feat(dsh): remote managed-agent as on-demand tool + bl managed-agent run

Rework the managed-agent integration so a dsh user can, in plain
language, have a Bailian cloud agent created and run a task — no
hand-written agents.yaml, no prior apply.

New `bl managed-agent run --prompt <task> [--instructions] [--model]
[--agent]`: one step that idempotently materializes a cloud agent + its
environment, then opens a session and streams the result. It mirrors the
OpenAgentPack webui backend's ensure+run recipe (resolveProjectConfigFrom
Object → syncAgentResourcesWithStateBackend → readProjectRuntime +
startSessionRun) from an in-memory config, reusing the existing
credential spine in _engine/credentials.ts. State persists under the bl
config dir (~/.bailian/managed-agent/<agent>/), never the user's cwd, so
repeat runs with the same --agent reuse the materialized agent. Unlike
apply it provisions without --yes, since running is the intent.

dsh side: replace the SubagentProvider with a plain tool
`bailian_run_remote_task` (packages/dsh/src/tool-managed-agent). The
subagent seam did not fit: in the web profile every tool-subagent row is
disabled in the host plane (delegation lives in agent presets), a
provider fixes one agent identity in config, and the default numeric
maxDepth would fail-mount a no-depthLimit provider. As a tool the model
calls it directly and fills `instructions` from the user's intent, so the
remote agent's role is defined per task. Enabled by default — it creates
nothing at load, only on invocation.

LLM row: configure the base bundle's existing llm-pi-ai row instead of
mounting a second pi-ai instance (a second instance re-declares pi-ai's
global configurable-provider catalog and fails boot on a duplicate
amazon-bedrock). TokenPlan reads a dedicated BAILIAN_TOKENPLAN_API_KEY,
not DASHSCOPE_API_KEY: TokenPlan (sk-sp-) and pay-as-you-go (sk-ws-) keys
401 each other's endpoints, so sharing one var would silently break
whichever plugin lost.

Note: the ensure+run happy path could not be verified end-to-end on the
available account — agentstudio returns 404 there, and the existing
`managed-agent apply` 404s identically against the same endpoint/key, so
the failure is account/service provisioning, not this change. Command
wiring, dry-run, config assembly, credential injection and URL
construction were all verified.

Co-Authored-By: Claude <noreply@anthropic.com>

packages/cli/src/commands.ts

@@ -102,6 +102,7 @@ import {
   managedAgentValidate,
   managedAgentPlan,
   managedAgentApply,
+  managedAgentRun,
   managedAgentDestroy,
   managedAgentStateList,
   managedAgentStateShow,
@@ -225,6 +226,7 @@ export const commands: Record<string, AnyCommand> = {
   "managed-agent validate": managedAgentValidate,
   "managed-agent plan": managedAgentPlan,
   "managed-agent apply": managedAgentApply,
+  "managed-agent run": managedAgentRun,
   "managed-agent destroy": managedAgentDestroy,
   "managed-agent state list": managedAgentStateList,
   "managed-agent state show": managedAgentStateShow,

packages/commands/src/commands/managed-agent/_engine/inline-runtime.ts

@@ -0,0 +1,122 @@
+import { mkdirSync } from "node:fs";
+import { dirname, join } from "node:path";
+import {
+  type BackendRuntimeInput,
+  LocalFileStateBackend,
+  resolveProjectConfigFromObject,
+} from "@openagentpack/sdk";
+import { getConfigDir } from "bailian-cli-core";
+import {
+  assertProviderCredentials,
+  type CredentialHost,
+  injectProviderCredentials,
+  normalizeInterpolatedProviderBlocks,
+  prepareProviderEnv,
+  scrubCredentialEnv,
+} from "./credentials.ts";
+import { type HostContext, installSdkTransport } from "./transport.ts";
+
+/** Default agent identity `bl managed-agent run` materializes and reuses. */
+export const DEFAULT_INLINE_AGENT = "dsh-remote-runner";
+
+/** Default model for the materialized agent. */
+export const DEFAULT_INLINE_MODEL = "qwen3.8-max";
+
+/** Default role when the caller supplies no `--instructions`. */
+export const DEFAULT_INLINE_INSTRUCTIONS = "You are a helpful assistant. Complete the task.";
+
+/** Environment name declared in the inline config; one cloud env per agent. */
+const INLINE_ENVIRONMENT = "cloud";
+
+export interface InlineAgentOptions {
+  agentName: string;
+  instructions: string;
+  model: string;
+  /** Override the persisted state location (defaults under the bl config dir). */
+  statePath?: string;
+}
+
+/**
+ * Slugify an agent name into a filesystem- and project-id-safe token. The state
+ * for each distinct agent lives in its own directory so repeat runs reuse the
+ * same materialized remote agent.
+ */
+function slugify(agentName: string): string {
+  const slug = agentName
+    .toLowerCase()
+    .replace(/[^a-z0-9._-]+/g, "-")
+    .replace(/^-+|-+$/g, "");
+  return slug.length > 0 ? slug : "agent";
+}
+
+/** Where a materialized agent's state is persisted (not the user's cwd). */
+export function inlineStatePath(agentName: string): string {
+  return join(getConfigDir(), "managed-agent", slugify(agentName), "state.json");
+}
+
+/**
+ * The minimal in-memory project config that materializes into one cloud agent.
+ * `providers.bailian` carries empty `api_key`/`base_url` placeholders so
+ * {@link injectProviderCredentials} fills them from bl's auth chain (it only
+ * writes fields the block already declares).
+ */
+export function buildInlineConfig(opts: InlineAgentOptions): Record<string, unknown> {
+  return {
+    version: "1",
+    providers: {
+      bailian: { api_key: "", base_url: "" },
+    },
+    defaults: { provider: "bailian" },
+    environments: {
+      [INLINE_ENVIRONMENT]: {
+        description: "Bailian CLI cloud environment",
+        config: { type: "cloud", networking: { type: "unrestricted" } },
+      },
+    },
+    agents: {
+      [opts.agentName]: {
+        description: opts.agentName,
+        model: opts.model,
+        instructions: opts.instructions,
+        environment: INLINE_ENVIRONMENT,
+        provider: "bailian",
+      },
+    },
+  };
+}
+
+/**
+ * Build the `BackendRuntimeInput` shared by ensure (`syncAgentResourcesWith
+ * StateBackend`) and run (`readProjectRuntime` + `startSessionRun`). Mirrors the
+ * credential spine of {@link buildAgentRuntime} but sources config from an
+ * in-memory object instead of a file, so no `agents.yaml` or `apply` is required.
+ */
+export async function buildInlineBackendInput(
+  host: HostContext & CredentialHost,
+  opts: InlineAgentOptions,
+): Promise<BackendRuntimeInput> {
+  installSdkTransport(host);
+  prepareProviderEnv();
+
+  const rawConfig = buildInlineConfig(opts);
+  const { config, projectName } = await resolveProjectConfigFromObject(rawConfig, {
+    projectName: slugify(opts.agentName),
+  });
+
+  normalizeInterpolatedProviderBlocks(config.providers);
+  injectProviderCredentials(config.providers, host);
+  scrubCredentialEnv();
+  assertProviderCredentials(config.providers);
+
+  const statePath = opts.statePath ?? inlineStatePath(opts.agentName);
+  mkdirSync(dirname(statePath), { recursive: true });
+  const stateBackend = new LocalFileStateBackend({ statePath });
+
+  return {
+    projectName,
+    config,
+    stateBackend,
+    stateScope: { projectId: slugify(opts.agentName) },
+    providers: config.providers,
+  };
+}

