朱子纯 / erp-workflow-plugin

cb9da34e 转 coding.mjs Workflow 时把 auto-continue Stop 钩子连同 17 个旧
coding 子技能一起删了，导致逐 REQ 手动编码会话每个 REQ 完就停下问 continue。
本提交恢复该机制：

- hooks/scripts/auto-continue.sh（Stop）：回合中途结束默认 block 并喂「继续
  推进」，使各 REQ 子任务连跑、用户无需手敲 continue；仅当 transcript 出现
  [ERP-HALT] 或连续空转达上限(8)/无 >4000B 增长时放行停下。
- hooks/scripts/reset-stall-counter.sh（PostToolUse）：有工具调用即清零空转
  计数，真推进时无限续跑。
- hooks/hooks.json：接线 Stop + PostToolUse（不含已删的 log-cross-module.sh）。
- coding-start 步骤5「Workflow 已启动」横幅补 [ERP-HALT]：Workflow 转后台后
  主循环干净停下等通知，不被钩子误推去重启 coding-start。
- .gitattributes：强制 *.sh / hooks.json 用 LF，避免 autocrlf 把 CRLF 注入
  shebang 与钩子输出的 JSON。

钩子在会话启动时加载，需重启已运行会话才生效（开发仓即运行源，无需 commit
即生效，commit 仅为持久化）。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

P1#2 — validate-ddl 支持增量 migration：
- parseDDL 新增第三遍 mergeAlterStatements：ALTER TABLE ADD COLUMN /
  ADD [UNIQUE] INDEX|KEY / ADD PRIMARY KEY 并入对应表（仅追加式，
  MODIFY/CHANGE/DROP/RENAME 不处理；并集与文件顺序无关）。
- CLI 接受多个 DDL 文件：node validate-ddl docs/03 sql/migrations/V*.sql，
  CREATE + 各 V_n 的 ALTER ADD 并集 ↔ docs/03 累积 SSoT 做 4 维比对。
- add-req SKILL 新增「增量 DDL 校验(fail-closed)」闸门，替换原「多 migration 不适用」说明。
- 7 个新测试（含 DEFAULT 字面量里的 ALTER 文本不被误吃）；52/52 通过。

P1#3 — coding 首跑自动建需求台账基线：
- coding-start 透传 pluginRoot；coding.mjs ensureLedgerBaseline()
  首跑若 .req-ledger.json 缺失则 commit 基线（幂等、best-effort、缺 pluginRoot 静默跳过），
  使日后 /add-req 能识别增量、免空跑。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

中断/halt 后重跑能复盘上次状态，不只恢复进度。编码阶段保持全自动静默不变。

- coding.mjs：新增 resumeJournalPromptM + recordResume（best-effort，写失败绝不阻断/
  掩盖主流程）。主循环结束（halt 或全完成）时向 docs/superpowers/RESUME.md 追加一条：
  halt 原因 + 本次自主默认假设(decisions 摘要) + 待跑模块 + 下一步
- coding-start：新增步骤 3.5，重跑时读 RESUME.md 末尾条目向用户复盘上次 halt/完成状态
- 进度真值仍是 git tag（milestone/req-done）+ committed 工件——硬中断后 Router 据此续跑；
  RESUME.md 补的是跨会话会丢失的「为何停/做过哪些假设」定向信息（GSD continue-here.md 类比）
- 硬中断(进程被杀)到不了写入点：那种情况无 halt 原因可记，且 tag+工件已够续跑，无信息损失

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- coding-start / plan-start：去掉已失效的「git 就绪由本入口预检/校验」表述（git 闸已移入 coding.mjs 运行时）
- skeleton-gen：删除「A2 不再单独出 docs/06/07/09」「原 docs/09 并入」等历史演进注释
- downstream-gen：描述去掉「A6 已并入此处」历史包袱
- CLAUDE-template：补全被截断的 migration 规约残句（旧 V_n.sql 保持原样、永不回改）
- README：修正悬空的「🧭 通用工作准则」引用、删「4 条通用准则」、模板计数 26→19、coding-start git 就绪描述

行为门 v3（docs/design/2026-06-05-frontend-behavior-stage-gate.md）：
- 行为验收从 per-FE review approve 子门迁回阶段末尾一次（phase Behavior，
  featureLoop 后、testGate 前），保留 fix 循环（BEHAVIOR_STAGE_MAX=3 轮，
  fix 后全量前端单测复验再重跑门）
- req-done/<FE> 语义降为「仅静态 review 过」；行为绿改由 reportPrompt 校验
  阶段级证据（module-reports/frontend-phase-behavior-r*-a*.md 最后一份非 RED）
- build-failed 取消「兄弟未实现」短路（阶段末尾全 FE 已实现）；断言作用域 =
  全部 FE spec「行为验收作用域」小节并集，缺小节记 scope-missing 阻断 green
- 新增样式层 styleIssues（颜色 token 比对 + layout sanity 共 6 kind），
  降维并入 behaviorHard 与交互硬问题同口径 fix；环境仲裁透传 riders 计数

前端测试目录隔离（对齐后端 src/main↔src/test 物理分离）：
- 锁定约定：单测 frontend/tests/** 镜像 src/（smoke 归 tests/__smoke__/ 且
  以 .test.* 结尾），e2e 在 frontend/e2e/，frontend/src/ 禁任何测试产物；
  vitest include 统一限定 tests/**/*.test.*
- 五层防线：docs-04 模板 §2.1 锁定约定 / planPrompt+tddPrompt 硬护栏 /
  fe-skeleton 单测基线 / code-reviewer 新增第 8 维「测试文件隔离」
- legacy 守卫：frontend/src/ 内已存在 colocated 测试时绝不收窄 include
  （防旧单测静默停跑），骨架幂等检测同步豁免，留人工迁移

经两轮多代理对抗审计（34 agents），确认项均已修复。

Replaces the phase-level read-only behavior-gate with a per-FE acceptance dimension: each FE is approved only when the code-reviewer approves AND runtime behavior verification is green. Behavior defects (dead control / sentinel text mismatch) become fixable must-fix that drive verify->fix->re-verify, not halts.

