@tea-agent/loop-agent 0.40.0-next.9 → 0.41.0-next.1

package/AGENTS.md

@@ -24,7 +24,7 @@
 - 机器校验契约真源：`docs/init-surface.manifest.json`、`docs/architecture/runtime-boundaries.md`、`src/cli/command-definitions.ts`、`skills/loop-agent/`、`scripts/check-*.sh`；本文件只指路。
 - 本仓库既是源项目也是 init 默认模板；新增能力必须判断 npm 内置 vs `loop-agent init` 投影。
 - 委托写入前必须结构化 `task.json.allowedPaths` / `forbiddenPaths`，并审查 DAG writer `writeSet`。
-- Shell 搜索优先 `rg`，按名找文件优先 `fd`；脚本用 Git Bash / 兼容 Bash。
+- Shell 搜索优先 `rg`，按名找文件优先 `fd`；脚本用 Git Bash / 兼容 Bash。rg 短标志禁止粘写：rg 的 `-r` 是 `--replace`（grep 里才是递归，rg 递归默认开启），`rg -rln pat` 会把匹配文本静默替换成 `ln` 且无报错；行号搜索用 `rg -n pat`，仅列文件用 `rg -l pat`，输出呈 `文件名:内容`（单冒号无行号）即为替换污染，须重跑。
 - 用 loop-agent 迭代本仓库时，控制器必须来自已发布 npm 包（记录实际版本）；启动后不要中途升级；不要用工作区 `npm link` / `npm run dev` 控制可能改 CLI/runtime/package 的任务。
 - 反复出现的约束固化为文档、脚本、检查、测试或模板；禁止占位实现（除非 contract 标明脚手架）。
 - 非微小实现：`exec-plan` 不能替代 Agent DAG；除非用户要求 one-shot 或计划记录 escape hatch，否则改实现前完成 `task advance`（PRD + 结构化路径边界）与 writeSet gate 审查。长跑 DAG 启动一律走 `prepareDagExecution → runDag → operationId` 由 Console operation 后台持有；**禁止**在前台 Bash 直接执行 `task advance --approve-gate`（会阻塞会话且超时杀 runner）、禁止把长跑命令管道到 `tail`/`head`、禁止自行拼接 `nohup`/`Start-Process`/`start`。

package/CHANGELOG.md

@@ -2,36 +2,45 @@
 
 ## [Unreleased]
 
+- 对话中的长时间 DAG 执行现在会保留一条实时状态：持续显示总运行时间、当前阶段、完成步数和最近一次有效进展；重复等待不再堆叠，点击状态可以直接查看对应 DAG。模型仍会在阶段变化、重要发现、失败恢复和结束时，用自然语言说明刚完成什么、下一步做什么。
+
+## [0.40.0] - 2026-08-31
+
 ### 重点更新
 
