@zhuoyuezs/ml-platform 0.1.9 → 0.1.11

package/README.md

@@ -1,6 +1,6 @@
 # ML Platform
 
-ML Platform 的 JavaScript CLI 和 `feature-management` Agent Skill。CLI 与 Skill 由同一个 npm
+ML Platform 的 JavaScript CLI、`feature-management` 与 `model-lifecycle-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
@@ -17,8 +17,8 @@ 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 Skill 安装到
-`~/.codex/skills/feature-management/`，供当前用户的 Codex 使用。安装后重启终端和 Agent 会话。
+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
+`~/.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：
 
@@ -103,6 +103,10 @@ ml-platform --help
 受控构建，不要求本地存在 Catalog。`apply`、`resolve-manifest` 和带本地 JSON 路径的
 `build-dataset` 用于新契约的编写和发布流程。
 
+Parameter JSON 可同时声明离线 `source` 和可选的实时 `online_source`。通过
+`ml-platform add-parameter <spec.json>` 或 Catalog `apply` 正常发布即可；客户端原样提交
+字段，离线构建使用 `source`，`fetch-inference-data` 使用 `online_source`。
+
 ## 迁移旧 Skill
 
 目标目录存在旧 Python Skill、实验 symlink 或其他未托管 Skill 时，确认后显式备份：

package/checksums.json

@@ -2,28 +2,28 @@
   "files": [
     {
       "path": "runtime/business-client/README.md",
-      "sha256": "sha256:02e6561b80cbef3cce2fd690750d863d8afb9c45bf5ccfcce3e3e2b082ff0761",
-      "size_bytes": 3780
+      "sha256": "sha256:b823c0cafc47478fa1297c7c2e61e15b537db155586dd012b40d678d788c3ce3",
+      "size_bytes": 5158
     },
     {
       "path": "runtime/business-client/package-lock.json",
-      "sha256": "sha256:d5e7102ec51df3192ae901908f36cf0380aba5a0421b9824877403ab034d6740",
+      "sha256": "sha256:b460e297ced5aa90210bc73c7ee2e8b73940674b4e1f62ffe3915fe4024eb089",
       "size_bytes": 381
     },
     {
       "path": "runtime/business-client/package.json",
-      "sha256": "sha256:9e815ffe0487eb216447f4a4b31d248fe37392e08f0fd4bae2aba9c4f4285547",
+      "sha256": "sha256:a60a2018c8714b09fcf386601cee82baa9384648526a36f95fbd7480cabcfef5",
       "size_bytes": 501
     },
     {
       "path": "runtime/business-client/src/catalog.js",
-      "sha256": "sha256:8dcc4c4baff7f292a6fb09d6f82693b5ce7da16b9f0a9e03b22e1f18e893240f",
-      "size_bytes": 53659
+      "sha256": "sha256:717df5f630e85da89fe8b315d6dabc6549af7984f93a6167b494e52ee3697803",
+      "size_bytes": 53795
     },
     {
       "path": "runtime/business-client/src/cli.js",
-      "sha256": "sha256:6ee4ed29a9df9f462502ccc0e10d5d9109bf00b52b8c27c9808730c7d8520592",
-      "size_bytes": 44480
+      "sha256": "sha256:5a31e407dba79205e56d2ed1e74260213d0b89c6bb5a665ec0f52e92927ebcaf",
+      "size_bytes": 44897
     },
     {
       "path": "runtime/business-client/src/config.js",
@@ -37,8 +37,8 @@
     },
     {
       "path": "skills/feature-management/SKILL.md",
-      "sha256": "sha256:5b3ce57d84f11188579227dd42617075f87534ac113a3f37b3543c85c0ff0b3d",
-      "size_bytes": 31114
+      "sha256": "sha256:e5b927611cd401bc5e6160a2d358369efd0c4f65d3fe5cbf21bbab9adb6bd978",
+      "size_bytes": 31991
     },
     {
       "path": "skills/feature-management/agents/openai.yaml",
@@ -97,8 +97,8 @@
     },
     {
       "path": "skills/feature-management/references/contracts.md",
-      "sha256": "sha256:d9ff95edb3ac50cb590df640c5034d45ddd5b758431a4ae2e69eb807bb659314",
-      "size_bytes": 28449
+      "sha256": "sha256:4e10ba13399de17561a10b1dead4e2fcc57e0962cf81eba5131d12d66c4ce013",
+      "size_bytes": 29965
     },
     {
       "path": "skills/feature-management/references/operator-authoring.md",
@@ -117,8 +117,8 @@
     },
     {
       "path": "skills/model-lifecycle-management/SKILL.md",
