CLI reference

Nine commands. The ordering below is a cost gradient — demo and cost are free, try is capped at $2, and only compare, sections and apply spend real money. Every spending command shows the estimate before it runs.

Command	Cost	What
`demo`	$0	Replay a bundled recorded run.
`cost`	$0	Static per-section token cost.
`init`	$0	Scaffold `.ablate/tasks/`.
`try`	≤ $2	Smallest real run against a live agent.
`compare`	~$28	The paired A/B, with a confidence interval.
`sections`	~$110	Screen + confirm section attribution.
`apply`	$0	Perform a deletion, with an archive.
`report`	$0	Re-analyze a stored run.
`mine`	—	Not yet implemented (post-v1).

Universal flags

Flags shared across the spending commands (compare, sections, try):

--tasks DIR (default .ablate/tasks) — directory of task YAML files.
--tasks-only IDS — run only these task ids or globs, e.g. 'fix-*,add-cli-flag'.
--tasks-tag TAGS — run only tasks carrying any of these tags, e.g. 'smoke'.
--model ID / --cheap — target model; --cheap maps to Haiku 4.5.
--concurrency N — parallel trials.
--max-spend USD (default 25, 2 for try) — hard ceiling, checked against real reported costs mid-run, not just the estimate.
--margin PCT — equivalence margin, e.g. 0.03 for ±3pp. Applies to every metric.
--alpha A (default 0.05) — significance level.
--seed N — shuffle and bootstrap seed. Not agent sampling; nothing can seed that.
--yes — skip the confirm prompt. Never bypasses --max-spend.