gsd-lite 0.4.1 → 0.5.0

package/.claude-plugin/marketplace.json

@@ -13,7 +13,7 @@
       "name": "gsd",
       "source": "./",
       "description": "AI orchestration tool — GSD management shell + Superpowers quality core. 5 commands, 4 agents, 5 workflows, MCP server, context monitoring.",
-      "version": "0.4.1",
+      "version": "0.5.0",
       "keywords": [
         "orchestration",
         "mcp",

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "gsd",
-  "version": "0.4.1",
+  "version": "0.5.0",
   "description": "AI orchestration tool for Claude Code — GSD management shell + Superpowers quality core",
   "author": {
     "name": "sdsrss",

package/agents/executor.md

@@ -55,7 +55,10 @@ tools: Read, Write, Edit, Bash, Grep, Glob
   "decisions": ["[DECISION] use optimistic locking by version column"],
   "blockers": [],
   "contract_changed": true,
-  "evidence": ["ev:test:users-update", "ev:typecheck:phase-2"]
+  "evidence": [
+    {"id": "ev:test:users-update", "scope": "task:2.3"},
+    {"id": "ev:typecheck:phase-2", "scope": "task:2.3"}
+  ]
 }
 `contract_changed` 判定指南:
 - 改了函数/方法签名 (参数、返回类型) → true

package/agents/reviewer.md

@@ -92,7 +92,7 @@ Minor = 建议修复 (命名/风格)
 ```json
 {
   "scope": "task | phase",
-  "scope_id": "2.3 | phase-2",
+  "scope_id": "2.3 (task scope: string ID) | 2 (phase scope: number ID)",
   "review_level": "L2 | L1-batch",
   "spec_passed": true,
   "quality_passed": false,
@@ -107,7 +107,10 @@ Minor = 建议修复 (命名/风格)
   "minor_issues": [],
   "accepted_tasks": [],
   "rework_tasks": ["2.3", "2.4"],
-  "evidence": ["ev:test:phase-2", "ev:lint:phase-2"]
+  "evidence": [
+    {"id": "ev:test:phase-2", "scope": "task:2.3"},
+    {"id": "ev:lint:phase-2", "scope": "task:2.3"}
+  ]
 }
 ```
 

package/commands/prd.md

@@ -51,99 +51,9 @@ argument-hint: File path to requirements doc, or inline description text
 - 使用 references/questioning.md 的提问技巧 (如可用)
 - 用户回答后，可适当追问直到需求清晰
 
-<!-- 以下 STEP 5-12 同 start.md -->
+<!-- STEP 5-12: 共享执行流程 — 修改 workflows/execution-flow.md 即同步所有入口 -->
 
-## STEP 5: 智能判断是否需要研究
-
-- 新项目 / 涉及新技术栈 → 必须研究
-- 简单 bug 修复 / 已有研究且未过期 → 跳过
-- 用户明确要求 → 研究
-- 需要时 → 派发 researcher 子代理 → 展示关键发现
-- 不需要 → 跳过，进入下一步
-
-## STEP 6: 深度思考
-
-- 如有 sequential-thinking MCP → 调用深入思考
-- 无则跳过，不影响流程
-
-## STEP 7: 生成分阶段计划
-
-- phase 负责管理与验收，task 负责执行
-- 每阶段控制在 5-8 个 task (便于 phase-level 收口)
-- 每个 task = 原子化 todo (含文件、操作、验证条件)
-- 每个 task 补充元数据: `requires` / `review_required` / `research_basis`
-- 审查级别按影响面判定: L0(无运行时语义变化) / L1(普通) / L2(高风险)
-- 标注可并行任务组 [PARALLEL] (当前仅作未来升级标记)
-
-## STEP 8: 计划自审
-
-轻量替代 plan-checker:
-- 检查: 是否有遗漏的需求点？
-- 检查: 阶段划分是否合理？(phase 过大则拆分)
-- 检查: 任务依赖关系是否正确？
-- 检查: 验证条件是否可执行？
-- 如属高风险项目 → 升级为增强计划审查:
-
-<enhanced_plan_review>
-触发条件: 涉及 auth / payment / security / public API / DB migration / 核心架构变更
-
-审查维度:
-1. 需求覆盖: 原始需求的每个要点是否都映射到了至少一个 task？
-2. 风险排序: 高风险 task 是否排在前面？(fail-fast 原则)
-3. 依赖安全: L2 task 的下游是否都用了 gate:accepted？
-4. 验证充分: 涉及 auth/payment 的 task 是否都有明确的安全验证条件？
-5. 陷阱规避: research/PITFALLS.md 中的每个陷阱是否都有对应的防御 task 或验证条件？
-
-输出: pass / revise (附具体修正建议)
-轮次: 最多 2 轮自审修正；2 轮后仍有问题 → 标注风险展示给用户
-</enhanced_plan_review>
-
-→ 自审修正后再展示给用户
-
-<HARD-GATE id="plan-confirmation">
-## STEP 9: 展示计划，等待用户确认
-
-- 展示完整分阶段计划
-- 用户指出问题 → 调整 → 再展示
-- 用户确认 → 继续
-
-⛔ 不得在用户确认前执行 STEP 10-12。未确认 = 不写文件、不执行代码。
-</HARD-GATE>
-
-## STEP 10: 生成文档
-
-- 创建 .gsd/ 目录
-- 写入 state.json + plan.md + phases/*.md
-- 初始化 `workflow_mode` / `current_task` / `current_review` / phase 状态与 handoff 信息
-- 如有研究: 写入 .gsd/research/
-
-<HARD-GATE id="docs-written">
-□ state.json 已写入且包含所有 canonical fields
-□ plan.md 已写入
-□ phases/*.md 已写入 (每个 phase 一个文件)
-□ 所有 task 都有 lifecycle / level / requires / review_required
-→ 全部满足才可继续
-</HARD-GATE>
-
-## STEP 11 — 自动执行主路径
-
-进入执行主循环。phase = 管理边界，task = 执行边界。
-
-<execution_loop>
-参考 `references/execution-loop.md` 获取完整 9 步执行循环规范 (11.1-11.9) 及依赖门槛语义。
-
-编排器必须严格按照该参考文档中的步骤顺序执行:
-加载 phase → 选择 task → 构建上下文 → 派发 executor → 处理结果 → 审查 → phase handoff → 批量更新 → 上下文检查
-</execution_loop>
-
-## STEP 12 — 全部完成
-
-全部 phase 完成后，输出最终报告:
-- 项目总结
-- 各阶段完成情况
-- 关键 decisions 汇总
-- 验证 evidence 汇总
-- 遗留问题 / 后续建议 (如有)
+使用 Read 工具读取 `workflows/execution-flow.md`，严格按照其中 STEP 5-12 执行。
 
