per finding; per-file coverage stays 100% on both packages.

packages/sdk/sdk-client/README.i18n.yaml

@@ -1,6 +1,6 @@
 # Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
 # 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
+#   pnpm run verify-translation-pairing --write packages/sdk/sdk-client/README.md
-README.md: 82e344014ac01120986ee2b4e07ac21192506ff7
+README.md: 3945f911990fa3df362581b0fb37389110bdd386
-README.zh.md: 6e67f57ab92a75be3600cbdd0ce6bf21832fd2c8
+README.zh.md: 3814b88aab1b10c96fdf809565994f29b8e8026b

packages/sdk/sdk-client/README.md

@@ -20,7 +20,7 @@ const result = await harness.run('say hi')
 console.log(result.status, result.finalResponse)
 ```
 
-The subprocess starts lazily on first use and stays owned by the instance across `run()` calls; `close()` (or `await using`) is required so the child is always reaped. `start()` memoizes the `initialize` handshake (cwd + provider/model route); a failed handshake closes the runtime and resets, so a later call may retry. `session(id?)` opens a named or fresh session handle; `run(input, { sessionId?, onNotification? })` sends one prompt turn and settles when the paired `session.finished` arrives, returning a `TurnResult`: `status` (`ok`/`error` as the deployment maps it), the structured `reason` (`TurnEndReason`), `finalResponse` (last assistant message text), plus every `session.event` envelope and raw notification observed for that session tree, in wire order. Model-level failure is a `status: 'error'` result, never a rejection; rejections mean transport loss, timeout, or protocol violation.
+The subprocess starts lazily on first use and stays owned by the instance across `run()` calls; `close()` (or `await using`) is required so the child is always reaped. `start()` memoizes the `initialize` handshake (the workspace cwd — resolved absolute before it crosses the wire — plus the provider/model route); a failed handshake reaps the runtime and swaps in a fresh client, so a later call retries with a new subprocess (until `close()`, which is terminal). `session(id?)` opens a named or fresh session handle; `run(input, { sessionId?, onNotification? })` sends one prompt turn and settles when the paired `session.finished` arrives, returning a `TurnResult`: `status` (`ok`/`error` as the deployment maps it), the structured `reason` (`TurnEndReason`), `finalResponse` (last assistant message text), plus every `session.event` envelope and raw notification observed for that session tree, in wire order. Model-level failure is a `status: 'error'` result, never a rejection; rejections mean transport loss, timeout, or protocol violation.
 
 ## HarnessClient
 

packages/sdk/sdk-client/README.zh.md

@@ -20,7 +20,7 @@ const result = await harness.run('say hi')
 console.log(result.status, result.finalResponse)
 ```
 
-子进程在首次使用时惰性启动，并在多次 `run()` 之间持续归实例所有；必须 `close()`（或 `await using`），子进程才总能被收割。`start()` 记忆化 `initialize` 握手（cwd + provider/model 路由）；握手失败会关闭运行时并复位，后续调用可以重试。`session(id?)` 打开具名或全新的会话句柄；`run(input, { sessionId?, onNotification? })` 发送一个 prompt 回合，在配对的 `session.finished` 到达时尘埃落定，返回 `TurnResult`：`status`（按部署映射的 `ok`/`error`）、结构化 `reason`（`TurnEndReason`）、`finalResponse`（最后一条助手消息文本），以及该会话树内按线序观察到的全部 `session.event` 封套与原始通知。模型层失败是 `status: 'error'` 的结果，绝不是拒绝；拒绝意味着传输丢失、超时或协议违例。
+子进程在首次使用时惰性启动，并在多次 `run()` 之间持续归实例所有；必须 `close()`（或 `await using`），子进程才总能被收割。`start()` 记忆化 `initialize` 握手（工作区 cwd——在跨越线之前解析为绝对路径——加 provider/model 路由）；握手失败会收割运行时并换入全新客户端，后续调用用新子进程重试（直到终结性的 `close()`）。`session(id?)` 打开具名或全新的会话句柄；`run(input, { sessionId?, onNotification? })` 发送一个 prompt 回合，在配对的 `session.finished` 到达时尘埃落定，返回 `TurnResult`：`status`（按部署映射的 `ok`/`error`）、结构化 `reason`（`TurnEndReason`）、`finalResponse`（最后一条助手消息文本），以及该会话树内按线序观察到的全部 `session.event` 封套与原始通知。模型层失败是 `status: 'error'` 的结果，绝不是拒绝；拒绝意味着传输丢失、超时或协议违例。
 
