AK-Benchbenchmark report

docs/CLI.md

CLI

Every command from the plan's CLI sketch is implemented. Global options:

--config/-c, --project-root, --verbose/-v, --quiet/-q, --version.

Nearly every command accepts --json; where a command can fail, a non-zero

exit code is part of the interface.

Interface facts

CommandNotes
akbench versionAK-Bench and key dependency versions
akbench initcreate directories, configuration and schema
akbench doctor [--strict]environment diagnostics; exit 1 on failures
akbench config show [--all] / config validateeffective configuration, paths, database URL
akbench descriptor list/validate/showdescriptor registry
akbench category listtask categories and whether they are enabled
akbench dataset list/validate/show/preparetasks, repositories, repeatable workspaces
akbench agent list/inspect/runregistered agents and one agent run
akbench runtime list/health/models/config/stateruntimes and their configuration triad
akbench evaluate taskjudge a produced workspace
akbench run single TASK --agent A [--store]one orchestrated run
akbench experiment list/validate/planvalidate and expand an experiment
akbench experiment run FILE [--store] [--only N]execute the matrix
akbench plugin list/status/verifyplugin descriptors and their observed state
akbench metrics catalog/runthe metric catalog and the metrics of recorded runs
akbench knowledge plan/readinessthe memory plan and the readiness gate
akbench history experiments/runs/show/metrics/verifyread the recorded history back
akbench report all/experiment/taskwrite the static HTML reports
akbench export experiment UIDwrite one experiment out, screened

Conventions that are part of the interface

  • --json prints a machine-readable document on stdout and nothing else that

is not part of it; human output goes through the console renderer.

  • A command that cannot do its job exits non-zero. A command that can do part

of it says which part it could not (e.g. history verify lists what does not

match, then exits 1).

  • Nothing is stubbed. An unimplemented capability is absent, not simulated:

doctor and the commands report what is actually present.

  • --store on the run and experiment commands records the run, its manifest

and its artifacts in the history; without it, nothing is written there.

Intentionally absent

experiment resume and experiment cancel need a scheduler with a queue of

runs in flight; there is none yet, so the commands are absent rather than

accepting arguments and doing nothing.