-      "sha256": "sha256:98437e919b6d76822b2a3521713400b2c73ae42fbe5086f2e0b864aadd6cbf55",
-      "size_bytes": 3464
+      "sha256": "sha256:faa4f8c0dbae7d1c1e2ca94499ce9b13cc8b52e8229154c05ac6282ea5e47788",
+      "size_bytes": 3610
     },
     {
       "path": "skills/model-lifecycle-management/agents/openai.yaml",
@@ -137,18 +137,18 @@
     },
     {
       "path": "skills/model-lifecycle-management/references/packaging.md",
-      "sha256": "sha256:8435283061d740e57aafc90fa402cd89adcd8e2e957805b3d435ea4ba1f8721c",
-      "size_bytes": 3142
+      "sha256": "sha256:0992f2f9104cb0d65a6325756108211d3f94856e05eca12640e9792c07e1108f",
+      "size_bytes": 3250
     },
     {
       "path": "skills/model-lifecycle-management/references/training-contracts.md",
-      "sha256": "sha256:1d1fa51ef7deb6abf178cf019eb69e8ee6a93ecb018029780f01a33696473dd6",
-      "size_bytes": 8229
+      "sha256": "sha256:b6d72f37adf40c7d3832dbe46645cd416485411dcd48ea7b93f6072d458922c9",
+      "size_bytes": 11825
     },
     {
       "path": "skills/model-lifecycle-management/references/training.md",
-      "sha256": "sha256:370d5e663292ce1bdcf0eb88c133c07a77f57dd31e21748906c92aa3407c4e95",
-      "size_bytes": 4516
+      "sha256": "sha256:f7bb40919da78b99cad13d4f1f05fef136552409df5ba6e2838e1e6b4a2435a5",
+      "size_bytes": 5508
     }
   ],
   "schema_version": "data_platform.ml_platform_checksums/v1"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zhuoyuezs/ml-platform",
-  "version": "0.1.9",
+  "version": "0.1.11",
   "description": "ML 数据平台 business CLI 与 Agent Skills 管理工具",
   "bin": {
     "ml-platform": "scripts/main.js"
@@ -19,7 +19,8 @@
   "scripts": {
     "test": "node --test test/*.test.js",
     "api:check": "node ../../scripts/check_api_contract.js",
-    "pack": "npm pack"
+    "pack": "npm pack",
+    "prepublishOnly": "node scripts/verify-release-package.js"
   },
   "engines": {
     "node": ">=18"

package/release.json

@@ -13,8 +13,8 @@
     "entrypoint": "src/cli.js",
     "name": "ml-platform",
     "path": "runtime/business-client",
-    "sha256": "sha256:fb2ebe76ddba0ccfcd052741e8e290e25fc2897df7656d48267e82fe6be06d1c",
-    "version": "0.7.0"
+    "sha256": "sha256:95558e7845b8a177c6090d4a960640469db658e88d9147ccc9170eadf7ca3410",
+    "version": "0.7.1"
   },
   "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": "0.1.9",
+  "release_version": "0.1.11",
   "runtime_requirements": {
     "node": ">=18",
     "os": [
@@ -37,16 +37,16 @@
   "skills": {
     "feature-management": {
       "path": "skills/feature-management",
-      "requires_cli": ">=0.7.0 <0.8.0",
-      "revision": "0.1.9",
-      "sha256": "sha256:82a49a3f636d4e1a4b5adf28cbcab533168e3aa5fc3014f096149055e7745624"
+      "requires_cli": ">=0.7.1 <0.8.0",
+      "revision": "0.1.11",
+      "sha256": "sha256:182b6f2df4fe58c946e482dfaf416c7d5b78c6fde9c8c8d7293103246618cb4e"
     },
     "model-lifecycle-management": {
       "path": "skills/model-lifecycle-management",
-      "requires_cli": ">=0.7.0 <0.8.0",
-      "revision": "0.1.9",
-      "sha256": "sha256:b84505ee632967d019e97f89292c0041f8b722bd320e7d68d170e41131edaff7"
+      "requires_cli": ">=0.7.1 <0.8.0",
+      "revision": "0.1.11",
+      "sha256": "sha256:61ba17f955b36d1c2e8329a10454d040ee7eca5afdb90ff0914d3f85b6e2d203"
     }
   },
-  "source_commit": "5f855cd27f7951b739424704ea7dc15d714aeb5d"
+  "source_commit": "374d9935e9a7267ae6485be140a8d99b2af57b44"
 }

