@zhuoyuezs/ml-platform 0.1.16 → 0.1.18

package/DEVELOPMENT.md

@@ -9,7 +9,7 @@
 - Business CLI 与 Skill 是并列组件；CLI 不属于 Skill。
 - Skill 直接调用正式命令 `ml-platform`，不嵌入 CLI runtime。
 - 正式 npm 包只有 `@zhuoyuezs/ml-platform`，正式 bin 只有 `ml-platform`。
-- `business-client-js/` 是 CLI 唯一源码，`skills/feature-management/` 是 Skill 唯一源码。
+- `business-client-js/` 是 CLI 唯一源码，仓库 `skills/` 下三个受支持 Skill 是各自唯一源码。
 - `npx @zhuoyuezs/ml-platform@<version> install` 是面向用户的统一安装入口；默认目标是 Codex
   用户级 Skill，不要求额外安装参数。
 - 管理命令默认输出简洁结果；Agent 和 CI 显式使用 `--json` 获取稳定结构化输出。
@@ -27,6 +27,8 @@
 ```text
 business-client-js/
 skills/feature-management/
+skills/model-lifecycle-management/
+skills/model-deployment-management/
 npm/ml-platform/
 npm/scripts/build_ml_platform_release.py
 agent_poc/codex-host/
@@ -41,7 +43,7 @@ package/
 │   ├── main.js
 │   └── lib.js
 ├── runtime/business-client/
-├── skills/feature-management/
+├── skills/<skill-name>/
 ├── release-policy.json
 ├── release.json
 ├── checksums.json
@@ -76,7 +78,7 @@ Host 副本、日志、会话数据、attic、凭证和实验文件。
 
 完成标准：正式 CLI 不需要 Python、`uv` 或第三方 npm runtime 依赖，旧 Python client 可删除。
 
-### Step 2：重建纯 feature-management Skill（已完成）
+### Step 2：重建纯 Skills（已完成）
 
 - 保留 `SKILL.md`、`agents/`、`references/` 和 `assets/catalog-template/`；
 - npm staging 时排除 `assets/wheels/`、Skill installer 和 CLI launcher；这些旧 Python
@@ -94,7 +96,7 @@ Host 副本、日志、会话数据、attic、凭证和实验文件。
 只有 Step 1 和 Step 2 都完成后，才开始 npm dispatcher/installer：
 
 - 定义 `release.json` v2 和 `checksums.json` 覆盖范围；
-- 确定 `runtime/business-client` 与 `skills/feature-management` 并列目录；
+- 确定 `runtime/business-client` 与 `skills/<skill-name>` 并列目录；
 - 定义 `ml-platform` 安装管理命令和业务命令的无歧义分发；
 - 定义 `npx ... install` 如何建立持久 CLI，禁止引用一次性 npx cache；
 - 定义 CLI home、shim、PATH 检查和跨平台失败行为；
@@ -106,7 +108,7 @@ Host 副本、日志、会话数据、attic、凭证和实验文件。
 ### Step 4：重构 release builder（已完成）
 
 - 从 `business-client-js` 复制唯一 client tree 到 `runtime/business-client`；
-- 从 canonical Skill 复制纯 Skill 到 `skills/feature-management`；
+- 从仓库 canonical `skills/<skill-name>/` 复制全部纯 Skill；
 - 生成 release v2 和 checksums；
 - 拒绝多份 client、Skill runtime、旧 wheel 和未知文件；
 - 执行 npm package allowlist 和可重现 digest 检查；
@@ -134,7 +136,7 @@ Host 副本、日志、会话数据、attic、凭证和实验文件。
 
 ### Step 7：Skill + CLI 端到端发布验收（已完成）
 
-- `git diff --check -- npm business-client-js skills/feature-management agent_poc/codex-host`；
+- `git diff --check -- npm business-client-js skills agent_poc/codex-host`；
 - JS business-client tests；
 - JavaScript OpenAPI contract 与 Catalog tests；
 - Skill validator 和 forward tests；

package/README.md

@@ -1,6 +1,7 @@
 # ML Platform
 
-ML Platform 的 JavaScript CLI、`feature-management` 与 `model-lifecycle-management` Agent Skill。CLI 与 Skill 由同一个 npm
+ML Platform 的 JavaScript CLI、`feature-management`、`model-lifecycle-management` 与
+`model-deployment-management` Agent Skill。CLI 与 Skill 由同一个 npm
 release 安装和升级，运行时不需要 Python 或 `uv`。
 
 ## 环境要求
@@ -17,7 +18,7 @@ release 安装和升级，运行时不需要 Python 或 `uv`。
 npx @zhuoyuezs/ml-platform@latest install
 ```
 
-这条命令不需要额外参数。默认安装持久化的 `ml-platform` 命令，并将两个 Skill 安装到
+这条命令不需要额外参数。默认安装持久化的 `ml-platform` 命令，并将三个 Skill 安装到
 `~/.codex/skills/`，供当前用户的 Codex 使用。安装后重启终端和 Agent 会话。
 