 ## HarnessClient
 

packages/sdk/sdk-client/src/api.ts

@@ -8,9 +8,10 @@
  */
 
 import { randomUUID } from 'node:crypto'
+import { resolve } from 'node:path'
 import type { SessionEvent, TurnEndReason } from '@deepseek-ai/dsh-session'
-import { HarnessClient } from './client.ts'
+import { HarnessClient, isRecord, SdkProtocolError } from './client.ts'
-import type { ContentBlock, DeepSeekHarnessOptions, HarnessNotification, TurnResult } from './types.ts'
+import type { ContentBlock, DeepSeekHarnessOptions, HarnessClientOptions, HarnessNotification, TurnResult } from './types.ts'
 
 /**
  * Reusable SDK for running DeepSeek Harness agent turns in a runtime
@@ -19,33 +20,52 @@ import type { ContentBlock, DeepSeekHarnessOptions, HarnessNotification, TurnRes
  * child is reaped.
  */
 export class DeepSeekHarness implements AsyncDisposable {
-  /** The underlying JSON-RPC client (exposed for low-level access). */
+  private clientInstance: HarnessClient
-  readonly client: HarnessClient
+  private readonly launch: HarnessClientOptions
   private readonly cwd: string
   private readonly provider: string
   private readonly model: string
   private initialized: Promise<void> | undefined
+  private closed = false
 
   /** @param options - runtime launch spec plus the session route (cwd/provider/model). */
   constructor(options: DeepSeekHarnessOptions) {
-    this.client = new HarnessClient(options.launch)
+    this.launch = options.launch
-    this.cwd = options.cwd ?? options.launch.cwd ?? process.cwd()
+    this.clientInstance = new HarnessClient(options.launch)
+    // Absolute before the handshake: the child spawns relative to THIS
+    // process's cwd, but the wire cwd is resolved again inside the child — a
+    // relative value would double-resolve (e.g. `worker` → `worker/worker`).
+    this.cwd = resolve(options.cwd ?? options.launch.cwd ?? process.cwd())
     this.provider = options.provider ?? 'deepseek'
     this.model = options.model ?? 'deepseek-v4-flash'
   }
 
   /**
-   * Start the subprocess and perform the `initialize` handshake once.
+   * The underlying JSON-RPC client (exposed for low-level access). A failed
+   * handshake reaps its runtime and swaps in a fresh instance, so do not
+   * cache this across a failed {@link start}.
+   * @returns the client currently owning the runtime subprocess.
+   */
+  get client(): HarnessClient {
+    return this.clientInstance
+  }
+
+  /**
+   * Start the subprocess and perform the `initialize` handshake once. On
+   * failure the runtime is reaped and a fresh client replaces it
+   * (`HarnessClient.close` is permanent), so a later call retries with a new
+   * subprocess — unless {@link close} already ended this harness.
    * @returns settlement of the (memoized) handshake.
    */
   start(): Promise<void> {
     this.initialized ??= (async () => {
       try {
-        this.client.start()
+        this.clientInstance.start()
-        await this.client.initialize({ cwd: this.cwd, provider: this.provider, model: this.model })
+        await this.clientInstance.initialize({ cwd: this.cwd, provider: this.provider, model: this.model })
       } catch (error) {
         this.initialized = undefined
-        await this.client.close()
+        await this.clientInstance.close()
+        if (!this.closed) this.clientInstance = new HarnessClient(this.launch)
         throw error
       }
     })()
@@ -73,11 +93,13 @@ export class DeepSeekHarness implements AsyncDisposable {
   }
 
   /**
-   * Shut down and reap the runtime subprocess. Idempotent.
+   * Shut down and reap the runtime subprocess. Idempotent and terminal —
+   * a closed harness no longer retries a failed handshake.
    * @returns settlement of the complete teardown.
    */
   close(): Promise<void> {
-    return this.client.close()
+    this.closed = true
+    return this.clientInstance.close()
   }
 