 </process>
 

package/commands/resume.md

@@ -212,6 +212,79 @@ description: Resume project execution from saved state with workspace validation
 
 所有展示数据从 canonical fields 实时推导，不使用 derived fields。
 
+## STEP 5: 自动执行循环
+
+<HARD-GATE id="auto-execution-loop">
+STEP 3 完成初次恢复后，进入自动执行循环。这是编排器的核心 —— 不要停在某一步等待用户，除非遇到终止条件。
+
+```
+循环入口:
+  1. 调用 MCP tool `orchestrator-resume` 获取 action
+  2. 根据 action 分派:
+
+     dispatch_executor:
+       → 使用 Agent tool 派发 executor 子代理 (subagent_type: gsd:executor)
+       → 传入 orchestrator 返回的 executor_context
+       → 收到 executor 结果后 → 调用 MCP tool `orchestrator-handle-executor-result`
+       → 回到步骤 1
+
+     dispatch_reviewer:
+       → 使用 Agent tool 派发 reviewer 子代理 (subagent_type: gsd:reviewer)
+       → 传入 review_targets / current_review
+       → 收到 reviewer 结果后 → 调用 MCP tool `orchestrator-handle-reviewer-result`
+       → 回到步骤 1
+
+     dispatch_researcher:
+       → 使用 Agent tool 派发 researcher 子代理 (subagent_type: gsd:researcher)
+       → 传入过期的研究信息
+       → 收到 researcher 结果后 → 调用 MCP tool `orchestrator-handle-researcher-result`
+       → 回到步骤 1
+
+     dispatch_debugger:
+       → 使用 Agent tool 派发 debugger 子代理 (subagent_type: gsd:debugger)
+       → 传入 debug_target
+       → 收到 debugger 结果后 → 调用 MCP tool `orchestrator-handle-debugger-result`
+       → 回到步骤 1
+
+     trigger_review:
+       → 直接派发 reviewer，scope 和 targets 从 action 响应中获取
+       → 回到步骤 1
+
+     complete_phase:
+       → 调用 MCP tool `phase-complete`，传入 phase_id + run_verify: true
+       → 回到步骤 1 (编排器会自动推进到下一 phase)
+
+     retry_executor:
+       → 重新调用 orchestrator-resume 获取更新后的 executor 上下文
+       → 回到步骤 1
+
+     rollback_to_dirty_phase:
+       → 编排器已自动回滚 current_phase，输出回滚通知
+       → 回到步骤 1
+
+     continue_execution:
+       → 直接回到步骤 1
+
+  3. 终止条件 — 遇到以下 action 时退出循环:
+
+     idle              → 输出 "无可执行任务"，停止
+     awaiting_user     → 展示 blockers / drift 信息，等待用户输入
+     await_manual_intervention → 展示需要人工干预的信息，停止
+     noop (completed)  → 展示完成报告，停止
+     await_recovery_decision (failed) → 展示失败信息和恢复选项，停止
+
+  4. 上下文安全阀:
+     每次循环迭代前检查上下文健康度
+     remaining <= 35% → 保存状态 + 输出 "请 /clear 后 /gsd:resume" → 退出循环
+```
+
+**关键原则:**
+- 循环是连续的: dispatch → handle result → resume → dispatch → ...
+- 不在中间步骤停下来等用户确认（除非是终止条件）
+- 每次 handle result 后立即 resume，让编排器决定下一步
+- Phase 审查通过后 → complete_phase → 自动推进下一 phase → 继续执行
+</HARD-GATE>
+
 </process>
 
 <EXTREMELY-IMPORTANT>
@@ -221,4 +294,5 @@ description: Resume project execution from saved state with workspace validation
 - awaiting_user / reconcile_workspace / replan_required 模式下不自动执行代码
 - 只有编排器写 state.json，子代理不直接写
 - 上下文 < 35% → 保存状态 + workflow_mode = awaiting_clear + 停止执行
+- **进入自动执行循环后，不要在循环中间停下来等用户 — 让编排器驱动**
 </EXTREMELY-IMPORTANT>