 只在当前项目使用 Skill 时，在项目根目录运行：
@@ -26,7 +27,7 @@ npx @zhuoyuezs/ml-platform@latest install
 npx @zhuoyuezs/ml-platform@latest install --scope project
 ```
 
-此时 Skill 安装到当前目录的 `.agents/skills/feature-management/`。如果命令不在项目根目录运行，
+此时 Skill 安装到当前目录的 `.agents/skills/<skill-name>/`。如果命令不在项目根目录运行，
 再使用 `--project-dir /absolute/path/to/project` 显式指定项目。
 
 ### 安装参数
@@ -124,7 +125,8 @@ npx --yes --registry=https://registry.npmjs.org/ \
 ## 包内容
 
 - 一份 JavaScript business CLI；
-- 两份纯 Skill：`feature-management` 与 `model-lifecycle-management`；
+- 三份纯 Skill：`feature-management`、`model-lifecycle-management` 与
+  `model-deployment-management`；
 - 显式安装、升级、状态、校验和回滚逻辑。
 
 不包含 Codex CLI、Python business-client wheel、Agent Host、平台服务端、模型训练或 Kubernetes

package/checksums.json

@@ -2,17 +2,17 @@
   "files": [
     {
       "path": "runtime/business-client/README.md",
-      "sha256": "sha256:b823c0cafc47478fa1297c7c2e61e15b537db155586dd012b40d678d788c3ce3",
-      "size_bytes": 5158
+      "sha256": "sha256:69a21686d3b27682ad9864646f59353ca188454851941e63a88175951f516450",
+      "size_bytes": 7326
     },
     {
       "path": "runtime/business-client/package-lock.json",
-      "sha256": "sha256:a8da8027de475825949a661bc0d5674903e686d72ac945dcd964febfc5cf4667",
+      "sha256": "sha256:386e8b46ae8e6d7fd83e247b0412560710bee263c2545b593c5442b8f1363765",
       "size_bytes": 381
     },
     {
       "path": "runtime/business-client/package.json",
-      "sha256": "sha256:e1f916820a4bb19929a76f2690970068d17bb12bf640b89aae007eef50ef7b95",
+      "sha256": "sha256:83e60780d91abc78420a533ba0f8ea184f31886f963ca6f1554da819530a6335",
       "size_bytes": 501
     },
     {
@@ -22,8 +22,8 @@
     },
     {
       "path": "runtime/business-client/src/cli.js",
-      "sha256": "sha256:4f135325b57725c614d266285013df799cdd9a2096d8fbe0a9e9608dae2bcdb7",
-      "size_bytes": 44947
+      "sha256": "sha256:566ce8f224b74f03fcf168f5f2c8102505f1041b85862e8c407ce9103c96424f",
+      "size_bytes": 53638
     },
     {
       "path": "runtime/business-client/src/config.js",
@@ -37,8 +37,8 @@
     },
     {
       "path": "skills/feature-management/SKILL.md",
-      "sha256": "sha256:e5b927611cd401bc5e6160a2d358369efd0c4f65d3fe5cbf21bbab9adb6bd978",
-      "size_bytes": 31991
+      "sha256": "sha256:88898ea83ff339512d5512537db485a2764420abb67afceb9dd090cc780c00e0",
+      "size_bytes": 33021
     },
     {
       "path": "skills/feature-management/agents/openai.yaml",
@@ -92,13 +92,13 @@
     },
     {
       "path": "skills/feature-management/references/commands.md",
-      "sha256": "sha256:c748f1877f6d2ac1f6ecc29e45cd270997cbd35b498a190decddb66c74b840ab",
-      "size_bytes": 13751
+      "sha256": "sha256:ffd106cd785fe0d07ee20f51b7f55ca520da2fc6367bc5bc1ec0218600146d6e",
+      "size_bytes": 13905
     },
     {
       "path": "skills/feature-management/references/contracts.md",
-      "sha256": "sha256:4e10ba13399de17561a10b1dead4e2fcc57e0962cf81eba5131d12d66c4ce013",
-      "size_bytes": 29965
+      "sha256": "sha256:ef582d4ee9ade16de2a5f97b8bf9546969b6f2859eb020986bf25952993df402",
+      "size_bytes": 30483
     },
     {
       "path": "skills/feature-management/references/operator-authoring.md",
@@ -115,10 +115,35 @@
       "sha256": "sha256:2c7a894330a649248e7f52c944422bcd7a3aba4558f7984d129dfd21c8736cf3",
       "size_bytes": 4540
     },
+    {
+      "path": "skills/model-deployment-management/SKILL.md",
+      "sha256": "sha256:a78f492b95258e2111e8f410f4bf731cb2fa71997f39d236d9bf75dda3e0ad18",
+      "size_bytes": 3082
+    },
+    {
+      "path": "skills/model-deployment-management/agents/openai.yaml",
+      "sha256": "sha256:55c1a33d59cf08e27e7c806157576a9f8cec079a0fc6775f299794304ae8b6e1",
+      "size_bytes": 238
+    },
+    {
+      "path": "skills/model-deployment-management/references/contracts.md",
+      "sha256": "sha256:49a2595106b9b0cb5a305acbd285cdccad2bf4aa7627d6fb6b966783372ea724",
+      "size_bytes": 3543
+    },
+    {
+      "path": "skills/model-deployment-management/references/operations.md",
+      "sha256": "sha256:4407d87c1debbc865095b2857a6ece7a728d04b25fd8772d4a79a0361b0724ea",
+      "size_bytes": 3474
+    },
+    {
+      "path": "skills/model-deployment-management/references/troubleshooting.md",
+      "sha256": "sha256:fcdd4b244d670513fe14a3f343af2074a24cac0e3b07645ad455658564314e95",
+      "size_bytes": 2852
+    },
     {
       "path": "skills/model-lifecycle-management/SKILL.md",
-      "sha256": "sha256:faa4f8c0dbae7d1c1e2ca94499ce9b13cc8b52e8229154c05ac6282ea5e47788",
-      "size_bytes": 3610
+      "sha256": "sha256:61a89145ef33a2d58c58d8f352006590a47cf9587705568dec1c8ffe7ed8938e",
+      "size_bytes": 4014
     },
     {
       "path": "skills/model-lifecycle-management/agents/openai.yaml",
@@ -137,18 +162,18 @@
     },
     {
       "path": "skills/model-lifecycle-management/references/packaging.md",
-      "sha256": "sha256:0992f2f9104cb0d65a6325756108211d3f94856e05eca12640e9792c07e1108f",
-      "size_bytes": 3250
+      "sha256": "sha256:ad4e519f895d121202f99ed038a0ac6a6db3d97bdf95473e942399163d7062a3",
+      "size_bytes": 5640
     },
     {
       "path": "skills/model-lifecycle-management/references/training-contracts.md",
-      "sha256": "sha256:b6d72f37adf40c7d3832dbe46645cd416485411dcd48ea7b93f6072d458922c9",
-      "size_bytes": 11825
+      "sha256": "sha256:28fe32b2673db8f6cb1e02053bf5709826786e48dbe9951eac7bb4d98368d0c4",
+      "size_bytes": 12644
     },
     {
       "path": "skills/model-lifecycle-management/references/training.md",
-      "sha256": "sha256:f7bb40919da78b99cad13d4f1f05fef136552409df5ba6e2838e1e6b4a2435a5",
-      "size_bytes": 5508
+      "sha256": "sha256:97c20f02b1179dc7b7527cb7b5e11978778ccb8b76cdf88c229bcfdb5786f045",
+      "size_bytes": 5629
     }
   ],
   "schema_version": "data_platform.ml_platform_checksums/v1"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zhuoyuezs/ml-platform",
-  "version": "0.1.16",
+  "version": "0.1.18",
   "description": "ML 数据平台 business CLI 与 Agent Skills 管理工具",
   "bin": {
     "ml-platform": "scripts/main.js"

package/release.json

@@ -13,8 +13,8 @@
     "entrypoint": "src/cli.js",
     "name": "ml-platform",
     "path": "runtime/business-client",
-    "sha256": "sha256:262508ceeef693cdd580e2952f49161d860cee257fbc87ee46c95979cf1285d1",
-    "version": "0.7.6"
+    "sha256": "sha256:4304c1a26444c60408541217bc252e871a2950d1e3d6259dc0bf6b27d294bf09",
+    "version": "0.7.8"
   },
   "policy_sha256": "sha256:02fbde0134696b78c0c365d6e0b76c9814227208da5b2b3d4e5dbd82dc4a3606",
   "public_schema_versions": [
@@ -25,7 +25,7 @@
     "ml_data_platform.feature_set/v1",
     "ml_data_platform.dataset_manifest/v1"
   ],
-  "release_version": "0.1.16",
+  "release_version": "0.1.18",
   "runtime_requirements": {
     "node": ">=18",
     "os": [
@@ -37,16 +37,22 @@
   "skills": {
     "feature-management": {
       "path": "skills/feature-management",
-      "requires_cli": ">=0.7.6 <0.8.0",
-      "revision": "0.1.16",
-      "sha256": "sha256:182b6f2df4fe58c946e482dfaf416c7d5b78c6fde9c8c8d7293103246618cb4e"
+      "requires_cli": ">=0.7.8 <0.8.0",
+      "revision": "0.1.18",
+      "sha256": "sha256:d32f0306f129d7bf76eb037fe239c84dd3c29c64b61600799d6e40320784bb10"
+    },
+    "model-deployment-management": {
+      "path": "skills/model-deployment-management",
+      "requires_cli": ">=0.7.8 <0.8.0",
+      "revision": "0.1.18",
+      "sha256": "sha256:b09f84d5270406d918cb5fe5818f84424ff41c1167c1b18f8d2b18f2e9b4a047"
     },
     "model-lifecycle-management": {
       "path": "skills/model-lifecycle-management",
-      "requires_cli": ">=0.7.6 <0.8.0",
-      "revision": "0.1.16",
-      "sha256": "sha256:61ba17f955b36d1c2e8329a10454d040ee7eca5afdb90ff0914d3f85b6e2d203"
+      "requires_cli": ">=0.7.8 <0.8.0",
+      "revision": "0.1.18",
+      "sha256": "sha256:7c79a6393e302579e2aa48c6e20754aa992776d0f8936f897bb0feeafc84e508"
     }
   },
-  "source_commit": "49bc1482ab3a58874e40e308640db26e13e6ef12"
+  "source_commit": "ec29c26"
 }

package/runtime/business-client/README.md

@@ -22,8 +22,15 @@ ml-platform get-dataset <dataset_id> <dataset_version> --project <project>
 ml-platform resolve-dataset <dataset_id> <dataset_version> --project <project>
 ml-platform build-registered-dataset <dataset_id> <dataset_version> \
   --project <project> --wait
+ml-platform register-model-artifact <artifact_id> --evaluation-id <evaluation_result_or_summary_id>
 ```
 
