feat: code review fixture

Author: gujieye

README.md

@@ -38,7 +38,7 @@ Equip your AI Agent out-of-the-box with these capabilities, composable across co
 - **MCP integration** — Orchestrate Bailian MCP servers: list services, inspect tools, and invoke any tool directly from the terminal
 - **Web search** — Real-time internet retrieval for up-to-date, accurate answers
 - **Model recommendation** — Describe your scenario and get best-fit model suggestions; supports scoped search, model comparison, and alternative discovery
-- **Fine-tuning & deployment** — Upload datasets, create SFT/LoRA/DPO/CPT jobs (`finetune create`), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy create`)
+- **Fine-tuning & deployment** — Upload datasets, create text/audio/image fine-tune jobs (`finetune text|audio|image create`; text covers SFT/LoRA/DPO/CPT), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy text|audio|image create`)
 - **Console capabilities** — Browse Bailian apps (`app list`), check free-tier quota (`usage free`), view model usage statistics (`usage stats`), manage workspaces (`workspace list`), and manage rate limits (`quota list/request/check/history`)
 - **Local file auto-upload** — Every URL parameter accepts a local path; uploaded to free temp storage with 48-hour validity
 
@@ -114,10 +114,10 @@ bl auth login --console
 
 # Fine-tune & deploy — a one-shot train-to-serve workflow
 bl dataset upload --file ./train.jsonl                 # Upload a .jsonl dataset (validated first)
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
 bl finetune watch --job-id ft-xxx --output json       # Non-blocking status probe (exit 0/1/3 = done/failed/running)
 bl finetune capability --model qwen3-8b               # Which training types a model supports
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
 
 # Browse apps / free-tier quota / usage statistics / workspaces
 bl app list

README.zh.md

@@ -38,7 +38,7 @@ _专为 AI Agent 打造，每个命令均可作为结构化工具调用。_
 - **MCP 集成** — 统一调度百炼 MCP 服务：列出服务、查看工具、直接在终端调用任意工具
 - **联网搜索** — 实时互联网信息检索，提升回答准确性及时效性
 - **模型推荐** — 描述你的场景，智能推荐最适合的模型；支持限定范围搜索、模型对比和替代发现
-- **微调与部署** — 上传数据集、创建 SFT/LoRA/DPO/CPT 调优任务（`finetune create`）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy create`）
+- **微调与部署** — 上传数据集、创建文本/音频/图像调优任务（`finetune text|audio|image create`；文本涵盖 SFT/LoRA/DPO/CPT）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy text|audio|image create`）
 - **控制台能力** — 浏览百炼应用（`app list`），查询模型免费额度（`usage free`），查看模型用量统计（`usage stats`），管理业务空间（`workspace list`），管理限流与提额（`quota list/request/check/history`）
 - **本地文件自动上传** — 所有 URL 参数同时支持本地路径，免费临时存储 48 小时
 
@@ -112,10 +112,10 @@ bl auth login --console
 
 # 微调与部署 — 从训练到服务的一站式流程
 bl dataset upload --file ./train.jsonl                 # 上传 .jsonl 数据集（先校验）
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
 bl finetune watch --job-id ft-xxx --output json       # 非阻塞状态探测（退出码 0/1/3 = 成功/失败/进行中）
 bl finetune capability --model qwen3-8b               # 查询模型支持哪些训练方式
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
 
 # 浏览应用 / 免费额度 / 用量统计 / 业务空间
 bl app list

packages/cli/README.md

@@ -38,7 +38,7 @@ Equip your AI Agent out-of-the-box with these capabilities, composable across co
 - **MCP integration** — Orchestrate Bailian MCP servers: list services, inspect tools, and invoke any tool directly from the terminal
 - **Web search** — Real-time internet retrieval for up-to-date, accurate answers
 - **Model recommendation** — Describe your scenario and get best-fit model suggestions; supports scoped search, model comparison, and alternative discovery
-- **Fine-tuning & deployment** — Upload datasets, create SFT/LoRA/DPO/CPT jobs (`finetune create`), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy create`)
+- **Fine-tuning & deployment** — Upload datasets, create text/audio/image fine-tune jobs (`finetune text|audio|image create`; text covers SFT/LoRA/DPO/CPT), probe job status non-blockingly (`finetune watch`), query per-model training capability (`finetune capability`), and deploy trained models as endpoints (`deploy text|audio|image create`)
 - **Console capabilities** — Browse Bailian apps (`app list`), check free-tier quota (`usage free`), view model usage statistics (`usage stats`), manage workspaces (`workspace list`), and manage rate limits (`quota list/request/check/history`)
 - **Local file auto-upload** — Every URL parameter accepts a local path; uploaded to free temp storage with 48-hour validity
 
