aoa-decision

Intent

Provide one front door for durable decision work without loading three procedures into every request. Keep owner-authored decision records authoritative; indexes, retrievers, KAG packets, caches, and session evidence may locate or contextualize a record but never supply its rationale.

Trigger boundary

Select this family only after the request fits its frontmatter description. Treat an undecided option, ordinary documentation edit, current authority map, or non-decision generated/runtime drift as a different task. Lack of evidence or write authority never becomes implicit approval.

Inputs

• one decision intent and the target owner, subject, ID, path, or mismatch
• exact owner routes, records, template, index contract, and effect authority

for the exact effect, target, and scope required by the selected mode

Outputs

• one typed result from references/contract.yaml with owner source, actual

effects, verification, skipped checks, uncertainty, and stop line

Procedure

Choose exactly one mode:

1. Read references/contract.yaml, then read the selected mode reference to

EOF. Do not load another mode or a retired child bundle.

1. Use an exact supplied owner root, decision ID, path, or source surface

before discovery. Read the owner's route law before any source or generated write. If that law defines a disposable or non-VCS owner root, do not probe Git; use its named files, builder checks, and direct hashes or comparisons.

1. Use a decision retriever only as an optional narrowing aid. Verify every

rationale, status, supersession, and owner claim in the authored decision record.

1. Execute one mode against one target. A handoff to another mode is a later

selection, not an implicit continuation.

1. Stop as blocked_missing_input before target reads or effects when a

required owner source, accepted decision field, index contract, builder, or effect authority is unavailable. For a write, bind authorization to the exact effect, target, and scope. A current Operator or human authorization that already matches all three satisfies the approval gate without a repeated question; a new or changed effect, target, or scope requires fresh authorization.

1. Return the selected mode, exact source refs, result state, actual effects,

verification, skipped checks, residual uncertainty, and stop line.

Mode: find

Read and follow references/find.md.

Mode: record

Read and follow references/record.md.

Mode: correct

Read and follow references/correct.md.

Contracts

• Keep one mode and one target active at a time.
• Keep decision acceptance, rationale, and record authority with the named

owner; this skill structures work but does not make the decision.

• Require explicit current authority for record and correct effects.
• Rebuild generated decision indexes from owner source; never patch them to

hide source or builder debt.

• Keep session evidence as optional historical context, never decision truth.
• Keep techniques as optional provenance, never runtime dependencies.

Risks and anti-patterns

• Loading retired child skills or several mode references.
• Trusting an index title, KAG packet, cache, or session narrative as rationale.
• Recording a trivial diff, inferring rejected options, correcting only a

generated view, or silently rewriting semantic history.

• Copying a sibling owner's rationale, ID scheme, or template as local law.

Verification

• Confirm one mode, one target, owner-source grounding, and actual effects.
• Confirm source-first correction and owner-builder use where applicable.
• Distinguish source validity, generated parity, external retrieval freshness,

and broader architectural correctness.

• Report skipped checks and unresolved owner or federation debt explicitly.

Adaptation points

Let each repository supply its decision home, ID and status vocabulary, template, index contract and builder, validation commands, supersession law, and optional retriever binding.