   /**
@@ -128,16 +150,26 @@ export class HarnessSession {
 
     const subscription = client.subscribeSessionTree(this.id)
     const collect = (notification: HarnessNotification): void => {
-      notifications.push(notification)
-      options?.onNotification?.(notification)
       if (notification.method === 'session.event' && notification.params.sessionId === this.id) {
-        events.push(notification.params.event as SessionEvent)
+        // Wire boundary: the envelope feeds the typed TurnResult, so a
+        // malformed runtime surfaces as a protocol error, not as type-invalid
+        // data (or a TypeError out of finalResponse).
+        const event = validatedSessionEvent(notification.params.event)
+        notifications.push(notification)
+        options?.onNotification?.(notification)
+        events.push(event)
+        return
       }
       if (notification.method === 'session.finished' && notification.params.sessionId === this.id) {
+        reason = validatedTurnEndReason(notification.params.reason)
+        notifications.push(notification)
+        options?.onNotification?.(notification)
         status = notification.params.status === 'ok' ? 'ok' : 'error'
-        reason = notification.params.reason as TurnEndReason | undefined
         finished = true
+        return
       }
+      notifications.push(notification)
+      options?.onNotification?.(notification)
     }
     const accepted = client.prompt(this.id, contentBlocks)
     // Drain concurrently so observers see progress while the prompt request
@@ -175,6 +207,32 @@ export function normalizeInput(input: string | ContentBlock[]): ContentBlock[] {
   return typeof input === 'string' ? [{ type: 'text', text: input }] : input
 }
 
+/** Validate a wire `session.event` envelope to the shape the typed result exposes. */
+function validatedSessionEvent(value: unknown): SessionEvent {
+  if (!isRecord(value) || typeof value.type !== 'string') {
+    throw new SdkProtocolError(`session.event carried no event envelope: ${JSON.stringify(value)}`)
+  }
+  // The one variant this module reads into (finalResponse) must carry
+  // kind-tagged content blocks; other variants pass through under their
+  // envelope shape.
+  if (value.type === 'assistant/message') {
+    const content = isRecord(value.data) ? value.data.content : undefined
+    if (!Array.isArray(content) || !content.every(block => isRecord(block) && typeof block.type === 'string')) {
+      throw new SdkProtocolError(`assistant/message event carried malformed content: ${JSON.stringify(value)}`)
+    }
+  }
+  return value as unknown as SessionEvent
+}
+
+/** Validate a wire `session.finished` reason (absent, or a kind-tagged record). */
+function validatedTurnEndReason(value: unknown): TurnEndReason | undefined {
+  if (value === undefined) return undefined
+  if (!isRecord(value) || typeof value.kind !== 'string') {
+    throw new SdkProtocolError(`session.finished carried a malformed reason: ${JSON.stringify(value)}`)
+  }
+  return value as unknown as TurnEndReason
+}
+
 /**
  * Extract the concatenated text of the last assistant message.
  * @param events - the turn's `session.event` payloads in wire order.

packages/sdk/sdk-client/src/client.ts

@@ -81,8 +81,9 @@ export class NotificationSubscription implements AsyncIterable<HarnessNotificati
 
   /**
    * Await the next matching notification.
-   * @returns the notification; rejects once the runtime is closed or the
+   * @returns the notification; after the runtime died, drains what was
-   * subscription itself is closed.
+   * already delivered and then rejects; after {@link close}, rejects
+   * immediately (the queue is dropped).
    */
   next(): Promise<HarnessNotification> {
     const queued = this.state.queue.shift()
@@ -104,11 +105,15 @@ export class NotificationSubscription implements AsyncIterable<HarnessNotificati
   /** Detach from the client; queued items drop and pending waiters reject. */
   close(): void {
     this.unsubscribe()
+    // The drop is part of this method's contract; a runtime-death fail() keeps
+    // the queue so already-delivered notifications remain drainable.
+    this.state.queue.length = 0
     this.fail(new TransportClosedError('notification subscription closed'))
   }
 
   /**
    * Reject pending and future waits (delivery stops; the first failure wins).
+   * Already-queued notifications remain drainable via {@link next}/{@link tryNext}.
    * @param error - the terminal failure delivered to waiters.
    */
   fail(error: Error): void {
@@ -117,11 +122,22 @@ export class NotificationSubscription implements AsyncIterable<HarnessNotificati
   }
 
