@tea-agent/loop-agent 0.39.0-next.2 → 0.39.0-next.20

package/CHANGELOG.md

@@ -2,17 +2,88 @@
 
 ## [Unreleased]
 
+### 变更
+
+- npm `next` 渠道发布改为提交数节流：main push 仅当距上次 next 发布的 source commit ≥ 5 个 commit 时才自动发布（新增 `scripts/next-publish-gate.mjs`，读已发布 tarball build-stamp 对比 HEAD）；不足阈值时 workflow 绿色跳过，手动 `workflow_dispatch` 不受限（切 release 线前强制出包 / 紧急补发）。发布文档真源 §4.1/§5.1/§7/§12 同步。
+
 ### 新增
 
+- backend-test 适配 A–D：可配置 `backendTest` 布局（`testRoot`/`markdownRel`/`scriptRel`）、用户公共前置文档引用与模块顶部 `user_shared_setup`（不写 conftest）、查询轴 Scenario Partitions 确定性展开与 fail-closed 门禁、`mode=gap-fill` 13 节点增量 DAG。Slot ID 以声明的 Partition ID 为真源（`TP-<Partition ID>-...`）。
+- 前端 plan 唯一机器契约化（JSON-only）：`frontend-plan-pi` / `frontend-plan-revision-pi` 输出收敛为「一行导语 + 唯一 fenced `json` 契约 + 唯一 `openspec-citations` 引用块」，彻底移除「可选 Markdown 说明」；plan.md（`contracts/frontend-plan.md`）改由 prewrite gate 用纯函数 `renderFrontendPlanMarkdown` 确定性渲染（不新增 DAG 节点/不新增 Pi 调用）
+- 契约 schema 新增 5 个可选字段（`implementationSteps` / `stylingStrategy` / `dependencyPolicy` / `residualRisks` / `realIntegrationGap`），`schemaVersion` 保持 1；`docs/templates/frontend-implementation-contract.schema.json` 同步
+- revision 输出改为 RFC 7386 merge-patch delta + 引用块：gate 配置新增 `revisionPatch`，生效时在 `planFallbackFromNodeIds[0]` 原始契约上应用 `applyFrontendImplementationContractPatch` 后进入既有分析/规范化/物化管线；旧 DAG 缺字段按整契约处理
+- canonical contract 序列化（UTF-8）超过 64KB → `retryable-invalid` + 新增 `contract-too-large` failureCode（体积守卫兜底）
+- `design-review` / `final-design-review` 提示改为消费契约 JSON（注入契约字段摘要），不再要求读取 plan prose
+- Console「模型与认证」页新增模型矩阵编辑器：展示并允许编辑 `harness.json executors.pi` 的 LOW/MED/HIGH 三档模型；输入框从 readiness 结构化投影 `piModelMatrix` 预填（provider/model 或裸 model id），本地内联校验（空串 / `default` 哨兵 / 畸形 `provider/`、`/model`、`a//b`）非法时禁用保存；保存经既有 `POST /api/operator/v1/operations` 派发 `savePiModelMatrix`，复用 HTTP 层 `evaluateMutationGate`（boot capability cookie + `x-loop-console-confirmation` 头 + Origin/Host/Sec-Fetch-Site），读全文档后原子写 tmp+fsync+rename（保留 `$schema` 与其余字段），成功后自动重测 readiness 并对活动 Chat session 触发 `/sessions/:id/reload`（409 `PI_SESSION_BUSY` 给出可恢复提示）
+- `GET /api/operator/v1/readiness` 的 probe 新增 `piModelMatrix` 结构化投影（向后兼容，不改动既有 `modelsMessage`）
+- 组件选型规范合规（ADR 0016）：契约 schema 新增可选 `uiComponentChoices[]`（`{ purpose, component, decision(specified|reuse-existing|new), specReference{path,section,line}|null, rationale }`），`specified` 必须给非空 `specReference`、`new` 不得携带、空串拒绝；`schemaVersion` 保持 1；`docs/templates/frontend-implementation-contract.schema.json` 同步
+- `frontend-prewrite-gate` 新增三个 `retryable-invalid` failureCode：`component-spec-reference-invalid`（specReference.path 非受支持规范路径）、`component-spec-not-cited`（未在 openspec-citations 引用块引用或未真实读取）、`component-choices-missing`（组件/主题候选桶非空且契约有 UI 可见工作而 `uiComponentChoices` 缺失/空）
+- gate 配置新增可选 `componentSpecCandidatePaths`（旧 DAG 缺字段兼容）；`init-hybrid` 从 `designEvidence.classified.component ∪ theme ∪ rule.components` 冻结注入，capability context 强调组件/主题桶语义，plan/revision/design-review/final-design-review 提示加入硬规则与 BAD/GOOD 示例
+- `frontend-plan.md` 新增 `Component Selection (spec conformance)` 章节（纯函数确定性渲染，缺省 `_(not specified)_`）
+
 - `frontend-prewrite-gate` 的 OpenSpec 读取校验支持 `task.json.frontendOpenspec` 配置：`policy` 缺省 `cited`（候选 = `requiredReadPaths` ∪ 任务源显式引用的 openspec 路径），`scan-strict` 保留止血任务后的全量必读语义；`requiredReadPaths` 中非 openspec 支持路径在生成期确定性失败
 - 确定性引用块协议：生效 plan 输出在 fenced `json` 契约块后追加 `openspec-citations` fenced 块（每行一个 JSON `{"path","section","line"}`）；gate 解析并与生效 plan/review 节点的成功 read 事件核验，新增 `openspec-citation-block-unparseable` / `openspec-not-cited` / `openspec-citation-not-read` 三个 retryable-invalid failureCode（防捏造、防遗漏）
 - `src/shared/openspec-spec.ts` 新增 `extractTaskSourceOpenspecPaths` 纯函数（反引号内联、markdown 链接目标、裸路径 token → 归一化 → spec 过滤 → 治理子树排除 → 去重排序）
 - 生成期把 `openspecPolicy` 与候选来源元数据（`openspecCandidateSources.declared` / `taskSourceCited` / `scanStrict`）冻结进 prewrite gate 配置；`task advance` 的 advisories 说明当前模式与候选来源
 
