@tea-agent/loop-agent 0.35.4-beta.0 → 0.36.0

package/dist/workflows/dag/init-hybrid.js

@@ -730,6 +730,7 @@ function resolveImplementPaths(taskConfig, options = {}) {
         const allowed = [...taskConfig.allowedPaths];
         const forbidden = mergeForbiddenPaths(taskConfig);
         const repoRoot = options.repoRoot;
+        let companions = [];
         // Only consult the catalog when writer paths can touch docs/* so non-docs
         // temp fixtures without a catalog keep working; docs writers still fail closed.
         const mayNeedDocIndex = allowed.some((entry) => {
@@ -742,6 +743,7 @@ function resolveImplementPaths(taskConfig, options = {}) {
                 paths: allowed,
                 forbiddenPaths: forbidden,
             });
+            companions = closure.companions;
             const explicitAllowedPaths = new Set(allowed.map((entry) => entry.replace(/\\/g, "/").replace(/^\.\//, "").replace(/\/+$/, "")));
             const missingCompanions = closure.companions.filter((companion) => !explicitAllowedPaths.has(companion));
             if (missingCompanions.length > 0) {
@@ -751,11 +753,13 @@ function resolveImplementPaths(taskConfig, options = {}) {
         return {
             allowedPaths: allowed,
             writeSet: [...allowed],
+            docIndexCompanions: companions,
         };
     }
     return {
         allowedPaths: ["**"],
         writeSet: [IMPLEMENT_WRITESET_PLACEHOLDER],
+        docIndexCompanions: [],
     };
 }
 function shellQuote(value) {
@@ -1802,6 +1806,15 @@ export function buildStandardHybridDagFromTask(sources) {
                 subtask_prompt: [
                     "Implement the approved plan with minimal focused changes.",
                     "Stay within writeSet. Do not write root artifacts/** unless artifacts paths are explicitly declared in writeSet.",
+                    "Apply changes by calling the write/edit tools so files change on disk. Do not paste the new code in your response instead of writing it: the runner verifies the working-tree diff mechanically, and a response-only change is an empty diff that fails this node.",
+                    "Edit existing files with the structured edit/write tools. NEVER rewrite Markdown (or any file with quoting/backticks/indentation-sensitive content) via bash sed/awk/echo redirection: escaping mistakes silently corrupt the file and self-repair loops burn the run. If a bash text-substitution already damaged a file, restore the damaged region with the edit tool against the original content instead of stacking more substitutions.",
+                    "The plan node already designed the changes — that design is your input, not your output. Your job is to write those changes into the files in writeSet. The task is not done until the files on disk contain the changes.",
+                    "If the task only adds tests/docs that use existing code as-is, forbidden implementation paths (e.g. src/**) do not block you: you never need to write them. Write the tests/docs and report changed.",
+                    ...(implementPaths.docIndexCompanions.length > 0
+                        ? [
+                            `Doc index sync is MANDATORY: ${implementPaths.docIndexCompanions.join(", ")} are catalog index files for this writeSet. When you add, rename, or remove any indexed file, you MUST also update ${implementPaths.docIndexCompanions.join(" and ")} in the same run (append or fix the matching index entry following the existing line style). Verification runs check-doc-index and fails the run when a new file is missing from the index. Skipping this sync is not "minimal change" — it is an incomplete delivery.`,
+                        ]
+                        : []),
                     writerDeliveryContract(taskConfig),
                     sourceContext,
                 ]
@@ -2565,7 +2578,7 @@ async function buildFrontendHybridDagFromTask(sources) {
                 allowedPaths: readOnlyPaths,
                 forbiddenPaths,
                 skills: FRONTEND_IMPLEMENTATION_SKILLS,
-                outputContract: "Markdown implementation plan with Requirement Coverage, Implementation Steps, Target Files, UI State Handling, Styling / Component Strategy, Interaction Notes, Mock / API Strategy, Dependency Policy, Verification Plan, Real Integration Gap, and Residual Risks, ending with exactly ONE fenced json object (\`\`\`json ... \`\`\`) conforming to frontend-implementation-contract-v1. This fenced block is the single authoritative implementation contract the prewrite gate materializes; it must appear exactly once and must not contain or be followed by any raw JSON or extra fenced block. No file writes.",
+                outputContract: "Markdown implementation plan with Requirement Coverage, Implementation Steps, Target Files, UI State Handling, Styling / Component Strategy, Interaction Notes, Mock / API Strategy, Dependency Policy, Verification Plan, Real Integration Gap, and Residual Risks, ending with exactly ONE fenced json object (```json ... ```) conforming to frontend-implementation-contract-v1. This fenced block is the single authoritative implementation contract the prewrite gate materializes; it must appear exactly once and must not contain or be followed by any raw JSON or extra fenced block. No file writes.",
                 subtask_prompt: [
                     "Based on frontend-contract-pi, frontend-scout-pi, task sources, and the generation-time Mock capability evidence, return a minimal frontend implementation plan.",
                     "Select the Mock / API strategy inside the plan and structured contract. Carry endpoint/fixture mapping, explicit activation, production-default-off rule, verification commands, and Real Integration Gap into both outputs.",
@@ -2623,7 +2636,7 @@ async function buildFrontendHybridDagFromTask(sources) {
                 allowedPaths: readOnlyPaths,
                 forbiddenPaths,
                 skills: FRONTEND_IMPLEMENTATION_SKILLS,
-                outputContract: "When the initial design review requests revision, return a complete Markdown revision plan followed by exactly ONE fenced json object (\`\`\`json ... \`\`\`) conforming to frontend-implementation-contract-v1. This fenced block is the single authoritative implementation contract the prewrite gate materializes; it must appear exactly once and must not contain or be followed by any raw JSON or extra fenced block. Do NOT include multiple fenced JSON blocks; only the single authoritative contract JSON block is accepted. No file writes.",