package/runtime/business-client/README.md

@@ -27,6 +27,10 @@ ml-platform build-registered-dataset <dataset_id> <dataset_version> \
 `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。
 
+同一 Parameter 需要从历史库训练、从实时库取数时，在 Parameter JSON 同时填写 `source` 与
+可选的 `online_source`，然后照常使用 `add-parameter` 或 Catalog `apply`；CLI 原样透传该
+字段，不需要额外参数。离线构建读取 `source`，实时因果读取自动选择 `online_source`。
+
 Registry 资产详情和版本发现使用与 `dev` OpenAPI 一致的命令：
 
 ```bash
@@ -44,6 +48,34 @@ 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
+`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
+同样提供整天的所有未来步：
+
+```json
+{
+  "input_mode": "inline",
+  "cutoff_time": "2026-09-09T00:00:00+08:00",
+  "horizon": [{"duration": "PT30M"}],
+  "inputs": {
+    "timestamps": ["...", "2026-09-09T00:00:00+08:00"],
+    "values": [{"storage_soc": 51.2, "storage_power": 4.0}, {"storage_soc": 52.0, "storage_power": 3.8}],
+    "future_covariates": [
+      {"weather_temperature": 28.0, "peak_valley": 1},
+      {"weather_temperature": 27.5, "peak_valley": 1}
+    ]
+  }
+}
+```
+
+```bash
+ml-platform predict-model /absolute/path/to/request.json \
+  --serving-url https://storage-forecast.example
+```
+
+训练历史必须使用各个历史 cutoff 当时可取得的天气预报，不得以事后实测天气替代。
+
 Parameter、Feature、FeatureSet 的删除命令执行软删除，并支持带 actor、reason 和幂等键的
 恢复；调用方应在任何变更前查询直接与传递引用。
 

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

@@ -1,12 +1,12 @@
 {
   "name": "@zhuoyuezs/ml-platform-business-client",
-  "version": "0.7.0",
+  "version": "0.7.1",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@zhuoyuezs/ml-platform-business-client",
-      "version": "0.7.0",
+      "version": "0.7.1",
       "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": "0.7.0",
+  "version": "0.7.1",
   "private": true,
   "description": "JavaScript implementation of the ML Platform business client",
   "bin": {

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

@@ -103,7 +103,7 @@ function normalizeRelationshipRule(value, relationshipId) { const label = `param
 function normalizeParameterRelationship(value, parentProject) { exactKeys(value, ["id", "inputs", "alignment", "rules", "description"], ["id", "inputs", "rules"], "parameter relationship"); const id = identifier(value.id, "parameter relationship id"); const inputs = objectList(value.inputs, "parameter relationship inputs").map((item) => { exactKeys(item, ["alias", "parameter", "version", "project"], ["alias", "parameter", "version"], "parameter relationship input"); validateReferenceProject(item, parentProject, "parameter relationship input"); return { alias: identifier(item.alias, "parameter relationship input alias"), parameter: identifier(item.parameter, "parameter relationship input parameter"), version: identifier(item.version, "parameter relationship input version"), project: parentProject }; }); if (inputs.length < 2) throw new Error("parameter relationship requires at least two inputs"); unique(inputs.map((item) => item.alias), "parameter relationship input aliases"); unique(inputs.map(parameterReferenceKey), "parameter relationship inputs"); const alignmentValue = object(defaultIfAbsent(value, "alignment", {}), "parameter relationship alignment"); exactKeys(alignmentValue, ["method", "tolerance"], [], "parameter relationship alignment"); const tolerance = optionalDuration(alignmentValue.tolerance, "parameter relationship alignment tolerance"); if (tolerance != null && durationSeconds(tolerance) === 0) throw new Error("parameter relationship alignment tolerance must be positive"); const alignment = { method: choice(defaultIfAbsent(alignmentValue, "method", "last_before_or_at"), ["exact", "last_before_or_at"], "parameter relationship alignment method"), tolerance }; const rules = objectList(value.rules, "parameter relationship rules").map((item) => normalizeRelationshipRule(item, id)); if (!rules.length) throw new Error("parameter relationship rules must not be empty"); unique(rules.map((item) => item.id), `parameter relationship ${id} rule ids`); const aliases = new Set(inputs.map((item) => item.alias)); for (const rule of rules) { if (rule.stage !== "aligned_grid") throw new Error(`parameter relationship ${id} rule ${rule.id} must use stage=aligned_grid`); const missing = relationshipAliases(rule.check).filter((item) => !aliases.has(item)); if (missing.length) throw new Error(`parameter relationship ${id} rule ${rule.id} references undeclared aliases: ${missing.join(", ")}`); } return { id, inputs, alignment, rules, description: pythonString(defaultIfAbsent(value, "description", "")) }; }