@@ -114,10 +114,10 @@ bl auth login --console
 
 # Fine-tune & deploy — a one-shot train-to-serve workflow
 bl dataset upload --file ./train.jsonl                 # Upload a .jsonl dataset (validated first)
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # Local paths auto-upload
 bl finetune watch --job-id ft-xxx --output json       # Non-blocking status probe (exit 0/1/3 = done/failed/running)
 bl finetune capability --model qwen3-8b               # Which training types a model supports
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # Deploy the trained model as an endpoint
 
 # Browse apps / free-tier quota / usage statistics / workspaces
 bl app list

packages/cli/README.zh.md

@@ -38,7 +38,7 @@ _专为 AI Agent 打造，每个命令均可作为结构化工具调用。_
 - **MCP 集成** — 统一调度百炼 MCP 服务：列出服务、查看工具、直接在终端调用任意工具
 - **联网搜索** — 实时互联网信息检索，提升回答准确性及时效性
 - **模型推荐** — 描述你的场景，智能推荐最适合的模型；支持限定范围搜索、模型对比和替代发现
-- **微调与部署** — 上传数据集、创建 SFT/LoRA/DPO/CPT 调优任务（`finetune create`）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy create`）
+- **微调与部署** — 上传数据集、创建文本/音频/图像调优任务（`finetune text|audio|image create`；文本涵盖 SFT/LoRA/DPO/CPT）、非阻塞探测任务状态（`finetune watch`）、按模型查训练能力（`finetune capability`），并把训练好的模型部署为推理服务（`deploy text|audio|image create`）
 - **控制台能力** — 浏览百炼应用（`app list`），查询模型免费额度（`usage free`），查看模型用量统计（`usage stats`），管理业务空间（`workspace list`），管理限流与提额（`quota list/request/check/history`）
 - **本地文件自动上传** — 所有 URL 参数同时支持本地路径，免费临时存储 48 小时
 
@@ -112,10 +112,10 @@ bl auth login --console
 
 # 微调与部署 — 从训练到服务的一站式流程
 bl dataset upload --file ./train.jsonl                 # 上传 .jsonl 数据集（先校验）
-bl finetune create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
+bl finetune text create --model qwen3-8b --datasets ./train.jsonl --training-type sft-lora  # 本地路径自动上传
 bl finetune watch --job-id ft-xxx --output json       # 非阻塞状态探测（退出码 0/1/3 = 成功/失败/进行中）
 bl finetune capability --model qwen3-8b               # 查询模型支持哪些训练方式
-bl deploy create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
+bl deploy text create --model qwen3-8b --name my-svc --plan mu  # 把训练好的模型部署为推理服务
 
 # 浏览应用 / 免费额度 / 用量统计 / 业务空间
 bl app list

packages/cli/tests/e2e/deploy.e2e.test.ts

@@ -87,9 +87,11 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: deploy (offline)", () => {
     expect(data.body.capacity).toBe(2);
   });
 
-  test("deploy audio create --dry-run 构造 lora 部署请求体", async () => {
-    // deploy create does not inspect modality; the split only changes the path.
-    // The audio subcommand shares the full flag surface and run logic.
+  test("deploy audio create --dry-run 默认 plan=mu（CosyVoice 部署契约）", async () => {
+    // Audio (CosyVoice TTS) outputs deploy model-unit-billed: the modality fixes
+    // the default plan to `mu` (text/image stay `lora`). In dry-run the mu
+    // strategy skips the catalog lookup, so deploy_spec is omitted and capacity
+    // falls back to 1 with billing_method POST_PAY.
     const { stdout, stderr, exitCode } = await runCli([
       "deploy",
       "audio",
@@ -103,10 +105,15 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: deploy (offline)", () => {
       "json",
     ]);
     expect(exitCode, stderr).toBe(0);
-    const data = parseStdoutJson<{ action: string; body: { plan: string; name: string } }>(stdout);
+    const data = parseStdoutJson<{
+      action: string;
+      body: { plan: string; name: string; billing_method?: string; capacity?: number };
+    }>(stdout);
     expect(data.action).toBe("deploy.create");
-    expect(data.body.plan).toBe("lora");
+    expect(data.body.plan).toBe("mu");
     expect(data.body.name).toBe("my-tts");
+    expect(data.body.billing_method).toBe("POST_PAY");
+    expect(data.body.capacity).toBe(1);
   });
 
