dsh-runnpu 0.1.0

package/compute-mode/preset/skills/runnpu/examples/01-dev-environment.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# 场景：给用户建一个带 Jupyter + 持久工作盘的开发环境，交付入口 URL。
+set -euo pipefail
+
+PROJECT=main
+NAME=nb-demo
+
+# 1. 建负载：1 张整卡 + 50Gi 工作盘 + Jupyter；昇腾镜像 root 运行所以 baseline
+runnpu workload create --project "$PROJECT" --name "$NAME" --type workspace \
+  --image m.daocloud.io/quay.io/ascend/vllm-ascend:v0.22.1rc1 \
+  --xpu 1 --security baseline --jupyter \
+  --workspace-pool shared-cephfs --workspace-gib 50 \
+  --wait
+
+# 2. 拿入口（tools[].url 平台登录态直达；没 url 说明还没 ready 或没权限）
+runnpu workload get "$PROJECT/$NAME" | python3 -c '
+import json,sys
+d=json.load(sys.stdin)
+for t in d["summary"].get("tools") or []:
+    print(t["name"], "ready:", t["ready"], "url:", t.get("url","-"))'
+
+# 3.（可选）验证卡真的可用
+runnpu workload exec "$PROJECT/$NAME" <<'SH'
+npu-smi info | head -8
+exit
+SH

package/compute-mode/preset/skills/runnpu/examples/02-training-job.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# 场景：提交一个单机训练任务，带重试与超时，完成后自动清理；轮询到终态并交付结果。
+set -euo pipefail
+
+PROJECT=main
+NAME=train-demo
+
+runnpu workload create --project "$PROJECT" --name "$NAME" --type training-job \
+  --image m.daocloud.io/quay.io/ascend/vllm-ascend:v0.22.1rc1 --xpu 1 --security baseline \
+  --mount datavolume:/data,name=dv-main-trainset,ro \
+  --mount newvolume:/output,pool=shared-cephfs,size=20Gi,persistency=project \
+  --command bash --command -c --arg "python /data/train.py --out /output" \
+  --backoff-limit 2 --active-deadline-seconds 7200 --ttl-after-finished 86400
+
+# 训练可能超过 --wait 的 10 分钟上限，所以自己轮询
+while :; do
+  st=$(runnpu workload get "$PROJECT/$NAME" | python3 -c 'import json,sys; print(json.load(sys.stdin)["summary"]["status"])')
+  case "$st" in completed|failed|rejected|stopped) break;; esac
+  sleep 30
+done
+echo "终态：$st"
+runnpu workload get "$PROJECT/$NAME" | python3 -c '
+import json,sys; d=json.load(sys.stdin)
+p=d.get("progress") or {}
+print("succeeded:",p.get("succeeded"),"failed:",p.get("failed"),"retries:",p.get("retries_used"))'
+[ "$st" = failed ] && runnpu workload logs "$PROJECT/$NAME" --tail 100

package/compute-mode/preset/skills/runnpu/examples/03-inference-service.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+# 场景：部署 vLLM 推理服务（权重在项目 PVC 里），nodePort 对外，实测端点可用。
+set -euo pipefail
+
+PROJECT=main
+NAME=qwen-svc
+# 权重 PVC 必须带 npu.run/project-id=<项目slug> 标签（否则挂载被拒为"不存在"）
+WEIGHTS_PVC=model-weights
+
+runnpu workload create --project "$PROJECT" --name "$NAME" --type inference \
+  --image m.daocloud.io/quay.io/ascend/vllm-ascend:v0.22.1rc1 --xpu 1 --security baseline \
+  --mount "pvc:/models,name=$WEIGHTS_PVC,ro" \
+  --command bash --command -c \
+  --arg "python -m vllm.entrypoints.openai.api_server --model /models/qwen --served-model-name qwen --port 8000" \
+  --port http:8000,expose=nodePort --replicas 1 --wait
+
+# nodePort 在工具观测里；ready 可能误报 false（run-npu#553），以实测为准
+NP=$(runnpu workload get "$PROJECT/$NAME" | python3 -c '
+import json,sys
+for t in json.load(sys.stdin)["summary"].get("tools") or []:
+    if t.get("node_port"): print(t["node_port"]); break')
+NODE_IP=$(runnpu node list | python3 -c 'import json,sys; print(json.load(sys.stdin)["items"][0]["ip"])' 2>/dev/null || echo "<节点IP>")
+echo "端点：http://$NODE_IP:$NP/v1"
+curl -s --max-time 10 "http://$NODE_IP:$NP/v1/models" | head -c 300 || echo "（模型加载中，稍后重试）"

package/compute-mode/preset/skills/runnpu/examples/04-fractional-vnpu.sh

@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+# 场景：分卡（vNPU）——单张卡切 25% 算力 + 16GiB 显存。
+# 只支持单卡切分：--xpu 必须为 1（ADR-0030 §6.3 起 CRD 不支持多卡分卡）。
+set -euo pipefail
+
+runnpu workload create --project main --name frac-demo --type workspace \
+  --image m.daocloud.io/quay.io/ascend/vllm-ascend:v0.22.1rc1 --security baseline \
+  --xpu 1 --xpu-core 25 --xpu-memory-mib 16384 \
+  --command bash --command -c --arg "sleep infinity" --wait
+
+# 配额按份额扣：1 × 25 核 = 0.25 卡当量
+runnpu project quota get main | python3 -c '
+import json,sys; d=json.load(sys.stdin)
+for u in d.get("cluster",{}).get("usage",[]) or []: print(u)'

package/compute-mode/preset/skills/runnpu/examples/05-troubleshoot.sh

@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+# 场景：负载不 running，按决策树定位（只读，随便跑）。
+set -euo pipefail
+W=${1:?用法: 05-troubleshoot.sh <project>/<name>}
+
+echo "== 1. 状态与原因"
+runnpu workload get "$W" | python3 -c '
+import json,sys; d=json.load(sys.stdin)["summary"]
+print(d["status"], "-", (d.get("reason") or {}).get("text",""))'
+
+echo "== 2. 事件（调度/准入/镜像拉取）"
+runnpu workload events "$W" | python3 -c '
+import json,sys
+for e in json.load(sys.stdin)["items"][:10]: print(e.get("time",""), e.get("reason",""), e.get("message","")[:100])'
+
+echo "== 3. 容器日志尾部（CrashLoop 时看 symbol lookup error → 镜像 LD 问题，见 troubleshooting.md）"
+runnpu workload logs "$W" --tail 30 2>/dev/null || echo "（还没有容器日志）"

package/compute-mode/preset/skills/runnpu/references/admin.md

@@ -0,0 +1,60 @@
+# 管理面：项目 / 配额 / 成员 / 用户 / 设置 / 审计
+
+## 项目与配额
+
+```bash
+runnpu project list / get <slug>
+runnpu project create --slug asr --name 语音识别 --pool ascend910b4 --guaranteed 4 --limit 4 \
+  --storage-pool shared-cephfs --storage-gib 200
+runnpu project delete asr --yes            # 项目里不能有未终止负载；负载卡 Terminating 时会失败（见 troubleshooting）
+runnpu project quota get asr               # 声明的配额 + 集群侧实际（cluster 段空 = 配额没落到集群）
+runnpu project quota set asr --pool ascend910b4=8,4 --storage-pool shared-cephfs=500
+runnpu project quota set asr --priority normal --workload-type workspace
+runnpu project repair asr --action resync  # 配额重新物化；release_grant_finalizer 是破窗动作，必须 --yes
+```
+
+- 单位是**卡**：`--pool <池>=<保底>,<可借增量>`；借用部分可被抢占。
+- `quota set` 是**整份替换**：给了 `--pool` 就替换全部池段——改一个池前先 `quota get`，把要保留的一并传入。
+- 建项目前先 `runnpu cluster get` 看池剩余；集群总卡被占满时新项目配额会被拒（负载报 QuotaGrantNotFound）。
+
+## 项目成员
+
+```bash
+runnpu project members list asr
+runnpu project members add asr --user alice@example.com --role 项目管理员   # 角色名 runnpu api GET /roles
+runnpu project members remove asr --user alice@example.com --yes           # 默认移除全部角色；--role 只删一个
+```
+
+## 用户与 Access Key
+
+```bash
+runnpu user list [--q alice] [--type robot] [--status disabled]
+runnpu user create --type local --email alice@example.com --display-name Alice   # 初始密码只返回一次，原样转给用户
+runnpu user create --type robot --name ci-bot --display-name "CI 机器人" --owner admin@example.com
+runnpu user disable|enable alice@example.com
+runnpu user access-key create ci-bot --name pipeline --expires-in-days 90        # 明文 token 只返回一次
+runnpu user access-key list|revoke …
+runnpu access-key list|create|revoke …    # 我自己的
+```
+
+- API token **不能**再签发 token（403）：`access-key create` 只在会话登录下可用，机器人 key 由管理员签发。
+- 禁用账号后其登录报 401 `account_disabled`；已有会话/token 同样被拒。
+
+## 系统设置与审计
+
+```bash
+runnpu settings get
+runnpu settings set storage.show_unnamed_pools true    # 点路径 → 嵌套 PATCH，自动解析 bool/数字
+runnpu audit list --limit 20 [--actor a@b.com] [--action workload.] [--target-kind user]   # --action 前缀匹配
+```
+
+## 未封装端点
+
+```bash
+runnpu api GET /projects/form-meta        # 池容量、其它项目已分配
+runnpu api GET '/overview/trends?range=24h'
+runnpu api GET /roles
+runnpu api POST /workloads -d '{...}'     # CLI 表达不了的字段（探针/autoscaling/configmap items…）
+```
+
+路径 = `api/openapi.yaml` 的 path 去掉 `/api/v1` 前缀。