 function normalizeDataset(payload) {
   object(payload, "dataset"); payload = Object.fromEntries(Object.entries(payload).filter(([name]) => name !== "deleted_at"));
-  exactKeys(payload, ["schema_version", "project", "dataset_id", "dataset_version", "mode", "read_policy", "snapshot_id", "as_of", "time_range", "parameters", "feature_set", "preprocess", "output", "source_read", "realtime_fetch", "abnormal_windows", "prediction", "target", "parameter_relationships", "parameter_missing_policies", "rowset", "rowset_splits", "endpoint_policy"], ["dataset_id", "time_range", "feature_set"], "dataset");
+  exactKeys(payload, ["schema_version", "project", "dataset_id", "dataset_version", "mode", "read_policy", "snapshot_id", "as_of", "time_range", "parameters", "feature_set", "preprocess", "output", "source_read", "realtime_fetch", "abnormal_windows", "prediction", "target", "label_materializations", "parameter_relationships", "parameter_missing_policies", "rowset", "rowset_splits", "endpoint_policy"], ["dataset_id", "time_range", "feature_set"], "dataset");
   const parentProject = identifier(defaultIfAbsent(payload, "project", "default"), "dataset project");
   const readPolicy = choice(defaultIfAbsent(payload, "read_policy", "snapshot"), ["snapshot", "as_of", "latest"], "dataset read_policy");
   const snapshotId = payload.snapshot_id == null ? null : identifier(payload.snapshot_id, "dataset snapshot_id");
@@ -122,6 +122,7 @@ function normalizeDataset(payload) {
   if (missingPolicies.length) output.parameter_missing_policies = missingPolicies;
   for (const name of ["source_read", "realtime_fetch", "abnormal_windows", "prediction", "rowset", "rowset_splits", "endpoint_policy"]) if (payload[name] != null) output[name] = payload[name];
   if (payload.target != null) output.target = normalizeTarget(payload.target, parentProject);
+  if (payload.label_materializations != null) output.label_materializations = payload.label_materializations;
   const relationships = objectList(defaultIfAbsent(payload, "parameter_relationships", []), "dataset parameter_relationships").map((item) => normalizeParameterRelationship(item, parentProject)); unique(relationships.map((item) => item.id), "dataset parameter_relationship ids"); if (relationships.length) output.parameter_relationships = relationships;
   return output;
 }

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

@@ -28,7 +28,7 @@ const BUSINESS_COMMANDS = new Set([
   "add-parameter", "add-feature", "add-feature-set", "add-operator", "publish-operator",
   "add-dataset", "update-dataset", "delete-dataset", "list-dataset-artifacts",
   "register-operator", "show-operator-specs", "delete-operator", "run-operator",
-  "get-job", "wait-job", "retry-job", "cancel-job", "get-dataset-artifact", "download-dataset-artifact", "fetch-inference-data", "fetch-inference-context", "apply",
+  "get-job", "wait-job", "retry-job", "cancel-job", "get-dataset-artifact", "download-dataset-artifact", "fetch-inference-data", "fetch-inference-context", "predict-model", "apply",
   "register-trainer", "list-trainer-definitions", "get-trainer-definition", "register-training-run", "validate-training-run", "submit-training-job", "get-training-job", "retry-training-job", "cancel-training-job",
   "register-evaluation-config", "get-evaluation-config", "register-metric-definition", "get-metric-definition",
   "register-executable-package", "verify-executable-package", "publish-executable-package",
@@ -105,6 +105,7 @@ const COMMAND_USAGE = {
   "download-dataset-artifact": "download-dataset-artifact DATASET_ID MANIFEST_HASH [--project PROJECT] --out-dir PATH [--file PATH] [--force]",
   "fetch-inference-data": "fetch-inference-data MANIFEST --cutoff-time TIMESTAMP [--max-workers N] [--max-source-lag-hours HOURS] [--no-validate-freshness] [--allow-missing]",
   "fetch-inference-context": "fetch-inference-context REQUEST_JSON",
