Job detail¶

The job detail page is the leaf view. It is a single scrollable layout. The route is /jobs/{batch_id}/{job_id} in the SPA, served by JobDetail.vue.

Layout¶

Telemetry strip — status / elapsed / GPU util / GPU mem peak / latest loss
Loss curve — ECharts line plot, fed by the SSE stream
Log tail — the most recent log_line rows, if any
Action bar — Stop · Rerun · Share · Copy command

Telemetry strip¶

Compact tiles populated from the latest events:

Tile	Source
Status	latest event
Elapsed	`job_start` to last activity
GPU util	latest `resource_snapshot` (mean across visible GPUs)
GPU mem peak	rolling max from snapshots
Latest loss	last `job_epoch` train/val loss

The dashboard idle detector flags this job when GPU util has stayed below 5% for ARGUS_IDLE_JOB_THRESHOLD_MIN minutes (default 10) — typical "dataloader stalled, GPU has been idle for an hour" footgun.

Loss curve¶

ECharts line plot streamed via SSE. By default train_loss and val_loss are drawn (any **extra numeric metric on job.epoch becomes another toggleable series).

Log tail¶

GET /api/batches/{batch_id}/log-lines powers this panel. Logs are not stored forever — the retention sweeper purges log_line events older than ARGUS_RETENTION_LOG_LINE_DAYS days (default 14). Summary rows survive indefinitely.

Action bar¶

Button	Endpoint	Notes
Stop	`POST /api/batches/{batch_id}/stop`	Sets a flag; SDK reads on next poll (10 s default)
Rerun	`POST /api/batches/{batch_id}/rerun`	Sends a `kind=rerun` to the origin host's `argus-agent` (shipped in Sibyl). Disabled if no agent online.
Share	shares API	Mints a public read-only link (covers the whole batch)
Copy command	client-side	Copies the recorded `env_snapshot.command`

Env snapshot¶

An Environment panel below the action bar shows what the SDK captured on batch_start:

Git commit (short SHA)
Command line, working directory, hostname

This is what gets re-executed on a rerun.

Failed jobs¶

If the run ended in job_failed, the failure reason is shown above the log tail.