+                outputContract: "When the initial design review requests revision, return a complete Markdown revision plan followed by exactly ONE fenced json object (```json ... ```) conforming to frontend-implementation-contract-v1. This fenced block is the single authoritative implementation contract the prewrite gate materializes; it must appear exactly once and must not contain or be followed by any raw JSON or extra fenced block. Do NOT include multiple fenced JSON blocks; only the single authoritative contract JSON block is accepted. No file writes.",
                 subtask_prompt: [
                     "Consume frontend-plan-pi (original plan) and frontend-design-review-pi (first design review findings).",
                     "This node runs only when frontend-design-review-pi emitted VERDICT: request-revision. Produce a complete revised implementation plan that addresses every Required Plan Correction from the design findings.",
@@ -2770,6 +2783,12 @@ async function buildFrontendHybridDagFromTask(sources) {
                     "frontend-prewrite-gate-shell confirmed the effective plan/review, requirement coverage, Mock policy, and contract. Stay within writeSet and preserve unrelated files.",
                     "For native, browser-intercept, or request-adapter, implement contract-aligned fixtures/states and a dev/test-only activation boundary in this same writer. For not-needed, do not add Mock files or a framework and state the positive reason.",
                     "Do not write root artifacts/** unless explicitly included in writeSet. Do not claim Browser/visual verification.",
+                    "Edit existing files with the structured edit/write tools. NEVER rewrite Markdown (or any file with quoting/backticks/indentation-sensitive content) via bash sed/awk/echo redirection: escaping mistakes silently corrupt the file and self-repair loops burn the run.",
+                    ...(implementPaths.docIndexCompanions.length > 0
+                        ? [
+                            `Doc index sync is MANDATORY: ${implementPaths.docIndexCompanions.join(", ")} are catalog index files for this writeSet. When you add, rename, or remove any indexed file, you MUST also update ${implementPaths.docIndexCompanions.join(" and ")} in the same run (append or fix the matching index entry following the existing line style). Verification runs check-doc-index and fails the run when a new file is missing from the index.`,
+                        ]
+                        : []),
                     writerDeliveryContract(taskConfig),
                 ]
                     .filter((value) => Boolean(value))
@@ -2825,6 +2844,11 @@ async function buildFrontendHybridDagFromTask(sources) {
                     "The repair assessment and canonical implementation contract are complete inputs for this phase. Do not re-open task sources, OpenSpec, AI workspace, plan/revision, or design-review prose, and do not repeat repository-wide discovery.",
                     "Do not change lint/type/test config, do not add .skip/.only, do not comment out real requests, do not default-enable Mock, do not add dependencies.",
                     "Do not re-plan requirements or expand allowed paths. Browser/visual remain not-run.",
+                    ...(implementPaths.docIndexCompanions.length > 0
+                        ? [
+                            `Doc index sync is MANDATORY: ${implementPaths.docIndexCompanions.join(", ")} are catalog index files for this writeSet. When your repair adds, renames, or removes any indexed file, update ${implementPaths.docIndexCompanions.join(" and ")} in the same run; verification runs check-doc-index and fails the run on a missing index entry.`,
+                        ]
+                        : []),
                     writerDeliveryContract(taskConfig),
                 ]
                     .filter((value) => Boolean(value))
@@ -4714,6 +4738,53 @@ function buildMultiPerspectiveReviewNodes(input) {
         ].join("\n\n"),
     }));
     const reviewIds = reviewNodes.map((n) => n.id);
+    if (input.structuredVerdictGate) {
+        const aggregateNodeId = `${input.nodePrefix}aggregate-pi`;
+        const aggregateNode = {
+            id: aggregateNodeId,
+            depends_on: reviewIds,
+            role: "reviewer",
+            executor: "pi",
+            complexity: "HIGH",
+            writePolicy: "read-only",
+            allowedPaths: input.allowedPaths,
+            forbiddenPaths: commonForbiddenPaths(input.sources),
+            outputContract: "Plain Markdown; first non-empty line is exactly VERDICT: pass only when every upstream perspective emitted VERDICT: pass, otherwise VERDICT: request-revision. No file writes.",
+            subtask_prompt: [
+                `Aggregate ${input.gateLabel} without changing any review finding.`,
+                `Inspect every upstream result: ${reviewIds.join(", ")}.`,
+                "First non-empty line must be exactly VERDICT: pass only when every upstream first verdict is VERDICT: pass; otherwise use VERDICT: request-revision.",
+                "List each upstream node and its observed verdict as evidence. Do not invent or override a missing/request-revision verdict.",
+                "Read-only: do not modify files.",
+            ].join("\n\n"),
+        };
+        const gateNode = {
+            id: input.gateId,
+            depends_on: [aggregateNodeId],
+            role: "verifier",
+            executor: "shell",
+            complexity: "LOW",
+            writePolicy: "read-only",
+            allowedPaths: commonReadOnlyPaths(input.sources),
+            forbiddenPaths: commonForbiddenPaths(input.sources),
+            outputContract: `Structured verdict gate: exit 0 only when ${aggregateNodeId} emits VERDICT: pass.`,
+            subtask_prompt: `Structured gate for ${input.gateLabel}: all perspectives must pass before downstream apply/promote.`,
+            shell: {
+                commands: [],
+                verdictGate: {
+                    fromNodeId: aggregateNodeId,
+                    accept: ["pass"],
+                    routingAccept: ["request-revision"],
+                    source: "verdict-line",
+                    lineMode: "first-verdict-line",
+                    label: input.gateLabel,
+                },
+                cwd: ".",
+                timeoutMs: 60000,
+            },
+        };
+        return [...reviewNodes, aggregateNode, gateNode];
+    }
     const aggregateScript = buildMultiPerspectiveReviewAggregateScript(reviewIds, input.gateLabel);
     const gateNode = {
         id: input.gateId,
@@ -4973,6 +5044,7 @@ function buildKnowledgeSyncMultiReviewNodes(sources, featureId) {
             `Primary draft path: ${knowledgeSyncDraftRelPath(featureId)}.`,
             "Apply is blocked until all perspectives pass. Do not approve fabricated verification pass.",
         ],
+        structuredVerdictGate: true,
     });
 }
 function buildKnowledgeSyncApplyNode(sources, featureId) {

