package-optimizer

Package Optimizer

Turn existing package evidence into one bounded, reviewable proposal. The optimizer is proposal-only: it never edits a package, changes a profile, or starts a benchmark run.

Scope

Situation	Action
One package has current evaluator evidence	Propose one disposition.
A bounded family shares the same evidence/profile	Propose one family plan with a per-package row.
Evidence is missing, stale, structurally invalid, or from another profile	Return inconclusive; name the unresolved package decision and request the smallest recertification needed.
User asks to apply a proposal	Require explicit approval, preserve before artifacts, then hand the edit to the normal package workflow.
User asks for a broad benchmark before identifying a package decision	Refuse the expansion; identify the package decision first.

Required inputs

1. Package scope — one type/name package or an explicitly bounded family.
2. Current evidence — static conformance and relevant eval-case result or a recorded behavioral result. Cite exact commands, artifacts, or case IDs.
3. Capability profile — model/client/tool-surface identity when behavioral evidence depends on one. Structural-only evidence is labelled as such.
4. Decision question — what behavior should be retained, simplified, or strengthened.

Do not infer effectiveness from package length, heading count, or static score.

Evidence validity

Reject evidence when any condition holds:

• package path, version, or evaluator case differs from the proposed scope;
• a behavioral result has no model/client/tool-surface identity;
• a profile is older than a recorded package change or cannot be compared to the requested profile;
• the evidence omits a failure, safety outcome, or relevant evaluator result;
• an M4 recertification report pools model targets or lacks complete cells.

A rejection uses the single output format below with Disposition: inconclusive, Proposed change: none, and Approval required: yes. Its Unresolved package decision must name the package scope and decision question. A recertification request repeats that exact field; do not request a benchmark without it.

Procedure

1. Resolve the package scope and read its definition, evaluator cases, and current conformance output.
2. State the decision question and evidence class: structural, behavioral, or profile-scoped behavioral.
3. Verify evidence validity. Stop with inconclusive on any invalid input.
4. Select exactly one disposition:
   • retain — evidence supports the current contract.
   • simplify — evidence identifies redundant detail and the retained contract/evaluator proves the smaller scope.
   • strengthen — a documented failure requires a concrete contract addition.
   • retire — explicit deprecation or replacement evidence supports removal.
   • inconclusive — evidence cannot support a safe change.
5. Produce a proposal. For non-retain dispositions, name exact sections or files to change and the evaluator behavior that must remain true.
6. Stop. Do not edit. Require the user to explicitly approve the proposal.

Approval handoff

An approval must name the package, disposition, and proposal ID. The downstream applying workflow—not this skill—must:

1. Save the original package artifact and evaluator evidence.
2. Apply only the approved package/family change.
3. Run static conformance and relevant eval cases.
4. Record before/after artifacts, verification output, profile identity, and any regression.
5. Revert the bounded package change on regression. Do not broaden the scope.

Output format

## Package Optimization Proposal

Scope: <type/name or bounded family>
Decision question: <question>
Evidence class: <structural | behavioral | profile-scoped behavioral>
Evidence:
- <command/artifact/case and observed result>
Capability profile: <identity | not applicable>

Disposition: <retain | simplify | strengthen | retire | inconclusive>
Rationale: <evidence-backed explanation>
Unresolved package decision: <scope + decision question | not applicable>

Proposal ID: <stable scope + evidence identifier>
Proposed change: <none | exact files/sections and intended behavior>
Preservation check: <existing evaluator/eval case>
Approval required: yes

Smallest next action: <concrete action>

Guardrails

• Never edit without explicit approval.
• Never optimize all packages at once.
• Never turn missing evidence into a simplification recommendation.
• Never claim token, cost, or quality improvement without recorded evidence.
• Prefer retain or inconclusive over speculative change.