   /**
-   * Deliver one notification to a waiter or the queue when the filter matches.
+   * Deliver one notification to a waiter or the queue when the filter
+   * matches. A throwing filter fails only THIS subscription (detached, the
+   * throw becomes its terminal error) — it never disturbs sibling
+   * subscriptions or the transport's read loop, mirroring the Python client.
    * @param notification - the wire notification to deliver.
    */
   push(notification: HarnessNotification): void {
-    if (this.state.filter !== undefined && !this.state.filter(notification)) return
+    let matches: boolean
+    try {
+      matches = this.state.filter === undefined || this.state.filter(notification)
+    } catch (error) {
+      this.unsubscribe()
+      this.fail(error instanceof Error ? error : new Error(String(error)))
+      return
+    }
+    if (!matches) return
     const waiter = this.state.waiters.shift()
     if (waiter !== undefined) waiter.resolve(notification)
     else this.state.queue.push(notification)
@@ -273,22 +289,18 @@ export class HarnessClient {
     const transport = this.transport
     /* v8 ignore next -- start() either sets the transport or throws */
     if (transport === undefined) throw new TransportClosedError('DeepSeek Harness runtime is not running')
-    const pending = transport.request(method, params ?? {})
     const timeout = timeoutMs ?? this.options.requestTimeoutMs
     try {
-      if (timeout === undefined) return await pending
+      if (timeout === undefined) return await transport.request(method, params ?? {})
-      let timer: NodeJS.Timeout | undefined
+      // The abort signal makes the timeout an abandonment: the transport drops
+      // its pending entry, so repeated bounded requests against a hung method
+      // retain no per-call state (the server-side work still runs to close).
+      const abandon = new AbortController()
+      const timer = setTimeout(() => {
+        abandon.abort(new RequestTimeoutError(`${method} timed out after ${timeout}ms waiting for the DeepSeek Harness runtime`))
+      }, timeout)
       try {
-        return await Promise.race([
+        return await transport.request(method, params ?? {}, abandon.signal)
-          pending,
-          new Promise<never>((_, reject) => {
-            timer = setTimeout(() => {
-              // The abandoned wire promise settles on close; keep it handled.
-              pending.catch(() => {})
-              reject(new RequestTimeoutError(`${method} timed out after ${timeout}ms waiting for the DeepSeek Harness runtime`))
-            }, timeout)
-          }),
-        ])
       } finally {
         clearTimeout(timer)
       }
@@ -303,12 +315,18 @@ export class HarnessClient {
   /**
    * Subscribe to server notifications.
    * @param filter - optional predicate; omitted means every notification.
-   * @returns the subscription handle; close it to stop delivery.
+   * @returns the subscription handle; close it to stop delivery. After
+   * {@link close} or runtime death the handle is born failed — there is no
+   * producer left, so `next()` rejects instead of waiting forever.
    */
   subscribe(filter?: NotificationFilter): NotificationSubscription {
     const id = String(this.subscriptionSerial++)
     const state: SubscriptionState = { queue: [], waiters: [], filter, failure: undefined }
     const subscription = new NotificationSubscription(state, () => { this.subscriptions.delete(id) })
+    if (this.closeTask !== undefined || this.exitCode !== undefined || this.spawnError !== undefined) {
+      subscription.fail(this.closedError('DeepSeek Harness runtime closed'))
+      return subscription
+    }
     this.subscriptions.set(id, subscription)
     return subscription
   }