+  "predict-model": "predict-model REQUEST_JSON --serving-url URL",
   "register-trainer": "register-trainer DEFINITION_JSON",
   "list-trainer-definitions": "list-trainer-definitions",
   "get-trainer-definition": "get-trainer-definition NAME VERSION [--project PROJECT]",
@@ -451,6 +452,10 @@ async function runBusinessCli(argv) {
     const request = { manifest: readDataset(manifest), cutoff_time: cutoffTime }; if (Object.keys(realtime).length) request.realtime_fetch = realtime; result = await client(options).post("/inference-data/fetch", request);
   } else if (options.command === "fetch-inference-context") {
     result = await client(options).post("/inference-data/context", readSpec(positional(rest, "request JSON")));
+  } else if (options.command === "predict-model") {
+    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 === "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))) {

package/scripts/verify-release-package.js

@@ -0,0 +1,28 @@
+#!/usr/bin/env node
+"use strict";
+
+const path = require("node:path");
+
+const { verifyPackage } = require("./lib");
+
+function verifyReleasePackage(packageRoot = path.resolve(__dirname, "..")) {
+  try {
+    return verifyPackage(packageRoot);
+  } catch (error) {
+    throw new Error(
+      `拒绝发布未验证的 ML Platform 包：${error.message}。`
+      + " 请只发布 build_ml_platform_release.py 生成的 .tgz 归档。",
+    );
+  }
+}
+
+if (require.main === module) {
+  try {
+    verifyReleasePackage();
+  } catch (error) {
+    process.stderr.write(`${error.message}\n`);
+    process.exitCode = 1;
+  }
+}
+
+module.exports = { verifyReleasePackage };

package/skills/feature-management/SKILL.md

@@ -258,6 +258,15 @@ Parameter name.
 
 Create one stable, versioned source contract for each independently readable value. Keep credentials out of JSON. Use only supported direct adapters (`postgresql_direct`, `mysql_direct`, `sqlserver_direct`, and `influxdb_direct`) and explicit source identity. The server supplies database DSNs from managed Secrets; never add them to CLI arguments or Catalog JSON. Put source data validity in `quality_rules`; put dataset-specific transformations in preprocess Operators.
 
+When historical and live systems expose the same value, keep one Parameter and add an optional
+`online_source`. Batch/as-of reads always use `source`; realtime causal reads use `online_source`
+when present and otherwise fall back to `source`. Both sources must produce the same standard long
+table and preserve value, unit, event-time, frequency, quality, availability, and missingness
+semantics. Do not duplicate the Feature or Operator. Realtime lineage identifies the selected source
+and includes the offline source; the response is not replayable through the offline manifest when an
+override is active. The executable JSON shape is in
+[references/contracts.md](references/contracts.md).
+
 When the approved source contract supplies SQL, copy its query text and named
 parameters exactly into `source.sql`/`source.params`. Do not retype, beautify,
 rename an identifier, change a join key, or "simplify" an approved query while
@@ -404,9 +413,11 @@ Confirm the cutoff and prediction horizon. The command must call
 publish a DatasetArtifact.
 
 Verify that the response reports the requested cutoff and derived target time,
-`uses_post_cutoff_data=false`, `contract.manifest_hash`, a replayable `as_of`
-manifest, and per-Parameter freshness evidence. Report tolerated missing or
-gap-filled inputs as degraded freshness, not as a normal read.
+`uses_post_cutoff_data=false`, `contract.manifest_hash`, and per-Parameter
+freshness evidence. A response without `online_source` should include a
+replayable `as_of` manifest; when an online source override is active, `replay.supported`
+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.
 
 ## Verify The Artifact
 

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

@@ -61,6 +61,7 @@ Required shape:
   "unit": "kPa",
   "expected_frequency": "5s",
   "source": {},
+  "online_source": null,
   "time_semantics": {
     "event_time_field": "event_time",
     "ingested_at_field": "ingested_at",
@@ -92,6 +93,38 @@ Required shape:
 }
 ```
 
