diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6cb07645d..09369cdb7 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -20,6 +20,24 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
 - **Parameter validation and messages** — Help and version queries validate supplied flags. Only explicitly declared flag names are accepted, and parsing errors support English and Chinese messages based on the configured language.
 - **File upload timeout** — Use `--timeout` to set the transfer timeout when uploading local files to temporary OSS storage.
 
+## [2.2.2] - 2026-10-14
+
+### Added
+
+- **Model compression lifecycle** — Added `bl finetune compress templates/create/list/get/watch/logs/cancel/delete` for quantizing supported full-parameter fine-tuned models, including dry-run previews, status polling, log retrieval, and safe deletion confirmation.
+- **Pi agent support** — `bl config agent` and `bl mcp connect` / `bl mcp disconnect` now support Pi, writing its native configuration while preserving existing settings and redacting credentials in output.
+- **Global flags before the command path** — Global flags such as `--config` can now precede the command (`bl --config token-plan auth status`) as well as follow it; both select the Profile for that invocation only, without changing `active_config`. Command-specific and credential flags must still follow the command path.
+- **Decision model inference** — Added `bl decision` for TypeSafe System One requests against workspace-scoped endpoints, including `choice` / `noul` / `score` validation, JSON and text output, training-sample reuse, Beijing/Singapore routing, and dry-run previews.
+
+### Changed
+
+- **Decision dataset validation** — Aligned the `decision` schema with the current API: `state` may be a string, object, or array; `instructions` is optional; and question criteria now enforce API limits.
+- **Stricter flag validation** — Only flags defined for a command are accepted. Unknown flags, missing values, and invalid values now return usage exit code 2 even when the invocation would otherwise show help, and parser errors follow the Profile language. A bare root or command-group invocation still shows help and returns 0. `--timeout` uses the same positive-finite rule on the command line and in resolved settings.
+
+### Fixed
+
+- **Temporary file upload timeout** — Uploads of local files to temporary OSS storage now honor `--timeout` instead of a fixed 120-second limit.
+
 ## [2.2.1] - 2026-10-12
 
 ### Fixed
diff --git a/CHANGELOG.zh.md b/CHANGELOG.zh.md
index 36a480b0a..b804b4df6 100644
--- a/CHANGELOG.zh.md
+++ b/CHANGELOG.zh.md
@@ -20,6 +20,24 @@
 - **参数校验与提示** —— 帮助和版本查询会校验已传入的参数；仅接受命令支持的参数名，参数解析错误支持按配置语言显示中英文提示。
 - **文件上传超时** —— `--timeout` 可设置本地文件上传至临时 OSS 的传输超时时间。
 
+## [2.2.2] - 2026-10-14
+
+### 新增
+
+- **模型压缩完整生命周期** —— 新增 `bl finetune compress templates/create/list/get/watch/logs/cancel/delete`，支持对符合条件的全参调优模型执行量化压缩，并覆盖 dry-run 预览、状态轮询、日志查询和安全删除确认。
+- **支持 Pi Agent** —— `bl config agent` 与 `bl mcp connect` / `bl mcp disconnect` 新增支持 Pi，写入其原生配置并保留已有设置，输出中的凭据会脱敏。
+- **全局参数可前置** —— `--config` 等全局参数现在既可放在命令之后，也可放在命令之前（`bl --config token-plan auth status`）；两种写法都只对本次调用选择 Profile，不修改 `active_config`。命令专属参数和凭证域参数仍须放在命令路径之后。
+- **决策模型推理** —— 新增 `bl decision`，通过工作空间专属 Endpoint 调用 TypeSafe System One，支持 `choice` / `noul` / `score` 校验、JSON 与文本输出、复用训练样本、北京/新加坡地域路由及 dry-run 预览。
+
+### 变更
+
+- **决策数据集校验** —— 与最新 API 对齐：`state` 可为字符串、对象或数组，`instructions` 改为可选，并按接口限制校验各类问题的 criteria。
+- **更严格的参数校验** —— 只接受命令已定义的参数。未知参数、缺少参数值或参数值非法时，即使本次调用原本会显示帮助，也返回用法错误退出码 2，解析错误提示跟随 Profile 语言；不带参数的根命令或命令组仍显示帮助并返回 0。`--timeout` 在命令行和最终生效配置中使用同一套“正有限数”校验规则。
+
+### 修复
+
+- **临时文件上传超时** —— 本地文件上传到临时 OSS 存储时会遵循 `--timeout`，不再固定为 120 秒。
+
 ## [2.2.1] - 2026-10-12
 
 ### 修复
diff --git a/README.md b/README.md
index 938a2a305..b35b055b0 100644
--- a/README.md
+++ b/README.md
@@ -121,6 +121,7 @@ Once installed, just describe your task to your AI Agent — no need to assemble
 | Monitoring & alerts      | "Show my model call stats, failures and logs, and create an alert rule."                                                              |
 | Throughput reservations  | "List my throughput reservations and capacity instances, then scale, renew or release capacity and wait for the operation to finish." |
 | Model selection          | "Recommend a model for image understanding and customer support."                                                                     |
+| Decision model           | "Use the Bailian decision model to route this ticket and assess whether it needs escalation."                                         |
 | About Bailian CLI        | "Tell me what Bailian CLI can do for me, and suggest how to use it for my needs."                                                     |
 
 > More examples and scenarios: [Aliyun Model Studio CLI Site](https://bailian.console.aliyun.com/cli?spm=a2ty07.github)
diff --git a/README.zh.md b/README.zh.md
index 08aba3dad..4cdebf186 100644
--- a/README.zh.md
+++ b/README.zh.md
@@ -120,6 +120,7 @@ irm https://bailian.aliyun.com/cli/install.ps1 | iex
 | 监控与告警       | “查看我的模型调用统计、失败明细和调用日志，并创建一条告警规则。”             |
 | 吞吐预留         | “查看我的吞吐预留及容量实例，并对容量进行扩缩、续订或释放，再等待操作完成。” |
 | 模型选型         | “推荐一个适合图片理解和智能客服的模型。”                                     |
+| 决策模型         | “用百炼决策模型给这张工单分流，并判断是否需要升级处理。”                     |
 | 了解 Bailian CLI | “介绍一下 Bailian CLI 能帮我完成哪些任务，并根据我的需求推荐使用方式。”      |
 
 > 更多案例与使用场景：[阿里云百炼 CLI 官方主页](https://bailian.console.aliyun.com/cli?spm=a2ty07.github)
diff --git a/packages/cli/README.md b/packages/cli/README.md
index 938a2a305..b35b055b0 100644
--- a/packages/cli/README.md
+++ b/packages/cli/README.md
@@ -121,6 +121,7 @@ Once installed, just describe your task to your AI Agent — no need to assemble
 | Monitoring & alerts      | "Show my model call stats, failures and logs, and create an alert rule."                                                              |
 | Throughput reservations  | "List my throughput reservations and capacity instances, then scale, renew or release capacity and wait for the operation to finish." |
 | Model selection          | "Recommend a model for image understanding and customer support."                                                                     |
+| Decision model           | "Use the Bailian decision model to route this ticket and assess whether it needs escalation."                                         |
 | About Bailian CLI        | "Tell me what Bailian CLI can do for me, and suggest how to use it for my needs."                                                     |
 
 > More examples and scenarios: [Aliyun Model Studio CLI Site](https://bailian.console.aliyun.com/cli?spm=a2ty07.github)
diff --git a/packages/cli/README.zh.md b/packages/cli/README.zh.md
index 08aba3dad..4cdebf186 100644
--- a/packages/cli/README.zh.md
+++ b/packages/cli/README.zh.md
@@ -120,6 +120,7 @@ irm https://bailian.aliyun.com/cli/install.ps1 | iex
 | 监控与告警       | “查看我的模型调用统计、失败明细和调用日志，并创建一条告警规则。”             |
 | 吞吐预留         | “查看我的吞吐预留及容量实例，并对容量进行扩缩、续订或释放，再等待操作完成。” |
 | 模型选型         | “推荐一个适合图片理解和智能客服的模型。”                                     |
+| 决策模型         | “用百炼决策模型给这张工单分流，并判断是否需要升级处理。”                     |
 | 了解 Bailian CLI | “介绍一下 Bailian CLI 能帮我完成哪些任务，并根据我的需求推荐使用方式。”      |
 
 > 更多案例与使用场景：[阿里云百炼 CLI 官方主页](https://bailian.console.aliyun.com/cli?spm=a2ty07.github)
diff --git a/packages/cli/src/commands.ts b/packages/cli/src/commands.ts
index 2a8749489..25c26bd24 100644
--- a/packages/cli/src/commands.ts
+++ b/packages/cli/src/commands.ts
@@ -14,6 +14,7 @@ import {
   videoTaskGet,
   videoDownload,
   visionDescribe,
+  decisionPredict,
   configShow,
   configSet,
   configList,
@@ -154,6 +155,14 @@ import {
   finetuneWatch,
   finetuneCapability,
   finetunePrice,
+  finetuneCompressTemplates,
+  finetuneCompressCreate,
+  finetuneCompressList,
+  finetuneCompressGet,
+  finetuneCompressWatch,
+  finetuneCompressLogs,
+  finetuneCompressCancel,
+  finetuneCompressDelete,
   deployTextCreate,
   deployAudioCreate,
   deployImageCreate,
@@ -298,6 +307,7 @@ export const commands: Record<string, AnyCommand> = {
   "video task get": videoTaskGet,
   "video download": videoDownload,
   "vision describe": visionDescribe,
+  decision: decisionPredict,
   "config show": configShow,
   "config set": configSet,
   "config list": configList,
@@ -440,6 +450,14 @@ export const commands: Record<string, AnyCommand> = {
   "finetune watch": finetuneWatch,
   "finetune capability": finetuneCapability,
   "finetune price": finetunePrice,
+  "finetune compress templates": finetuneCompressTemplates,
+  "finetune compress create": finetuneCompressCreate,
+  "finetune compress list": finetuneCompressList,
+  "finetune compress get": finetuneCompressGet,
+  "finetune compress watch": finetuneCompressWatch,
+  "finetune compress logs": finetuneCompressLogs,
+  "finetune compress cancel": finetuneCompressCancel,
+  "finetune compress delete": finetuneCompressDelete,
   "deploy text create": deployTextCreate,
   "deploy audio create": deployAudioCreate,
   "deploy image create": deployImageCreate,
diff --git a/packages/commands/src/commands/dataset/upload.ts b/packages/commands/src/commands/dataset/upload.ts
index 48677ba2c..d22284cba 100644
--- a/packages/commands/src/commands/dataset/upload.ts
+++ b/packages/commands/src/commands/dataset/upload.ts
@@ -6,6 +6,7 @@ import {
   formatIssue,
   MAX_DATASET_BYTES,
   MAX_CPT_BYTES,
+  MAX_DECISION_DATASET_BYTES,
   MAX_MEDIA_ZIP_BYTES,
   BailianError,
   ExitCode,
@@ -18,9 +19,10 @@ const UPLOAD_FLAGS = {
     type: "string",
     valueHint: "<path>",
     description: {
-      "en-US": "Local dataset file (.jsonl or .zip; ≤200MB SFT/DPO, ≤300MB CPT, ≤2GB media zip)",
+      "en-US":
+        "Local dataset file (.jsonl or .zip; ≤200MB SFT/DPO, ≤300MB CPT/decision, ≤2GB media zip)",
       "zh-CN":
-        "本地数据集文件（.jsonl 或 .zip；SFT/DPO 不超过 200MB，CPT 不超过 300MB，媒体 ZIP 不超过 2GB）",
+        "本地数据集文件（.jsonl 或 .zip；SFT/DPO 不超过 200MB，CPT/决策模型不超过 300MB，媒体 ZIP 不超过 2GB）",
     },
     required: true,
   },
@@ -37,9 +39,9 @@ const UPLOAD_FLAGS = {
     valueHint: "<s>",
     description: {
       "en-US":
-        'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record.',
+        'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "decision" (state/questions), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record.',
       "zh-CN":
-        '记录 Schema："chatml"（SFT）、"dpo"（chosen/rejected）、"cpt"（原始文本）、"tts"（音频）、"image"（图片生成）或 "video"（视频生成）。默认逐条自动识别。',
+        '记录 Schema："chatml"（SFT）、"dpo"（chosen/rejected）、"cpt"（原始文本）、"decision"（state/questions 决策数据）、"tts"（音频）、"image"（图片生成）或 "video"（视频生成）。默认逐条自动识别。',
     },
   },
   noValidate: {
@@ -65,12 +67,13 @@ export default defineCommand({
   },
   auth: "apiKey",
   usageArgs:
-    "--file <path> [--purpose <name>] [--schema <chatml|dpo|cpt|tts|image|video>] [--no-validate] [--full-validate]",
+    "--file <path> [--purpose <name>] [--schema <chatml|dpo|cpt|decision|tts|image|video>] [--no-validate] [--full-validate]",
   flags: UPLOAD_FLAGS,
   exampleArgs: [
     "--file train.jsonl",
     "--file dpo.jsonl --schema dpo",
     "--file cpt.jsonl --schema cpt",
+    "--file decision.jsonl --schema decision",
     "--file audio.zip --schema tts",
     "--file eval.jsonl --purpose evaluation",
     "--file train.jsonl --full-validate",
@@ -79,21 +82,21 @@ export default defineCommand({
   notes: [
     {
       "en-US":
-        'Supports .jsonl (text) and .zip (audio/image/video archives with a data.jsonl manifest). Six record schemas are recognized: chatml = {messages:[...]} (SFT); dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."} (continual pre-training, raw text); tts = {wav_fn:"train/xxx.wav", text:"..."} (audio fine-tuning); image = {img_path:"..."} (image generation); video = {first_frame_path:...} (video generation).',
+        'Supports .jsonl (text/decision) and .zip (audio/image/video archives with a data.jsonl manifest). Seven record schemas are recognized: chatml = {messages:[...]} (SFT); dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."}; decision = {state:{content}, questions:{...}}; tts = {wav_fn:"train/xxx.wav", text:"..."}; image = {img_path:"..."}; video = {first_frame_path:...}.',
       "zh-CN":
-        '支持 .jsonl（文本）和 .zip（包含 data.jsonl 清单的音频、图片或视频归档）。可识别六种记录 Schema：chatml = {messages:[...]}（SFT）；dpo = {messages:[...], chosen, rejected}；cpt = {text:"..."}（持续预训练，原始文本）；tts = {wav_fn:"train/xxx.wav", text:"..."}（音频微调）；image = {img_path:"..."}（图片生成）；video = {first_frame_path:...}（视频生成）。',
+        '支持 .jsonl（文本/决策）和 .zip（包含 data.jsonl 清单的音频、图片或视频归档）。可识别七种记录 Schema：chatml = {messages:[...]}（SFT）；dpo = {messages:[...], chosen, rejected}；cpt = {text:"..."}；decision = {state:{content}, questions:{...}}；tts = {wav_fn:"train/xxx.wav", text:"..."}；image = {img_path:"..."}；video = {first_frame_path:...}。',
     },
     {
       "en-US":
-        "With no --schema, a record carrying wav_fn is validated as TTS, img_path as image, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML.",
+        "With no --schema, state/questions is validated as decision data, wav_fn as TTS, img_path as image, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML.",
       "zh-CN":
-        "未指定 --schema 时，包含 wav_fn 的记录按 TTS 验证，包含 img_path 的按 image 验证，包含 chosen/rejected 的按 DPO 验证，仅含 text（无 messages）的按 CPT 验证，其他记录按 ChatML 验证。",
+        "未指定 --schema 时，state/questions 记录按 decision 验证，包含 wav_fn 的按 TTS 验证，包含 img_path 的按 image 验证，包含 chosen/rejected 的按 DPO 验证，仅含 text（无 messages）的按 CPT 验证，其他记录按 ChatML 验证。",
     },
     {
       "en-US":
-        "Upload cap: 200MB SFT/DPO text, 300MB CPT, 2GB media zip. Upload uses the OpenAI-compatible /compatible-mode/v1/files endpoint so the purpose tag is persisted (the DashScope-native /api/v1/files drops it).",
+        "Upload cap: 200MB SFT/DPO text, 300MB CPT/decision data, 2GB media zip. For decision files over 200MB, pass --schema decision so the correct cap is applied. Upload uses the OpenAI-compatible /compatible-mode/v1/files endpoint so the purpose tag is persisted.",
       "zh-CN":
-        "上传上限：SFT/DPO 文本 200MB、CPT 300MB、媒体 ZIP 2GB。上传使用 OpenAI 兼容的 /compatible-mode/v1/files Endpoint，以便保留 purpose 标签；DashScope 原生 /api/v1/files 会丢弃该标签。",
+        "上传上限：SFT/DPO 文本 200MB、CPT/决策模型数据 300MB、媒体 ZIP 2GB。决策数据超过 200MB 时请显式传入 --schema decision，以应用正确上限。上传使用 OpenAI 兼容的 /compatible-mode/v1/files 接口，以确保持久化 purpose 标签。",
     },
   ],
   async run(ctx) {
@@ -101,13 +104,15 @@ export default defineCommand({
     const filePath = flags.file;
     const purpose = flags.purpose || "fine-tune";
     const schema = parseDatasetSchemaFlag(flags.schema);
-    // Size caps differ per training type: SFT/DPO 200MB, CPT 300MB, media ZIP 2GB.
+    // 大小上限按平台数据类型区分；自动识别仍采用保守的文本上限，避免放宽已有格式。
     const isMediaSchema = schema === "image" || schema === "video";
     const maxBytes = isMediaSchema
       ? MAX_MEDIA_ZIP_BYTES
-      : schema === "cpt"
-        ? MAX_CPT_BYTES
-        : MAX_DATASET_BYTES;
+      : schema === "decision"
+        ? MAX_DECISION_DATASET_BYTES
+        : schema === "cpt"
+          ? MAX_CPT_BYTES
+          : MAX_DATASET_BYTES;
 
     if (!flags.noValidate) {
       const result = await validateDataset(filePath, {
diff --git a/packages/commands/src/commands/dataset/validate.ts b/packages/commands/src/commands/dataset/validate.ts
index 055a21037..fa34fcff5 100644
--- a/packages/commands/src/commands/dataset/validate.ts
+++ b/packages/commands/src/commands/dataset/validate.ts
@@ -2,6 +2,7 @@ import {
   defineCommand,
   validateDataset,
   parseDatasetSchemaFlag,
+  MAX_DECISION_DATASET_BYTES,
   BailianError,
   ExitCode,
   type FlagsDef,
@@ -30,9 +31,9 @@ const VALIDATE_FLAGS = {
     valueHint: "<s>",
     description: {
       "en-US":
-        'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record.',
+        'Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "decision" (state/questions), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record.',
       "zh-CN":
-        '记录 Schema："chatml"（SFT）、"dpo"（chosen/rejected）、"cpt"（原始文本）、"tts"（音频）、"image"（图片生成）或 "video"（视频生成）。默认逐条自动识别。',
+        '记录 Schema："chatml"（SFT）、"dpo"（chosen/rejected）、"cpt"（原始文本）、"decision"（state/questions 决策数据）、"tts"（音频）、"image"（图片生成）或 "video"（视频生成）。默认逐条自动识别。',
     },
   },
 } satisfies FlagsDef;
@@ -44,12 +45,13 @@ export default defineCommand({
   },
   // 纯本地校验，不触网、不需 API key（与 `pipeline validate` 一致）。
   auth: "none",
-  usageArgs: "--file <path> [--full-validate] [--schema <chatml|dpo|cpt|tts|image|video>]",
+  usageArgs: "--file <path> [--full-validate] [--schema <chatml|dpo|cpt|decision|tts|image|video>]",
   flags: VALIDATE_FLAGS,
   exampleArgs: [
     "--file train.jsonl",
     "--file dpo.jsonl --schema dpo",
     "--file cpt.jsonl --schema cpt",
+    "--file decision.jsonl --schema decision",
     "--file audio.zip --schema tts",
     "--file wan-i2v-training-dataset.zip --schema video",
     "--file eval.jsonl --full-validate",
@@ -64,15 +66,15 @@ export default defineCommand({
     },
     {
       "en-US":
-        'Schemas: chatml = {messages:[...]} (SFT); dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."} (continual pre-training, raw text); tts = {wav_fn:"train/xxx.wav", text:"..."} (audio fine-tuning); image = {img_path:"..."} (image generation); video = {first_frame_path:"...", video_path:"..."} (video generation, i2v first-frame or kf2v first+last-frame with last_frame_path).',
+        'Schemas: chatml = {messages:[...]}; dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."}; decision = {state:{content}, questions:{...}}; tts = {wav_fn:"train/xxx.wav", text:"..."}; image = {img_path:"..."}; video = {first_frame_path:"...", video_path:"..."}.',
       "zh-CN":
-        'Schema：chatml = {messages:[...]}（SFT）；dpo = {messages:[...], chosen, rejected}；cpt = {text:"..."}（持续预训练，原始文本）；tts = {wav_fn:"train/xxx.wav", text:"..."}（音频微调）；image = {img_path:"..."}（图片生成）；video = {first_frame_path:"...", video_path:"..."}（视频生成，支持 i2v 首帧或通过 last_frame_path 指定 kf2v 首尾帧）。',
+        'Schema：chatml = {messages:[...]}；dpo = {messages:[...], chosen, rejected}；cpt = {text:"..."}；decision = {state:{content}, questions:{...}}；tts = {wav_fn:"train/xxx.wav", text:"..."}；image = {img_path:"..."}；video = {first_frame_path:"...", video_path:"..."}。',
     },
     {
       "en-US":
-        "With no --schema, a record carrying wav_fn is validated as TTS, img_path as image, first_frame_path/video_path as video, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML. Pass --schema to require a specific shape on every record.",
+        "With no --schema, state/questions is validated as decision data, wav_fn as TTS, img_path as image, first_frame_path/video_path as video, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML.",
       "zh-CN":
-        "未指定 --schema 时，包含 wav_fn 的记录按 TTS 验证，包含 img_path 的按 image 验证，包含 first_frame_path/video_path 的按 video 验证，包含 chosen/rejected 的按 DPO 验证，仅含 text（无 messages）的按 CPT 验证，其他记录按 ChatML 验证。使用 --schema 要求每条记录符合指定结构。",
+        "未指定 --schema 时，state/questions 记录按 decision 验证，包含 wav_fn 的按 TTS 验证，包含 img_path 的按 image 验证，包含 first_frame_path/video_path 的按 video 验证，包含 chosen/rejected 的按 DPO 验证，仅含 text（无 messages）的按 CPT 验证，其他记录按 ChatML 验证。",
     },
     {
       "en-US":
@@ -98,7 +100,11 @@ export default defineCommand({
       return;
     }
 
