本文へ移動
cccskills
無料GitHub で公開

pui-package-budget

Apply one independently reviewable numeric Proto UI package-budget increase for an already accepted capability, using canonical cost evidence and existing development authorization. Use for standalone ceiling transactions, not feature implementation, measurement-policy changes, spending limits, publication or release.

インストール方法を見る

含まれるファイル(2)

  • SKILL.md7.0 KB
  • agents/openai.yaml311 B

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Apply one numeric package-budget transaction

Own the standalone numeric mutation described by pui-dev under "Decide package-budget ceilings". This leaf changes tracked feature-branch content only. It does not perform a GitHub write or gain permission from an observed CI failure.

  1. Require a current capability-envelope, the traced authority-map, the measured candidate-change, evidence-report, and implementation-authorization from the current bounded request or an active standing scope. The authority map must identify the already accepted capability and exact affected package entries. The report comes from pui-validate on that accepted capability candidate; use the validation route to collect missing evidence before this mutation. A read-only governance or CI report does not replace those inputs.
  2. Inspect the factual measurements before editing. evidence-report must identify the canonical before/after CI runs for the baseline and accepted capability candidate, their revisions, affected entries, minified artifact hashes and sizes, measured gzip costs, current ceilings, observed failures and growth attribution. A proposed ceiling or its margin is not a prerequisite; pui-validate reports measured facts without deciding that value. Record Node, esbuild, platform/architecture, zlib and build/compression parameters. Compare unchanged source or equivalent isolated artifacts when the environments differ; local diagnostics do not overrule canonical CI.
  3. Choose the proposed ceiling in this leaf from those measurements and the accepted capability's bounded cost. The proposed value and resulting headroom rationale are outputs of this leaf, not decisions that the input report or a human must already supply. Preserve evidence of investigated accidental closure, duplicate Runtime copies, dead code and avoidable eager inclusion. Justify accepted capability cost without weakening semantics or raising a number merely to obtain green CI. For related changes, identify and measure the integrated combination; do not sum isolated deltas or pass off an old feature-only result as the final combination. Missing or stale measurements mean collect/refresh them before selecting a number, not ask for approval of an unsupported value.
  4. Apply the smallest bounded change to numeric ceiling literals in scripts/analysis/package-budgets.mjs, associated explanatory comments when needed, and the transaction's supporting evidence record. Do not change entry points, bundling, compression, external dependencies, diagnostic profiles, pass/fail behavior, implementation code or unrelated limits. Route those different scopes through their own accepted work and owning leaves. The whole-entry gate remains blocking; diagnostic profiles cannot replace it.
  5. Return a separately reviewable commit or focused PR candidate linked to its accepted capability. The candidate-change must include the numeric diff and its supporting transaction record, identifying the exact changed entries/literals, old and proposed ceilings, resulting headroom rationale, evidence references/digests, attribution, combination revisions, and the rollback boundary. Preserve prior red runs and alternatives as history. State candidate-bound checks as passed, failed or unrun; a candidate awaiting canonical CI is not accepted evidence.
  6. Hand the candidate to pui-validate for the final candidate's canonical blocking package-budget check, including a fresh combination when related integration changes invalidate prior measurements. Keep the transaction record in the candidate and require a final-candidate evidence-report before fresh-context pui-review with review-input. A v1 handoff replaces its singleton report; a v2 handoff may retain distinct historical reports with their revisions/results alongside the new report. Retention never turns prior failed, partial or unrun evidence into a current pass. Fresh independent review, trusted CI/DCO and exact-head integration remain separate conditions. Only evidence-backed numeric choice is delegated. Real unresolved product direction and privileged operations retain their normal decision boundaries; pui-govern stays read-only.

Explicit handoff

Entry into this leaf requires a digest on every incoming candidate-change, including all distinct candidate materials retained in a v2 handoff. Other leaves retain their existing artifact-digest requirements.