+`resolve-dataset` 返回两个不同的身份：`manifest_hash` 是解析后的 DatasetArtifact
+缓存/构建键，`registered_manifest_hash` 是注册 DatasetManifest 的 canonical 契约哈希。
+构造 `FeatureRetrievalSpec.manifest.manifest_hash` 时必须使用后者；客户端不需要、也不应
+自行复算 Python canonical hash。模型评估结果通过 `register-model-artifact` 交给服务端治理
+校验后注册，不要绕过 CLI 直接调用 HTTP 注册接口。
+
 `resolve-manifest` 和 `build-dataset <manifest.json>` 仍用于尚未注册的本地契约。Catalog
 normalizer 保留并校验非默认 `project`；`apply --project` 与 Catalog 内项目不一致时拒绝执行。
 
@@ -48,9 +55,37 @@ ml-platform get-parameter-lifecycle <name> <version> --project <project>
 ml-platform fetch-inference-context /absolute/path/to/request.json
 ```
 
-已部署模型的标准推理使用 Serving URL，不使用控制面 `--api-url`。Chronos-2 v2 的
+受治理部署通过同一业务 CLI 完成预检、创建、状态等待、晋级、回滚和下线。正常业务操作
+不使用 `kubectl`，也不创建、修改或删除 NetworkPolicy：
+
+```bash
+ml-platform validate-deployment /absolute/path/to/deployment.json \
+  --evaluation-id <pass-evaluation-id>
+ml-platform create-deployment /absolute/path/to/deployment.json \
+  --evaluation-id <pass-evaluation-id>
+ml-platform wait-deployment <deployment-id> --wait-timeout-seconds 1800
+ml-platform get-deployment <deployment-id>
+ml-platform get-deployment-events <deployment-id> --limit 100 --offset 0
+```
+
+`READY` 必须同时包含 `ReplicasReady=true`、`ArtifactsVerified=true` 和
+`SmokePredictionPassed=true`。`get-deployment` 返回控制器实际观察到的 endpoint 元数据；
+本地 CLI 不推测 Kubernetes Service 名称。部署后验证优先使用控制面代理：
+
+```bash
+ml-platform predict-deployment <deployment-id> /absolute/path/to/request.json
+```
+
+部署文件可用 `inference_policy` 配置默认执行 horizon 和半开输出窗口。例如 15 分钟频率、
+当天 00:00 cutoff、只返回次日全天时，配置 `horizon=2d` 和 `[1d,2d)`；模型必须具备
+192 点输出能力，响应返回 96 点。请求可省略 `horizon` 和 `output_window` 使用部署默认值，
+也可携带其中任意一项覆盖；覆盖值仍受模型长度、频率和窗口边界校验。
+
+直接访问已知外部 Serving URL 时仍可使用 `predict-model`。Chronos-2 v2 的
 `known_future` 协变量放在 inline 请求的 `inputs.future_covariates` 中；每个预测步一行，
-行数必须等于模型的 `prediction_length`。下例为 15 分钟频率、预测 30 分钟；预测明天时
+行数必须等于模型的 `prediction_length`。`feature_lookup` 模式不得传该字段；Serving 会从
+部署绑定的 FeatureSet 在线读取完整未来 Feature 窗口，并只使用
+`available_at <= cutoff_time` 的修订。下例为 15 分钟频率、预测 30 分钟；预测明天时
 同样提供整天的所有未来步：
 
 ```json

package/runtime/business-client/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@zhuoyuezs/ml-platform-business-client",
-  "version": "0.7.6",
+  "version": "0.7.8",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@zhuoyuezs/ml-platform-business-client",
-      "version": "0.7.6",
+      "version": "0.7.8",
       "license": "UNLICENSED",
       "bin": {
         "ml-platform": "src/cli.js"

package/runtime/business-client/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zhuoyuezs/ml-platform-business-client",
-  "version": "0.7.6",
+  "version": "0.7.8",
   "private": true,
   "description": "JavaScript implementation of the ML Platform business client",
   "bin": {

package/runtime/business-client/src/cli.js

@@ -34,7 +34,11 @@ const BUSINESS_COMMANDS = new Set([
   "register-executable-package", "verify-executable-package", "publish-executable-package",
   "validate-evaluation-run", "submit-evaluation", "get-evaluation-run", "get-evaluation-job", "retry-evaluation-job", "cancel-evaluation-job", "list-evaluation-attempts",
   "get-evaluation-result", "get-evaluation-result-metrics", "get-evaluation-result-artifacts", "summarize-evaluation-run", "get-evaluation-summary",
-  "create-model-package", "get-model-package", "cancel-model-package", "validate-model-artifact",
+  "create-model-package", "get-model-package", "cancel-model-package", "validate-model-artifact", "register-model-artifact",
+  "list-deployment-targets", "validate-deployment", "create-deployment", "list-deployments",
+  "get-deployment", "wait-deployment", "reconcile-deployment", "approve-deployment",
+  "promote-deployment", "rollback-deployment", "suspend-deployment", "resume-deployment",
+  "undeploy-deployment", "get-deployment-events", "predict-deployment",
 ]);
 
 const COMMAND_USAGE = {
@@ -138,6 +142,22 @@ const COMMAND_USAGE = {
   "get-model-package": "get-model-package PACKAGE_ID [--package-version VERSION]",
   "cancel-model-package": "cancel-model-package PACKAGE_ID [--package-version VERSION]",
   "validate-model-artifact": "validate-model-artifact ARTIFACT_ID",
+  "register-model-artifact": "register-model-artifact ARTIFACT_ID --evaluation-id EVALUATION_ID",
+  "list-deployment-targets": "list-deployment-targets",
+  "validate-deployment": "validate-deployment SPEC_JSON --evaluation-id ID",
+  "create-deployment": "create-deployment SPEC_JSON --evaluation-id ID",
+  "list-deployments": "list-deployments [--service SERVICE] [--environment ENV] [--state STATE] [-q QUERY] [--limit N] [--offset N]",
+  "get-deployment": "get-deployment DEPLOYMENT_ID",
+  "wait-deployment": "wait-deployment DEPLOYMENT_ID [--poll-interval SECONDS] [--wait-timeout-seconds SECONDS]",
+  "reconcile-deployment": "reconcile-deployment DEPLOYMENT_ID",
+  "approve-deployment": "approve-deployment DEPLOYMENT_ID --revision N --approver NAME --decision approved|rejected [--reason TEXT]",
+  "promote-deployment": "promote-deployment DEPLOYMENT_ID --stage canary|production --actor NAME --reason TEXT [--approval-id ID]",
+  "rollback-deployment": "rollback-deployment DEPLOYMENT_ID --actor NAME --reason TEXT [--to-revision N] [--to-deployment-id ID]",
+  "suspend-deployment": "suspend-deployment DEPLOYMENT_ID --actor NAME --reason TEXT",
+  "resume-deployment": "resume-deployment DEPLOYMENT_ID --actor NAME --reason TEXT",
+  "undeploy-deployment": "undeploy-deployment DEPLOYMENT_ID --actor NAME --reason TEXT",
+  "get-deployment-events": "get-deployment-events DEPLOYMENT_ID [--limit N] [--offset N]",
+  "predict-deployment": "predict-deployment DEPLOYMENT_ID REQUEST_JSON",
   apply: "apply CATALOG_DIRECTORY [--dry-run] [--project PROJECT]",
 };
 