-- backend-test 的 pytest repair 现在继续使用 run-owned plan，不再搜索测试目录 README；correspondence finding 会直接指向 docstring、参数 ID、符号、脚本或 payload 的修复位置，collection/scenario repair 与 generated-local Python 写边界保持一致，并在单次 Pi 会话达到 24 轮或 80K 最大 usage snapshot 前受控返回 partial failure。
-- Operator Chat 现在可以在同一个页面管理多个工作区；每个会话始终使用创建它的工作区，切换对话时地址栏保持简洁稳定。
-- 对话消息流整体调整为更紧凑、清晰的布局，思考过程、工具执行、阶段性结果和最终回答会按真实发生顺序展示。
-- 上下文过大时会自动压缩并继续原请求，消息流会依次展示触发原因、压缩进度和压缩结果；手动压缩也使用同一套展示方式。
-- 对话进行中可以连续提交多条消息，并支持编辑、删除或立即插入当前回复；刷新页面或重启 Console 后，未发送内容仍会保留。
+- **Console 对话页「轨迹」视图**：仿 DSH ui-trajectory 把对话按「用户 → 过程 → AI 回复」渲染为统一单元格列表，每行带类型标签与图标，思考/工具/超长消息可展开收起，支持一键导出自包含 HTML。轨迹顶部新增「会话注入」区，展示真实注入 Pi 会话的系统提示词全文（含字符数与注入上下文 suffix）与模型可见工具目录，离线导出同样包含；每轮新增遥测头行（步骤数、LLM 时长、首 token 延迟、解码吞吐、工具调用数与耗时、上下行 token 与成本），切换历史会话立即显示。同时修复刷新后工具耗时全部变成 0ms/消失（会话恢复现在回填每次工具调用的真实起止时间）、窄窗口耗时列被整体隐藏、展开后内容仍截断在 420 字符、多工作区下轨迹数据恒为空等问题。
+- **上下文压力处理默认开启**：Operator Chat 与 DAG Pi 节点在发送前接近模型窗口时主动压缩一次、识别网关「请求上下文过大」响应后压缩并继续原请求、对 OpenAI 格式的大工具输出做线路压缩；消息流按顺序展示触发原因、压缩进度与结果。Anthropic 缺省 prompt cache 形态保持 no-op 不破坏缓存；`HARNESS_CONTEXT_PRESSURE_MODE` / `HARNESS_CONSOLE_SIFT` / `HARNESS_DAG_SIFT` 可整体或分路关闭；可选依赖 `@agent-context/sift@0.0.1-alpha.6` 精确 pin，不投影 init。
+- **Console「Pi 插件」管理**：从「设置」提升为与「设置」同级的独立 tab，内置经 npm 校验的插件目录（6303 个，展示 5562 个可安装、含中文说明），支持搜索、每页条数、分页与一键安装/卸载；插件版本改为从 npm 配置的仓库地址拉取（内网镜像可用），安装时提供版本下拉、默认装最新，已安装插件显示实际版本，`@version` 形式的已装包也能正确识别；进度动画改为 JS 驱动，不受系统「减少动态效果」影响。
+- **Operator Chat 持续目标模式**：用 `/goal` 或输入框上方的目标条创建跨多回合推进的目标，可随时暂停、继续、编辑或清除；长任务在安全轮数内自动续行，优先处理排队消息；上下文压缩后当前目标自动回到模型上下文；刷新后仍可见，重启 Console 不会未经确认自行续跑。配套新增任务清单与结构化提问卡片，长任务进度与需要用户决策的内容更易理解。
+- **多工作区与消息流重构**：同一页面管理多个工作区，会话稳定归属创建它的工作区，会话列表改为独立全局目录（移除工作区不丢会话）；消息流整体更紧凑，思考、工具执行、阶段性结果与最终回答按真实发生顺序展示；对话进行中可连续提交多条消息（编辑/删除/立即插入），排队消息由输入框上方队列面板统一管理；右侧辅助栏统一承载文件、上下文、调用记录、运行活动与当前 DAG，宽度可调。
+- **CLI 可用性修复**：`--help` 现在直接显示真实 usage（此前 raw-args 命令只在解析出错时才抛出）：`task advance` / `task status` / `task source prepare` 与 `dag resume|doctor|report|rerun|...` 的帮助顶部即列出完整参数契约；`task source prepare`（确定性任务源投影）从不可达死代码转正为可用子命令。
+- **backend-test 执行与判定系列修复**：pytest 7.x 环境改用 run-owned Python launcher 精确执行已授权 items（不再把 `@<targets-file>` 当普通路径导致全部无法执行）；payload readiness 修复空值标记误判 UNSAFE、本地 transport payload 假阴性、同模块安全 payload helper 绑定、`\u0020` 空白表示误判、超长边界 observed 持久化等系列假阳/假阴性；Scenario Partition 的 exact slot 清单正确下发 Markdown writer（普通别名不再替代 partition slots）、补集槽位合同统一；payload/fixture/Markdown→pytest 多类缺陷合并进唯一一次有界 repair；文档审计不再扫描 `.scratch/` 临时证据阻断发布前验证。
+
+### 改进
+
+- 初始化选择模型时识别公司内部 `wizard-local` 的 AICoder 型号梯度（default/plus/pro/max），作为 LOW/MED/HIGH 参考但不机械分档。
+- 右侧文件工作台打开 HTML 默认沙箱预览（同项目相对资源可用、脚本与外部网络阻断），超大源码显示前 512 KiB 并明确提示截断。
+- 工作区列表支持拖动排序并自动保存；多个 Console 同时运行时安全合并工作区增删并实时同步。
+- 设置页用自然中文展示模型状态；DAG 检查器简化节点说明与提示词重启面板，长内容更易定位。
+- 输入框整合模型与推理等级选择，只展示有真实数据的轮次、耗时、速度与缓存统计。
+- 修复 Console「更新并初始化」被观测顶栏遮挡（入口改到操作 → 设置）、内嵌观测样式泄漏与「提示词重启」入口丢失、宽表格/窄屏多处显示异常、部分 Windows 偶发读取失败等。
+
+## [0.39.3] - 2026-08-28
 