package/commands/start.md

@@ -44,133 +44,8 @@ argument-hint: Optional feature or project description
     └── 否 → 追问
 ```
 
-## STEP 5 — 智能研究判断
+<!-- STEP 5-12: 共享执行流程 — 修改 workflows/execution-flow.md 即同步所有入口 -->
 
-判断是否需要研究:
-```
-├── 新项目                    → 必须研究
-├── 涉及新技术栈              → 必须研究
-├── 简单 bug 修复 / 小功能    → 跳过研究
-├── 已有 .gsd/research/ 且未过期 → 跳过研究
-├── 用户明确要求               → 研究
-└── 已有研究但需求方向变了     → 增量研究 (只研究新方向)
-```
-
-需要研究时:
-1. 派发 `researcher` 子代理 (新鲜上下文)
-2. 研究输出写入 `.gsd/research/` (STACK.md, ARCHITECTURE.md, PITFALLS.md, SUMMARY.md)
-3. 向用户展示关键发现: 技术栈推荐 + 陷阱警告 + ⭐ 推荐方案
-
-不需要时: 跳过，直接进入 STEP 6。
-
-## STEP 6 — 深度思考
-
-如有 `sequential-thinking` MCP 可用 → 调用深入思考:
-- 输入: 需求摘要 + 代码库分析 + 研究结果 (如有)
-- 目的: 在生成计划前进行系统性架构思考
-
-如无 `sequential-thinking` MCP → 降级为内联思考，继续。
-
-## STEP 7 — 生成分阶段计划
-
-生成 plan.md + phases/*.md:
-- **phase** 负责管理与验收，**task** 负责执行
-- 每阶段控制在 **5-8 个 task** (便于 phase-level 收口)
-- 每个 task = 原子化 todo (含文件、操作、验证条件)
-- 每个 task 补充元数据:
-  - `requires` — 依赖列表 (含 gate 类型)
-  - `review_required` — 是否需要审查
-  - `research_basis` — 引用的 research decision id
-- 审查级别按影响面判定:
-  - **L0** — 无运行时语义变化 (docs/config/style)
-  - **L1** — 普通编码任务 (默认)
-  - **L2** — 高风险 (auth/payment/public API/DB migration/核心架构)
-- 标注可并行任务组 `[PARALLEL]` (当前仅作未来升级标记)
-
-## STEP 8 — 计划自审
-
-轻量自审 (编排器自身执行，不派发子代理):
-
-### 基础审查 (所有项目)
-- [ ] 是否有遗漏的需求点？
-- [ ] 阶段划分是否合理？(phase 过大则拆分)
-- [ ] 任务依赖关系是否正确？
-- [ ] 验证条件是否可执行？
-
-### 增强审查 (高风险项目)
-
-触发条件: 项目涉及 auth / payment / security / public API / DB migration / 核心架构变更
-
-维度:
-1. **需求覆盖:** 原始需求的每个要点是否都映射到了至少一个 task？
-2. **风险排序:** 高风险 task 是否排在前面？(fail-fast 原则)
-3. **依赖安全:** L2 task 的下游是否都用了 `gate:accepted`？
-4. **验证充分:** 涉及 auth/payment 的 task 是否都有明确的安全验证条件？
-5. **陷阱规避:** `research/PITFALLS.md` 中的每个陷阱是否都有对应的防御 task 或验证条件？
-
-输出: `pass` / `revise` (附具体修正建议)
-轮次: 最多 2 轮自审修正；2 轮后仍有问题 → 标注风险展示给用户
-
-→ 自审修正后再展示给用户。
-
-<HARD-GATE id="plan-confirmation">
-## STEP 9 — 用户确认计划
-
-展示计划给用户，等待确认:
-- 用户指出问题 → 调整计划 → 重新展示
-- 用户确认 → 继续
-
-⛔ 不得在用户确认前执行 STEP 10-12。未确认 = 不写文件、不执行代码。
-</HARD-GATE>
-
-<HARD-GATE id="docs-written">
-## STEP 10 — 生成文档
-
-1. 创建 `.gsd/` 目录
-2. 写入 `state.json`:
-   - 初始化 `workflow_mode: "executing_task"`
-   - 初始化 `current_phase: 1`
-   - 初始化 `current_task: null` (由执行循环填充)
-   - 初始化 `current_review: null`
-   - 初始化所有 phase lifecycle = `pending` (第一个 = `active`)
-   - 初始化所有 task lifecycle = `pending`
-   - 初始化 phase_handoff 信息
-   - 初始化 `decisions: []`
-   - 初始化 `context.remaining_percentage`
-3. 写入 `plan.md` — 项目总览索引 (不含 task 级细节)
-4. 写入 `phases/*.md` — 每阶段详细 task 规格 (source of truth)
-5. 如有研究: 确认 `.gsd/research/` 已写入
-
-规则:
-- `plan.md` 是只读索引: 生成后不再修改 (除非 replan)
-- `phases/*.md` 是 task 规格的唯一 source of truth
-- `plan.md` 不包含 task 级细节，避免与 `phases/*.md` 重复
-
-□ state.json 已写入且包含所有 canonical fields
-□ plan.md 已写入
-□ phases/*.md 已写入 (每个 phase 一个文件)
-□ 所有 task 都有 lifecycle / level / requires / review_required
-→ 全部满足才可继续
-</HARD-GATE>
-
-## STEP 11 — 自动执行主路径
-
-进入执行主循环。phase = 管理边界，task = 执行边界。
-
-<execution_loop>
-参考 `references/execution-loop.md` 获取完整 9 步执行循环规范 (11.1-11.9) 及依赖门槛语义。
-
-编排器必须严格按照该参考文档中的步骤顺序执行:
-加载 phase → 选择 task → 构建上下文 → 派发 executor → 处理结果 → 审查 → phase handoff → 批量更新 → 上下文检查
-</execution_loop>
-
-## STEP 12 — 最终报告
-
-全部 phase 完成后，输出最终报告:
-- 项目总结
-- 各阶段完成情况
-- 关键 decisions 汇总
-- 验证 evidence 汇总
-- 遗留问题 / 后续建议 (如有)
+使用 Read 工具读取 `workflows/execution-flow.md`，严格按照其中 STEP 5-12 执行。
 
 </process>

package/hooks/gsd-statusline.cjs

@@ -56,8 +56,8 @@ process.stdin.on('end', () => {
     }
 
     // Context window display (USED percentage scaled to usable context)
-    // Claude Code reserves ~16.5% for autocompact buffer
-    const AUTO_COMPACT_BUFFER_PCT = 16.5;
+    // Claude Code reserves ~16.5% for autocompact buffer (configurable via env)
+    const AUTO_COMPACT_BUFFER_PCT = Number(process.env.GSD_AUTOCOMPACT_BUFFER) || 16.5;
     let ctx = '';
     if (remaining != null) {
       const usableRemaining = Math.max(0, ((remaining - AUTO_COMPACT_BUFFER_PCT) / (100 - AUTO_COMPACT_BUFFER_PCT)) * 100);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gsd-lite",
-  "version": "0.4.1",
+  "version": "0.5.0",
   "description": "AI orchestration tool for Claude Code — GSD management shell + Superpowers quality core",
   "type": "module",
   "bin": {

package/src/schema.js

@@ -19,7 +19,7 @@ export const WORKFLOW_MODES = [
 
 export const TASK_LIFECYCLE = {
   pending:              ['running', 'blocked'],
-  running:              ['checkpointed', 'blocked', 'failed'],
+  running:              ['checkpointed', 'blocked', 'failed', 'accepted'], // accepted: auto-accept for L0/review_required=false (atomic, skips checkpointed)
   checkpointed:         ['accepted', 'needs_revalidation'],
   accepted:             ['needs_revalidation'],
   blocked:              ['pending'],