package/compute-mode/preset/skills/runnpu/references/data-sources.md

@@ -0,0 +1,35 @@
+# 数据源与数据卷
+
+`--mount <类型>:<挂载点>[,键=值…][,ro]`，可重复。**整段创建后不可修改**（CRD 拒绝任何增删改），
+"再挂一个盘"= 重建负载。挂载点必须是非根绝对路径、彼此不重叠。
+
+| 类型 | 必填键 | 可选键 | 备注 |
+|---|---|---|---|
+| `pvc` | `name=` | `subpath=` | ⚠️ **PVC 必须带 `npu.run/project-id=<项目slug>` 标签**，否则被拒为 `PersistentVolumeClaimNotFound`（授权门槛，不是真的不存在；`kubectl label pvc <名> npu.run/project-id=<slug>`） |
+| `datavolume`（别名 `dv`） | `name=`（**集群卷名 cr_name**，`runnpu datavolume list` 查，不是显示名） | | 归属项目可读写；共享清单内 / 全平台共享**必须加 `,ro`**（否则 422 `cross_project_readonly`） |
+| `newvolume` | `pool=` `size=` | `persistency=workload\|project` `accessmode=` | 计入项目存储配额；`project` 持久卷在负载删除后保留 |
+| `nfs` | `server=` `path=` | | 直接投影，用前确认节点能挂 |
+| `s3` | `bucket=` `credential=` | `prefix=` | credential 是同项目登记的凭据名 |
+| `git` | `url=` | `revision=` `depth=` `credential=` | |
+| `hostpath` | `path=` | `kind=` | 高危：需 `workloads:hostpath` 权限 + 项目策略白名单（前缀+kind+读写位），写入还要策略放行 |
+| `emptydir` | | `size=` `medium=disk\|memory` | 只有 `medium=memory` 计内存 |
+| `configmap` / `secret` | `name=` | | 逐键投影（items[]）CLI 表达不了，走 `runnpu api` |
+
+## 受管卷（newvolume / 工作区）要点
+
+- `--workspace-pool` + `--workspace-gib` 是 `newvolume:/workspace,…` 的顺手写法。
+- 卷根由平台初始化为 `root:20000` mode 770 + ACL，容器进程带补充组 20000——**restricted 非 root 也可写**。
+- 容量计入项目在该池的存储配额（`runnpu project quota get` 看余量；不足 422 `exceeds_quota`）。
+- `persistency=workload` 随负载删除；`project` 保留为项目持久卷（删除走回收流程，可能有延迟）。
+
+## 数据卷命令组
+
+```bash
+runnpu datavolume list [--project X]     # 「集群卷名」列 = mount 时的 name
+runnpu datavolume create --project main --name 训练集A --pool shared-cephfs --capacity-gib 100 \
+  [--share proj-b --share proj-c] [--reclaim-policy retain|delete]
+runnpu datavolume update <集群卷名> [--name …] [--share …整体替换]
+runnpu datavolume delete <集群卷名> --yes    # 有活跃绑定 409 has_active_bindings；--force 跳过
+```
+
+`state` 从 creating 收敛到 ready 才能被挂；`cleanup_blocked` / `conflict` 是一等异常态，看 `state_note`。

package/compute-mode/preset/skills/runnpu/references/templates-assets.md

@@ -0,0 +1,61 @@
+# 模板与资产速查
+
+模板 = **整张创建表单**的值拷贝；资产 = **单个配置段**的值拷贝。两者都是「值」不是「链接」：
+展开时把值抄进创建请求，之后改模板不影响已建负载。
+
+## 模板
+
+```bash
+runnpu template list [--type training-env] [--scope cluster|project] [--project X] [-q 关键词] [--all]
+runnpu template get <名称或ID>                       # payload 就是创建请求的字段
+runnpu template create --name N --type T --payload @tpl.json [--description D] [--scope project --project X]
+runnpu template update <名称或ID> [--name …] [--payload @tpl.json]     # PUT 整体替换，没给的字段用现值兜底
+runnpu template delete <名称或ID> --yes
+```
+
+- 列表里 **★ 开头是系统预置**：只读，不能改删，只能照抄一份自己的。
+- `--payload` 支持 `@文件` / `-`（stdin）/ 字面 JSON。改现成的最省事：
+  `runnpu template get X -o json | jq .payload > tpl.json`，改完 `--payload @tpl.json`。
+- `--type` 创建后不可改（类型是模板身份的一部分）。
+- 权限不足时 `list` 返回**空列表而不是 403**，所以「空」要说成「没有或没权限」。
+
+### 用模板创建负载
+
+```bash
+runnpu workload create --project main --name nb6 --from-template <名称或ID> [--xpu 2] [--mount …] --wait
+```
+
+合并规则（CLI 保证，出错会明确报错而不是静默）：
+
+| 情况 | 行为 |
+|---|---|
+| 命令行没写的字段 | **保留模板值**——`--xpu` 的默认值 1 不会覆盖模板里的 4 |
+| 命令行显式写了的字段 | 覆盖模板值。**`--xpu 0`、`--jupyter=false` 这种显式零值也算覆盖** |
+| `--mount` 与模板预设的数据源 | 取**并集**；挂载点重复直接报错（不静默覆盖，否则用户以为模板的盘还在） |
+| `--type` 与模板类型不一致 | 报错。换模板，不要改类型 |
+| 模板里有 CLI 不认识的字段 | 报错并提示升级 CLI——**宁可报错也不丢字段** |
+
+模板的 `pending_fields` 是「使用者提交时要补」的字段：**问用户，不要编**。
+
+## 资产
+
+```bash
+runnpu asset list [--kind environment|compute|data-source|credential] [--scope …] [-q 关键词]
+runnpu asset get <名称或ID>                          # 凭证的密文永不回显
+runnpu asset create --kind K --name N [--subtype S] [--payload @a.json] [--secret KEY=VALUE ...]
+runnpu asset update <名称或ID> [--payload @a.json] [--secret KEY=VALUE ...]
+runnpu asset delete <名称或ID> --yes
+```
+
+- `--secret` 的值支持 `@文件`（推荐：避免密码/私钥进 shell history）：
+
+  ```bash
+  runnpu asset create --kind credential --subtype ssh-key --name 部署密钥 \
+      --payload '{}' --secret 'ssh-privatekey=@~/.ssh/id_ed25519'
+  ```
+
+- 凭证是唯一会离开 `ready` 的资产——它要在项目命名空间里落成 Secret，那一步可能失败；
+  `list` 的状态列会带出 `state_note`。
+- **没有 `data-volume` 大类**：数据卷背后是真实的一块盘，是独立一等资源，走 `runnpu datavolume`。
+- 资产怎么用：environment（镜像 + env）/ compute（池 + 卡数）按段并入创建参数；
+  credential / data-source 在挂载里按名引用（如 `--mount s3:/data,bucket=b,credential=<资产名>`）。