-    const result = await validateDataset(filePath, { fullValidate: flags.fullValidate, schema });
+    const result = await validateDataset(filePath, {
+      fullValidate: flags.fullValidate,
+      schema,
+      maxBytes: schema === "decision" ? MAX_DECISION_DATASET_BYTES : undefined,
+    });
 
     if (settings.quiet) {
       emitBare(result.valid ? "ok" : "fail");
diff --git a/packages/commands/src/commands/decision/predict.ts b/packages/commands/src/commands/decision/predict.ts
new file mode 100644
index 000000000..509a264e4
--- /dev/null
+++ b/packages/commands/src/commands/decision/predict.ts
@@ -0,0 +1,178 @@
+import {
+  defineCommand,
+  detectOutputFormat,
+  DECISION_REGIONS,
+  DEFAULT_DECISION_MODEL,
+  DEFAULT_DECISION_REGION,
+  type DecisionRequest,
+  type DecisionResponse,
+} from "bailian-cli-core";
+import { emitBare, emitRequestId, emitResult } from "bailian-cli-runtime";
+import {
+  buildDecisionInput,
+  decisionInputFlagsError,
+  formatDecisionAnswers,
+  formatDecisionStats,
+  resolveDecisionEndpoint,
+} from "./shared.ts";
+
+export default defineCommand({
+  description: {
+    "en-US": "Run the decision model: typed choice / noul / score answers with probabilities",
+    "zh-CN": "调用决策模型：一次返回 choice / noul / score 判定结果及概率分布",
+  },
+  auth: "apiKey",
+  usageArgs:
+    "(--state <text> | --state-file <path>) (--questions <json> | --questions-file <path>) | --request-file <path>",
+  flags: {
+    model: {
+      type: "string",
+      valueHint: "<model>",
+      description: {
+        "en-US": `Decision model ID (default: ${DEFAULT_DECISION_MODEL})`,
+        "zh-CN": `决策模型 ID（默认：${DEFAULT_DECISION_MODEL}）`,
+      },
+    },
+    state: {
+      type: "string",
+      valueHint: "<text>",
+      description: {
+        "en-US": "Business state to decide on, sent as a string",
+        "zh-CN": "待决策的业务状态，按字符串发送",
+      },
+    },
+    stateFile: {
+      type: "string",
+      valueHint: "<path>",
+      description: {
+        "en-US":
+          "Read state from a file (.json is parsed as an object/array, others as text; - for stdin)",
+        "zh-CN": "从文件读取 state（.json 解析为对象/数组，其他按文本读取；- 表示 stdin）",
+      },
+    },
+    questions: {
+      type: "string",
+      valueHint: "<json>",
+      description: {
+        "en-US": "Question table JSON, keyed by question ID",
+        "zh-CN": "问题表 JSON，键为问题 ID",
+      },
+    },
+    questionsFile: {
+      type: "string",
+      valueHint: "<path>",
+      description: {
+        "en-US": 'Read the question table from a JSON file ({...} or {"questions":{...}})',
+        "zh-CN": '从 JSON 文件读取问题表（{...} 或 {"questions":{...}}）',
+      },
+    },
+    requestFile: {
+      type: "string",
+      valueHint: "<path>",
+      description: {
+        "en-US": "Full request JSON {state, questions}; label/target are stripped",
+        "zh-CN": "完整请求 JSON {state, questions}；自动去掉 label/target",
+      },
+    },
+    workspaceId: {
+      type: "string",
+      valueHint: "<id>",
+      description: {
+        "en-US": "Workspace ID for the endpoint host (or set BAILIAN_WORKSPACE_ID)",
+        "zh-CN": "用于拼接请求域名的 Workspace ID（也可设置 BAILIAN_WORKSPACE_ID）",
+      },
+    },
+    region: {
+      type: "string",
+      valueHint: "<region>",
+      choices: DECISION_REGIONS,
+      description: {
+        "en-US": `Endpoint region: ${DECISION_REGIONS.join(", ")} (default: ${DEFAULT_DECISION_REGION})`,
+        "zh-CN": `请求地域：${DECISION_REGIONS.join("、")}（默认：${DEFAULT_DECISION_REGION}）`,
+      },
+    },
+  },
+  exampleArgs: [
+    {
+      "en-US":
+        '--workspace-id <id> --state "Is this review asking for a refund?" --questions \'{"refund":{"type":"noul","instructions":"Asking for a refund?"}}\'',
+      "zh-CN":
+        '--workspace-id <id> --state "这段评论是否在索要退款？" --questions \'{"refund":{"type":"noul","instructions":"是否在索要退款？"}}\'',
+    },
+    "--state-file ticket.json --questions-file workload.json",
+    "--request-file sample.json --model decision-model-preview",
+    "--region ap-southeast-1 --request-file sample.json",
+    "--request-file sample.json --dry-run",
+  ],
+  notes: [
+    {
+      "en-US":
+        "Endpoint: configured --base-url, DASHSCOPE_BASE_URL, or profile base_url takes precedence; otherwise https://{workspace}.{region}.maas.aliyuncs.com/compatible-mode/v1/systemone, with workspace from --workspace-id, BAILIAN_WORKSPACE_ID, then config workspace_id.",
+      "zh-CN":
+        "请求地址：已配置的 --base-url、DASHSCOPE_BASE_URL 或 Profile base_url 优先；否则为 https://{workspace}.{region}.maas.aliyuncs.com/compatible-mode/v1/systemone，工作空间按 --workspace-id、BAILIAN_WORKSPACE_ID、配置项 workspace_id 解析。",
+    },
+    {
+      "en-US":
+        "Question types: choice (criteria: option → description, 1-255), noul (optional criteria with true/false descriptions), score (criteria: levels from low to high, 2-255; 3-7 recommended). Keep questions per request at 16 or fewer; latency grows roughly linearly.",
+      "zh-CN":
+        "问题类型：choice（criteria 为选项 → 描述，1-255 项）、noul（可选 criteria，含 true/false 描述）、score（criteria 为从低到高的等级，2-255 级，建议 3-7 级）。建议每次请求不超过 16 个问题，延迟随问题数近线性增长。",
+    },
+    {
+      "en-US":
+        "Output defaults to the full JSON response. Use --output text for one human-readable line per question; the request ID and usage summary are written to stderr.",
+      "zh-CN":
+        "默认输出完整 JSON 响应。使用 --output text 可按问题逐行输出易读结果；请求 ID 与用量摘要写入 stderr。",
+    },
+    {
+      "en-US":
+        "Training samples can be reused directly: label/target are stripped before sending. Sampling parameters (temperature, max_tokens, ...) are not supported by this API.",
+      "zh-CN":
+        "训练样本可直接复用：发送前自动去掉 label/target。该接口不支持 temperature、max_tokens 等采样参数。",
+    },
+  ],
+  validate: (flags) => {
+    const inputError = decisionInputFlagsError(flags);
+    if (typeof inputError === "string" || !inputError) return inputError;
+    return `${inputError["en-US"]} / ${inputError["zh-CN"]}`;
+  },
+  async run(ctx) {
+    const { settings, flags } = ctx;
+    const input = buildDecisionInput(flags, ctx.localize);
+    const body: DecisionRequest = {
+      model: flags.model || input.model || DEFAULT_DECISION_MODEL,
+      state: input.state,
+      questions: input.questions,
+    };
+    const endpoint = resolveDecisionEndpoint({
+      workspaceId: flags.workspaceId || settings.workspaceId,
+      region: flags.region ?? DEFAULT_DECISION_REGION,
+      binName: ctx.identity.binName,
+      client: ctx.client,
+      localize: ctx.localize,
+    });
+    const format = settings.outputExplicit ? detectOutputFormat(settings.output) : "json";
+
+    if (settings.dryRun) {
+      emitResult({ endpoint, request: body }, format);
+      return;
+    }
+
+    const response = await ctx.client.requestJson<DecisionResponse>({
+      path: endpoint,
+      method: "POST",
+      body,
+    });
+
+    if (format !== "text") {
+      emitResult(response, format);
+      return;
+    }
+
+    emitBare(formatDecisionAnswers(response));
+    if (!settings.quiet) {
+      const stats = formatDecisionStats(response);
+      if (stats) process.stderr.write(`${stats}\n`);
+    }
+    emitRequestId(response.request_id, settings.quiet);
+  },
+});
diff --git a/packages/commands/src/commands/decision/shared.ts b/packages/commands/src/commands/decision/shared.ts
new file mode 100644
index 000000000..c2d4a83f1
--- /dev/null
+++ b/packages/commands/src/commands/decision/shared.ts
@@ -0,0 +1,270 @@
+import { extname } from "path";
+import {
+  BailianError,
+  ExitCode,
+  decisionPath,
+  inspectDecisionQuestions,
+  inspectDecisionState,
+  isDecisionObject,
+  readTextFromPathOrStdin,
+  stripDecisionTrainingFields,
+  unwrapDecisionQuestions,
+  workspaceMaaSBaseUrl,
+  type Client,
+  type DecisionAnswer,
+  type DecisionQuestion,
+  type DecisionResponse,
+  type DecisionState,
+  type LocalizedText,
+} from "bailian-cli-core";
+
+type Localize = (text: LocalizedText) => string;
+
+export interface DecisionInputFlags {
+  state?: string;
+  stateFile?: string;
+  questions?: string;
+  questionsFile?: string;
+  requestFile?: string;
+}
+
+export interface DecisionInput {
+  state: DecisionState;
+  questions: Record<string, DecisionQuestion>;
+  /** `--request-file` 内携带的 model，`--model` 未指定时使用。 */
+  model?: string;
+}
+
+function usageError(localize: Localize, text: LocalizedText, hint?: LocalizedText): BailianError {
+  return new BailianError(localize(text), ExitCode.USAGE, hint ? localize(hint) : undefined);
+}
+
+function parseJson(raw: string, source: string, localize: Localize): unknown {
+  try {
+    return JSON.parse(raw);
+  } catch (error) {
+    const reason = error instanceof Error ? error.message : String(error);
+    throw usageError(localize, {
+      "en-US": `${source} is not valid JSON: ${reason}`,
+      "zh-CN": `${source} 不是合法的 JSON：${reason}`,
+    });
+  }
+}
+
+function countProvided(...values: Array<string | undefined>): number {
+  return values.filter((value) => value !== undefined).length;
+}
+
+/** 返回输入来源组合错误；供 command validate 在鉴权前执行。 */
+export function decisionInputFlagsError(flags: DecisionInputFlags): LocalizedText | undefined {
+  if (flags.requestFile !== undefined) {
+    if (countProvided(flags.state, flags.stateFile, flags.questions, flags.questionsFile) > 0) {
+      return {
+        "en-US":
+          "--request-file cannot be combined with --state, --state-file, --questions, or --questions-file.",
+        "zh-CN":
+          "--request-file 不能与 --state、--state-file、--questions、--questions-file 同时使用。",
+      };
+    }
+    return undefined;
+  }
+
+  const stateSources = countProvided(flags.state, flags.stateFile);
+  if (stateSources !== 1) {
+    return {
+      "en-US":
+        stateSources === 0
+          ? "Provide --state, --state-file, or --request-file."
+          : "Use only one of --state or --state-file.",
+      "zh-CN":
+        stateSources === 0
+          ? "请提供 --state、--state-file 或 --request-file。"
+          : "--state 与 --state-file 只能选择一个。",
+    };
+  }
+
+  const questionSources = countProvided(flags.questions, flags.questionsFile);
+  if (questionSources !== 1) {
+    return {
+      "en-US":
+        questionSources === 0
+          ? "Provide --questions or --questions-file."
+          : "Use only one of --questions or --questions-file.",
+      "zh-CN":
+        questionSources === 0
+          ? "请提供 --questions 或 --questions-file。"
+          : "--questions 与 --questions-file 只能选择一个。",
+    };
+  }
+  return undefined;
+}
+
+/** 校验输入来源组合：--request-file 独占；否则 state 与 questions 各选一种来源。 */
+export function validateDecisionInputFlags(flags: DecisionInputFlags, localize: Localize): void {
+  const inputError = decisionInputFlagsError(flags);
+  if (inputError) throw usageError(localize, inputError);
+}
+
+function readStateFile(path: string, localize: Localize): unknown {
+  const raw = readTextFromPathOrStdin(path);
+  return extname(path).toLowerCase() === ".json" ? parseJson(raw, path, localize) : raw;
+}
+
+function readRawInput(
+  flags: DecisionInputFlags,
+  localize: Localize,
+): { state: unknown; questions: unknown; model?: string } {
+  if (flags.requestFile !== undefined) {
+    const parsed = parseJson(
+      readTextFromPathOrStdin(flags.requestFile),
+      flags.requestFile,
+      localize,
+    );
+    if (!isDecisionObject(parsed)) {
+      throw usageError(localize, {
+        "en-US": `${flags.requestFile} must be a JSON object with "state" and "questions".`,
+        "zh-CN": `${flags.requestFile} 必须是包含 "state" 和 "questions" 的 JSON 对象。`,
+      });
+    }
+    return {
+      state: parsed.state,
+      questions: parsed.questions,
+      model: typeof parsed.model === "string" && parsed.model ? parsed.model : undefined,
+    };
+  }
+
+  const state =
+    flags.stateFile !== undefined ? readStateFile(flags.stateFile, localize) : flags.state;
+  const questions =
+    flags.questionsFile !== undefined
+      ? parseJson(readTextFromPathOrStdin(flags.questionsFile), flags.questionsFile, localize)
+      : parseJson(flags.questions ?? "", "--questions", localize);
+  return { state, questions: unwrapDecisionQuestions(questions) };
+}
+
+/**
+ * 按 flag 组装推理输入：读取文件、解包问题文件、去掉训练字段（label / target），
+ * 再按决策模型 API 规则做本地结构校验；不通过时抛 USAGE 错误，不发请求。
+ */
+export function buildDecisionInput(flags: DecisionInputFlags, localize: Localize): DecisionInput {
+  validateDecisionInputFlags(flags, localize);
+  const raw = readRawInput(flags, localize);
+  const questions = stripDecisionTrainingFields(raw.questions);
+
+  const issues = [...inspectDecisionState(raw.state), ...inspectDecisionQuestions(questions)];
+  if (issues.length > 0) {
+    throw usageError(
+      localize,
+      {
+        "en-US": "Invalid decision request:",
+        "zh-CN": "决策请求不合法：",
+      },
+      {
+        "en-US": issues.map((issue) => `- ${issue.message["en-US"]}`).join("\n"),
+        "zh-CN": issues.map((issue) => `- ${issue.message["zh-CN"]}`).join("\n"),
+      },
+    );
+  }
+
+  return {
+    state: raw.state as DecisionState,
+    questions: questions as Record<string, DecisionQuestion>,
+    model: raw.model,
+  };
+}
+
+/**
+ * 解析请求地址：显式配置的 Base URL（--base-url / DASHSCOPE_BASE_URL / Profile base_url）优先；
+ * 否则按 --workspace-id → BAILIAN_WORKSPACE_ID → 配置项 workspace_id 拼 workspace 域名。
+ */
+export function resolveDecisionEndpoint(ctx: {
+  workspaceId?: string;
+  region: string;
+  binName: string;
+  client: Pick<Client, "url">;
+  localize: Localize;
+}): string {
+  return ctx.client.url(decisionPath(), () => {
+    if (!ctx.workspaceId) {
+      throw usageError(
+        ctx.localize,
+        {
+          "en-US": "Workspace ID is required for the decision model endpoint.",
+          "zh-CN": "调用决策模型需要 Workspace ID。",
+        },
+        {
+          "en-US": `Pass --workspace-id, set BAILIAN_WORKSPACE_ID env, or configure: ${ctx.binName} config set workspace_id <id>. You can also override the endpoint with --base-url.`,
+          "zh-CN": `请传入 --workspace-id、设置环境变量 BAILIAN_WORKSPACE_ID，或执行：${ctx.binName} config set workspace_id <id>。也可以用 --base-url 覆盖请求地址。`,
+        },
+      );
+    }
+    return workspaceMaaSBaseUrl(ctx.workspaceId, ctx.region);
+  });
+}
+
+function formatProbability(value: unknown): string {
+  return typeof value === "number" ? String(Number(value.toFixed(4))) : String(value);
+}
+
+function formatProbabilities(probabilities: Record<string, number> | undefined): string {
+  if (!probabilities) return "";
+  return Object.entries(probabilities)
+    .map(([optionName, probability]) => `${optionName}=${formatProbability(probability)}`)
+    .join(" ");
+}
+
+function topLevel(answer: DecisionAnswer): string | undefined {
+  const entries = Object.entries(answer.probabilities ?? {});
+  if (entries.length === 0) return undefined;
+  const [levelKey] = entries.reduce((best, entry) => (entry[1] > best[1] ? entry : best));
+  const description = answer.legend?.[levelKey];
+  return description ? `${levelKey} ${description}` : levelKey;
+}
+
+function formatAnswer(answer: DecisionAnswer): string {
+  if (answer.type === "choice" && answer.choice !== undefined) {
+    return [
+      `choice=${answer.choice}`,
+      answer.confidence !== undefined ? `confidence=${formatProbability(answer.confidence)}` : "",
+      formatProbabilities(answer.probabilities),
+    ]
+      .filter(Boolean)
+      .join("  ");
+  }
+  if (answer.type === "noul" && answer.noul !== undefined) {
+    return `noul P(yes)=${formatProbability(answer.noul)}`;
+  }
+  if (answer.type === "score" && answer.score !== undefined) {
+    const level = topLevel(answer);
+    return [
+      `score=${formatProbability(answer.score)}`,
+      answer.confidence !== undefined ? `confidence=${formatProbability(answer.confidence)}` : "",
+      level ? `top=${level}` : "",
+    ]
+      .filter(Boolean)
+      .join("  ");
+  }
+  return JSON.stringify(answer);
+}
+
+/** 文本模式：每个问题一行；未知结构回落为该答案的 JSON。 */
+export function formatDecisionAnswers(response: DecisionResponse): string {
+  const answers = Object.entries(response.answers ?? {});
+  if (answers.length === 0) return JSON.stringify(response);
+  const idWidth = Math.max(...answers.map(([questionId]) => questionId.length));
+  return answers
+    .map(
+      ([questionId, answer]) =>
+        `${questionId.padEnd(idWidth)}  ${isDecisionObject(answer) ? formatAnswer(answer) : JSON.stringify(answer)}`,
+    )
+    .join("\n");
+}
+
+/** 文本模式统计行（写 stderr）：input_tokens / latency_ms。 */
+export function formatDecisionStats(response: DecisionResponse): string | undefined {
+  const parts = [
+    response.usage?.input_tokens !== undefined ? `input_tokens=${response.usage.input_tokens}` : "",
+    response.latency_ms !== undefined ? `latency_ms=${response.latency_ms}` : "",
+  ].filter(Boolean);
+  return parts.length > 0 ? parts.join("  ") : undefined;
+}
diff --git a/packages/commands/src/commands/finetune/compress-cancel.ts b/packages/commands/src/commands/finetune/compress-cancel.ts
new file mode 100644
index 000000000..6d15ea310
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-cancel.ts
@@ -0,0 +1,27 @@
+import { defineCommand, cancelCompressJob } from "bailian-cli-core";
+import { emitBare, emitResult } from "bailian-cli-runtime";
+import { COMPRESS_JOB_ID_FLAGS } from "./compress-shared.ts";
+
+export default defineCommand({
+  description: { "en-US": "Cancel a compression job", "zh-CN": "取消压缩任务" },
+  auth: "apiKey",
+  usageArgs: "--job-id <id>",
+  flags: COMPRESS_JOB_ID_FLAGS,
+  exampleArgs: ["--job-id quant-xxx", "--job-id quant-xxx --dry-run"],
+  notes: [
+    {
+      "en-US": "Only PENDING, QUEUING, or RUNNING jobs can be cancelled.",
+      "zh-CN": "只有 PENDING、QUEUING 或 RUNNING 状态的任务可以取消。",
+    },
+  ],
+  async run(ctx) {
+    const jobId = ctx.flags.jobId;
+    if (ctx.settings.dryRun) {
+      emitResult({ action: "finetune.compress.cancel", job_id: jobId }, "json");
+      return;
+    }
+    const response = await cancelCompressJob(ctx.client, jobId);
+    if (ctx.settings.quiet) emitBare(jobId);
+    else emitResult(response, "json");
+  },
+});
diff --git a/packages/commands/src/commands/finetune/compress-create.ts b/packages/commands/src/commands/finetune/compress-create.ts
new file mode 100644
index 000000000..331b1a524
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-create.ts
@@ -0,0 +1,116 @@
+import {
+  defineCommand,
+  createCompressJob,
+  type CreateCompressJobRequest,
+  type FlagsDef,
+} from "bailian-cli-core";
+import { emitBare, emitResult } from "bailian-cli-runtime";
+import { parseCommaSeparated, parseCompressHyperParameters } from "./compress-shared.ts";
+
+const CREATE_FLAGS = {
+  model: {
+    type: "string",
+    valueHint: "<model>",
+    description: {
+      "en-US": "Source custom model ID (required)",
+      "zh-CN": "源自定义模型 ID（必填）",
+    },
+    required: true,
+  },
+  templateId: {
+    type: "string",
+    valueHint: "<id>",
+    description: { "en-US": "Compression template ID (required)", "zh-CN": "压缩模板 ID（必填）" },
+    required: true,
+  },
+  jobName: {
+    type: "string",
+    valueHint: "<name>",
+    description: {
+      "en-US": "Unique job name (max 50 characters)",
+      "zh-CN": "唯一任务名称（最多 50 字符）",
+    },
+  },
+  description: {
+    type: "string",
+    valueHint: "<text>",
+    description: {
+      "en-US": "Job description (max 200 characters)",
+      "zh-CN": "任务描述（最多 200 字符）",
+    },
+  },
+  hparam: {
+    type: "array",
+    valueHint: "<key=value>",
+    description: {
+      "en-US": "Hyper-parameter override; repeat for multiple values",
+      "zh-CN": "超参数覆盖；可重复传入多个值",
+    },
+  },
+  calibrationFiles: {
+    type: "string",
+    valueHint: "<file-id,...>",
+    description: {
+      "en-US": "Comma-separated calibration dataset file IDs",
+      "zh-CN": "逗号分隔的校准数据集文件 ID",
+    },
+  },
+  suffix: {
+    type: "string",
+    valueHint: "<suffix>",
+    description: {
+      "en-US": "Output model suffix (1–8 lowercase letters or digits)",
+      "zh-CN": "产出模型后缀（1–8 个小写字母或数字）",
+    },
+  },
+} satisfies FlagsDef;
+
+export default defineCommand({
+  description: { "en-US": "Create a model compression job", "zh-CN": "创建模型压缩任务" },
+  auth: "apiKey",
+  usageArgs: "--model <model> --template-id <id> [flags]",
+  flags: CREATE_FLAGS,
+  validate(flags) {
+    if (flags.jobName !== undefined && flags.jobName.length > 50)
+      return "--job-name must be at most 50 characters. / --job-name 最多 50 个字符。";
+    if (flags.description !== undefined && flags.description.length > 200)
+      return "--description must be at most 200 characters. / --description 最多 200 个字符。";
+    if (flags.suffix !== undefined && !/^[a-z0-9]{1,8}$/.test(flags.suffix))
+      return "--suffix must contain 1–8 lowercase letters or digits. / --suffix 必须为 1–8 个小写字母或数字。";
+    return undefined;
+  },
+  exampleArgs: [
+    "--model qwen3.5-flash-2026-02-23-ft-xxx --template-id quant-flash-nvfp4-mlp-nomtp",
+    "--model qwen3.5-flash-2026-02-23-ft-xxx --template-id quant-flash-nvfp4-mlp-nomtp --suffix test --hparam calib_input=true",
+  ],
+  notes: [
+    {
+      "en-US": "Run `finetune compress templates` first to get a compatible template ID.",
+      "zh-CN": "请先运行 `finetune compress templates` 获取兼容的模板 ID。",
+    },
+    {
+      "en-US": "Model compression APIs are currently available only in the Beijing region.",
+      "zh-CN": "模型压缩 API 当前仅在北京 Region 开放。",
+    },
+  ],
+  async run(ctx) {
+    const hyperParameters = parseCompressHyperParameters(ctx.flags.hparam);
+    const calibrationFileIds = parseCommaSeparated(ctx.flags.calibrationFiles);
+    const body: CreateCompressJobRequest = {
+      model: ctx.flags.model,
+      template_id: ctx.flags.templateId,
+      ...(ctx.flags.jobName ? { job_name: ctx.flags.jobName } : {}),
+      ...(ctx.flags.description ? { job_description: ctx.flags.description } : {}),
+      ...(Object.keys(hyperParameters).length ? { hyper_parameters: hyperParameters } : {}),
+      ...(calibrationFileIds ? { custom_calibration_file_ids: calibrationFileIds } : {}),
+      ...(ctx.flags.suffix ? { output_model_suffix: ctx.flags.suffix } : {}),
+    };
+    if (ctx.settings.dryRun) {
+      emitResult({ action: "finetune.compress.create", request: body }, "json");
+      return;
+    }
+    const response = await createCompressJob(ctx.client, body);
+    if (ctx.settings.quiet) emitBare(response.output?.job_id ?? "");
+    else emitResult(response, "json");
+  },
+});
diff --git a/packages/commands/src/commands/finetune/compress-delete.ts b/packages/commands/src/commands/finetune/compress-delete.ts
new file mode 100644
index 000000000..fe1bc14bd
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-delete.ts
@@ -0,0 +1,35 @@
+import { defineCommand, deleteCompressJob } from "bailian-cli-core";
+import { emitBare, emitResult } from "bailian-cli-runtime";
+import { COMPRESS_JOB_ID_FLAGS } from "./compress-shared.ts";
+
+export default defineCommand({
+  description: { "en-US": "Delete a compression job record", "zh-CN": "删除压缩任务记录" },
+  auth: "apiKey",
+  risk: {
+    level: "high",
+    message: {
+      "en-US": "This permanently deletes the compression job record and cannot be undone.",
+      "zh-CN": "该操作会永久删除压缩任务记录，且无法撤销。",
+    },
+  },
+  usageArgs: "--job-id <id>",
+  flags: COMPRESS_JOB_ID_FLAGS,
+  exampleArgs: ["--job-id quant-xxx", "--job-id quant-xxx --dry-run", "--job-id quant-xxx --yes"],
+  notes: [
+    {
+      "en-US":
+        "Only terminal jobs can be deleted. Deleting a job record does not delete its quantized output model.",
+      "zh-CN": "只有终态任务可以删除；删除任务记录不会删除其产出的量化模型。",
+    },
+  ],
+  async run(ctx) {
+    const jobId = ctx.flags.jobId;
+    if (ctx.settings.dryRun) {
+      emitResult({ action: "finetune.compress.delete", job_id: jobId }, "json");
+      return;
+    }
+    const response = await deleteCompressJob(ctx.client, jobId);
+    if (ctx.settings.quiet) emitBare(jobId);
+    else emitResult(response, "json");
+  },
+});
diff --git a/packages/commands/src/commands/finetune/compress-get.ts b/packages/commands/src/commands/finetune/compress-get.ts
new file mode 100644
index 000000000..8d0279284
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-get.ts
@@ -0,0 +1,19 @@
+import { defineCommand, getCompressJob } from "bailian-cli-core";
+import { emitResult } from "bailian-cli-runtime";
+import { COMPRESS_JOB_ID_FLAGS } from "./compress-shared.ts";
+
+export default defineCommand({
+  description: { "en-US": "Get a compression job", "zh-CN": "查询压缩任务详情" },
+  auth: "apiKey",
+  usageArgs: "--job-id <id>",
+  flags: COMPRESS_JOB_ID_FLAGS,
+  exampleArgs: ["--job-id quant-xxx"],
+  async run(ctx) {
+    const jobId = ctx.flags.jobId;
+    if (ctx.settings.dryRun) {
+      emitResult({ action: "finetune.compress.get", job_id: jobId }, "json");
+      return;
+    }
+    emitResult(await getCompressJob(ctx.client, jobId), "json");
+  },
+});
diff --git a/packages/commands/src/commands/finetune/compress-list.ts b/packages/commands/src/commands/finetune/compress-list.ts
new file mode 100644
index 000000000..4539654f7
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-list.ts
@@ -0,0 +1,148 @@
+import { defineCommand, listCompressJobs, type FlagsDef } from "bailian-cli-core";
+import { emitResult } from "bailian-cli-runtime";
+
+const LIST_FLAGS = {
+  status: {
+    type: "string",
+    valueHint: "<status>",
+    description: { "en-US": "Filter by job status", "zh-CN": "按任务状态筛选" },
+  },
+  model: {
+    type: "string",
+    valueHint: "<model>",
+    description: { "en-US": "Filter by source model ID", "zh-CN": "按源模型 ID 筛选" },
+  },
+  templateId: {
+    type: "string",
+    valueHint: "<id>",
+    description: { "en-US": "Filter by template ID", "zh-CN": "按模板 ID 筛选" },
+  },
+  quantSpec: {
+    type: "string",
+    valueHint: "<spec>",
+    description: { "en-US": "Filter by quantization specification", "zh-CN": "按量化规格筛选" },
+  },
+  quantMethod: {
+    type: "string",
+    valueHint: "<method>",
+    description: { "en-US": "Filter by quantization method", "zh-CN": "按量化算法筛选" },
+  },
+  startTime: {
+    type: "string",
+    valueHint: "<time>",
+    description: {
+      "en-US": "Filter jobs starting no earlier than this time",
+      "zh-CN": "筛选开始时间不早于该时间的任务",
+    },
+  },
+  endTime: {
+    type: "string",
+    valueHint: "<time>",
+    description: {
+      "en-US": "Filter jobs ending no later than this time",
+      "zh-CN": "筛选结束时间不晚于该时间的任务",
+    },
+  },
+  search: {
+    type: "string",
+    valueHint: "<keyword>",
+    description: { "en-US": "Search job ID or name", "zh-CN": "搜索任务 ID 或名称" },
+  },
+  searchField: {
+    type: "string",
+    valueHint: "<field>",
+    choices: ["job_id", "job_name"] as const,
+    description: {
+      "en-US": "Limit --search to job_id or job_name",
+      "zh-CN": "将 --search 限制为 job_id 或 job_name",
+    },
+  },
+  sortOrder: {
+    type: "string",
+    valueHint: "<order>",
+    choices: ["asc", "desc"] as const,
+    description: {
+      "en-US": "Sort order by create_time (default: desc)",
+      "zh-CN": "按 create_time 排序（默认：desc）",
+    },
+  },
+  page: {
+    type: "number",