packages/sdk/sdk-client/tests/fake-runtime.ts

@@ -16,6 +16,16 @@
  * - `FAKE_MALFORMED`: `initialize` returns `{}` (no serverInfo); `prompt` returns `{}` (no accepted).
  * - `FAKE_MALFORMED_PROMPT`: `initialize` is normal; only `prompt` returns `{}` (no accepted).
  * - `FAKE_INIT_ERROR`: `initialize` answers a JSON-RPC error response with code 7.
+ * - `FAKE_INIT_ERROR_ONCE_FILE`: fail `initialize` (code 7) only when this
+ *   marker file does NOT exist yet, creating it — so the first runtime
+ *   process fails the handshake and a respawned one succeeds (retry probe).
+ * - `FAKE_ECHO_CWD_IN_INIT`: reply `serverInfo.version` = this process's cwd
+ *   (wire-visible spawn-cwd probe).
+ * - `FAKE_MALFORMED_EVENT`: the turn's `session.event` carries a number as
+ *   the event; `FAKE_MALFORMED_MESSAGE`: assistant/message content is not an
+ *   array; `FAKE_MESSAGE_WITHOUT_DATA`: assistant/message with no data
+ *   member; `FAKE_MALFORMED_REASON`: `session.finished` reason is a bare
+ *   string (wire-validation probes).
  * - `FAKE_HANG_INIT`: never answer `initialize` (mid-handshake cancel probe).
  * - `FAKE_INIT_READY` + `FAKE_INIT_GO`: touch the READY file when `initialize`
  *   arrives, then poll for the GO file before answering (deterministic
@@ -78,8 +88,20 @@ function assistantText(): string {
 
 function runTurn(sessionId: string): void {
   const text = assistantText()
+  if (env.FAKE_MALFORMED_EVENT !== undefined) {
+    notify('session.event', { sessionId, event: 42 })
+    return
+  }
   event(sessionId, 'turn/start', { turn: 0 })
   event(sessionId, 'assistant/chunk', { turn: 0, step: 0, chunk: { type: 'text-delta', index: 0, text } })
+  if (env.FAKE_MALFORMED_MESSAGE !== undefined) {
+    event(sessionId, 'assistant/message', { turn: 0, step: 0, content: 'not-an-array' })
+    return
+  }
+  if (env.FAKE_MESSAGE_WITHOUT_DATA !== undefined) {
+    notify('session.event', { sessionId, event: { type: 'assistant/message', seq: seq++, time: 0 } })
+    return
+  }
   event(sessionId, 'assistant/message', {
     turn: 0,
     step: 0,
@@ -110,7 +132,9 @@ function runTurn(sessionId: string): void {
   notify('session.finished', {
     sessionId,
     status: env.FAKE_STATUS ?? 'ok',
-    ...(reasonKind === 'none' ? {} : { reason: { kind: reasonKind } }),
+    ...(env.FAKE_MALFORMED_REASON !== undefined
+      ? { reason: 'not-a-record' }
+      : reasonKind === 'none' ? {} : { reason: { kind: reasonKind } }),
   })
 }
 
@@ -144,10 +168,19 @@ reader.on('line', (line) => {
         write({ jsonrpc: '2.0', id: frame.id, error: { code: 7, message: 'scripted init failure', data: { hint: 'fake' } } })
         return
       }
+      if (env.FAKE_INIT_ERROR_ONCE_FILE !== undefined && !existsSync(env.FAKE_INIT_ERROR_ONCE_FILE)) {
+        writeFileSync(env.FAKE_INIT_ERROR_ONCE_FILE, 'failed-once\n')
+        write({ jsonrpc: '2.0', id: frame.id, error: { code: 7, message: 'scripted first-boot failure' } })
+        return
+      }
       if (env.FAKE_MALFORMED !== undefined) {
         respond({})
         return
       }
+      if (env.FAKE_ECHO_CWD_IN_INIT !== undefined) {
+        respond({ serverInfo: { name: 'deepseek-harness-sdk-runtime', version: process.cwd() } })
+        return
+      }
       respond({ serverInfo: { name: 'deepseek-harness-sdk-runtime', version: '0.0.1' } })
       return
     case 'session/prompt': {

packages/sdk/sdk-client/tests/sdk-client.spec.ts

@@ -5,9 +5,9 @@
  * and session-tree scoping, error surfaces, timeouts, and the dispose ladder.
  */
 
-import { mkdtemp, readFile, rm, stat } from 'node:fs/promises'
+import { mkdir, mkdtemp, readFile, realpath, rm, stat } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
-import { join } from 'node:path'
+import { isAbsolute, join, relative, resolve as resolvePath } from 'node:path'
 import { fileURLToPath } from 'node:url'
 import { afterEach, describe, expect, it } from 'vitest'
 import {
@@ -122,6 +122,31 @@ describe('DeepSeekHarness', () => {
     expect(records).toEqual([{ cwd: dir, provider: 'custom-provider', model: 'custom-model' }])
   })
 
