clu
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| execA | 在远程连接上执行单条 shell 命令。危险命令与集群登录节点重操作会被服务端拒绝。 Args: connection: 连接 label(见 connection_list) command: 要执行的 shell 命令 timeout: 超时秒数(默认 60) |
| jobsB | 列出 Slurm 集群上当前用户的作业(job_id/状态/分区/耗时/资源,结构化)。 |
| statusB | Slurm 集群概览:分区(节点/CPU/内存/GPU)与队列状态计数。 |
| resourcesA | 查询集群资源余量:每分区 CPU/GPU 的总量/已占/空闲与节点数。 一开始就限定查看范围(agent 与人类同款口径): flt=cpu_free(有空闲 CPU 的节点)、flt=cpu_gpu_free(CPU 和 GPU 都有空闲)、 flt=fully_idle(CPU 与 GPU 全空闲);gpu_only=true 只看有 GPU 的节点(可叠加)。 partition 限定分区。默认紧凑(分区聚合);带任一过滤或 include_nodes 才返回节点明细。不要自己拼 sinfo/scontrol。 Args: connection: cluster 连接 label partition: 只看该分区 flt: cpu_free | cpu_gpu_free | fully_idle gpu_only: 只看有 GPU 的节点(与其他条件叠加) include_nodes: 返回全部节点明细(大集群建议配合过滤) |
| submitB | 提交 Slurm 批处理作业:脚本文本经 base64 上传后 sbatch,返回 job_id。 |
| logB | 读取作业 stdout(自动定位 StdOut 路径并 tail)。 |
| cancelA | 取消指定作业(scancel 单个作业)。scancel --all 被安全门禁止。 |
| ledger_queryB | 检索执行留痕(跨会话记忆):按连接/类型/关键词/作业号过滤,时间倒序。 |
| allocA | 抢占资源并直接执行(绝不只返回命令):tmux 内 salloc + srun bash, 分配后 attach 即入计算节点 shell,断线不丢资源。 抢占选择自动完成(无需先查询):resource 缺省 = GPU 优先、无空闲 GPU 回退 CPU; resource=gpu / cpu 只抢该类(无则失败并附当前余量摘要); partition 显式指定则直接使用。返回 granted/pending/failed 与 job_id/node。 Args: connection: cluster 连接 label purpose: 会话用途名(如 train-a100),用于 tmux 命名与检索 resource: gpu | cpu;缺省为自动(GPU 优先回退 CPU) gpus: GPU 卡数(默认 1;resource 含 gpu 时生效) cpus: CPU 核数(cpu 资源必给则默认 4;也可与 GPU 同时指定) partition: 显式分区(跳过自动选择) wait: 等待授予秒数(超时会话保留,session_info 随时查) |
| session_listB | 列出 tmux 托管会话(alloc 资源会话),含所属连接/用途/创建时间。 |
| session_infoB | 查看某个托管会话详情:分配状态、job_id、节点、日志尾部。 |
| session_releaseB | 释放托管会话:杀 tmux → salloc 退出 → 资源归还集群。 |
| uploadB | 上传本地文件到远端(SFTP,远端父目录自动创建)。 |
| downloadC | 下载远端文件到本地(SFTP,本地父目录自动创建)。 |
| connection_listA | 列出全部受管连接及连接池健康状态(label/kind/host/state)。 |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 15 tools
Most tools have clearly distinct purposes (exec vs submit vs alloc, session_* vs jobs, upload vs download). Minor overlap exists between status and resources for cluster overview, and between log and session_info for log tail, but descriptions help differentiate.
Naming mixes short imperative verbs (exec, submit, alloc, cancel, log) with prefixed noun patterns (ledger_query, session_list, connection_list). The set is readable but not a single consistent verb_noun convention.
15 tools is well within the typical 3–15 range and each tool maps to a distinct capability (remote exec, job lifecycle, resource query, session management, file transfer, connection listing). No obvious redundant tools.
Core workflows are covered: command execution, job submission/monitoring/cancel, resource allocation, session management, and file transfer. Minor gaps include no dedicated job detail/history tool (e.g., sacct) and no connection create/delete management, though workarounds via exec exist.