+`online_source` 是可选的实时来源覆盖。离线 Dataset build、`snapshot`、`as_of` 和历史回放
+只读取 `source`；实时因果读取在声明时读取 `online_source`，否则回退到 `source`。两者必须
+表达同一个 Parameter，产生相同语义的标准长表（`timestamp, metric_name, value, unit`），
+并保持值、单位、事件时间、频率、质量和可用时间语义一致。它们可以使用不同的受支持
+adapter；凭据不进入 JSON。Feature/Operator/FeatureSet 不需要复制一份。
+
+例如：
+
+```json
+{
+  "source": {
+    "adapter": "postgresql_direct",
+    "schema": "process_data",
+    "table": "pressure_history",
+    "time_column": "event_time",
+    "value_column": "pressure",
+    "metric_name": "hot_blast_pressure"
+  },
+  "online_source": {
+    "adapter": "influxdb_direct",
+    "measurement": "blast_furnace_live",
+    "field": "pressure",
+    "metric_name": "hot_blast_pressure"
+  }
+}
+```
+
+实时 lineage 会标出 `source_role=online_source` 并保留 `offline_source`。`online_source`
+参与 Parameter hash；因此变更它必须发布新的 Parameter 版本及受影响的下游依赖。使用在线
+覆盖时，响应中的 replay 会明确标记为不支持，因为离线 manifest 只会重新读取 `source`，
+不能伪造线上读法的回放。
+
 Supported source modes are `direct_column`, `sql`, and `derived`. Prefer `direct_column` for ordinary business parameters.
 