packages/commands/src/commands/managed-agent/run.ts

@@ -0,0 +1,137 @@
+import {
+  BailianError,
+  defineCommand,
+  detectOutputFormat,
+  ExitCode,
+  type FlagsDef,
+} from "bailian-cli-core";
+import { emitResult } from "bailian-cli-runtime";
+import {
+  readProjectRuntime,
+  startSessionRun,
+  startSessionRunPolling,
+  syncAgentResourcesWithStateBackend,
+} from "@openagentpack/sdk";
+import { CREDENTIALS_NOTE } from "./_engine/config-loader.ts";
+import { withStdoutProtected } from "./_engine/console-capture.ts";
+import { withAgentErrors } from "./_engine/errors.ts";
+import {
+  buildInlineBackendInput,
+  DEFAULT_INLINE_AGENT,
+  DEFAULT_INLINE_INSTRUCTIONS,
+  DEFAULT_INLINE_MODEL,
+} from "./_engine/inline-runtime.ts";
+import { renderCollectedEvents, streamAndRenderEvents } from "./_engine/session-render.ts";
+
+const RUN_FLAGS = {
+  prompt: {
+    type: "string",
+    valueHint: "<text>",
+    description: "Task to run (required)",
+    required: true,
+  },
+  instructions: {
+    type: "string",
+    valueHint: "<text>",
+    description: "Role/system instructions for the remote agent (default: generic assistant)",
+  },
+  model: {
+    type: "string",
+    valueHint: "<id>",
+    description: `Model for the remote agent (default: ${DEFAULT_INLINE_MODEL})`,
+  },
+  agent: {
+    type: "string",
+    valueHint: "<name>",
+    description: `Agent identity to create/reuse (default: ${DEFAULT_INLINE_AGENT})`,
+  },
+  noStream: {
+    type: "switch",
+    description: "Use polling instead of SSE streaming",
+  },
+} satisfies FlagsDef;
+
+export default defineCommand({
+  description: "Provision (if needed) a cloud agent and run a task in one step",
+  auth: "apiKey",
+  usageArgs: "--prompt <text> [--instructions <text>] [--model <id>] [--agent <name>]",
+  flags: RUN_FLAGS,
+  exampleArgs: [
+    '--prompt "Summarize the latest AI news"',
+    '--prompt "Audit this dependency tree" --instructions "You are a security expert" --model qwen3.8-max',
+  ],
+  notes: [
+    ...CREDENTIALS_NOTE,
+    "Unlike `apply`, this creates/updates the cloud agent + environment on demand without --yes. The first run provisions cloud resources (may incur cost and take longer to start); later runs with the same --agent reuse them.",
+  ],
+  async run(ctx) {
+    const { settings, flags } = ctx;
+    const format = detectOutputFormat(settings.output);
+    const asJson = format === "json";
+
+    const agentName = flags.agent ?? DEFAULT_INLINE_AGENT;
+    const model = flags.model ?? DEFAULT_INLINE_MODEL;
+    const instructions = flags.instructions ?? DEFAULT_INLINE_INSTRUCTIONS;
+
+    if (settings.dryRun) {
+      emitResult(
+        {
+          would_run: {
+            prompt: flags.prompt,
+            agent: agentName,
+            model,
+            instructions,
+            mode: flags.noStream ? "polling" : "streaming",
+          },
+        },
+        format,
+      );
+      return;
+    }
+
+    await withAgentErrors(() =>
+      withStdoutProtected(async () => {
+        const input = await buildInlineBackendInput(ctx, { agentName, instructions, model });
+
+        // Ensure the remote agent + its cloud environment exist. Idempotent:
+        // a repeat run with the same agent name reuses the materialized state.
+        if (!asJson) process.stderr.write(`Ensuring cloud agent "${agentName}"…\n`);
+        const sync = await syncAgentResourcesWithStateBackend(input, agentName, {
+          policy: "force",
+          quiet: true,
+        });
+        if (sync.status !== "completed") {
+          const detail =
+            sync.error ??
+            sync.diagnostics.find((diag) => diag.severity === "error")?.message ??
+            `provisioning ended with status "${sync.status}"`;
+          throw new BailianError(
+            `Failed to provision cloud agent "${agentName}": ${detail}`,
+            ExitCode.GENERAL,
+          );
+        }
+
+        // Run the task inside a runtime bound to the just-materialized state.
+        await readProjectRuntime(input, async (runtime) => {
+          if (flags.noStream) {
+            const run = await startSessionRunPolling(runtime, flags.prompt, { agent: agentName });
+            if (!asJson) process.stderr.write(`Session created: ${run.session.id}\n`);
+            renderCollectedEvents(run, asJson, {
+              session_id: run.session.id,
+              provider: run.provider,
+              agent: run.agentName,
+            });
+          } else {
+            const run = await startSessionRun(runtime, flags.prompt, { agent: agentName });
+            if (!asJson) process.stderr.write(`Session created: ${run.session.id}\n`);
+            await streamAndRenderEvents(run.events, asJson, {
+              session_id: run.session.id,
+              provider: run.provider,
+              agent: run.agentName,
+            });
+          }
+        });
+      }),
+    );
+  },
+});

