Skip to content

nisavid/artifact-customs

v1.0.0MIT

Clear, adopt, maintain, replace, and retire exact third-party software components under explicit policy.

GitHubOpen Website

Artifact Customs

Artifact Customs governs inbound third-party software components from first clearance through retirement. It treats every component as an exact external identity or byte sequence crossing a declared trust boundary.

The root plugin.json is the canonical Agent Plugins v1 package identity. Codex reads its com.openai interface extension and the standard skills; Claude Code uses the .claude-plugin/plugin.json identity projection and the same skills. Each skill carries ordinary local projections of the shared contracts it reads so standard clients do not depend on package-sibling access. Release validation requires every projection to match its single authored package-level source byte for byte.

The plugin has three public skills:

  • assessing-third-party-components freezes and evaluates a candidate without mutation;
  • adopting-third-party-components establishes or explicitly revises the policy for a new trust boundary; and
  • maintaining-third-party-components updates, investigates, replaces, seals, or retires a component under an existing policy.

The scope includes packages and transitive dependencies, vendored source, wheels and binaries, GitHub Actions pins, container-image digests, toolchain and CLI pins, and agent-plugin pins. It excludes models and datasets, credentials, SaaS-vendor governance, outbound builds and releases, general incident response, deployment, generic Git and pull-request work, and generic code review.

Artifact Customs does not duplicate those neighboring owners. It calls rolecasting:delegating-cross-agent-work for bounded dispatch, tricritical:review, tricritical:adjudicate, and tricritical:loop for independent review through revision, versionkeeping:checkpointing-and-publishing-git-work for Git and push mechanics, and mergecraft:publishing-reviewable-prs plus mergecraft:getting-prs-merged for pull-request publication and merge closeout.

Manual repository-and-PR invocation is permanent. Scheduled tasks in Codex via ChatGPT, Claude Code via Claude Desktop, or another automation harness invoke the same envelope and receive the same receipts. During onboarding, first inspect for any existing or possibly existing dependency or component maintenance schedule or analogous maintenance process. When one is found or suspected, offer best-effort integration or alignment from available evidence; if its suitability is uncertain, also offer a context-sensitive standalone cadence. When none is found, recommend a context-sensitive cadence. The operator always chooses activation, cadence, and autonomyMode; Artifact Customs remains scheduler-neutral.

Every request, lock, receipt, and idempotence key binds the coarse mode, exact requested lifecycle action, caller identity, autonomyMode, candidate identity, policy identity, and authority identity. maintain must name exactly one action: update, advisory, replace, seal, retire, or pull-request.

Offer exactly three autonomy modes. High-confidence is the recommended mode, not a default or implicit consent: scheduled invocation and deployment state must record the operator-selected autonomyMode. report-only never upgrades autonomously and can research/report/recommend; high-confidence upgrades only after adequate verification, security review, and a high-confidence safe judgment, otherwise escalating with a comprehensive report; and confidence-forward/deferred upgrades when confident but defers uncertainty to a later research cycle for ecosystem evidence. Treat compromised-artifact risk fail-closed. No autonomy mode expands mutation authority beyond the named existing policy.

Development evals, fixtures, tests, and generated content locks live in the repository-level evals/, tests/, and release/ trees rather than this runtime plugin root.