package/compute-mode/preset/skills/runnpu/references/training-runs.md

@@ -0,0 +1,56 @@
+# 训练类负载的平台侧事实（RunNPU / 昇腾）
+
+> 训练**方法论**（跑通链路的纪律、看曲线诊断、自动超参研究、长程与并行）不在本文件——
+> 见 `training-research` skill（训练模式会话里可用）。本文件只放**换个平台就作废**的事实：
+> 这个平台怎么传文件、这套镜像有什么炸点、这个集群实测过什么。
+
+## 1. 工程进集群（本平台的通道）
+
+- **先起一个带 SSH 的训练环境**。SSH 与安全级、与 root **无关**（helper 是注入的原生二进制，
+  PID 1 非 root 时保持当前身份；run-npu#595 起 restricted 也支持，旧版 operator 才会以
+  `SshRootInitUnavailable` 拒绝）——`restricted` 下照样能开，**不要为了拿 SSH 把负载降到 baseline**。
+  SSH 会话身份继承容器主进程，传上去的文件属主与容器内进程一致。
+  root 镜像在 restricted 下要 `--run-as-uid`，那是安全级自身的要求，别和 SSH 绑一起。
+- ⚠️ **torch-npu 官方镜像做 training-env 必须带 `--arg sleep --arg infinity`**（平台已知问题：
+  该镜像 entrypoint 无命令时空参立退，PID 1 数秒退出 → 容器每 1-2 分钟重启，SSH 只在残存窗口
+  时好时坏，极易误判成网络/平台故障）。用 args、**不要用 `--command`**——保留 entrypoint 的
+  CANN 环境装配，SSH helper 的 env 快照才有内容。training-job 有真实训练命令，不受此影响。
+- 手工上传路径（无 `training_upload` 工具时，如算力管理模式）：连接参数**别自己拼**——
+  `runnpu workload endpoints <p>/<n>` 给出拼好的命令，结构化值（脚本用）在
+  `runnpu workload get <p>/<n> -o json` 的 `.ssh.{username,host,port}`。登录用户名固定是
+  `<负载名>.<命名空间>`（SSHPiper 路由标识；注意命名空间带 `proj-` 前缀，不是项目 slug 本身，
+  更不是平台用户名），如 `scp -P 32222 -r ./工程 job1.proj-main@<入口IP>:/workspace/`；
+  平台 SFTP 已验证可用。能力边界：exec / 交互 shell / sftp / 本地转发 `-L`/`-D` 都支持，
+  **远端监听 `-R` 未实现**。
+- 备选 exec 管道（无 SSH 时）：`tar czf - . | runnpu workload exec <p>/<n>` 里接
+  `tar xzf - -C /workspace`（stdin 必须以 `exit` 结尾）；只传代码，数据集级流量不过控制面。
+- 数据集进数据卷 / 工作盘（`--mount` / `--workspace-gib`），一次落盘、代码单独同步。
+
+## 2. 容器内运行（本部署实测）
+
+- 依赖安装走平台注入的 pip 镜像（负载 env 里已有）；装 opencv 报 libGL 缺失 →
+  改 `opencv-python-headless`，不要 apt 补一堆图形库。
+- **pip 装到 `--target /workspace/pylibs`** 并 `export PYTHONPATH=/workspace/pylibs`：
+  容器层随重启丢失，/workspace 工作盘持久——否则重启一次就要重装一遍。
+- 本集群到 PyPI 的出口带宽实测只有几十 KiB/s：大依赖预置进数据卷/工作盘。
+- 平台日志通道：`runnpu workload logs <p>/<n> --tail`（监控双信号源里的信号源 B）。
+
+## 3. 两个"看起来成功了"的假象（本平台实测踩过，2026-09-02）
+
+- **job 秒级 completed + 跟踪平台无 run**：大概率是**覆盖了镜像 entrypoint**
+  （`--command bash -c …`）丢掉 CANN 环境装配 → torch_npu 不可用 → 设备探测降级 CPU
+  秒跑完、PYTHONPATH 也可能没带上。提交 training-job 时**保留 entrypoint、用 args 传命令**
+  （与 training-env 的 sleep infinity 同一个原理）；手动按模板 payload 拼 job 时，
+  模板里的 `LD_LIBRARY_PATH` 救援 env 必须原样带上。正常 NPU 训练首个 step 要编译算子
+  （分钟级），秒级完成先怀疑设备降级，别急着重提。
+- **卡够 ≠ 调度得上**：`CardInsufficientMemory` 事件 + 有空闲卡数，通常是**分卡（vNPU）碎片化**
+  ——平台外 Pod 按显存切片占卡，凑不出整卡。查逐卡占用（node metrics / overview 逐卡网格），
+  必要时先停自己占整卡的 env 给 job 让路（先 job 后 env，反过来 env 会把唯一整卡占住）。
+
+## 4. 昇腾特有炸点（详见 troubleshooting.md「镜像兼容性」）
+
+- 官方 torch-npu 镜像的 LD 冲突：预置模板已带救援 `LD_LIBRARY_PATH`，自拼镜像时照抄那条 env。
+- CUDA 工程迁移三件套：`import torch, torch_npu` + `from torch_npu.contrib import transfer_to_npu`
+  必须在训练框架 import **之前**；AMP 自检会探 CUDA，训练参数要显式关掉（如 ultralytics `amp=False`）；
+  dataloader `workers` 从小起步。
+- 容器内 `npu-smi` 报 libruntime 错链时的前缀写法也在 troubleshooting.md。

package/compute-mode/preset/skills/runnpu/references/troubleshooting.md

@@ -0,0 +1,75 @@
+# 排查决策树与已知集群行为
+
+## 决策树
+
+```
+runnpu workload get <p>/<n>  → status + reason.text 先读完
+│
+├─ rejected            → 终态，不会自愈。reason.text 说明原因：
+│   ├─ QuotaGrantNotFound   → 项目配额没落到集群（建项目太快 / 集群配额被占满）。
+│   │                         runnpu project quota get <slug> 看 cluster 段；配额恢复后**删掉重建**负载
+│   ├─ PersistentVolumeClaimNotFound → pvc 挂载缺 npu.run/project-id 标签（见 data-sources），或真不存在
+│   ├─ QuotaResourceUnauthorized / AcceleratorPoolUnavailable → 池没在项目配额里 / 池发现不可用
+│   └─ 其它 reason 原样转述，别猜
+├─ queued              → 配额不够在排队：runnpu project quota get 看 used vs guaranteed（+limit 可借）
+├─ pending / admitted 很久 → runnpu workload events：
+│   ├─ Pulling image…       → 大镜像拉取，等；--wait 超时 ≠ 失败
+│   ├─ FilteringFailed / 0 nodes fit → 卡不够或分卡份额放不下
+│   └─ 无事件               → runnpu workload yaml 看 CR conditions
+├─ running 但工具/端口不通 → 下面「入口就绪」
+├─ CrashLoopBackOff    → runnpu workload logs：
+│   ├─ symbol lookup error（libdrvdsmi/libruntime）→ 下面「镜像兼容性」
+│   └─ 其它应用错误原样转述
+└─ Terminating 很久    → 下面「删除卡住」
+```
+
+## 镜像兼容性（昇腾镜像 + HAMi 注入）
+
+平台用 HAMi 软切时会向**每个进程**注入 preload 库，它要求容器的 `LD_LIBRARY_PATH` 能同时解析到
+CANN `libruntime.so` 和与宿主驱动一致的 dsmi/hal 链。已知结论：
+
+- ✅ `vllm-ascend` 镜像（平台预置列表里的）：路径顺序兼容，直接用；
+- ❌ **ascendhub 官方 `torch-npu` 镜像**：烘死的 LD 路径与注入冲突，容器里 **bash 都起不来**
+  （`symbol lookup error: libdrvdsmi_host.so: drvSetDeviceInfo`，CrashLoopBackOff）。
+  **救援配方**（创建时加一个 env，实测 Running + npu-smi + torch_npu 全通）：
+
+  ```
+  --env "LD_LIBRARY_PATH=/usr/local/Ascend/driver/lib64/driver:/usr/local/Ascend/driver/lib64/common:/usr/local/Ascend/driver/lib64:/usr/local/Ascend/ascend-toolkit/latest/lib64"
+  ```
+
+  要点：宿主驱动目录在前，镜像 toolkit 目录兜 `libruntime.so`；**只留驱动路径会缺 libruntime.so 照样炸**。
+  其它镜像炸在同一错误时，同思路调整（驱动在前 + 该镜像的 CANN lib64 在后）。
+- ❌ busybox/alpine 等非 glibc 镜像与注入不兼容，连通性验证用 `ubuntu:24.04`。
+- 容器内 `npu-smi` 报 `libruntime.so: cannot open` 是另一回事（宿主二进制错链镜像 CANN）：
+  exec 时前缀 `LD_LIBRARY_PATH=/usr/local/Ascend/driver/lib64:…/driver:…/common npu-smi info` 即可。
+
+## 入口就绪的口径
+
+- 工具 `url` 只在 ready 且有权限时返回；`internal_url` 是集群内地址（浏览器打不开）。
+- `node_port` 是**分配观测**：不含节点地址、不代表可达、可能在 ready=false 时就有。访问 = `http://<任一节点IP>:<node_port>`。
+- **nodePort 工具的 ready 有已知误报**（run-npu#553：EndpointUnavailable 但实际可达）：
+  Pod Running 后直接 `curl --max-time 5 http://<节点IP>:<node_port>/` 实测，别等 ready。
+
+## 删除卡住（run-npu#549）
+
+带**工具 Service / 受管卷 / 项目持久卷**的负载删除后可能长期 Terminating（operator teardown 对已删对象
+的 404 处理缺陷）。表现：`kubectl` 里 RunWorkload 带 deletionTimestamp 不消失、PVC 卡 Terminating。
+
+- **不要**反复 delete、不要循环 --force；
+- 如实告知用户是已知上游问题，需等待修复或运维用 finalizer 手段清理（运维操作，不在本 Skill 权限内做）；
+- 项目删除因此失败时（`has_running_workloads` / 卡 Terminating）同理。
+
+训练任务删除后残留的 Completed Pod（ownerReferences 为空）是另一个已知无害残留（run-npu#551）。
+
+## 配额相关
+
+- 建项目后配额（QuotaGrant）落集群有延迟且可能被拒（集群卡总量被其它项目占满时）：
+  建完先 `runnpu project quota get <slug>` 确认 `cluster` 段非空再建负载。
+- `queued` 时算账：单位是卡；分卡按份额折算（25 核 = 0.25 卡）；`limit` 是可借增量，借用的部分可被抢占。
+- `runnpu project repair <slug> --action resync` 可触发配额重新物化（管理员）。
+
+## 指标与日志
+
+- `metrics.available=false` / 值为 null = 数据源未接入（Prometheus 的 NPU 指标、ClickHouse 日志可能没部署），
+  说"未接入"，不是 0，原因在 `metrics.reason`。
+- `runnpu workload logs` 是容器 stdout；`events` 是 K8s Event + CR 条件（最近 1 小时）。