   test("deploy scale --dry-run 转发 capacity", async () => {

packages/commands/src/commands/deploy/create.ts

@@ -4,7 +4,8 @@ import {
   createDeployment,
   pickPlanStrategy,
   STRATEGIES,
-  DEFAULT_DEPLOY_PLAN,
+  defaultDeployPlan,
+  type DeployModality,
   type CreateDeploymentRequest,
   type CreatePlanFlags,
   type CommandContext,
@@ -66,7 +67,8 @@ const CREATE_USAGE =
   "--model <model_name> --name <display_name> [--plan <plan>] [--deploy-spec <id>] [--capacity <n>] [--billing-method <m>] [--input-tpm <n>] [--output-tpm <n>] [--thinking-output-tpm <n>]";
 
 const CREATE_NOTES = [
-  "Plan defaults to `lora` (Token-billed). Pass --plan to override.",
+  "Plan defaults to `lora` (Token-billed) for text/image and `mu` (model-unit-",
+  "billed) for audio (CosyVoice TTS). Pass --plan to override.",
   "For plan=ptu (Token-billed, provisioned throughput), --input-tpm and",
   "--output-tpm are required (the platform rejects creation without an",
   "explicit ptu_capacity despite the doc listing defaults).",
@@ -77,22 +79,23 @@ const CREATE_NOTES = [
   "After creation, status starts at PENDING and transitions to RUNNING.",
   "Invoke the deployed model with: bl text chat --model <deployed_model>",
   "WARNING: --model is overloaded across commands and refers to DIFFERENT",
-  "values. `bl deploy create --model` takes the exported model_name (e.g.",
-  "`qwen3-8b-ft-...`), but the create response also returns a `deployed_model`",
-  "field (the deployment instance id, e.g. `qwen3-8b-5ecb5f068d79`). The",
-  "inference call `bl text chat --model` must use the `deployed_model` from",
-  "the create response — NOT the `model_name` you passed to `deploy create`.",
-  "Do not reuse the value across the two commands.",
+  "values. `bl deploy <modality> create --model` takes the exported model_name",
+  "(e.g. `qwen3-8b-ft-...`), but the create response also returns a",
+  "`deployed_model` field (the deployment instance id, e.g.",
+  "`qwen3-8b-5ecb5f068d79`). The inference call `bl text chat --model` must use",
+  "the `deployed_model` from the create response — NOT the `model_name` you",
+  "passed to `deploy <modality> create`. Do not reuse the value across the two",
+  "commands.",
 ];
 
 /**
  * Shared `deploy <modality> create` flag validation. Plan support is
- * server-catalog-driven, not modality-scoped, so validation is identical for
- * every modality: reject an unknown --plan, then defer to the plan strategy's
- * required-flag check.
+ * server-catalog-driven, so validation is identical for every modality: resolve
+ * the effective plan (modality-specific default when --plan is omitted), reject
+ * an unknown --plan, then defer to the plan strategy's required-flag check.
  */
-function validateCreate(flags: CreatePlanFlags): string | undefined {
-  const plan = flags.plan || DEFAULT_DEPLOY_PLAN;
+function validateCreate(modality: DeployModality, flags: CreatePlanFlags): string | undefined {
+  const plan = flags.plan || defaultDeployPlan(modality);
   const strategy = STRATEGIES[plan];
   if (!strategy) {
     return `Unsupported plan "${plan}". Supported plans: ${Object.keys(STRATEGIES).join(", ")}.`;
@@ -102,20 +105,23 @@ function validateCreate(flags: CreatePlanFlags): string | undefined {
 
