feat(scheduler-agent): Phase 1 write operations — submit/cancel/hold/release + templates

Add write operations to the scheduler-agent MCP server (Phase 1):

- schedulerclient: SubmitJob/CancelJob/HoldJob/ReleaseJob + JobSpec/
  SubmitJobResult/JobControlRequest/ControlResult types + doPost helper
  (HTTP 200 + envelope-code dual check). TDD: 12 new tests.
- mcp-server: 4 write tools (scheduler_submit/cancel/hold/release) with
  destructive annotations; scheduler_submit has dry_run review mode +
  general_serial (通用串行脚本) job template registry. Fix the 6 read-only
  tools' misleading destructive annotations. TDD: 14 new tests.

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
lei.chen 2026-08-01 09:54:38 +00:00
parent 2e299f4a76
commit 35416d5cf4
9 changed files with 1447 additions and 6 deletions

View file

@ -29,11 +29,14 @@ func main() {
s := mcps.NewMCPServer(
"scheduler-agent",
"0.1.0",
"0.2.0",
mcps.WithResourceCapabilities(true, true),
mcps.WithLogging(),
)
// Read-only tools are annotated as such so the MCP client does not prompt
// for confirmation on them (NewTool defaults to destructive=true).
// ── scheduler_list_jobs ──
s.AddTool(mcp.NewTool("scheduler_list_jobs",
mcp.WithDescription("List jobs in the cluster. Supports filtering by queue, owner, status, and pagination."),
@ -42,6 +45,8 @@ func main() {
mcp.WithString("status", mcp.Description("Filter by status: pending|running|succeeded|failed|cancelled|suspended|unknown")),
mcp.WithNumber("limit", mcp.Description("Max results to return")),
mcp.WithString("page_token", mcp.Description("Pagination token from a prior response")),
mcp.WithReadOnlyHintAnnotation(true),
mcp.WithDestructiveHintAnnotation(false),
), app.HandleListJobs)
// ── scheduler_get_job ──
@ -51,16 +56,22 @@ func main() {
mcp.Description("Job ID"),
mcp.Required(),
),
mcp.WithReadOnlyHintAnnotation(true),
mcp.WithDestructiveHintAnnotation(false),
), app.HandleGetJob)
// ── scheduler_list_queues ──
s.AddTool(mcp.NewTool("scheduler_list_queues",
mcp.WithDescription("List all queues/partitions and their current state (active, inactive, draining)."),
mcp.WithReadOnlyHintAnnotation(true),
mcp.WithDestructiveHintAnnotation(false),
), app.HandleListQueues)
// ── scheduler_list_nodes ──
s.AddTool(mcp.NewTool("scheduler_list_nodes",
mcp.WithDescription("List all compute nodes with their resource usage and status."),
mcp.WithReadOnlyHintAnnotation(true),
mcp.WithDestructiveHintAnnotation(false),
), app.HandleListNodes)
// ── scheduler_job_logs ──
@ -72,13 +83,68 @@ func main() {
),
mcp.WithNumber("tail", mcp.Description("Number of lines from the end to return (0 = all)")),
mcp.WithNumber("offset_bytes", mcp.Description("Byte offset to start reading from")),
mcp.WithReadOnlyHintAnnotation(true),
mcp.WithDestructiveHintAnnotation(false),
), app.HandleJobLogs)
// ── scheduler_cluster_metrics ──
s.AddTool(mcp.NewTool("scheduler_cluster_metrics",
mcp.WithDescription("Get a snapshot of overall cluster resource usage and utilization."),
mcp.WithReadOnlyHintAnnotation(true),
mcp.WithDestructiveHintAnnotation(false),
), app.HandleClusterMetrics)
// ── write tools (Phase 1) ──
// Write tools are annotated destructive so the MCP client (e.g. Claude
// Code) prompts the user for confirmation before executing them — the
// design doc's "写操作默认 require confirm" at the layer MCP supports it.
// ── scheduler_submit ──
s.AddTool(mcp.NewTool("scheduler_submit",
mcp.WithDescription("Submit a job. REVIEW REQUIRED: call with dry_run=true first to inspect the resolved job spec, then re-call with dry_run=false to actually submit."),
mcp.WithString("template", mcp.Description("Job template (default 'general_serial' 通用串行脚本)."),
mcp.Enum(internal.TemplateGeneralSerial), mcp.DefaultString(internal.TemplateGeneralSerial)),
mcp.WithBoolean("dry_run", mcp.Description("Review mode: validate and return the resolved job spec without submitting."), mcp.DefaultBool(false)),
mcp.WithString("command", mcp.Description("Shell command to run (required)."), mcp.Required()),
mcp.WithString("name", mcp.Description("Job name; array syntax 'myjob[1-10]' supported.")),
mcp.WithString("queue", mcp.Description("Queue (default 'normal').")),
mcp.WithString("project", mcp.Description("Project name.")),
mcp.WithNumber("cores", mcp.Description("Number of processors.")),
mcp.WithString("walltime", mcp.Description("Walltime limit 'HH:MM', e.g. '02:30'.")),
mcp.WithString("memory", mcp.Description("Memory limit, e.g. '4 GB', '2048 MB'.")),
mcp.WithString("cwd", mcp.Description("Working directory.")),
mcp.WithString("env", mcp.Description("Comma-separated KEY=VALUE environment variables.")),
mcp.WithString("out_file", mcp.Description("Standard output file.")),
mcp.WithString("err_file", mcp.Description("Standard error file.")),
mcp.WithString("description", mcp.Description("Job description.")),
mcp.WithReadOnlyHintAnnotation(false),
mcp.WithDestructiveHintAnnotation(true),
), app.HandleSubmitJob)
// ── scheduler_cancel ──
s.AddTool(mcp.NewTool("scheduler_cancel",
mcp.WithDescription("Cancel (kill) a job by ID. Destructive — you will be asked to confirm before execution."),
mcp.WithString("id", mcp.Description("Job ID to cancel"), mcp.Required()),
mcp.WithReadOnlyHintAnnotation(false),
mcp.WithDestructiveHintAnnotation(true),
), app.HandleCancelJob)
// ── scheduler_hold ──
s.AddTool(mcp.NewTool("scheduler_hold",
mcp.WithDescription("Hold (suspend/stop) a job by ID. Destructive — you will be asked to confirm before execution."),
mcp.WithString("id", mcp.Description("Job ID to hold"), mcp.Required()),
mcp.WithReadOnlyHintAnnotation(false),
mcp.WithDestructiveHintAnnotation(true),
), app.HandleHoldJob)
// ── scheduler_release ──
s.AddTool(mcp.NewTool("scheduler_release",
mcp.WithDescription("Release (resume) a held job by ID. Destructive — you will be asked to confirm before execution."),
mcp.WithString("id", mcp.Description("Job ID to release"), mcp.Required()),
mcp.WithReadOnlyHintAnnotation(false),
mcp.WithDestructiveHintAnnotation(true),
), app.HandleReleaseJob)
log.Printf("Starting scheduler-agent MCP server (base-url: %s)", baseURL)
if err := mcps.ServeStdio(s); err != nil {
log.Fatalf("server error: %v", err)