package/dist/workflows/dag/lifecycle.js

@@ -743,7 +743,6 @@ function findDoctorFailureNode(state) {
             nodeId: state.pausedByNodeId,
             status: node?.status,
             rawFailureCategory: node?.failureCategory,
-            skippedReason: node?.skippedReason,
         };
     }
     const errorEntry = Object.entries(state.nodes).find(([, node]) => node.status === "ERROR");
@@ -759,7 +758,6 @@ function findDoctorFailureNode(state) {
         nodeId: selected[0],
         status: selected[1].status,
         rawFailureCategory: selected[1].failureCategory,
-        skippedReason: selected[1].skippedReason,
     };
 }
 async function readRunOwnedBackendTestClassification(runDir) {
@@ -788,7 +786,6 @@ async function resolveDoctorFailureRouting(input) {
             normalizedFailureCategory: "unknown",
             nodeId: input.nodeId,
             productLineFailureCategory: classifiedCategory,
-            skippedReason: input.skippedReason,
         });
     }
     return routeDagFailure(input);
@@ -810,7 +807,6 @@ async function formatDagDoctorMarkdown(repoRoot, runId) {
         rawFailureCategory,
         normalizedFailureCategory: normalizedCategory,
         nodeId: failure.nodeId,
-        skippedReason: failure.skippedReason,
     });
     const evidence = failure.nodeId
         ? path.join(located.runDir, failure.nodeId, "result.summary.md")

package/dist/workflows/dag/report.js

@@ -441,7 +441,6 @@ export async function buildDagRunReportEntry(input) {
                 normalizedFailureCategory,
                 nodeId,
                 executor: node.executor,
-                skippedReason: node.skippedReason,
             });
             const followUp = node.status === "ERROR" || node.status === "SKIPPED"
                 ? recommendFollowUpForFailureCategory(node.failureCategory)
@@ -506,7 +505,6 @@ export async function buildDagRunReportEntry(input) {
             normalizedFailureCategory: normalizeDagFailureCategory(pausedNode.failureCategory, pausedNode.status),
             failureCategory: pausedNode.failureCategory,
             nodeId: input.state.pausedByNodeId,
-            skippedReason: pausedNode.skippedReason,
         }
         : firstActionableNode
             ? {
@@ -515,7 +513,6 @@ export async function buildDagRunReportEntry(input) {
                     normalizeDagFailureCategory(firstActionableNode.failureCategory, firstActionableNode.status),
                 failureCategory: firstActionableNode.failureCategory,
                 nodeId: firstActionableNode.nodeId,
-                skippedReason: input.state.nodes[firstActionableNode.nodeId]?.skippedReason,
             }
             : {
                 status: input.state.status,
@@ -532,9 +529,6 @@ export async function buildDagRunReportEntry(input) {
         rawFailureCategory: runRecoverySource.failureCategory,
         normalizedFailureCategory: runRecoverySource.normalizedFailureCategory,
         nodeId: "nodeId" in runRecoverySource ? runRecoverySource.nodeId : undefined,
-        skippedReason: "skippedReason" in runRecoverySource
-            ? runRecoverySource.skippedReason
-            : undefined,
     });
     const runFollowUp = input.state.status === "failed" ||
         input.state.status === "partial_failed"

package/docs/README.md

@@ -2,7 +2,7 @@
 
 `docs/` 是 loop-agent 源仓的治理文档根目录。新读者不需要先浏览所有文件，先按任务选择入口，再由目录 README 或 `docs/document-catalog.json` 判断文档身份。
 
-**维护日：2026-07-27** — 对齐 `@tea-agent/loop-agent@0.24.0`。版本事实以根 `CHANGELOG.md` 与 `package.json` 为准；能力短摘要见 `reports/current-capability-summary.md`。
+**维护日：2026-08-16** — 对齐 `@tea-agent/loop-agent@0.35.3`。版本事实以根 `CHANGELOG.md` 与 `package.json` 为准；能力短摘要见 `reports/current-capability-summary.md`。
 
 ## 我现在要做什么？
 
@@ -10,7 +10,7 @@
 | --- | --- | --- |
 | 理解当前能力与边界 | `reports/current-capability-summary.md` | `architecture/README.md`、`website/docs/overview/roadmap.md` |
 | 开始一个受治理功能变更 | `governance/feature-workflow.md` | `governance/verification-matrix.md`、`templates/exec-plan.md` |
-| 跑或恢复 Agent DAG | `runtime/agent-dag-runner.md` | `runtime/agent-dag-recovery-playbook.md`、`runtime/loop-agent-harness.md` |
+| 跑或恢复 Agent DAG | `runtime/agent-dag-runner.md` | `runtime/agent-dag-recovery-playbook.md`、`runtime/loop-agent-harness.md`；Operator Chat Phase 1/P0 的 meaningful wait 与安全 terminal summary 已实现，Phase 2/3 仍见 active design |
 | 判断架构边界 | `architecture/README.md` | `architecture/runtime-boundaries.md` |
 | 查设计输入而非实现证明 | `design/README.md` | 对应设计文档；视觉参考见 `design/visual-references/README.md`，归档设计见 `design/archive/README.md` |
 | 接手进行中任务 | `exec-plans/active/README.md` | 对应 active plan，再读 `progress/README.md` |
@@ -26,7 +26,7 @@
 | 选择 taskKind、治理 profile 和 Agent DAG 路径 | `governance/feature-workflow.md` |
 | 选择快速、标准或完整验证命令 | `governance/verification-matrix.md` |
 | 理解 CLI、runtime 和初始化能力边界 | `runtime/loop-agent-harness.md` |
-| 运行 DAG、查看报告或处理失败 | `runtime/agent-dag-runner.md`、`runtime/agent-dag-recovery-playbook.md` |
+| 运行 DAG、查看报告或处理失败 | `runtime/agent-dag-runner.md`、`runtime/agent-dag-recovery-playbook.md`；Operator Chat P0 closeout 见 `reports/feature/2026-08-14-operator-chat-dag-recovery-p0.md` |
 | 按 TDD、验证和调试纪律推进修复 | `governance/harness-methodology-tdd.md`、`governance/harness-methodology-verification.md`、`governance/harness-methodology-debugging.md` |
 | 执行真实 backend-test provider 验证 | `runtime/backend-test-live-campaign.md` |
 | 判断单任务 loop-agent Production Readiness | `operations/production-readiness.md` |

