/keel

name: keel
description: "Design, review, or govern load-bearing architecture for open choices, design judgments, or long-lived health; skip fixed-architecture execution. Applies to new/existing systems. Private/local/reversible module boundaries qualify; scale rigor to impact. Covers responsibility/authority, interfaces/contracts, dependency/state/recovery, structural seams, guards/exceptions, drift/rot, migration, retirement, deletion, and rewrite risk."

Keel

Keel discovers, judges, and governs load-bearing arrangements. It shapes the best-supported direction and keeps declared architecture enforceable, economical, migratable, and deletable as the system evolves.

Use *keel* at module, boundary, and system scale for responsibility between parts, authority or ownership, interfaces or contracts, dependency direction, state or recovery ownership, and seams that isolate future change. Use repository-native terms and treat applicable instructions, records, contracts, and entry paths as evidence rather than decoration.

State And Ownership

Inspect only enough of the request, governing sources, and representative call sites or contracts to select the current state. Transition only when evidence changes the architecture job:

• **Decision** — establish, change, or compare a load-bearing arrangement. This includes greenfield work and an authorized improvement to an existing target.
• **Judgment** — evaluate the adequacy, conformance, or feasibility of a supplied architecture, design, RFC, change set, or existing system. Keep the target fixed. If it is infeasible, finish Open with evidence and required decision authority; enter Decision only when selection or improvement is authorized.
• **Governance** — decide whether to restore, retain, migrate, retire, or remove a declared or evidenced architecture arrangement or control, or judge its long-lived health. Evidence must show a load-bearing obligation, recurring bypass, or architecture drift rather than ordinary maintenance. Enter Decision only when an authorized remedy requires a new load-bearing arrangement.
• **Exit** — the request and governing sources fix a feasible material architecture and the remaining task only implements, diagnoses, tunes, tests, or verifies within it. Re-enter only when evidence opens a materially different load-bearing Decision, Judgment, or Governance issue.

Private, local, reversible, and single-consumer choices can still require Decision or Judgment; scale evidence and ceremony to impact, reversibility, and uncertainty.

Focused workflows own their domain evidence, criteria, method, vocabulary, artifact, professional judgment, and completion condition. Keel integrates only the load-bearing architecture result under declared decision authority; execution workflows own mutation, work preservation, test mechanics, and verification.

Automatic or forced loading grants neither applicability nor action authority. On Exit, inspect the skill catalog when the host exposes one, select and completely read a suitable owning workflow when present, and otherwise continue with the general workflow. Stop using Keel-specific routes, references, and output framing.

Solve Constructively

Decision, Judgment, and Governance share a bounded evidence loop but do different work. Decision compares directions. Judgment tests the supplied target and transitions to Decision before proposing improvements. Governance compares applicable restoration, retention, migration, retirement, removal, and no-action responses.

1. **Frame and inspect** — state the current state's question, desired outcome, authority, hard constraints, and criteria. **Greenfield evidence** consists of user goals, explicit external constraints, provenance-bearing focused or domain findings, and operating context. Treat architecture shapes as candidates, label assumptions, and treat absent repository precedent as unavailable rather than inventing inherited architecture. Search only for evidence likely to change the framing, candidate set, or result. 2. **Form, test, or compare** — for Decision, derive the smallest set of materially distinct viable directions needed to expose the real tradeoff. For Judgment, test the supplied target against the same declared criteria without generating replacement directions. For Governance, compare only applicable interventions. Include a simpler or no-change result when credible; do not invent alternatives when constraints leave one path. Use a bounded model, characterization, conformance check, or reversible experiment when it is the cheapest decision-changing evidence. 3. **Integrate and return** — preserve focused-workflow findings, assess evidence and any real candidates on the same criteria, and return the best-supported direction, verdict, or governance action with its material tradeoffs and confidence. Surface a better option the user did not name when evidence supports it without silently changing the goal. 4. **Deepen only the supported result** — resolve numbered concerns only when their answers can still change the result, adoption, recovery, or retirement plan.

Stop when more search, options, or detail are unlikely to change the result or close a material risk. If new evidence reopens a viable direction, compare it rather than defending the incumbent.

Decision, Judgment, and Governance finish in one of two states; Exit hands off without this framing:

• **Closed** — the best-supported result, applicable authority, accountability, and material risks are resolved; no known unknown can still change the result.
• **Open** — give the strongest bounded result, recommendation, or experiment available, the decision-changing unknown, and its owner or ownership gap plus the next evidence or decision trigger. Alternatives are included only when real.

For a low-blast-radius internal choice, a concise comparison and recommendation may be complete. Do not invent compatibility, migration, recovery, guards, governance, or