c9a6777340ca79fbabc054c4712c8dc409cc5e3c
Based on real execution review (lookup-knowledge-integration), discovered that constraints are "soft" - rules are clear but lack enforcement mechanism. Core problem: Agent can skip stages despite explicit rules saying "must not skip" Changes: 1. Commit phase: add verifiable checkpoint - File completeness check: proposal ≥50 words, design ≥1 data structure, specs ≥3 requirements, tasks ≥5 items - Consistency check: proposal concepts → design mapping, design decisions → tasks implementation - Gate file: create .committed after passing all checks 2. Apply phase: add pre-gate check - Hard constraint: check .committed file existence - If missing: report, list gaps, ask user to fix or explicitly skip - Block implementation based on incomplete OpenSpec 3. Archive phase: add mandatory execution order - 5-step checklist: devflow files → index → .archive-ready → report → archive - Self-check: verify Step 1-3 before Step 4 - Prevent "handoff first, devflow forgotten" issue New gate files: - .committed: created by commit phase, checked by apply phase - .archive-ready: created by archive Step 3 Expected impact: - Commit skip rate: -100% (explicit checklist prevents fuzzy pass) - Apply on incomplete OpenSpec: -100% (gate blocks) - Archive devflow omission: -100% (mandatory order) Design principles: - Verifiability: "executable state" → "≥50 words + ≥1 structure + ≥3 requirements" - Gate files: soft judgment → file existence check - Mandatory order: advisory "should do" → 5-step checklist "must do" Complexity: +80 lines (phase-contracts.md +40, archive-rules.md +40) Approach: turn soft constraints into hard checks, no new stages Relation to v4.1: - v4.1: solve "insufficient research causes rework" (quality issue) - v4.2: solve "lack of enforcement causes stage skipping" (process issue) Docs: - phase-contracts.md: enhanced commit + apply phases - archive-rules.md: added mandatory execution order at top - workflow.md: added v4.2 evolution chapter - phase-contracts-v4.2-changelog.md: detailed change log - sm-flow-optimization-suggestions.md: source review
github-learn
个人知识学习与 agent workflow 实验仓库。
当前核心内容
| 路径 | 作用 |
|---|---|
knowledge/entries/knowledge_YYYYMMDD_*/ |
知识条目源文件与渲染 HTML |
knowledge-index.html |
知识索引页面,由 scripts/update-knowledge-index.sh 生成 |
scripts/ |
本仓库维护脚本 |
openspec/ |
当前 OpenSpec 变更工作区与归档 |
devflow/ |
人类可读项目档案、glossary、复合知识 |
.agents/skills/ |
Codex/agent 使用的本地运行时 skills,当前重点是 sm-flow |
.claude/skills/ |
Claude 版本 skills 与 OpenSpec skills |
skill-workbench/ |
skill 创建、设计文档、历史文档与验证沙箱 |
上下文入口
- 领域词汇表真理源:
devflow/glossary/CONTEXT.md - 根目录
CONTEXT.md仅保留兼容性重定向。
历史内容
skill-workbench/generated-skills/sm-flow/:本仓库新增的sm-flowskill 源产物,只放可安装 skill 本体。skill-workbench/generated-skills/skills/:早期 Explore Skill Family,只放 skill 本体。skill-workbench/docs/sm-flow/workflow.md:sm-flow的设计演进文档。skill-workbench/docs/explore-essence-follow/latest-design.md:explore、essence、follow三个学习型 skill 的最新设计总结。.agents/skills/sm-flow/:sm-flow的运行时安装副本,用于当前 Codex 会话加载。skill-workbench/history/changelog/:早期 skill family 变更记录。skill-workbench/history/docs/:早期设计、OpenSpec 与 agent 文档。skill-workbench/validation/project/、skill-workbench/validation/tmp2/:验证 skill 时使用的第三方源码和临时报告。skill-workbench/validation/skills-lock.json:验证安装 mattpocock skills 时生成的锁文件。
Languages
HTML
72.6%
Shell
13.9%
PowerShell
7%
Python
6.5%