@@ -291,6 +311,28 @@ async function waitForJob(api, jobId, pollInterval, timeoutSeconds) {
   }
 }
 
+function observedDeploymentState(payload) {
+  const state = payload?.observed_state?.observed_state ?? payload?.observed_state;
+  if (typeof state !== "string") throw new Error("deployment status response has no observed state");
+  return state;
+}
+
+async function waitForDeployment(api, deploymentId, pollInterval, timeoutSeconds) {
+  if (!(pollInterval > 0)) throw new Error("deployment poll interval must be positive");
+  if (timeoutSeconds !== undefined && !(timeoutSeconds > 0)) throw new Error("deployment wait timeout must be positive");
+  const deadline = timeoutSeconds === undefined ? Infinity : Date.now() + timeoutSeconds * 1000;
+  let previous;
+  while (true) {
+    const payload = await api.get(`/deployments/${encodeURIComponent(deploymentId)}`);
+    if (!payload || typeof payload !== "object" || Array.isArray(payload)) throw new Error("deployment status response must be a JSON object");
+    const state = observedDeploymentState(payload);
+    if (state !== previous) { process.stderr.write(`deployment ${deploymentId}: ${state}\n`); previous = state; }
+    if (["READY", "FAILED", "DEGRADED", "SUPERSEDED", "SUSPENDED", "UNDEPLOYED", "ROLLED_BACK"].includes(state)) return payload;
+    if (Date.now() >= deadline) throw new Error(`timed out waiting for deployment ${deploymentId}`);
+    await new Promise((resolve) => setTimeout(resolve, pollInterval * 1000));
+  }
+}
+
 async function runBusinessCli(argv) {
   if (argv.length === 0) throw new Error("a business command is required");
   if (argv.includes("--help") || argv.includes("-h")) {
@@ -390,9 +432,8 @@ async function runBusinessCli(argv) {
     const out = take(rest, "--out");
     if (out) { const target = resolvedPath(out); fs.mkdirSync(path.dirname(target), { recursive: true }); fs.writeFileSync(target, `${JSON.stringify(result, null, 2)}\n`); }
   } else if (options.command === "resolve-dataset") {
-    const datasetId = positional(rest, "dataset id"); const datasetVersion = positional(rest, "dataset version"); const project = take(rest, "--project", "default"); const api = client(options);
-    const manifest = normalizeRegistryDataset(await registeredDataset(api, project, datasetId, datasetVersion));
-    result = await api.post("/datasets/resolve", manifest);
+    const datasetId = positional(rest, "dataset id"); const datasetVersion = positional(rest, "dataset version"); const project = take(rest, "--project", "default");
+    result = await client(options).post(`/datasets/${encodeURIComponent(project)}/${encodeURIComponent(datasetId)}/${encodeURIComponent(datasetVersion)}/resolve`);
     const out = take(rest, "--out");
     if (out) { const target = resolvedPath(out); fs.mkdirSync(path.dirname(target), { recursive: true }); fs.writeFileSync(target, `${JSON.stringify(result, null, 2)}\n`); }
   } else if (options.command === "build-dataset") {
@@ -457,6 +498,91 @@ async function runBusinessCli(argv) {
     const request = readSpec(positional(rest, "request JSON")); const servingUrl = take(rest, "--serving-url");
     if (!servingUrl) throw new Error("--serving-url is required");
     result = await new PlatformApiClient(servingUrl, options.timeout).post("/v1/inference", request);
+  } else if (options.command === "list-deployment-targets") {
+    result = await client(options).get("/deployment-targets");
+  } else if (["validate-deployment", "create-deployment"].includes(options.command)) {
+    const deployment = readSpec(positional(rest, "deployment spec JSON"));
+    const evaluationId = take(rest, "--evaluation-id");
+    if (!evaluationId) throw new Error("--evaluation-id is required");
+    const endpoint = options.command === "validate-deployment" ? "/deployments/validate" : "/deployments";
+    result = await client(options).post(endpoint, { deployment, evaluation_id: evaluationId });
+  } else if (options.command === "list-deployments") {
+    const params = {
+      service: take(rest, "--service"),
+      environment: take(rest, "--environment"),
+      state: take(rest, "--state"),
+      q: take(rest, "--q", take(rest, "-q")),
+    };
+    boolean(rest, "--all");
+    result = await listCollection(client(options), "/deployments", params, rest);
+  } else if (options.command === "get-deployment") {
+    const deploymentId = positional(rest, "deployment id");
+    result = await client(options).get(`/deployments/${encodeURIComponent(deploymentId)}`);
+  } else if (options.command === "wait-deployment") {
+    const deploymentId = positional(rest, "deployment id");
+    const poll = finiteNumber(take(rest, "--poll-interval", 2), "--poll-interval");
+    const rawTimeout = take(rest, "--wait-timeout-seconds");
+    result = await waitForDeployment(
+      client(options),
+      deploymentId,
+      poll,
+      rawTimeout === undefined ? undefined : finiteNumber(rawTimeout, "--wait-timeout-seconds"),
+    );
+    if (observedDeploymentState(result) !== "READY") exitCode = 1;
+  } else if (options.command === "reconcile-deployment") {
+    const deploymentId = positional(rest, "deployment id");
+    result = await client(options).post(`/deployments/${encodeURIComponent(deploymentId)}/reconcile`);
+  } else if (options.command === "approve-deployment") {
+    const deploymentId = positional(rest, "deployment id");
+    const rawRevision = take(rest, "--revision");
+    const approver = take(rest, "--approver");
+    if (rawRevision === undefined) throw new Error("--revision is required");
+    if (!approver) throw new Error("--approver is required");
+    const decision = take(rest, "--decision", "approved");
+    if (!["approved", "rejected"].includes(decision)) throw new Error("--decision must be approved or rejected");
+    result = await client(options).post(`/deployments/${encodeURIComponent(deploymentId)}/approvals`, {
+      revision: integer(rawRevision, "--revision"), stage: "production", decision, approver,
+      reason: take(rest, "--reason", ""),
+    });
+  } else if (options.command === "promote-deployment") {
+    const deploymentId = positional(rest, "deployment id");
+    const stage = take(rest, "--stage"); const actor = take(rest, "--actor"); const reason = take(rest, "--reason");
+    if (!["canary", "production"].includes(stage)) throw new Error("--stage must be canary or production");
+    if (!actor) throw new Error("--actor is required");
+    if (!reason) throw new Error("--reason is required");
+    result = await client(options).post(`/deployments/${encodeURIComponent(deploymentId)}/promote`, {
+      stage, actor, reason, approval_id: take(rest, "--approval-id"),
+    });
+  } else if (options.command === "rollback-deployment") {
+    const deploymentId = positional(rest, "deployment id");
+    const actor = take(rest, "--actor"); const reason = take(rest, "--reason");
+    if (!actor) throw new Error("--actor is required");
+    if (!reason) throw new Error("--reason is required");
+    const rawRevision = take(rest, "--to-revision");
+    result = await client(options).post(`/deployments/${encodeURIComponent(deploymentId)}/rollback`, {
+      actor, reason,
+      to_revision: rawRevision === undefined ? undefined : integer(rawRevision, "--to-revision"),
+      to_deployment_id: take(rest, "--to-deployment-id"),
+    });
+  } else if (["suspend-deployment", "resume-deployment", "undeploy-deployment"].includes(options.command)) {
+    const deploymentId = positional(rest, "deployment id");
+    const actor = take(rest, "--actor"); const reason = take(rest, "--reason");
+    if (!actor) throw new Error("--actor is required");
+    if (!reason) throw new Error("--reason is required");
+    const action = options.command.split("-")[0];
+    result = await client(options).post(`/deployments/${encodeURIComponent(deploymentId)}/${action}`, { actor, reason });
+  } else if (options.command === "get-deployment-events") {
+    const deploymentId = positional(rest, "deployment id");
+    result = await listCollection(
+      client(options),
+      `/deployments/${encodeURIComponent(deploymentId)}/events`,
+      {},
+      rest,
+    );
+  } else if (options.command === "predict-deployment") {
+    const deploymentId = positional(rest, "deployment id");
+    const request = readSpec(positional(rest, "request JSON"));
+    result = await client(options).post(`/deployments/${encodeURIComponent(deploymentId)}/predict`, request);
   } else if (options.command === "get-model-runtime-release") {
     result = await client(options).get("/model-runtime-release");
   } else if (options.command.startsWith("list-") && LIFECYCLE_COLLECTIONS.has(options.command.slice(5))) {
@@ -506,6 +632,11 @@ async function runBusinessCli(argv) {
     const name = positional(rest, "package name"); const version = positional(rest, "package version"); const action = options.command.split("-")[0];
     result = await client(options).post(`/executable-packages/${encodeURIComponent(name)}/${encodeURIComponent(version)}/${action}`);
   } else if (options.command === "validate-model-artifact") result = await client(options).get(`/model-artifacts/${encodeURIComponent(positional(rest, "model artifact id"))}/validate`);
+  else if (options.command === "register-model-artifact") {
+    const artifactId = positional(rest, "model artifact id"); const evaluationId = take(rest, "--evaluation-id");
+    if (!evaluationId) throw new Error("--evaluation-id is required");
+    result = await client(options).post(`/models/${encodeURIComponent(artifactId)}/register`, { evaluation_id: evaluationId });
+  }
   else if (/^(get|cancel)-model-package$/.test(options.command)) {
     const packageId = positional(rest, "model package id"); const params = { package_version: take(rest, "--package-version") }; const endpoint = `/model-packages/${encodeURIComponent(packageId)}`;
     result = options.command === "get-model-package" ? await client(options).get(endpoint, params) : await client(options).post(`${endpoint}/cancel`, undefined, params);
@@ -529,7 +660,7 @@ async function listCollection(api, endpoint, params, rest) {
   return api.get(endpoint, { ...params, limit, offset });
 }
 
