agent-evallisted

# agent-eval — Agent/Skill Adversarial Evaluation Workflow ## Purpose Read the target asset **trying to break it**, bind every judgment to evidence (file:line · SkillSpector finding · reproduction), and produce severity-ranked defects plus a 0–10 scorecard. This is the execution engine of the `agent-evaluator` agent. > **Grounding docs**: [`reference/principles.md`](reference/principles.md), [`reference/kb/INDEX.md`](reference/kb/INDEX.md). > Don't trust scores at face value — SkillEvaluator and SkillSpector results are **triage input**. ## When to use - Gate a new/changed agent or skill before install/merge. - "Evaluate / review / security-check this agent (skill)." - Quality + security check after `install/check-manifest.sh` passes. ## Prerequisites and scope - `tier1` requires an installed, pinned `skillevaluator` CLI; use the wrapper help for its exact installation command. - `security`, `tier2`, and `tier3` require the scanners, providers, credentials, datasets, and sandbox described in [`kb/skillevaluator-usage.md`](reference/kb/skillevaluator-usage.md). Missing prerequisites are incomplete evidence, never a pass. - Keep reports outside the target repository. Do not run `validate --full` or `--autopilot`, because they can create target-local evaluation data. ## Instructions ### Identify the target - Agent: `agents/<name>.md` + sibling bundle (`agents/<bundle>/SOUL.md`, `reference/`). Confirm the bundle stem via the relative path in the definition body. - Skill