package/docs/architecture/evolution.md

@@ -1,73 +1,108 @@
-# 架构演进：当前 vs 未来
+# 架构演进：当前内核与未来边界
 
-本页区分 loop-agent **当前已实现**的架构能力与**未来规划**。当前事实以代码、发布 CLI、已完成计划为准；未来能力一律标「规划 / 未实现 / 前瞻」。权威源：`CHANGELOG.md`、`docs/reports/current-capability-summary.md`、ADR 0001–0005、`docs/exec-plans/completed/` 与 `docs/exec-plans/active/`。
+本文区分 `loop-agent` 当前已实现架构和第 3–6 月前瞻。当前事实以代码、`CHANGELOG.md`、`docs/reports/current-capability-summary.md` 与 completed/active plans 为准；路线以 `docs/design/active/六个月规划.md` 为准。
 
-**对照版本：`@tea-agent/loop-agent@0.24.0`（2026-07-27）**。细节版本条目见根 `CHANGELOG.md`；本表只保留架构层可读摘要。
+对照版本：`@tea-agent/loop-agent@0.35.3`（2026-08-16）。
 
-## 当前已实现（累计至 0.24.0）
+## 当前架构
 
-| 域 | 现状 | 权威入口 |
+```text
+Skills / AGENTS
+      ↓
+loop-agent CLI → application use cases → workflow runtime → Pi/Shell/Static executors
+      ↓                                  ↓
+task source / contracts              run-owned facts / artifacts / receipts
+
+agent-worker CLI / Console
+      ↓ frozen published controller subprocess
+loop-agent CLI
+```
+
+### 已实现能力
+
+| 域 | 当前事实 | 权威入口 |
 | --- | --- | --- |
-| 受治理 Agent runtime | **Pi-only**；`cursor-prompt` 仅显式 one-shot sidecar | ADR 0001、`CHANGELOG.md [0.10.0]` |
-| Agent DAG 主链 | `dag run-task` → generate → validate → `run-dag`；report/doctor/reconcile-run | `dag-execution.md`、`test/cli-contract.test.ts` |
-| Dynamic Workflow | `WorkflowSpec` → validate → compile → 同一 `run-dag`；agent-like 节点仅 `pi\|static` | `src/workflows/dynamic/**`、`website/docs/guides/dynamic-workflow.md` |
-| 版本化自举 | controller identity + run-owned skill snapshot + deterministic canary | `docs/reports/feature/2026-07-13-versioned-self-hosting-bootstrap.md` |
-| Feature 交付（M2） | review/run/approve-followup/delivery/closeout/verify-final | `docs/reports/feature/2026-07-12-m2-completion-audit.md`、`CHANGELOG.md [0.10.0]` |
-| Task Pool | 唯一根 `.harness/task-pool/`；**feature-scoped** state v2（`TaskPoolTaskRef`） | ADR 0002、ADR 0004、`CHANGELOG.md [0.13.0]` |
-| Ready Planner | priority 选择；`batch plan-ready` 零写入；`run-ready` 冻结 plan；产物门禁 | `CHANGELOG.md [0.13.0]` / `[0.16.0]` |
-| TaskSpec → workflow | 可选 `execution.workflow` → `taskKind`；typed Outcome；`fullstack-v1` Verification Bundle | completed `2026-07-19-taskspec-workflow-routing.md`、`CHANGELOG.md [0.16.0]` |
-| 专用 taskKind DAG | `frontend-implementation` / `frontend-test` / `backend-test` / knowledge* | `website/docs/guides/agent-dag.md`、`CHANGELOG.md [0.12.0]`–`[0.17.x]` |
-| Backend-test | **Markdown-first 8 节点**；Markdown 映射限定 pytest 执行，生成 Markdown 与分层 HTML/脱敏请求响应详情 | completed `2026-07-24-backend-test-scoped-execution-report-ui.md`；`CHANGELOG.md [0.21.0]` / `[0.22.0]` |
-| Frontend-test | 默认 short-chain RAG；内容/evidence 质量进入 advisory findings，路径与安全边界仍 fail-closed | `CHANGELOG.md [0.21.0]` / `[0.22.0]` |
-| Frontend-implementation | Contract / trace / repair / Mock assess；`ai_workspace` + 知识库 + `openspec` 规范发现；lint 基线债务隔离 | `CHANGELOG.md [0.22.0]` |
-| Inspect（原 Observe） | 统一 Console 的 `/inspect/` 只读运营面与富时间线；独立 `observe serve` 已硬下线 | `website/docs/guides/observe-ui.md`、`CHANGELOG.md [Unreleased]` |
-| Local Operator Console | `agent-worker console\|doctor`（loopback；默认打开系统浏览器，`--no-open` 禁止）；Operate + Inspect、Happy Path / Interview / split view / recovery CTA；**0.23.x 起含 General Operator Chat**（Pi SDK services session、operator-only tool surface）；**非**远端多用户 Console | unify design；ADR 0005；completed `2026-07-25-console-phase-4-general-operator-chat.md` |
-| Task Contract / operator surface | journaled `task contract *`；machine envelope；**DagSpec v4 `taskContractBinding`**（exclusive writer 强制） | `CHANGELOG.md [0.17.0]`–`[0.17.2]`；ADR 0005 |
-| DagSpec / repair | v3 `runtimeContract` + 显式 `repairNodeId`；新 writer 生成默认 v4 binding | `dag-execution.md`、`CHANGELOG.md [0.11.0]` / `[0.17.0]` |
-| 自适应 liveness | runner lease、Provider/tool/output 活动与 meaningful progress 分离；静默 stall 受控终止，退出未确认时禁止自动重试 | `dag-execution.md`、active `2026-07-23-dag-adaptive-liveness.md` |