packages/commands/src/index.ts

@@ -99,6 +99,7 @@ export { default as managedAgentInit } from "./commands/managed-agent/init.ts";
 export { default as managedAgentValidate } from "./commands/managed-agent/validate.ts";
 export { default as managedAgentPlan } from "./commands/managed-agent/plan.ts";
 export { default as managedAgentApply } from "./commands/managed-agent/apply.ts";
+export { default as managedAgentRun } from "./commands/managed-agent/run.ts";
 export { default as managedAgentDestroy } from "./commands/managed-agent/destroy.ts";
 export { default as managedAgentStateList } from "./commands/managed-agent/state-list.ts";
 export { default as managedAgentStateShow } from "./commands/managed-agent/state-show.ts";

packages/commands/tests/e2e/topic-routes.ts

@@ -177,6 +177,7 @@ export const MANAGED_AGENT_ROUTES: E2eRouteExports = {
   "managed-agent validate": "managedAgentValidate",
   "managed-agent plan": "managedAgentPlan",
   "managed-agent apply": "managedAgentApply",
+  "managed-agent run": "managedAgentRun",
   "managed-agent destroy": "managedAgentDestroy",
   "managed-agent state list": "managedAgentStateList",
   "managed-agent state rm": "managedAgentStateRm",

packages/dsh/README.md

@@ -4,23 +4,23 @@
 
 一个包提供 5 个插件行，外加对 base bundle 的 `llm-pi-ai` 行做一次配置覆盖：
 
-| row id                           | 能力                                                            | 默认 | 依赖                  |
-| -------------------------------- | --------------------------------------------------------------- | ---- | --------------------- |
-| `llm-pi-ai`（覆盖 base 行）      | 把百炼 TokenPlan 网关注册成 LLM provider（`bailian-tokenplan`） | 启用 | TokenPlan Key         |
-| `bailian-tool-vision`            | `bailian_vision_describe`：图片/视频理解                        | 启用 | `bl`                  |
-| `bailian-tool-image`             | `bailian_image_generate`：文生图                                | 启用 | `bl`                  |
-| `bailian-web-search-rag`         | 百炼知识库检索，注册为 `web_search` 的后端                      | 停用 | 按量付费 Key + 知识库 |
-| `bailian-memory`                 | 跨会话长期记忆（tools + 自动检索/落库）                         | 停用 | 按量付费 Key          |
-| `bailian-subagent-managed-agent` | 子 agent 跑在百炼托管运行时                                     | 停用 | `bl` + `agents.yaml`  |
+| row id                       | 能力                                                            | 默认 | 依赖                  |
+| ---------------------------- | --------------------------------------------------------------- | ---- | --------------------- |
+| `llm-pi-ai`（覆盖 base 行）  | 把百炼 TokenPlan 网关注册成 LLM provider（`bailian-tokenplan`） | 启用 | TokenPlan Key         |
+| `bailian-tool-vision`        | `bailian_vision_describe`：图片/视频理解                        | 启用 | `bl`                  |
+| `bailian-tool-image`         | `bailian_image_generate`：文生图                                | 启用 | `bl`                  |
+| `bailian-tool-managed-agent` | `bailian_run_remote_task`：按需在云端创建 agent 并跑任务        | 启用 | `bl` + 按量付费 Key   |
+| `bailian-web-search-rag`     | 百炼知识库检索，注册为 `web_search` 的后端                      | 停用 | 按量付费 Key + 知识库 |
+| `bailian-memory`             | 跨会话长期记忆（tools + 自动检索/落库）                         | 停用 | 按量付费 Key          |
 
-后三个默认停用是有意的：它们要么需要部署方特有的资源 ID，要么按次计费，不该在用户没配置时就生效。
+`web-search-rag` 与 `memory` 默认停用是有意的：它们要么需要部署方特有的资源 ID，要么按次计费，不该在用户没配置时就生效。`tool-managed-agent` 默认启用——它加载时不建任何资源，只有模型真正调用时才在云端创建 agent。
 
 ---
 
 ## 1. 前置条件
 
 - Node ≥ 22.19（`dsh` 的要求）
-- `bl`（vision / image / subagent 三个插件通过子进程调它）
+- `bl`（vision / image / 远程任务 三个工具通过子进程调它）
 