+### 改进
+
+- `frontend-plan-pi` 改为 runtime skeleton + editable RFC 7386 patch：`schemaVersion`、source binding、risk、writer target files 与生产 Mock 安全位由 DAG 冻结且不可被 plan/revision 覆盖；节点 validator 将补丁编译为 canonical 完整契约后再交给 review/prewrite，并在 `contracts/candidates/<nodeId>/` 保存每次 raw/extracted/normalized/validation 证据。任务提示与三类结构化重试（invalid-output / provider length 截断 / output-too-large）同步收敛为单一 patch + 唯一 `openspec-citations`；无法在声明文件中解析的可选 verification symbol 会被确定性移除并留下 normalization action，避免格式冲突、完整契约重复输出和说明性 symbol 延迟到 trace 才失败。
+- Operator Chat 恢复工具对齐真实能力：系统提示改用 `dagReport` / `dagRerunPlan` / `dagRerun` / `standaloneTaskRerun` / `workerTaskRetry` 真名，不再写 `operator_*` 或把 `dagResume` 当成模型可执行项；`dagDoctor` / `dagStatus` / `workerPoolDoctor` 改为 typed `runId`/`taskId`，不再让模型猜 CLI `args[]`。ADR 0013 自定义 base prompt 上限由 1900 统一调整为 2500（2026-08-18，废止 0.8×2498 比例），预算用于写清节点续跑 / 整单重生 / Worker 重排队分流，以及多轮迭代恢复纪律（先一句话总结上一轮错误、修根因不修症状、禁止原样重交、同错两轮后停下报告）。
+- Operator Chat 模型工具面收敛：`advanced` 分类的命令（workflow/loop/benchmark/knowledge-graph 等 8 个只读低频入口）不再进入模型工具白名单，模型注意力集中到高频工具；保留的 27 个 model-callable 官方 CLI 动作全部换上「做什么 + 何时用」语义描述，不再是无信息的 `Official CLI action coverage for ...` 模板；dispatcher 对官方尾部动作的识别从描述前缀匹配改为结构化 `OFFICIAL_TAIL_ACTION_IDS` 集合判断。审查收口：`knowledgeCurate` 纠正为 `mutation`/`advanced`（条件写盘，移出模型面，防经 read 通道传 `--output` 写盘）；`promoteRun`/`closeoutTask` 描述去掉 dispatch 不保证的生命周期断言，如实标注为 task-advance 别名并提示先核 run/任务状态。工具清单渲染修复：promptSnippet 去掉重复工具名前缀（SDK 渲染自带 `- name:`，原实现名字出现两遍）；`inspect`/`status`/`spineCheck`/`sourceInstructions` 四条占位描述换成语义描述；task-advance 家族（`newTask`/`dagRunTask`）标注与标准流 `taskAdvance` 的关系，避免模型在 5 个相似入口间选错。可调用性修复：`observeLink` 补上 4 个可选参数 schema（dagRunId/taskId/featureId/workerRunId，原空 schema + additionalProperties:false 导致模型传参必被拒、工具名存实亡）；Pi 工具参数 schema 现在携带注册表里的参数描述（Type.String/Boolean/Number/Array 注入 `description`，原实现丢弃参数描述、模型只能看到参数名）。
+
+- Operator Chat 对齐 pi-web：建会话不再传封闭 `tools` 白名单（避免 SDK 把扩展 MCP 工具滤出 registry），创建后 `bindExtensions` 发 `session_start`，再钉激活集。扩展注册的 MCP 工具可被模型调用；`apply_patch` / `full-tools` / `shell` / `coding-chat` 仍排除。
+- 对话侧栏 Plugins 右侧新增只读 MCP 弹窗：展示 `mcp.json` 配置的服务，以及当前会话真正注册/激活的 MCP 工具；没有配置也没有工具时保持真实空态。