 ### 新增
 
-- 右侧辅助栏统一承载文件、上下文、调用记录、运行活动和当前 DAG；点击消息流中的运行状态即可在同一侧栏查看运行、门禁与产物，并可自由调整侧栏宽度。
-- 新增任务清单与结构化提问卡片，让长任务的进度和需要用户选择的内容更容易理解。
-- 「对话 / 操作 / 观测」使用统一导航；操作和观测页面可以切换工作区查看对应数据。
+- 前端实现支持来源归属运行时注入，使用宽松校验，不符合格式也不会阻断现有流程，方便后续接入完整的来源追溯。
 
-### 改进
+## [0.39.2] - 2026-08-28
+
+### 重点更新
 
-- 后端测试的计划与生成阶段只读取任务明确绑定的需求、环境报告和本轮产物，减少无关仓库内容干扰结果，并让后续步骤继续使用同一份已确认计划。
-- Operator Chat 的任务清单工具描述补齐多步骤与 DAG 使用纪律：开始前按具体步骤规划、保持进行中约束、完成后立即更新完整列表，并用面向用户的阶段呈现 DAG 进度，简单任务自动跳过。
-- 正式发布会自动同步更新版本日志和文档站摘要；发布说明优先生成简洁自然的中文，生成服务不可用时也能安全完成发布。
-- 输入框整合模型与推理等级选择，并只展示有真实数据的轮次、耗时、速度和缓存统计。
-- 工作区、会话列表、文件标签和侧栏交互更加紧凑自然，折叠、切换和页面过渡更加流畅。
-- 设置页改用自然中文展示当前工作区的模型状态，减少内部状态和重复信息。
-- DAG 检查器简化节点说明、执行记录和提示词重启面板，长内容更容易阅读和定位。
+- DAG 检查器与提示词重启面板重做：节点备注列改为节点用途简介不再复读节点 ID，默认隐藏协议事件、需要时可一键展开；提示词重启只读面板撑满剩余高度，内容在卡片内滚动，避免长提示词被压缩或被全局滚动条逼出视区。
 
 ### 修复
 
-- 修复重命名会话时输入框没有回显当前标题的问题。
-- 修复停止回复后输入区恢复迟缓、排队消息偶发丢失，以及切换会话后旧状态残留的问题。
-- 修复宽表格、文件预览、滚动条和窄屏布局中的多处显示异常。
-- 修复部分 Windows 环境读取运行状态时的偶发失败，并提高前后端任务验证的可靠性。
+- 为 `views/dag.js` 补充缺失类型声明，修复 next 发布通道的 typecheck 阻断。
+- 记录 release/0.39 上 observe 面板改动的回移植入 0.39.2，方便追溯修复来源。
 
