ak test optimize

Use ak test optimize for the operation described below. The command help records its arguments, effects, and options.

Without a manifest at the --manifest path the command fails and asks you to configure trusted groups and full_suite. With --run, the command reloads the manifest and fails if it changed after planning.

Semantic modes use the test-impact consumer. A provider call happens only when you have enabled semantic_decision at user scope, set semantic_decision.consumers.test-impact to true, and made the selected provider's credential available (see providers and credentials). The grant is required even when you have no consumers map. See ak config prefs set.

Configure the group manifest

The repository owns .agentkit/test-impact.json. Each group declares its command argv, a repository-relative working directory, the paths it owns, dependency edges, an execution class, and a timeout. The manifest also declares a nonempty full_suite, which groups are mandatory or critical, and explicit required_checks mappings for active policy. coverage_complete is your declaration, not proof that the dependency data is complete.

Selection stays conservative. Mandatory and critical checks always apply. An unknown path, an incomplete mapping, or uncertain evidence broadens the plan to the full suite and safety checks. A missing policy mapping blocks execution instead of skipping a required check, and a documentation-only skip requires explicit allowlist evidence in the manifest. The default --semantic off path makes no network or provider call. Existing preflight and CI gates remain authoritative; this command is a local verification route.

Plan, run, and save a receipt

Without --run, the command only plans: it reads Git, the manifest, and policy, then reports selected and skipped groups with reasons, unresolved policy checks, and fallback reasons. Committed changes, both endpoints of a rename, and local dirty or untracked changes all contribute to the source fingerprint.

--run executes the trusted manifest commands sequentially with timeouts and process-tree cancellation. It runs only against the current checkout: the requested head must equal it, so a historical head can be planned but not executed. Source, manifest, and policy fingerprints are checked before execution, and a source change during the run invalidates the evidence. A working-directory symlink must stay inside the repository. A failing narrow group triggers broader checks and remains a failure. The commands are repository code you trust; --run is not a sandbox.

--output <new-file> creates a local JSON receipt and never replaces an existing file or symlink. The receipt records group results, required checks, broadening, attempt counts, duration, and the source revision. Cost is null when it was not observed, and raw subprocess output is not stored. Avoid tests that modify tracked source or leave new untracked files behind.

Command contract

Explain impact-selected checks, optionally execute trusted commands

Synopsis

Default is read-only planning. Supply a trusted manifest with group commands, ownership, full-suite groups and explicit policy mappings. Invalid optional evidence broadens; missing mandatory mappings block. --run executes repository code without a sandbox.

Effects: Planning reads Git and policy only. --run executes trusted commands in this repository; --output creates a local receipt. Semantic mode defaults off; an explicitly permitted mode may send compact facts and incur provider charges. Shadow never changes the effective plan.

Exit status: 0 executable plan or all selected checks passed 1 invalid inputs, blocked plan, failed or stale execution 2 invalid arguments

ak test optimize [flags]

Examples

  ak test optimize --base origin/dev --json
  ak test optimize --base origin/dev --run --json

Options

      --base string               Git comparison base (required)
      --evidence string           Optional bounded external evidence JSON
      --head string               Git comparison head; local changes also count (default "HEAD")
  -h, --help                      help for optimize
      --json                      Emit machine-readable JSON (implies --no-interactive)
      --manifest string           Repository-owned executable group manifest (default ".agentkit/test-impact.json")
      --no-interactive            Disable interactive prompts (CI-safe)
      --output string             Create a new local JSON result file; never overwrite
  -q, --quiet                     Suppress non-error output on stderr
      --run                       Execute trusted manifest commands; this is not a sandbox
      --semantic string           Optional semantic observations: off, shadow, or advisory; requires explicit user consent (default "off")
      --semantic-context string   Reviewed compact semantic context JSON bound to this request
  -V, --verbose                   Extra diagnostic output on stderr (loses to --quiet)
  -y, --yes                       Assume yes for all prompts