@@ -106,7 +106,7 @@ export function validateResearchDecisionIndex(decisionIndex, requiredIds = []) {
   return { valid: errors.length === 0, errors };
 }
 
-export function validateResearchArtifacts(artifacts, { decisionIds = [], volatility, expiresAt } = {}) {
+export function validateResearchArtifacts(artifacts) {
   const errors = [];
   if (!isPlainObject(artifacts)) {
     return { valid: false, errors: ['artifacts must be an object'] };
@@ -119,19 +119,6 @@ export function validateResearchArtifacts(artifacts, { decisionIds = [], volatil
     }
   }
 
-  const summary = typeof artifacts['SUMMARY.md'] === 'string' ? artifacts['SUMMARY.md'] : '';
-  if (volatility && !summary.includes(volatility)) {
-    errors.push('artifacts.SUMMARY.md must mention volatility');
-  }
-  if (expiresAt && !summary.includes(expiresAt)) {
-    errors.push('artifacts.SUMMARY.md must mention expires_at');
-  }
-  for (const id of decisionIds) {
-    if (!summary.includes(id)) {
-      errors.push(`artifacts.SUMMARY.md must mention decision id ${id}`);
-    }
-  }
-
   return { valid: errors.length === 0, errors };
 }
 
@@ -452,6 +439,13 @@ export function validateState(state) {
       if (!Number.isFinite(phase.done)) {
         errors.push(`Phase ${phase.id}: done must be a finite number`);
       }
+      // Cross-validate done against actual accepted tasks
+      if (Number.isFinite(phase.done) && Array.isArray(phase.todo)) {
+        const acceptedCount = phase.todo.filter(t => t.lifecycle === 'accepted').length;
+        if (phase.done !== acceptedCount) {
+          errors.push(`Phase ${phase.id}: done (${phase.done}) does not match accepted task count (${acceptedCount})`);
+        }
+      }
       if (!Array.isArray(phase.todo)) {
         errors.push(`Phase ${phase.id}: todo must be an array`);
         continue;

package/src/tools/orchestrator.js

@@ -4,7 +4,6 @@ import {
   read,
   storeResearch,
   update,
-  addEvidence,
   selectRunnableTask,
   buildExecutorContext,
   matchDecisionForBlocker,
@@ -278,6 +277,19 @@ function buildDecisionEntries(decisions, phaseId, taskId, existingCount = 0) {
     .filter(Boolean);
 }
 
+function buildErrorFingerprint(result) {
+  const parts = [];
+  if (result.blockers?.length > 0) {
+    const b = result.blockers[0];
+    parts.push(typeof b === 'string' ? b : (b.reason || b.type || ''));
+  }
+  if (result.files_changed?.length > 0) {
+    parts.push([...result.files_changed].sort().join(','));
+  }
+  const combined = parts.filter(Boolean).join('|');
+  return combined.length > 0 ? combined.slice(0, 120) : result.summary.slice(0, 80);
+}
+
 function getBlockedReasonFromResult(result) {
   const firstBlocker = (result.blockers || [])[0];
   if (!firstBlocker) return { blocked_reason: result.summary, unblock_condition: null };
@@ -705,15 +717,44 @@ export async function resumeWorkflow({ basePath = process.cwd(), _depth = 0 } =
         message: 'Project is paused. Confirm to resume execution.',
       };
     case 'planning':
-    case 'reconcile_workspace':
-    case 'replan_required':
-    case 'research_refresh_needed':
       return {
         success: true,
         action: 'await_manual_intervention',
         workflow_mode: state.workflow_mode,
-        message: `workflow_mode "${state.workflow_mode}" is recognized but not yet automated by the orchestrator`,
+        guidance: 'Complete planning and call state-init to initialize the project',
+        message: 'Project is in planning mode; complete the plan and initialize with state-init',
+      };
+    case 'reconcile_workspace': {
+      const reconGitHead = await getGitHead(basePath);
+      return {
+        success: true,
+        action: 'reconcile_workspace',
+        workflow_mode: state.workflow_mode,
+        expected_head: state.git_head,
+        actual_head: reconGitHead,
+        guidance: 'Workspace git HEAD has diverged. Verify changes and update git_head via state-update, then set workflow_mode to executing_task',
+        message: `Git HEAD mismatch: saved=${state.git_head}, current=${reconGitHead}`,
+      };
+    }
+    case 'replan_required':
+      return {
+        success: true,
+        action: 'replan_required',
+        workflow_mode: state.workflow_mode,
+        guidance: 'Plan files modified since last session. Review changes, update the plan if needed, then set workflow_mode to executing_task via state-update',
+        message: 'Plan artifacts modified since last session; review and re-align before resuming',
       };
+    case 'research_refresh_needed': {
+      const expiredResearch = collectExpiredResearch(state);
+      return {
+        success: true,
+        action: 'dispatch_researcher',
+        workflow_mode: state.workflow_mode,
+        expired_research: expiredResearch,
+        guidance: 'Research cache expired. Dispatch researcher sub-agent to refresh, then call orchestrator-handle-researcher-result',