+  it('resolves a relative launch cwd to an absolute workspace before the handshake', async () => {
+    // vitest workers forbid chdir, so derive a RELATIVE path from the real
+    // process cwd to a temp worker dir; resolution is lexical either way.
+    const dir = await tempDir('sdk-client-relcwd-')
+    const recordFile = join(dir, 'init.jsonl')
+    const inner = join(dir, 'worker')
+    await mkdir(inner)
+    const relativeCwd = relative(process.cwd(), inner)
+    expect(isAbsolute(relativeCwd)).toBe(false)
+    const harness = new DeepSeekHarness({
+      launch: fakeLaunch({ FAKE_RECORD_INIT: recordFile, FAKE_ECHO_CWD_IN_INIT: '1' }, { cwd: relativeCwd }),
+    })
+    cleanups.push(() => harness.close())
+    await harness.start()
+    const identity = await harness.client.initialize({ cwd: inner, provider: 'p', model: 'm' })
+    await harness.close()
+    // The child spawned under the temp worker dir (its physical cwd)...
+    expect(identity.serverInfo.version).toBe(await realpath(inner))
+    // ...and the handshake wire cwd went out ABSOLUTE, so the child cannot
+    // re-resolve a relative string into dir/worker/worker.
+    const records = (await readFile(recordFile, 'utf8')).trim().split('\n')
+      .map(line => (JSON.parse(line) as { cwd: string }).cwd)
+    expect(records).toEqual([resolvePath(relativeCwd), inner])
+  })
+
   it('propagates a JSON-RPC error response from initialize and closes the runtime', async () => {
     const harness = harnessWith({ FAKE_INIT_ERROR: '1' })
     const failure = await harness.run('boom').then(
@@ -134,6 +159,23 @@ describe('DeepSeekHarness', () => {
     await expect(harness.run('later')).rejects.toThrow()
   })
 
+  it('retries a failed handshake with a fresh runtime process', async () => {
+    const dir = await tempDir('sdk-client-retry-')
+    const marker = join(dir, 'first-boot-failed')
+    const harness = harnessWith({ FAKE_INIT_ERROR_ONCE_FILE: marker, FAKE_TEXT: 'second boot answer' })
+    const firstClient = harness.client
+    // First start: the scripted runtime fails the handshake and is reaped.
+    await expect(harness.start()).rejects.toThrow('scripted first-boot failure')
+    // Retry spawns a NEW subprocess through a fresh client (close is permanent).
+    const result = await harness.run('again')
+    expect(harness.client).not.toBe(firstClient)
+    expect(result.status).toBe('ok')
+    expect(result.finalResponse).toBe('second boot answer')
+    await harness.close()
+    // close() is terminal: a handshake failure after it must not respawn.
+    await expect(harness.run('after-close')).rejects.toThrow(TransportClosedError)
+  })
+
   it('rejects a malformed initialize result as a protocol error', async () => {
     const harness = harnessWith({ FAKE_MALFORMED: '1' })
     await expect(harness.run('bad')).rejects.toThrow(SdkProtocolError)
@@ -162,6 +204,22 @@ describe('HarnessClient', () => {
     await client.close()
   })
 
+  it('a timed-out request leaves no pending transport state', async () => {
+    const client = new HarnessClient(fakeLaunch({ FAKE_HANG_PROMPT: '1' }))
+    cleanups.push(() => client.close())
+    await client.initialize({ cwd: process.cwd(), provider: 'p', model: 'm' })
+    for (let round = 0; round < 3; round++) {
+      await expect(client.request('session/prompt', { sessionId: 's', contentBlocks: normalizeInput('x') }, 50))
+        .rejects.toThrow(RequestTimeoutError)
+    }
+    // Abandonment removed each pending entry at its timeout; a hung method
+    // retains nothing per call. (Private map read is the observable here —
+    // no wire surface reports transport bookkeeping.)
+    const transport = (client as unknown as { transport: { pending: Map<string, unknown> } }).transport
+    expect(transport.pending.size).toBe(0)
+    await client.close()
+  })
+
   it('applies the client-wide request timeout when no per-call bound is given', async () => {
     const client = new HarnessClient(fakeLaunch({ FAKE_HANG_PROMPT: '1' }, { requestTimeoutMs: 400 }))
     cleanups.push(() => client.close())
@@ -255,6 +313,10 @@ describe('HarnessClient', () => {
     expect(finished.method).toBe('session.finished')
     expect(finishedOnly.tryNext()).toBeUndefined()
 