+- 对话侧栏原 Packages 入口改名为 Plugins，与 pi-web / Pi 用户面一致；内部仍是 package inventory。
+- 对话进行中底部上下文占用百分比保持常驻：token 用量事件不再冲掉上一帧的 contextUsage，并在回合中持续刷新。
+- 上一轮大量 bash 结束后再发新消息，旧工具条不再整批挂到新气泡下面：无主工具不再丢进最新一组，历史按完整回合裁切，空助手气泡若还挂着工具会留下来当锚点。
+- 侧栏文件树点击目录行直接插入 `@目录/` 引用（箭头单独负责展开/收起）；输入框 @ 补全也列出目录，排在文件前面。
+- 对话进行中点「停止」立即安静下来（乐观停止，对齐 pi-web）：不再等 POST /abort 往返；后端 abort 等待从 8s 降到 2.5s，未终结时由状态对账收敛。
+- 侧栏文件树实时刷新：agent 回合结束、浏览器标签页回到前台、手动点「刷新」时重拉根目录+全部已展开目录（此前手动刷新只拉根目录，子目录永远旧缓存）。
+- Skills / Packages / Models 面板不再把 Pi 的 warning / collision 原文堆在列表上方：改成标题条一颗摘要芯片（如「14 条提示」），点开按「需要处理 / 名称不规范 / 同名未生效」分组，文案用自然语言，路径只显示末两段；名称不规范的技能在列表旁标「提示」，详情里说明「现在这样也能用」。加载失败仍单独成组，可点「再读一次」。
+- Skills / Packages 双栏改为左右各自滚动；资源弹窗挂到 `document.body`，按顶部对话导航的实时高度在剩余视口居中，不再被 sticky 顶栏挡住上沿。
+
+### 修复
+
+- Console 对话中的 Markdown 表格只约束横向滚动，不再截断纵向滚动链；鼠标或触控板停在表格上时仍可继续上下滚动对话。
+- scenario-param 确定性 rewrite 的 Python 语法校验：Windows 优先 `python`，并把 Microsoft Store `python3` 空壳（非 0 且无语法诊断）视为解释器不可用后回退，避免误回滚合法 rewrite。
+- `frontend-plan-revision-pi` 节点级校验与 merge-patch 输出协议矛盾修复：新增 schemaId `frontend-implementation-contract-revision-patch-v1` 与格式级宽松校验 `validateFrontendRevisionPatchNodeOutput`（至少一个 `openspec-citations` 块 + 剥离引用块后恰好一个 JSON 对象，不做完整契约 schema 校验），revision 节点 `structuredContractOutput.schemaId` 切换（plan-pi 仍为 v1）；node-execution 的 invalid-output retry 指令按 schemaId 分支，patch 协议提示 merge-patch delta + 引用块，消除「prompt 要求 patch、retry 要求完整契约」的 retryable 死循环；gate 仍是唯一权威合并/校验点（零改动）。ADR 0015 补「节点级校验」决策 7
+- `frontend-implementation-contract.schema.json` 的 `uiComponentChoices` items 补 `if/then` 决策语义（`specified` ⇒ `specReference` 必填非 null；`new` ⇒ `specReference` 为 null/缺失）；`frontend-plan-render.ts` 空 `section` 不再渲染尾随 `#`；ADR 0016 增补「门禁边界」小节
+- 前端 gate 内容驱动 revisionPatch 兼容 + verdict 归一化统一：旧 DAG（缺 `revisionPatch` 字段）但 revision 节点按协议输出 merge-patch 时，gate 按内容判定自动合并原始契约并走既有校验（不再因 `missing targets.files` 进入 retryable 死循环）；patch 形态但无法合并（fallback 缺失/自身、原始节点不可读或非完整契约、apply 失败）→ `blocked` + 新增 `revision-patch-unmergeable`（错误信息含重新生成 DAG 指引）；`normalizeVerdictCandidateLine` 增强 backtick 强调剥离，gate 的 `firstNonEmptyVerdictLine` 复用共享 normalize，runIf / output-protocol / gate 三处一致
+
 ### 文档
 