-| 运行恢复 | 节点内协议纠错重试、`dag rerun` 安全子图续跑、`dag rerun-task` 完整任务重跑与 Console Recovery | `CHANGELOG.md [0.20.1]`、active `2026-07-23-operator-rerun-hardening.md` |
-| Eval Lab | corpus / campaign / promote（人工门禁）；`autoPromote=false`；**非** RSI Level 1 | `CHANGELOG.md [0.15.0]`；active plan 可归档 |
-| 文档双树 | `website/docs/` 用法 vs `docs/` 治理；docs-converge | ADR 0003 |
-| 文档治理 | 任务式 Start Here、机器可读身份清单、目录默认身份与版本/索引语义检查 | 本目录 README、`docs/README.md`、`docs/document-catalog.json` |
-
-### 受治理 runtime 的边界（已实现、不变式）
-
-- DAG writer 固定 `implement-pi` / `repair-pi`；`executor: "cursor"`、`implement-cursor` / `repair-cursor`、Cursor worker、`cursor-fix` 已从受治理路径移除。
-- 完成权威 = shell verification；model verdict / Observe / 报告是 derived/advisory。
-- Worker 通过已发布 `loop-agent` 子进程执行，不 in-process import runtime kernel（governance 机器校验）。
-- Console / Compatibility 主会话不得绕过 CLI 直接改业务实现（ADR 0005）。
-
-## 未来规划（第 3–6 月，**未实现**）
-
-以下能力来自 `docs/design/active/六个月规划.md` 与相关设计输入（页首校准条之后的未交付 phase 为**设计输入**）。它们**当前不存在于产品声称的线上形态**，或仅有本地/部分替代：
-
-| 未来方向 | 状态 | 规划来源 |
+| 标准生命周期 | `task advance` 是唯一标准 mutation 入口；`task status` 只读；run 级 `dag` 命令保留为 advanced recovery/forensic | ADR 0008、`docs/governance/feature-workflow.md` |
+| DAG kernel | Contract/Scout/Plan/Implement/Verify/Closeout；dynamic workflow、pause/resume、doctor/report/rerun/reconcile | `dag-execution.md`、`agent-dag-runner.md` |
+| 写入与契约 | Pi-only writer、Task Contract、DagSpec v4 binding、writeSet review、writer admission、write guard | ADR 0001、`runtime-boundaries.md` |
+| 专用 taskKind | `frontend-implementation`、`backend-test`、`frontend-test`、knowledge 路径；显式 taskKind 贯穿 intake→task→DAG | `CHANGELOG.md [0.35.0]`–`[0.35.3]` |
+| Backend test | map-shard、execution readiness、generated-output conflict、collection/pytest 结构化失败、恢复加固 | completed plans、`CHANGELOG.md [0.30.0]`–`[0.31.1]` |
+| Frontend implementation | canonical prewrite result、两层 writer gate、统一 writer helper、rollback journal、child recovery、root manifest | completed `2026-08-13-frontend-implementation-dag-optimization.md` |
+| Frontend test | lean 默认 DAG、case map、evidence/quality report；安全边界 hard gate，内容质量 advisory | completed `2026-08-07-frontend-test-dag-lean-nodes.md` |
+| Worker / Feature | TaskSpec、feature-scoped Task Pool、Ready Planner、Outcome、Verification Bundle、Delivery/Closeout | `worker-and-feature.md`、current capability summary |
+| Worker readiness baseline | Attempt identity、immutable owner、lineage/revision、task-scoped lease、controller protocol、acceptance、reconcile | Agent Worker PR v1 Phase 1–7 / `0.32.0` |
+| Operator Surface | 裸 `agent-worker console`；Operate、`/inspect/`、Operator Chat、recovery CTA、intake bridge | ADR 0005、0011–0013 |
+| 自举与验证 | controller identity、skill snapshot、deterministic/live canary、verified tarball、`verify:tree` receipt | `runtime-boundaries.md`、verification matrix |
+
+### 当前不变量
+
+- `loop-agent` 拥有 repo-local DAG kernel；`agent-worker` 通过冻结的已发布 CLI 子进程执行，不复制内核。
+- 受治理 Agent writer 只有 Pi；`cursor-prompt` 只是显式 one-shot sidecar。
+- shell verification / Final Verification 是完成权威；模型 verdict 与 UI projection 不是。
+- Console 是本地 loopback 控制面；canonical mutation 仍经 sibling CLI/application contract。
+- controller identity 和 run-owned skill snapshot 分属不同冻结层，不能互相替代。
+- recovery 必须保留原始终态并生成显式 reconciliation/continuation 事实。
+
+完整 import 与 runtime 依赖规则见 `runtime-boundaries.md`，本文不复制其机器校验表。
+
+## 当前尚未 Ready
+
+| 项目 | 状态 |
+| --- | --- |
+| Agent Worker PR v1 Phase 8 | `NOT_READY`：burn-in、live Closeout、C4/C7 crash 未收口；live Pi canary 已验证 |
+| Fullstack release train | I40 Delivery/Closeout live；FC-B orphan recovery failed，overall=partial |
+| Operator Chat primary session | 主体能力已发布；event recovery / dogfood 门槛仍 open |
+| frontend-test | formal V7 smoke 与独立成功率报告 open |
+| backend-test | 真实 provider 五路径 smoke blocked-external |
+
+## 下一阶段目标架构
+