+    // A bare unbounded request with omitted params sends `{}` on the wire.
+    const identity = await client.request('initialize') as { serverInfo: { name: string } }
+    expect(identity.serverInfo.name).toBe('deepseek-harness-sdk-runtime')
+
     // Async iteration consumes queued items and then parks.
     const collected: string[] = []
     for await (const notification of all) {
@@ -269,6 +331,56 @@ describe('HarnessClient', () => {
     await client.close()
   })
 
+  it('contains a throwing filter to its own subscription', async () => {
+    const client = new HarnessClient(fakeLaunch())
+    cleanups.push(() => client.close())
+    await client.initialize({ cwd: process.cwd(), provider: 'p', model: 'm' })
+
+    const broken = client.subscribe(() => { throw new Error('filter exploded') })
+    // A non-Error throw is normalized rather than crashing dispatch.
+    const brokenNonError = client.subscribe(() => { throw 'string boom' })
+    const healthy = client.subscribe(n => n.method === 'session.finished')
+    await client.prompt('filter-contain', normalizeInput('go'))
+
+    // The sibling subscription and the read loop are undisturbed.
+    expect((await healthy.next()).method).toBe('session.finished')
+    // Each broken subscription failed with ITS OWN error and detached.
+    await expect(broken.next()).rejects.toThrow('filter exploded')
+    await expect(brokenNonError.next()).rejects.toThrow('string boom')
+    healthy.close()
+    await client.close()
+  })
+
+  it('close() drops queued notifications; runtime death keeps them drainable', async () => {
+    const client = new HarnessClient(fakeLaunch())
+    await client.initialize({ cwd: process.cwd(), provider: 'p', model: 'm' })
+    const closed = client.subscribe()
+    const drainable = client.subscribe()
+    await client.prompt('queue-drop', normalizeInput('go'))
+    expect(closed.tryNext()).toBeDefined()
+    closed.close()
+    // Manual close drops the rest of the queue outright.
+    expect(closed.tryNext()).toBeUndefined()
+    await expect(closed.next()).rejects.toThrow('notification subscription closed')
+    // Runtime teardown, by contrast, only stops FUTURE delivery: what was
+    // already delivered before close() stays drainable.
+    await client.close()
+    expect(drainable.tryNext()).toBeDefined()
+  })
+
+  it('subscriptions created after termination are born failed', async () => {
+    const client = new HarnessClient(fakeLaunch())
+    await client.initialize({ cwd: process.cwd(), provider: 'p', model: 'm' })
+    await client.close()
+    // No producer can ever feed this subscription; next() must not park forever.
+    await expect(client.subscribe().next()).rejects.toThrow(TransportClosedError)
+
+    const dead = new HarnessClient(fakeLaunch({ FAKE_EXIT_BEFORE_INIT: '1' }))
+    cleanups.push(() => dead.close())
+    await dead.initialize({ cwd: process.cwd(), provider: 'p', model: 'm' }).catch(() => {})
+    await expect(dead.subscribe().next()).rejects.toThrow(TransportClosedError)
+  })
+
   it('closes subscriptions with the runtime and rejects parked waiters', async () => {
     const client = new HarnessClient(fakeLaunch())
     await client.initialize({ cwd: process.cwd(), provider: 'p', model: 'm' })
@@ -310,6 +422,28 @@ describe('HarnessClient', () => {
   })
 })
 
+describe('wire payload validation', () => {
+  it('rejects a non-object session.event envelope as a protocol error', async () => {
+    const harness = harnessWith({ FAKE_MALFORMED_EVENT: '1' })
+    await expect(harness.run('bad-event')).rejects.toThrow(SdkProtocolError)
+  })
+
+  it('rejects an assistant/message without a content array as a protocol error', async () => {
+    const harness = harnessWith({ FAKE_MALFORMED_MESSAGE: '1' })
+    await expect(harness.run('bad-message')).rejects.toThrow(SdkProtocolError)
+  })
+
+  it('rejects an assistant/message without a data member as a protocol error', async () => {
+    const harness = harnessWith({ FAKE_MESSAGE_WITHOUT_DATA: '1' })
+    await expect(harness.run('no-data')).rejects.toThrow(SdkProtocolError)
+  })
+
+  it('rejects a malformed session.finished reason as a protocol error', async () => {
+    const harness = harnessWith({ FAKE_MALFORMED_REASON: '1' })
+    await expect(harness.run('bad-reason')).rejects.toThrow(SdkProtocolError)
+  })
+})
+
 describe('stderr tail bound', () => {
   it('keeps only the newest lines up to the limit', async () => {
     const manyLines = Array.from({ length: 450 }, (_, i) => `line-${i}`).join('\n')