+- 新增 `docs/decisions/0016-frontend-component-conformance.md`，并同步 `docs/runtime/frontend-implementation-workflow.md`、`docs/templates/frontend-implementation-contract.schema.json`、`skills/frontend-implementation/references/design-spec.md`、`skills/frontend-implementation/references/node-contracts.md`、`skills/frontend-design-review/references/review-checklist.md`、`CONTEXT.md`
+- 新增 `docs/decisions/0015-frontend-plan-json-only-contract.md`，并同步 `docs/runtime/frontend-implementation-workflow.md`、`docs/templates/frontend-implementation-contract.schema.json`、`skills/frontend-implementation/references/node-contracts.md`
 - 新增 `docs/decisions/0014-openspec-gate-claim-verification.md`，并同步 `docs/runtime/frontend-implementation-workflow.md`、`docs/templates/frontend-task-constraints.md`、`docs/templates/frontend-design-contract.md`、`skills/frontend-implementation/references/node-contracts.md`
 
+## [0.37.1-beta.0] - 2026-08-16
+
+### 重点更新
+
+- backend-test live campaign R115–R141 收口：全新 R141 达到 readiness PASS、23/23 correspondence exact、146/146 Test Points、77/77 scenario MATCH、traceability PASS，并唯一执行 82 个真实 API pytest body
+- 强化 backend-test correspondence、scenario-param、Markdown/pytest writer、collection repair、日志审计和 Windows pytest argfile 执行链，稳定区分 controller TestBug、generated TestBug 与真实 API ProductBug
+
+### 改进
+
+- backend-test correspondence 只分析目标请求实际绑定的 body/payload，忽略 `scenario` 元数据与 setup-only payload；对 `scenario["body"]` 保留字段 shape，并把 intentional invalid enum 交由 scenario-param 门禁判断；同时识别下划线前缀 pytest 参数集合与带说明的 `Payload Contract: none（…）`，避免零 eligible 的 payload/parameter 误阻断
+- backend-test correspondence 继续解析模块内 transport wrapper 的位置 payload 与 wrapper 构造 body，并清理 `Payload Enum` 值后的中文说明，避免 PUT/lifecycle 合法载荷误报；Markdown/pytest writer 同步禁止臆测响应 ID 类型，并在路径 ID 类型未声明时使用 create-delete 派生的合法 missing ID，避免把测试资产缺陷误分类为产品 400/404 缺陷
+- backend-test correspondence 可还原 `_apply_patch(_base_body(...), patch)` 等 transport-body 组合表达式，并把 setup wrapper payload 限定为有 body 契约时的 fallback；traceability 接受经递归脱敏和有界输出的结构化 request/response 日志，scenario-param 支持带显式例值的 string wrong-type；pytest writer/repair 同步禁止把 path/query/control `id` 混入 DTO patch
+- backend-test correspondence 识别模块级 `_post_case`/`_put_case` 参数工厂、factory TP 字面量与通过 `payload["field"] = value` 构造的 DTO shape，并避免空列表全局变量吞并后续参数集合；writer/repair 要求每个变体展开为直接 `pytest.param(..., id="TP-...")`，scenario readiness 不再依赖隐藏的动态参数行；traceability 允许窄 `urllib.error.URLError`/`httpx.RequestError` DELETE cleanup
+- backend-test correspondence 继续还原 `payload = _base_note_payload()` 形式的 builder-return dataflow，并把 intentional invalid enum 交给 scenario-param 而非 Case 聚合 payload gate；Markdown review 明确 GET/DELETE setup DTO 不得改写目标 no-body contract，pytest writer/repair 禁止把返回 scalar 的 helper 当 tuple 解包
+- backend-test collection repair 把 upstream 的 `Repair paths: testcase/...` 直接视为完整权威路径证据并读取/编辑该文件；禁止猜测根级 `contracts/**` 或 DAG run 目录，也禁止 read tool 已成功返回文件后反事实声称路径不存在并提前 blocked
+- backend-test 的 REPAIRABLE collection repair 现在强制 non-empty diff，禁止 `already-satisfied` 制造假修复；Markdown review 增加 `### 测试点` 与三类 binding union 的 exact-set 校验并删除 legacy alias；no-body Case 的 payload correspondence 忽略同一 journey 的 direct setup POST/PUT DTO
+- backend-test generated-local collection repair 从 HIGH 路由降为 MED，在保留 write guard 与 requireChangedFiles 的同时避开连续多轮 grok repair hallucinated blocked/changed-without-write，改由稳定 writer 路由执行有界单文件修复
+- writer outcome 协议确定性归一化显式中文 `实现结果：已变更|已满足|已阻塞` 到 canonical `IMPLEMENTATION_OUTCOME`，仍拒绝无标签 prose、未知值和冲突值，不根据 diff 猜测语义
+- scenario-param 识别 `"x" * _TITLE_MAX_LENGTH` 等长度表达式与 helper 常量名相乘形式
+- correspondence 忽略 pytest payload 中的 `expect`/`echo_*` 辅助键，避免把 DTO 外 helper 字段误判为 PAYLOAD_SHAPE_MISMATCH
+- scenario-param 解析 `bound=LEN100` 机器行、解引用 `_STATUS_ACTIVE` 等 Python 字符串常量，并把 `NOMINAL` 长度样本从 min/max 纠正为 nominal
+
+- scenario-param 观测识别 `_OMIT`/omit sentinel 为合法 `missing`，跳过 `"2xx"`/`"400"` 期望码位置参数，解码 Python `\t`/`\n` 空白字面量，识别 `_TITLE_LEN1` 等长度 helper 名，支持 `custom-literal:whitespace-only|whitespace-padded` 语义标签，`enum-invalid` 接受具体非法字面量，并在 `min`/`LEN1` 意图下优先读取 TP 编码长度与 `minLength` 边界
+- scenario-param 兼容 `empty` 意图下的 omit/`missing-key`，规范化 `custom-literal` 外包引号与可见空白符，并把 `max+1`/`min`/`max` 的裸数字位置参数按长度 token 比较（含 oversized 数字本身）；`custom-literal:trim` 视为 padded whitespace 语义标签
+- backend-test writer 合同明确：`enum-invalid` 必须使用具体非法字面量且禁止 `_OMIT`/missing-key；`custom-literal:trim|ACTIVE|ARCHIVED` 必须使用真实 padded/enum 字面量，禁止 `filter-active` 类描述性伪值
+- scenario-param 识别 `_*_MAX_LENGTH` / `_*_OVER_LENGTH` 等长度 helper 名，明确 `custom-literal:trim` 不能是全空白，并把无 pytest.param 的 request-level/health nominal 记为 MATCH 以免污染 eligibility
+
 ## [0.36.4-beta.0] - 2026-08-16
 