package/compute-mode/preset/skills/runnpu/references/workload-types.md

@@ -0,0 +1,47 @@
+# 四类负载：参数矩阵与生命周期
+
+
+|      | workspace 开发环境                           | training-env 训练环境 | training-job 训练任务                                                                                              | inference 推理服务                                  |
+| ---- | ---------------------------------------- | ----------------- | -------------------------------------------------------------------------------------------------------------- | ----------------------------------------------- |
+| 形态   | 常驻交互（Sandbox）                            | 常驻交互（Sandbox）     | 跑完即止（Job / 多机 JobSet）                                                                                          | 常驻服务（Deployment）                                |
+| 可抢占  | ✅（闲时可被回收）                                | ❌                 | 由优先级决定                                                                                                         | ❌                                               |
+| 工具   | `--jupyter` `--code-server` `--ssh`      | 同左                | 无                                                                                                              | `--port`（对外端口）                                  |
+| 专属参数 | `--idle/max-lifetime`（暂无 CLI flag，走 api） | 同左                | `--nodes` `--completions` `--parallelism` `--backoff-limit` `--active-deadline-seconds` `--ttl-after-finished` | `--port` `--replicas`（探针/伸缩/滚动策略走 `runnpu api`） |
+| 终态   | stopped / failed                         | 同左                | **completed** / failed                                                                                         | 无自然终态                                           |
+
+
+## 通用参数
+
+- `--xpu N`：整卡卡数 0..8（0 = 纯 CPU）。**分卡**：`--xpu-core 1..100` + `--xpu-memory-mib M` 成对给出，
+且 `--xpu` 必须为 1——**只支持单卡切分**，多卡分卡上游不支持（要多张卡就整卡申请，或建多个负载）。
+只给一半会被 422 拒（服务端不猜另一半）。
+- `--security restricted|baseline`：restricted 强制非 root（root 镜像必须给 `--run-as-uid`）。但注意多数昇腾官方镜像（root 运行）仍需要 baseline——那是镜像自身身份的要求，不是 SSH 的。
+- `--pool`：省略取项目第一个池；`runnpu cluster get` 查池名。
+- `--command` / `--arg` 可重复按序拼接：`--command bash --command -c --arg "…"`。
+
+## training-job 细节
+
+- 单机（`--nodes 1`，默认）：`--completions`（目标完成次数）、`--parallelism`（并行 Pod 数）可用；
+- 多机（`--nodes ≥ 2`）：**没有** completions/parallelism（规模由节点数决定，传了 422）；
+`--backoff-limit` 变成整组重启次数；`--active-deadline-seconds` 创建后不可变。
+- `--ttl-after-finished 0` 表示完成立即删；省略 = 不自动删（完成的负载保留，占记录不占卡）。
+- 进度看 `runnpu workload get`：`progress` 里 `succeeded / failed / retries_used`，多机另有 `ready_replicas`。
+- 已知小残留：任务完成后被删除时，Completed Pod 可能残留在 namespace（上游 run-npu#551，无害）。
+
+## inference 细节
+
+- `--port [名字:]<端口>[,expose=…][,node-port=N]` 可重复（≤8）；名字是 DNS label ≤15（要拼进入口 host），省略为 `http`。
+- `expose`：`clusterIP`（默认，仅集群内）/ `nodePort`（30000..32767，`node-port=` 可指定，否则集群分配）/
+`loadBalancer`（需集群有 LB）/ `externalUrl`（需平台配了入口域名）。
+- 推理端点一律 `access: public`——平台登录态挡不住 API 客户端，**鉴权由服务自身负责**（vLLM 加 `--api-key` 之类）。
+- `--replicas N` 副本数；探针（readiness/liveness）、autoscaling、max_surge/max_unavailable 目前 CLI 不表达，
+需要时 `runnpu api POST /workloads` 直接给 `service` 对象（字段见 api/openapi.yaml 的 ServiceSpecInput）。
+- 状态语义：`running` = Deployment 副本就绪；工具 `ready` 是**入口观测**，nodePort 场景有已知误报
+（run-npu#553：EndpointUnavailable 但端口实际可达）——以实测 `curl <节点IP>:<nodePort>` 为准。
+- 服务视图字段：`runnpu workload get` 的 `service` 段有 replicas / ready_replicas / available_replicas。
+
+## 不可修改的字段（想改 = 重建负载）
+
+类型、项目、规格（xpu/core/memoryMiB）、数据源整段、SSH 开关、多机节点数、
+batch 的 backoff/deadline、镜像与命令（可改的只有：暂停/恢复、部分运行时参数）。
+用户要"改配置"时先对照这张表，别许诺做不到的事。

package/cordis.patch.yml

@@ -0,0 +1,80 @@
+# dsh-runnpu bundle patch — HOST-PLANE rows only.
+#
+# What belongs here vs. in an agent preset:
+#   - HOST plane (this file): things shared across every session — the RunNPU /
+#     SwanLab connectors, the plugin's own HTTP routes the browser half calls,
+#     and the client roster row that serves the browser bundle.
+#   - AGENT plane (~/.dsh/.agent-presets/{compute,training}/agent.cordis.yml):
+#     what one session's agent gets — its tools, persona and skill roots. A
+#     preset is what decides whether a session is "compute mode" at all.
+#
+# The two `*/node` rows below own the per-mode HTTP surface. They inject
+# `webServer`, so they activate exactly when the shipped webserver row does.
+
+- insert:
+    # Shared connectors: ctx.runnpu (REST client for the RunNPU /api/v1 control
+    # panel) and ctx.swanlab (thin wrapper over the `swanlab api` CLI). Both are
+    # plain-value services read with ctx.get, so a consumer degrades with a
+    # stable reason when nothing is configured yet.
+    - id: runnpu-connectors
+      name: 'dsh-runnpu/shared/connectors'
+      config:
+        # Server address and token are NOT here: they are entered in the UI and
+        # stored through the credential seam. A value here is only a fallback
+        # default for a fresh install.
+        runnpu:
+          server: ''
+          tokenEnv: 'RUNNPU_TOKEN'
+        swanlab:
+          apiHost: 'https://api.swanlab.cn'
+          webHost: 'https://swanlab.cn'
+          apiKeyEnv: 'SWANLAB_API_KEY'
+          cli: 'swanlab'
+
+    # Compute mode's node half: connection/login routes, cluster reads, and the
+    # pinned-chart registry with its refresh loop.
+    - id: runnpu-compute-node
+      name: 'dsh-runnpu/compute/node'
+      config:
+        apiPath: '/runnpu-api'
+        refreshMs: 30000
+
+    # Training mode's node half: `.train/` ledger reads, SwanLab metric pulls
+    # for the metric strip, and the training-side config routes.
+    - id: runnpu-training-node
+      name: 'dsh-runnpu/training/node'
+      config:
+        apiPath: '/training-api'
+        refreshMs: 15000
+
+    # ── periodic reports ────────────────────────────────────────────────
+    #
+    # Session-local reminders, which is how 定期报告 is delivered: the agent
+    # calls `schedule_create` with an interval, and each occurrence comes back
+    # as an ordinary follow-up message that prompts it to run `cluster_report`.
+    #
+    # HOST plane, not the preset, and the reason is mechanical: the plugin
+    # observes `agent/created` and installs on ROOT agents published AFTER it
+    # loads. A preset row loads with the agent it configures, which is already
+    # too late — the tools would never appear. Mounting it here also gives both
+    # modes reminders from one row.
+    #
+    # ⚠️ A session that was already running when this row loaded does NOT have
+    # the reminder tools. After adding this, start a NEW session.
+    #
+    # `time-context` is what lets the model turn "每天早上九点" into an explicit
+    # instant: Schedule itself never reads a browser, session or process time
+    # zone, and refuses a local time without an explicit IANA zone.
+    - id: time-context
+      name: '@deepseek-ai/dsh-time-context'
+
+    - id: schedule
+      name: '@deepseek-ai/dsh-schedule'
+
+    # Client roster row: the client-modules node half scans this entry, reads
+    # `dsh.client` + exports["./client"] from package.json, and serves the
+    # browser bundle (dsh ≥ 0.1.5: /plugins/??dsh-runnpu/client.js&rev=…). One row per
+    # package — the bundle carries BOTH modes' UI and branches on the session's
+    # `agentPreset` projection.
+    - id: runnpu-modes-ui
+      name: 'dsh-runnpu'