   ```sh
   npm install -g bailian-cli
@@ -105,7 +105,9 @@ Web UI 在 http://127.0.0.1:3080。
 | `deepseek-v4-pro`        | 否                 |
 | `deepseek-v4-flash-0731` | 否                 |
 
-**两个工具** — `bailian_vision_describe`、`bailian_image_generate`。
+**三个工具** — `bailian_vision_describe`、`bailian_image_generate`、`bailian_run_remote_task`。
+
+前两个走 TokenPlan（vision/image）。`bailian_run_remote_task` 见 [§3.1](#31-远程任务-bailian_run_remote_task)——它默认启用但用的是**按量付费 Key + dashscope 端点**，与 TokenPlan 那两个不同。
 
 ### 关于看图，有个坑值得知道
 
@@ -118,6 +120,24 @@ DeepSeek 那两个模型在 TokenPlan 网关上传图**不报错但也看不见*
 
 `bailian_image_generate` 同理：模型能看图时返回内联图片，不能看图时降级为返回落盘路径，你可以接着用 vision 工具读它。文件不会被删除，正是为了这个衔接。
 
+### 3.1 远程任务（`bailian_run_remote_task`）
+
+把一个任务甩到百炼云端的托管 agent 上跑，不占本地会话。**无需预先写 `agents.yaml` 或 `apply`**：工具首次被调用时，`bl managed-agent run` 会在你的账号里幂等创建一个 agent + cloud environment，之后复用。
+
+- 模型自己按用户意图填 `instructions`（远程 agent 的角色），`task` 是要它做的事。例如你说「在云端帮我审计这个依赖树，它该懂安全」→ 模型调 `bailian_run_remote_task(task="审计依赖树", instructions="你是安全专家")`。
+- **前提**：这条路走的是 managed-agent（agentstudio）服务，需要**按量付费 Key**（`sk-ws-`）+ dashscope 端点，且账号已开通 managed-agent。TokenPlan Key 不适用。若 `DASHSCOPE_API_KEY`/端点没配好，首次调用会返回 `Bailian API 404`。
+- 首次会创建云资源（可能计费、启动有延迟）；同名 agent 后续复用。默认 agent 名 `dsh-remote-runner`，可在配置里改。
+
+需要非默认的 agent 名 / 模型时：
+
+```yaml
+- id: bailian-tool-managed-agent
+  config:
+    agent: my-runner
+    model: qwen3.8-max
+    timeoutMs: 600000
+```
+
 ---
 
 ## 4. 开启可选插件
@@ -181,30 +201,7 @@ dsh 自身没有跨会话记忆（`ctx.compaction` 只在单会话内压缩上
     autoPersist: false
 ```
 
-### 托管子 agent
-
-把 dsh 的子 agent 派发到百炼云端运行，不占本地资源。
-
-先在工作目录准备好清单并应用：
-
-```sh
-bl managed-agent init
-bl managed-agent apply --yes
-```
-
-再开启：
-
-```yaml
-- id: bailian-subagent-managed-agent
-  disabled: false
-  config:
-    file: agents.yaml
-    agent: assistant
-    provider: bailian
-    timeoutMs: 600000
-```
-
-两个已知限制：CLI 在会话结束时才一次性输出 JSON，所以**中途没有增量进度**，被取消时也拿不到部分输出；`prepareContinuable` 未实现，即只支持一次性委派，不支持多轮续聊。
+> 远程任务（`bailian_run_remote_task`）默认启用，配置见 [§3.1](#31-远程任务-bailian_run_remote_task)。
 
 ---
 
@@ -226,7 +223,7 @@ bl auth status
 - **文生图**：让模型生成一张图
 - **RAG**：问一个只有知识库里才有答案的问题
 - **记忆**：会话 A 告诉它一个事实 → 关掉 → 新开会话 B 提问，看是否命中
-- **子 agent**：派发一个子任务
+- **远程任务**：说「在云端帮我跑一个任务：<something>，它该擅长 <role>」→ 确认模型调用 `bailian_run_remote_task`（`instructions` 由模型按 role 填）→ 首次触发云端创建 → 返回远程会话结果（需按量付费 Key + 已开通 agentstudio）
 
 ---
 

packages/dsh/cordis.patch.yml

@@ -76,6 +76,12 @@
     - id: bailian-tool-image
       name: bailian-cli-dsh/tool-image
 
+    # Enabled by default: the tool creates no resources at load time. It only
+    # provisions a cloud agent when the model actually calls it, and reuses it
+    # after — no deployment-specific ID to configure up front.
+    - id: bailian-tool-managed-agent
+      name: bailian-cli-dsh/tool-managed-agent
+
     # Disabled by default: the knowledge base to query is deployment-specific,
     # and an enabled provider with no agentId would make `web_search` ambiguous
     # for everyone. Set workspaceId + agentId and flip `disabled` to use it.
@@ -89,9 +95,3 @@
       name: bailian-cli-dsh/memory
       disabled: true
       config: {}
-
-    # Disabled by default: requires an applied `agents.yaml` in the workspace.
-    - id: bailian-subagent-managed-agent
-      name: bailian-cli-dsh/subagent-managed-agent
-      disabled: true
-      config: {}