+        message: 'Research has expired and must be refreshed before execution can resume',
+      };
+    }
     default:
       return {
         error: true,
@@ -746,55 +787,47 @@ export async function handleExecutorResult({ result, basePath = process.cwd() }
   if (result.outcome === 'checkpointed') {
     const reviewLevel = reclassifyReviewLevel(task, result);
     const isL0 = reviewLevel === 'L0';
+    const autoAccept = isL0 || task.review_required === false;
 
     const current_review = !isL0 && reviewLevel === 'L2' && task.review_required !== false
       ? { scope: 'task', scope_id: task.id, stage: 'spec' }
       : null;
     const workflow_mode = current_review ? 'reviewing_task' : 'executing_task';
 
-    // First persist: checkpoint the task (running → checkpointed)
+    // Single atomic persist: auto-accept goes directly running → accepted,
+    // otherwise running → checkpointed (awaiting review)
+    const taskPatch = {
+      id: task.id,
+      lifecycle: autoAccept ? 'accepted' : 'checkpointed',
+      checkpoint_commit: result.checkpoint_commit,
+      files_changed: result.files_changed || [],
+      evidence_refs: result.evidence || [],
+      level: reviewLevel,
+      blocked_reason: null,
+      unblock_condition: null,
+      debug_context: null,
+    };
+    const phasePatch = { id: phase.id, todo: [taskPatch] };
+    // done is auto-recomputed by update() — no manual increment needed
+
+    // Bundle evidence into the same atomic persist to prevent inconsistency
+    const evidenceUpdates = {};
+    for (const ev of (result.evidence || [])) {
+      if (ev && typeof ev === 'object' && typeof ev.id === 'string' && typeof ev.scope === 'string') {
+        evidenceUpdates[ev.id] = ev;
+      }
+    }
+
     const persistError = await persist(basePath, {
       workflow_mode,
       current_task: null,
       current_review,
       decisions,
-      phases: [{
-        id: phase.id,
-        todo: [{
-          id: task.id,
-          lifecycle: 'checkpointed',
-          checkpoint_commit: result.checkpoint_commit,
-          files_changed: result.files_changed || [],
-          evidence_refs: result.evidence || [],
-          level: reviewLevel,
-          blocked_reason: null,
-          unblock_condition: null,
-          debug_context: null,
-        }],
-      }],
+      phases: [phasePatch],
+      ...(Object.keys(evidenceUpdates).length > 0 ? { evidence: evidenceUpdates } : {}),
     });
     if (persistError) return persistError;
 
-    // Store structured evidence entries
-    for (const ev of (result.evidence || [])) {
-      if (ev && typeof ev === 'object' && typeof ev.id === 'string' && typeof ev.scope === 'string') {
-        await addEvidence({ id: ev.id, data: ev, basePath });
-      }
-    }
-
-    // Auto-accept: L0 tasks or tasks with review_required: false
-    const autoAccept = isL0 || task.review_required === false;
-    if (autoAccept) {
-      const acceptError = await persist(basePath, {
-        phases: [{
-          id: phase.id,
-          done: (phase.done || 0) + 1,
-          todo: [{ id: task.id, lifecycle: 'accepted' }],
-        }],
-      });
-      if (acceptError) return acceptError;
-    }
-
     return {
       success: true,
       action: current_review ? 'dispatch_reviewer' : 'continue_execution',
@@ -841,7 +874,7 @@ export async function handleExecutorResult({ result, basePath = process.cwd() }
   const retry_count = (task.retry_count || 0) + 1;
   const error_fingerprint = typeof result.error_fingerprint === 'string' && result.error_fingerprint.length > 0
     ? result.error_fingerprint
-    : result.summary.slice(0, 80);
+    : buildErrorFingerprint(result);
   const shouldDebug = retry_count >= MAX_DEBUG_RETRY;
   const current_review = shouldDebug
     ? {
@@ -989,8 +1022,6 @@ export async function handleReviewerResult({ result, basePath = process.cwd() }
   }
 
   const taskPatches = [];
-  let doneIncrement = 0;
-  let doneDecrement = 0;
 
   // Accept tasks
   for (const taskId of (result.accepted_tasks || [])) {
@@ -998,7 +1029,6 @@ export async function handleReviewerResult({ result, basePath = process.cwd() }
     if (!task) continue;
     if (task.lifecycle === 'checkpointed') {
       taskPatches.push({ id: taskId, lifecycle: 'accepted' });
-      doneIncrement += 1;
     }
   }
 
@@ -1007,19 +1037,10 @@ export async function handleReviewerResult({ result, basePath = process.cwd() }
     const task = getTaskById(phase, taskId);
     if (!task) continue;
     if (task.lifecycle === 'checkpointed' || task.lifecycle === 'accepted') {
-      if (task.lifecycle === 'accepted') doneDecrement += 1;
       taskPatches.push({ id: taskId, lifecycle: 'needs_revalidation', evidence_refs: [] });
     }
   }
 