 For `sql`, `source.sql` must contain one statement without a trailing semicolon,
@@ -415,8 +448,10 @@ writing parquet, or publishing a DatasetArtifact. It derives
 `target_time = cutoff_time + prediction.horizon`, filters every source to
 `timestamp <= cutoff_time`, and derives each Parameter read window from the
 selected Operators' `input_schema.history_requirements`. The response carries
-`contract.manifest_hash` and a replayable `read_policy=as_of` manifest instead
-of an artifact.
+`contract.manifest_hash`; it also carries a replayable `read_policy=as_of` manifest
+unless a Parameter's `online_source` override is active, in which case
+`replay.supported=false` is returned because an offline manifest would read the
+offline `source` instead.
 
 Keep the default strict: `allow_missing=false`. Tail-edge and bounded internal
 gap-fill tolerance may return a row, but the response must mark

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

@@ -1,11 +1,11 @@
 ---
 name: model-lifecycle-management
-description: Manage ITSMP model training, evaluation, and immutable model packaging through the deployed platform API. Use when a user asks to list or inspect registered TrainerDefinitions, register a trainer, validate or submit a training run, inspect/retry/cancel training jobs, configure or run governed evaluation, inspect evaluation evidence, validate a ModelArtifact, or create and inspect a ModelPackage. Do not use for feature catalog authoring, dataset builds, model deployment, Kubernetes administration, or image release engineering.
+description: Manage ITSMP model training, evaluation, immutable model packaging, and contract-bound inference requests. Use when a user asks to list or inspect registered TrainerDefinitions, register a trainer, validate or submit a training run, inspect/retry/cancel training jobs, configure or run governed evaluation, inspect evaluation evidence, validate a ModelArtifact, create and inspect a ModelPackage, or invoke a deployed model. Do not use for feature catalog authoring, dataset builds, model deployment, Kubernetes administration, or image release engineering.
 ---
 
