ADR-0151 - Skill, agent, and artifact economy

Context

The product accumulated overlapping workflow skills, commodity implementation procedures, custom agent definitions, and artifact types that did not improve the core contract. Every extra surface increased activation ambiguity, maintenance cost, and user ceremony.

Decision

1. The skill catalog is deliberately small. Universal methods are bulletproof, demolition, dissect, disrespec, revolver, triple-check, promote, and remember. Canonical artifact authors are sus-spec, sus-task, sus-review, sus-inventory, sus-change-plan, sus-audit, and sus-research.
2. Each skill stands alone. No alias ships. No skill requires another skill, names catalog internals, or relies on a repository-local mirror. bulletproof owns active evidence generation. sus-spec owns planning, authoring, and checking a spec. sus-research covers technical and market questions without weakening source discipline.
3. Implementation stays native. Suspec ships no implementation, debugging, testing, security-review, Git, documentation, PRD, RFC, or bug-report skill. The harness and project tools execute work from the artifact or request.
4. Fresh subagents replace a custom catalog. Suspec ships no agent definitions, projections, delegation hooks, or role catalog. A harness may create a fresh native subagent when isolation or independence matters; the standalone skill carries the method.
5. Suspec owns these artifact types. They are spec, task, review, inventory, change-plan, audit, research, and inspection. Evidence receipts and run notes are untyped sidecars. Other durable records belong to the project or harness that owns them.
6. Artifacts serve action. Ordinary conversation and direct action create no Suspec artifact. A requested or workflow-required artifact lives in the neutral workspace until its work ends; the user then chooses Delete, Leave, or Promote for the complete transient set.

Narrowed decisions

• ADR-0114 is superseded. Suspec keeps no exhaustive active/retired/relocated status registry. Exact topology and stale-name gates inspect current repository surfaces directly; docs/artifact-registry.md remains only a surface-ownership index.
• ADR-0030, ADR-0061, and ADR-0144: the current owned artifact set is the list above.
• ADR-0092, ADR-0098, and ADR-0099: the custom agent catalog and projections retire; native fresh subagents provide isolation.
• ADR-0140, ADR-0146, and ADR-0149: skills remain the product, narrowed to the standalone catalog above.

Consequences

• Skill selection is smaller and non-overlapping.
• Users can install any skill independently.
• Suspec conditions work without replacing native implementation or delegation.
• Current documentation and checks recognize only the owned artifact set.

Status

Accepted (2026-07-13).