+    valueHint: "<n>",
+    description: { "en-US": "Page number (default: 1)", "zh-CN": "页码（默认：1）" },
+  },
+  pageSize: {
+    type: "number",
+    valueHint: "<n>",
+    description: {
+      "en-US": "Results per page (default: 10, max: 100)",
+      "zh-CN": "每页数量（默认：10，最多：100）",
+    },
+  },
+} satisfies FlagsDef;
+
+export default defineCommand({
+  description: { "en-US": "List model compression jobs", "zh-CN": "列出模型压缩任务" },
+  auth: "apiKey",
+  usageArgs: "[filters] [--page <n>] [--page-size <n>]",
+  flags: LIST_FLAGS,
+  validate(flags) {
+    if (flags.page !== undefined && (!Number.isInteger(flags.page) || flags.page < 1))
+      return "--page must be an integer of at least 1. / --page 必须是大于等于 1 的整数。";
+    if (
+      flags.pageSize !== undefined &&
+      (!Number.isInteger(flags.pageSize) || flags.pageSize < 1 || flags.pageSize > 100)
+    )
+      return "--page-size must be an integer between 1 and 100. / --page-size 必须是 1–100 的整数。";
+    if (flags.searchField && !flags.search)
+      return "--search-field requires --search. / --search-field 必须与 --search 一起使用。";
+    return undefined;
+  },
+  exampleArgs: [
+    "",
+    "--status SUCCEEDED --page-size 20",
+    "--quant-method gptq",
+    "--search qwen3 --search-field job_name --sort-order asc",
+  ],
+  async run(ctx) {
+    const params = {
+      status: ctx.flags.status,
+      model: ctx.flags.model,
+      templateId: ctx.flags.templateId,
+      quantSpec: ctx.flags.quantSpec,
+      quantMethod: ctx.flags.quantMethod,
+      startTime: ctx.flags.startTime,
+      endTime: ctx.flags.endTime,
+      searchKey: ctx.flags.search,
+      selectKey: ctx.flags.searchField,
+      sortOrder: ctx.flags.sortOrder,
+      pageNo: ctx.flags.page,
+      pageSize: ctx.flags.pageSize,
+    };
+    if (ctx.settings.dryRun) {
+      emitResult({ action: "finetune.compress.list", ...params }, "json");
+      return;
+    }
+    const response = await listCompressJobs(ctx.client, params);
+    const output = response.output;
+    const items = (output?.jobs ?? []).map((job) => ({
+      job_id: job.job_id ?? "",
+      job_name: job.job_name ?? "",
+      status: job.status ?? "",
+      model: job.model ?? "",
+      template_name: job.template_name ?? "",
+      quantized_output: job.quantized_output ?? "",
+      create_time: job.create_time ?? "",
+    }));
+    emitResult(
+      {
+        items,
+        total: output?.total,
+        page_no: output?.page_no,
+        page_size: output?.page_size,
+        request_id: response.request_id,
+      },
+      "json",
+    );
+  },
+});
diff --git a/packages/commands/src/commands/finetune/compress-logs.ts b/packages/commands/src/commands/finetune/compress-logs.ts
new file mode 100644
index 000000000..1a0544ad7
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-logs.ts
@@ -0,0 +1,90 @@
+import { defineCommand, getCompressJobLogs, type FlagsDef } from "bailian-cli-core";
+import { emitResult } from "bailian-cli-runtime";
+import { COMPRESS_JOB_ID_FLAGS } from "./compress-shared.ts";
+
+const LOG_FLAGS = {
+  ...COMPRESS_JOB_ID_FLAGS,
+  offset: {
+    type: "number",
+    valueHint: "<n>",
+    description: {
+      "en-US": "Skip the first N lines (default: 0)",
+      "zh-CN": "跳过前 N 行（默认：0）",
+    },
+  },
+  lines: {
+    type: "number",
+    valueHint: "<n>",
+    description: {
+      "en-US": "Number of lines to read (default: 100, max: 1000)",
+      "zh-CN": "读取行数（默认：100，最多：1000）",
+    },
+  },
+  tail: {
+    type: "number",
+    valueHint: "<n>",
+    description: { "en-US": "Read the last N log lines", "zh-CN": "读取最后 N 行日志" },
+  },
+} satisfies FlagsDef;
+
+export default defineCommand({
+  description: { "en-US": "Get compression job logs", "zh-CN": "获取压缩任务日志" },
+  auth: "apiKey",
+  usageArgs: "--job-id <id> [--offset <n>] [--lines <n> | --tail <n>]",
+  flags: LOG_FLAGS,
+  validate(flags) {
+    if (flags.offset !== undefined && (!Number.isInteger(flags.offset) || flags.offset < 0))
+      return "--offset must be a non-negative integer. / --offset 必须是非负整数。";
+    if (
+      flags.lines !== undefined &&
+      (!Number.isInteger(flags.lines) || flags.lines < 1 || flags.lines > 1000)
+    )
+      return "--lines must be an integer between 1 and 1000. / --lines 必须是 1–1000 的整数。";
+    if (
+      flags.tail !== undefined &&
+      (!Number.isInteger(flags.tail) || flags.tail < 1 || flags.tail > 1000)
+    )
+      return "--tail must be an integer between 1 and 1000. / --tail 必须是 1–1000 的整数。";
+    if (flags.tail !== undefined && (flags.offset !== undefined || flags.lines !== undefined))
+      return "--tail cannot be combined with --offset or --lines. / --tail 不能与 --offset 或 --lines 同时使用。";
+    return undefined;
+  },
+  exampleArgs: [
+    "--job-id quant-xxx",
+    "--job-id quant-xxx --offset 0 --lines 50",
+    "--job-id quant-xxx --tail 20",
+  ],
+  async run(ctx) {
+    const jobId = ctx.flags.jobId;
+    if (ctx.settings.dryRun) {
+      emitResult(
+        {
+          action: "finetune.compress.logs",
+          job_id: jobId,
+          offset: ctx.flags.offset,
+          line: ctx.flags.lines,
+          tail: ctx.flags.tail,
+        },
+        "json",
+      );
+      return;
+    }
+    if (ctx.flags.tail !== undefined) {
+      const first = await getCompressJobLogs(ctx.client, jobId, { offset: 0, line: 1 });
+      const total = first.output?.total ?? 0;
+      const offset = Math.max(0, total - ctx.flags.tail);
+      emitResult(
+        await getCompressJobLogs(ctx.client, jobId, { offset, line: ctx.flags.tail }),
+        "json",
+      );
+      return;
+    }
+    emitResult(
+      await getCompressJobLogs(ctx.client, jobId, {
+        offset: ctx.flags.offset,
+        line: ctx.flags.lines,
+      }),
+      "json",
+    );
+  },
+});
diff --git a/packages/commands/src/commands/finetune/compress-shared.ts b/packages/commands/src/commands/finetune/compress-shared.ts
new file mode 100644
index 000000000..94e6fe9db
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-shared.ts
@@ -0,0 +1,83 @@
+import { BailianError, ExitCode, type FlagsDef } from "bailian-cli-core";
+
+export const COMPRESS_JOB_ID_FLAGS = {
+  jobId: {
+    type: "string",
+    valueHint: "<id>",
+    description: {
+      "en-US": "Compression job ID (required)",
+      "zh-CN": "压缩任务 ID（必填）",
+    },
+    required: true,
+  },
+} satisfies FlagsDef;
+
+export const COMPRESS_TERMINAL_STATUSES = new Set(["SUCCEEDED", "FAILED", "CANCELED"]);
+
+/** Parse repeatable KEY=VALUE entries without coercing values. */
+export function parseCompressHyperParameters(
+  entries: string[] | undefined,
+): Record<string, string> {
+  const result: Record<string, string> = {};
+  for (const entry of entries ?? []) {
+    const separator = entry.indexOf("=");
+    if (separator <= 0) {
+      throw new BailianError(
+        "--hparam values must use KEY=VALUE format. / --hparam 必须使用 KEY=VALUE 格式。",
+        ExitCode.USAGE,
+      );
+    }
+    const key = entry.slice(0, separator).trim();
+    if (!key) {
+      throw new BailianError(
+        "--hparam requires a non-empty key. / --hparam 的键不能为空。",
+        ExitCode.USAGE,
+      );
+    }
+    result[key] = entry.slice(separator + 1);
+  }
+  return result;
+}
+
+export function parseCommaSeparated(value: string | undefined): string[] | undefined {
+  if (!value) return undefined;
+  const items = [
+    ...new Set(
+      value
+        .split(",")
+        .map((entry) => entry.trim())
+        .filter(Boolean),
+    ),
+  ];
+  return items.length ? items : undefined;
+}
+
+export function formatCompressFailure(
+  jobId: string,
+  status: string,
+  message?: string,
+): BailianError {
+  return new BailianError(
+    message ||
+      `Compression job ${jobId} ended in status ${status}. / 压缩任务 ${jobId} 已以 ${status} 状态结束。`,
+    ExitCode.GENERAL,
+  );
+}
+
+export function sleepWithSignal(milliseconds: number, signal: AbortSignal): Promise<void> {
+  return new Promise((resolve, reject) => {
+    if (signal.aborted) {
+      reject(new Error("aborted"));
+      return;
+    }
+    const onAbort = () => {
+      clearTimeout(timer);
+      reject(new Error("aborted"));
+    };
+    const timer = setTimeout(() => {
+      signal.removeEventListener("abort", onAbort);
+      resolve();
+    }, milliseconds);
+    signal.addEventListener("abort", onAbort, { once: true });
+  });
+}
diff --git a/packages/commands/src/commands/finetune/compress-templates.ts b/packages/commands/src/commands/finetune/compress-templates.ts
new file mode 100644
index 000000000..10a0cea0c
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-templates.ts
@@ -0,0 +1,59 @@
+import { defineCommand, listCompressTemplates, type FlagsDef } from "bailian-cli-core";
+import { emitResult } from "bailian-cli-runtime";
+
+const TEMPLATE_FLAGS = {
+  model: {
+    type: "string",
+    valueHint: "<model>",
+    description: {
+      "en-US": "Filter by custom model ID or base model name",
+      "zh-CN": "按自定义模型 ID 或基础模型名称筛选",
+    },
+  },
+} satisfies FlagsDef;
+
+export default defineCommand({
+  description: {
+    "en-US": "List quantizable models and their compression templates",
+    "zh-CN": "列出可量化模型及其压缩模板",
+  },
+  auth: "apiKey",
+  usageArgs: "[--model <model>]",
+  flags: TEMPLATE_FLAGS,
+  exampleArgs: ["", "--model qwen3.5-flash-2026-02-23-ft-xxx"],
+  notes: [
+    {
+      "en-US":
+        "Only supported full-parameter fine-tuned models are returned. LoRA and already-quantized models are excluded.",
+      "zh-CN": "仅返回支持压缩的全参调优模型；LoRA 模型和已量化模型不会返回。",
+    },
+    {
+      "en-US":
+        "Template IDs live in output.custom_models[].templates[].template_id. output.base_models only lists quantizable base model names and carries no templates, so an empty custom_models array means no template ID is available yet.",
+      "zh-CN":
+        "模板 ID 位于 output.custom_models[].templates[].template_id；output.base_models 只列出可量化的基础模型名称，不含模板，因此 custom_models 为空数组时说明当前还拿不到任何模板 ID。",
+    },
+    {
+      "en-US":
+        "An empty templates[].hyper_parameters array means that template exposes no tunable hyper-parameters; omit --hparam when creating the compression job.",
+      "zh-CN":
+        "templates[].hyper_parameters 为空数组表示该模板没有可调超参，创建压缩任务时不要传 --hparam。",
+    },
+    {
+      "en-US":
+        "template_name and description are localized server-side from the CLI language setting (default en-US), sent as the lang query parameter. Run `config set --key language --value zh-CN` to get Chinese text.",
+      "zh-CN":
+        "template_name 和 description 由服务端按 CLI 的 language 设置本地化（默认 en-US，作为 lang 查询参数下发）；需要中文执行 `config set --key language --value zh-CN`。",
+    },
+  ],
+  async run(ctx) {
+    const model = ctx.flags.model || undefined;
+    const lang = ctx.localize({ "en-US": "en-US", "zh-CN": "zh-CN" });
+    if (ctx.settings.dryRun) {
+      emitResult({ action: "finetune.compress.templates", model, lang }, "json");
+      return;
+    }
+    const response = await listCompressTemplates(ctx.client, { model, lang });
+    emitResult(response, "json");
+  },
+});
diff --git a/packages/commands/src/commands/finetune/compress-watch.ts b/packages/commands/src/commands/finetune/compress-watch.ts
new file mode 100644
index 000000000..76d59abac
--- /dev/null
+++ b/packages/commands/src/commands/finetune/compress-watch.ts
@@ -0,0 +1,151 @@
+import {
+  BailianError,
+  ExitCode,
+  defineCommand,
+  getCompressJob,
+  type FlagsDef,
+} from "bailian-cli-core";
+import { emitBare, emitResult } from "bailian-cli-runtime";
+import {
+  COMPRESS_JOB_ID_FLAGS,
+  COMPRESS_TERMINAL_STATUSES,
+  formatCompressFailure,
+  sleepWithSignal,
+} from "./compress-shared.ts";
+
+const DEFAULT_INTERVAL_SEC = 10;
+const WATCH_FLAGS = {
+  ...COMPRESS_JOB_ID_FLAGS,
+  follow: {
+    type: "switch",
+    description: {
+      "en-US": "Poll until the job reaches a terminal state",
+      "zh-CN": "持续轮询直至任务进入终态",
+    },
+  },
+  interval: {
+    type: "number",
+    valueHint: "<sec>",
+    description: {
+      "en-US": "Polling interval with --follow (default: 10, min: 1)",
+      "zh-CN": "使用 --follow 时的轮询间隔（默认：10，最小：1）",
+    },
+  },
+  pollTimeout: {
+    type: "number",
+    valueHint: "<sec>",
+    description: {
+      "en-US": "Stop polling after this many seconds",
+      "zh-CN": "在指定秒数后停止轮询",
+    },
+  },
+} satisfies FlagsDef;
+
+export default defineCommand({
+  description: {
+    "en-US": "Probe or follow a compression job's status",
+    "zh-CN": "查询或持续跟踪压缩任务状态",
+  },
+  auth: "apiKey",
+  usageArgs: "--job-id <id> [--follow] [--interval <sec>] [--poll-timeout <sec>]",
+  flags: WATCH_FLAGS,
+  validate(flags) {
+    if (flags.interval !== undefined && (!Number.isFinite(flags.interval) || flags.interval < 1))
+      return "--interval must be at least 1 second. / --interval 必须至少为 1 秒。";
+    if (
+      flags.pollTimeout !== undefined &&
+      (!Number.isFinite(flags.pollTimeout) || flags.pollTimeout < 1)
+    )
+      return "--poll-timeout must be at least 1 second. / --poll-timeout 必须至少为 1 秒。";
+    return undefined;
+  },
+  exampleArgs: [
+    "--job-id quant-xxx",
+    "--job-id quant-xxx --follow",
+    "--job-id quant-xxx --follow --interval 30 --poll-timeout 7200",
+  ],
+  notes: [
+    {
+      "en-US": "Without --follow, the command performs one non-blocking status probe.",
+      "zh-CN": "不使用 --follow 时仅执行一次非阻塞状态查询。",
+    },
+    {
+      "en-US": "On success, quantized_output can be passed to `deploy text create --model-name`.",
+      "zh-CN": "成功后可将 quantized_output 传给 `deploy text create --model-name`。",
+    },
+  ],
+  async run(ctx) {
+    const jobId = ctx.flags.jobId;
+    const intervalSec = ctx.flags.interval ?? DEFAULT_INTERVAL_SEC;
+    if (ctx.settings.dryRun) {
+      emitResult(
+        {
+          action: "finetune.compress.watch",
+          job_id: jobId,
+          follow: ctx.flags.follow,
+          interval: intervalSec,
+          timeout: ctx.flags.pollTimeout,
+        },
+        "json",
+      );
+      return;
+    }
+    if (!ctx.flags.follow) {
+      const response = await getCompressJob(ctx.client, jobId);
+      const status = String(response.output?.status ?? "UNKNOWN").toUpperCase();
+      if (ctx.settings.quiet) emitBare(status);
+      else
+        emitResult(
+          {
+            job_id: jobId,
+            status,
+            terminal: COMPRESS_TERMINAL_STATUSES.has(status),
+            quantized_output: response.output?.quantized_output,
+            request_id: response.request_id,
+          },
+          "json",
+        );
+      if (COMPRESS_TERMINAL_STATUSES.has(status) && status !== "SUCCEEDED")
+        throw formatCompressFailure(jobId, status, response.output?.error?.message);
+      return;
+    }
+
+    const controller = new AbortController();
+    const onSigint = () => controller.abort();
+    process.on("SIGINT", onSigint);
+    const startedAt = Date.now();
+    try {
+      while (true) {
+        const response = await getCompressJob(ctx.client, jobId, controller.signal);
+        const status = String(response.output?.status ?? "UNKNOWN").toUpperCase();
+        if (COMPRESS_TERMINAL_STATUSES.has(status)) {
+          if (ctx.settings.quiet) emitBare(status);
+          else emitResult(response, "json");
+          if (status !== "SUCCEEDED")
+            throw formatCompressFailure(jobId, status, response.output?.error?.message);
+          return;
+        }
+        if (
+          ctx.flags.pollTimeout !== undefined &&
+          (Date.now() - startedAt) / 1000 >= ctx.flags.pollTimeout
+        ) {
+          throw new BailianError(
+            `Watching compression job ${jobId} timed out. / 等待压缩任务 ${jobId} 超时。`,
+            ExitCode.TIMEOUT,
+          );
+        }
+        await sleepWithSignal(intervalSec * 1000, controller.signal);
+      }
+    } catch (error) {
+      if (controller.signal.aborted) {
+        process.stderr.write(
+          `\n${ctx.localize({ "en-US": "Interrupted.", "zh-CN": "已中断。" })}\n`,
+        );
+        return;
+      }
+      throw error;
+    } finally {
+      process.off("SIGINT", onSigint);
+    }
+  },
+});
diff --git a/packages/commands/src/index.ts b/packages/commands/src/index.ts
index 518408083..21de894ff 100644
--- a/packages/commands/src/index.ts
+++ b/packages/commands/src/index.ts
@@ -17,6 +17,7 @@ export { default as videoRef } from "./commands/video/ref.ts";
 export { default as videoTaskGet } from "./commands/video/task-get.ts";
 export { default as videoDownload } from "./commands/video/download.ts";
 export { default as visionDescribe } from "./commands/vision/describe.ts";
+export { default as decisionPredict } from "./commands/decision/predict.ts";
 export { default as configShow } from "./commands/config/show.ts";
 export { default as configSet } from "./commands/config/set.ts";
 export { default as configList } from "./commands/config/list.ts";
@@ -159,6 +160,14 @@ export { default as finetuneExport } from "./commands/finetune/export.ts";
 export { default as finetuneWatch } from "./commands/finetune/watch.ts";
 export { default as finetuneCapability } from "./commands/finetune/capability.ts";
 export { default as finetunePrice } from "./commands/finetune/price.ts";
+export { default as finetuneCompressTemplates } from "./commands/finetune/compress-templates.ts";
+export { default as finetuneCompressCreate } from "./commands/finetune/compress-create.ts";
+export { default as finetuneCompressList } from "./commands/finetune/compress-list.ts";
+export { default as finetuneCompressGet } from "./commands/finetune/compress-get.ts";
+export { default as finetuneCompressWatch } from "./commands/finetune/compress-watch.ts";
+export { default as finetuneCompressLogs } from "./commands/finetune/compress-logs.ts";
+export { default as finetuneCompressCancel } from "./commands/finetune/compress-cancel.ts";
+export { default as finetuneCompressDelete } from "./commands/finetune/compress-delete.ts";
 export {
   deployTextCreate,
   deployAudioCreate,
diff --git a/packages/commands/tests/decision.test.ts b/packages/commands/tests/decision.test.ts
new file mode 100644
index 000000000..c2d5c082f
--- /dev/null
+++ b/packages/commands/tests/decision.test.ts
@@ -0,0 +1,229 @@
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { afterEach, beforeEach, describe, expect, test } from "vite-plus/test";
+import {
+  inspectDecisionQuestions,
+  inspectDecisionState,
+  stripDecisionTrainingFields,
+  unwrapDecisionQuestions,
+  type LocalizedText,
+} from "bailian-cli-core";
+import {
+  buildDecisionInput,
+  formatDecisionAnswers,
+  formatDecisionStats,
+  resolveDecisionEndpoint,
+} from "../src/commands/decision/shared.ts";
+
+const localize = (text: LocalizedText) => (typeof text === "string" ? text : text["en-US"]);
+
+function createUrlResolver(baseUrl: string, usesDefaultBaseUrl: boolean) {
+  return {
+    url(path: string, defaultBaseUrl?: () => string): string {
+      const resolvedBaseUrl = usesDefaultBaseUrl && defaultBaseUrl ? defaultBaseUrl() : baseUrl;
+      return resolvedBaseUrl + path;
+    },
+  };
+}
+
+function issueCodes(issues: Array<{ code: string }>): string[] {
+  return issues.map((issue) => issue.code);
+}
+
+let workDirectory: string;
+
+beforeEach(() => {
+  workDirectory = mkdtempSync(join(tmpdir(), "bl-decision-unit-"));
+});
+
+afterEach(() => {
+  rmSync(workDirectory, { recursive: true, force: true });
+});
+
+describe("decision question rules", () => {
+  test("state 接受非空字符串、对象和数组", () => {
+    expect(inspectDecisionState("text")).toEqual([]);
+    expect(inspectDecisionState({ ticket_id: "T-1" })).toEqual([]);
+    expect(inspectDecisionState([{ role: "user" }])).toEqual([]);
+    for (const invalidState of ["  ", {}, [], 1, null, undefined]) {
+      expect(issueCodes(inspectDecisionState(invalidState))).toEqual(["DECISION_INVALID_STATE"]);
+    }
+  });
+
+  test("choice 选项数为 1-255", () => {
+    const options = (count: number) =>
+      Object.fromEntries(Array.from({ length: count }, (_option, index) => [`o${index}`, "desc"]));
+    expect(inspectDecisionQuestions({ team: { type: "choice", criteria: options(1) } })).toEqual(
+      [],
+    );
+    expect(inspectDecisionQuestions({ team: { type: "choice", criteria: options(255) } })).toEqual(
+      [],
+    );
+    expect(
+      issueCodes(inspectDecisionQuestions({ team: { type: "choice", criteria: options(256) } })),
+    ).toEqual(["DECISION_INVALID_CHOICE_CRITERIA"]);
+    expect(
+      issueCodes(inspectDecisionQuestions({ team: { type: "choice", criteria: {} } })),
+    ).toEqual(["DECISION_INVALID_CHOICE_CRITERIA"]);
+  });
+
+  test("score 等级数为 2-255", () => {
+    const levels = (count: number) => Array.from({ length: count }, () => "level");
+    expect(
+      issueCodes(inspectDecisionQuestions({ severity: { type: "score", criteria: levels(1) } })),
+    ).toEqual(["DECISION_INVALID_SCORE_CRITERIA"]);
+    expect(inspectDecisionQuestions({ severity: { type: "score", criteria: levels(2) } })).toEqual(
+      [],
+    );
+    expect(
+      inspectDecisionQuestions({ severity: { type: "score", criteria: levels(255) } }),
+    ).toEqual([]);
+    expect(
+      issueCodes(inspectDecisionQuestions({ severity: { type: "score", criteria: levels(256) } })),
+    ).toEqual(["DECISION_INVALID_SCORE_CRITERIA"]);
+  });
+
+  test("noul criteria 可选，填写时只能含 true/false", () => {
+    expect(inspectDecisionQuestions({ refund: { type: "noul" } })).toEqual([]);
+    expect(
+      inspectDecisionQuestions({ refund: { type: "noul", criteria: { true: "是", false: "否" } } }),
+    ).toEqual([]);
+    expect(
+      issueCodes(
+        inspectDecisionQuestions({ refund: { type: "noul", criteria: { maybe: "不确定" } } }),
+      ),
+    ).toEqual(["DECISION_INVALID_NOUL_CRITERIA"]);
+  });
+
+  test("instructions 可选，类型与问题类型受校验", () => {
+    expect(
+      issueCodes(
+        inspectDecisionQuestions({
+          refund: { type: "noul", instructions: 1 },
+          other: { type: "summary" },
+        }),
+      ),
+    ).toEqual(["DECISION_INVALID_INSTRUCTIONS", "DECISION_INVALID_QUESTION_TYPE"]);
+  });
+
+  test("解包 {questions:{...}}，但不误解包 ID 为 questions 的问题", () => {
+    const table = { refund: { type: "noul" } };
+    expect(unwrapDecisionQuestions({ questions: table, name: "workload" })).toEqual(table);
+    const idNamedQuestions = { questions: { type: "noul" } };
+    expect(unwrapDecisionQuestions(idNamedQuestions)).toBe(idNamedQuestions);
+  });
+
+  test("去掉 label/target 且不修改原对象", () => {
+    const original = { refund: { type: "noul", label: true, target: { true: 1, false: 0 } } };
+    expect(stripDecisionTrainingFields(original)).toEqual({ refund: { type: "noul" } });
+    expect(original.refund).toHaveProperty("label");
+  });
+});
+
+describe("decision command helpers", () => {
+  test("state-file：.json 解析为对象，其他扩展名按文本", () => {
+    const jsonPath = join(workDirectory, "ticket.json");
+    const textPath = join(workDirectory, "ticket.txt");
+    writeFileSync(jsonPath, JSON.stringify({ ticket_id: "T-1001" }));
+    writeFileSync(textPath, "订单未到账");
+    const questions = JSON.stringify({ refund: { type: "noul" } });
+
+    expect(buildDecisionInput({ stateFile: jsonPath, questions }, localize).state).toEqual({
+      ticket_id: "T-1001",
+    });
+    expect(buildDecisionInput({ stateFile: textPath, questions }, localize).state).toBe(
+      "订单未到账",
+    );
+  });
+
+  test("questions-file 兼容裸问题表与包裹格式", () => {
+    const barePath = join(workDirectory, "bare.json");
+    const wrappedPath = join(workDirectory, "wrapped.json");
+    const table = { refund: { type: "noul" } };
+    writeFileSync(barePath, JSON.stringify(table));
+    writeFileSync(wrappedPath, JSON.stringify({ questions: table }));
+
+    expect(buildDecisionInput({ state: "x", questionsFile: barePath }, localize).questions).toEqual(
+      table,
+    );
+    expect(
+      buildDecisionInput({ state: "x", questionsFile: wrappedPath }, localize).questions,
+    ).toEqual(table);
+  });
+
+  test("request-file 携带的 model 被读出", () => {
+    const requestPath = join(workDirectory, "request.json");
+    writeFileSync(
+      requestPath,
+      JSON.stringify({ model: "my-ft", state: "x", questions: { refund: { type: "noul" } } }),
+    );
+    expect(buildDecisionInput({ requestFile: requestPath }, localize).model).toBe("my-ft");
+  });
+
+  test("非法 JSON 与非法问题定义抛 USAGE 错误", () => {
+    expect(() => buildDecisionInput({ state: "x", questions: "{" }, localize)).toThrow(
+      /not valid JSON/,
+    );
+    expect(() =>
+      buildDecisionInput({ state: "x", questions: JSON.stringify({ q: { type: "x" } }) }, localize),
+    ).toThrow(/Invalid decision request/);
+  });
+
+  test("Endpoint：默认按 workspace + region 拼接，显式 Base URL 优先", () => {
+    const base = { binName: "bl", localize };
+    expect(
+      resolveDecisionEndpoint({
+        ...base,
+        workspaceId: "ws_test",
+        region: "ap-southeast-1",
+        client: createUrlResolver("https://dashscope.aliyuncs.com", true),
+      }),
+    ).toBe("https://ws_test.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/systemone");
+    expect(
+      resolveDecisionEndpoint({
+        ...base,
+        region: "cn-beijing",
+        client: createUrlResolver("https://example.test", false),
+      }),
+    ).toBe("https://example.test/compatible-mode/v1/systemone");
+    expect(() =>
+      resolveDecisionEndpoint({
+        ...base,
+        region: "cn-beijing",
+        client: createUrlResolver("https://dashscope.aliyuncs.com", true),
+      }),
+    ).toThrow(/Workspace ID/);
+  });
+
+  test("文本输出：每题一行，score 显示概率最高的等级", () => {
+    const response = {
+      answers: {
+        department: {
+          type: "choice",
+          choice: "billing",
+          confidence: 0.88,
+          probabilities: { billing: 0.94, technical: 0.06 },
+        },
+        escalate: { type: "noul", noul: 0.99 },
+        severity: {
+          type: "score",
+          score: 2.25,
+          confidence: 0.91,
+          legend: { "0": "轻微", "1": "部分受影响", "2": "核心不可用", "3": "严重" },
+          probabilities: { "0": 0, "1": 0.01, "2": 0.73, "3": 0.26 },
+        },
+        future: { type: "rank", rank: [1, 2] },
+      },
+      usage: { input_tokens: 125 },
+      latency_ms: 52.9,
+    };
+    expect(formatDecisionAnswers(response).split("\n")).toEqual([
+      "department  choice=billing  confidence=0.88  billing=0.94 technical=0.06",
+      "escalate    noul P(yes)=0.99",
+      "severity    score=2.25  confidence=0.91  top=2 核心不可用",
+      'future      {"type":"rank","rank":[1,2]}',
+    ]);
+    expect(formatDecisionStats(response)).toBe("input_tokens=125  latency_ms=52.9");
+  });
+});
diff --git a/packages/commands/tests/e2e/dataset.e2e.test.ts b/packages/commands/tests/e2e/dataset.e2e.test.ts
index a7cb0dd4f..2c733106c 100644
--- a/packages/commands/tests/e2e/dataset.e2e.test.ts
+++ b/packages/commands/tests/e2e/dataset.e2e.test.ts
@@ -31,6 +31,7 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: dataset (offline)", () => {
     ]);
     expect(exitCode, stderr).toBe(0);
     expect(stderr).toMatch(/--file|jsonl/i);
+    expect(stderr).toMatch(/decision/i);
   });
 