package/dist/chunk-JLZMOSQC.mjs

@@ -0,0 +1 @@
+var h=["overview.trends","node.metrics","workload.metrics"],k=["utilization","allocation","storage","workloads"],w={cpu:{label:"CPU \u5229\u7528\u7387",unit:"%",rows:[["cpu_pct","CPU \u5229\u7528\u7387"]]},memory:{label:"\u5185\u5B58\u5229\u7528\u7387",unit:"%",rows:[["memory_pct","\u5185\u5B58\u5229\u7528\u7387"]]},net:{label:"\u7F51\u7EDC\u541E\u5410",unit:"B/s",rows:[["net_rx_bps","\u63A5\u6536"],["net_tx_bps","\u53D1\u9001"]]},disk:{label:"\u78C1\u76D8\u541E\u5410",unit:"B/s",rows:[["disk_read_bps","\u8BFB"],["disk_write_bps","\u5199"]]}},_=240;function y(s,t){if(s==="workload.metrics"){let e={};for(let n of t?.groups??[])e[n.key]={label:n.title??n.key,unit:n.unit,rows:n.series??[]};return e}if(s==="node.metrics"){let e={};for(let[n,r]of Object.entries(w)){let u=r.rows.filter(([l])=>Array.isArray(t?.[l])).map(([l,f])=>({key:l,label:f,unit:r.unit,data:t[l]}));u.length>0&&(e[n]={label:r.label,unit:r.unit,rows:u})}return e}let o={};for(let e of k)Array.isArray(t?.[e])&&(o[e]={label:e,unit:t[e][0]?.unit,rows:t[e]});return o}function x(s,t){return Object.keys(y(s,t))}function A(s,t={}){let o=t.query??"overview.trends",e=s?.timestamps;if(!Array.isArray(e))return{series:[],note:"\u8BE5\u67E5\u8BE2\u6CA1\u6709\u65F6\u95F4\u8F74\uFF0C\u65E0\u6CD5\u753B\u6210\u66F2\u7EBF"};let n=y(o,s),r=Object.keys(n);if(r.length===0)return{series:[],note:"\u8FD9\u6B21\u54CD\u5E94\u91CC\u6CA1\u6709\u4EFB\u4F55\u53EF\u7ED8\u5236\u7684\u6307\u6807\u7EC4"};let u=t.group??r[0],l=n[u]??n[r[0]],f=n[u]===void 0,m=e.map(i=>new Date(i).getTime()),b=t.series?.length?new Set(t.series):void 0,d=Math.max(1,Math.ceil(m.length/_)),g=(l.rows??[]).filter(i=>b?b.has(i.key):!0).slice(0,5).map(i=>({key:i.key,name:i.label??i.key,unit:i.unit??l.unit,points:(i.data??[]).map((a,c)=>[m[c],a]).filter((a,c)=>c%d===0).filter(([a])=>Number.isFinite(a)).map(([a,c])=>[a,typeof c=="number"&&Number.isFinite(c)?c:null])})),p=[];return f&&p.push(`\u6CA1\u6709 ${u} \u8FD9\u4E00\u7EC4\uFF0C\u753B\u7684\u662F ${r[0]}\uFF08\u53EF\u7528\uFF1A${r.join(" / ")}\uFF09`),s?.metrics?.available===!1&&p.push(`\u6307\u6807\u6E90\u672A\u63A5\u5165\uFF08${s.metrics.reason??"\u539F\u56E0\u672A\u7ED9\u51FA"}\uFF09\uFF0C\u8FD9\u7EC4\u66F2\u7EBF\u6CA1\u6709\u6570\u636E`),{series:g,group:n[u]?u:r[0],note:p.length>0?p.join("\uFF1B"):void 0}}function O(s){let t=s?.points??[];for(let o=t.length-1;o>=0;o-=1){let e=t[o]?.[1];if(e!=null)return e}}export{h as a,k as b,y as c,x as d,A as e,O as f};

package/dist/chunk-JUBIWGKW.mjs

@@ -0,0 +1,4 @@
+import{randomUUID as K}from"node:crypto";import{readFile as H,writeFile as g,mkdir as v,chmod as R,readdir as P,rename as U,rm as _}from"node:fs/promises";import{existsSync as w}from"node:fs";import{homedir as C}from"node:os";import{join as s,resolve as A}from"node:path";var L=".train",x={apiHost:"https://api.swanlab.cn",webHost:"https://swanlab.cn"},S={runnpuToken:"RUNNPU_TOKEN",swanlabApiKey:"SWANLAB_API_KEY"};function N(){let e=process.env.DSH_HOME||s(C(),".dsh");return s(e,"runnpu-modes")}async function f(e){try{return JSON.parse(await H(e,"utf8"))}catch{return}}async function m(e,n,r){await v(s(e,".."),{recursive:!0}),await g(e,`${JSON.stringify(n,void 0,2)}
+`,{encoding:"utf8",mode:r}),r!==void 0&&await R(e,r)}function l(e){return s(A(e),L)}async function I(e){if(!e)return{config:void 0,secrets:void 0,dir:void 0};let n=l(e);return{dir:n,config:await f(s(n,"config.json")),secrets:await f(s(n,"secrets.json"))}}async function D(e){let n=l(e);await v(s(n,"runs"),{recursive:!0});let r=s(n,".gitignore");return w(r)||await g(r,`secrets.json
+`,"utf8"),n}async function G(e,n){let r=await D(e),i={...await f(s(r,"config.json"))??{},...n,updatedAt:new Date().toISOString()};return await m(s(r,"config.json"),i),i}async function J(e,n){let r=await D(e),t=s(r,"secrets.json"),o={...await f(t)??{}};for(let[a,c]of Object.entries(n))c===void 0||c===""?delete o[a]:o[a]=c;return await m(t,o,384),Object.keys(o)}async function h(e,n,r){let t=r?.[n];if(typeof t=="string"&&t!=="")return{value:t,source:"project"};let i=e.get?.("credentials");if(i)try{let a=await i.resolve(n);if(a?.value)return{value:a.value,source:`credentials:${a.source}`}}catch{}let o=process.env[n];if(o)return{value:o,source:"env"}}async function V(e,n,r,t){if(t.scope==="project"){if(!t.cwd)throw new Error("project scope needs a workspace");return await J(t.cwd,{[n]:r}),"project"}let i=e.get?.("credentials");if(!i)throw new Error("credential provider unavailable; use project scope");return await i.set(n,r),"credentials"}async function E(e,n,r){let{config:t,secrets:i,dir:o}=await I(r),a=await f(s(N(),"connections.json"))??{},c=t?.runnpu?.server||a.runnpu?.server||n?.runnpu?.server||process.env.RUNNPU_SERVER||"",p=await h(e,S.runnpuToken,i),u=t?.swanlab?.mode||a.swanlab?.mode||"cloud",d=u==="cloud",j=t?.swanlab?.apiHost||a.swanlab?.apiHost||(d?x.apiHost:""),O=t?.swanlab?.webHost||a.swanlab?.webHost||(d?x.webHost:""),b=await h(e,S.swanlabApiKey,i);return{workspace:r??void 0,trainDir:o,runnpu:{server:c,configured:!!(c&&p),tokenSource:p?.source,token:p?.value},swanlab:{mode:u,apiHost:j,webHost:O,configured:!!(j&&b),apiKeySource:b?.source,apiKey:b?.value}}}async function Y(e,n,r){let t=await E(e,n,r);return{workspace:t.workspace,trainDir:t.trainDir,runnpu:{server:t.runnpu.server,configured:t.runnpu.configured,tokenSource:t.runnpu.tokenSource},swanlab:{mode:t.swanlab.mode,apiHost:t.swanlab.apiHost,webHost:t.swanlab.webHost,configured:t.swanlab.configured,apiKeySource:t.swanlab.apiKeySource}}}async function Q(e,n,r){return await E(e,n,r)}async function X(e){let n=s(N(),"connections.json"),t={...await f(n)??{},...e};return await m(n,t),t}async function Z(e){if(!e)return[];let n=s(l(e),"runs"),r;try{r=await P(n)}catch{return[]}let t=[];for(let i of r.filter(o=>o.endsWith(".json")).sort().reverse()){let o=await f(s(n,i));o&&t.push(o)}return t}function T(e){return`${String(e).padStart(4,"0")}.json`}var k=Object.freeze(["swanlabRun","status","endedAt","swanlabUpload"]),y=new Map;function $(e,n){let r=A(e),i=(y.get(r)??Promise.resolve()).then(n),o=i.catch(()=>{});return y.set(r,o),o.then(()=>{y.get(r)===o&&y.delete(r)}),i}async function F(e,n){await v(s(e,".."),{recursive:!0});let r=`${e}.${K().slice(0,8)}.tmp`;await g(r,`${JSON.stringify(n,void 0,2)}
+`,"utf8");try{await U(r,e)}catch(t){throw await _(r,{force:!0}),t}}async function ee(e,n,r){return await $(e,async()=>{let t=s(l(e),"runs",T(n)),i=r??{},o=Object.keys(i).filter(u=>!k.includes(u)),a=await f(t);if(!a)return{path:t,run:void 0,changed:[],ignored:o};let c={...a},p=[];for(let u of k){if(!(u in i))continue;let d=i[u];d===void 0||d===a[u]||(c[u]=d,p.push(u))}return p.length===0?{path:t,run:a,changed:p,ignored:o}:(await F(t,c),{path:t,run:c,changed:p,ignored:o})})}async function ne(e){if(e)return await f(s(l(e),"metrics.json"))}async function te(e){if(!e)return!1;if(w(l(e))||w(s(e,"train.py")))return!0;for(let n of["requirements.txt","pyproject.toml"])try{let r=await H(s(e,n),"utf8");if(/\b(torch|transformers|deepspeed|swanlab|accelerate|ultralytics)\b/i.test(r))return!0}catch{}for(let n of["configs","config"])if(w(s(e,n,"accelerate.yaml"))||w(s(e,n,"deepspeed.json")))return!0;return!1}export{x as a,S as b,N as c,l as d,I as e,D as f,G as g,J as h,V as i,Y as j,Q as k,X as l,Z as m,ee as n,ne as o,te as p};