 > 前端 DAG 规范上下文确定性落地（beta，发布到 `beta` dist-tag）。
@@ -407,6 +478,61 @@
 - 会话创建期间侧栏与 Landing 页共享 busy 状态反馈，抑制双击导致的重复创建
 - 删除正在初始化的会话不会被迟到的后台初始化任务复活，初始化失败统一返回 HTTP 503 错误码
 
+## [0.33.7-beta.2] - 2026-08-13
+
+### Fixed
+
+- scenario-param 解析 `bound=LEN100`、解引用 `_STATUS_ACTIVE` 等 Python 字符串常量，识别 `"x" * _TITLE_MAX_LENGTH` 长度表达式
+- correspondence 忽略 pytest payload 中的 `expect`/`echo_*` 辅助键，避免假 PAYLOAD_SHAPE_MISMATCH
+
+### Changed
+
+- my-webapp 五轮 backend-test 全流程 live campaign 归档（R1–R5 均 finished；R5 correspondence PASS）
+
+## [0.33.7-beta.1] - 2026-08-12
+
+### Fixed
+
+- scenario-param 识别 `_*_MAX_LENGTH` / `_*_OVER_LENGTH` 等长度 helper 名，明确 `custom-literal:trim` 不能是全空白，并把无 pytest.param 的 request-level/health nominal 记为 MATCH 以免污染 eligibility
+- backend-test writer 合同明确：`enum-invalid` 必须使用具体非法字面量且禁止 `_OMIT`/missing-key；`custom-literal:trim|ACTIVE|ARCHIVED` 必须使用真实 padded/enum 字面量；max/min/max+1 应用长度表达式/helper
+
+### Changed
+
+- 归档 R108–R110 live campaign 证据与 residual 收紧结果
+
+## [0.33.7-beta.0] - 2026-08-11
+
+### 重点更新
+
+- backend-test 收口 source-fidelity P0：绑定 source coverage、有界 Markdown/pytest 同步、item-level eligibility 与单次 eligible 业务执行；my-webapp R107 16 节点 finished
+
+### 新增
+
+- backend-test 新增绑定 source references 的来源覆盖事实、Markdown 一轮增量同步和 pytest correspondence 一轮有界修复，在业务测试体执行前保留 initial/effective/residual 证据
+- backend-test 新增 item-level execution eligibility：只授权 payload-safe 的 pytest items，excluded items 与真实业务失败分开展示
+
+### 改进
+
+- 将 backend-test Markdown/pytest map 的默认总 token 预算从 `600000` 提升到 `3000000`，降低多模块生成时过早触发 `dynamic map token budget exhausted` 的概率；仍保持预算耗尽 fail-closed
+- Coverage Matrix 接受 `R400001` 等错误码 Rule Key，并保留单元格内 `ACTIVE|ARCHIVED` 未转义枚举管道符，避免合法行被拆成 10 列后整表丢弃
+- scenario-param 按 TP 精确匹配多行 `场景意图`，保留 `ACTIVE` 大小写字面量，识别 `"a" * N` 长度表达式与空白 `empty`，request-level `nominal` 允许 `None` 查询过滤；避免把合法 payload 误判为 MISMATCH 并错误排除 eligible items
+- backend-test 支持 evidence-only 元用例：Markdown 显式声明 `脚本/primary symbol=无` 且仅绑定 assertion/cross-cutting TP 时，不再强制 1:1 pytest 符号，也不再把其算作 MISSING_PYTEST/SCRIPT_MISMATCH
+- generated pytest 内的 `ScopeMismatch`/setup fixture 失败若完全归因于 `testcase/**` 映射脚本，则归为 REPAIRABLE 并进入唯一一轮 collection repair，不再直接 BLOCKED
+- scenario-param 观测跳过 `pytest.param("TP-...", value, id="TP-...")` 的前导 TP 标签，并支持长度意图用数字位置参数表达
+
+### 修复
+
+- 修复 readiness 错误信息截断与无效 Coverage Matrix 状态语义，避免超长错误或 advisory 矩阵状态误阻断后续节点
+- 修复 scenario 排除项未绑定 payload 字段、request-level 不确定性被误升级为硬失败的问题
+
+## [0.33.6-beta.0] - 2026-08-11
+
+### 修复
+
+- 修复 backend-test correspondence scanner 将下一 Case 的分隔注释误归入前一个 pytest function region，造成虚假的 `MULTIPLE_PYTEST`、script mismatch 和额外 Test Point；现在按函数实际缩进边界隔离 body，并只从精确 `Case-ID:` 元数据或 primary symbol 绑定 Case
+- 修复 backend-test 将 README 其他区域的 Markdown 链接误提取为业务模块、允许 `a401606` 等 opaque/hash-like stem 和异常多模块进入 map 的问题；现在只解析唯一 `## Module Index`，非法、mixed 或超过 8 个模块时 README-only 有界重试并由 manifest fail-closed
+- 修复 backend-test map 总 token 预算耗尽后只生成部分 Markdown/pytest 资产但 parent 仍可能显示成功的问题；budget-blocked child 现在使 backend map 明确 ERROR，并保留 expected/finished/blocked/token 事实
+
 ## [0.33.6] - 2026-08-11
 
