Skip to content

plugin-trust-draft.md

Latest commit

 

History

History
56 lines (41 loc) · 2.36 KB

plugin-trust-draft.md

File metadata and controls

56 lines (41 loc) · 2.36 KB

HOL-HCS-CODEX-PLUGIN-TRUST-DRAFT

Scope

This local draft defines trust attribution for top-level Codex plugins using the same HCS-style structure as HCS-28: single-score adapters, explicit weights, explicit contribution modes, and deterministic denominator rules.

Subject model

The scored subject is one top-level Codex plugin rooted at a plugin directory.

Scoring profile

  • Profile id: hol-codex-plugin-trust/baseline
  • Profile version: 0.1
  • Execution modes:
    • read mode: includeExternal=false
    • refresh mode: reserved for future external provenance checks

Adapter contract

Each adapter emits exactly one score key named <adapterId>.score.

Baseline adapter catalog

Adapter ID Weight Contribution Mode
verification.manifest-integrity 0.35 universal
verification.interface-integrity 0.25 conditional
verification.path-safety 0.20 universal
verification.marketplace-alignment 0.20 conditional
security.disclosure 0.15 universal
security.license 0.10 universal
security.secret-hygiene 0.35 universal
security.mcp-safety 0.20 conditional
security.approval-hygiene 0.20 universal
metadata.documentation 0.15 universal
metadata.manifest-metadata 0.2625 universal
metadata.discoverability 0.15 universal
metadata.provenance 0.1875 universal
operations.action-pinning 0.2625 universal
operations.permission-scope 0.15 universal
operations.untrusted-checkout 0.1875 universal
operations.update-automation 0.15 universal

Interpretation

SECURITY.md now has one explicit weight: security.disclosure = 0.15. It no longer sits inside an opaque grouped adapter. Contributors can see its contribution directly relative to stronger signals such as security.secret-hygiene = 0.35, verification.manifest-integrity = 0.35, and operations.action-pinning = 0.2625.

Aggregation

The scanner uses the same aggregation pattern as HCS-28:

  1. clamp emitted scores to [0,100]
  2. materialize missing universal adapter scores as 0
  3. include conditional adapters only when their subject actually exposes the corresponding surface
  4. compute the final plugin trust score as the weighted mean of the included adapter totals