-module.exports = { BUSINESS_COMMANDS, fetchAll, listCollection, parse, runBusinessCli, usage, waitForJob };
+module.exports = { BUSINESS_COMMANDS, fetchAll, listCollection, parse, runBusinessCli, usage, waitForDeployment, waitForJob };
 
 if (require.main === module) {
   runBusinessCli(process.argv.slice(2)).then((code) => { process.exitCode = code; }).catch((error) => {

package/skills/feature-management/SKILL.md

@@ -205,11 +205,14 @@ platform source checkout or a local Catalog directory for a matching JSON file.
    target from a similar local filename.
 4. Run `get-dataset <dataset_id> <dataset_version> --project <project>` and
    inspect the returned registered contract.
-5. Run `resolve-dataset` for a read-only dependency and manifest-hash check.
+5. Run `resolve-dataset` for a read-only dependency and identity check. Use its
+   `registered_manifest_hash` for `FeatureRetrievalSpec.manifest.manifest_hash`;
+   use its `manifest_hash` only as the resolved DatasetArtifact/build key.
 6. Submit exactly one `build-registered-dataset ... --wait` only after the user
    authorizes that exact registered version.
-7. Use the returned Job ID with `get-job`, then use the returned manifest hash
-   with `get-dataset-artifact` after success.
+7. Use the returned Job ID with `get-job`, then use the resolved `manifest_hash`
+   with `get-dataset-artifact` after success. Keep the separate
+   `registered_manifest_hash` for FeatureRetrievalSpec binding.
 
 Local `resolve-manifest` and `build-dataset <file>` remain authoring tools for a
 new or changed Catalog contract before it is registered. They are not required
@@ -419,6 +422,17 @@ replayable `as_of` manifest; when an online source override is active, `replay.s
 must be false because the offline manifest cannot reproduce that source. Report
 tolerated missing or gap-filled inputs as degraded freshness, not as a normal read.
 
+Model-serving `feature_lookup` may additionally request declared known-future
+Features from the same FeatureSet. Do not emulate that by passing model inputs
+through this Skill. Author their Parameters with trustworthy availability
+evidence: future event times are valid, but only revisions with
+`available_at <= cutoff_time` may be used. `available_at` may therefore be
+earlier than `event_time` for forecasts and schedules. The platform reads only
+the requested known-future Features and their Parameter dependency closure;
+unrelated historical Parameters in the same DatasetManifest do not need a
+future-visibility contract. Every Parameter inside the requested closure still
+fails closed when trustworthy availability is absent.
+
 ## Verify The Artifact
 
 Treat Registry metadata inspection and downloaded-file inspection as separate
@@ -461,7 +475,7 @@ Require all of the following before reporting success:
   windows tiling the fetch range, and any retries are accounted for;
 - missing values and warnings are reported, not silently repaired.
 
-The manifest hash returned by `resolve-manifest` or `resolve-dataset` is the artifact key: it stays
+The `manifest_hash` returned by `resolve-manifest` or `resolve-dataset` is the artifact key: it stays
 the same after the build, so it is usable for cache reuse and polling before
 submission. Distinguish the two lookup failures rather than treating both as "not
 ready" — `422` with `invalid_artifact_key` means the key can never name an
@@ -480,7 +494,7 @@ new / unchanged assets by type
 planned versus resolved dependency versions and any explicitly retained old versions
 Operator package filename and SHA-256
 FeatureSet and ordered feature count
-dataset id, version, manifest hash, and Job id
+ dataset id, version, resolved manifest hash, registered manifest hash when applicable, and Job id
 row and column counts
 rowset candidate / policy-applied / eligible / ineligible counts when endpoint_policy is declared
 validation status, warnings, and missing-data summary

package/skills/feature-management/references/commands.md

@@ -220,7 +220,8 @@ DatasetManifest hash. The initial implementation is sequential, so keep
 `--max-parallelism 1`. Completed partitions are checkpointed and resumed by
 default; use `--no-resume` only to force a clean execution.
 