   test("dataset validate 通过合法 JSONL", async () => {
diff --git a/packages/commands/tests/e2e/decision.e2e.test.ts b/packages/commands/tests/e2e/decision.e2e.test.ts
new file mode 100644
index 000000000..170c48eca
--- /dev/null
+++ b/packages/commands/tests/e2e/decision.e2e.test.ts
@@ -0,0 +1,195 @@
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { afterEach, beforeEach, describe, expect, test } from "vite-plus/test";
+import { parseStdoutJson, runCommandE2e, runCommandHelp } from "./helpers.ts";
+import { DECISION_ROUTES } from "./topic-routes.ts";
+
+let isolatedDirectory: string;
+
+beforeEach(() => {
+  isolatedDirectory = mkdtempSync(join(tmpdir(), "bl-decision-e2e-"));
+  writeFileSync(join(isolatedDirectory, "config.json"), "{}");
+});
+
+afterEach(() => {
+  rmSync(isolatedDirectory, { recursive: true, force: true });
+});
+
+const EMPTY_MODEL_ENV = {
+  DASHSCOPE_API_KEY: "",
+  DASHSCOPE_BASE_URL: "",
+  BAILIAN_WORKSPACE_ID: "",
+};
+
+function runIsolatedDecision(args: string[], envOverrides: NodeJS.ProcessEnv = {}) {
+  return runCommandE2e(DECISION_ROUTES, ["decision", ...args], {
+    BAILIAN_CONFIG_DIR: isolatedDirectory,
+    HOME: isolatedDirectory,
+    USERPROFILE: isolatedDirectory,
+    ...EMPTY_MODEL_ENV,
+    ...envOverrides,
+  });
+}
+
+const REFUND_QUESTIONS = JSON.stringify({
+  refund: { type: "noul", instructions: "是否在索要退款？" },
+});
+
+interface DryRunOutput {
+  endpoint?: string;
+  request?: {
+    model?: string;
+    state?: unknown;
+    questions?: Record<string, Record<string, unknown>>;
+  };
+}
+
+/** 决策模型推理：help / 缺参 / 本地校验 / dry-run 均离线。 */
+describe("e2e: decision", () => {
+  test("decision --help 列出输入与地域 flag", async () => {
+    const { stderr, exitCode } = await runCommandHelp(DECISION_ROUTES, ["decision", "--help"]);
+    expect(exitCode, stderr).toBe(0);
+    expect(stderr).toMatch(/--request-file/);
+    expect(stderr).toMatch(/--questions-file/);
+    expect(stderr).toMatch(/--workspace-id/);
+    expect(stderr).toMatch(/--region/);
+  });
+
+  test("缺少 state 时报参数错误", async () => {
+    const { stderr, exitCode } = await runIsolatedDecision(["--questions", REFUND_QUESTIONS]);
+    expect(exitCode).toBe(2);
+    expect(stderr).toMatch(/--state|--request-file/);
+  });
+
+  test("--request-file 不能与 --state 混用", async () => {
+    const { stderr, exitCode } = await runIsolatedDecision([
+      "--request-file",
+      "sample.json",
+      "--state",
+      "x",
+    ]);
+    expect(exitCode).toBe(2);
+    expect(stderr).toMatch(/--request-file/);
+  });
+
+  test("本地校验拒绝非法问题定义", async () => {
+    const { stderr, exitCode } = await runIsolatedDecision([
+      "--state",
+      "x",
+      "--questions",
+      JSON.stringify({ severity: { type: "score", criteria: ["only"] } }),
+      "--workspace-id",
+      "ws_test",
+      "--dry-run",
+    ]);
+    expect(exitCode).toBe(2);
+    expect(stderr).toMatch(/criteria/);
+  });
+
+  test("默认 Base URL 下要求 workspace", async () => {
+    const { stderr, exitCode } = await runIsolatedDecision([
+      "--state",
+      "x",
+      "--questions",
+      REFUND_QUESTIONS,
+      "--dry-run",
+    ]);
+    expect(exitCode).toBe(2);
+    expect(stderr).toMatch(/Workspace ID/i);
+  });
+
+  test("dry-run 输出 workspace Endpoint 与请求体", async () => {
+    const { stdout, stderr, exitCode } = await runIsolatedDecision([
+      "--workspace-id",
+      "ws_test",
+      "--state",
+      "这段评论是否在索要退款？",
+      "--questions",
+      REFUND_QUESTIONS,
+      "--dry-run",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<DryRunOutput>(stdout);
+    expect(data.endpoint).toBe(
+      "https://ws_test.cn-beijing.maas.aliyuncs.com/compatible-mode/v1/systemone",
+    );
+    expect(data.request?.model).toBe("decision-model-preview");
+    expect(data.request?.state).toBe("这段评论是否在索要退款？");
+  });
+
+  test("dry-run 支持新加坡地域与 BAILIAN_WORKSPACE_ID", async () => {
+    const { stdout, stderr, exitCode } = await runIsolatedDecision(
+      [
+        "--region",
+        "ap-southeast-1",
+        "--state",
+        "x",
+        "--questions",
+        REFUND_QUESTIONS,
+        "--dry-run",
+        "--output",
+        "json",
+      ],
+      { BAILIAN_WORKSPACE_ID: "ws_env" },
+    );
+    expect(exitCode, stderr).toBe(0);
+    expect(parseStdoutJson<DryRunOutput>(stdout).endpoint).toBe(
+      "https://ws_env.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/systemone",
+    );
+  });
+
+  test("显式 Base URL 优先于 workspace 域名", async () => {
+    const { stdout, stderr, exitCode } = await runIsolatedDecision([
+      "--base-url",
+      "https://example.test",
+      "--state",
+      "x",
+      "--questions",
+      REFUND_QUESTIONS,
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    expect(parseStdoutJson<DryRunOutput>(stdout).endpoint).toBe(
+      "https://example.test/compatible-mode/v1/systemone",
+    );
+  });
+
+  test("--request-file 复用训练样本并去掉 label/target", async () => {
+    const samplePath = join(isolatedDirectory, "sample.json");
+    writeFileSync(
+      samplePath,
+      JSON.stringify({
+        state: { content: "订单支付后超过 24 小时仍未到账" },
+        questions: {
+          department: {
+            type: "choice",
+            instructions: "应该由哪个团队处理？",
+            criteria: { billing: "支付和账单问题", technical: "产品故障问题" },
+            label: "billing",
+            target: { billing: 0.96, technical: 0.04 },
+          },
+        },
+      }),
+    );
+    const { stdout, stderr, exitCode } = await runIsolatedDecision([
+      "--workspace-id",
+      "ws_test",
+      "--request-file",
+      samplePath,
+      "--model",
+      "my-decision-ft",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<DryRunOutput>(stdout);
+    expect(data.request?.model).toBe("my-decision-ft");
+    expect(data.request?.state).toEqual({ content: "订单支付后超过 24 小时仍未到账" });
+    expect(data.request?.questions?.department).not.toHaveProperty("label");
+    expect(data.request?.questions?.department).not.toHaveProperty("target");
+  });
+});
diff --git a/packages/commands/tests/e2e/finetune-compress.e2e.test.ts b/packages/commands/tests/e2e/finetune-compress.e2e.test.ts
new file mode 100644
index 000000000..8a0a37a0a
--- /dev/null
+++ b/packages/commands/tests/e2e/finetune-compress.e2e.test.ts
@@ -0,0 +1,99 @@
+import { describe, expect, test } from "vite-plus/test";
+import { isDashScopeE2EReady, parseStdoutJson, runCommandE2e, runCommandHelp } from "./helpers.ts";
+import { FINETUNE_ROUTES } from "./topic-routes.ts";
+
+describe.skipIf(!isDashScopeE2EReady())("e2e: finetune compress (offline)", () => {
+  test("compress group lists the lifecycle commands", async () => {
+    const { stderr, exitCode } = await runCommandHelp(FINETUNE_ROUTES, [
+      "finetune",
+      "compress",
+      "--help",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    expect(stderr).toMatch(/templates|create|list|get|watch|logs|cancel|delete/);
+  });
+
+  test("create help exposes required flags", async () => {
+    const { stderr, exitCode } = await runCommandHelp(FINETUNE_ROUTES, [
+      "finetune",
+      "compress",
+      "create",
+      "--help",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    expect(stderr).toMatch(/--model/);
+    expect(stderr).toMatch(/--template-id/);
+  });
+
+  test("create dry-run builds the compression request", async () => {
+    const { stdout, stderr, exitCode } = await runCommandE2e(FINETUNE_ROUTES, [
+      "finetune",
+      "compress",
+      "create",
+      "--model",
+      "qwen3.5-flash-2026-02-23-ft-xxx",
+      "--template-id",
+      "quant-flash-nvfp4-mlp-nomtp",
+      "--job-name",
+      "compression-test",
+      "--hparam",
+      "calib_input=true",
+      "--calibration-files",
+      "file-aaa,file-bbb,file-aaa",
+      "--suffix",
+      "test1",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    const data = parseStdoutJson<{
+      action: string;
+      request: Record<string, unknown>;
+    }>(stdout);
+    expect(data.action).toBe("finetune.compress.create");
+    expect(data.request).toMatchObject({
+      model: "qwen3.5-flash-2026-02-23-ft-xxx",
+      template_id: "quant-flash-nvfp4-mlp-nomtp",
+      job_name: "compression-test",
+      hyper_parameters: { calib_input: "true" },
+      custom_calibration_file_ids: ["file-aaa", "file-bbb"],
+      output_model_suffix: "test1",
+    });
+  });
+
+  test("create rejects an invalid output suffix", async () => {
+    const { stderr, exitCode } = await runCommandE2e(FINETUNE_ROUTES, [
+      "finetune",
+      "compress",
+      "create",
+      "--model",
+      "qwen-ft-xxx",
+      "--template-id",
+      "template-xxx",
+      "--suffix",
+      "Upper_Case",
+      "--dry-run",
+    ]);
+    expect(exitCode).not.toBe(0);
+    expect(stderr).toMatch(/suffix/i);
+  });
+
+  test("delete dry-run does not require confirmation", async () => {
+    const { stdout, stderr, exitCode } = await runCommandE2e(FINETUNE_ROUTES, [
+      "finetune",
+      "compress",
+      "delete",
+      "--job-id",
+      "quant-xxx",
+      "--dry-run",
+      "--output",
+      "json",
+    ]);
+    expect(exitCode, stderr).toBe(0);
+    expect(parseStdoutJson(stdout)).toMatchObject({
+      action: "finetune.compress.delete",
+      job_id: "quant-xxx",
+    });
+  });
+});
diff --git a/packages/commands/tests/e2e/topic-routes.ts b/packages/commands/tests/e2e/topic-routes.ts
index 700152507..7ae05a84a 100644
--- a/packages/commands/tests/e2e/topic-routes.ts
+++ b/packages/commands/tests/e2e/topic-routes.ts
@@ -16,6 +16,8 @@ export const UPDATE_ROUTES: E2eRouteExports = {
 
 export const TEXT_CHAT_ROUTES: E2eRouteExports = { "text chat": "textChat" };
 
+export const DECISION_ROUTES: E2eRouteExports = { decision: "decisionPredict" };
+
 export const CONFIG_ROUTES: E2eRouteExports = {
   "config show": "configShow",
   "config set": "configSet",
@@ -228,6 +230,14 @@ export const FINETUNE_ROUTES: E2eRouteExports = {
   "finetune export": "finetuneExport",
   "finetune watch": "finetuneWatch",
   "finetune capability": "finetuneCapability",
+  "finetune compress templates": "finetuneCompressTemplates",
+  "finetune compress create": "finetuneCompressCreate",
+  "finetune compress list": "finetuneCompressList",
+  "finetune compress get": "finetuneCompressGet",
+  "finetune compress watch": "finetuneCompressWatch",
+  "finetune compress logs": "finetuneCompressLogs",
+  "finetune compress cancel": "finetuneCompressCancel",
+  "finetune compress delete": "finetuneCompressDelete",
 };
 
 export const CONSOLE_FLAGS_DRY_RUN_ROUTES: E2eRouteExports = {
diff --git a/packages/core/src/client/endpoints.ts b/packages/core/src/client/endpoints.ts
index ce8c1bb41..de68dd7fb 100644
--- a/packages/core/src/client/endpoints.ts
+++ b/packages/core/src/client/endpoints.ts
@@ -11,6 +11,11 @@ export function responsesPath(): string {
   return "/compatible-mode/v1/responses";
 }
 
+// ---- Decision model (TypeSafe System One, workspace-based host) ----
+export function decisionPath(): string {
+  return "/compatible-mode/v1/systemone";
+}
+
 // ---- Image Generation (DashScope) ----
 /** Async image API used by wan2.6-t2i / wan2.6-image (T2I) and similar message-format models. */
 export function imagePath(): string {
@@ -108,10 +113,10 @@ export function knowledgeRetrievePath(): string {
   return "/api/v1/indices/rag/index/retrieve";
 }
 
-// ---- Workspace-scoped MaaS endpoints (cn-beijing only) ----
+// ---- Workspace-scoped MaaS endpoints (cn-beijing by default) ----
 
-export function workspaceMaaSBaseUrl(workspaceId: string): string {
-  return `https://${workspaceId}.cn-beijing.maas.aliyuncs.com`;
+export function workspaceMaaSBaseUrl(workspaceId: string, region = "cn-beijing"): string {
+  return `https://${workspaceId}.${region}.maas.aliyuncs.com`;
 }
 
 // ---- Knowledge Search (新版 RAG 检索, workspace-based host) ----
@@ -219,6 +224,33 @@ export function finetuneExportPath(jobId: string, checkpoint: string): string {
   return `/api/v1/fine-tunes/${encodeURIComponent(jobId)}/export/${encodeURIComponent(checkpoint)}`;
 }
 
+// ---- Model Compression (DashScope /api/v1/fine-tunes/compress) ----
+
+/** GET /api/v1/fine-tunes/compress/templates */
+export function finetuneCompressTemplatesPath(): string {
+  return "/api/v1/fine-tunes/compress/templates";
+}
+
+/** Create (POST) and list (GET) endpoint. */
+export function finetuneCompressJobsPath(): string {
+  return "/api/v1/fine-tunes/compress/jobs";
+}
+
+/** Single-job get / delete endpoint. */
+export function finetuneCompressJobPath(jobId: string): string {
+  return `/api/v1/fine-tunes/compress/jobs/${encodeURIComponent(jobId)}`;
+}
+
+/** POST /api/v1/fine-tunes/compress/jobs/{job_id}/cancel */
+export function finetuneCompressCancelPath(jobId: string): string {
+  return `/api/v1/fine-tunes/compress/jobs/${encodeURIComponent(jobId)}/cancel`;
+}
+
+/** GET /api/v1/fine-tunes/compress/jobs/{job_id}/logs */
+export function finetuneCompressLogsPath(jobId: string): string {
+  return `/api/v1/fine-tunes/compress/jobs/${encodeURIComponent(jobId)}/logs`;
+}
+
 // ---- Model Deployments (DashScope /api/v1/deployments) ----
 
 /** POST (create) and GET (list) endpoint. */
diff --git a/packages/core/src/client/index.ts b/packages/core/src/client/index.ts
index c1063462a..26b6a9d6b 100644
--- a/packages/core/src/client/index.ts
+++ b/packages/core/src/client/index.ts
@@ -28,6 +28,7 @@ export {
   sandboxTemplatePath,
   sandboxTemplateBuildStatusPath,
   responsesPath,
+  decisionPath,
   agentStudioHost,
   securityOverviewEndpoint,
   securityAgentLogsEndpoint,
diff --git a/packages/core/src/dataset/index.ts b/packages/core/src/dataset/index.ts
index b6e581c23..ea214c9fb 100644
--- a/packages/core/src/dataset/index.ts
+++ b/packages/core/src/dataset/index.ts
@@ -8,6 +8,7 @@ export {
   listSupportedFormats,
   MAX_DATASET_BYTES,
   MAX_CPT_BYTES,
+  MAX_DECISION_DATASET_BYTES,
   MAX_MEDIA_ZIP_BYTES,
   parseDatasetSchemaFlag,
   formatIssue,
diff --git a/packages/core/src/dataset/validate/common.ts b/packages/core/src/dataset/validate/common.ts
index 212fe7ba3..f9784307a 100644
--- a/packages/core/src/dataset/validate/common.ts
+++ b/packages/core/src/dataset/validate/common.ts
@@ -24,6 +24,9 @@ export const MAX_DATASET_BYTES = 200 * 1024 * 1024;
  */
 export const MAX_CPT_BYTES = 300 * 1024 * 1024;
 