 ## [0.39.1] - 2026-08-25
 

package/README.md

@@ -54,7 +54,7 @@ loop-agent init doctor --repo-root .
 loop-agent inspect --repo-root .
 ```
 
-`init model-catalog` 会通过 Pi runtime 与 OpenCode 的非 verbose 列表发现本机可用模型，只输出白名单模型元数据、Pi setup 状态和 provider 凭据就绪状态，不输出原始配置或凭据。已有 LOW/MED/HIGH 只要仍存在于 Pi 就保持不变；只有已移除或缺失的档位才由模型结合项目、能力、成本和上下文窗口重选。如果 Pi 没有模型定义，初始化模型会先向用户补齐 provider/model、API/base URL、上下文窗口、最大输出等无法推断的信息并协助配置；如果只是凭据缺失，则保留 harness 选择并修复认证。初始化会保留已有用户内容，并补充 `AGENTS.md`、`harness.json`、`ai_workspace/loop-agent/`、`.agents/skills/`、`.harness/` 和验证脚本等治理入口。
+`init model-catalog` 会通过 Pi runtime 与 OpenCode 的非 verbose 列表发现本机可用模型，只输出白名单模型元数据、Pi setup 状态和 provider 凭据就绪状态，不输出原始配置或凭据。已有 LOW/MED/HIGH 只要仍存在于 Pi 就保持不变；只有已移除或缺失的档位才由模型结合项目、能力、成本和上下文窗口重选。对于 `wizard-local` provider，目录会识别 `aicoder-default < aicoder-plus < aicoder-pro < aicoder-max` 的家族内相对能力顺序，但不会把它机械绑定为固定三档。如果 Pi 没有模型定义，初始化模型会先向用户补齐 provider/model、API/base URL、上下文窗口、最大输出等无法推断的信息并协助配置；如果只是凭据缺失，则保留 harness 选择并修复认证。初始化会保留已有用户内容，并补充 `AGENTS.md`、`harness.json`、`ai_workspace/loop-agent/`、`.agents/skills/`、`.harness/` 和验证脚本等治理入口。
 
 控制器升级后，旧项目可能仍使用过期的 repo-local skills、模板、managed blocks 和治理说明。普通安全仓库命令结束时会自动检测目标项目是否需要对齐：非 TTY 只向 stderr 输出提示且不写入目标；TTY 且无 human decisions、无活跃 DAG/Worker 时，可经明确 `y/yes` 同意后应用 deterministic safe actions；也可手动运行统一入口：
 
@@ -126,7 +126,7 @@ agent-worker console doctor --repo .
 
 ```
 
-裸 `agent-worker console` 不在终端询问升级。Operate 右上角提供非阻塞的更新入口，默认选择 npm `latest` 稳定版，也可由操作者显式选择 `next` 预发布版；更新前会先检查活跃运行和全局安装来源。`console doctor`、帮助和版本命令均不触发更新检查。
+裸 `agent-worker console` 不在终端询问升级。操作 → 设置提供非阻塞的「更新并初始化」入口，默认选择 npm `latest` 稳定版，也可由操作者显式选择 `next` 预发布版；更新前会先检查活跃运行和全局安装来源。`console doctor`、帮助和版本命令均不触发更新检查。
 
 访问 Console <http://127.0.0.1:8790/>；**Operator Chat Official** 是 Console 内的主引导交互，默认打开「对话」工作区。构建时设置 `VITE_OPERATOR_CHAT_DEFAULT_LANDING=tasks` 可改回 Tasks；URL hash（`#/tasks`）与 localStorage 仍优先于默认值。顶部「检视」以及 `/inspect/` 提供只读 Dashboard、DAG、时间线与 artifacts；`loop-agent …` / `agent-worker …` CLI direct path 仍是受治理 fallback。当前状态为 **code ready, dogfood pending/unmet**，不宣称已替代 openCode 或其他外部主会话。
 

package/dist/application/dag/generate-task-dag.js

@@ -109,8 +109,12 @@ function findForbiddenOverlapsForPacket(task) {
     const overlaps = [];
     for (const writeSetEntry of task.writeSet ?? []) {
         for (const forbiddenPath of task.forbiddenPaths ?? []) {
-            if (pathMatchesPattern(writeSetEntry, forbiddenPath) ||
-                pathMatchesPattern(forbiddenPath, writeSetEntry)) {
+            // A narrower forbidden child may intentionally carve a protected
+            // file out of a broader node writeSet. The runtime write guard gives
+            // forbiddenPaths precedence; report a packet conflict only when the
+            // writeSet entry itself is inside the forbidden boundary. Console G2
+            // separately rechecks task-global forbidden paths with live scope.
+            if (pathMatchesPattern(writeSetEntry, forbiddenPath)) {
                 overlaps.push({ writeSetEntry, forbiddenPath });
             }
         }