 /**
  * Shared `deploy <modality> create` implementation. deploy create takes a model
- * by name and a billing plan — it does NOT inspect data modality, so the run
- * logic is identical across text / audio / image. The modality split only
- * changes the command path, description and examples; the parameter surface and
- * run logic are unchanged.
+ * by name and a billing plan — it does NOT inspect data modality for the request
+ * body, so the run logic is identical across text / audio / image. The modality
+ * only fixes the default plan (audio → mu, text/image → lora) and the command
+ * path / description / examples.
  *
  * Plan-specific behaviour (required flags / body assembly / auto-pick) lives in
  * core `plans.ts` (`PlanStrategy` + `STRATEGIES`). This file only handles the
  * shared envelope: dispatch, dry-run, and result formatting.
  */
-async function runCreate(ctx: CommandContext<typeof CREATE_FLAGS>): Promise<void> {
+async function runCreate(
+  modality: DeployModality,
+  ctx: CommandContext<typeof CREATE_FLAGS>,
+): Promise<void> {
   const { identity, settings, flags } = ctx;
   const model = flags.model as string;
   const name = flags.name as string;
-  const plan = (flags.plan as string | undefined) || DEFAULT_DEPLOY_PLAN;
+  const plan = (flags.plan as string | undefined) || defaultDeployPlan(modality);
   const format = detectOutputFormat(settings.output);
 
   // Plan-specific behaviour is owned by core `plans.ts`. The strategy resolves
@@ -175,24 +181,24 @@ export const deployTextCreate = defineCommand({
     "--model qwen3-8b --name my-qwen3 --plan mu --deploy-spec MU1 --capacity 2",
   ],
   notes: CREATE_NOTES,
-  validate: (flags) => validateCreate(flags),
-  run: (ctx) => runCreate(ctx),
+  validate: (flags) => validateCreate("text", flags),
+  run: (ctx) => runCreate("text", ctx),
 });
 
-/** `bl deploy audio create` — deploy an audio (TTS) model. */
+/** `bl deploy audio create` — deploy an audio (TTS) model. Defaults to plan=mu. */
 export const deployAudioCreate = defineCommand({
   description: "Create an audio (TTS) model deployment",
   auth: "apiKey",
   usageArgs: CREATE_USAGE,
   flags: CREATE_FLAGS,
   exampleArgs: [
     "--model my-cosyvoice-ft --name my-tts",
-    "--model my-cosyvoice-ft --name my-tts-mu --plan mu",
+    "--model my-cosyvoice-ft --name my-tts --deploy-spec dps-xxxx --capacity 1",
     "--model my-cosyvoice-ft --name my-tts --dry-run",
   ],
   notes: CREATE_NOTES,
-  validate: (flags) => validateCreate(flags),
-  run: (ctx) => runCreate(ctx),
+  validate: (flags) => validateCreate("audio", flags),
+  run: (ctx) => runCreate("audio", ctx),
 });
 
 /** `bl deploy image create` — deploy an image generation model. */
@@ -207,6 +213,6 @@ export const deployImageCreate = defineCommand({
     "--model my-wan-ft --name my-wan --dry-run",
   ],
   notes: CREATE_NOTES,
-  validate: (flags) => validateCreate(flags),
-  run: (ctx) => runCreate(ctx),
+  validate: (flags) => validateCreate("image", flags),
+  run: (ctx) => runCreate("image", ctx),
 });

packages/commands/src/commands/deploy/models.ts

@@ -73,8 +73,8 @@ export default defineCommand({
     //   - custom (fine-tuned): top-level supported_plans: string[]
     //   - base (catalog):      plans: [{plan, templates?, cu_specs?}]
     // For json: surface the deployment-relevant fields preserved as a tree, so
-    // downstream tooling can drive `bl deploy create --template-id <…>` without
-    // a second round-trip. For text: keep the compact one-line summary.
+    // downstream tooling can drive `bl deploy <modality> create --deploy-spec <…>`
+    // without a second round-trip. For text: keep the compact one-line summary.
     if (format === "json") {
       const items = models.map((model) => {
         const out: Record<string, unknown> = {
@@ -92,7 +92,7 @@ export default defineCommand({
               planEntry.cu_specs = plan.cu_specs;
             }
             if (plan.templates && plan.templates.length > 0) {
-              // Pull the top 6 fields most useful for `bl deploy create`.
+              // Pull the top 6 fields most useful for `bl deploy <modality> create`.
               // Drop noisy/redundant: template_source, template_type,
               // template_version, deploy_spec (typically == template_id).
               planEntry.templates = plan.templates.map((template) => {