package/dist/chunk-JZXLCA2E.mjs

@@ -0,0 +1 @@
+var d=Object.defineProperty;var e=(c,a)=>{for(var b in a)d(c,b,{get:a[b],enumerable:!0})};export{e as a};

package/dist/chunk-KQOFBYGE.mjs

@@ -0,0 +1,9 @@
+import{randomBytes as m,timingSafeEqual as b}from"node:crypto";var y=m(32).toString("base64url"),N="x-runnpu-nonce",E="nonce",a=0,d;function _(r){let t=r.get("webServer");t&&r.effect(()=>(a+=1,a===1&&(d=t.tapIndex(e=>e.replace("</head>",`<script>globalThis.__RUNNPU_NONCE__=${JSON.stringify(y)}</script></head>`))),()=>{a-=1,!(a>0)&&(d?.(),d=void 0)}),"runnpu: index nonce injection")}function p(r){if(typeof r!="string"||r.length===0)return!1;let t=Buffer.from(r,"utf8"),e=Buffer.from(y,"utf8");return t.byteLength===e.byteLength&&b(t,e)}function u(r,t){let e=r.headers?.[N];if(p(Array.isArray(e)?e[0]:e))return!0;let n=t??new URL(r.url??"/","http://127.0.0.1");return p(n.searchParams.get(E)??"")}function x(r,t,e){let n=JSON.stringify(e);r.writeHead(t,{"content-type":"application/json; charset=utf-8","cache-control":"no-store","content-length":Buffer.byteLength(n)}),r.end(n)}function h(r,t,e,n,o){x(r,t,{error:e,status:t,message:n,details:o})}async function H(r,t=256*1024){let e=[],n=0;for await(let o of r){if(n+=o.length,n>t)throw Object.assign(new Error(`request body exceeds ${t} bytes`),{status:413,code:"body_too_large"});e.push(o)}if(n!==0)try{return JSON.parse(Buffer.concat(e).toString("utf8"))}catch(o){throw Object.assign(new Error(`request body is not valid JSON: ${o.message}`),{status:400,code:"bad_json"})}}function S(r){let t=r.headers.origin;if(!t)return!0;try{let e=new URL(t);return e.hostname==="127.0.0.1"||e.hostname==="localhost"||e.hostname==="[::1]"}catch{return!1}}function J(r){return async(t,e)=>{try{let n=new URL(t.url??"/","http://127.0.0.1"),o=r[t.method??"GET"];if(!o){e.writeHead(405,{allow:Object.keys(r).join(", ")}).end();return}if(t.method!=="GET"&&t.method!=="HEAD"&&!S(t)){h(e,403,"forbidden_origin","cross-origin request refused");return}if(!u(t,n)){h(e,401,"unauthenticated","\u7F3A\u5C11\u6216\u65E0\u6548\u7684\u63D2\u4EF6\u8BF7\u6C42\u4EE4\u724C");return}await o(t,e,n)}catch(n){if(e.headersSent){e.end();return}h(e,n?.status&&Number.isInteger(n.status)?n.status:500,n?.code??"internal_error",String(n?.message??n),n?.details)}}}var f=class{#e=new Set;subscribe(t,e){e.writeHead(200,{"content-type":"text/event-stream","cache-control":"no-cache, no-transform",connection:"keep-alive"}),e.write(`: connected
+
+`);let n={res:e};this.#e.add(n);let o=()=>{this.#e.delete(n);try{e.end()}catch{}};return t.on("close",o),o}publish(t,e){let n=`event: ${t}
+data: ${JSON.stringify(e)}
+
+`;for(let o of this.#e)try{o.res.write(n)}catch{this.#e.delete(o)}}closeAll(){for(let t of this.#e)try{t.res.end()}catch{}this.#e.clear()}get size(){return this.#e.size}};function v(r,t,e){try{r.write(`event: ${t}
+data: ${JSON.stringify(e)}
+
+`)}catch{}}function R({refresh:r,intervalMs:t}){let e=new f;async function n(){await r()}function o(i,c,s){if(i.method!=="GET"){c.writeHead(405,{allow:"GET"}).end();return}if(!u(i)){c.writeHead(401,{"content-type":"application/json"}).end('{"error":"unauthenticated"}');return}if(e.subscribe(i,c),s?.onClose&&i.on("close",s.onClose),!!s?.seed)try{s.seed((l,w)=>v(c,l,w))}catch{}}function g(i,c){i.effect(()=>{let s=!1,l=setInterval(()=>{e.size!==0&&(s||(s=!0,n().catch(()=>{}).finally(()=>{s=!1})))},t);return()=>clearInterval(l)},`${c}: refresh loop`),i.effect(()=>()=>e.closeAll(),`${c}: sse teardown`)}return{hub:e,handler:o,install:g}}export{_ as a,x as b,h as c,H as d,J as e,R as f};