+未来能力不直接塞入当前 runtime，而是增加两个外层边界：
+
+```text
+Organization Control Plane
+  Feature Graph / Queue / Policy / Approval / Remote Adapters
+                     ↓ assignment + lease
+Isolated Execution Plane
+  Workspace Manager / published loop-agent / repo-local agent-worker projection
+                     ↓ immutable refs + receipts
+Evidence & Trust Plane
+  Artifact Manifest / Git & CI Receipts / Audit / Metrics / SLO
+```
+
+### 演进顺序
+
+| 阶段 | 新边界 | 不能破坏的旧边界 |
 | --- | --- | --- |
-| 远程 PR / CI 自动创建与写回 | 规划 / 未实现 | `docs/design/active/六个月规划.md`（第 3 个月起） |
-| 线上 / 云 Worker | 规划 / 未实现 | 同上 |
-| 云 Task Pool / SQL / Orchestrator | 规划 / 未实现 | 同上（第 2 月原始设计已调整为本地 Feature 闭环） |
-| 多仓库平台 | 规划 / 未实现 | 同上 |
-| 组织级服务 | 规划 / 未实现 | 同上 |
-| Web Console（远端多用户） | 规划 / 未实现；**不等于**已发布的本地 loopback Operator Console | 同上；对照 `docs/design/archive/2026-07-21-local-operator-console-from-pi-web.md` |
-| fullstack Delivery / Final Verification live 闭环 | 进行中（active train） | `docs/exec-plans/active/2026-07-19-fullstack-dogfood-016x-release-train.md` |
-| Dynamic Workflow runtime limits 强执法、更广 profile | 设计输入 | `docs/design/active/dynamic-workflow-dag-engine-roadmap.md` |
-| Loop 与 Dynamic Workflow 更深双向集成 / 自动恢复 | 设计输入 | 同上；当前已有基础 `workflow` action |
-| Operator Chat 作为主操作会话（裸 bash 收敛、真 human gate、durable session/turn、Operation 闭环、完整 CLI action coverage、默认入口切换） | 设计输入 / 未达主会话门槛 | `docs/design/active/operator-chat-primary-session-roadmap.md`；Phase 4 General Operator Chat 本体已实现归档（`docs/exec-plans/completed/2026-07-25-console-phase-4-general-operator-chat.md`） |
-| Eval Lab auto-promote / RSI Level 1 产品声称 | **禁止**写成已交付 | Eval Lab design + active plan |
-
-> 本地 Operator Console 统一面（实际首发 0.19.0）已发布：单仓库、loopback、随 `@tea-agent/loop-agent` 同包、canonical mutation 只经 sibling CLI，Inspect 复用 GET-only Observe read model。它不是远端 Web Console，也不能把多租户/云编排需求偷渡进本地 Console。
-
-> 注意：`docs/design/active/dynamic-workflow-dag-engine-roadmap.md` 含 2026-07-04 历史叙述；文中凡把 Cursor 写成受治理 executor 或 `loop` 的 `cursor-fix` 动作，均为**历史叙述**，现状以 Pi-only + 显式 `cursor-prompt` sidecar 为准。
-
-## 已收敛为 archive / 历史基线（非未来）
-
-- 第 1–2 月规划已收敛为 archive/reports 指针，不在本文展开：`docs/design/archive/2026-07-12-第二月规划.md`。
-- `docs/reports/2026-07-02-repository-analysis.md` 自 2026-07-14 起冻结为**历史基线快照**（允许有限再采样），不当第二 CHANGELOG。
-- 活能力短摘要在 `docs/reports/current-capability-summary.md`。
-- 专题证据按类型落在 `docs/reports/{feature,init-evolution,merge,fullstack-dogfood,dogfood}/`。
-
-## 文档本体的演进边界
-
-- 本目录新增文档随发布包发布（`package.json` `files` + manifest `packageRequired` 显式条目），但 **不**投影到目标项目 init surface；目标项目 init 仍只投影语言无关的 `runtime-boundaries.md`。
-- `runtime-boundaries.md` 是边界真源；本目录其他文档交叉引用，不复制其 import 方向表 / governance-hook 表 / 版本化自举边界表。
-- 未来如有能力落地，应先改 `src/` / CLI / ADR / completed plan，再回写本目录的「已实现」表与站上 `website/docs/overview/*` 短索引。
+| 远程交付 | Git/PR/CI adapter、head-bound receipt、approval | repo-local writer gate 与 verification authority |
+| 隔离 Worker | workspace lease、fencing、credential/network policy、artifact projection | frozen controller 与 DAG run facts |
+| 跨仓库 Feature | repo-qualified TaskRef、contract hash、saga/compensation | 每仓库独立 Task/PR/CI/terminal state |
+| 生产治理 | versioned policy、audit、SLO、semi-auto lane | owner opt-in、可撤销授权、fail-closed |
+
+### 未来合同
+
+- `AssignmentEnvelope`
+- `WorkspaceLease`
+- `ControllerAttestation`
+- `GitMutationReceipt`
+- `CiRunReceipt`
+- `ArtifactManifest`
+- `PolicyDecision`
+- `FeatureCloseout`
+
+这些对象必须先有 schema、fixture、verifier、失败语义和 ADR，再接真实 provider。
+
+## 仍属前瞻、禁止写成已交付
+
+- 自动远程 push / PR / CI / merge。
+- 线上队列、容器 Worker Pool、多租户与跨 region 调度。
+- 组织级凭据代理、策略服务与审计服务。
+- 跨仓库原子交付；目标模型是 saga 与补偿，不是伪全局事务。
+- Eval Lab auto-promote、默认自动 merge、自动发布或 RSI Level 1。
+
+## 文档演进规则
+
+1. 新能力先由代码、测试、live campaign、completed plan 或正式报告证明。
+2. 证明成立后更新 current capability summary 和本页“当前架构”。
+3. 未证明部分只保留在 active design/exec-plan，不能通过修改措辞升级状态。
+4. 使用者短索引同步到 `website/docs/overview/roadmap.md`，避免复制长版本流水账。