-  // Snapshot accepted task IDs before propagation (for done counter adjustment).
-  // Note: rework_tasks patches above are NOT yet applied in-memory, so tasks demoted
-  // by the rework loop are still 'accepted' here. The guard below
-  // `!taskPatches.some(p => p.id === task.id)` prevents double-counting.
-  const acceptedBeforePropagation = new Set(
-    (phase.todo || []).filter(t => t.lifecycle === 'accepted').map(t => t.id),
-  );
-
   // Propagation for critical issues with invalidates_downstream
   for (const issue of (result.critical_issues || [])) {
     if (issue.invalidates_downstream && issue.task_id) {
@@ -1031,18 +1052,15 @@ export async function handleReviewerResult({ result, basePath = process.cwd() }
   for (const task of (phase.todo || [])) {
     if (task.lifecycle === 'needs_revalidation' && !taskPatches.some((p) => p.id === task.id)) {
       taskPatches.push({ id: task.id, lifecycle: 'needs_revalidation', evidence_refs: [] });
-      if (acceptedBeforePropagation.has(task.id)) {
-        doneDecrement += 1;
-      }
     }
   }
 
   const hasCritical = (result.critical_issues || []).length > 0;
   const reviewStatus = hasCritical ? 'rework_required' : 'accepted';
 
+  // done is auto-recomputed by update() — no manual tracking needed
   const phaseUpdates = {
     id: phase.id,
-    done: Math.max(0, (phase.done || 0) + doneIncrement - doneDecrement),
     phase_review: {
       status: reviewStatus,
       ...(hasCritical
@@ -1063,20 +1081,22 @@ export async function handleReviewerResult({ result, basePath = process.cwd() }
 
   const workflowMode = 'executing_task';
 
+  // Bundle evidence into the same atomic persist
+  const evidenceUpdates = {};
+  for (const ev of (result.evidence || [])) {
+    if (ev && typeof ev === 'object' && typeof ev.id === 'string' && typeof ev.scope === 'string') {
+      evidenceUpdates[ev.id] = ev;
+    }
+  }
+
   const persistError = await persist(basePath, {
     workflow_mode: workflowMode,
     current_review: null,
     phases: [phaseUpdates],
+    ...(Object.keys(evidenceUpdates).length > 0 ? { evidence: evidenceUpdates } : {}),
   });
   if (persistError) return persistError;
 
-  // Store evidence entries if provided
-  for (const ev of (result.evidence || [])) {
-    if (ev && typeof ev === 'object' && typeof ev.id === 'string' && typeof ev.scope === 'string') {
-      await addEvidence({ id: ev.id, data: ev, basePath });
-    }
-  }
-
   return {
     success: true,
     action: hasCritical ? 'rework_required' : 'review_accepted',

package/src/tools/state.js

@@ -42,6 +42,16 @@ export function setLockPath(lockPath) {
   _fileLockPath = lockPath;
 }
 
+/**
+ * Ensure _fileLockPath is set from a known state path.
+ * Must be called before withStateLock in all mutation paths.
+ */
+function ensureLockPathFromStatePath(statePath) {
+  if (!_fileLockPath && statePath) {
+    _fileLockPath = join(dirname(statePath), 'state.lock');
+  }
+}
+
 function withStateLock(fn) {
   const p = _mutationQueue.then(() => {
     if (_fileLockPath) {
@@ -80,6 +90,7 @@ export async function init({ project, phases, research, force = false, basePath
   }
   const gsdDir = join(basePath, '.gsd');
   const statePath = join(gsdDir, 'state.json');
+  ensureLockPathFromStatePath(statePath);
 
   return withStateLock(async () => {
     // Guard: reject re-initialization unless force is set
@@ -129,6 +140,9 @@ export async function init({ project, phases, research, force = false, basePath
       ...state.phases.map((phase) => join(phasesDir, `phase-${phase.id}.md`)),
     ];
     const mtimes = await Promise.all(trackedFiles.map(async (filePath) => (await stat(filePath)).mtimeMs));
+    // Math.ceil is required: mtimeMs has sub-millisecond precision (float), but
+    // Date.toISOString() truncates to milliseconds. Without ceil, the stored timestamp
+    // can be slightly less than the file's actual mtime, causing false plan-drift detection.
     state.context.last_session = new Date(Math.ceil(Math.max(...mtimes))).toISOString();
     await writeJson(statePath, state);
 
@@ -195,8 +209,7 @@ export async function update({ updates, basePath = process.cwd() } = {}) {
   if (!statePath) {
     return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: 'No .gsd directory found' };
   }
-  // C-2: Initialize cross-process lock path on first mutation
-  if (!_fileLockPath) _fileLockPath = join(dirname(statePath), 'state.lock');
+  ensureLockPathFromStatePath(statePath);
 
   return withStateLock(async () => {
     const result = await readJson(statePath);
@@ -242,6 +255,12 @@ export async function update({ updates, basePath = process.cwd() } = {}) {
 
     // Deep merge phases by ID instead of shallow replace [I-1]
     const merged = { ...state, ...updates };
+
+    // Deep merge evidence by key (preserves existing entries)