-Record the returned `job_id`, `dataset_id`, and `manifest_hash`. If a separate wait is needed:
+Record the returned `job_id`, `dataset_id`, and resolved `manifest_hash`; for a registered Dataset
+also retain `registered_manifest_hash` for FeatureRetrievalSpec binding. If a separate wait is needed:
 
 ```bash
 ml-platform --profile server wait-job <job_id>
@@ -266,7 +267,8 @@ ml-platform --profile server fetch-inference-context /tmp/inference-context.json
 ```
 
 The `FeatureRetrievalSpec` must bind an existing inference-mode DatasetManifest
-by project, dataset id, version, and manifest hash; its FeatureSet and contract
+by project, dataset id, version, and the `registered_manifest_hash` returned by
+`resolve-dataset`; its FeatureSet and contract
 signature must match the resolved online contract. Context mode also fixes
 `context_length`, `frequency`, `context_end`, `stride`, input-adapter hash, and
 temporal-binding hash. Do not replace these immutable deployment bindings with

package/skills/feature-management/references/contracts.md

@@ -138,6 +138,12 @@ supported field.
 当前严格 `as_of` 只接受 `exact` 或有合同支撑的 `upper_bound`；修改这些语义需要发布
 新的 Parameter 业务版本。
 
+未来预报、排程、日历等 known-future 数据允许 `available_at < event_time`。这不是时钟错误：
+`event_time` 是未来目标时刻，`available_at` 是该修订何时已可读取。模型
+`feature_lookup` 会读取 cutoff 后的未来事件窗口，但仍严格过滤
+`available_at <= cutoff_time`。因此这类 Parameter 应优先使用能表达预报签发时间或计划发布
+时间的 `source_field`，不能用事后实测值替代，也不能用调用方请求绕过 availability 合同。
+
 An optional `rounding` section states the fixed-point precision the platform
 enforces at the Parameter boundary:
 

package/skills/model-deployment-management/SKILL.md

@@ -0,0 +1,53 @@
+---
+name: model-deployment-management
+description: Manage governed ITSMP model deployments and post-deployment predictions through the ml-platform business CLI. Use when a user asks to validate or create a DeploymentSpec, inspect targets or rollout state, wait for readiness, record approval, promote, roll back, suspend, resume, undeploy, inspect deployment events, or invoke a deployed model. Do not use for training, evaluation, packaging, image release engineering, or Kubernetes administration.
+---
+
+# Model Deployment Management
+
+Use the `ml-platform` executable installed with the same npm release as this
+Skill. Normal deployment operations go through the ITSMP control plane; do not
+use `kubectl` to create, promote, roll back, suspend, resume, or undeploy a
+business model.
+
+```text
+REGISTERED ModelArtifact + PASS EvaluationResult + READY ModelPackage
+  -> validated DeploymentSpec -> DeploymentRevision -> READY -> prediction
+```
+
+Read [references/contracts.md](references/contracts.md) before authoring or
+validating a DeploymentSpec. Read
+[references/operations.md](references/operations.md) immediately before a
+deployment mutation or prediction. Read
+[references/troubleshooting.md](references/troubleshooting.md) when a
+deployment does not reach `READY` or an operation is rejected.
+
+## Required Gates
+
+- Verify `ml-platform version`, `ml-platform show-config`, and
+  `ml-platform --profile server health` before control-plane work.
+- Use only a `READY` ModelPackage image with an immutable `@sha256:` digest and
+  the package-bound FeatureRetrievalSpec and smoke request.
+- When `DeploymentSpec.inference_policy` is configured, treat its horizon and
+  half-open output window as request defaults. Prediction requests may omit or
+  override `horizon` and `output_window`; every effective combination must fit
+  the model capacity and sampling grid. In `feature_lookup` mode, never send
+  `future_covariates`; the platform retrieves the complete known-future Feature
+  window visible at the cutoff. Inline mode still supplies those rows in inputs.
+- Run `validate-deployment` before `create-deployment`. Validation is read-only
+  and does not authorize creation.
+- Obtain explicit authorization for the exact deployment and operation before
+  create, reconcile, approval, promotion, rollback, suspend, resume, or
+  undeploy. Retry requires renewed authorization when the previous mutation
+  returned an ambiguous result.
+- Treat `READY` as valid only when `ReplicasReady`, `ArtifactsVerified`, and
+  `SmokePredictionPassed` are all true for the current revision.
+- Use `predict-deployment`; it routes through the control plane to the recorded
+  deployment endpoint. Do not infer Kubernetes Service names or require the
+  user to reach a cluster-local URL.
+- Never create, modify, or delete NetworkPolicy. Cluster operators own that
+  boundary.
+
+Report the deployment ID, revision, immutable image digest, desired and
+observed state, failed conditions, endpoint reachability, and event evidence.
+Do not claim a deployment is usable merely because its Pod is running.

package/skills/model-deployment-management/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "模型部署管理"
+  short_description: "管理 ITSMP 模型部署、晋级、回滚与部署后推理"
+  default_prompt: "使用 $model-deployment-management 管理受治理模型部署及部署后推理。"

package/skills/model-deployment-management/references/contracts.md

@@ -0,0 +1,82 @@
+# Deployment Contracts
+
+## DeploymentSpec
+
+Use a JSON object accepted by `ml-platform validate-deployment` with these
+load-bearing fields:
+
+```json
+{
+  "schema_version": "ml_data_platform.deployment/v1",
+  "deployment_id": "deployment_forecast_staging_01",
+  "service": "forecast-service",
+  "environment": "staging",
+  "model_artifact_id": "artifact_01",
+  "package_id": "package_01",
+  "package_version": "1",
+  "image": "registry.example/model-serving@sha256:<64-hex>",
+  "target": {"type": "kubernetes", "namespace": "itsmp-model-serving"},
+  "feature_retrieval": {},
+  "inference_policy": {
+    "horizon": {"duration": "2d"},
+    "output_window": {"start_offset": "1d", "end_offset": "2d"},
+    "allow_request_override": true
+  },
+  "replicas": 1,
+  "resources": {},
+  "rollout": {"strategy": "canary", "traffic_percent": 10},
+  "stage": "staging"
+}
+```
+
+Do not reconstruct `feature_retrieval` or `feature_lookup_test` from memory.
+They must match the registered artifact lineage and the tested ModelPackage.
+The server normalizes an omitted lookup test from the package during
+validation and creation.
+
+`inference_policy` is optional. When present, its horizon and offsets must be
+whole-minute durations aligned to the model sampling frequency. The half-open
+window `[cutoff+start_offset, cutoff+end_offset)` must be non-empty, end within
+the configured horizon, and fit the model's maximum sequence output. A 15-minute
+model using `horizon=2d` and `[1d,2d)` therefore needs at least 192 output steps
+and returns 96 steps. This is a response projection, not extra model capacity.
+
+The deployment values are defaults. Calls may omit `horizon` and
+`output_window`, override either one, or override both. Serving validates the
+effective values again; an output window must remain non-empty, aligned, inside
+the effective horizon, and within model capacity. Set
+`allow_request_override=false` only when the endpoint must reject non-equivalent
+request values.
+
+Known-future covariates still cover the model's complete prediction length,
+including points outside a shorter returned window. In `feature_lookup` mode,
+Serving requests those Feature columns from the deployment-bound FeatureSet;
+the platform permits future event times only when `available_at <= cutoff_time`.
+The prediction caller must not send top-level `future_covariates`. Inline mode
+continues to carry them in `inputs.future_covariates`.
+
+The first deployment starts at `staging`. A replacement for an existing
+service/environment declares the active leaf in `previous_deployment_id`.
+Production is reached only by staging to canary to production promotion.
+
+## Revision State
+
+`DeploymentSpec` is immutable. Its inference policy and full deployment contract
+hash are copied into each immutable revision. Promotion, rollback, suspension,
+resumption, and undeploy append revisions that retain that policy. Reconcile
+writes mutable observed state separately.
+
+`READY` requires all three current-revision conditions:
+
+- `ReplicasReady=true`
+- `ArtifactsVerified=true`
+- `SmokePredictionPassed=true`
+
+The detail response includes `endpoints`. Cluster-local endpoints have
+`externally_reachable=false`; a Gateway endpoint is externally reachable only
+when rollout hostname metadata exists. `available=true` means the current
+revision passed all READY gates.
+
+Successful smoke verification is cached durably by revision, image digest,
+FeatureRetrieval contract hash, deployment contract hash, and frozen lookup-test
+hash. A new revision or changed identity must run smoke again.