package/docs/templates/frontend-task-constraints.md

@@ -18,19 +18,13 @@ TODO
 
 ## Mock 约束（数据型任务）
 
-本段决定生成期 `allowedMockStrategies` 与冻结的 Mock 验证命令。只有当生成期**能证明任务确实包含 Mock 且有确定性验证命令**时，DAG 才会放行 `native` / `browser-intercept` / `request-adapter`；否则自动收敛为 `not-needed`，design review 若按项目规范改选 `native`，会在 prewrite 被 `mock-strategy-outside-allowed` 拦截、`implement` 被跳过。
-
-- **何时必须填**：项目规范（`openspec/project-specs/**`、`ai_workspace/**`、mock 规则或类似 DEC-* 决策）要求/建议 Mock；或需求涉及远端接口且后端未就绪。规范要求 Mock 时，优先 `task.json.frontendMock.policy: "required"`。
-- **命令要探索、不要硬编码**：到项目 `package.json` scripts 里找实际存在的 Mock 相关脚本（如 `mock`、`mock:*`、`dev:mock`，或名字含 mock 的脚本），结合既有 service root 的 handler/fixture/bootstrap 启动方式，确定一条**真实存在、确定、可自终止**（启动→断言→退出 0）的验证命令。常驻 dev server 必须包装成自终止脚本（start→assert→stop），否则 verify shell 会超时 fail-closed。
-- **写入 task.json**：把探索到的命令原样写进 `frontendMock.verifyCommands`（`label` 唯一、`command` 与项目脚本一致）。`label` 会进入冻结命令集，implementation contract 的 `verificationTarget.commandLabel` 必须逐字引用它。
-- **命令来源白名单**：只允许项目 `package.json` 已有脚本、Mock capability seed、或本段声明的 `verifyCommands`；plan/design 阶段不能发明 shell 命令。
-- **Mock/API/schema 规范路径**：TODO
-- **既有 Mock service root、handler/fixture/bootstrap**：TODO
-- **既有 browser/e2e interception 或 request adapter/DI seam**：TODO
-- **production 禁用边界**：TODO
-- **真实请求默认路径与 Mock 显式启用方式**：TODO
-- **`task.json.frontendMock.policy`**：`auto | required | disabled`（规范强制 Mock 用 `required`）
-- **被拦截时怎么修**：`mock-strategy-outside-allowed` / `no authorized Mock verification commands` 是**生成期契约问题，不是 plan 问题**——补 `frontendMock.verifyCommands`（或 `policy: "required"` + 命令）后**重新生成 DAG** 再跑，plan-revision 无法修复它。
+- Mock/API/schema 规范路径：TODO
+- 既有 Mock service root、handler/fixture/bootstrap：TODO
+- 既有 browser/e2e interception 或 request adapter/DI seam：TODO
+- 启动、健康检查和专项验证命令：TODO
+- production 禁用边界：TODO
+- 真实请求默认路径与 Mock 显式启用方式：TODO
+- `task.json.frontendMock.policy`：`auto | required | disabled`
 
 ## allowedPaths
 

package/docs/templates/init-managed-agents.md

@@ -33,7 +33,8 @@
 | **允许** | 已发布 `loop-agent` / `agent-worker` CLI；只读 status/doctor/report/inspect/observe；准备 `source/*` 与 `task.json` 边界；human gate；shell 验证与 handoff。 |
 | **禁止** | 绕过 CLI 用宿主 Edit/Write/ApplyPatch 直接改业务实现；CLI/DAG 失败后「救火改文件」；用聊天自述代替 shell 验证。 |
 | **失败时** | 先 `dag report` / `dag doctor`；优先 `dag rerun --from-node` 安全子图续跑；仅契约/源真变或 plan 不合格时同 task 重 advance / `dag rerun-task`；paused 用 approve→`dag resume`。 |
-| **实现写入** | 业务代码 **只** 经受治理 DAG writer（`implement-pi` / `repair-pi`）经 `task advance`（批准 writeSet gate 后长跑）。 |
+| **实现写入** | 业务代码 **只** 经受治理 DAG writer（`implement-pi` / `repair-pi`）经 `task advance`（批准 writeSet gate 后长跑）由 Console operation 后台持有；主会话**不得**在前台 Bash 直接执行 `task advance --approve-gate`，不得把长跑命令管道到 `tail`/`head`，不得自行拼接 `nohup` / `Start-Process` / `start`。 |
+| **契约漂移** | 直接编辑 `source/需求.md` / `执行约束.md` 后契约会进入 `externally-modified`，此时 advance/dagRunTask 拒用新内容（DIRTY_SOURCE / hash 不一致）。合法收编路径：`task contract adopt --task <id> --expected-revision <n> --expected-observed-hash <hash> --request-id <id> --request-payload-sha256 <sha>`（hash 从 `task status --json` 的 `contract.observedCanonicalHash` 取）；或改 PRD 后重新 importPrd 走完整重投影。禁止手工改 `lifecycle.json` / `.transactions/` 绕过。 |
 