packages/dsh/package.json

@@ -33,6 +33,10 @@
       "types": "./src/tool-image/index.ts",
       "default": "./src/tool-image/index.ts"
     },
+    "./tool-managed-agent": {
+      "types": "./src/tool-managed-agent/index.ts",
+      "default": "./src/tool-managed-agent/index.ts"
+    },
     "./web-search-rag": {
       "types": "./src/web-search-rag/index.ts",
       "default": "./src/web-search-rag/index.ts"
@@ -41,10 +45,6 @@
       "types": "./src/memory/index.ts",
       "default": "./src/memory/index.ts"
     },
-    "./subagent-managed-agent": {
-      "types": "./src/subagent-managed-agent/index.ts",
-      "default": "./src/subagent-managed-agent/index.ts"
-    },
     "./cordis.patch.yml": "./cordis.patch.yml",
     "./package.json": "./package.json"
   },
@@ -54,9 +54,9 @@
       ".": "./dist/index.mjs",
       "./tool-vision": "./dist/tool-vision/index.mjs",
       "./tool-image": "./dist/tool-image/index.mjs",
+      "./tool-managed-agent": "./dist/tool-managed-agent/index.mjs",
       "./web-search-rag": "./dist/web-search-rag/index.mjs",
       "./memory": "./dist/memory/index.mjs",
-      "./subagent-managed-agent": "./dist/subagent-managed-agent/index.mjs",
       "./cordis.patch.yml": "./cordis.patch.yml",
       "./package.json": "./package.json"
     },

packages/dsh/src/subagent-managed-agent/index.ts

@@ -1,205 +0,0 @@
-/**
- * `bailian-cli-dsh/subagent-managed-agent`: runs child agents on Bailian's
- * hosted managed-agent runtime instead of in this process, through
- * `bl managed-agent session run`.
- *
- * Out-of-process delegation is an established shape here — `subagent-acp` and
- * `subagent-codex` do the same over their own transports. Going through the
- * CLI keeps `agents.yaml` resolution, provider selection, and SSE decoding in
- * one place.
- *
- * Two honest limits. The CLI buffers the whole session and emits it at exit,
- * so no incremental progress reaches the parent and a cancelled run yields no
- * partial output. And `prepareContinuable` is deliberately absent: method
- * presence IS the continuable capability, and multi-turn continuation is not
- * wired up yet.
- *
- * @module bailian-cli-dsh/subagent-managed-agent
- */
-import type { Context } from "@deepseek-ai/cordis";
-import type { ContentBlock } from "@deepseek-ai/dsh-llm";
-import { SessionId } from "@deepseek-ai/dsh-session";
-import type {
-  ResolvedSubagentStartRequest,
-  SubagentCapabilities,
-  SubagentProvider,
-  SubagentResult,
-  SubagentRun,
-} from "@deepseek-ai/dsh-subagent";
-import type {} from "@deepseek-ai/dsh-fs";
-import z from "@deepseek-ai/schemastery";
-import { runBlJson } from "../shared/bl.ts";
-
-/** Cordis plugin name used by loader diagnostics. */
-export const name = "bailian-subagent-managed-agent";
-
-/** Seams this plugin registers into. */
-export const inject = ["subagents", "subprocess", "fs"];
-
-/** Registry name callers select this transport by. */
-export const BAILIAN_MANAGED_AGENT_PROVIDER = "bailian-managed-agent";
-
-export interface Config {
-  /** Manifest passed as `--file`; must already be applied. */
-  file?: string;
-  /** Agent name within the manifest. */
-  agent?: string;
-  /** Backing provider understood by `bl managed-agent`. */
-  provider?: string;
-  /** Cooperative budget for one hosted run. */
-  timeoutMs?: number;
-}
-
-export const Config: z<Config> = z.object({
-  file: z.string().description("Path to agents.yaml; defaults to the CLI's own default."),
-  agent: z.string().description("Agent name declared in the manifest."),
-  provider: z.string().description("Managed-agent backing provider."),
-  timeoutMs: z.natural().description("Cooperative timeout budget in milliseconds."),
-});
-
-const DEFAULT_MANIFEST = "agents.yaml";
-const DEFAULT_TIMEOUT_MS = 600_000;
-
-/** A one-shot transport supports none of the start-time features. */
-const CAPABILITIES: SubagentCapabilities = {
-  outputSchema: false,
-  depthLimit: false,
-  toolFilter: false,
-  persona: false,
-};
-
-interface SessionEvent {
-  type?: string;
-  content?: unknown;
-  role?: string;
-}
-
-interface SessionRunResponse {
-  session_id?: string;
-  events?: readonly SessionEvent[];
-}
-
-function promptText(blocks: readonly ContentBlock[]): string {
-  return blocks
-    .filter((block): block is Extract<ContentBlock, { type: "text" }> => block.type === "text")
-    .map((block) => block.text)
-    .join("\n")
-    .trim();
-}
-
-/** Assistant-visible text of a finished hosted session. */
-function assistantOutput(events: readonly SessionEvent[]): ContentBlock[] {
-  const text = events
-    .filter((event) => event.type === "message" && typeof event.content === "string")
-    .map((event) => event.content as string)
-    .join("\n")
-    .trim();
-  return text.length > 0 ? [{ type: "text", text }] : [];
-}
-
-function isAbort(error: unknown): boolean {
-  return error instanceof DOMException && error.name === "AbortError";
-}
-
-class BailianManagedAgentProvider implements SubagentProvider {
-  readonly name = BAILIAN_MANAGED_AGENT_PROVIDER;
-  readonly capabilities = CAPABILITIES;
-  readonly inheritsParentContext = false;
-
-  constructor(
-    private readonly ctx: Context,
-    private readonly config: Config,
-  ) {}
-
-  async start(request: ResolvedSubagentStartRequest): Promise<SubagentRun> {
-    const cwd = request.parent.session.header.cwd ?? process.cwd();
-    const manifest = this.config.file ?? DEFAULT_MANIFEST;
-
-    // Pre-publication: a missing manifest is the common misconfiguration and
-    // deserves a start-time rejection rather than a failed run.
-    const target = await this.ctx.fs.resolve(manifest, { cwd, signal: request.signal });
-    const info = await this.ctx.fs.stat(target, request.signal);
-    if (info === undefined) {
-      throw new Error(
-        `bailian-managed-agent: no manifest at "${target.displayPath}". Create one with ` +
-          `\`bl managed-agent init\` and apply it with \`bl managed-agent apply --yes\`.`,
-      );
-    }
-
-    const prompt = promptText(request.prompt);
-    if (prompt.length === 0) {
-      throw new Error("bailian-managed-agent: the prompt carried no text content.");
-    }
-
-    const argv = ["managed-agent", "session", "run", "--prompt", prompt, "--file", manifest];
-    if (this.config.agent !== undefined) argv.push("--agent", this.config.agent);
-    if (this.config.provider !== undefined) argv.push("--provider", this.config.provider);
-
-    const controller = new AbortController();
-    const abort = (): void => controller.abort();
-    request.signal.addEventListener("abort", abort, { once: true });
-
-    // The seam has no deadline of its own — cancellation arrives only through
-    // the caller's signal — so the transport owns one, or a wedged hosted
-    // session never settles.
-    const deadline = AbortSignal.timeout(this.config.timeoutMs ?? DEFAULT_TIMEOUT_MS);
-    const combined = AbortSignal.any([controller.signal, deadline]);
-
-    // Ownership transfers on fulfillment, so every later failure settles
-    // through `result` — which must not reject for child-level problems.
-    const result = this.execute(argv, cwd, combined, deadline).finally(() => {
-      request.signal.removeEventListener("abort", abort);
-    });
-
-    let disposal: Promise<void> | undefined;
-    return {
-      id: SessionId(`bailian-managed-agent:${crypto.randomUUID()}`),
-      localAgent: undefined,
-      result,
-      dispose: (): Promise<void> => {
-        disposal ??= (async (): Promise<void> => {
-          controller.abort();
-          await result;
-        })();
-        return disposal;
-      },
-    };
-  }
-
-  private async execute(
-    argv: readonly string[],
-    cwd: string,
-    signal: AbortSignal,
-    deadline: AbortSignal,
-  ): Promise<SubagentResult> {
-    try {
-      const response = await runBlJson<SessionRunResponse>(this.ctx, argv, {
-        cwd,
-        signal,
-        graceMs: 10_000,
-      });
-      return { output: assistantOutput(response.events ?? []), stopReason: "completed" };
-    } catch (error) {
-      if (deadline.aborted) {
-        return {
-          output: [
-            {
-              type: "text",
-              text: `the hosted session exceeded ${this.config.timeoutMs ?? DEFAULT_TIMEOUT_MS}ms and was terminated`,
-            },
-          ],
-          stopReason: "error",
-        };
-      }
-      // The CLI emits its JSON envelope only at exit, so a cancelled run has
-      // no partial output to salvage.
-      if (isAbort(error) || signal.aborted) return { output: [], stopReason: "aborted" };
-      const reason = error instanceof Error ? error.message : String(error);
-      return { output: [{ type: "text", text: reason }], stopReason: "error" };
-    }
-  }
-}
-
-export function apply(ctx: Context, config: Config): void {
-  ctx.subagents.registerProvider(new BailianManagedAgentProvider(ctx, config));
-}

