Projects and batches¶

Argus organises events in a three-level hierarchy:

Project ─┬── Batch ─┬── Job ─┬── Epoch metric
         │          │        ├── Resource snapshot
         │          │        └── Log line
         │          └── Batch metadata (env_snapshot, owner, command)
         └── Project-level settings (notification recipients, share links)

A project is what the SDK passes as source_project=. A batch is what one with Reporter(...) opens. Jobs are runs inside a batch (several for a sweep, one for a single training).

Projects¶

Listing & creating¶

Sidebar → Projects. Members see projects they belong to; admins see everything. New project asks for a slug (immutable) and display name.

Project detail¶

Tabs:

Tab	What you find
Overview	Recent batches, summary counters, owners
Batches	Filterable batch list
Settings	Display name, notification recipients, share links, members

Visibility¶

A project is private (members only) or public (read-only to anyone, including unauthenticated visitors). Members can be added by the project owner or any admin.

Notification recipients¶

Each project routes its email notifications to a list of recipients, not just the creator (#116). Edit in Settings → Recipients. See Notifications for triggers.

Batches¶

Listing¶

GET /api/batches is the underlying endpoint. The Batches page filters by project, status, host, owner, and tag. URLs are deep-linkable.

Batch detail¶

The header shows status, elapsed time, ETA (when computable), owner, host, and the recorded command. The body is dominated by the JobMatrix (see Job matrix) for sweeps, or a single-job summary card.

Actions on a batch¶

Action	Endpoint	Effect
Star	(front-end pin store)	Pinned to your starred list
Stop	`POST /api/batches/{id}/stop`	Sets a stop flag; SDK exits cleanly within ~10 s
Rerun	`POST /api/batches/{id}/rerun`	Asks `argus-agent` on the origin host to relaunch the recorded command
Share	(`shares` API)	Mints a read-only public link
Delete	`DELETE /api/batches/{id}`	Soft delete; sweeper purges per retention policy
Bulk delete	`POST /api/batches/bulk-delete`	Multi-select
Export CSV	`GET /api/batches/{id}/export.csv`	Per-job final metrics

Stop is always available to the owner. Rerun requires an online agent on the origin host (the agent is shipped by the Sibyl package — see Argus Agent).

Health & ETA¶

GET /api/batches/{id}/health flags a stalled batch when its newest event is older than ARGUS_STALL_TIMEOUT_MIN minutes (default 15). The UI surfaces this as an amber dot.

GET /api/batches/{id}/eta and /jobs/eta-all compute ETAs from completed epoch durations. The UI hides the value until enough data points have been collected to keep the estimate stable.