packages/sdk/sdk-protocol/src/transport.ts

@@ -109,15 +109,47 @@ export class JsonRpcLineTransport implements JsonRpcTransportPeer {
     this.notificationHandler = handler
   }
 
-  request(method: string, params: object): Promise<unknown> {
+  /**
+   * Send a request and await its response.
+   * @param method - the JSON-RPC method name.
+   * @param params - the request parameters object.
+   * @param signal - optional abandonment signal: aborting removes the pending
+   * entry (no state is retained for a response that may never come) and
+   * rejects with the signal's reason.
+   * @returns the result; rejects per {@link JsonRpcTransportPeer.request}.
+   */
+  request(method: string, params: object, signal?: AbortSignal): Promise<unknown> {
     const id = `req_${randomUUID().replaceAll('-', '')}`
     const message = { jsonrpc: '2.0', id, method, params }
     return new Promise((resolve, reject) => {
-      this.pending.set(id, { resolve, reject })
+      let detach = (): void => {}
+      if (signal !== undefined) {
+        if (signal.aborted) {
+          reject(abortError(signal.reason))
+          return
+        }
+        const onAbort = (): void => {
+          this.pending.delete(id)
+          reject(abortError(signal.reason))
+        }
+        signal.addEventListener('abort', onAbort, { once: true })
+        detach = () => { signal.removeEventListener('abort', onAbort) }
+      }
+      this.pending.set(id, {
+        resolve: (value) => {
+          detach()
+          resolve(value)
+        },
+        reject: (error) => {
+          detach()
+          reject(error)
+        },
+      })
       try {
         this.write(message)
       } catch (error) {
         this.pending.delete(id)
+        detach()
         reject(error instanceof Error ? error : new Error(String(error)))
       }
     })
@@ -240,3 +272,8 @@ export class JsonRpcLineTransport implements JsonRpcTransportPeer {
 function objectParams(params: unknown): Record<string, unknown> {
   return params && typeof params === 'object' && !Array.isArray(params) ? params as Record<string, unknown> : {}
 }
+
+/** Normalize an abort reason into the rejection Error (a non-Error reason is stringified). */
+function abortError(reason: unknown): Error {
+  return reason instanceof Error ? reason : new Error(`JSON-RPC request aborted: ${String(reason)}`)
+}

packages/sdk/sdk-protocol/tests/transport.spec.ts

@@ -60,6 +60,29 @@ describe('JsonRpcLineTransport', () => {
     b.close()
   })
 
+  it('rejects immediately on a pre-aborted signal without registering pending state', async () => {
+    const { b } = transportPair()
+    b.start()
+    const controller = new AbortController()
+    controller.abort(new Error('already gone'))
+    await expect(b.request('never-sent', {}, controller.signal)).rejects.toThrow('already gone')
+    expect((b as unknown as { pending: Map<string, unknown> }).pending.size).toBe(0)
+    b.close()
+  })
+
+  it('abandons a pending request on abort, stringifying a non-Error reason', async () => {
+    const { b } = transportPair()
+    b.start()
+    const controller = new AbortController()
+    const pending = b.request('never-answered', {}, controller.signal)
+    controller.abort('plain-string-reason')
+    await expect(pending).rejects.toThrow('JSON-RPC request aborted: plain-string-reason')
+    // The abandonment removed the pending entry — nothing is retained for a
+    // response that may never come.
+    expect((b as unknown as { pending: Map<string, unknown> }).pending.size).toBe(0)
+    b.close()
+  })
+
   it('preserves structured error data from an error response frame', async () => {
     const { aToB, bToA, b } = transportPair()
     b.start()