package/dist/chunk-XMZG2NP7.mjs

@@ -0,0 +1,2 @@
+import{a as $,e as v}from"./chunk-JLZMOSQC.mjs";import{c as R}from"./chunk-JUBIWGKW.mjs";import{readFile as ne,writeFile as ie,mkdir as oe,rename as ae,rm as ue}from"node:fs/promises";import{join as D}from"node:path";import{randomUUID as G}from"node:crypto";var S=["line","area","bar","table"],y={projects:{items:{label:"\u9879\u76EE",at:t=>t?.items},usage:{label:"\u9879\u76EE \xD7 \u8D44\u6E90\u6C60 \u7528\u91CF",at:t=>A(t,"usage",K)},storage:{label:"\u9879\u76EE \xD7 \u5B58\u50A8\u6C60 \u7528\u91CF",at:t=>A(t,"storage",W)},queues:{label:"\u9879\u76EE\u961F\u5217\u51C6\u5165",at:t=>A(t,"queues",z)}},workloads:{items:{label:"\u8D1F\u8F7D",at:t=>t?.items}},nodes:{items:{label:"\u8282\u70B9",at:t=>t?.items}},cluster:{items:{label:"\u96C6\u7FA4\u7EC4\u4EF6",at:t=>t?.items}},overview:{projects:{label:"\u9879\u76EE\u914D\u989D\u4E0E\u7528\u91CF",at:t=>t?.projects},idle:{label:"\u95F2\u7F6E\u5360\u7528",at:t=>t?.idle},queued:{label:"\u6392\u961F\u8D1F\u8F7D",at:t=>t?.queued},nodes:{label:"\u8282\u70B9",at:t=>t?.nodes}}},F=Object.fromEntries(Object.entries(y).map(([t,e])=>[t,Object.entries(e).map(([r,s])=>({key:r,label:s.label}))]));function A(t,e,r){let s=[];for(let n of t?.items??[])for(let a of n?.cluster?.[e]??[])s.push({project:n.slug,project_name:n.name,...r(a)});return s}function K(t){let e=t?.guaranteed??0,r=t?.used??0;return{pool:t?.pool,device_type:t?.device_type,guaranteed:e,used:r,over_guarantee:r-e}}function W(t){let e=t?.limit_gib??0,r=t?.remaining_gib,s=typeof r=="number"&&e>0?e-r:void 0;return{pool:t?.pool,limit_gib:t?.limit_gib,persistent_gib:t?.persistent_gib,runtime_gib:t?.runtime_gib,remaining_gib:r,used_gib:s,used_pct:s===void 0||e<=0?void 0:Math.round(s/e*1e3)/10}}function z(t){return{queue:t?.name,pool:t?.pool,admitted:t?.admitted,reason:t?.reason}}function Q(t,e,r){if(t==="overview"&&e==="idle"){let s=r?.workload;return{workload:typeof s=="string"?s:`${s?.project??"?"}/${s?.name??"?"}`,type:typeof s=="object"?s?.display_type:void 0,idle_cards:r?.idle_cards,idle_hours:r?.idle_seconds===void 0?void 0:Math.round(r.idle_seconds/3600*10)/10,source:r?.source}}return r}var I={"workloads/items":["project","name","status","xpu_count","pods_ready","pods_total","pool"],"overview/queued":["project","name","xpu_count","pool","created_at"],"overview/projects":["slug","quota_cards","used_cards","running","queued"],"overview/idle":["workload","idle_cards","idle_hours","source"],"overview/nodes":["name","status","cards_total","cards_allocated","cards_free"],"nodes/items":["name","status","pool","cards_total","cards_allocated","cards_free","avg_util"],"projects/items":["slug","name","member_count","created_at"],"projects/usage":["project","pool","guaranteed","used","over_guarantee"],"projects/storage":["project","pool","limit_gib","used_gib","used_pct","remaining_gib"],"projects/queues":["project","queue","pool","admitted","reason"],"cluster/items":["name","status","ready","total","version"]},H={"overview/projects":{dimension:"slug",measures:["quota_cards","used_cards"]},"overview/idle":{dimension:"workload",measures:["idle_cards"]},"overview/queued":{dimension:"name",measures:["xpu_count"]},"overview/nodes":{dimension:"name",measures:["cards_allocated","cards_free"]},"workloads/items":{dimension:"name",measures:["xpu_count"]},"nodes/items":{dimension:"name",measures:["cards_allocated","cards_free"]},"projects/items":{dimension:"slug",measures:["member_count"]},"projects/usage":{dimension:"project",measures:["guaranteed","used"]},"projects/storage":{dimension:"project",measures:["used_gib","remaining_gib"]},"projects/queues":{dimension:"project",measures:[]},"cluster/items":{dimension:"name",measures:["ready","total"]}},B={project:"\u9879\u76EE",project_name:"\u9879\u76EE\u540D",slug:"\u9879\u76EE",name:"\u540D\u79F0",workload:"\u8D1F\u8F7D",status:"\u72B6\u6001",phase:"\u9636\u6BB5",pool:"\u8D44\u6E90\u6C60",type:"\u7C7B\u578B",display_type:"\u7C7B\u578B",xpu_count:"\u5361\u6570",quota_cards:"\u914D\u989D(\u5361)",used_cards:"\u5DF2\u7528(\u5361)",running:"\u8FD0\u884C",queued:"\u6392\u961F",idle_cards:"\u95F2\u7F6E(\u5361)",idle_hours:"\u95F2\u7F6E(\u5C0F\u65F6)",source:"\u5224\u5B9A\u6765\u6E90",pods_ready:"\u5C31\u7EEA Pod",pods_total:"\u603B Pod",member_count:"\u6210\u5458\u6570",cards_total:"\u603B\u5361",cards_allocated:"\u5DF2\u5206\u914D",cards_free:"\u7A7A\u95F2",cards_healthy:"\u5065\u5EB7",cards_allocatable:"\u53EF\u5206\u914D",avg_util:"\u5E73\u5747\u5229\u7528\u7387",cpu_cores:"CPU \u6838",ready:"\u5C31\u7EEA",total:"\u603B\u6570",version:"\u7248\u672C",arch:"\u67B6\u6784",os:"\u7CFB\u7EDF",guaranteed:"\u4FDD\u5E95(\u5361)",used:"\u5B9E\u9645\u7528(\u5361)",over_guarantee:"\u8D85\u51FA\u4FDD\u5E95",limit_gib:"\u914D\u989D(GiB)",used_gib:"\u5DF2\u7528(GiB)",used_pct:"\u5DF2\u7528\u6BD4\u4F8B",remaining_gib:"\u5269\u4F59(GiB)",persistent_gib:"\u6301\u4E45(GiB)",runtime_gib:"\u8FD0\u884C\u65F6(GiB)",queue:"\u961F\u5217",admitted:"\u5DF2\u51C6\u5165",reason:"\u539F\u56E0",created_at:"\u521B\u5EFA\u65F6\u95F4",admitted_at:"\u51C6\u5165\u65F6\u95F4",started_at:"\u5F00\u59CB\u65F6\u95F4",last_activity_at:"\u6700\u8FD1\u6D3B\u52A8",updated_at:"\u66F4\u65B0\u65F6\u95F4",internal_ip:"\u5185\u7F51 IP",owner_name:"\u6240\u6709\u8005",image:"\u955C\u50CF",device_type:"\u8BBE\u5907\u578B\u53F7",chip_model:"\u82AF\u7247"},J=t=>t===null||["string","number","boolean"].includes(typeof t);function X(t,e){if(e==null||e==="")return{text:"\u2014",align:"right"};if(typeof e=="boolean")return{text:e?"\u662F":"\u5426",align:"center"};if(typeof e=="number")return t.endsWith("_bytes")?{text:Y(e),align:"right"}:t.endsWith("_pct")||t==="avg_util"?{text:`${Number(e.toPrecision(3))}%`,align:"right"}:{text:String(Number(e.toPrecision(6))),align:"right"};if(t.endsWith("_at")&&/^\d{4}-\d{2}-\d{2}T/.test(e)){let r=new Date(e);if(Number.isFinite(r.getTime())){let s=n=>String(n).padStart(2,"0");return{text:`${s(r.getMonth()+1)}-${s(r.getDate())} ${s(r.getHours())}:${s(r.getMinutes())}`,align:"right"}}}return{text:String(e),align:"left"}}function Y(t){let e=["B","KiB","MiB","GiB","TiB"],r=t,s=0;for(;Math.abs(r)>=1024&&s<e.length-1;)r/=1024,s+=1;return`${Number(r.toPrecision(3))} ${e[s]}`}function T(t,e,r){let s=y[t];if(!s)return;let n=r&&s[r]?r:Object.keys(s)[0],a=s[n]?.at(e);if(Array.isArray(a))return{source:n,rows:a.map(i=>Q(t,n,i))}}function Z(t){let e=[];for(let r of t.slice(0,20))for(let[s,n]of Object.entries(r??{}))J(n)&&!e.includes(s)&&e.push(s);return e}function ee(t,e,r,s){let n=new Set(r),a=Array.isArray(s)?s:[],i=a.filter(c=>!n.has(c)),o=a.filter(c=>n.has(c));if(o.length>0)return{columns:o,dropped:i};let u=I[`${t}/${e}`];if(u){let c=u.filter(m=>n.has(m));if(c.length>0)return{columns:c,dropped:i}}return{columns:r.slice(0,8),dropped:i}}var M=50,N=t=>Number.isFinite(t)?Math.max(1,Math.min(200,t)):M;function te(t,e,r){if(!e)return t;let s=r===!1?1:-1;return[...t].sort((n,a)=>{let i=n?.[e],o=a?.[e];return typeof i=="number"&&typeof o=="number"?(i-o)*s:String(i??"").localeCompare(String(o??""))