 # Model Lifecycle Management
 
-Use the `ml-platform` executable installed with the same npm release as this Skill. Verify `ml-platform version`, `ml-platform show-config`, and server health before API operations. Never use a platform source checkout or a second client runtime.
+Use the `ml-platform` executable installed with the same npm release as this Skill. Verify `ml-platform version`, `ml-platform show-config`, and server health before control-plane API operations. Use `predict-model` with the deployed model's Serving URL; it is distinct from the control-plane API URL. Never use a platform source checkout or a second client runtime.
 
 ```text
 DatasetArtifact -> TrainingRun -> TrainingJob -> ModelArtifact

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

@@ -25,8 +25,10 @@ runtime contract, not repeated package versions. Unknown request keys return 422
 fields. Inspect the returned resolved package.
 
 `input_modes` must be unique, include `inline`, and may additionally include
-`feature_lookup` only when supported by the selected runtime and artifact lineage
-has `online_eligible=true`. Do not infer online eligibility from training success.
+`feature_lookup` or `tabular_forecast` only when supported by the selected
+runtime. `feature_lookup` still requires artifact lineage
+`online_eligible=true`; `tabular_forecast` requires a direct tabular model
+signature. Do not infer either capability from training success.
 Omit rather than use an empty list to request default inline behavior.
 
 Always query the exact package ID and version; states are `PACKAGING`, `READY`,

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

@@ -68,6 +68,66 @@ Non-temporal regression can omit it. Do not change task kind just to evade check
 `series_keys` must identify entity-key columns when multiple series coexist;
 an empty list treats the entire artifact as one series.
 
+## Direct tabular forecasting with known-future features
+
+Tree and other matrix-based models can forecast a variable target grid without
+pretending that future observations are available. Set the tabular adapter to
+`forecast_mode="direct"`. The dataset's `TemporalBinding.prediction_time_column`
+is the historical `cutoff_time`; `label_time_column` is the row's future
+`target_time`. Each row is one `(cutoff_time, target_time)` pair. Values such as
+weather forecasts, schedules, calendar fields and `lead_hours` are ordinary
+feature columns; list the forecast columns in `known_future_columns`.
+
+```json
+{
+  "input_adapter": {
+    "kind": "tabular",
+    "forecast_mode": "direct",
+    "known_future_columns": ["weather_temperature", "weather_wind"]
+  },
+  "data": {
+    "temporal": {
+      "prediction_time_column": "cutoff_time",
+      "label_time_column": "target_time",
+      "frequency": "1h",
+      "horizon": "240h",
+      "series_keys": ["market"]
+    }
+  }
+}
+```
+
+At serving time use `input_mode="tabular_forecast"` and send one row per
+target. `target_time` must be later than the cutoff and aligned to the model
+frequency. When known-future columns are declared, every row must include
+`available_at <= cutoff_time`; this prevents using a forecast revision that did
+not exist when the prediction was made.
+
+```json
+{
+  "input_mode": "tabular_forecast",
+  "cutoff_time": "2026-09-09T10:30:00+08:00",
+  "targets": [
+    {
+      "target_time": "2026-09-10T00:00:00+08:00",
+      "available_at": "2026-09-09T08:00:00+08:00",
+      "values": {
+        "weather_temperature": 21.3,
+        "weather_wind": 2.8,
+        "lead_hours": 13.5
+      }
+    }
+  ]
+}
+```
+
+The serving runtime calls the same plugin `predict()` with a two-dimensional
+batch and returns one result carrying each `target_time`. XGBoost and LightGBM
+already implement this batch matrix operation; the platform adapter owns the
+cutoff, target-grid and availability checks. Training must use historical
+forecast snapshots selected by `available_at <= cutoff_time`, never later actual
+weather observations.
+
 Inspect `data_schema.json`, `resolved_manifest.json` and the feature Parquet via
 `download-dataset-artifact DATASET_ID MANIFEST_HASH --project PROJECT --out-dir DIR`
 (repeat `--file` for selective downloads). The FeatureSet reference must match
@@ -79,6 +139,41 @@ in `required_inputs` and at least one `online_inputs` entry, which must be a
 subset; it must have an inverse. Do not use target transforms to fix incorrect
 dataset label clocks or to invent unsupported transforms.
 
+## Chronos-2 v2 targets and covariates
+
+Use the v2 adapter when Chronos-2 should jointly forecast multiple storage
+channels or consume covariates. The target and output-label arrays are ordered,
+and `data.labels` must use that same output-label order:
+
+```json
+{
+  "labels": ["future_storage_soc", "future_storage_power"],
+  "forecast_target_adapter": {
+    "schema_version": "ml_data_platform.forecast_target_adapter/v2",
+    "kind": "continuous_target_series_with_covariates",
+    "targets": [
+      {"input_column": "storage_soc", "output_label_column": "future_storage_soc"},
+      {"input_column": "storage_power", "output_label_column": "future_storage_power"}
+    ],
+    "covariates": [
+      {"column": "recent_load", "role": "past_only"},
+      {"column": "weather_temperature", "role": "known_future"},
+      {"column": "peak_valley", "role": "known_future"}
+    ],
+    "target_gap_policy": "reject",
+    "point_output": {"statistic": "median", "step": "horizon"}
+  }
+}
+```
+
+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.
+
 ## TrainingJobRequest and identity propagation
 
 After validation and authorized registration, use the returned `spec_hash`

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

@@ -13,16 +13,17 @@ FeatureSet order and evaluation runtime before the first Job. See
   for device`, preserve the rejected identity, device and release ID. Use `get-model-runtime-release` to find the
   approved binding; if the deployed release lacks discovery, request its contract;
   do not try similar names or re-register trainers to bypass approval.
-- The current platform Chronos2Plugin requires `[sample, context, 1]`, exactly
-  one channel. This is a platform adapter restriction, not a universal statement
-  about Chronos-2. Six-feature sequences are incompatible. Do not silently drop
-  business inputs: select a suitable trainer or agree a univariate contract.
+- Chronos-2 supports the legacy v1 single-target adapter and the v2
+  covariate-aware adapter. v2 jointly forecasts ordered target channels and
+  accepts numeric `past_only` and `known_future` covariates. Do not silently
+  drop business inputs: declare every target and covariate in the adapter.
 - Keep labels outside FeatureSet order, verify temporal horizon against
   `label_time - prediction_time`, and validate the complete TrainingRun before
   registration/submission. Validation may not catch undeclared plugin limits or
   missing dependencies in the eventual Worker.
-- For a continuous future trajectory, use TrainingRun v3 with
-  `forecast_target_adapter.kind=continuous_target_series`. Its fixed maximum
+- For a continuous future trajectory, use TrainingRun v3. Use
+  `continuous_target_series` for legacy v1 or
+  `continuous_target_series_with_covariates` for v2. Its fixed maximum
   prediction length comes from temporal horizon divided by frequency; do not
   put `prediction_length` in trainer-specific parameters.
 
@@ -83,3 +84,24 @@ ml-platform --profile server cancel-training-job <job-id>
 ```
 
 Retry and cancellation require exact-Job authorization. Never change a run contract while retrying; create a new immutable run when semantic inputs change.
+
+## Chronos-2 future-known inputs
+
+For storage forecasting, tomorrow's weather forecast and peak/valley schedule
+are `known_future` covariates. The v2 adapter binds their Feature columns and
+the ordered target/output-label pairs; see [training-contracts.md](training-contracts.md)
+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`:
+
+```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`.