ADR-0161 - Semantic skill contracts and maintained evidence

Context

Independent installation forbids sibling dependencies. Pure isolation, however, left adjacent methods disconnected even when a harness could discover the next capability from loaded skill metadata. Skill descriptions also drifted between selection metadata and body procedure, while catalog doctrine changed faster than its design evidence.

Implicit activation is harness- and model-dependent. It may be tested and encouraged; it may not be claimed as a portable guarantee.

Decision

1. Descriptions select. Every description states the action and target, a positive Use when trigger, and genuine near misses in exactly three sentences. It contains no self-name, procedure, sequencing, evidence rules, or completion mechanics.
2. Bodies execute. Method is mandatory. Other standard sections appear only when they carry behavior. Instructions use hard imperatives and omit filler, default behavior, repeated rationale, and internal catalog vocabulary.
3. Every skill remains complete alone. No body names, requires, invokes, or assumes a sibling. Load-bearing behavior stays local.
4. Adjacent methods compose semantically. A skill may state a useful next job in language that matches another skill's description. The sentence must remain executable when the companion is absent. Intended handoffs are tested with each skill alone and together; results remain local evidence, not a universal activation claim.
5. Design docs are current contracts. The skills repository documents activation, body anatomy, execution, scope, self-containment, existence, and evidence. Gates bind those documents to the shipped catalog.
6. Evidence changes with doctrine. docs/sources.md is updated in the same change whenever an externally grounded skill or documentation rule changes. Primary and official sources lead. Secondary evidence is labeled and bounded. Dead or irrelevant sources leave with the claim.

Narrowed decisions

• ADR-0149: standalone operation stands; semantic wording may expose an optional adjacent job without creating a dependency.
• ADR-0157: ruthless economy governs descriptions and maintainer documentation as well as bodies.
• ADR-0113: executable skills remain self-contained; design docs own catalog contracts and supporting evidence.

Consequences

• Installation order never determines correctness.
• Harnesses may discover coherent method chains without explicit sibling calls.
• Descriptions stop wasting preload context on procedure.
• Documentation drift and unsupported design folklore become gate failures.

Status

Accepted (2026-07-15). Narrows ADR-0113, ADR-0149, and ADR-0157.