ak skills evolution evaluate

Re-run the offline, deterministic evaluation of a recorded candidate and refresh its evaluation slot.

Use ak skills evolution evaluate to rebuild a candidate on its base and run the same evaluation that ak skills improve --candidate ran when it recorded the candidate. No model, network, or script runs. A live model routing benchmark is a separate, optional check and never gates apply.

Usage

ak skills evolution evaluate <candidateId> [--kits-dir <dir>] [--dry-run] [--json]

Exactly one candidate reference is required.

Flag	Default	Description
`--kits-dir <dir>`	`$AGENTKIT_KITS_DIR` or `./kits`	Foundation kits directory the base is rebuilt from.
`--dry-run`	`false`	Print the evaluation and write nothing.

Shared flags are described in CLI conventions.

What it evaluates

The candidate is rebuilt on its base: the foundation skill from --kits-dir, or the running personal kit composition it stacks on. A base that moved since the candidate was recorded fails with evolve.base_changed.

The evaluation routes the evidence tasks with no skill, with the base skill, and with the candidate through the frozen lexical router, scores failures and corrections for the terms the skill text covers, routes the other skills of the kits directory as negative controls, and runs the functional and security checks. The verdict is improved, inconclusive, regressed, or blocked.

Effects

The command reads the candidate record, the kits directory, and the running personal kit records and skill store entries under the AgentKit home. Without --dry-run it replaces the evaluation slot of AGENTKIT_HOME/state/skill-evolve/candidates/<id>/candidate.json under the kit lifecycle lock. It never changes a skill, a kit, or a personal kit.

Output

Text prints the candidate ID, the verdict with its method, router, and reasons, the task counts, the routing and coverage figures of each variant over all tasks and over the held-out tasks, the negative controls, every check that is not a plain pass, and each permission change.

JSON uses kind=skills.evolution.evaluate:

schema_version: 1
kind: skills.evolution.evaluate
data.dryRun
data.evaluation: schema, state, verdict, reasons[], method, router,
                 candidateId, evidenceDigest, baseTreeDigest, compositionDigest,
                 tasks, variants[], holdout[], gain[],
                 functional[], security[], permissionDelta[],
                 negativeControls, sideEffects, digest

Each check has id, candidate, and optionally baseline and detail; a check status is pass, fail, or review. Variants are none, baseline, and candidate.

Exit behavior

Exit	Meaning
`0`	Success, including a dry run.
`1`	The candidate store, the personal kit, or the skill store could not be read or written, or a record failed verification (`store.corrupt`).
`2`	Invalid flags or arguments.
`4`	Another AgentKit operation holds the kit lifecycle lock (`evolve.store_busy`).
`5`	No recorded candidate matches, or the skill's base cannot be resolved (`evolve.base_unresolved`).
`6`	The base moved since the candidate was recorded (`evolve.base_changed`).
`7`	Redaction or the content gate could not be prepared (`evolve.redaction_failed`).