- reviewWithFixLoop (frontend only, via if(fe)): at the approve gate, behaviorSubGate boots this FE's full stack + seeds sentinels, enumerates this FE's routes, two-tier asserts. Hard issues with a locator -> fixPrompt -> functional reverify -> next behaviorRound; soft text (i18n/literal/semantic) -> adjudicate(continue); behaviorRound bounded by BEHAVIOR_FE_MAX=3, env race by BEHAVIOR_ATTEMPT_MAX=2. Backend featureLoop branch unchanged.

- New runFrontendSkeleton stage (before featureLoop(frontend)): App shell + full lazy router + FeStub placeholders + shared nav, so the app is buildable at every mid-phase point; tdd swaps FeStub->real component per FE. Idempotent via fe-skeleton-done tag.

- BEHAVIOR_GATE_SCHEMA gains build-failed envError kind (sibling-FE-unimpl short-circuit, not a bug) + locator-not-resolvable coverage reason; deriveSpec emits a per-FE route-scope section, reviewer validates it.

- Removed phase-level runBehaviorGate + 'Behavior' phase; kept phase-level testGate (regression). REVIEW_HARD_ROUNDS 8->10.

- Safety: test-DB naming guard pushed into scripts-setup-test-db.mjs template (fail-closed unless name contains test/_dev/_local or ALLOW_NONTEST_DROP=1) + 3 tests.

- agentType stays erp-workflow:code-reviewer. v1 design doc marked SUPERSEDED; v2 design at docs/design/2026-06-02-frontend-behavior-in-review-loop.md.

Verified: wrapped syntax check SYNTAX_OK, 87/87 lib tests pass, no orphan refs, no time/random builtins, top-level return intact. Not yet run end-to-end against a real ERP project.

New 'Behavior' stage between Gate and Milestone, frontend-phase only, after testGate green and before report/milestone. Verifies every interactive control actually works and every text region shows the right content, independent of the tests the tdd agent wrote.

- BEHAVIOR_GATE_SCHEMA / behaviorGateContract() / behaviorGatePrompt() / runBehaviorGate(); reportPrompt now gates the milestone on behavior-gate evidence (frontend-phase-behavior-gate-r*.md, last attempt must be non-RED).

- Two-tier failure: interaction defects (incl. binding-garbage) flake-retry once then hard-halt via adjudicate(allowContinue:false); text issues split by source (sentinel=objective -> no continue; i18n/literal/semantic=adjudicable). Convergence loop re-runs the env + interaction hard gates after any text-layer retry so a refreshed result can't slip a green past the hard gates.

- Full-stack seeded run: test-DB name guard (deterministic halt, not adjudicated), strict 4-phase ordering (empty DB -> boot backend so Flyway builds schema -> seed -> boot frontend), auth bootstrap via storageState, router-config-driven route discovery, ephemeral .tmp/behavior-gate runner with finally teardown, type-legal per-field-unique sentinels.

- agentType uses the plugin-namespaced 'erp-workflow:code-reviewer' (a bare 'code-reviewer' is ambiguous with feature-dev:code-reviewer); README + agents/code-reviewer.md aligned (frontmatter name: stays bare).

- Design: docs/design/2026-06-02-frontend-behavior-gate.md. README + coding-start banner updated.

…n-check + projectRoot

git 就绪（默认分支/工作树干净/功能分支创建与合并）改由 coding.mjs runBranchSetup 运行时把守；
入口只校验 Plan 完成态（docs/08 §一 全勾）并取 projectRoot 绝对路径传入 Workflow。
相应收窄 allowed-tools（去掉 git status/branch/checkout 授权）。

- delete frontend-scope-lock skill + docs/06 templates (static + module-pagelist)
- A5 downstream-gen gains prototype/ gate + FE-list derivation -> docs/08 § 三
- plan-start: drop A6 dispatch row, gate 5->4 items; coding-start A0~A6 -> A0~A5
- Design Tokens single source = src/styles/tokens.css (code-reviewer, coding.mjs)
- plugin.json skills 9->8; docs-09 tree drops 06-UI; docs-08 drops A6 section

Visual clarity: the file tree now matches the README's phase model.
- skills/plan/   — 8 skills (plan-start entry + A0~A6)
- skills/coding/ — 1 skill (coding-start thin entry)

Slugs unchanged. /erp-workflow:<name> still resolves identically because
each SKILL.md declares its own `name:` frontmatter (verified for all 9).

Plugin loader treats the default `./skills/` scan as flat, so the nested
layout is opted into via an explicit path array in plugin.json — the only
mechanism Claude Code provides for nested skill discovery.

Changes:
- mv skills/{plan-start,project-init,scope-lock,skeleton-gen,
       db-design-gen,db-init,downstream-gen,frontend-scope-lock}
     → skills/plan/<same>/
- mv skills/coding-start → skills/coding/coding-start/
- .claude-plugin/plugin.json: skills[] now lists all 9 explicit paths
- README.md: directory tree + Skill 清单 heading updated to reflect grouping

No code touches any skill path (only README did, and only in the tree
diagram). lib tests still pass (62/62).