packages/dsh/src/tool-managed-agent/index.ts

@@ -0,0 +1,131 @@
+/**
+ * `bailian-cli-dsh/tool-managed-agent`: run a task on a Bailian-hosted managed
+ * agent, provisioned on demand, through `bl managed-agent run`.
+ *
+ * A plain tool rather than a `SubagentProvider`: in dsh's `web` profile every
+ * `tool-subagent` row is disabled in the host plane (delegation tools live in
+ * agent presets), and a subagent provider fixes one agent identity in config —
+ * neither fits "the model describes an intent and a remote agent is created for
+ * it". As a tool the model calls it directly and fills `instructions` from the
+ * user's intent, so the remote agent's role is defined per task.
+ *
+ * The CLI does ensure+run in one step: it materializes (idempotently) a cloud
+ * agent + environment under the given `agent` name on first use and reuses them
+ * after, so no `agents.yaml` or prior `apply` is required. First use provisions
+ * cloud resources — it may incur cost and take longer to start.
+ *
+ * @module bailian-cli-dsh/tool-managed-agent
+ */
+import type { Context } from "@deepseek-ai/cordis";
+import { defineTool } from "@deepseek-ai/dsh-tools";
+import type {} from "@deepseek-ai/dsh-tools";
+import z from "@deepseek-ai/schemastery";
+import { runBlJson } from "../shared/bl.ts";
+
+/** Cordis plugin name used by loader diagnostics. */
+export const name = "bailian-tool-managed-agent";
+
+/** Seams this plugin registers into. */
+export const inject = ["tools", "subprocess"];
+
+/** Default agent identity provisioned and reused across calls. */
+const DEFAULT_AGENT = "dsh-remote-runner";
+
+export interface Config {
+  /** Agent identity to create/reuse; distinct names get distinct remote agents. */
+  agent?: string;
+  /** Model for the remote agent. */
+  model?: string;
+  /** Cooperative budget; first-run provisioning of a cloud environment is slow. */
+  timeoutMs?: number;
+}
+
+export const Config: z<Config> = z.object({
+  agent: z.string().description("Remote agent identity to create/reuse."),
+  model: z.string().description("Model for the remote agent."),
+  timeoutMs: z.natural().description("Cooperative timeout budget in milliseconds."),
+});
+
+const DEFAULT_TIMEOUT_MS = 600_000;
+
+/** The `bl managed-agent run --output json` envelope: a session-event list. */
+interface SessionRunResponse {
+  session_id?: string;
+  agent?: string;
+  events?: readonly { type?: string; content?: unknown; role?: string }[];
+}
+
+/** Assistant-visible text of a finished remote session. */
+function assistantText(response: SessionRunResponse): string {
+  return (response.events ?? [])
+    .filter((event) => event.type === "message" && typeof event.content === "string")
+    .map((event) => event.content as string)
+    .join("\n")
+    .trim();
+}
+
+export function apply(ctx: Context, config: Config): void {
+  ctx.tools.register(
+    defineTool({
+      name: "bailian_run_remote_task",
+      description:
+        "Run a task on a Bailian-hosted cloud agent. Use for long-running or isolated work you " +
+        "want executed remotely rather than in this session. A remote agent is created on demand " +
+        "(and reused) — describe the role it should play through `instructions`, and the concrete " +
+        "task through `task`. Returns the remote agent's final answer.",
+      parameters: {
+        task: {
+          type: "string",
+          required: true,
+          description: "The concrete task for the remote agent to carry out.",
+        },
+        instructions: {
+          type: "string",
+          description:
+            "Role/system instructions defining what the remote agent is good at. " +
+            "Defaults to a generic assistant.",
+        },
+        model: {
+          type: "string",
+          description: "Override the configured model for this task.",
+        },
+      },
+      output: {
+        schema: {
+          type: "object",
+          additionalProperties: false,
+          properties: {
+            answer: { type: "string", required: true },
+            sessionId: { type: "string", required: true },
+          },
+        },
+        render: (_args, value) => [{ type: "text", text: value.answer }],
+      },
+      timeoutMs: config.timeoutMs ?? DEFAULT_TIMEOUT_MS,
+      async execute(args, exec) {
+        const argv = [
+          "managed-agent",
+          "run",
+          "--prompt",
+          args.task,
+          "--agent",
+          config.agent ?? DEFAULT_AGENT,
+        ];
+        if (args.instructions !== undefined) argv.push("--instructions", args.instructions);
+        const model = args.model ?? config.model;
+        if (model !== undefined) argv.push("--model", model);
+
+        const response = await runBlJson<SessionRunResponse>(ctx, argv, {
+          cwd: exec.agent?.session.header.cwd ?? process.cwd(),
+          signal: exec.signal,
+        });
+
+        const answer = assistantText(response);
+        if (answer.length === 0) {
+          throw new Error("the remote agent produced no assistant output.");
+        }
+        return { answer, sessionId: response.session_id ?? "" };
+      },
+    }),
+  );
+}

