ADR 0017: No always-loaded skills

Status

Accepted

Context

Two concerns are candidates for loading on every task: task-file lifecycle and promotion discipline (manage-task), and routing and forbidden-flow rules (documentation-gatekeeper). A skill authored to load unconditionally is the wrong primitive. Skills load on demand and cost context every time they fire; content that must be present on every task is persistent project context, which belongs in the always-available entry file, not in a skill paid for on each activation [LOSTMID].

An always-loaded skill also blurs the contract: an agent can't tell which skills matched the work from its description and which were force-fed, and consumers can't vendor selectively if two skills are secretly mandatory.

Decision

There are no always-loaded skills. manage-task and documentation-gatekeeper are removed as skills. Content is re-homed by kind:

• Persistent context (facts, conventions, commands every task needs) → AGENTS.md.
• Multi-step procedures (the discipline for a kind of work) → on-demand skills that self-activate when their description matches (see 0020).

Concretely:

• manage-task's task-file lifecycle and promotion discipline now live in the task templates (.agents/templates/task-base.md and the per-skill references/task-template.md) plus the process docs — the discipline is exercised where the task file is actually filled in, not preloaded as standing instructions.
• documentation-gatekeeper's routing and forbidden-flow rules become framework concept docs describing recommended routing. They are guidance, not an enforced gate (see 0020 and the superseding note on 0002).

Consequences

• Positive: every loaded skill earned its place by matching the work; context isn't spent on dormant procedures.
• Positive: selective vendoring is honest — no hidden mandatory skills.
• Negative: the lifecycle/routing discipline is now distributed (templates + docs) rather than centralised in one skill — mitigated by keeping each task template carry its own slice and by the process docs being the single narrative.

Alternatives rejected

• Keep the two skills but mark them "auto-load". Reintroduces the wrong primitive — per-task context cost for content that is really persistent (belongs in AGENTS.md) or really on-demand (belongs in a matching skill).
• Fold both into AGENTS.md wholesale. Over-stuffs the persistent file with multi-step procedures that most tasks don't need; procedures belong in on-demand skills, only the facts belong in AGENTS.md.