Agent Plugin Eval

Agent Plugin Evaluation

Treat the portable Agent Plugins specification as the authority. A client-native manifest (e.g., .codex-plugin/plugin.json, .claude/settings.json, .cursor/mcp.json) does not replace the required root plugin.json.

Parameters

Safety boundary

Audit untrusted repositories statically. Do not run bundled executables, hooks, install scripts, package managers, MCP servers, or networked tests unless the user explicitly authorizes execution. Redact suspected secret values; report only their location and kind. A secret-like key or value is a suspicion, not confirmation: do not assign the FAIL gate without corroborating evidence such as a recognized live credential format, a trusted secret scanner, repository history/provenance, or user confirmation. Never test a credential against a service merely to confirm it.

Workflow

1. Resolve the plugin root. Use a local target in place. For a Git URL, shallow-clone into a mktemp -d directory. A plugin root contains root plugin.json; if a repo has zero or multiple candidates, report the ambiguity instead of guessing. Done when every target maps to one explicit plugin root.
2. Load the governing rules. Read references/spec-checklist.md and references/rubric.md. For Agent Plugins 1.0.0, use the bundled snapshot. For another declared version, or when the user asks for the latest spec, browse the canonical specification and schemas at agent-plugins.org and record the evaluated version and retrieval date. The normative text wins if it conflicts with JSON Schema.
3. Inventory every package path. Include dotfiles, symlinks, immediate skill children, extension namespaces, executable files, and files ignored by Git. Resolve every symlink and package-relative path against the plugin root. Done when every discovered path is accounted for as portable core, client extension, supporting file, or containment violation.
4. Run the deterministic scan. Execute python3 scripts/inspect_plugin.py <plugin-root> --json. Treat its output as evidence leads, not the final judgment. Confirm each reported issue in the source and add file:line or JSON-pointer evidence. Never weaken a normative finding merely because a client happens to accept it.
5. Review components completely. Inspect every immediate skills/*/SKILL.md and every mcpServers entry. Validate Agent Skills against their own specification. Assess instructions, resources, scripts, MCP configuration, extension isolation, cohesion, and practical utility. If skill-evaluation is available, it may deepen individual skill-quality analysis, but it never replaces this plugin-level rubric.
6. Classify conformance before scoring. Use the exact failure boundaries in references/spec-checklist.md: PASS, PARTIAL, or FAIL. Keep client compatibility separate from portable conformance. A client-specific feature may be excellent for that client and still add zero portable coverage.
7. Score with cite-or-cut. Score all applicable rubric criteria from references/rubric.md. Every score needs specific evidence; every N/A needs a reason. Run scripts/score.py for the weighted result and gate cap; do not calculate it by hand. Done when all criteria and all findings are reconciled with the conformance status.
8. Answer in the requested language. Read references/output-template.md and emit that structure. Lead with the verdict, distinguish blockers from recommendations, and provide concrete fixes. When compare is set, evaluate both independently before computing deltas; never force the same N/A set on both plugins.

Gates and scoring

• PASS: no normative violation found; no score cap.
• PARTIAL: non-fatal manifest deviation or invalid/skipped component; final score capped at 59.
• FAIL: fatal manifest/package-root failure, root-manifest escape, or confirmed embedded credential; final score capped at 39.
• Keep the uncapped score visible so authors can distinguish design quality from release-blocking conformance.

Invoke the calculator with one criterion:score:weight triple per criterion:

bashCopy code
python3 scripts/score.py --gate partial 1:90:3 2:80:3 3:NA:2

Evaluation with Jev (Optional)

When TypeSafe Jev is available, use it for subjective quality criteria. Jev provides calibrated probability judgments that augment the deterministic checks.

When to use Jev

Discovery protocol

pythonCopy code
from typesafe import jev_available

if jev_available():
    from typesafe import Score, Noul
    # Use Jev for subjective criteria
else:
    # Fall back to heuristic scoring

Questions and integration

Questions are defined in scripts/jev_questions.json:

• 7 Score questions: Axis 4 (UX coherence, documentation clarity, error handling) and Axis 2 (validation, schema design, naming, API elegance)
• 20 Noul questions: Gates (G1-G4), secrets (4), quality checklist (7), component validity (4)

Score results (0.0-1.0) are averaged per axis and scaled to the rubric (0-25). Noul results provide categorical classifications for gates and checklists.

See references/jev-integration.md for full integration patterns and code examples.

Output format

When Jev is used, the scorecard includes a jev section:

jsonCopy code
{
  "jev": {
    "available": true,
    "quality_scores": { "ux_coherence": 0.72, ... },
    "gate_classifications": { "G1": {"label": "conformant", "passed": true} },
    "secret_findings": { "requires_review": false }
  }
}

Gotchas

• The v1 portable core contains exactly Agent Skills and MCP servers. Hooks, commands, agents, apps, marketplaces, and distribution policy are client-specific unless placed in a valid extension namespace.
• Missing optional skills/ or mcp.json is not an error. A present path of the wrong filesystem kind is an invalid component type.
• Unknown root manifest fields are schema violations but have the spec's narrow non-fatal handling; most other manifest schema violations reject the whole plugin.
• One invalid skill or MCP server must not be reported as if every independent component were invalid.
• ${PLUGIN_ROOT} and ${PLUGIN_DATA} expand only in MCP args, env values, and cwd; never in command, URLs, or headers.
• A high-quality client-native plugin can still fail the portable standard when root plugin.json is absent. Report both facts without averaging them away.
• Keep possible credentials labeled “suspected” and redacted. A heuristic hit alone lowers the security score and demands remediation review, but does not become a confirmed-credential FAIL gate.

Final quality gate

• Every target resolved to exactly one root
• Every file, symlink, skill, MCP server, and extension inspected
• Every normative violation mapped to its correct failure boundary
• Every score cited and every N/A justified
• Suspected secrets redacted
• Score produced by scripts/score.py
• Comparison deltas use independently computed scores