+    if (updates.evidence && isPlainObject(updates.evidence)) {
+      merged.evidence = { ...(state.evidence || {}), ...updates.evidence };
+    }
+
     if (updates.phases && Array.isArray(updates.phases)) {
       merged.phases = state.phases.map(oldPhase => {
         const newPhase = updates.phases.find(p => p.id === oldPhase.id);
@@ -276,6 +295,21 @@ export async function update({ updates, basePath = process.cwd() } = {}) {
       }
     }
 
+    // Recompute `done` from actual accepted tasks (prevents counter drift)
+    if (updates.phases && Array.isArray(updates.phases)) {
+      for (const phase of merged.phases) {
+        if (Array.isArray(phase.todo)) {
+          phase.done = phase.todo.filter(t => t.lifecycle === 'accepted').length;
+        }
+      }
+    }
+
+    // Auto-prune evidence when entries exceed limit
+    if (merged.evidence && Object.keys(merged.evidence).length > MAX_EVIDENCE_ENTRIES) {
+      const gsdDir = dirname(statePath);
+      await _pruneEvidenceFromState(merged, merged.current_phase, gsdDir);
+    }
+
     // Use incremental validation for simple updates (no phases changes)
     const validation = !updates.phases
       ? validateStateUpdate(state, updates)
@@ -336,11 +370,12 @@ export async function phaseComplete({
   if (!statePath) {
     return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: 'No .gsd directory found' };
   }
+  ensureLockPathFromStatePath(statePath);
 
   return withStateLock(async () => {
     const result = await readJson(statePath);
     if (!result.ok) {
-      return { error: true, message: result.error };
+      return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: result.error };
     }
     const state = result.data;
 
@@ -484,11 +519,12 @@ export async function addEvidence({ id, data, basePath = process.cwd() }) {
   if (!statePath) {
     return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: 'No .gsd directory found' };
   }
+  ensureLockPathFromStatePath(statePath);
 
   return withStateLock(async () => {
     const result = await readJson(statePath);
     if (!result.ok) {
-      return { error: true, message: result.error };
+      return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: result.error };
     }
     const state = result.data;
 
@@ -563,11 +599,12 @@ export async function pruneEvidence({ currentPhase, basePath = process.cwd() })
   if (!statePath) {
     return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: 'No .gsd directory found' };
   }
+  ensureLockPathFromStatePath(statePath);
 
   return withStateLock(async () => {
     const result = await readJson(statePath);
     if (!result.ok) {
-      return { error: true, message: result.error };
+      return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: result.error };
     }
     const state = result.data;
 
@@ -831,16 +868,27 @@ export function reclassifyReviewLevel(task, executorResult) {
 const MIN_TOKEN_LENGTH = 2;
 const MIN_OVERLAP = 2;
 
+// High-frequency words too generic for meaningful keyword matching
+const STOPWORDS = new Set([
+  'the', 'and', 'for', 'with', 'this', 'that', 'from', 'have', 'not',
+  'but', 'are', 'was', 'been', 'will', 'can', 'should', 'would', 'could',
+  'use', 'using', 'need', 'needs', 'into', 'also', 'when', 'then',
+  'than', 'more', 'some', 'does', 'did', 'its', 'has', 'all', 'any',
+  'error', 'data', 'type', 'value', 'file', 'code', 'function',
+  'return', 'null', 'true', 'false', 'undefined', 'object', 'string',
+  'number', 'array', 'list', 'map', 'set', 'key', 'name',
+]);
+
 /**
  * Tokenize a string into lowercase tokens, splitting on whitespace and punctuation.
- * Filters out short tokens (< MIN_TOKEN_LENGTH).
+ * Filters out short tokens (< MIN_TOKEN_LENGTH) and stopwords.
  */
 function tokenize(text) {
   if (!text) return [];
   return text
     .toLowerCase()
     .split(/[\s,.:;!?()[\]{}<>/\\|@#$%^&*+=~`'"，。：；！？（）【】、]+/)
-    .filter(t => t.length >= MIN_TOKEN_LENGTH);
+    .filter(t => t.length >= MIN_TOKEN_LENGTH && !STOPWORDS.has(t));
 }
 
 /**
@@ -949,11 +997,7 @@ export async function storeResearch({ result, artifacts, decision_index, basePat
     return { error: true, code: ERROR_CODES.VALIDATION_FAILED, message: `Invalid researcher result: ${resultValidation.errors.join('; ')}` };
   }
 
-  const artifactsValidation = validateResearchArtifacts(artifacts, {
-    decisionIds: result.decision_ids,
-    volatility: result.volatility,
-    expiresAt: result.expires_at,
-  });
+  const artifactsValidation = validateResearchArtifacts(artifacts);
   if (!artifactsValidation.valid) {
     return { error: true, code: ERROR_CODES.VALIDATION_FAILED, message: `Invalid research artifacts: ${artifactsValidation.errors.join('; ')}` };
   }
@@ -967,11 +1011,12 @@ export async function storeResearch({ result, artifacts, decision_index, basePat