*s})}function re(t,e,r={}){let s=S.includes(r.kind)?r.kind:"line";if(s==="line"||s==="area"){let d=v(e,{query:t,group:r.group,series:r.series});return{kind:s,series:d.series,group:d.group,note:d.note}}let n=T(t,e,r.source);if(!n){let d=Object.keys(y).join(" / ");return{kind:s,columns:[],rows:[],note:`${t} \u6CA1\u6709\u53EF\u5217\u8868\u5316\u7684\u6570\u636E\uFF08\u652F\u6301\uFF1A${d}\uFF09`}}let{source:a,rows:i}=n;if(i.length===0)return{kind:s,source:a,columns:[],rows:[],note:"\u8FD9\u6B21\u67E5\u8BE2\u6CA1\u6709\u8FD4\u56DE\u4EFB\u4F55\u884C"};let o=N(r.limit),u=te(i,r.sort,r.desc),c=u.slice(0,o),m=u.length>c.length?u.length-c.length:0,f=[];m>0&&f.push(`\u53EA\u663E\u793A\u524D ${c.length} \u884C\uFF0C\u53E6\u6709 ${m} \u884C\u672A\u663E\u793A`);let h=Z(i);if(r.sort&&!h.includes(r.sort)&&f.push(`\u6CA1\u6709 ${r.sort} \u8FD9\u4E00\u5217\uFF0C\u6392\u5E8F\u6CA1\u6709\u751F\u6548\uFF08\u53EF\u7528\uFF1A${h.join(" / ")}\uFF09`),s==="bar"){let d=h.filter(l=>i.slice(0,20).some(w=>typeof w?.[l]=="number")),g=r.dimension&&h.includes(r.dimension)?r.dimension:["slug","name","workload","project"].find(l=>h.includes(l))??h[0],p=Array.isArray(r.measures)?r.measures:[],k=p.filter(l=>d.includes(l)),O=p.filter(l=>!d.includes(l)),j=k.length>0?k.slice(0,5):d.slice(0,2);return j.length===0?{kind:s,source:a,categories:[],series:[],note:"\u8FD9\u7EC4\u6570\u636E\u91CC\u6CA1\u6709\u53EF\u7528\u4F5C\u67F1\u9AD8\u7684\u6570\u503C\u5217"}:(O.length>0&&f.push(`\u6CA1\u6709 ${O.join("/")} \u8FD9\u4E9B\u6570\u503C\u5217\uFF0C\u753B\u7684\u662F ${j.join("/")}\uFF08\u53EF\u7528\uFF1A${d.join(" / ")}\uFF09`),{kind:s,source:a,categories:c.map(l=>String(l?.[g]??"\u2014")),series:j.map(l=>({key:l,name:B[l]??l,data:c.map(w=>typeof w?.[l]=="number"?w[l]:null)})),dimension:g,note:f.length>0?f.join("\uFF1B"):void 0})}let{columns:b,dropped:E}=ee(t,a,h,r.columns);return E.length>0&&f.push(`\u6CA1\u6709 ${E.join("/")} \u8FD9\u4E9B\u5217\uFF0C\u663E\u793A\u7684\u662F ${b.join("/")}\uFF08\u53EF\u7528\uFF1A${h.join(" / ")}\uFF09`),{kind:"table",source:a,columns:b.map(d=>({key:d,label:B[d]??d})),rows:c.map(d=>b.map(g=>{let p=X(g,d?.[g]);return{text:p.text,align:p.align}})),total:i.length,note:f.length>0?f.join("\uFF1B"):void 0}}function se(t){return{kind:"line",...t?.view??{}}}function _(t,e={}){let r=S.includes(e.kind)?e.kind:"line";if(U(r))return{kind:r};let s=y[t],n={kind:r};s&&(n.source=e.source&&s[e.source]?e.source:Object.keys(s)[0]);let a=r==="bar"?H[`${t}/${n.source}`]:void 0,i=e.dimension??a?.dimension;i!==void 0&&(n.dimension=i),e.sort!==void 0&&(n.sort=e.sort);let o=N(e.limit);o!==M&&(n.limit=o),e.desc===!1&&(n.desc=!1);let u=r==="table"?I[`${t}/${n.source}`]:void 0,c=Array.isArray(e.columns)&&e.columns.length>0?e.columns:u;Array.isArray(c)&&c.length>0&&(n.columns=[...c]);let m=Array.isArray(e.measures)&&e.measures.length>0?e.measures:a?.measures;return Array.isArray(m)&&m.length>0&&(n.measures=[...m]),n}var U=t=>t===void 0||t==="line"||t==="area";var L=["1h","6h","24h","7d"],P=16,ce=["kind","source","dimension","sort","desc","limit"];function C(t){if(t?.range===void 0)return t??{};let{range:e,...r}=t;return r}function x(t){let e=Object.entries(t.params??{}).filter(([o])=>o!=="range").filter(([,o])=>o!==void 0).sort(([o],[u])=>o<u?-1:o>u?1:0).map(([o,u])=>`${o}=${String(u)}`).join("&"),r=Array.isArray(t.series)&&t.series.length>0?[...t.series].sort().join(","):"",s=_(t.query,t.view),n=ce.filter(o=>s[o]!==void 0).map(o=>`${o}=${String(s[o])}`).join(","),a=Array.isArray(s.columns)?s.columns.join("+"):"",i=Array.isArray(s.measures)?s.measures.join("+"):"";return`${t.query}?${e}#${r}@${n}|${a}|${i}`}function q(t,e){let r=Object.entries({...t.params??{},range:e}).filter(([s])=>s!=="group").filter(([,s])=>s!==void 0).sort(([s],[n])=>s<n?-1:s>n?1:0).map(([s,n])=>`${s}=${String(n)}`).join("&");return`${t.query}?${r}`}var V=class{#n;#e=[];#s=new Map;#t;#o=new Set;#r="24h";#a=0;#m=Promise.resolve();constructor(e=D(R(),"pins.json")){this.#n=e,this.#t=this.#f()}#u(e){let r=this.#m.then(e);return this.#m=r.catch(()=>{}),r}async#f(){try{let e=JSON.parse(await ne(this.#n,"utf8"));L.includes(e?.range)&&(this.#r=e.range);let r=Array.isArray(e?.pins)?e.pins:[],s=new Set,n=!1;this.#e=r.map(a=>{let i=a;i.params?.range!==void 0&&(n=!0,i={...i,params:C(i.params)});let o=_(i.query,i.view);return JSON.stringify(o)!==JSON.stringify(i.view)&&(n=!0,i={...i,view:o}),i}).filter(a=>{let i=x(a);return s.has(i)?(n=!0,!1):(s.add(i),!0)}),(n||this.#e.length!==r.length)&&await this.#i()}catch{this.#e=[]}}async#i(){await oe(D(this.#n,".."),{recursive:!0});let e=`${JSON.stringify({range:this.#r,pins:this.#e},void 0,2)}
+`,r=`${this.#n}.${G().slice(0,8)}.tmp`;await ie(r,e,"utf8");try{await ae(r,this.#n)}catch(s){throw await ue(r,{force:!0}),s}}async ready(){return await this.#t,this}async list(){return await this.#t,this.#e.map(e=>({...e,latest:this.#s.get(e.id)}))}async range(){return await this.#t,this.#r}async setRange(e,r,s){if(await this.#t,!L.includes(e))throw new Error(`\u4E0D\u652F\u6301\u7684\u65F6\u95F4\u8303\u56F4\uFF1A${e}`);if(!await this.#u(async()=>e===this.#r?!1:(this.#r=e,this.#s.clear(),this.#a+=1,await this.#i(),!0)))return{range:e,results:[]};let a=await this.refreshAll(r,s);return this.#l(),{range:e,results:a}}async findEquivalent(e){await this.#t;let r=x(e);return this.#e.find(s=>x(s)===r)}async add(e,r){return await this.#t,this.#u(async()=>{let s=C(e.params),n=await this.findEquivalent({...e,params:s});if(n!==void 0)return{pin:n,created:!1};if(this.#e.length>=P){let i=new Error(`\u56FA\u5B9A\u56FE\u8868\u5DF2\u8FBE\u4E0A\u9650\uFF08${P} \u5F20\uFF09\uFF0C\u5148\u53D6\u6D88\u56FA\u5B9A\u4E00\u5F20\u518D\u9489\u65B0\u7684`);throw i.code="pin_limit",i.status=409,i}let a={id:G(),title:e.title||"\u672A\u547D\u540D\u56FE\u8868",query:e.query,params:s,series:e.series,view:_(e.query,e.view),workspace:e.workspace,provenance:{source:"runnpu",createdAt:new Date().toISOString(),...r}};return this.#e=[...this.#e,a],await this.#i(),this.#l(),{pin:a,created:!0}})}async remove(e){return await this.#t,this.#u(async()=>{let r=this.#e.length;return this.#e=this.#e.filter(s=>s.id!==e),this.#s.delete(e),this.#e.length===r?!1:(await this.#i(),this.#l(),!0)})}async refresh(e,r,s){let n=await this.#c(e,r,s);return n===void 0?{id:e.id,at:Date.now(),ok:!1,stale:!0,error:"\u65F6\u95F4\u8303\u56F4\u521A\u521A\u5207\u6362\uFF0C\u8FD9\u6B21\u53D6\u6570\u5DF2\u4F5C\u5E9F\uFF0C\u7A0D\u540E\u4F1A\u6309\u65B0\u7A97\u53E3\u91CD\u65B0\u53D6\u6570"}:this.#d(e,n)}async#c(e,r,s){let n=this.#a,a=$.includes(e.query)?{...e.params,range:this.#r}:{...e.params},i;try{i={ok:!0,data:await r.query(e.workspace,e.query,a,s)}}catch(o){i={ok:!1,error:o.message,code:o.code}}return n===this.#a?i:void 0}#d(e,r){let s={id:e.id,at:Date.now(),...r};return this.#s.set(e.id,s),s}async refreshAll(e,r){await this.#t;let s=new Map;for(let a of this.#e){let i=q(a,this.#r),o=s.get(i);o?o.push(a):s.set(i,[a])}return(await Promise.all([...s.values()].map(async a=>{let i=await this.#c(a[0],e,r);return i===void 0?[]:a.map(o=>this.#d(o,i))}))).flat()}async requestCount(){return await this.#t,new Set(this.#e.map(e=>q(e,this.#r))).size}async refreshMissing(e,r){await this.#t;let s=this.#e.filter(i=>this.#s.get(i.id)===void 0);if(s.length===0)return[];let n=new Map;for(let i of s){let o=q(i,this.#r),u=n.get(o);u?u.push(i):n.set(o,[i])}return(await Promise.all([...n.values()].map(async i=>{let o=await this.#c(i[0],e,r);return o===void 0?[]:i.map(u=>this.#d(u,o))}))).flat()}latestOf(e){return this.#s.get(e)}onChange(e){return this.#o.add(e),()=>this.#o.delete(e)}#l(){for(let e of[...this.#o])try{e()}catch{}}};export{S as a,F as b,T as c,re as d,se as e,_ as f,U as g,L as h,P as i,x as j,V as k};