Skip to main content

gentle-ai-rdd-defect-workflow

Trigger: RDD, receipt-driven development, review authority, receipt/lineage, correction/recovery, delivery gate/kill switch, bounded review defects. Guide work.

ソース情報

リポジトリ
Gentleman-Programming/gentle-shell
ソースの最終更新活動
2026年8月21日 18:20
検出された SKILL.md の言語
英語
スター
1,486
フォーク
234

インストール方法

デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。

ソースファイルを確認

インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。

SKILL.md を表示中

SKILL.md
ソースの指示 · 読み取り専用プレビュー
name
gentle-ai-rdd-defect-workflow
description
Trigger: RDD, receipt-driven development, review authority, receipt/lineage, correction/recovery, delivery gate/kill switch, bounded review defects. Guide work.
license
Apache-2.0
metadata
{"author":"gentleman-programming","version":"1.0"}
## Activation Contract Load when the frontmatter trigger terms apply to a defect workflow. This skill guides public collaboration. It does not grant issue approval, label, review, exception, or merge authority. ## Hard Rules - Review and Judgment Day evidence is review-only. Pi never mints delivery authority: ordinary commit, push, PR, and release always follow repository policy, regardless of RDD mode. - Require an approved issue (`status:approved`) and clean current `main` reproduction before implementation. Audit existing PRs for supersession or conflict; stop or narrow stale claims. - Group by causal authority invariant. Use one issue and one PR or explicit chain per independent invariant and rollback boundary. Split independent causes; never merge a superseded or conflicting authority line. - Inventory every operator flow claimed by the issue or PR, including entry, mode, environment, expectation, and negative controls. Require one truthful black-box bench journey per CLI or lifecycle flow, or actual runtime E2E proof when the core bench cannot represent it. Synthetic proxy coverage never proves another runtime. - Use CodeGraph-first impact mapping, a dedicated worktree, and behavior-first tests. Run source-mutating normalization before candidate freeze. - Forecast authored changes before edits. The hard limit is 400 additions plus deletions; above it, STOP for a chain or explicit maintainer-approved exception. - When RDD is enabled, bind review receipts, lineage, correction, and recovery evidence to the exact candidate. Keep bounded review defects in one correction transaction; never treat that evidence as delivery authority. - Require independent read-only candidate validation before publication. Validation cannot edit source or authority; findings require a new candidate. - Keep communication humane and evidence-based. Repository labels and workflow metadata are maintainer-owned, never evidence of contributor blame. ## Decision Gates | Condition | Action | | --- | --- | | Any RDD mode | Review evidence remains review-only; ordinary commit, push, PR, and release follow repository policy with no Pi delivery authority. | | Issue gate or reproduction fails | Wait, stop, or narrow with evidence. | | Invariant or rollback is independent | Separate issue and authoritative PR line. | | Core bench fits / does not fit | Bench journey / actual runtime E2E; never proxy. | | Forecast exceeds 400 lines | Chain or approved exception before edits. | ## Execution Steps 1. Check mode, approval, PR conflicts, and current-main reproduction. 2. Name invariant and rollback; isolate the worktree; CodeGraph-map code, tests, evidence, docs, distribution, and registration. 3. Inventory flows and controls; add failing tests and the smallest correction. 4. Normalize, enforce budget, run tests, and record each flow's exact candidate, command, scenario, and result. 5. Freeze, validate read-only, and give the verdict, evidence, and one humane next action. ## Output Contract Return `rdd_mode`, `issue_pr`, `causal_invariant`, `operator_flows`, `journey_runtime_evidence`, `changed_line_budget`, `tests`, `rollback`, and `unresolved_authority_decisions`. Identify approved and superseded/conflicting authority lines; every flow, negative control, and candidate-bound proof; additions plus deletions and chain/exception; test results; independent rollback; and unresolved maintainer decisions. ## References No supporting files. Current repository policy remains authoritative.
GitHubで見る