0007: Workflow scope and reader

Status: accepted Date: 2026-07-14

Context

The corpus was framed as a library of skill-design patterns. That framing made the material approachable, but it also confused one packaging mechanism with the system being designed. Contracts, context flow, recursive dispatch, verification, effects, persistence, and control flow apply to agentic workflows whether or not they are packaged as skills.

The stated audience, "developers," was also too broad to govern assumed knowledge, comparisons, or the depth of the material.

Decision

The unit of analysis is the agentic workflow. A workflow is agentic when an LLM makes at least one runtime judgment that materially shapes a subsequent action, artifact, routing decision, or effect. It need not be autonomous, multi-agent, or tool-using.

Skills remain the familiar on-ramp and an important packaging/reuse mechanism; they do not define the architecture's scope.

The primary reader is an experienced software engineer who has built at least one agentic workflow and wants to reason about its structure more critically. The intended outcome is the ability to inspect a workflow, locate the decisions that genuinely need LLM judgment, select and compose patterns, define seams, and anticipate context, failure, and effect costs before implementation.

The corpus deliberately relates agentic design to traditional system design. For borrowed ideas, the canonical spec records both the invariant that still holds and the LLM-specific delta; flavor prose turns that relationship into the teaching journey.

"Complete" means a stable, structurally coherent baseline: foundations and domain boundaries are settled, admitted patterns have defined moves and relationships, and known contradictions are resolved. It does not mean the catalog is closed. Later patterns may extend the tree without casually redefining its ontology or governing claims.

Consequences

• Skill-specific nodes generalize to workflow-level claims when their mechanics are not actually tied to skill packaging.
• Skills remain useful examples and artifacts, but no longer act as the root unit against which every pattern is defined.
• The repository title and Path: root remain provisionally unchanged until a naming pass chooses durable wording; this ADR settles scope, not branding.
• Corpus-wide ontology or foundation changes after the first complete baseline require an explicit decision and a blast-radius sweep.
• Concrete compositional examples remain deferred until the spec lane reaches that baseline.