+/** 决策模型训练数据单文件上限，按平台文档为 300MB。 */
+export const MAX_DECISION_DATASET_BYTES = 300 * 1024 * 1024;
+
 /**
  * Image / video ZIP size cap — 2 GB per the platform docs. Used by
  * `bl dataset upload` for media schemas and by the `sft-lora` training
@@ -88,12 +91,20 @@ export function emptyStats(): ValidationStats {
 export function parseDatasetSchemaFlag(value: string | undefined): DatasetSchema | undefined {
   if (value === undefined || value.trim() === "") return undefined;
   const v = value.trim();
-  if (v === "chatml" || v === "dpo" || v === "cpt" || v === "tts" || v === "image" || v === "video")
+  if (
+    v === "chatml" ||
+    v === "dpo" ||
+    v === "cpt" ||
+    v === "decision" ||
+    v === "tts" ||
+    v === "image" ||
+    v === "video"
+  )
     return v;
   throw new BailianError(
-    `Unsupported --schema "${value}". Supported: chatml, dpo, cpt, tts, image.`,
+    `Unsupported --schema "${value}". Supported: chatml, dpo, cpt, decision, tts, image, video. / 不支持的 --schema；可选值：chatml、dpo、cpt、decision、tts、image、video。`,
     ExitCode.USAGE,
-    `Omit --schema to auto-detect per record (chosen/rejected → DPO, text → CPT, wav_fn → TTS, img_path → image, else ChatML).`,
+    `Omit --schema to auto-detect per record (state/questions → decision, chosen/rejected → DPO, text → CPT, wav_fn → TTS, img_path → image, first_frame_path/video_path → video, else ChatML). / 省略 --schema 可逐条自动识别。`,
   );
 }
 
diff --git a/packages/core/src/dataset/validate/index.ts b/packages/core/src/dataset/validate/index.ts
index 169a32c7b..43e6418c6 100644
--- a/packages/core/src/dataset/validate/index.ts
+++ b/packages/core/src/dataset/validate/index.ts
@@ -7,6 +7,7 @@ export {
 export {
   MAX_DATASET_BYTES,
   MAX_CPT_BYTES,
+  MAX_DECISION_DATASET_BYTES,
   MAX_MEDIA_ZIP_BYTES,
   parseDatasetSchemaFlag,
 } from "./common.ts";
diff --git a/packages/core/src/dataset/validate/schemas/decision.ts b/packages/core/src/dataset/validate/schemas/decision.ts
new file mode 100644
index 000000000..04934ad24
--- /dev/null
+++ b/packages/core/src/dataset/validate/schemas/decision.ts
@@ -0,0 +1,227 @@
+/**
+ * 决策模型训练记录 Schema。
+ *
+ * 每条记录由输入状态 `state` 和问题定义 `questions` 组成。问题支持：
+ * - choice：封闭选项分类；
+ * - noul：是非判断；
+ * - score：从低到高的离散评分。
+ *
+ * `state` 与问题定义的结构规则与推理 API 共用（见 `decision/questions.ts`）；
+ * 训练侧额外要求 `label` 硬标签，`target` 是可选软标签概率分布。
+ */
+import {
+  inspectDecisionQuestion,
+  inspectDecisionQuestionId,
+  inspectDecisionQuestionTable,
+  inspectDecisionState,
+  isDecisionObject,
+  type DecisionIssue,
+} from "../../../decision/questions.ts";
+import { makeIssue } from "../common.ts";
+import type { ValidationIssue } from "../types.ts";
+import type { RecordSchemaSpec } from "./types.ts";
+
+const TARGET_SUM_TOLERANCE = 1e-6;
+
+function decisionIssue(
+  code: string,
+  message: string,
+  lineNo: number,
+  path: string,
+): ValidationIssue {
+  return makeIssue("error", code, message, { line: lineNo, path });
+}
+
+function toValidationIssue(issue: DecisionIssue, lineNo: number): ValidationIssue {
+  return decisionIssue(
+    issue.code,
+    `${issue.message["en-US"]} / ${issue.message["zh-CN"]}`,
+    lineNo,
+    issue.path,
+  );
+}
+
+function validateTarget(
+  target: unknown,
+  expectedKeys: string[],
+  lineNo: number,
+  path: string,
+): ValidationIssue[] {
+  if (!isDecisionObject(target)) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_TARGET",
+        `"${path}" must be an object containing a complete probability distribution. / "${path}" 必须是包含完整概率分布的对象。`,
+        lineNo,
+        path,
+      ),
+    ];
+  }
+
+  const targetKeys = Object.keys(target).sort();
+  const sortedExpectedKeys = [...expectedKeys].sort();
+  if (
+    targetKeys.length !== sortedExpectedKeys.length ||
+    targetKeys.some((targetKey, index) => targetKey !== sortedExpectedKeys[index])
+  ) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_TARGET_KEYS",
+        `"${path}" must contain exactly these keys: ${sortedExpectedKeys.join(", ")}. / "${path}" 必须且只能包含这些键：${sortedExpectedKeys.join(", ")}。`,
+        lineNo,
+        path,
+      ),
+    ];
+  }
+
+  let probabilitySum = 0;
+  for (const targetKey of sortedExpectedKeys) {
+    const probability = target[targetKey];
+    if (
+      typeof probability !== "number" ||
+      !Number.isFinite(probability) ||
+      probability < 0 ||
+      probability > 1
+    ) {
+      return [
+        decisionIssue(
+          "DECISION_INVALID_TARGET_PROBABILITY",
+          `"${path}.${targetKey}" must be a finite number in [0, 1]. / "${path}.${targetKey}" 必须是 [0, 1] 范围内的有限数字。`,
+          lineNo,
+          `${path}.${targetKey}`,
+        ),
+      ];
+    }
+    probabilitySum += probability;
+  }
+
+  if (Math.abs(probabilitySum - 1) > TARGET_SUM_TOLERANCE) {
+    return [
+      decisionIssue(
+        "DECISION_TARGET_SUM_NOT_ONE",
+        `Probabilities in "${path}" must sum to 1 (got ${probabilitySum}). / "${path}" 中的概率之和必须为 1（当前为 ${probabilitySum}）。`,
+        lineNo,
+        path,
+      ),
+    ];
+  }
+
+  return [];
+}
+
+/** 问题定义已通过共享校验后，再校验 choice 的 label / target。 */
+function validateChoiceLabels(
+  question: Record<string, unknown>,
+  lineNo: number,
+  path: string,
+): ValidationIssue[] {
+  const optionKeys = Object.keys(question.criteria as Record<string, unknown>);
+  if (typeof question.label !== "string" || !optionKeys.includes(question.label)) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_CHOICE_LABEL",
+        `"${path}.label" must be one of: ${optionKeys.join(", ")}. / "${path}.label" 必须是以下选项之一：${optionKeys.join(", ")}。`,
+        lineNo,
+        `${path}.label`,
+      ),
+    ];
+  }
+
+  return question.target === undefined
+    ? []
+    : validateTarget(question.target, optionKeys, lineNo, `${path}.target`);
+}
+
+function validateNoulLabels(
+  question: Record<string, unknown>,
+  lineNo: number,
+  path: string,
+): ValidationIssue[] {
+  if (typeof question.label !== "boolean") {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_NOUL_LABEL",
+        `"${path}.label" must be true or false. / "${path}.label" 必须是 true 或 false。`,
+        lineNo,
+        `${path}.label`,
+      ),
+    ];
+  }
+
+  return question.target === undefined
+    ? []
+    : validateTarget(question.target, ["true", "false"], lineNo, `${path}.target`);
+}
+
+function validateScoreLabels(
+  question: Record<string, unknown>,
+  lineNo: number,
+  path: string,
+): ValidationIssue[] {
+  const levelCount = (question.criteria as unknown[]).length;
+  if (
+    typeof question.label !== "number" ||
+    !Number.isInteger(question.label) ||
+    question.label < 0 ||
+    question.label >= levelCount
+  ) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_SCORE_LABEL",
+        `"${path}.label" must be an integer from 0 to ${levelCount - 1}. / "${path}.label" 必须是 0 到 ${levelCount - 1} 之间的整数。`,
+        lineNo,
+        `${path}.label`,
+      ),
+    ];
+  }
+
+  const scoreKeys = Array.from({ length: levelCount }, (_level, levelIndex) => String(levelIndex));
+  return question.target === undefined
+    ? []
+    : validateTarget(question.target, scoreKeys, lineNo, `${path}.target`);
+}
+
+function inspectDecisionRecord(record: Record<string, unknown>, lineNo: number): ValidationIssue[] {
+  const issues: ValidationIssue[] = inspectDecisionState(record.state).map((issue) =>
+    toValidationIssue(issue, lineNo),
+  );
+
+  const questions = record.questions;
+  const tableIssues = inspectDecisionQuestionTable(questions);
+  if (tableIssues.length > 0) {
+    issues.push(...tableIssues.map((issue) => toValidationIssue(issue, lineNo)));
+    return issues;
+  }
+
+  for (const [questionId, questionValue] of Object.entries(questions as Record<string, unknown>)) {
+    const idIssues = inspectDecisionQuestionId(questionId);
+    if (idIssues.length > 0) {
+      issues.push(...idIssues.map((issue) => toValidationIssue(issue, lineNo)));
+      continue;
+    }
+
+    const path = `questions.${questionId}`;
+    const definitionIssues = inspectDecisionQuestion(questionValue, path);
+    if (definitionIssues.length > 0) {
+      issues.push(...definitionIssues.map((issue) => toValidationIssue(issue, lineNo)));
+      continue;
+    }
+
+    const question = questionValue as Record<string, unknown>;
+    if (question.type === "choice") {
+      issues.push(...validateChoiceLabels(question, lineNo, path));
+    } else if (question.type === "noul") {
+      issues.push(...validateNoulLabels(question, lineNo, path));
+    } else {
+      issues.push(...validateScoreLabels(question, lineNo, path));
+    }
+  }
+
+  return issues;
+}
+
+export const decisionSchema: RecordSchemaSpec = {
+  name: "decision",
+  detect: (record) => !("messages" in record) && ("questions" in record || "state" in record),
+  inspect: inspectDecisionRecord,
+};
diff --git a/packages/core/src/dataset/validate/schemas/index.ts b/packages/core/src/dataset/validate/schemas/index.ts
index 19cdeb80b..6652c9236 100644
--- a/packages/core/src/dataset/validate/schemas/index.ts
+++ b/packages/core/src/dataset/validate/schemas/index.ts
@@ -16,17 +16,18 @@ import type { RecordSchemaSpec } from "./types.ts";
 import { chatmlSchema } from "./chatml.ts";
 import { cptSchema } from "./cpt.ts";
 import { dpoSchema } from "./dpo.ts";
+import { decisionSchema } from "./decision.ts";
 import { ttsSchema } from "./tts.ts";
 import { imageSchema } from "./image.ts";
 import { videoSchema } from "./video.ts";
 
-// Order matters: TTS (wav_fn), image (img_path), video (first_frame_path/
-// video_path), DPO (chosen/rejected) and CPT (text) before ChatML (the catch-
-// all fallback). Each keys off a distinguishing field so they partition cleanly.
+// 顺序决定自动识别优先级：所有特定 Schema 必须位于 ChatML 兜底项之前。
+// decision 仅识别 questions，或不含 messages 的 state，避免影响扩展了 state 的 ChatML 记录。
 export const RECORD_SCHEMAS: RecordSchemaSpec[] = [
   ttsSchema,
   imageSchema,
   videoSchema,
+  decisionSchema,
   dpoSchema,
   cptSchema,
   chatmlSchema,
diff --git a/packages/core/src/dataset/validate/types.ts b/packages/core/src/dataset/validate/types.ts
index ea8cfced0..d88ad1214 100644
--- a/packages/core/src/dataset/validate/types.ts
+++ b/packages/core/src/dataset/validate/types.ts
@@ -24,9 +24,11 @@ export interface ValidateOpts {
    *                   Every record MUST carry `chosen` + `rejected`.
    *   - `"cpt"`    — `{text: "..."}` (continual pre-training). Raw text only,
    *                   no `messages[]`.
-   *   - `undefined` — auto-detect per record: a record with `chosen` or
-   *                   `rejected` is validated as DPO, one with `text` (and no
-   *                   `messages`) as CPT, otherwise as ChatML.
+   *   - `"decision"` — `{state:{content}, questions:{...}}`（决策模型训练数据）。
+   *   - `undefined` — auto-detect per record: a record with `state`/`questions`
+   *                   is validated as decision data, one with `chosen` or
+   *                   `rejected` as DPO, one with `text` (and no `messages`) as
+   *                   CPT, otherwise as ChatML.
    * `finetune create` sets this from `--training-type` (dpo* → "dpo",
    * cpt → "cpt") so a malformed dataset fails at validate time, not on the
    * platform ten minutes in.
@@ -42,7 +44,7 @@ export interface ValidateOpts {
 }
 
 /** The schemas a `.jsonl` record can be validated against. */
-export type DatasetSchema = "chatml" | "dpo" | "cpt" | "tts" | "image" | "video";
+export type DatasetSchema = "chatml" | "dpo" | "cpt" | "decision" | "tts" | "image" | "video";
 
 export type ValidationSeverity = "error" | "warning";
 