 ### 重点更新

package/dist/build-stamp.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": 1,
   "version": "0.37.0",
-  "gitSha": "4e502c40653f25c38302f1104e367d49bc950791",
-  "builtAt": "2026-08-17T08:09:21.899Z"
+  "gitSha": "a2ca9f2fc2fa0e3eee17e28fd1ad187819d922b7",
+  "builtAt": "2026-08-19T15:13:16.805Z"
 }

package/dist/commands/init-upgrade.js

@@ -938,6 +938,80 @@ function changedWorkspacePaths(before, after) {
         .filter((relativePath) => before[relativePath] !== after[relativePath])
         .sort();
 }
+/**
+ * Generated-output prefixes treated as verification noise when git ignore
+ * rules are unavailable (non-git target). In a git repo the authoritative
+ * filter is `.gitignore` via `git check-ignore`; this list only covers the
+ * common case where a verification command such as `npm run build` writes
+ * untracked artifacts (build/dist/coverage/...) into the workspace.
+ */
+const WORKSPACE_NOISE_PREFIXES = [
+    "node_modules/",
+    "build/",
+    "dist/",
+    "coverage/",
+    ".turbo/",
+    ".next/",
+    ".nuxt/",
+    ".cache/",
+    ".parcel-cache/",
+    ".vite/",
+    "out/",
+    "target/",
+    "__pycache__/",
+];
+function isWorkspaceNoisePath(relative) {
+    if (relative === ".git" || relative.startsWith(".git/"))
+        return true;
+    if (relative === RUN_ROOT || relative.startsWith(`${RUN_ROOT}/`))
+        return true;
+    return WORKSPACE_NOISE_PREFIXES.some((prefix) => relative === prefix.slice(0, -1) || relative.startsWith(prefix));
+}
+/**
+ * Filters changed workspace paths down to meaningful (non-generated) changes
+ * so a verification command that legitimately writes git-ignored artifacts is
+ * not misjudged as an unauthorized workspace mutation. Git ignore rules are
+ * authoritative when the target is a git repo; otherwise the built-in
+ * generated-output prefixes apply.
+ */
+async function filterWorkspaceNoise(repoRoot, relativePaths) {
+    const builtinFiltered = relativePaths.filter((relative) => !isWorkspaceNoisePath(relative));
+    if (builtinFiltered.length === 0)
+        return builtinFiltered;
+    const ignored = await gitIgnoredPaths(repoRoot, builtinFiltered);
+    if (ignored === null)
+        return builtinFiltered;
+    return builtinFiltered.filter((relative) => !ignored.has(relative));
+}
+/**
+ * Returns the subset of paths matched by git ignore rules, or null when the
+ * target is not a git repository (or git is unavailable), in which case the
+ * caller falls back to the built-in noise prefixes.
+ */
+async function gitIgnoredPaths(repoRoot, relativePaths) {
+    return await new Promise((resolve) => {
+        execFile("git", ["-C", repoRoot, "check-ignore", ...relativePaths], {
+            timeout: 15_000,
+            maxBuffer: 4 * 1024 * 1024,
+            windowsHide: true,
+        }, (error, stdout) => {
+            if (error) {
+                // `git check-ignore` exits 1 when no path is ignored: that is a
+                // successful lookup with an empty result set. Any other failure
+                // (e.g. 128 outside a repo) means the git filter is unavailable.
+                if (error.code === 1 && stdout.length === 0)
+                    resolve(new Set());
+                else
+                    resolve(null);
+                return;
+            }
+            resolve(new Set(stdout
+                .split("\n")
+                .filter(Boolean)
+                .map((entry) => entry.split(path.sep).join("/"))));
+        });
+    });
+}
 function isFrozenVerificationSurface(value) {
     return (isRecord(value) &&
         value.schemaVersion === 1 &&
@@ -1587,7 +1661,7 @@ async function validateContinueMerge(directory, state, anchor) {
         await assertSafeMergePath(state.repoRoot, allowedPath);
     }
     const current = await snapshotWorkspace(state.repoRoot);
-    const changed = changedWorkspacePaths(authoritativeGuard.baselineWorkspace, current);
+    const rawChanged = changedWorkspacePaths(authoritativeGuard.baselineWorkspace, current);
     const pending = await readPendingAcceptance({
         directory,
         state,
@@ -1602,6 +1676,10 @@ async function validateContinueMerge(directory, state, anchor) {
             pending,
         })
         : undefined;
+    // Only the out-of-bound drift check treats git-ignored/generated-output
+    // changes as noise; allowed-path changes are merge content and must stay
+    // intact even when the target is git-ignored (e.g. `.agents/skills/*`).
+    const changed = await filterWorkspaceNoise(state.repoRoot, rawChanged);
     const unexpected = changed.filter((relativePath) => {
         if (allowed.has(relativePath))
             return false;
@@ -1611,7 +1689,7 @@ async function validateContinueMerge(directory, state, anchor) {
     if (unexpected.length > 0) {
         throw new Error(`init upgrade merge write-guard blocked out-of-bound workspace changes: ${unexpected.join(", ")}`);
     }
-    const changedAllowed = changed.filter((relativePath) => allowed.has(relativePath));
+    const changedAllowed = rawChanged.filter((relativePath) => allowed.has(relativePath));
     if (changedAllowed.length === 0) {
         if (pending) {
             throw new Error("pending merge acceptance cannot authorize a merge without an allowed workspace change");