   if (!statePath) {
     return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: 'No .gsd directory found' };
   }
+  ensureLockPathFromStatePath(statePath);
 
   return withStateLock(async () => {
     const current = await readJson(statePath);
     if (!current.ok) {
-      return { error: true, message: current.error };
+      return { error: true, code: ERROR_CODES.NO_PROJECT_DIR, message: current.error };
     }
 
     const state = current.data;
@@ -1026,6 +1071,13 @@ export async function storeResearch({ result, artifacts, decision_index, basePat
       state.workflow_mode = inferWorkflowModeAfterResearch(state);
     }
 
+    // Recompute done after applyResearchRefresh may have invalidated tasks
+    for (const phase of (state.phases || [])) {
+      if (Array.isArray(phase.todo)) {
+        phase.done = phase.todo.filter(t => t.lifecycle === 'accepted').length;
+      }
+    }
+
     const validation = validateState(state);
     if (!validation.valid) {
       return { error: true, code: ERROR_CODES.VALIDATION_FAILED, message: `State validation failed: ${validation.errors.join('; ')}` };

package/workflows/execution-flow.md

@@ -0,0 +1,147 @@
+# 共享执行流程 (STEP 5-12)
+
+> 由 start.md 和 prd.md 共享引用。修改此文件即同步两个入口。
+
+## STEP 5 — 智能研究判断
+
+判断是否需要研究:
+```
+├── 新项目                    → 必须研究
+├── 涉及新技术栈              → 必须研究
+├── 简单 bug 修复 / 小功能    → 跳过研究
+├── 已有 .gsd/research/ 且未过期 → 跳过研究
+├── 用户明确要求               → 研究
+└── 已有研究但需求方向变了     → 增量研究 (只研究新方向)
+```
+
+需要研究时:
+1. 派发 `researcher` 子代理 (新鲜上下文)
+2. 研究输出写入 `.gsd/research/` (STACK.md, ARCHITECTURE.md, PITFALLS.md, SUMMARY.md)
+3. 向用户展示关键发现: 技术栈推荐 + 陷阱警告 + ⭐ 推荐方案
+
+不需要时: 跳过，直接进入 STEP 6。
+
+## STEP 6 — 深度思考
+
+如有 `sequential-thinking` MCP 可用 → 调用深入思考:
+- 输入: 需求摘要 + 代码库分析 + 研究结果 (如有)
+- 目的: 在生成计划前进行系统性架构思考
+
+如无 `sequential-thinking` MCP → 降级为内联思考，继续。
+
+## STEP 7 — 生成分阶段计划
+
+生成 plan.md + phases/*.md:
+- **phase** 负责管理与验收，**task** 负责执行
+- 每阶段控制在 **5-8 个 task** (便于 phase-level 收口)
+- 每个 task = 原子化 todo (含文件、操作、验证条件)
+- 每个 task 补充元数据:
+  - `requires` — 依赖列表 (含 gate 类型)
+  - `review_required` — 是否需要审查
+  - `research_basis` — 引用的 research decision id
+- 审查级别按影响面判定:
+  - **L0** — 无运行时语义变化 (docs/config/style)
+  - **L1** — 普通编码任务 (默认)
+  - **L2** — 高风险 (auth/payment/public API/DB migration/核心架构)
+- 标注可并行任务组 `[PARALLEL]` (当前仅作未来升级标记)
+
+## STEP 8 — 计划自审
+
+轻量自审 (编排器自身执行，不派发子代理):
+
+### 基础审查 (所有项目)
+- [ ] 是否有遗漏的需求点？
+- [ ] 阶段划分是否合理？(phase 过大则拆分)
+- [ ] 任务依赖关系是否正确？
+- [ ] 验证条件是否可执行？
+
+### 增强审查 (高风险项目)
+
+触发条件: 项目涉及 auth / payment / security / public API / DB migration / 核心架构变更
+
+维度:
+1. **需求覆盖:** 原始需求的每个要点是否都映射到了至少一个 task？
+2. **风险排序:** 高风险 task 是否排在前面？(fail-fast 原则)
+3. **依赖安全:** L2 task 的下游是否都用了 `gate:accepted`？
+4. **验证充分:** 涉及 auth/payment 的 task 是否都有明确的安全验证条件？
+5. **陷阱规避:** `research/PITFALLS.md` 中的每个陷阱是否都有对应的防御 task 或验证条件？
+
+输出: `pass` / `revise` (附具体修正建议)
+轮次: 最多 2 轮自审修正；2 轮后仍有问题 → 标注风险展示给用户
+
+→ 自审修正后再展示给用户。
+
+<HARD-GATE id="plan-confirmation">
+## STEP 9 — 用户确认计划
+
+展示计划给用户，等待确认:
+- 用户指出问题 → 调整计划 → 重新展示
+- 用户确认 → 继续
+
+⛔ 不得在用户确认前执行 STEP 10-12。未确认 = 不写文件、不执行代码。
+</HARD-GATE>
+
+<HARD-GATE id="docs-written">
+## STEP 10 — 生成文档
+
+1. 调用 `state-init` MCP 工具初始化项目:
+   ```
+   state-init({
+     project: "<项目名>",
+     phases: [
+       {
+         name: "<阶段名>",
+         tasks: [
+           { name: "<任务名>", level: "L1", review_required: true, requires: [] },
+           ...
+         ]
+       },
+       ...
+     ]
+   })
+   ```
+   ⚠️ 必须使用 `state-init` MCP 工具，禁止手写 state.json — 工具自动生成 id/lifecycle/phase_review/phase_handoff，内置 schema 校验和循环依赖检测。
+2. 写入 `plan.md` — 项目总览索引 (不含 task 级细节)
+3. 写入 `phases/*.md` — 每阶段详细 task 规格 (source of truth)
+4. 如有研究: 确认 `.gsd/research/` 已写入
+
+规则:
+- `plan.md` 是只读索引: 生成后不再修改 (除非 replan)
+- `phases/*.md` 是 task 规格的唯一 source of truth
+- `plan.md` 不包含 task 级细节，避免与 `phases/*.md` 重复
+
+□ state-init 调用成功 (返回 success: true)
+□ plan.md 已写入
+□ phases/*.md 已写入 (每个 phase 一个文件)
+□ 所有 task 都有 lifecycle / level / requires / review_required
+→ 全部满足才可继续
+</HARD-GATE>
+
+## STEP 11 — 自动执行主路径
+
+进入执行主循环。phase = 管理边界，task = 执行边界。
+
+<execution_loop>
+参考 `references/execution-loop.md` 获取完整 9 步执行循环规范 (11.1-11.9) 及依赖门槛语义。
+
+编排器必须严格按照该参考文档中的步骤顺序执行:
+加载 phase → 选择 task → 构建上下文 → 派发 executor → 处理结果 → 审查 → phase handoff → 批量更新 → 上下文检查
+
+**自动执行循环:** 进入执行后，持续循环直到遇到终止条件:
+1. 调用 `orchestrator-resume` 获取 action
+2. 按 action 派发对应子代理 (executor/reviewer/researcher/debugger)
+3. 收到结果后调用对应 `orchestrator-handle-*-result`
+4. 回到步骤 1
+5. 终止: action ∈ {idle, awaiting_user, completed, failed, await_manual_intervention}
+
+不要在循环中间停下来等用户确认 — 让编排器驱动。`complete_phase` action → 调 `phase-complete` MCP tool → 自动推进下一 phase。
+</execution_loop>
+
+## STEP 12 — 最终报告
+
+全部 phase 完成后，输出最终报告:
+- 项目总结
+- 各阶段完成情况
+- 关键 decisions 汇总
+- 验证 evidence 汇总
+- 遗留问题 / 后续建议 (如有)