packages/dsh/vite.config.ts

@@ -8,9 +8,9 @@ export default defineConfig({
       "src/index.ts",
       "src/tool-vision/index.ts",
       "src/tool-image/index.ts",
+      "src/tool-managed-agent/index.ts",
       "src/web-search-rag/index.ts",
       "src/memory/index.ts",
-      "src/subagent-managed-agent/index.ts",
     ],
     minify: true,
     dts: {

skills/bailian-managed-agent/reference/index.md

@@ -9,31 +9,32 @@ Use this index for the skill-scoped quick index and global flags.
 
 ## Quick index
 
-| Command                           | Description                                                   | Detail                               |
-| --------------------------------- | ------------------------------------------------------------- | ------------------------------------ |
-| `bl managed-agent apply`          | Apply planned changes to create/update/delete agent resources | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent destroy`        | Destroy all managed agent resources tracked in state          | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent init`           | Create a new agents.yaml template                             | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent plan`           | Show what changes would be applied to agent infrastructure    | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent session create` | Create a new session for an agent                             | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent session delete` | Delete a session                                              | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent session events` | List event history for a session                              | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent session get`    | Get details of a session                                      | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent session list`   | List sessions from the provider                               | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent session run`    | Create a session, send a message, and stream the response     | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent session send`   | Send a message to an existing session and stream the response | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent skill-list`     | List skills from the provider's skill catalog                 | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent state import`   | Import an existing remote resource into agents state          | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent state list`     | List resources tracked in agents state                        | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent state rm`       | Remove a resource from state without destroying it remotely   | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent state show`     | Show details of a resource in agents state                    | [managed-agent.md](managed-agent.md) |
-| `bl managed-agent validate`       | Validate an agents.yaml configuration (offline)               | [managed-agent.md](managed-agent.md) |
+| Command                           | Description                                                    | Detail                               |
+| --------------------------------- | -------------------------------------------------------------- | ------------------------------------ |
+| `bl managed-agent apply`          | Apply planned changes to create/update/delete agent resources  | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent destroy`        | Destroy all managed agent resources tracked in state           | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent init`           | Create a new agents.yaml template                              | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent plan`           | Show what changes would be applied to agent infrastructure     | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent run`            | Provision (if needed) a cloud agent and run a task in one step | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent session create` | Create a new session for an agent                              | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent session delete` | Delete a session                                               | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent session events` | List event history for a session                               | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent session get`    | Get details of a session                                       | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent session list`   | List sessions from the provider                                | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent session run`    | Create a session, send a message, and stream the response      | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent session send`   | Send a message to an existing session and stream the response  | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent skill-list`     | List skills from the provider's skill catalog                  | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent state import`   | Import an existing remote resource into agents state           | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent state list`     | List resources tracked in agents state                         | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent state rm`       | Remove a resource from state without destroying it remotely    | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent state show`     | Show details of a resource in agents state                     | [managed-agent.md](managed-agent.md) |
+| `bl managed-agent validate`       | Validate an agents.yaml configuration (offline)                | [managed-agent.md](managed-agent.md) |
 
 ## By group
 
-| Group           | Commands                                                                                                                                                                                                                                 | Reference                            |
-| --------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------ |
-| `managed-agent` | `apply`, `destroy`, `init`, `plan`, `session create`, `session delete`, `session events`, `session get`, `session list`, `session run`, `session send`, `skill-list`, `state import`, `state list`, `state rm`, `state show`, `validate` | [managed-agent.md](managed-agent.md) |
+| Group           | Commands                                                                                                                                                                                                                                        | Reference                            |
+| --------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------ |
+| `managed-agent` | `apply`, `destroy`, `init`, `plan`, `run`, `session create`, `session delete`, `session events`, `session get`, `session list`, `session run`, `session send`, `skill-list`, `state import`, `state list`, `state rm`, `state show`, `validate` | [managed-agent.md](managed-agent.md) |
 
