Class contracts

Every skill declares a class: in its frontmatter. That declaration is not decoration — it selects which rules check.sh holds the skill to.

What each class must prove

Class	Must contain	Why
review	NJ_AGENTS_CI or a CONVENTIONS.md §5 reference	A review skill with no CI mode cannot gate a pipeline
review + subclass: gate	A BLOCK verdict	A gate that cannot block is not a gate
authoring	§A3 and §A4	Must propose the commit, and say where its artifact lands
pm	§P2 and a paste-ready-markdown fallback	Must work when no tracker is connected
workflow	A statement that it never runs git	The class's whole promise

gate vs scan

Review class splits in two, and the distinction is load-bearing:

• gate reviews a diff and answers "is this safe to push?" — PASS / WARN / BLOCK, and a hook can act on the verdict.
• scan sweeps the whole repo for accumulated debt and returns candidates. There is no sensible BLOCK for "you have 12 unused exports."

Only gates are held to the verdict tokens.

That split is read from the subclass: key — never from a list of skill names. A hardcoded allow-list would silently exempt skill #24, which is precisely the forget-to-update failure the harness exists to catch.

Spawning skills carry two extra obligations

Any skill that spawns subagents — whatever its class — also follows CONVENTIONS-orchestration.md:

§C — cost. State the cost shape before the first dispatch, cap fix rounds at 2, halt on any signal to stop. The user should never discover the cost mid-run.

§R — progress. Announce the roster before dispatch, then mark each agent as it lands. This binds the skill, not the agent: a subagent returns once, at the end, with no channel to stream status — so the spawning skill is the only place that can make a multi-agent run legible.

Both are enforced by spawn detection, not a skill list.

Behavioural contracts

Four contracts are asserted against real runs rather than text:

Contract	Asserted as
Review class leaves no files	File listing and git status byte-identical before and after
No skill ever commits	HEAD unchanged, nothing staged
review-secrets blocks with no scanner	Zero agents spawned — structural, not a grep for "BLOCK"
Authoring writes exactly one artifact	One new path, at the declared §A4 location, existing content preserved

The third is worth dwelling on. Asserting "the output contains BLOCK" would pass if a summary merely mentioned the word. Asserting "no subagent was spawned" proves the diff was never shared with anything — which is the actual promise.