package/skills/model-deployment-management/references/operations.md

@@ -0,0 +1,111 @@
+# Deployment Operations
+
+## Initialize
+
+```bash
+command -v ml-platform
+ml-platform version
+ml-platform show-config
+ml-platform --profile server health
+ml-platform --profile server list-deployment-targets
+```
+
+## Validate And Create
+
+Validation and creation are separate authorization points:
+
+```bash
+ml-platform --profile server validate-deployment /absolute/path/deployment.json \
+  --evaluation-id <pass-evaluation-id>
+
+ml-platform --profile server create-deployment /absolute/path/deployment.json \
+  --evaluation-id <pass-evaluation-id>
+
+ml-platform --profile server wait-deployment <deployment-id> \
+  --poll-interval 5 --wait-timeout-seconds 1800
+```
+
+Use `get-deployment` to inspect the full revision history and endpoints. Use
+`list-deployments` for discovery and `get-deployment-events` for paginated
+audit evidence.
+
+## Promotion And Approval
+
+```bash
+ml-platform --profile server promote-deployment <deployment-id> \
+  --stage canary --actor <actor> --reason <reason>
+
+ml-platform --profile server approve-deployment <deployment-id> \
+  --revision <canary-revision> --approver <approver> \
+  --decision approved --reason <reason>
+
+ml-platform --profile server promote-deployment <deployment-id> \
+  --stage production --actor <actor> --reason <reason> \
+  --approval-id <approval-id>
+```
+
+Do not approve production before the canary revision is `READY` and the user
+has authorized that exact revision.
+
+## Rollback And Offline Operations
+
+```bash
+ml-platform --profile server rollback-deployment <deployment-id> \
+  --to-revision <revision> --actor <actor> --reason <reason>
+
+ml-platform --profile server suspend-deployment <deployment-id> \
+  --actor <actor> --reason <reason>
+
+ml-platform --profile server resume-deployment <deployment-id> \
+  --actor <actor> --reason <reason>
+
+ml-platform --profile server undeploy-deployment <deployment-id> \
+  --actor <actor> --reason <reason>
+```
+
+Suspend is recoverable. Undeploy removes runtime resources but retains revision
+history; it is not equivalent to resume.
+
+## Prediction
+
+With a deployment-level inference policy, the request may use both deployment
+defaults by omitting `horizon` and `output_window`:
+
+```json
+{
+  "input_mode": "feature_lookup",
+  "cutoff_time": "2026-09-14T00:00:00+08:00",
+  "entity": {}
+}
+```
+
+To override the revision defaults, add either or both fields. This example keeps
+the deployment horizon but returns its first day:
+
+```json
+{
+  "input_mode": "feature_lookup",
+  "cutoff_time": "2026-09-14T00:00:00+08:00",
+  "entity": {},
+  "output_window": {"start_offset": "0min", "end_offset": "1d"}
+}
+```
+
+Do not send `future_covariates` in `feature_lookup` mode. Serving derives the
+declared known-future columns and full model prediction length from the
+ModelArtifact, then the platform retrieves visible future Feature revisions.
+Only those Features and their Parameter dependency closure are read; unrelated
+historical Parameters in the same inference Manifest are not evaluated as
+future inputs. Requested future Parameters must still provide trustworthy
+`available_at <= cutoff_time` evidence.
+For inline mode only, supply one row per model prediction step in
+`inputs.future_covariates`.
+
+```bash
+ml-platform --profile server predict-deployment <deployment-id> \
+  /absolute/path/request.json
+```
+
+Prediction is allowed only for a current `READY` revision with all readiness
+conditions true. This command uses the control-plane proxy and does not require
+a cluster-local Serving URL.

package/skills/model-deployment-management/references/troubleshooting.md

@@ -0,0 +1,50 @@
+# Deployment Troubleshooting
+
+Start with read-only control-plane evidence:
+
+```bash
+ml-platform --profile server get-deployment <deployment-id>
+ml-platform --profile server get-deployment-events <deployment-id> \
+  --limit 100 --offset 0
+```
+
+Interpret conditions before considering a reconcile:
+
+- `ReplicasReady=false`: inspect scheduler, image pull, resource quota, and Pod
+  readiness through the platform/cluster diagnostic boundary.
+- `ArtifactsVerified=false`: the Serving `/model` identity does not match the
+  revision, or the endpoint is unavailable.
+- `SmokePredictionPassed=false`: preserve the frozen package smoke request and
+  error. Do not replace it with a current-time request to force success.
+- `request horizon cannot override deployment inference_policy` or
+  `request output_window cannot override deployment inference_policy`: the
+  revision has `allow_request_override=false`; use equivalent values or create
+  a reviewed revision whose policy permits the required request behavior.
+- `inference_policy output window must end within the configured horizon`:
+  override `horizon` and `output_window` together, or keep the deployment
+  defaults. A shorter horizon cannot retain a later default window.
+- `feature_lookup obtains known future covariates from online retrieval`: remove
+  top-level `future_covariates`. Verify the columns belong to the bound
+  FeatureSet and their Parameters provide trustworthy `available_at` values.
+- `availability_not_supported` or `available_at_missing` during future lookup:
+  publish a corrected Parameter/Feature dependency version with a source-field
+  or justified upper-bound availability contract; do not substitute caller data.
+- `inference_policy horizon exceeds model maximum prediction length`: retrain
+  or select a model artifact with enough sequence steps. For `[1d,2d)` at
+  15-minute frequency, the model needs 192 steps, not only the returned 96.
+- An unexpected output point count usually indicates frequency/target-time
+  drift or an incomplete runtime trajectory. Inspect `output_window.boundary`,
+  UTC `start_time`/`end_time`, `point_count`, and the model sampling interval.
+- `SUPERSEDED`: operate on the active deployment leaf.
+- `SUSPENDED`: resume only with explicit authorization and a reason.
+- `UNDEPLOYED`: create a governed deployment or rollback as supported; resume
+  is invalid.
+
+`reconcile-deployment` is a mutation because it may update Kubernetes
+resources and execute verification. Obtain authorization for the exact
+deployment before invoking it. Repeated successful reconcile should not rerun
+the same expensive smoke or append identical events.
+
+Use Kubernetes commands only when the user explicitly requests platform
+diagnostics or administration. Do not use them as an alternate deployment
+workflow, and never create, modify, or delete NetworkPolicy.

package/skills/model-lifecycle-management/SKILL.md

@@ -32,7 +32,12 @@ to discover fields. Preserve release/error evidence when deployment differs.
 - Before submission, show the exact JSON path, immutable identities, expected workload, and target API. Obtain explicit authorization. Retry and cancel require separate authorization for the exact Job.
 - Never use `latest` image tags. During development TrainerDefinition stays at `v1`; rebinding it follows the reviewed release procedure and requires confirming no active Job references the old definition.
 - A succeeded Job is insufficient evidence. Training requires a valid ModelArtifact and signature; evaluation requires immutable Result/Summary evidence and coverage; packaging requires `READY`, an immutable image digest, manifest, and passing test report.
+- Register a ModelArtifact through `register-model-artifact` only after a platform-generated PASS evaluation/result or summary has complete evidence; do not bypass the CLI with a direct registration HTTP request.
 - Do not register caller-computed evaluation results. Workers produce predictions, metrics, decisions, and summaries from frozen DatasetArtifact rowsets.
 - Keep credentials, registry secrets, Kubernetes details, and model binaries out of request JSON and reports.
 