 ## Global flags
 

skills/bailian-managed-agent/reference/managed-agent.md

@@ -7,25 +7,26 @@ Index: [index.md](index.md)
 
 ## Commands in this group
 
-| Command                           | Description                                                   |
-| --------------------------------- | ------------------------------------------------------------- |
-| `bl managed-agent apply`          | Apply planned changes to create/update/delete agent resources |
-| `bl managed-agent destroy`        | Destroy all managed agent resources tracked in state          |
-| `bl managed-agent init`           | Create a new agents.yaml template                             |
-| `bl managed-agent plan`           | Show what changes would be applied to agent infrastructure    |
-| `bl managed-agent session create` | Create a new session for an agent                             |
-| `bl managed-agent session delete` | Delete a session                                              |
-| `bl managed-agent session events` | List event history for a session                              |
-| `bl managed-agent session get`    | Get details of a session                                      |
-| `bl managed-agent session list`   | List sessions from the provider                               |
-| `bl managed-agent session run`    | Create a session, send a message, and stream the response     |
-| `bl managed-agent session send`   | Send a message to an existing session and stream the response |
-| `bl managed-agent skill-list`     | List skills from the provider's skill catalog                 |
-| `bl managed-agent state import`   | Import an existing remote resource into agents state          |
-| `bl managed-agent state list`     | List resources tracked in agents state                        |
-| `bl managed-agent state rm`       | Remove a resource from state without destroying it remotely   |
-| `bl managed-agent state show`     | Show details of a resource in agents state                    |
-| `bl managed-agent validate`       | Validate an agents.yaml configuration (offline)               |
+| Command                           | Description                                                    |
+| --------------------------------- | -------------------------------------------------------------- |
+| `bl managed-agent apply`          | Apply planned changes to create/update/delete agent resources  |
+| `bl managed-agent destroy`        | Destroy all managed agent resources tracked in state           |
+| `bl managed-agent init`           | Create a new agents.yaml template                              |
+| `bl managed-agent plan`           | Show what changes would be applied to agent infrastructure     |
+| `bl managed-agent run`            | Provision (if needed) a cloud agent and run a task in one step |
+| `bl managed-agent session create` | Create a new session for an agent                              |
+| `bl managed-agent session delete` | Delete a session                                               |
+| `bl managed-agent session events` | List event history for a session                               |
+| `bl managed-agent session get`    | Get details of a session                                       |
+| `bl managed-agent session list`   | List sessions from the provider                                |
+| `bl managed-agent session run`    | Create a session, send a message, and stream the response      |
+| `bl managed-agent session send`   | Send a message to an existing session and stream the response  |
+| `bl managed-agent skill-list`     | List skills from the provider's skill catalog                  |
+| `bl managed-agent state import`   | Import an existing remote resource into agents state           |
+| `bl managed-agent state list`     | List resources tracked in agents state                         |
+| `bl managed-agent state rm`       | Remove a resource from state without destroying it remotely    |
+| `bl managed-agent state show`     | Show details of a resource in agents state                     |
+| `bl managed-agent validate`       | Validate an agents.yaml configuration (offline)                |
 
 ## Command details
 
@@ -170,6 +171,43 @@ bl managed-agent plan --provider bailian
 bl managed-agent plan --no-refresh
 ```
 
+### `bl managed-agent run`
+
+| Field           | Value                                                                                          |
+| --------------- | ---------------------------------------------------------------------------------------------- |
+| **Name**        | `managed-agent run`                                                                            |
+| **Description** | Provision (if needed) a cloud agent and run a task in one step                                 |
+| **Usage**       | `bl managed-agent run --prompt <text> [--instructions <text>] [--model <id>] [--agent <name>]` |
+
+#### Flags
+
+| Flag                    | Type   | Required | Description                                                                |
+| ----------------------- | ------ | -------- | -------------------------------------------------------------------------- |
+| `--prompt <text>`       | string | yes      | Task to run (required)                                                     |
+| `--instructions <text>` | string | no       | Role/system instructions for the remote agent (default: generic assistant) |
+| `--model <id>`          | string | no       | Model for the remote agent (default: qwen3.8-max)                          |
+| `--agent <name>`        | string | no       | Agent identity to create/reuse (default: dsh-remote-runner)                |
+| `--no-stream`           | switch | no       | Use polling instead of SSE streaming                                       |
+| `--api-key <key>`       | string | no       | API key                                                                    |
+| `--base-url <url>`      | string | no       | API base URL                                                               |
+
+#### Notes
+
+- Bailian credentials come from bl's auth chain: --api-key > DASHSCOPE_API_KEY > `bl auth login` (active config profile).
+- Other providers read the env vars referenced in agents.yaml (e.g. ${ANTHROPIC_API_KEY}), including .env and ~/.agents/config.json.
+- Resolved credentials are injected into the SDK in-memory and cleared from the environment; they never persist in process env.
+- Unlike `apply`, this creates/updates the cloud agent + environment on demand without --yes. The first run provisions cloud resources (may incur cost and take longer to start); later runs with the same --agent reuse them.
+
+#### Examples
+
+```bash
+bl managed-agent run --prompt "Summarize the latest AI news"
+```
+
+```bash
+bl managed-agent run --prompt "Audit this dependency tree" --instructions "You are a security expert" --model qwen3.8-max
+```
+
 ### `bl managed-agent session create`
 
 | Field           | Value                                                                                                       |