 **永远不要**：`loop-agent` / `agent-worker` 失败 ⇒ 主会话直接改仓库实现；也勿因 provider 抖动或只读节点失败**新建无关 task-id**。
 
@@ -109,6 +110,8 @@ loop-agent task status <task-id> --json
 # 有 plan 时收尾：loop-agent plan complete <plan-id> --summary "..."
 ```
 
+长跑 DAG 采用 `prepareDagExecution → runDag → operationId` 异步 operation 路径（`--approve-gate` 不在前台 Bash 中同步等待）；监督退避 60 → 90 → 120 → 180 秒，状态变化后重置为 60 秒，疑似 stall 用 30–60 秒复查。
+
 `--verify` 命令应取项目 `AGENTS.md` / `__LOOP_AGENT_GOVERNANCE_ROOT__/verification-matrix.md` 登记的验证命令（不要假定 `npm run typecheck` 存在）；`--verify` 可选，省略时自动从 package.json scripts 或既有 managed `task.json.verifyCommands` 推导建议。
 
 `source/需求.md` 与 `source/执行约束.md` 仍必需（M8/M9），但默认由 `task advance --prd` 投影生成，而不是主会话手写。`plan create` 不是 `task advance` 的硬依赖。写入前同步 `task.json.allowedPaths` / `task.json.forbiddenPaths` 并审查 writer `writeSet`。高级任意 DagSpec 才用 `dag validate|execute|report`，不进入标准 happy path。
@@ -139,7 +142,7 @@ live run 先用 `loop-agent dag status --run-id <run-id>` 看 lifecycle 与 live
 
 **失败默认恢复序**：`dag report`/`dag doctor` → 优先 `dag rerun --from-node <node> --plan` 再带 `--plan-hash`（provider 抖动、plan/review/verify 安全下游；writer/decision/fingerprint 不合格勿硬跑）→ paused：approve→`dag resume` → 契约/源真变或 R1 不合格：同 task `task advance` / `dag rerun-task`（禁无理由新建 task-id）→ Worker-owned：`agent-worker task retry`。
 
-Operator 须监控 live run 至终态（FINISHED / FAILED / partial_failed）或 Decision Gate 需要 approve；可在节点/rank 变化、verify/closeout、stall 或需 approve 时简短汇报（告知非请求确认）。判活须组合 runner heartbeat、session events 与 `dag doctor` liveness/provider meaningful progress（heartbeat alone ≠ progress）。bounded writer 运行期间不得并发修改工作区（write guard / write-guard 会把越界 diff 错误归因到 writer）；只读 status/doctor/report 与 approve/reject/resume 仍允许。恢复：report/doctor → **优先** `dag rerun --from-node` → 必要时 rerun-task/同 task advance → shell verify。**禁止**把主会话直接 Edit 业务代码当作恢复手段。
+Operator 须监控 live run 至终态（FINISHED / FAILED / partial_failed）或 Decision Gate 需要 approve；监督退避 60 → 90 → 120 → 180 秒（状态变化后重置 60 秒，疑似 stall 用 30–60 秒复查），长跑 DAG 由 Console operation 后台持有（`prepareDagExecution → runDag → operationId`），禁止前台 Bash 直接 `task advance --approve-gate`、禁止管道 `tail`/`head`、禁止自行拼接 `nohup` / `Start-Process` / `start`；可在节点/rank 变化、verify/closeout、stall 或需 approve 时简短汇报（告知非请求确认）。判活须组合 runner heartbeat、session events 与 `dag doctor` liveness/provider meaningful progress（heartbeat alone ≠ progress）。bounded writer 运行期间不得并发修改工作区（write guard / write-guard 会把越界 diff 错误归因到 writer）；只读 status/doctor/report 与 approve/reject/resume 仍允许。恢复：report/doctor → **优先** `dag rerun --from-node` → 必要时 rerun-task/同 task advance → shell verify。**禁止**把主会话直接 Edit 业务代码当作恢复手段。
 
 ### 运行态与验证
 

package/harness.json

@@ -79,9 +79,9 @@
   },
   "executors": {
     "pi": {
-      "description": "Pi 负责规划、评审、诊断；当 DAG toolProfile=write 时也可做有界写入。模型按复杂度三档配置，格式为 provider/model 字符串（例：wizard-local/grok-4.5）；斜杠前为 Pi provider，后为 modelId，勿只写裸 modelId。",
+      "description": "Pi 负责规划、评审、诊断；当 DAG toolProfile=write 时也可做有界写入。模型按复杂度三档配置，支持 provider/model 字符串或带 thinking 的对象；斜杠前为 Pi provider，后为 modelId，勿只写裸 modelId。",
       "LOW": "wizard-local/minimax-m3",
-      "MED": "wizard-local/glm-5.3",
+      "MED": "wizard-local/grok-4.6",
       "HIGH": "wizard-local/gpt-5.6-sol"
     }
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tea-agent/loop-agent",
-  "version": "0.35.4-beta.0",
+  "version": "0.36.0",
   "type": "module",
   "bin": {
     "loop-agent": "bin/loop-agent.js",

package/skills/loop-agent/SKILL.md

@@ -24,6 +24,7 @@ references:
 ## 初始化升级路由
 
 - 写入型「**loop-agent初始化更新**」「loop-agent 初始化更新」「更新 loop-agent 初始化内容」「升级后对齐」统一运行 `loop-agent init upgrade --repo-root . --json`；处理 controller 返回的单文件 `allowedPaths`，以同一 run `--continue` 推进至终态，不停在中间状态。
+- 升级可能因 gitignore 迁移评估暂停：若 `.agents/` 存在已跟踪内容，需先审查是否有应团队共享的自定义内容；存在 staged 变更或 git 查询失败时需先解决 blocker，再重跑评估。建议的 `git rm -r --cached --ignore-unmatch` 只修改 index、保留工作区文件，且由用户执行而非 controller。
 - 「检查初始化更新」「初始化更新校验」「只检查，不要修改」仅运行 `loop-agent init check-update --repo-root . --markdown`，不得创建 run 或写入。
 - facts 位于 `.harness/init-upgrades/<run-id>/`；禁止业务源码和宽泛 glob。默认管理项目 `.opencode/plugins/`、`.pi/extensions/`、`.pi/settings.json`；Pi trust 后加载，默认不读写 `~/.pi/agent/settings.json`。
 

package/skills/loop-agent/references/command-reference.md

@@ -152,6 +152,8 @@ fresh init 将精确引用写到 `executors.pi.LOW/MED/HIGH`。fresh `loop-agent
 
 `init upgrade` 是写入型升级入口：首次运行冻结 controller identity、检查 npm latest、扫描 surface、自动执行安全动作，并把 run facts 持久化到 `.harness/init-upgrades/<run-id>/`。`--status` 与 `--report` 严格只读；`--continue` 会重新核验 identity、当前 hash 与不变量。版本选择、无法判定冲突、凭据、破坏性删除或 active writer 才进入 human decision。普通语义冲突返回具体单文件 `allowedPaths` merge task，主会话必须合并后继续，不能把 `needs-model-merge` 作为完成。默认管理项目级 OpenCode/Pi recovery 与 `.pi/settings.json` nested merge，保留显式 disabled，Pi trust 后才加载；默认不读写 `~/.pi/agent/settings.json`。旧 `init reconcile` 保留为兼容低层诊断。
 
+升级在 managed block 收敛后生成只读 gitignore 迁移评估（`.harness/init-upgrades/<run-id>/gitignore-migration.json`）：inventory 已跟踪的 `.harness/**` 与 `.agents/**`，建议 index-only 命令 `git rm -r --cached --ignore-unmatch .harness` / `.agents`（只改 index、保留工作区文件；`scripts/**` 与 `ai_workspace/loop-agent/**` 永不出现）。`.agents` 已跟踪内容会触发 human-decision 暂停等待审查；非 Git 仓库、git 查询失败或存在 staged 条目时 `status: blocked` 且无命令建议。controller 不自动修改 Git index。
+
 ### 查看或复制内置示例
 
 ```bash