ADR-0051: Curated skill routing and external method adaptation

Status

Accepted

Context

Intentloom already distributes portable Agent Skills and governs external extensions, but its initial catalog contains mostly narrow execution and review procedures. It lacks an explicit task router, an interactive feature-discovery procedure, a completion verification gate, and a skill-specific external extension review.

External projects such as obra/superpowers and mattpocock/skills demonstrate useful brainstorming, debugging, test-first, planning, and review methods. Their complete plugin workflows also include assumptions that Intentloom cannot adopt as canonical behavior: mandatory invocation for trivial work, provider-specific setup, automatic hooks or updates, remote visual assets, unconditional commits, and broad subagent orchestration.

Direct installation would create a second workflow authority, make generated behavior provider-dependent, and weaken the managed-extension boundary.

Decision

Intentloom will curate useful methods into first-party, provider-neutral skills instead of bundling or requiring either external plugin.

1. Project-owned instructions, accepted specifications, ADRs, capability policy, and explicit user decisions remain authoritative. A skill supplies procedure; it never grants authority.
2. aif-task-router selects the smallest useful route and explains non-obvious skill choices before non-trivial work.
3. High-interaction feature discovery is recommended only when ambiguity, alternatives, or risk justify it. Clear bounded work proceeds without a mandatory interview.
4. aif-feature-discovery, aif-verification-gate, and aif-extension-review close the initial catalog gaps. Existing debugging, testing, planning-review, and code-review skills adopt compatible methods without copying external provider behavior.
5. Canonical skills remain under catalog/skills/ and are emitted through the existing Claude, Codex, Cursor, and Copilot adapters. No provider plugin is a runtime dependency.
6. External skills and plugins remain untrusted managed-extension candidates. Adoption requires a pinned source, integrity and license evidence, capability review, evaluation, explicit approval, and rollback.
7. Automatic plugin installation, hooks, telemetry, remote assets, background mutation, commits, pushes, merges, releases, and permission expansion remain prohibited.

The initial routing modes are:

direct | clarify | discover | diagnose | plan | implement | review | adopt

The selected route and skill set must be explainable. Progressive skill loading continues to use catalog metadata, execution contract, and full procedure levels.

Consequences

Positive

• Intentloom gains structured brainstorming and clarification without forcing it on every task.
• The same curated behavior reaches supported agents through deterministic adapters.
• External methods can improve the framework without becoming supply-chain or workflow dependencies.
• Verification, licensing, capability, and human-approval boundaries remain explicit.

Negative

• Curated skills must be evaluated and maintained independently from upstream projects.
• Future upstream improvements are not inherited automatically.
• Rich provider-specific plugin features remain unavailable until an optional, reviewed adapter has a demonstrated consumer.

Compatibility

This decision adds canonical skills and refines existing procedures. It does not change schema versions, adapter layout, CLI commands, capability grants, or project mutation behavior. Existing generated projects adopt the new files only through their normal preview, diff, conflict, and synchronization workflow.