diff --git a/packages/core/src/decision/index.ts b/packages/core/src/decision/index.ts
new file mode 100644
index 000000000..51d3f7cec
--- /dev/null
+++ b/packages/core/src/decision/index.ts
@@ -0,0 +1,2 @@
+export * from "./questions.ts";
+export * from "./request.ts";
diff --git a/packages/core/src/decision/questions.ts b/packages/core/src/decision/questions.ts
new file mode 100644
index 000000000..29e83e841
--- /dev/null
+++ b/packages/core/src/decision/questions.ts
@@ -0,0 +1,232 @@
+/**
+ * 决策模型（TypeSafe System One）问题定义与 state 的结构校验。
+ *
+ * 推理请求（`bl decision`）与训练数据校验（`decision` record schema）共用这里的
+ * 规则；训练侧在此之上再校验 `label` / `target`。规则以决策模型 API 文档为准：
+ * - state：非空字符串、对象或数组；
+ * - instructions：可选字符串；
+ * - choice.criteria：选项名 → 描述，1–255 项；
+ * - score.criteria：从低到高的等级描述数组，2–255 级；
+ * - noul.criteria：可选，只能包含 `true` / `false` 描述。
+ */
+
+export const DECISION_QUESTION_TYPES = ["choice", "noul", "score"] as const;
+export type DecisionQuestionType = (typeof DECISION_QUESTION_TYPES)[number];
+
+export const MAX_DECISION_CHOICE_OPTIONS = 255;
+export const MIN_DECISION_SCORE_LEVELS = 2;
+export const MAX_DECISION_SCORE_LEVELS = 255;
+
+const NOUL_CRITERIA_KEYS = new Set(["true", "false"]);
+
+export interface DecisionIssue {
+  code: string;
+  /** 双语文案：命令层用 `ctx.localize` 渲染，训练校验拼成 `英文 / 中文`。 */
+  message: { readonly "en-US": string; readonly "zh-CN": string };
+  path: string;
+}
+
+export function isDecisionObject(value: unknown): value is Record<string, unknown> {
+  return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+
+function isNonEmptyString(value: unknown): value is string {
+  return typeof value === "string" && value.trim().length > 0;
+}
+
+function decisionIssue(
+  code: string,
+  path: string,
+  english: string,
+  chinese: string,
+): DecisionIssue {
+  return { code, path, message: { "en-US": english, "zh-CN": chinese } };
+}
+
+export function inspectDecisionState(state: unknown, path = "state"): DecisionIssue[] {
+  const isValid =
+    isNonEmptyString(state) ||
+    (Array.isArray(state) && state.length > 0) ||
+    (isDecisionObject(state) && Object.keys(state).length > 0);
+  if (isValid) return [];
+  return [
+    decisionIssue(
+      "DECISION_INVALID_STATE",
+      path,
+      `"${path}" must be a non-empty string, object, or array.`,
+      `"${path}" 必须是非空字符串、对象或数组。`,
+    ),
+  ];
+}
+
+function inspectChoiceCriteria(criteria: unknown, path: string): DecisionIssue[] {
+  const criteriaPath = `${path}.criteria`;
+  if (!isDecisionObject(criteria)) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_CHOICE_CRITERIA",
+        criteriaPath,
+        `"${criteriaPath}" must be an option-to-description object.`,
+        `"${criteriaPath}" 必须是“选项名到描述”的对象。`,
+      ),
+    ];
+  }
+  const optionCount = Object.keys(criteria).length;
+  if (optionCount < 1 || optionCount > MAX_DECISION_CHOICE_OPTIONS) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_CHOICE_CRITERIA",
+        criteriaPath,
+        `"${criteriaPath}" must contain 1-${MAX_DECISION_CHOICE_OPTIONS} options (got ${optionCount}).`,
+        `"${criteriaPath}" 必须包含 1-${MAX_DECISION_CHOICE_OPTIONS} 个选项（当前 ${optionCount} 个）。`,
+      ),
+    ];
+  }
+  for (const [optionName, description] of Object.entries(criteria)) {
+    if (!isNonEmptyString(optionName) || !isNonEmptyString(description)) {
+      return [
+        decisionIssue(
+          "DECISION_INVALID_CHOICE_CRITERIA",
+          `${criteriaPath}.${optionName}`,
+          `Choice option names and descriptions in "${criteriaPath}" must be non-empty strings.`,
+          `"${criteriaPath}" 中的选项名和描述必须是非空字符串。`,
+        ),
+      ];
+    }
+  }
+  return [];
+}
+
+function inspectScoreCriteria(criteria: unknown, path: string): DecisionIssue[] {
+  const criteriaPath = `${path}.criteria`;
+  if (!Array.isArray(criteria) || criteria.some((level) => !isNonEmptyString(level))) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_SCORE_CRITERIA",
+        criteriaPath,
+        `"${criteriaPath}" must be an array of non-empty level descriptions, from low to high.`,
+        `"${criteriaPath}" 必须是从低到高排列、由非空等级描述组成的数组。`,
+      ),
+    ];
+  }
+  if (criteria.length < MIN_DECISION_SCORE_LEVELS || criteria.length > MAX_DECISION_SCORE_LEVELS) {
+    return [
+      decisionIssue(
+        "DECISION_INVALID_SCORE_CRITERIA",
+        criteriaPath,
+        `"${criteriaPath}" must contain ${MIN_DECISION_SCORE_LEVELS}-${MAX_DECISION_SCORE_LEVELS} levels (got ${criteria.length}).`,
+        `"${criteriaPath}" 必须包含 ${MIN_DECISION_SCORE_LEVELS}-${MAX_DECISION_SCORE_LEVELS} 个等级（当前 ${criteria.length} 个）。`,
+      ),
+    ];
+  }
+  return [];
+}
+
+function inspectNoulCriteria(criteria: unknown, path: string): DecisionIssue[] {
+  if (criteria === undefined) return [];
+  const criteriaPath = `${path}.criteria`;
+  const isValid =
+    isDecisionObject(criteria) &&
+    Object.entries(criteria).every(
+      ([criterionKey, description]) =>
+        NOUL_CRITERIA_KEYS.has(criterionKey) && typeof description === "string",
+    );
+  if (isValid) return [];
+  return [
+    decisionIssue(
+      "DECISION_INVALID_NOUL_CRITERIA",
+      criteriaPath,
+      `"${criteriaPath}" is optional; when provided it must be an object with only "true" / "false" string descriptions.`,
+      `"${criteriaPath}" 为可选项；填写时必须是只包含 "true" / "false" 字符串描述的对象。`,
+    ),
+  ];
+}
+
+/** 校验单个问题定义（不含训练专用的 label / target）。 */
+export function inspectDecisionQuestion(question: unknown, path: string): DecisionIssue[] {
+  if (!isDecisionObject(question)) {
+    return [
+      decisionIssue(
+        "DECISION_QUESTION_NOT_OBJECT",
+        path,
+        `"${path}" must be an object.`,
+        `"${path}" 必须是对象。`,
+      ),
+    ];
+  }
+
+  const issues: DecisionIssue[] = [];
+  if (question.instructions !== undefined && typeof question.instructions !== "string") {
+    issues.push(
+      decisionIssue(
+        "DECISION_INVALID_INSTRUCTIONS",
+        `${path}.instructions`,
+        `"${path}.instructions" must be a string when provided.`,
+        `"${path}.instructions" 填写时必须是字符串。`,
+      ),
+    );
+  }
+
+  if (question.type === "choice") {
+    issues.push(...inspectChoiceCriteria(question.criteria, path));
+  } else if (question.type === "noul") {
+    issues.push(...inspectNoulCriteria(question.criteria, path));
+  } else if (question.type === "score") {
+    issues.push(...inspectScoreCriteria(question.criteria, path));
+  } else {
+    issues.push(
+      decisionIssue(
+        "DECISION_INVALID_QUESTION_TYPE",
+        `${path}.type`,
+        `"${path}.type" must be choice, noul, or score.`,
+        `"${path}.type" 必须是 choice、noul 或 score。`,
+      ),
+    );
+  }
+  return issues;
+}
+
+/** 校验问题表本身是否为非空对象。 */
+export function inspectDecisionQuestionTable(
+  questions: unknown,
+  path = "questions",
+): DecisionIssue[] {
+  if (isDecisionObject(questions) && Object.keys(questions).length > 0) return [];
+  return [
+    decisionIssue(
+      "DECISION_INVALID_QUESTIONS",
+      path,
+      `"${path}" must be a non-empty object keyed by question ID.`,
+      `"${path}" 必须是以问题 ID 为键的非空对象。`,
+    ),
+  ];
+}
+
+export function inspectDecisionQuestionId(questionId: string, path = "questions"): DecisionIssue[] {
+  if (isNonEmptyString(questionId)) return [];
+  return [
+    decisionIssue(
+      "DECISION_INVALID_QUESTION_ID",
+      path,
+      "Question IDs must be non-empty strings.",
+      "问题 ID 必须是非空字符串。",
+    ),
+  ];
+}
+
+/** 校验完整问题表：表结构、每个问题 ID 与问题定义。 */
+export function inspectDecisionQuestions(questions: unknown, path = "questions"): DecisionIssue[] {
+  const tableIssues = inspectDecisionQuestionTable(questions, path);
+  if (tableIssues.length > 0) return tableIssues;
+
+  const issues: DecisionIssue[] = [];
+  for (const [questionId, question] of Object.entries(questions as Record<string, unknown>)) {
+    const idIssues = inspectDecisionQuestionId(questionId, path);
+    if (idIssues.length > 0) {
+      issues.push(...idIssues);
+      continue;
+    }
+    issues.push(...inspectDecisionQuestion(question, `${path}.${questionId}`));
+  }
+  return issues;
+}
diff --git a/packages/core/src/decision/request.ts b/packages/core/src/decision/request.ts
new file mode 100644
index 000000000..3fd62702e
--- /dev/null
+++ b/packages/core/src/decision/request.ts
@@ -0,0 +1,80 @@
+/**
+ * 决策模型推理请求 / 响应类型与请求构造。
+ *
+ * 协议：TypeSafe System One，`POST {workspace host}/compatible-mode/v1/systemone`。
+ */
+import type { DecisionQuestionType } from "./questions.ts";
+import { isDecisionObject } from "./questions.ts";
+
+export const DEFAULT_DECISION_MODEL = "decision-model-preview";
+
+/** 决策模型开放的 workspace 域名地域。 */
+export const DECISION_REGIONS = ["cn-beijing", "ap-southeast-1"] as const;
+export type DecisionRegion = (typeof DECISION_REGIONS)[number];
+export const DEFAULT_DECISION_REGION: DecisionRegion = "cn-beijing";
+
+/** 训练专用字段，推理请求中必须去掉。 */
+const TRAINING_ONLY_QUESTION_FIELDS = ["label", "target"] as const;
+
+export type DecisionState = string | Record<string, unknown> | unknown[];
+
+export interface DecisionQuestion {
+  type: DecisionQuestionType;
+  instructions?: string;
+  criteria?: Record<string, string> | string[];
+  [key: string]: unknown;
+}
+
+export interface DecisionRequest {
+  model: string;
+  state: DecisionState;
+  questions: Record<string, DecisionQuestion>;
+}
+
+export interface DecisionAnswer {
+  type?: string;
+  choice?: string;
+  noul?: number;
+  score?: number;
+  confidence?: number;
+  probabilities?: Record<string, number>;
+  legend?: Record<string, string>;
+  [key: string]: unknown;
+}
+
+export interface DecisionResponse {
+  model?: string;
+  request_id?: string;
+  answers?: Record<string, DecisionAnswer>;
+  usage?: { input_tokens?: number; [key: string]: unknown };
+  latency_ms?: number;
+  [key: string]: unknown;
+}
+
+/**
+ * 兼容两种问题文件：裸问题表 `{ id: {...} }`，或外层包了一层的 `{ "questions": {...} }`
+ * （如示例包里的 workload.json / 训练样本）。当 `questions` 本身像一个问题定义
+ * （带字符串 `type`）时，视为 ID 恰好叫 "questions" 的裸问题表，不做解包。
+ */
+export function unwrapDecisionQuestions(parsed: unknown): unknown {
+  if (!isDecisionObject(parsed)) return parsed;
+  const nested = parsed.questions;
+  if (isDecisionObject(nested) && typeof nested.type !== "string") return nested;
+  return parsed;
+}
+
+/** 去掉每个问题上的训练字段（label / target），返回新对象，不修改入参。 */
+export function stripDecisionTrainingFields(questions: unknown): unknown {
+  if (!isDecisionObject(questions)) return questions;
+  const stripped: Record<string, unknown> = {};
+  for (const [questionId, question] of Object.entries(questions)) {
+    if (!isDecisionObject(question)) {
+      stripped[questionId] = question;
+      continue;
+    }
+    const definition = { ...question };
+    for (const field of TRAINING_ONLY_QUESTION_FIELDS) delete definition[field];
+    stripped[questionId] = definition;
+  }
+  return stripped;
+}
diff --git a/packages/core/src/finetune/compress.ts b/packages/core/src/finetune/compress.ts
new file mode 100644
index 000000000..747a49b14
--- /dev/null
+++ b/packages/core/src/finetune/compress.ts
@@ -0,0 +1,172 @@
+/**
+ * Model compression (quantization) HTTP API wrappers.
+ *
+ * Thin functions over `requestJson` for DashScope `/api/v1/fine-tunes/compress`.
+ * They return the parsed body verbatim (snake_case) so callers decide how to
+ * surface fields.
+ */
+import {
+  finetuneCompressTemplatesPath,
+  finetuneCompressJobsPath,
+  finetuneCompressJobPath,
+  finetuneCompressCancelPath,
+  finetuneCompressLogsPath,
+} from "../client/endpoints.ts";
+import type { Client } from "../client/client.ts";
+import type {
+  CreateCompressJobRequest,
+  CompressJobResponse,
+  ListCompressJobsResponse,
+  ListCompressTemplatesResponse,
+  GetCompressJobLogsResponse,
+  CompressJobActionResponse,
+} from "./types.ts";
+
+/** Append defined, non-empty params as a query string. */
+function withQuery(base: string, params: Record<string, string | number | undefined>): string {
+  const qs = new URLSearchParams();
+  for (const [key, value] of Object.entries(params)) {
+    if (value !== undefined && value !== "") qs.set(key, String(value));
+  }
+  return qs.toString() ? `${base}?${qs.toString()}` : base;
+}
+
+export interface ListCompressTemplatesParams {
+  /** Filter by custom model ID, or a base model name to list all derived custom models. */
+  model?: string;
+  /** Response language: `zh-CN` | `en-US`. */
+  lang?: string;
+  signal?: AbortSignal;
+}
+
+/** GET /api/v1/fine-tunes/compress/templates */
+export async function listCompressTemplates(
+  client: Client,
+  params: ListCompressTemplatesParams = {},
+): Promise<ListCompressTemplatesResponse> {
+  return client.requestJson<ListCompressTemplatesResponse>({
+    path: withQuery(finetuneCompressTemplatesPath(), { model: params.model, lang: params.lang }),
+    method: "GET",
+    signal: params.signal,
+  });
+}
+
+/** POST /api/v1/fine-tunes/compress/jobs */
+export async function createCompressJob(
+  client: Client,
+  body: CreateCompressJobRequest,
+  signal?: AbortSignal,
+): Promise<CompressJobResponse> {
+  return client.requestJson<CompressJobResponse>({
+    path: finetuneCompressJobsPath(),
+    method: "POST",
+    body,
+    signal,
+  });
+}
+
+export interface ListCompressJobsParams {
+  status?: string;
+  model?: string;
+  templateId?: string;
+  quantSpec?: string;
+  quantMethod?: string;
+  startTime?: string;
+  endTime?: string;
+  searchKey?: string;
+  /** `job_id` | `job_name`; narrows `searchKey` to one field. */
+  selectKey?: string;
+  sortOrder?: string;
+  pageNo?: number;
+  pageSize?: number;
+  signal?: AbortSignal;
+}
+
+/** GET /api/v1/fine-tunes/compress/jobs */
+export async function listCompressJobs(
+  client: Client,
+  params: ListCompressJobsParams = {},
+): Promise<ListCompressJobsResponse> {
+  const path = withQuery(finetuneCompressJobsPath(), {
+    status: params.status,
+    model: params.model,
+    template_id: params.templateId,
+    quant_spec: params.quantSpec,
+    quant_method: params.quantMethod,
+    start_time: params.startTime,
+    end_time: params.endTime,
+    search_key: params.searchKey,
+    select_key: params.selectKey,
+    sort_order: params.sortOrder,
+    page_no: params.pageNo,
+    page_size: params.pageSize,
+  });
+  return client.requestJson<ListCompressJobsResponse>({
+    path,
+    method: "GET",
+    signal: params.signal,
+  });
+}
+
+/** GET /api/v1/fine-tunes/compress/jobs/{job_id} */
+export async function getCompressJob(
+  client: Client,
+  jobId: string,
+  signal?: AbortSignal,
+): Promise<CompressJobResponse> {
+  return client.requestJson<CompressJobResponse>({
+    path: finetuneCompressJobPath(jobId),
+    method: "GET",
+    signal,
+  });
+}
+
+export interface GetCompressJobLogsParams {
+  /** Skip the first N lines. */
+  offset?: number;
+  /** Number of lines to read (server cap: 1000). */
+  line?: number;
+  signal?: AbortSignal;
+}
+
+/** GET /api/v1/fine-tunes/compress/jobs/{job_id}/logs */
+export async function getCompressJobLogs(
+  client: Client,
+  jobId: string,
+  params: GetCompressJobLogsParams = {},
+): Promise<GetCompressJobLogsResponse> {
+  return client.requestJson<GetCompressJobLogsResponse>({
+    path: withQuery(finetuneCompressLogsPath(jobId), {
+      offset: params.offset,
+      line: params.line,
+    }),
+    method: "GET",
+    signal: params.signal,
+  });
+}
+
+/** POST /api/v1/fine-tunes/compress/jobs/{job_id}/cancel */
+export async function cancelCompressJob(
+  client: Client,
+  jobId: string,
+  signal?: AbortSignal,
+): Promise<CompressJobActionResponse> {
+  return client.requestJson<CompressJobActionResponse>({
+    path: finetuneCompressCancelPath(jobId),
+    method: "POST",
+    signal,
+  });
+}
+
+/** DELETE /api/v1/fine-tunes/compress/jobs/{job_id} */
+export async function deleteCompressJob(
+  client: Client,
+  jobId: string,
+  signal?: AbortSignal,
+): Promise<CompressJobActionResponse> {
+  return client.requestJson<CompressJobActionResponse>({
+    path: finetuneCompressJobPath(jobId),
+    method: "DELETE",
+    signal,
+  });
+}
diff --git a/packages/core/src/finetune/index.ts b/packages/core/src/finetune/index.ts
index 3b82e5831..4b3f261d3 100644
--- a/packages/core/src/finetune/index.ts
+++ b/packages/core/src/finetune/index.ts
@@ -1,5 +1,6 @@
 export * from "./types.ts";
 export * from "./api.ts";
+export * from "./compress.ts";
 export * from "./capability.ts";
 export * from "./preflight.ts";
 export * from "./profiles/index.ts";
diff --git a/packages/core/src/finetune/types.ts b/packages/core/src/finetune/types.ts
index 82241e419..416ff39d1 100644
--- a/packages/core/src/finetune/types.ts
+++ b/packages/core/src/finetune/types.ts
@@ -200,3 +200,128 @@ export interface GetFineTuneResponse {
   output?: FineTuneJob;
   data?: FineTuneJob;
 }
+
+// ---- Model Compression (DashScope /api/v1/fine-tunes/compress) ----
+
+/** A tunable hyper-parameter declared by a compression template. */
+export interface CompressHyperParameter {
+  name?: string;
+  /** `number` (data_range / step) or `string` (support_values). */
+  type?: string;
+  display_name?: string;
+  description?: string;
+  defaultValue?: string;
+  recommend_value?: string;
+  required?: boolean;
+  support_values?: string[];
+  data_range?: string[];
+  step?: number;
+  [k: string]: unknown;
+}
+
+/** A compression template bound to a quantizable model. */
+export interface CompressTemplate {
+  template_id?: string;
+  template_name?: string;
+  description?: string;
+  /** Currently always `quantization`. */
+  compress_type?: string;
+  hyper_parameters?: CompressHyperParameter[];
+  [k: string]: unknown;
+}
+
+/** A quantizable custom model and the templates it supports. */
+export interface CompressModel {
+  model?: string;
+  model_name?: string;
+  base_model?: string;
+  templates?: CompressTemplate[];
+  [k: string]: unknown;
+}
+
+/** GET /api/v1/fine-tunes/compress/templates response. */
+export interface ListCompressTemplatesResponse {
+  request_id?: string;
+  output?: {
+    base_models?: string[];
+    custom_models?: CompressModel[];
+    [k: string]: unknown;
+  };
+}
+
+/** POST /api/v1/fine-tunes/compress/jobs request body. */
+export interface CreateCompressJobRequest {
+  /** Source custom (full-parameter fine-tuned) model ID. */
+  model: string;
+  template_id: string;
+  job_name?: string;
+  job_description?: string;
+  /** Key-value overrides; only the keys to override. */
+  hyper_parameters?: Record<string, unknown>;
+  /** Calibration dataset file IDs; only effective when `calib_input=true`. */
+  custom_calibration_file_ids?: string[];
+  /** Output model name suffix (≤ 8 chars, lowercase letters and digits). */
+  output_model_suffix?: string;
+}
+
+/** A compression job record as returned by the platform. */
+export interface CompressJob {
+  job_id?: string;
+  job_name?: string;
+  job_description?: string;
+  /** PENDING | QUEUING | RUNNING | CANCELING | SUCCEEDED | FAILED | CANCELED */
+  status?: string;
+  model?: string;
+  base_model?: string;
+  template_id?: string;
+  template_name?: string;
+  template_description?: string;
+  training_type?: string;
+  compress_type?: string;
+  hyper_parameters?: Record<string, unknown>;
+  custom_calibration_file_ids?: string[];
+  /** Quantized output model ID (only set on SUCCEEDED). */
+  quantized_output?: string | null;
+  create_time?: string;
+  start_time?: string | null;
+  end_time?: string | null;
+  error?: { code?: string; message?: string; [k: string]: unknown } | null;
+  group?: string;
+  /** GPU seconds; present on SUCCEEDED / CANCELED. */
+  usage?: number;
+  [k: string]: unknown;
+}
+
+/** POST /api/v1/fine-tunes/compress/jobs and GET /jobs/{job_id} response. */
+export interface CompressJobResponse {
+  request_id?: string;
+  output?: CompressJob;
+}
+
+/** GET /api/v1/fine-tunes/compress/jobs response. */
+export interface ListCompressJobsResponse {
+  request_id?: string;
+  output?: {
+    total?: number;
+    page_no?: number;
+    page_size?: number;
+    jobs?: CompressJob[];
+    [k: string]: unknown;
+  };
+}
+
+/** GET /api/v1/fine-tunes/compress/jobs/{job_id}/logs response. */
+export interface GetCompressJobLogsResponse {
+  request_id?: string;
+  output?: {
+    total?: number;
+    logs?: string[];
+    [k: string]: unknown;
+  };
+}
+
+/** POST /jobs/{job_id}/cancel and DELETE /jobs/{job_id} response. */
+export interface CompressJobActionResponse {
+  request_id?: string;
+  output?: { status?: string; [k: string]: unknown };
+}
diff --git a/packages/core/src/index.ts b/packages/core/src/index.ts
index 3d1b15693..4fd88e068 100644
--- a/packages/core/src/index.ts
+++ b/packages/core/src/index.ts
@@ -11,6 +11,7 @@ export * from "./config/index.ts";
 export * from "./output/index.ts";
 export * from "./files/index.ts";
 export * from "./dataset/index.ts";
+export * from "./decision/index.ts";
 export * from "./finetune/index.ts";
 export * from "./deploy/index.ts";
 export * from "./speech/index.ts";
diff --git a/packages/core/tests/dataset-validate.test.ts b/packages/core/tests/dataset-validate.test.ts
index b478cca65..5a3f3c294 100644
--- a/packages/core/tests/dataset-validate.test.ts
+++ b/packages/core/tests/dataset-validate.test.ts
@@ -144,6 +144,196 @@ describe("validateDataset — DPO schema", () => {
   });
 });
 
+describe("validateDataset — decision schema", () => {
+  const DECISION_OK = JSON.stringify({
+    state: { content: "发票抬头写错了，请重新开票。" },
+    questions: {
+      department: {
+        type: "choice",
+        instructions: "应该由哪个团队处理？",
+        criteria: { billing: "支付和账单问题", technical: "产品故障问题" },
+        label: "billing",
+        target: { billing: 0.96, technical: 0.04 },
+      },
+      escalate: {
+        type: "noul",
+        instructions: "是否需要立即通知值班人员？",
+        label: false,
+        target: { true: 0.1, false: 0.9 },
+      },
+      severity: {
+        type: "score",
+        instructions: "这个问题有多严重？",
+        criteria: ["轻微", "一般", "严重"],
+        label: 0,
+        target: { "0": 1, "1": 0, "2": 0 },
+      },
+    },
+  });
+
+  test("valid decision record passes under auto-detect and --schema decision", async () => {
+    const datasetPath = file("decision_ok.jsonl", [DECISION_OK]);
+    const automatic = await validateDataset(datasetPath, { fullValidate: true });
+    expect(automatic.valid).toBe(true);
+    const explicit = await validateDataset(datasetPath, {
+      fullValidate: true,
+      schema: "decision",
+    });
+    expect(explicit.valid).toBe(true);
+  });
+
+  test("missing state and empty questions are rejected", async () => {
+    const datasetPath = file("decision_missing.jsonl", [JSON.stringify({ questions: {} })]);
+    const result = await validateDataset(datasetPath, { fullValidate: true });
+    expect(result.valid).toBe(false);
+    expect(codes(result).errors).toEqual(
+      expect.arrayContaining(["DECISION_INVALID_STATE", "DECISION_INVALID_QUESTIONS"]),
+    );
+  });
+
+  test("choice criteria and label must use the declared option set", async () => {
+    const datasetPath = file("decision_choice_bad.jsonl", [
+      JSON.stringify({
+        state: { content: "x" },
+        questions: {
+          department: {
+            type: "choice",
+            instructions: "team?",
+            criteria: { billing: "账单" },
+            label: "technical",
+          },
+        },
+      }),
+    ]);
+    const result = await validateDataset(datasetPath, { fullValidate: true });
+    expect(codes(result).errors).toContain("DECISION_INVALID_CHOICE_LABEL");
+  });
+
+  test("noul label must be boolean", async () => {
+    const datasetPath = file("decision_noul_bad.jsonl", [
+      JSON.stringify({
+        state: { content: "x" },
+        questions: {
+          escalate: { type: "noul", instructions: "notify?", label: "false" },
+        },
+      }),
+    ]);
+    const result = await validateDataset(datasetPath, { fullValidate: true });
+    expect(codes(result).errors).toContain("DECISION_INVALID_NOUL_LABEL");
+  });
+
+  test("score label must be an in-range integer", async () => {
+    const datasetPath = file("decision_score_bad.jsonl", [
+      JSON.stringify({
+        state: { content: "x" },
+        questions: {
+          severity: {
+            type: "score",
+            instructions: "severity?",
+            criteria: ["low", "high"],
+            label: 2,
+          },
+        },
+      }),
+    ]);
+    const result = await validateDataset(datasetPath, { fullValidate: true });
+    expect(codes(result).errors).toContain("DECISION_INVALID_SCORE_LABEL");
+  });
+
+  test("target must cover all options and sum to one", async () => {
+    const datasetPath = file("decision_target_bad.jsonl", [
+      JSON.stringify({
+        state: { content: "x" },
+        questions: {
+          department: {
+            type: "choice",
+            instructions: "team?",
+            criteria: { billing: "账单", technical: "技术" },
+            label: "billing",
+            target: { billing: 0.8 },
+          },
+          severity: {
+            type: "score",
+            instructions: "severity?",
+            criteria: ["low", "high"],
+            label: 0,
+            target: { "0": 0.8, "1": 0.3 },
+          },
+        },
+      }),
+    ]);
+    const result = await validateDataset(datasetPath, { fullValidate: true });
+    expect(codes(result).errors).toEqual(
+      expect.arrayContaining(["DECISION_INVALID_TARGET_KEYS", "DECISION_TARGET_SUM_NOT_ONE"]),
+    );
+  });
+
+  test("API-aligned relaxations: string state, optional instructions, noul criteria", async () => {
+    const datasetPath = file("decision_relaxed.jsonl", [
+      JSON.stringify({
+        state: "这段评论是否在索要退款？",
+        questions: {
+          refund: {
+            type: "noul",
+            criteria: { true: "索要退款", false: "未索要退款" },
+            label: true,
+          },
+        },
+      }),
+      JSON.stringify({
+        state: [{ role: "user", content: "我要退款" }],
+        questions: {
+          severity: { type: "score", criteria: ["low", "high"], label: 1 },
+        },
+      }),
+    ]);
+    const result = await validateDataset(datasetPath, { fullValidate: true, schema: "decision" });
+    expect(result.valid).toBe(true);
+  });
+
+  test("criteria bounds follow the API: score needs >= 2 levels, noul only true/false", async () => {
+    const datasetPath = file("decision_bounds_bad.jsonl", [
+      JSON.stringify({
+        state: "x",
+        questions: {
+          severity: { type: "score", criteria: ["only"], label: 0 },
+          refund: { type: "noul", criteria: { maybe: "不确定" }, label: true },
+          team: {
+            type: "choice",
+            instructions: 1,
+            criteria: { billing: "账单" },
+            label: "billing",
+          },
+        },
+      }),
+    ]);
+    const result = await validateDataset(datasetPath, { fullValidate: true, schema: "decision" });
+    expect(codes(result).errors).toEqual(
+      expect.arrayContaining([
+        "DECISION_INVALID_SCORE_CRITERIA",
+        "DECISION_INVALID_NOUL_CRITERIA",
+        "DECISION_INVALID_INSTRUCTIONS",
+      ]),
+    );
+  });
+
+  test("ChatML record with extension state/questions fields remains ChatML", async () => {
+    const datasetPath = file("chatml_with_decision_extensions.jsonl", [
+      JSON.stringify({
+        state: { trace_id: "trace-1" },
+        questions: { source: "legacy-metadata" },
+        messages: [
+          { role: "user", content: "hi" },
+          { role: "assistant", content: "hello" },
+        ],
+      }),
+    ]);
+    const result = await validateDataset(datasetPath, { fullValidate: true });
+    expect(result.valid).toBe(true);
+    expect(codes(result).errors).not.toContain("DECISION_INVALID_QUESTIONS");
+  });
+});
+
 describe("validateDataset — CPT schema", () => {
   const CPT_OK = '{"text":"The quick brown fox jumps over the lazy dog."}';
 
@@ -594,10 +784,11 @@ describe("parseDatasetSchemaFlag", () => {
     expect(parseDatasetSchemaFlag("  ")).toBeUndefined();
   });
 
-  test("chatml / dpo / cpt / tts / image / video pass through", () => {
+  test("chatml / dpo / cpt / decision / tts / image / video pass through", () => {
     expect(parseDatasetSchemaFlag("chatml")).toBe("chatml");
     expect(parseDatasetSchemaFlag("dpo")).toBe("dpo");
     expect(parseDatasetSchemaFlag("cpt")).toBe("cpt");
+    expect(parseDatasetSchemaFlag("decision")).toBe("decision");
     expect(parseDatasetSchemaFlag("tts")).toBe("tts");
     expect(parseDatasetSchemaFlag("image")).toBe("image");
     expect(parseDatasetSchemaFlag("video")).toBe("video");
diff --git a/packages/core/tests/finetune-compress-api.test.ts b/packages/core/tests/finetune-compress-api.test.ts
new file mode 100644
index 000000000..c54c007a6
--- /dev/null
+++ b/packages/core/tests/finetune-compress-api.test.ts
@@ -0,0 +1,64 @@
+import { describe, expect, test, vi } from "vite-plus/test";
+import {
+  cancelCompressJob,
+  createCompressJob,
+  deleteCompressJob,
+  getCompressJob,
+  getCompressJobLogs,
+  listCompressJobs,
+  listCompressTemplates,
+  type Client,
+} from "../src/index.ts";
+
+function mockClient(response: unknown = {}) {
+  const requestJson = vi.fn<Client["requestJson"]>().mockResolvedValue(response);
+  return { client: { requestJson } as unknown as Client, requestJson };
+}
+
+describe("model compression API", () => {
+  test("serializes template and job-list query parameters", async () => {
+    const templates = mockClient();
+    await listCompressTemplates(templates.client, { model: "qwen ft/?", lang: "en-US" });
+    expect(templates.requestJson).toHaveBeenCalledWith({
+      path: "/api/v1/fine-tunes/compress/templates?model=qwen+ft%2F%3F&lang=en-US",
+      method: "GET",
+      signal: undefined,
+    });
+
+    const jobs = mockClient();
+    await listCompressJobs(jobs.client, {
+      status: "SUCCEEDED",
+      model: "qwen-ft",
+      templateId: "tpl-1",
+      quantSpec: "w4a16",
+      quantMethod: "gptq",
+      startTime: "2026-04-01",
+      endTime: "2026-04-30",
+      searchKey: "name",
+      selectKey: "job_name",
+      sortOrder: "asc",
+      pageNo: 2,
+      pageSize: 20,
+    });
+    expect(jobs.requestJson.mock.calls[0]?.[0].path).toBe(
+      "/api/v1/fine-tunes/compress/jobs?status=SUCCEEDED&model=qwen-ft&template_id=tpl-1&quant_spec=w4a16&quant_method=gptq&start_time=2026-04-01&end_time=2026-04-30&search_key=name&select_key=job_name&sort_order=asc&page_no=2&page_size=20",
+    );
+  });
+
+  test("maps create, get, logs, cancel, and delete endpoints", async () => {
+    const { client, requestJson } = mockClient();
+    await createCompressJob(client, { model: "qwen-ft", template_id: "tpl-1" });
+    await getCompressJob(client, "quant /?#");
+    await getCompressJobLogs(client, "quant /?#", { offset: 10, line: 50 });
+    await cancelCompressJob(client, "quant /?#");
+    await deleteCompressJob(client, "quant /?#");
+
+    expect(requestJson.mock.calls.map(([options]) => [options.method, options.path])).toEqual([
+      ["POST", "/api/v1/fine-tunes/compress/jobs"],
+      ["GET", "/api/v1/fine-tunes/compress/jobs/quant%20%2F%3F%23"],
+      ["GET", "/api/v1/fine-tunes/compress/jobs/quant%20%2F%3F%23/logs?offset=10&line=50"],
+      ["POST", "/api/v1/fine-tunes/compress/jobs/quant%20%2F%3F%23/cancel"],
+      ["DELETE", "/api/v1/fine-tunes/compress/jobs/quant%20%2F%3F%23"],
+    ]);
+  });
+});
diff --git a/skills/bailian-cli/SKILL.md b/skills/bailian-cli/SKILL.md
index e8586934b..5c35400b5 100644
--- a/skills/bailian-cli/SKILL.md
+++ b/skills/bailian-cli/SKILL.md
@@ -6,7 +6,7 @@ metadata:
     bins: ["bl"]
 description: >-
   阿里云百炼 / Aliyun Bailian / DashScope 资源管理与 `bl` CLI hub：
-  应用调用（bl app）、应用记忆、知识库检索、模型目录/模型列表、用量/额度/配额、免费额度、
+  应用调用（bl app）、决策模型推理（bl decision）、应用记忆、知识库检索、模型目录/模型列表、用量/额度/配额、免费额度、
   工作空间、MCP 市场、pipeline、文件上传、console API、登录鉴权与配置、
   Agent skill 安装/列表/更新/卸载（bl skill add|list|update|remove，百炼 skill registry）。
   用户点名百炼 / DashScope / `bl`，或继续既有 `bl` 工作流时直接使用。
@@ -58,38 +58,39 @@ Do not guess flags — use the reference files or `--help`.
 
 Use this table only after the decision table in [`bailian-protocol`](../bailian-protocol/SKILL.md#provider-selection-and-consent) has routed the request to `bl` (class 4, or class 2 after the user picks Bailian). Hub-owned intents only — for media / fine-tune / agents.yaml / Sandbox, soft hand-off to the domain skill.
 
