[heydryft]

Adds jobs to the CLI: one command on its own GPU machine, its output streamed back, and the machine stopped when the command ends (RUN-486).

Commands

• runcrate run --gpu A4000 [--gpu-count 1] [--volume <id>] [--max-runtime 2h] [--template t] [--image img] [--env K=V ...] [--name n] [--region r] [-d/--detach] -- <command...>
  • One argument after -- is sent as is (-- 'make && ./train'); several are joined with spaces, quoting the ones that need it (-- python -c 'print(1)' runs as typed). Flags stop at the first argument, so the command's own flags pass through.
  • --max-runtime: Go duration or bare minutes, rounded up to whole minutes, 1m to 24h, default 1h (printed with a hint to change it).
  • --env KEY with no value copies it from the local shell.
  • Prints the job, then follows logs (follow=true long polls with the cursor), writing each chunk to stdout verbatim; status notes and the final line (status, exit code, runtime, cost) go to stderr.
  • Exit code: succeeded 0, failed the command's code (or 1), timed_out 124, cancelled 130.
  • Ctrl-C detaches and never cancels: prints runcrate jobs logs -f <id> and runcrate jobs cancel <id>, exits 0. --detach prints the job ID and returns.
  • Network errors, 429 and 5xx while following are retried with backoff (1s up to 15s); the client is rebuilt before each poll so long follows pick up refreshed sign-in tokens.
• runcrate jobs list [--status running,queued] [--limit 20] [-q] [--json]
• runcrate jobs status <id> [--json]: exit code, error, cost, instance, env keys, volume, timestamps.
• runcrate jobs logs <id> [-f]: without -f, pages from after=0 until the cursor stops; with -f, follows like run (same detach and exit codes).
• runcrate jobs cancel <id>

API client

• internal/api/jobs.go: CreateJob, ListJobs, GetJob, GetJobLogs, CancelJob. Follow requests carry a 40 s deadline through the request context; other calls keep the client's timeout.
• checkHTTPStatus now returns a typed *StatusError (same message text as before) with code and details. A 402 BUDGET_REACHED reads "this environment's monthly budget is reached; an owner or admin can raise it in the dashboard". A launch failure after creation names details.job_id.

Before merging / releasing

• Needs the app PR deployed first: feat/jobs in aeonmindai/runcrate_app (RUN-486), which adds /api/v1/jobs.
• Needs a release (vX.Y.Z tag) after merge so runcrate update and the installers pick it up.

Tests
go vet ./..., go test ./... (also -race -count=5 on api and commands) and go build ./... pass. New tests use httptest fakes: max-runtime parsing, command joining, the create body (gpu_type, max_runtime_minutes, env object, storage_id), the follow loop (chunks in order across responses, cursor honored, null-seq marker printed, retry on 503, stop on 404, stop on done), exit codes per status, and Ctrl-C detach aborting an open long poll without calling /cancel.

[heydryft]

App side: aeonmindai/runcrate_app#169 (adds /api/v1/jobs). Merge and deploy that first, then release this.