 Report immutable input identities, returned run/job/package IDs, status, validation or gate failures, and the next authorized action. Never claim deployment; it is outside this Skill.
+
+Route deployment validation, creation, rollout state, approval, promotion,
+rollback, offline operations, and deployment-bound prediction to the separate
+`model-deployment-management` Skill.

package/skills/model-lifecycle-management/references/packaging.md

@@ -31,6 +31,69 @@ runtime. `feature_lookup` still requires artifact lineage
 signature. Do not infer either capability from training success.
 Omit rather than use an empty list to request default inline behavior.
 
+Register a validated artifact only after a platform-generated evaluation result or
+summary is `PASS`, `COMPLETE`, and belongs to the same artifact:
+
+```bash
+ml-platform --profile server register-model-artifact <artifact-id> \
+  --evaluation-id <evaluation-result-or-summary-id>
+```
+
+The API performs the complete registration gate. Do not replace this command with
+a direct `POST /models/{artifact_id}/register` request.
+
+For `feature_lookup`, the package request must bind both the exact retrieval
+contract and a sanitized real lookup test. They are immutable package identity,
+not namespace Secret/ConfigMap settings:
+
+```json
+{
+  "model_artifact_id": "replace_registered_model_artifact_id",
+  "package_version": "1.0.0",
+  "input_modes": ["inline", "feature_lookup"],
+  "feature_retrieval": {
+    "type": "data_platform",
+    "manifest": {
+      "project": "replace_project",
+      "dataset_id": "replace_inference_manifest",
+      "dataset_version": "v1",
+      "manifest_hash": "sha256:<registered-manifest-hash>"
+    },
+    "feature_set": {
+      "project": "replace_project",
+      "name": "replace_feature_set",
+      "version": "v1"
+    },
+    "contract_signature": "sha256:<artifact-contract-signature>",
+    "request": {
+      "mode": "row",
+      "input_adapter_hash": "sha256:<artifact-input-adapter-hash>",
+      "temporal_binding_hash": "sha256:<training-temporal-binding-hash>",
+      "entity_keys": {}
+    },
+    "freshness_policy": "strict"
+  },
+  "feature_lookup_test": {
+    "input_mode": "feature_lookup",
+    "cutoff_time": "2026-09-11T00:00:00+08:00",
+    "horizon": [{"duration": "PT10M"}],
+    "entity": {}
+  }
+}
+```
+
+The inference DatasetManifest must exist with the supplied hash and use
+`mode=inference`. Its FeatureSet, stable contract signature, adapter/temporal
+hashes, entity keys, and row/context parameters must match the ModelArtifact.
+The package-bound signature remains the Serving binding identity. Online fetch
+responses may also report an independently resolved execution signature for
+diagnostics; database principals are excluded from new stable signatures, so
+credential rotation or distinct read-only workload accounts do not alter the
+model data contract.
+The test cutoff must include a timezone. A legacy feature-lookup package missing
+either package-bound field remains readable but cannot be rerun or deployed;
+create a new package version.
+
 Always query the exact package ID and version; states are `PACKAGING`, `READY`,
 `FAILED`, `CANCELLED`. A client timeout does not prove packaging failed. Retain the
 known identity and query it before attempting another creation.

package/skills/model-lifecycle-management/references/training-contracts.md

@@ -169,10 +169,12 @@ and `data.labels` must use that same output-label order:
 Every target and covariate must be a numeric Feature. `past_only` is available
 only through the cutoff. `known_future` must be represented in training by the
 forecast or schedule available at that historic cutoff, never by a later actual
-observation. At runtime callers provide one finite numeric value for every
-`known_future` column and every prediction step. The current contract validates
-the names, values, and row count; retain forecast issue/availability lineage in
-the upstream Feature data when revisions must be audited.
+observation. At runtime inline callers provide one finite numeric value for
+every `known_future` column and every model prediction step. In
+`feature_lookup` mode, callers provide none: Serving requests the declared
+Feature columns from the deployment-bound FeatureSet, and the platform selects
+future event rows whose revisions satisfy `available_at <= cutoff_time`.
+Retain forecast issue/availability lineage in the upstream Feature data.
 
 ## TrainingJobRequest and identity propagation
 
@@ -243,6 +245,16 @@ aligned horizons up to that maximum and returns the corresponding trajectory
 prefix; it rejects unaligned or longer horizons. A one-step model remains a
 scalar response.
 
+When a deployment must use cutoff at today's 00:00 and return only tomorrow's
+00:00 through the following 00:00, train for the complete interval from cutoff
+to the output end. At 15-minute frequency, deployment horizon `2d` means
+`prediction_length=192`; output window `[1d,2d)` then projects the second 96
+points. Chronos-2 known-future covariates must still contain all 192 rows. In
+`feature_lookup` mode the platform retrieves those rows even when the request
+overrides the returned window; inline requests must still carry them. A
+150-step artifact cannot serve this contract and must be retrained rather than
+having its response window stretched at deployment time.
+
 `gap_policy=fill` requires `padding_policy=edge` or `zero`; non-fill requires
 `padding_policy=none`. Filling changes model inputs and needs an intended
 business policy. `skip` and padding are not automatic fixes for rejected grids.

package/skills/model-lifecycle-management/references/training.md

@@ -93,15 +93,18 @@ the ordered target/output-label pairs; see [training-contracts.md](training-cont
 for the complete JSON fragment.
 
 Training records must contain the forecast that was available at each historic
-cutoff, not actual weather observed later. At online inference, pass one numeric
-row per forecast step using the Serving runtime, not `fetch-inference-data`:
+cutoff, not actual weather observed later. Invoke the Serving runtime, not
+`fetch-inference-data`:
 
 ```bash
 ml-platform predict-model /absolute/path/to/inference-request.json \
   --serving-url https://storage-forecast.example
 ```
 
-For `input_mode=inline`, put those rows in `inputs.future_covariates`. For
-`input_mode=feature_lookup`, put them in the top-level `future_covariates`.
-Their column names must exactly match the adapter's `known_future` covariates,
-and their count must equal the trained `prediction_length`.
+For `input_mode=inline`, put one row per trained prediction step in
+`inputs.future_covariates`; names must exactly match the adapter's
+`known_future` columns. For `input_mode=feature_lookup`, do not send future
+rows. Serving asks the platform for the declared columns and full
+`prediction_length`; the Features must belong to the deployment-bound
+FeatureSet, and their source revisions must satisfy
+`available_at <= cutoff_time` even though their event times are later.