| name | role-candidate-score |
| description | Score and filter candidate sources before turning them into Agent Roles; use after role research when comparing skills, plugins, repositories, tools, prompts, or workflow sources for relevance, authority, license, host fit, security, testability, and Role-shape fit. |
Role Candidate Score
Use this skill after research has produced candidate sources. Its job is to
make selection explicit before source ingestion, blueprinting, or scaffolding.
Inputs
- Research brief and candidate list.
- Research evidence with source locators, access dates, inspected files/pages,
authority labels, extracted facts, and confidence.
- User priorities: speed, stability, license safety, host fit, novelty,
local-only use, or other constraints.
- Existing inventory output when a local source tree has already been
inspected.
Hard Gates
Reject or require a user decision before vendoring, copying, or adapting when
any gate is hit:
- license is unknown or incompatible for copied or vendored content;
- source contains secrets, provider sessions, runtime state, or hidden install
state;
- depended-on tooling has no maintainer signal;
- host fit is poor or adapter behavior would be misleading;
- testability is weak and cannot be made explicit;
- Role-shape fit is unclear: one Role, multiple Roles, or topology recipe.
Direct vendoring is a positive option, not a fallback, when the hard gates pass
and the user wants upstream behavior carried as-is. Score it separately from
synthesized or referenced_only treatment.
Scoring Dimensions
Score each candidate on a 0 to 3 scale:
- relevance to requested Role tasks;
- source authority and maintenance;
- license/provenance safety;
- Agent Roles host fit;
- runtime boundary cleanliness;
- dependency risk;
- security posture;
- testability;
- Role-shape fit.
Apply user-priority overrides explicitly instead of hiding them in prose.
Confidence must be high, medium, low, or unknown.
Output
Use templates/candidate-scorecard.md when a persisted artifact is useful.
Return:
- scorecard table;
- hard-gate results;
- recommended candidate or split recommendation;
- recommended copy treatment (
vendored_intact, vendored_modified,
synthesized, referenced_only, or excluded);
- rejected candidates with reasons, or an explanation that only one source
was available;
- next action: inspect selected source, produce blueprint, ask the user, or
stop.