When blocked before a numeric edit, return the incoming candidate-change unchanged with the existing report, nextSkillId: null, and explicit blocker/refresh work in the notes. Validate that terminal using pnpm agent:skill -- --handoff <output.json> --prior-handoff <received.json>. Keep the actual received handoff: it must select this leaf and retain the same format, mode and v2 binding. Preserve every required input artifact's reference, digest and metadata; the candidate needs a digest. A changed candidate or added transaction material cannot use this stop. State that no budget edit or approval occurred; do not fabricate a transaction or a passed check.

Return exactly one handoff conforming to internal/agent-operations/schemas/skill-handoff.schema.json. Preserve capability-envelope, authority-map and implementation-authorization references and any digests, and carry the incoming evidence-report until validation refreshes it. Include the registered candidate-change, containing both the numeric edit and its supporting transaction record. After any numeric edit, set nextSkillId to pui-validate; the incoming report does not validate the changed candidate and cannot justify a direct review handoff. Use null with the unchanged incoming candidate when blocked before an edit. Do not load another leaf. An actual transaction record is evidence of a numeric edit, not a review packet, published approval or mutation authorization for integration. A blocked unchanged candidate makes no such edit claim.

The registry declares allowedNextSkillIds: [pui-validate], which the handoff validator enforces for every nonterminal return from this leaf. A terminal return instead requires the validated received handoff and unchanged required inputs; notes alone never establish that no edit occurred. Adding review-input or previously published review artifacts cannot turn the retained pre-mutation report into a direct review or integration route. After running the required candidate-bound checks, pui-validate owns the refreshed report and its normal independent-review handoff. The resolver enforces these structural relationships; it does not authenticate the contents of opaque evidence references or a forged source identity.

Communicate with the user in the user's current language. Keep paths, entry names, measurements, revisions and evidence links exact.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Implement one governed Proto UI Adapter-profile realization slice. Use when target and runtime range, supported or omitted Module role, Host Capability fidelity, lifecycle translation, failure behavior, compatibility boundary, and evidence scope are fixed by current authority. Use pui-adapter-assess for research or readiness assessment and pui-regression first for a governed failing behavior.

日本語の概要は準備中です。原文の説明を表示しています。

Proto-UI/Proto-UI362026年10月13日 更新

Assess one bounded Proto UI Adapter question without changing repository or external state. Use for new-Adapter research, target compatibility, catalog readiness, reviewed Module support or omission evidence, Host Capability fidelity, lifecycle translation, dependency risk, or implementation readiness. Do not catalog, implement, repair, or approve the Adapter slice.

日本語の概要は準備中です。原文の説明を表示しています。

Proto-UI/Proto-UI362026年10月13日 更新

Measure the active Agent model with pinned offline ModelTrace sampling and return an unsigned closed-set attribution record. Use before Agent-originated commits, Issue/PR creation or material updates, comments, reviews, and current collaboration writes in either execution mode. Do not treat a fingerprint, declared model label, or record reference as authenticated identity, permission, independent review, or acceptance.

日本語の概要は準備中です。原文の説明を表示しています。

Proto-UI/Proto-UI362026年10月13日 更新

Run Proto UI's dynamic comprehension assessment and derive an unsigned local task-fit result. Use when an Agent needs review calibration or a fresh autonomous task and review ceiling. Do not use a self-result as identity, permission, or acceptance evidence.

日本語の概要は準備中です。原文の説明を表示しています。

Proto-UI/Proto-UI362026年10月13日 更新

Investigate a bounded Proto UI Contract, Prototype, Module, Host Capability, Adapter, schema, or ownership question and turn a genuinely unresolved direction into one semantic decision packet. Use for spec brainstorming, admission questions, competing semantic models, uncataloged behavior, or unresolved cross-layer boundaries.

日本語の概要は準備中です。原文の説明を表示しています。

Proto-UI/Proto-UI362026年10月13日 更新

pui-ci

無料

Diagnose one bounded Proto UI Actions or CI failure from an exact workflow run and revision, then route the exact repair, validation, or recheck transition. Use to locate the owning failure and distinguish infrastructure from change failures. This leaf remains read-only.

日本語の概要は準備中です。原文の説明を表示しています。

Proto-UI/Proto-UI362026年10月13日 更新

Proto-UI のスキルをすべて見る

このスキルの問題を報告する