Fitness — read-only goal measurement

Inspect the active goals document and run only the caller-selected measurement, validation, drift, history, export, or meta-goal command.

Renamed from goals (2026-07-29): the semantic skill is fitness; the ao goals CLI command family is a separate product surface and keeps its name. A thin goals compatibility alias resolves to this skill.

Measurement stays trustworthy only because it cannot mutate what it measures; the moment a fitness report edits a goal, the next report measures the editor, not the project.

Named failure mode — advice creep: a measurement report that ends with "you should…" has silently become work selection.

Anti-pattern: padding the report with recommendations to look helpful. Corrective: return the numbers, the evidence gaps, and checked/not-checked scope, and let the caller decide.

Boundary

• Prefer GOALS.md when both Markdown and legacy YAML exist.
• Preserve stable directive and gate identities in the report.
• Every measured gate must name its executable check and observed outcome.
• Do not add, remove, prioritize, recommend, apply, prune, migrate, or otherwise

mutate goals.

• Do not translate a fitness gap into work selection or a next action.
• No subcommand edits the goals source through its own logic. measure,

drift, and export persist a best-effort JSON snapshot under the fixed derived path .agents/ao/goals/baselines/. render --out <file> writes a Gherkin spec to whatever path the caller names — the CLI does not constrain it, so never point --out at the goals source or any non-derived file.

Commands

All eight subcommands read the goals source without mutating it. The snapshot write (fixed derived path) and the render --out write (caller-chosen path, caller's responsibility) are the only side effects.

bash
ao goals measure --json
ao goals validate --json
ao goals drift
ao goals history
ao goals export
ao goals meta --json
ao goals scenarios
ao goals render          # append --out <file> to write the spec instead of stdout

Run the requested command once. Return the command, exit code, goal-level results, aggregate measurement, missing evidence, and checked/not-checked scope. Then stop.