-| User intent                                         | Command                                                                                                        | Notes                                                                                                                                  |
-| --------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
-| Explicit Bailian model chat / text execution        | `bl text chat`                                                                                                 | Default `qwen3.8-max`                                                                                                                  |
-| Bailian agent / workflow                            | `bl app call`                                                                                                  | Needs `--app-id`                                                                                                                       |
-| Find app by name                                    | `bl app list` then `bl app call`                                                                               | Console auth                                                                                                                           |
-| Bailian app memory CRUD (not host-agent memory)     | `bl memory *`                                                                                                  | [`reference/memory.md`](reference/memory.md)                                                                                           |
-| Bailian knowledge base RAG                          | `bl knowledge search` / `chat`                                                                                 | API key + agent/workspace IDs                                                                                                          |
-| Upload a file as a step of a Bailian workflow       | `bl file upload`                                                                                               | When you need `oss://` URL explicitly; not for generic hosting                                                                         |
-| Bailian model selection / recommendation            | `bl advisor recommend`                                                                                         | Intent → candidate recall → LLM ranking                                                                                                |
-| Bailian model catalog / pricing / params            | `bl model list`                                                                                                | No auth (public catalog); `--model <family>` for detail, `--enrich` for input params                                                   |
-| Find a model by keyword / modality                  | `bl model search`                                                                                              | Weighted ranking; `--input-modality` / `--output-modality`; offline models hidden by default                                           |
-| SDK sample code for calling a model                 | `bl model code`                                                                                                | `--sdk openai\|dashscope`, `--lang`; text output is the snippet alone, safe to redirect                                                |
-| Install / list / update / remove registry skills    | `bl skill add` / `list` / `update` / `remove`                                                                  | Bailian skill registry; see [`reference/skill.md`](reference/skill.md)                                                                 |
-| Bailian MCP marketplace discovery / call            | `bl mcp list` / `tools` / `call`                                                                               | —                                                                                                                                      |
-| Register or remove a Bailian MCP in an Agent        | `bl mcp connect` / `disconnect`                                                                                | Codex, Claude Code, Cursor, Qoder, Qoder Work, QwenWork, Qwen Code, Gemini, OpenCode, OpenClaw, DeepSeek Harness, ZCode, WorkBuddy, Pi |
-| Bailian pipeline workflow (a step in a bl flow)     | `bl pipeline run` / `validate`                                                                                 | JSON/YAML workflow definitions                                                                                                         |
-| Bailian rate limits / quota                         | `bl quota list` / `check` / `request`                                                                          | Console auth; class 2 — ask which product first if unnamed                                                                             |
-| Bailian free tier / usage stats                     | `bl usage free` / `stats` / `freetier`                                                                         | Console auth; class 2 — ask which product first if unnamed                                                                             |
-| Bailian Token Plan quota usage                      | `bl usage token-plan`                                                                                          | Console auth; class 2 — ask which product first if unnamed                                                                             |
-| Token Plan harness entitlement quota                | `bl token-plan harness-quota`                                                                                  | Console auth; issued quota usage + still-issuing harnesses                                                                             |
-| Bailian Coding Plan quota usage                     | `bl usage coding-plan`                                                                                         | Console auth; class 2 — ask which product first if unnamed                                                                             |
-| Model monitoring (calls, latency, errors, delivery) | `bl monitor overview` / `models` / `metrics` / `errors` / `delivery *`                                         | Console auth; see [`reference/monitor.md`](reference/monitor.md)                                                                       |
-| Model call logs / traces                            | `bl log audit *` / `bl log inference *` / `bl log status` / `bl log trace *`                                   | Console auth; audit and inference logs have separate APIs and switches; see [`reference/log.md`](reference/log.md)                     |
-| Model alert rules / templates / history             | `bl alert list` / `create` / `update` / `delete` / `enable` / `disable` / `history` / `metrics` / `template *` | Console auth; contacts are managed in the CloudMonitor console; see [`reference/alert.md`](reference/alert.md)                         |
-| Console API (advanced)                              | `bl console call`                                                                                              | Console auth                                                                                                                           |
-| Bailian workspace listing                           | `bl workspace list`                                                                                            | Console auth                                                                                                                           |
-| Switch CLI Help / Quick Start language              | `bl config set --key language --value zh-CN`                                                                   | Use `en-US` to switch back; follows the active config profile                                                                          |
-| Image / video / speech / omni / vision              | → skill `bailian-gen`                                                                                          | Fallback: `bl image\|video\|speech\|omni\|vision --help`                                                                               |
-| Dataset / fine-tune / deploy                        | → skill `bailian-finetune`                                                                                     | Fallback: `bl dataset\|finetune\|deploy --help`                                                                                        |
-| agents.yaml IaC / managed-agent sessions            | → skill `bailian-managed-agent`                                                                                | Fallback: `bl managed-agent --help`; `apply`/`destroy` need `--yes` after `plan`                                                       |
-| Bailian Sandbox instance / template lifecycle       | → skill `bailian-sandbox`                                                                                      | Fallback: `bl sandbox --help`                                                                                                          |
-| Web search (model-aware routing)                    | → skill `bailian-web-search`                                                                                   | Token Plan vs MCP path + fallback; fallback: `bl search web --help`                                                                    |
+| User intent                                         | Command                                                                                                        | Notes                                                                                                                              |
+| --------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- |
+| Explicit Bailian model chat / text execution        | `bl text chat`                                                                                                 | Default `qwen3.8-max`                                                                                                              |
+| Typed decision model inference                      | `bl decision`                                                                                                  | Workspace endpoint; returns `choice` / `noul` / `score` values and probabilities, not generated text                               |
+| Bailian agent / workflow                            | `bl app call`                                                                                                  | Needs `--app-id`                                                                                                                   |
+| Find app by name                                    | `bl app list` then `bl app call`                                                                               | Console auth                                                                                                                       |
+| Bailian app memory CRUD (not host-agent memory)     | `bl memory *`                                                                                                  | [`reference/memory.md`](reference/memory.md)                                                                                       |
+| Bailian knowledge base RAG                          | `bl knowledge search` / `chat`                                                                                 | API key + agent/workspace IDs                                                                                                      |
+| Upload a file as a step of a Bailian workflow       | `bl file upload`                                                                                               | When you need `oss://` URL explicitly; not for generic hosting                                                                     |
+| Bailian model selection / recommendation            | `bl advisor recommend`                                                                                         | Intent → candidate recall → LLM ranking                                                                                            |
+| Bailian model catalog / pricing / params            | `bl model list`                                                                                                | No auth (public catalog); `--model <family>` for detail, `--enrich` for input params                                               |
+| Find a model by keyword / modality                  | `bl model search`                                                                                              | Weighted ranking; `--input-modality` / `--output-modality`; offline models hidden by default                                       |
+| SDK sample code for calling a model                 | `bl model code`                                                                                                | `--sdk openai\|dashscope`, `--lang`; text output is the snippet alone, safe to redirect                                            |
+| Install / list / update / remove registry skills    | `bl skill add` / `list` / `update` / `remove`                                                                  | Bailian skill registry; see [`reference/skill.md`](reference/skill.md)                                                             |
+| Bailian MCP marketplace discovery / call            | `bl mcp list` / `tools` / `call`                                                                               | —                                                                                                                                  |
+| Register or remove a Bailian MCP in an Agent        | `bl mcp connect` / `disconnect`                                                                                | Codex, Claude Code, Cursor, Qoder, Qoder Work, QwenWork, Qwen Code, Gemini, OpenCode, OpenClaw, DeepSeek Harness, ZCode, WorkBuddy |
+| Bailian pipeline workflow (a step in a bl flow)     | `bl pipeline run` / `validate`                                                                                 | JSON/YAML workflow definitions                                                                                                     |
+| Bailian rate limits / quota                         | `bl quota list` / `check` / `request`                                                                          | Console auth; class 2 — ask which product first if unnamed                                                                         |
+| Bailian free tier / usage stats                     | `bl usage free` / `stats` / `freetier`                                                                         | Console auth; class 2 — ask which product first if unnamed                                                                         |
+| Bailian Token Plan quota usage                      | `bl usage token-plan`                                                                                          | Console auth; class 2 — ask which product first if unnamed                                                                         |
+| Token Plan harness entitlement quota                | `bl token-plan harness-quota`                                                                                  | Console auth; issued quota usage + still-issuing harnesses                                                                         |
+| Bailian Coding Plan quota usage                     | `bl usage coding-plan`                                                                                         | Console auth; class 2 — ask which product first if unnamed                                                                         |
+| Model monitoring (calls, latency, errors, delivery) | `bl monitor overview` / `models` / `metrics` / `errors` / `delivery *`                                         | Console auth; see [`reference/monitor.md`](reference/monitor.md)                                                                   |
+| Model call logs / traces                            | `bl log audit *` / `bl log inference *` / `bl log status` / `bl log trace *`                                   | Console auth; audit and inference logs have separate APIs and switches; see [`reference/log.md`](reference/log.md)                 |
+| Model alert rules / templates / history             | `bl alert list` / `create` / `update` / `delete` / `enable` / `disable` / `history` / `metrics` / `template *` | Console auth; contacts are managed in the CloudMonitor console; see [`reference/alert.md`](reference/alert.md)                     |
+| Console API (advanced)                              | `bl console call`                                                                                              | Console auth                                                                                                                       |
+| Bailian workspace listing                           | `bl workspace list`                                                                                            | Console auth                                                                                                                       |
+| Switch CLI Help / Quick Start language              | `bl config set --key language --value zh-CN`                                                                   | Use `en-US` to switch back; follows the active config profile                                                                      |
+| Image / video / speech / omni / vision              | → skill `bailian-gen`                                                                                          | Fallback: `bl image\|video\|speech\|omni\|vision --help`                                                                           |
+| Dataset / fine-tune / deploy                        | → skill `bailian-finetune`                                                                                     | Fallback: `bl dataset\|finetune\|deploy --help`                                                                                    |
+| agents.yaml IaC / managed-agent sessions            | → skill `bailian-managed-agent`                                                                                | Fallback: `bl managed-agent --help`; `apply`/`destroy` need `--yes` after `plan`                                                   |
+| Bailian Sandbox instance / template lifecycle       | → skill `bailian-sandbox`                                                                                      | Fallback: `bl sandbox --help`                                                                                                      |
+| Web search (model-aware routing)                    | → skill `bailian-web-search`                                                                                   | Token Plan vs MCP path + fallback; fallback: `bl search web --help`                                                                |
 
 Flags, usage, and examples: see hub [`reference/`](reference/index.md) or `bl <command> --help` — do not guess flags. Domain command details live in the owning skill's `reference/`.
 
@@ -99,6 +100,7 @@ Flags, usage, and examples: see hub [`reference/`](reference/index.md) or `bl <c
 
 ```bash
 bl text chat --message "Write a poem about spring in Chinese"
+bl decision --workspace-id <id> --state "Is this requesting a refund?" --questions '{"refund":{"type":"noul"}}'
 bl app list --output json
 bl app call --app-id <code> --prompt "Hello"
 bl usage stats
diff --git a/skills/bailian-cli/reference/decision.md b/skills/bailian-cli/reference/decision.md
new file mode 100644
index 000000000..f6bcdd247
--- /dev/null
+++ b/skills/bailian-cli/reference/decision.md
@@ -0,0 +1,67 @@
+# `bl decision` commands
+
+> Auto-generated from `packages/cli/src/commands.ts`. Do not edit by hand.
+> Regenerate: `pnpm --filter bailian-cli run generate:reference`.
+
+Index: [index.md](index.md)
+
+## Commands in this group
+
+| Command       | Authentication | Description                                                                    |
+| ------------- | -------------- | ------------------------------------------------------------------------------ |
+| `bl decision` | API Key        | Run the decision model: typed choice / noul / score answers with probabilities |
+
+## Command details
+
+### `bl decision`
+
+| Field              | Value                                                                                                                          |
+| ------------------ | ------------------------------------------------------------------------------------------------------------------------------ |
+| **Name**           | `decision`                                                                                                                     |
+| **Description**    | Run the decision model: typed choice / noul / score answers with probabilities                                                 |
+| **Authentication** | API Key                                                                                                                        |
+| **Usage**          | `bl decision (--state <text> \| --state-file <path>) (--questions <json> \| --questions-file <path>) \| --request-file <path>` |
+
+#### Flags
+
+| Flag                                    | Type   | Required | Description                                                                              |
+| --------------------------------------- | ------ | -------- | ---------------------------------------------------------------------------------------- |
+| `--model <model>`                       | string | no       | Decision model ID (default: decision-model-preview)                                      |
+| `--state <text>`                        | string | no       | Business state to decide on, sent as a string                                            |
+| `--state-file <path>`                   | string | no       | Read state from a file (.json is parsed as an object/array, others as text; - for stdin) |
+| `--questions <json>`                    | string | no       | Question table JSON, keyed by question ID                                                |
+| `--questions-file <path>`               | string | no       | Read the question table from a JSON file ({...} or {"questions":{...}})                  |
+| `--request-file <path>`                 | string | no       | Full request JSON {state, questions}; label/target are stripped                          |
+| `--workspace-id <id>`                   | string | no       | Workspace ID for the endpoint host (or set BAILIAN_WORKSPACE_ID)                         |
+| `--region <cn-beijing\|ap-southeast-1>` | string | no       | Endpoint region: cn-beijing, ap-southeast-1 (default: cn-beijing)                        |
+| `--api-key <key>`                       | string | no       | API key                                                                                  |
+| `--base-url <url>`                      | string | no       | API base URL                                                                             |
+
+#### Notes
+
+- Endpoint: configured --base-url, DASHSCOPE_BASE_URL, or profile base_url takes precedence; otherwise https://{workspace}.{region}.maas.aliyuncs.com/compatible-mode/v1/systemone, with workspace from --workspace-id, BAILIAN_WORKSPACE_ID, then config workspace_id.
+- Question types: choice (criteria: option → description, 1-255), noul (optional criteria with true/false descriptions), score (criteria: levels from low to high, 2-255; 3-7 recommended). Keep questions per request at 16 or fewer; latency grows roughly linearly.
+- Output defaults to the full JSON response. Use --output text for one human-readable line per question; the request ID and usage summary are written to stderr.
+- Training samples can be reused directly: label/target are stripped before sending. Sampling parameters (temperature, max_tokens, ...) are not supported by this API.
+
+#### Examples
+
+```bash
+bl decision --workspace-id <id> --state "Is this review asking for a refund?" --questions '{"refund":{"type":"noul","instructions":"Asking for a refund?"}}'
+```
+
+```bash
+bl decision --state-file ticket.json --questions-file workload.json
+```
+
+```bash
+bl decision --request-file sample.json --model decision-model-preview
+```
+
+```bash
+bl decision --region ap-southeast-1 --request-file sample.json
+```
+
+```bash
+bl decision --request-file sample.json --dry-run
+```
diff --git a/skills/bailian-cli/reference/index.md b/skills/bailian-cli/reference/index.md
index b3f9c402a..7be2022dc 100644
--- a/skills/bailian-cli/reference/index.md
+++ b/skills/bailian-cli/reference/index.md
@@ -39,6 +39,7 @@ Use this index for the skill-scoped quick index and global flags.
 | `bl config ui`                   | No Auth        | Open a local web UI to manage config profiles                                                                      | [config.md](config.md)         |
 | `bl config use`                  | No Auth        | Set the active config profile                                                                                      | [config.md](config.md)         |
 | `bl console call`                | Console        | Call a Bailian console API via the CLI gateway                                                                     | [console.md](console.md)       |
+| `bl decision`                    | API Key        | Run the decision model: typed choice / noul / score answers with probabilities                                     | [decision.md](decision.md)     |
 | `bl file upload`                 | API Key        | Upload a local file to DashScope temporary storage (48h)                                                           | [file.md](file.md)             |
 | `bl knowledge category add`      | API Key        | Create a data-center category                                                                                      | [knowledge.md](knowledge.md)   |
 | `bl knowledge category delete`   | API Key        | Delete a data-center category                                                                                      | [knowledge.md](knowledge.md)   |
@@ -157,6 +158,7 @@ Use this index for the skill-scoped quick index and global flags.
 | `auth`       | `generate-access-token`, `login`, `logout`, `status`                                                                                                                                                                                                                                                                                                                                                                                                                                        | [auth.md](auth.md)             |
 | `config`     | `agent`, `list`, `set`, `show`, `ui`, `use`                                                                                                                                                                                                                                                                                                                                                                                                                                                 | [config.md](config.md)         |
 | `console`    | `call`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      | [console.md](console.md)       |
+| `decision`   | `(root)`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | [decision.md](decision.md)     |
 | `file`       | `upload`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    | [file.md](file.md)             |
 | `knowledge`  | `category add`, `category delete`, `category list`, `chat`, `chunk add`, `chunk delete`, `chunk list`, `chunk update`, `collection create`, `collection get`, `create`, `delete`, `doc delete`, `doc import-oss`, `doc list`, `doc status`, `doc tag`, `doc upload`, `file delete`, `file get`, `file list`, `info`, `list`, `retrieve`, `search`, `service copy`, `service create`, `service delete`, `service deploy`, `service get`, `service list`, `service update`, `stats`, `update` | [knowledge.md](knowledge.md)   |
 | `log`        | `audit count`, `audit disable`, `audit enable`, `audit get`, `audit list`, `inference count`, `inference disable`, `inference enable`, `inference get`, `inference list`, `status`, `trace get`, `trace list`, `trace stats`                                                                                                                                                                                                                                                                | [log.md](log.md)               |
diff --git a/skills/bailian-finetune/SKILL.md b/skills/bailian-finetune/SKILL.md
index d2ce35a1f..8820d364b 100644
--- a/skills/bailian-finetune/SKILL.md
+++ b/skills/bailian-finetune/SKILL.md
@@ -8,7 +8,7 @@ description: >-
   阿里云百炼模型精调训练入口：用户要精调、微调、训练自己的模型（fine-tune，支持 SFT / SFT-LoRA / DPO / DPO-LoRA / CPT，
   覆盖文本、语音、图像）、校验或上传训练数据集、看训练进度和日志、挑 checkpoint、导出精调产物、
   把专属模型部署成服务时使用 `bl dataset` / `bl finetune` / `bl deploy`。链路是 validate 校验数据 →
-  upload 拿 file-id → finetune create 建任务 → watch 看进度 → export 导出 → deploy 上线，需要 API key；
+  upload 拿 file-id → finetune create 建任务 → watch 看进度 → export 导出 → 可选 compress 量化压缩 → deploy 上线，需要 API key；
   写操作先用 `--dry-run` 预览。反触发：用户点名火山方舟/ark 的精调不走本 skill；只是要选哪个模型走
   bailian-model-recommend；用现成模型生图生视频走 bailian-gen；百炼其他资源管理走 bailian-cli。
   官方安装：`bl skill init`（与共享协议 bailian-protocol 同装）。
@@ -27,7 +27,10 @@ description: >-
 4. Watch progress  bl finetune watch --job-id ft-xxx             # or get / logs
 5. Pick artifact   bl finetune checkpoints --job-id ft-xxx
 6. Export model    bl finetune export --job-id ft-xxx --checkpoint ckpt-N --model-name my-model
-7. Deploy service  bl deploy text|audio|image create --model-name my-model --display-name my-svc
+7. Compress model  bl finetune compress templates --model my-model       # optional; full-parameter models only
+                   bl finetune compress create --model my-model --template-id <template-id>
+                   bl finetune compress watch --job-id quant-xxx --follow
+8. Deploy service  bl deploy text|audio|image create --model-name <my-model|quantized-output> --display-name my-svc
 ```
 
 - Unsure which training methods a base model supports → `bl finetune capability --base-model <base>` or `--training-type sft|sft-lora|dpo|cpt`.
@@ -44,7 +47,9 @@ description: >-
 | Create a fine-tuning job               | `bl finetune text\|audio\|image create`                                                          |
 | Job list / detail / follow             | `bl finetune list` / `get` / `watch` / `logs`                                                    |
 | Artifacts and export                   | `bl finetune checkpoints` / `export`                                                             |
-| Cancel / delete a job                  | `bl finetune cancel` / `delete`                                                                  |
+| Quantize / compress a full model       | `bl finetune compress templates` / `create` / `watch` / `get` / `logs` / `list`                  |
+| Cancel / delete a compression job      | `bl finetune compress cancel` / `delete`                                                         |
+| Cancel / delete a fine-tuning job      | `bl finetune cancel` / `delete`                                                                  |
 | Trainable capability lookup            | `bl finetune capability`                                                                         |
 | Deploy / lifecycle                     | `bl deploy text\|audio\|image create`, `list` / `get` / `update` / `scale` / `delete` / `models` |
 | Query throughput reservations          | `bl deploy list --plan ptu` / `bl deploy get`                                                    |
@@ -68,7 +73,11 @@ bl dataset upload --file train.jsonl
 bl finetune text create --base-model qwen3-8b --training-type sft-lora --datasets file-xxx
 bl finetune watch --job-id ft-xxx
 bl finetune export --job-id ft-xxx --checkpoint ckpt-3 --model-name my-qwen-sft
-bl deploy text create --model-name my-qwen-sft --display-name my-svc
+# Optional quantization (Beijing region; full-parameter fine-tuned models only)
+bl finetune compress templates --model my-qwen-sft
+bl finetune compress create --model my-qwen-sft --template-id quant-xxx --suffix int4
+bl finetune compress watch --job-id quant-xxx --follow
+bl deploy text create --model-name <my-qwen-sft-or-quantized-output> --display-name my-svc
 ```
 
 ## Common hand-offs
diff --git a/skills/bailian-finetune/reference/dataset.md b/skills/bailian-finetune/reference/dataset.md
index f9dfccc4c..672708a4f 100644
--- a/skills/bailian-finetune/reference/dataset.md
+++ b/skills/bailian-finetune/reference/dataset.md
@@ -120,30 +120,30 @@ bl dataset list --output json
 
 ### `bl dataset upload`
 
-| Field              | Value                                                                                                                                   |
-| ------------------ | --------------------------------------------------------------------------------------------------------------------------------------- |
-| **Name**           | `dataset upload`                                                                                                                        |
-| **Description**    | Upload a dataset file (.jsonl or .zip) to Bailian                                                                                       |
-| **Authentication** | API Key                                                                                                                                 |
-| **Usage**          | `bl dataset upload --file <path> [--purpose <name>] [--schema <chatml\|dpo\|cpt\|tts\|image\|video>] [--no-validate] [--full-validate]` |
+| Field              | Value                                                                                                                                             |
+| ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Name**           | `dataset upload`                                                                                                                                  |
+| **Description**    | Upload a dataset file (.jsonl or .zip) to Bailian                                                                                                 |
+| **Authentication** | API Key                                                                                                                                           |
+| **Usage**          | `bl dataset upload --file <path> [--purpose <name>] [--schema <chatml\|dpo\|cpt\|decision\|tts\|image\|video>] [--no-validate] [--full-validate]` |
 
 #### Flags
 
-| Flag               | Type   | Required | Description                                                                                                                                                                          |
-| ------------------ | ------ | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| `--file <path>`    | string | yes      | Local dataset file (.jsonl or .zip; ≤200MB SFT/DPO, ≤300MB CPT, ≤2GB media zip)                                                                                                      |
-| `--purpose <name>` | string | no       | Dataset purpose tag (default: "fine-tune"; e.g. "evaluation")                                                                                                                        |
-| `--schema <s>`     | string | no       | Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record. |
-| `--no-validate`    | switch | no       | Skip the local JSONL pre-flight check (not recommended)                                                                                                                              |
-| `--full-validate`  | switch | no       | JSON.parse every line instead of sampling (slower)                                                                                                                                   |
-| `--api-key <key>`  | string | no       | API key                                                                                                                                                                              |
-| `--base-url <url>` | string | no       | API base URL                                                                                                                                                                         |
+| Flag               | Type   | Required | Description                                                                                                                                                                                                        |
+| ------------------ | ------ | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| `--file <path>`    | string | yes      | Local dataset file (.jsonl or .zip; ≤200MB SFT/DPO, ≤300MB CPT/decision, ≤2GB media zip)                                                                                                                           |
+| `--purpose <name>` | string | no       | Dataset purpose tag (default: "fine-tune"; e.g. "evaluation")                                                                                                                                                      |
+| `--schema <s>`     | string | no       | Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "decision" (state/questions), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record. |
+| `--no-validate`    | switch | no       | Skip the local JSONL pre-flight check (not recommended)                                                                                                                                                            |
+| `--full-validate`  | switch | no       | JSON.parse every line instead of sampling (slower)                                                                                                                                                                 |
+| `--api-key <key>`  | string | no       | API key                                                                                                                                                                                                            |
+| `--base-url <url>` | string | no       | API base URL                                                                                                                                                                                                       |
 
 #### Notes
 
-- Supports .jsonl (text) and .zip (audio/image/video archives with a data.jsonl manifest). Six record schemas are recognized: chatml = {messages:[...]} (SFT); dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."} (continual pre-training, raw text); tts = {wav_fn:"train/xxx.wav", text:"..."} (audio fine-tuning); image = {img_path:"..."} (image generation); video = {first_frame_path:...} (video generation).
-- With no --schema, a record carrying wav_fn is validated as TTS, img_path as image, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML.
-- Upload cap: 200MB SFT/DPO text, 300MB CPT, 2GB media zip. Upload uses the OpenAI-compatible /compatible-mode/v1/files endpoint so the purpose tag is persisted (the DashScope-native /api/v1/files drops it).
+- Supports .jsonl (text/decision) and .zip (audio/image/video archives with a data.jsonl manifest). Seven record schemas are recognized: chatml = {messages:[...]} (SFT); dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."}; decision = {state:{content}, questions:{...}}; tts = {wav_fn:"train/xxx.wav", text:"..."}; image = {img_path:"..."}; video = {first_frame_path:...}.
+- With no --schema, state/questions is validated as decision data, wav_fn as TTS, img_path as image, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML.
+- Upload cap: 200MB SFT/DPO text, 300MB CPT/decision data, 2GB media zip. For decision files over 200MB, pass --schema decision so the correct cap is applied. Upload uses the OpenAI-compatible /compatible-mode/v1/files endpoint so the purpose tag is persisted.
 
 #### Examples
 
@@ -159,6 +159,10 @@ bl dataset upload --file dpo.jsonl --schema dpo
 bl dataset upload --file cpt.jsonl --schema cpt
 ```
 
