CLI

Read jobs¶

List the jobs under a root, open one, compare several, list the workspaces.

The reading commands

nullius runs      list <root>... [--json]
nullius runs      show <root> <job id> [<trial>] [--json]
nullius compare   <root> <job id> <job id>... [--matched] [--json]
nullius workspace list [--home DIR] [--json]
nullius standard traces <dir> [--dry-run]

Inspect and compare saved jobs¶

List the jobs under a root.
List
```
nullius runs list ./jobs
```

Open one; compare two on their shared tasks.

Show and compare

nullius runs show ./jobs run-a
nullius compare ./jobs run-a run-b --matched

--json: one document on stdout.

JSON

nullius runs show ./jobs run-a --json > /tmp/run-a.json

Reads only

None launches an evaluation.

Command	Gives
`nullius runs list <root>...`	The runs, as `/runs` lists them
`nullius runs show <root> <id>`	A job: trials, outcomes, counts, solve rates
`nullius runs show <root> <job> <trial>`	A trial: outcome, checks, usage, exception, steps. Tool arguments and observations stay on the page
`nullius compare <root> <id> <id>... [--matched]`	`/compare`'s cohort: whole, or matched on the tasks every job ran
`nullius workspace list [--home DIR]`	The workspaces, as `/workspaces` lists them

Read results

The same reads on the pages.
Exit codes and output

What every command promises.

Give past standard runs a trace¶

A standard run writes agent/trace.json as it ends. nullius standard traces gives an earlier job the same file, from what its trials kept. It only adds that file: an existing trace is never replaced, and Harbor jobs and controls are left alone.

See what it would write, then write it

nullius standard traces ./jobs --dry-run
nullius standard traces ./jobs

A trial that ended early kept no ACP events: its trace has the model calls, and its tool calls at their start, with no duration.