@@ -1834,16 +1912,8 @@ async function createVerificationWriteAudit(input) {
     const note = (scope, file) => {
         events.push(`${scope}=${file.split(path.sep).join("/")}`);
     };
-    const ignoredWorkspacePath = (candidate) => {
-        const relative = toRepoPath(canonicalRepoRoot, candidate);
-        return (relative === ".git" ||
-            relative.startsWith(".git/") ||
-            relative === "node_modules" ||
-            relative.startsWith("node_modules/") ||
-            relative === RUN_ROOT ||
-            relative.startsWith(`${RUN_ROOT}/`));
-    };
-    const install = async (scope, directory, ignore) => {
+    const ignoredWorkspacePath = (candidate) => isWorkspaceNoisePath(toRepoPath(canonicalRepoRoot, candidate));
+    const install = async (scope, directory, ignore, scopeRoot) => {
         const canonical = await realpath(directory);
         if (watched.has(canonical))
             return;
@@ -1867,11 +1937,11 @@ async function createVerificationWriteAudit(input) {
                 // traversal opens an existing child directory. Nested watchers own real
                 // mutations below that child; persistent changes are also snapshot-checked.
                 if (!existingDirectory) {
-                    note(scope, path.relative(directory, candidate) || ".");
+                    note(scope, path.relative(scopeRoot ?? directory, candidate) || ".");
                 }
                 // A parent rename can introduce a new directory. Installing its watcher
                 // closes the delayed-descendant gap even when the container event is noise.
-                void install(scope, candidate, ignore).catch((error) => {
+                void install(scope, candidate, ignore, scopeRoot).catch((error) => {
                     const code = error.code;
                     if (code !== "ENOENT" && code !== "ENOTDIR") {
                         failures.push(`watcher refresh failed for ${scope}: ${error instanceof Error ? error.message : String(error)}`);
@@ -1891,10 +1961,10 @@ async function createVerificationWriteAudit(input) {
             const child = path.join(canonical, entry.name);
             if (ignore?.(child) || !entry.isDirectory() || entry.isSymbolicLink())
                 continue;
-            await install(scope, child, ignore);
+            await install(scope, child, ignore, scopeRoot);
         }
     };
-    await install("workspace", canonicalRepoRoot, ignoredWorkspacePath);
+    await install("workspace", canonicalRepoRoot, ignoredWorkspacePath, canonicalRepoRoot);
     await install("controller", runRoot(input.repoRoot));
     await install("controller-anchor", anchor.anchorRoot);
     const piHomeEntry = await lstat(piHome).catch((error) => {
@@ -2194,17 +2264,25 @@ async function runVerificationCommand(input) {
         if (audit.failures.length)
             throw new Error(audit.failures.join("; "));
         await assertFrozenVerificationSurface(repoRoot, surface);
-        const changedWorkspace = changedWorkspacePaths(workspaceBefore, await snapshotWorkspace(repoRoot));
+        const changedWorkspace = await filterWorkspaceNoise(repoRoot, changedWorkspacePaths(workspaceBefore, await snapshotWorkspace(repoRoot)));
         const changedController = changedWorkspacePaths(controllerBefore, await snapshotDirectory(runRoot(repoRoot)));
         const changedAnchors = changedWorkspacePaths(anchorBefore, await controllerAnchorSnapshot(repoRoot, state));
         const changedPiHome = changedWorkspacePaths(piHomeBefore, await snapshotOptionalDirectory(path.join(os.homedir(), ".pi")));
-        if (descendantsObserved || audit.events.length || changedWorkspace.length || changedController.length || changedAnchors.length || changedPiHome.length) {
+        // Workspace watcher events are repo-root relative (see scopeRoot in
+        // createVerificationWriteAudit); drop git-ignored/generated-output noise
+        // before deciding whether the subprocess wrote the workspace.
+        const workspaceEvents = audit.events.filter((event) => event.startsWith("workspace="));
+        const remainingEvents = [
+            ...audit.events.filter((event) => !event.startsWith("workspace=")),
+            ...(await filterWorkspaceNoise(repoRoot, workspaceEvents.map((event) => event.slice("workspace=".length)))).map((relative) => `workspace=${relative}`),
+        ];
+        if (descendantsObserved || remainingEvents.length || changedWorkspace.length || changedController.length || changedAnchors.length || changedPiHome.length) {
             return {
                 ...result,
                 ok: false,
                 stderr: boundedOutput(`${result.stderr}\nverification write guard blocked changes: ${[
                     ...(descendantsObserved ? ["process-tree=descendant"] : []),
-                    ...(audit.events.length ? [`events=${audit.events.join(",")}`] : []),
+                    ...(remainingEvents.length ? [`events=${remainingEvents.join(",")}`] : []),
                     ...(changedWorkspace.length ? [`workspace=${changedWorkspace.join(",")}`] : []),
                     ...(changedController.length ? [`controller=${changedController.join(",")}`] : []),
                     ...(changedAnchors.length ? [`controller-anchor=${changedAnchors.join(",")}`] : []),