+```bash
+bl dataset upload --file decision.jsonl --schema decision
+```
+
 ```bash
 bl dataset upload --file audio.zip --schema tts
 ```
@@ -177,26 +181,26 @@ bl dataset upload --file train.jsonl --no-validate
 
 ### `bl dataset validate`
 
-| Field              | Value                                                                                                  |
-| ------------------ | ------------------------------------------------------------------------------------------------------ |
-| **Name**           | `dataset validate`                                                                                     |
-| **Description**    | Locally validate a dataset file (.jsonl or .zip) without uploading                                     |
-| **Authentication** | No Auth                                                                                                |
-| **Usage**          | `bl dataset validate --file <path> [--full-validate] [--schema <chatml\|dpo\|cpt\|tts\|image\|video>]` |
+| Field              | Value                                                                                                            |
+| ------------------ | ---------------------------------------------------------------------------------------------------------------- |
+| **Name**           | `dataset validate`                                                                                               |
+| **Description**    | Locally validate a dataset file (.jsonl or .zip) without uploading                                               |
+| **Authentication** | No Auth                                                                                                          |
+| **Usage**          | `bl dataset validate --file <path> [--full-validate] [--schema <chatml\|dpo\|cpt\|decision\|tts\|image\|video>]` |
 
 #### Flags
 
-| Flag              | Type   | Required | Description                                                                                                                                                                          |
-| ----------------- | ------ | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| `--file <path>`   | string | yes      | Local dataset file (.jsonl or .zip)                                                                                                                                                  |
-| `--full-validate` | switch | no       | JSON.parse every line instead of sampling (slower)                                                                                                                                   |
-| `--schema <s>`    | string | no       | Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record. |
+| Flag              | Type   | Required | Description                                                                                                                                                                                                        |
+| ----------------- | ------ | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| `--file <path>`   | string | yes      | Local dataset file (.jsonl or .zip)                                                                                                                                                                                |
+| `--full-validate` | switch | no       | JSON.parse every line instead of sampling (slower)                                                                                                                                                                 |
+| `--schema <s>`    | string | no       | Record schema: "chatml" (SFT), "dpo" (chosen/rejected), "cpt" (raw text), "decision" (state/questions), "tts" (audio), "image" (image generation), or "video" (video generation). Default auto-detects per record. |
 
 #### Notes
 
 - Default scan: every line gets a structural check, then ~160 lines (front 50, evenly spaced 100, last 10) are JSON.parsed against the active schema.
-- Schemas: chatml = {messages:[...]} (SFT); dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."} (continual pre-training, raw text); tts = {wav_fn:"train/xxx.wav", text:"..."} (audio fine-tuning); image = {img_path:"..."} (image generation); video = {first_frame_path:"...", video_path:"..."} (video generation, i2v first-frame or kf2v first+last-frame with last_frame_path).
-- With no --schema, a record carrying wav_fn is validated as TTS, img_path as image, first_frame_path/video_path as video, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML. Pass --schema to require a specific shape on every record.
+- Schemas: chatml = {messages:[...]}; dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."}; decision = {state:{content}, questions:{...}}; tts = {wav_fn:"train/xxx.wav", text:"..."}; image = {img_path:"..."}; video = {first_frame_path:"...", video_path:"..."}.
+- With no --schema, state/questions is validated as decision data, wav_fn as TTS, img_path as image, first_frame_path/video_path as video, chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML.
 - ZIP archives (.zip) are validated structurally (data.jsonl present, media references resolve) in addition to per-record content checks. Use --full-validate to JSON.parse every line.
 
 #### Examples
@@ -213,6 +217,10 @@ bl dataset validate --file dpo.jsonl --schema dpo
 bl dataset validate --file cpt.jsonl --schema cpt
 ```
 
+```bash
+bl dataset validate --file decision.jsonl --schema decision
+```
+
 ```bash
 bl dataset validate --file audio.zip --schema tts
 ```
diff --git a/skills/bailian-finetune/reference/finetune.md b/skills/bailian-finetune/reference/finetune.md
index 236260f63..e14af2242 100644
--- a/skills/bailian-finetune/reference/finetune.md
+++ b/skills/bailian-finetune/reference/finetune.md
@@ -7,22 +7,30 @@ Index: [index.md](index.md)
 
 ## Commands in this group
 
-| Command                    | Authentication | Description                                                                                                                     |
-| -------------------------- | -------------- | ------------------------------------------------------------------------------------------------------------------------------- |
-| `bl finetune audio create` | API Key        | Create an audio TTS model fine-tune job (sft-lora)                                                                              |
-| `bl finetune cancel`       | API Key        | Cancel a running fine-tune job                                                                                                  |
-| `bl finetune capability`   | No Auth        | Query fine-tune training capability — by model (which training types it supports) or by training type (which models support it) |
-| `bl finetune checkpoints`  | API Key        | List checkpoints produced by a fine-tune job                                                                                    |
-| `bl finetune delete`       | API Key        | Delete a fine-tune job record                                                                                                   |
-| `bl finetune export`       | API Key        | Publish a checkpoint as a deployable model                                                                                      |
-| `bl finetune get`          | API Key        | Get details of a single fine-tune job                                                                                           |
-| `bl finetune image create` | API Key        | Create an image generation model fine-tune job (sft-lora)                                                                       |
-| `bl finetune list`         | API Key        | List fine-tune jobs                                                                                                             |
-| `bl finetune logs`         | API Key        | Fetch training logs for a fine-tune job                                                                                         |
-| `bl finetune price`        | Console        | Estimate the training cost for a fine-tune job (token billing)                                                                  |
-| `bl finetune text create`  | API Key        | Create a text model fine-tune job (sft \| sft-lora \| dpo \| dpo-lora \| cpt)                                                   |
-| `bl finetune video create` | API Key        | Create a video generation model fine-tune job (Wan i2v/kf2v, efficient_sft)                                                     |
-| `bl finetune watch`        | API Key        | Probe a fine-tune job's status (default: single non-blocking fetch). Pass --follow to poll until terminal.                      |
+| Command                          | Authentication | Description                                                                                                                     |
+| -------------------------------- | -------------- | ------------------------------------------------------------------------------------------------------------------------------- |
+| `bl finetune audio create`       | API Key        | Create an audio TTS model fine-tune job (sft-lora)                                                                              |
+| `bl finetune cancel`             | API Key        | Cancel a running fine-tune job                                                                                                  |
+| `bl finetune capability`         | No Auth        | Query fine-tune training capability — by model (which training types it supports) or by training type (which models support it) |
+| `bl finetune checkpoints`        | API Key        | List checkpoints produced by a fine-tune job                                                                                    |
+| `bl finetune compress cancel`    | API Key        | Cancel a compression job                                                                                                        |
+| `bl finetune compress create`    | API Key        | Create a model compression job                                                                                                  |
+| `bl finetune compress delete`    | API Key        | Delete a compression job record                                                                                                 |
+| `bl finetune compress get`       | API Key        | Get a compression job                                                                                                           |
+| `bl finetune compress list`      | API Key        | List model compression jobs                                                                                                     |
+| `bl finetune compress logs`      | API Key        | Get compression job logs                                                                                                        |
+| `bl finetune compress templates` | API Key        | List quantizable models and their compression templates                                                                         |
+| `bl finetune compress watch`     | API Key        | Probe or follow a compression job's status                                                                                      |
+| `bl finetune delete`             | API Key        | Delete a fine-tune job record                                                                                                   |
+| `bl finetune export`             | API Key        | Publish a checkpoint as a deployable model                                                                                      |
+| `bl finetune get`                | API Key        | Get details of a single fine-tune job                                                                                           |
+| `bl finetune image create`       | API Key        | Create an image generation model fine-tune job (sft-lora)                                                                       |
+| `bl finetune list`               | API Key        | List fine-tune jobs                                                                                                             |
+| `bl finetune logs`               | API Key        | Fetch training logs for a fine-tune job                                                                                         |
+| `bl finetune price`              | Console        | Estimate the training cost for a fine-tune job (token billing)                                                                  |
+| `bl finetune text create`        | API Key        | Create a text model fine-tune job (sft \| sft-lora \| dpo \| dpo-lora \| cpt)                                                   |
+| `bl finetune video create`       | API Key        | Create a video generation model fine-tune job (Wan i2v/kf2v, efficient_sft)                                                     |
+| `bl finetune watch`              | API Key        | Probe a fine-tune job's status (default: single non-blocking fetch). Pass --follow to poll until terminal.                      |
 
 ## Command details
 
@@ -179,6 +187,292 @@ bl finetune checkpoints --job-id ft-xxx
 bl finetune checkpoints --job-id ft-xxx --output json
 ```
 
+### `bl finetune compress cancel`
+
+| Field              | Value                                       |
+| ------------------ | ------------------------------------------- |
+| **Name**           | `finetune compress cancel`                  |
+| **Description**    | Cancel a compression job                    |
+| **Authentication** | API Key                                     |
+| **Usage**          | `bl finetune compress cancel --job-id <id>` |
+
+#### Flags
+
+| Flag               | Type   | Required | Description                   |
+| ------------------ | ------ | -------- | ----------------------------- |
+| `--job-id <id>`    | string | yes      | Compression job ID (required) |
+| `--api-key <key>`  | string | no       | API key                       |
+| `--base-url <url>` | string | no       | API base URL                  |
+
+#### Notes
+
+- Only PENDING, QUEUING, or RUNNING jobs can be cancelled.
+
+#### Examples
+
+```bash
+bl finetune compress cancel --job-id quant-xxx
+```
+
+```bash
+bl finetune compress cancel --job-id quant-xxx --dry-run
+```
+
+### `bl finetune compress create`
+
+| Field              | Value                                                                    |
+| ------------------ | ------------------------------------------------------------------------ |
+| **Name**           | `finetune compress create`                                               |
+| **Description**    | Create a model compression job                                           |
+| **Authentication** | API Key                                                                  |
+| **Usage**          | `bl finetune compress create --model <model> --template-id <id> [flags]` |
+
+#### Flags
+
+| Flag                                | Type   | Required | Description                                           |
+| ----------------------------------- | ------ | -------- | ----------------------------------------------------- |
+| `--model <model>`                   | string | yes      | Source custom model ID (required)                     |
+| `--template-id <id>`                | string | yes      | Compression template ID (required)                    |
+| `--job-name <name>`                 | string | no       | Unique job name (max 50 characters)                   |
+| `--description <text>`              | string | no       | Job description (max 200 characters)                  |
+| `--hparam <key=value>`              | array  | no       | Hyper-parameter override; repeat for multiple values  |
+| `--calibration-files <file-id,...>` | string | no       | Comma-separated calibration dataset file IDs          |
+| `--suffix <suffix>`                 | string | no       | Output model suffix (1–8 lowercase letters or digits) |
+| `--api-key <key>`                   | string | no       | API key                                               |
+| `--base-url <url>`                  | string | no       | API base URL                                          |
+
+#### Notes
+
+- Run `finetune compress templates` first to get a compatible template ID.
+- Model compression APIs are currently available only in the Beijing region.
+
+#### Examples
+
+```bash
+bl finetune compress create --model qwen3.5-flash-2026-02-23-ft-xxx --template-id quant-flash-nvfp4-mlp-nomtp
+```
+
+```bash
+bl finetune compress create --model qwen3.5-flash-2026-02-23-ft-xxx --template-id quant-flash-nvfp4-mlp-nomtp --suffix test --hparam calib_input=true
+```
+
+### `bl finetune compress delete`
+
+| Field              | Value                                                                     |
+| ------------------ | ------------------------------------------------------------------------- |
+| **Name**           | `finetune compress delete`                                                |
+| **Description**    | Delete a compression job record                                           |
+| **Authentication** | API Key                                                                   |
+| **Usage**          | `bl finetune compress delete --job-id <id>`                               |
+| **Risk**           | `high`                                                                    |
+| **Risk message**   | This permanently deletes the compression job record and cannot be undone. |
+
+> **Agent safety:** Never add `--yes` automatically. On `type="requires_confirmation"`, stop and ask for explicit user confirmation of the same action and scope.
+
+#### Flags
+
+| Flag               | Type   | Required | Description                      |
+| ------------------ | ------ | -------- | -------------------------------- |
+| `--job-id <id>`    | string | yes      | Compression job ID (required)    |
+| `--yes`            | switch | no       | Confirm this high-risk operation |
+| `--api-key <key>`  | string | no       | API key                          |
+| `--base-url <url>` | string | no       | API base URL                     |
+
+#### Notes
+
+- Only terminal jobs can be deleted. Deleting a job record does not delete its quantized output model.
+
+#### Examples
+
+```bash
+bl finetune compress delete --job-id quant-xxx
+```
+
+```bash
+bl finetune compress delete --job-id quant-xxx --dry-run
+```
+
+```bash
+# Only after explicit user confirmation:
+bl finetune compress delete --job-id quant-xxx --yes
+```
+
+### `bl finetune compress get`
+
+| Field              | Value                                    |
+| ------------------ | ---------------------------------------- |
+| **Name**           | `finetune compress get`                  |
+| **Description**    | Get a compression job                    |
+| **Authentication** | API Key                                  |
+| **Usage**          | `bl finetune compress get --job-id <id>` |
+
+#### Flags
+
+| Flag               | Type   | Required | Description                   |
+| ------------------ | ------ | -------- | ----------------------------- |
+| `--job-id <id>`    | string | yes      | Compression job ID (required) |
+| `--api-key <key>`  | string | no       | API key                       |
+| `--base-url <url>` | string | no       | API base URL                  |
+
+#### Examples
+
+```bash
+bl finetune compress get --job-id quant-xxx
+```
+
+### `bl finetune compress list`
+
+| Field              | Value                                                                |
+| ------------------ | -------------------------------------------------------------------- |
+| **Name**           | `finetune compress list`                                             |
+| **Description**    | List model compression jobs                                          |
+| **Authentication** | API Key                                                              |
+| **Usage**          | `bl finetune compress list [filters] [--page <n>] [--page-size <n>]` |
+
+#### Flags
+
+| Flag                                | Type   | Required | Description                                    |
+| ----------------------------------- | ------ | -------- | ---------------------------------------------- |
+| `--status <status>`                 | string | no       | Filter by job status                           |
+| `--model <model>`                   | string | no       | Filter by source model ID                      |
+| `--template-id <id>`                | string | no       | Filter by template ID                          |
+| `--quant-spec <spec>`               | string | no       | Filter by quantization specification           |
+| `--quant-method <method>`           | string | no       | Filter by quantization method                  |
+| `--start-time <time>`               | string | no       | Filter jobs starting no earlier than this time |
+| `--end-time <time>`                 | string | no       | Filter jobs ending no later than this time     |
+| `--search <keyword>`                | string | no       | Search job ID or name                          |
+| `--search-field <job_id\|job_name>` | string | no       | Limit --search to job_id or job_name           |
+| `--sort-order <asc\|desc>`          | string | no       | Sort order by create_time (default: desc)      |
+| `--page <n>`                        | number | no       | Page number (default: 1)                       |
+| `--page-size <n>`                   | number | no       | Results per page (default: 10, max: 100)       |
+| `--api-key <key>`                   | string | no       | API key                                        |
+| `--base-url <url>`                  | string | no       | API base URL                                   |
+
+#### Examples
+
+```bash
+bl finetune compress list
+```
+
+```bash
+bl finetune compress list --status SUCCEEDED --page-size 20
+```
+
+```bash
+bl finetune compress list --quant-method gptq
+```
+
+```bash
+bl finetune compress list --search qwen3 --search-field job_name --sort-order asc
+```
+
+### `bl finetune compress logs`
+
+| Field              | Value                                                                                |
+| ------------------ | ------------------------------------------------------------------------------------ |
+| **Name**           | `finetune compress logs`                                                             |
+| **Description**    | Get compression job logs                                                             |
+| **Authentication** | API Key                                                                              |
+| **Usage**          | `bl finetune compress logs --job-id <id> [--offset <n>] [--lines <n> \| --tail <n>]` |
+
+#### Flags
+
+| Flag               | Type   | Required | Description                                       |
+| ------------------ | ------ | -------- | ------------------------------------------------- |
+| `--job-id <id>`    | string | yes      | Compression job ID (required)                     |
+| `--offset <n>`     | number | no       | Skip the first N lines (default: 0)               |
+| `--lines <n>`      | number | no       | Number of lines to read (default: 100, max: 1000) |
+| `--tail <n>`       | number | no       | Read the last N log lines                         |
+| `--api-key <key>`  | string | no       | API key                                           |
+| `--base-url <url>` | string | no       | API base URL                                      |
+
+#### Examples
+
+```bash
+bl finetune compress logs --job-id quant-xxx
+```
+
+```bash
+bl finetune compress logs --job-id quant-xxx --offset 0 --lines 50
+```
+
+```bash
+bl finetune compress logs --job-id quant-xxx --tail 20
+```
+
+### `bl finetune compress templates`
+
+| Field              | Value                                                   |
+| ------------------ | ------------------------------------------------------- |
+| **Name**           | `finetune compress templates`                           |
+| **Description**    | List quantizable models and their compression templates |
+| **Authentication** | API Key                                                 |
+| **Usage**          | `bl finetune compress templates [--model <model>]`      |
+
+#### Flags
+
+| Flag               | Type   | Required | Description                                  |
+| ------------------ | ------ | -------- | -------------------------------------------- |
+| `--model <model>`  | string | no       | Filter by custom model ID or base model name |
+| `--api-key <key>`  | string | no       | API key                                      |
+| `--base-url <url>` | string | no       | API base URL                                 |
+
+#### Notes
+
+- Only supported full-parameter fine-tuned models are returned. LoRA and already-quantized models are excluded.
+- Template IDs live in output.custom_models[].templates[].template_id. output.base_models only lists quantizable base model names and carries no templates, so an empty custom_models array means no template ID is available yet.
+- An empty templates[].hyper_parameters array means that template exposes no tunable hyper-parameters; omit --hparam when creating the compression job.
+- template_name and description are localized server-side from the CLI language setting (default en-US), sent as the lang query parameter. Run `config set --key language --value zh-CN` to get Chinese text.
+
+#### Examples
+
+```bash
+bl finetune compress templates
+```
+
+```bash
+bl finetune compress templates --model qwen3.5-flash-2026-02-23-ft-xxx
+```
+
+### `bl finetune compress watch`
+
+| Field              | Value                                                                                           |
+| ------------------ | ----------------------------------------------------------------------------------------------- |
+| **Name**           | `finetune compress watch`                                                                       |
+| **Description**    | Probe or follow a compression job's status                                                      |
+| **Authentication** | API Key                                                                                         |
+| **Usage**          | `bl finetune compress watch --job-id <id> [--follow] [--interval <sec>] [--poll-timeout <sec>]` |
+
+#### Flags
+
+| Flag                   | Type   | Required | Description                                          |
+| ---------------------- | ------ | -------- | ---------------------------------------------------- |
+| `--job-id <id>`        | string | yes      | Compression job ID (required)                        |
+| `--follow`             | switch | no       | Poll until the job reaches a terminal state          |
+| `--interval <sec>`     | number | no       | Polling interval with --follow (default: 10, min: 1) |
+| `--poll-timeout <sec>` | number | no       | Stop polling after this many seconds                 |
+| `--api-key <key>`      | string | no       | API key                                              |
+| `--base-url <url>`     | string | no       | API base URL                                         |
+
+#### Notes
+
+- Without --follow, the command performs one non-blocking status probe.
+- On success, quantized_output can be passed to `deploy text create --model-name`.
+
+#### Examples
+
+```bash
+bl finetune compress watch --job-id quant-xxx
+```
+
+```bash
+bl finetune compress watch --job-id quant-xxx --follow
+```
+
+```bash
+bl finetune compress watch --job-id quant-xxx --follow --interval 30 --poll-timeout 7200
+```
+
 ### `bl finetune delete`
 
 | Field              | Value                                                                             |
diff --git a/skills/bailian-finetune/reference/index.md b/skills/bailian-finetune/reference/index.md
index 756194009..7ebcee860 100644
--- a/skills/bailian-finetune/reference/index.md
+++ b/skills/bailian-finetune/reference/index.md
@@ -41,6 +41,14 @@ Use this index for the skill-scoped quick index and global flags.
 | `bl finetune cancel`             | API Key        | Cancel a running fine-tune job                                                                                                  | [finetune.md](finetune.md) |
 | `bl finetune capability`         | No Auth        | Query fine-tune training capability — by model (which training types it supports) or by training type (which models support it) | [finetune.md](finetune.md) |
 | `bl finetune checkpoints`        | API Key        | List checkpoints produced by a fine-tune job                                                                                    | [finetune.md](finetune.md) |
+| `bl finetune compress cancel`    | API Key        | Cancel a compression job                                                                                                        | [finetune.md](finetune.md) |
+| `bl finetune compress create`    | API Key        | Create a model compression job                                                                                                  | [finetune.md](finetune.md) |
+| `bl finetune compress delete`    | API Key        | Delete a compression job record                                                                                                 | [finetune.md](finetune.md) |
+| `bl finetune compress get`       | API Key        | Get a compression job                                                                                                           | [finetune.md](finetune.md) |
+| `bl finetune compress list`      | API Key        | List model compression jobs                                                                                                     | [finetune.md](finetune.md) |
+| `bl finetune compress logs`      | API Key        | Get compression job logs                                                                                                        | [finetune.md](finetune.md) |
+| `bl finetune compress templates` | API Key        | List quantizable models and their compression templates                                                                         | [finetune.md](finetune.md) |
+| `bl finetune compress watch`     | API Key        | Probe or follow a compression job's status                                                                                      | [finetune.md](finetune.md) |
 | `bl finetune delete`             | API Key        | Delete a fine-tune job record                                                                                                   | [finetune.md](finetune.md) |
 | `bl finetune export`             | API Key        | Publish a checkpoint as a deployable model                                                                                      | [finetune.md](finetune.md) |
 | `bl finetune get`                | API Key        | Get details of a single fine-tune job                                                                                           | [finetune.md](finetune.md) |
@@ -54,11 +62,11 @@ Use this index for the skill-scoped quick index and global flags.
 
 ## By group
 
-| Group      | Commands                                                                                                                                                                                                                                                                                                 | Reference                  |
-| ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------- |
-| `dataset`  | `delete`, `get`, `list`, `upload`, `validate`                                                                                                                                                                                                                                                            | [dataset.md](dataset.md)   |
-| `deploy`   | `audio create`, `capacity create`, `capacity delete`, `capacity get`, `capacity list`, `capacity renew`, `capacity scale`, `capacity unsubscribe`, `delete`, `get`, `image create`, `list`, `models`, `operation get`, `operation wait`, `overflow`, `pause`, `resume`, `scale`, `text create`, `update` | [deploy.md](deploy.md)     |
-| `finetune` | `audio create`, `cancel`, `capability`, `checkpoints`, `delete`, `export`, `get`, `image create`, `list`, `logs`, `price`, `text create`, `video create`, `watch`                                                                                                                                        | [finetune.md](finetune.md) |
+| Group      | Commands                                                                                                                                                                                                                                                                                                             | Reference                  |
+| ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------- |
+| `dataset`  | `delete`, `get`, `list`, `upload`, `validate`                                                                                                                                                                                                                                                                        | [dataset.md](dataset.md)   |
+| `deploy`   | `audio create`, `capacity create`, `capacity delete`, `capacity get`, `capacity list`, `capacity renew`, `capacity scale`, `capacity unsubscribe`, `delete`, `get`, `image create`, `list`, `models`, `operation get`, `operation wait`, `overflow`, `pause`, `resume`, `scale`, `text create`, `update`             | [deploy.md](deploy.md)     |
+| `finetune` | `audio create`, `cancel`, `capability`, `checkpoints`, `compress cancel`, `compress create`, `compress delete`, `compress get`, `compress list`, `compress logs`, `compress templates`, `compress watch`, `delete`, `export`, `get`, `image create`, `list`, `logs`, `price`, `text create`, `video create`, `watch` | [finetune.md](finetune.md) |
 
 ## Global flags