package/dist/executors/dag-pi-executor.js

@@ -10,6 +10,7 @@ import { redactPromptForLog, truncateOutput, } from "../shared/output-truncation
 import { GitStatusUnavailableError, pathsChangedDuringRun, readGitStatusPorcelain, recoverRootNulArtifact, snapshotGitStatusPathFingerprints, snapshotGitStatusPorcelain, validateShellWriteGuard, } from "./shell-write-guard.js";
 import { isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, INCOMPLETE_WRITE_SET_RETRY_CATEGORY, WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY, WRITER_EMPTY_DIFF_RETRY_CATEGORY, } from "../workflows/dag/retry-policy.js";
 import { assessBackendTestMdPlanCompleteness, assessBackendTestMdWriterCompleteness, assessBackendTestPytestPlanCompleteness, assessBackendTestPytestWriterCompleteness, assessBackendTestShardChildCompleteness, backendTestWriterProgressRoleForTask, classifyBackendTestWriterCompletenessFailure, isBackendTestCompletenessRetryCandidate, isBackendTestMdPlanTask, isBackendTestPytestPlanTask, isBackendTestShardChildTask, writeBackendTestWriterProgressArtifacts, } from "../workflows/dag/backend-test-writer-completeness.js";
+import { resolveBackendTestLayout } from "../workflows/dag/backend-test-layout.js";
 import { redactSecrets, truncateUtf8Preview } from "../shared/preview.js";
 /**
  * Writer classification for a length-stopped thinking-only attempt: the model
@@ -808,6 +809,13 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
         isBackendTestCompletenessRetryCandidate(input.task)) {
         try {
             const writerRole = backendTestWriterProgressRoleForTask(input.task.id);
+            const backendLayout = resolveBackendTestLayout(meta.spec.backendTestLayout
+                ? {
+                    testRoot: meta.spec.backendTestLayout.testRoot,
+                    markdownRel: meta.spec.backendTestLayout.markdownRel,
+                    scriptRel: meta.spec.backendTestLayout.scriptRel,
+                }
+                : undefined);
             const progress = isBackendTestShardChildTask(input.task)
                 ? await assessBackendTestShardChildCompleteness({
                     workspaceRoot: input.cwd,
@@ -815,12 +823,12 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                     writeSet: input.task.writeSet ?? [],
                 })
                 : isBackendTestMdPlanTask(input.task)
-                    ? await assessBackendTestMdPlanCompleteness(input.cwd)
+                    ? await assessBackendTestMdPlanCompleteness(input.cwd, backendLayout)
                     : isBackendTestPytestPlanTask(input.task)
-                        ? await assessBackendTestPytestPlanCompleteness(input.cwd)
+                        ? await assessBackendTestPytestPlanCompleteness(input.cwd, backendLayout)
                         : writerRole === "pytest-generate"
-                            ? await assessBackendTestPytestWriterCompleteness(input.cwd)
-                            : await assessBackendTestMdWriterCompleteness(input.cwd);
+                            ? await assessBackendTestPytestWriterCompleteness(input.cwd, backendLayout)
+                            : await assessBackendTestMdWriterCompleteness(input.cwd, backendLayout);
             await writeBackendTestWriterProgressArtifacts({
                 runDir: meta.runDir,
                 progress,
@@ -977,6 +985,15 @@ function parseWriterImplementationOutcome(text) {
     const lines = text.split(/\r?\n/);
     const candidates = [];
     for (const [lineIndex, line] of lines.entries()) {
+        const localizedOutcome = localizedWriterImplementationOutcome(line);
+        if (localizedOutcome) {
+            candidates.push({
+                lineIndex,
+                value: localizedOutcome,
+                outcome: localizedOutcome,
+            });
+            continue;
+        }
         const nextLine = lines[lineIndex + 1];
         const normalized = normalizeProtocolLine(line, WRITER_OUTCOME_PROTOCOL_LINE, nextLine);
         if (normalized === undefined)
@@ -1005,6 +1022,26 @@ function parseWriterImplementationOutcome(text) {
         candidates,
     };
 }
+function localizedWriterImplementationOutcome(line) {
+    const normalized = normalizedProtocolValue(line);
+    const match = /^(?:实现|实施)结果\s*[:：]\s*(已变更|已更改|已修改|已满足|无需变更|已阻塞|阻塞)$/u.exec(normalized);
+    if (!match)
+        return undefined;
+    switch (match[1]) {
+        case "已变更":
+        case "已更改":
+        case "已修改":
+            return "changed";
+        case "已满足":
+        case "无需变更":
+            return "already-satisfied";
+        case "已阻塞":
+        case "阻塞":
+            return "blocked";
+        default:
+            return undefined;
+    }
+}
 function isWriterImplementationOutcome(value) {
     return (value === "changed" || value === "already-satisfied" || value === "blocked");
 }
@@ -1043,6 +1080,7 @@ export function mapPiResultToDagNodeResult(result, firstProtocolLine) {
         sdkAttempted: result.sdkAttempted,
         tokensUsed: result.tokensUsed,
         parsedEvents: result.parsedEvents,
+        stopReason: readWriterThinkingExhaustionEvidence(result).stopReason,
     };
 }
 function canonicalizeProtocolFirstLine(assistantText, firstProtocolLine) {