package/dist/build-stamp.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": 1,
-  "version": "0.40.0-next.9",
-  "gitSha": "772282937d5cddad444b25c58265d2d368491287",
-  "builtAt": "2026-08-28T00:59:36.332Z"
+  "version": "0.41.0-next.1",
+  "gitSha": "082093199826919c9699608265a26e4840207421",
+  "builtAt": "2026-08-31T02:28:34.598Z"
 }

package/dist/cli/command-definitions.js

@@ -11,6 +11,7 @@ import { runInspectNext } from "../commands/inspect-next.js";
 import { runInstructions } from "../commands/instructions.js";
 import { runTaskAdvance } from "../commands/task-advance.js";
 import { runTaskStatus } from "../commands/task-status.js";
+import { runTaskSourcePrepare } from "../commands/task-source-prepare.js";
 import { runOperator } from "../commands/operator.js";
 import { runPlanList } from "../commands/plan-list.js";
 import { runPlanCheck, runPlanComplete, runPlanCreate, } from "../commands/plan.js";
@@ -100,7 +101,6 @@ const EVAL_SUBCOMMANDS = [
     "outer-loop",
     "ignition",
 ];
-const CLOSEOUT_SUBCOMMANDS = ["task"];
 const PLAN_SUBCOMMANDS = ["list", "create", "complete", "check"];
 const SPINE_SUBCOMMANDS = ["audit"];
 const HANDOFF_SUBCOMMANDS = ["check", "coverage"];
@@ -139,7 +139,7 @@ const DAG_SUBCOMMANDS = [
     "final-verification",
     "decision",
 ];
-const TASK_SUBCOMMANDS = ["advance", "status"];
+const TASK_SUBCOMMANDS = ["advance", "status", "source"];
 const OPERATOR_SUBCOMMANDS = ["capabilities"];
 const WORKFLOW_SUBCOMMANDS = [
     "list",
@@ -215,7 +215,7 @@ export const COMMAND_DEFINITIONS = [
         name: "init",
         adapter: "none",
         tier: "primary",
-        intent: "Initialize a target repository with loop-agent harness capabilities.",
+        intent: "Initialize a target repository with loop-agent harness capabilities. Run INSIDE the target repo, or pass the global --repo-root <path> BEFORE this subcommand (e.g. `loop-agent --repo-root . init --profile full --merge`).",
         usage: "init [instructions|model-catalog|doctor|check-update|update|reconcile|upgrade] [--profile full|minimal] [--merge] [--model provider/model | --low-model provider/model --med-model provider/model --high-model provider/model] [--json|--markdown] [--bootstrap-surface|--apply-safe] [--client-recovery=auto|project|user|off] [--run-id <id> --status|--continue|--report]",
         subcommands: [...INIT_SUBCOMMANDS],
         handler: async ({ repoRoot, subcommand, rest }) => {
@@ -248,8 +248,8 @@ export const COMMAND_DEFINITIONS = [
         name: "task",
         adapter: "required",
         tier: "primary",
-        intent: "Task lifecycle: advance (mutation) and status (read).",
-        usage: "task <advance|status> ...",
+        intent: "Task lifecycle: advance (mutation), status (read), source prepare (deterministic source files).",
+        usage: "task <advance|status|source> ...",
         subcommands: [...TASK_SUBCOMMANDS],
         handler: async ({ repoRoot, subcommand, rest }) => {
             if (subcommand === "advance") {
@@ -260,6 +260,10 @@ export const COMMAND_DEFINITIONS = [
                 await runTaskStatus(repoRoot, rest);
                 return;
             }
+            if (subcommand === "source") {
+                await runTaskSourcePrepare(repoRoot, rest);
+                return;
+            }
             throw new Error(formatSubcommandUsageError("task", TASK_SUBCOMMANDS));
         },
     },

package/dist/cli/program.js

@@ -38,8 +38,9 @@ import { runInstructions } from "../commands/instructions.js";
 import { runKnowledge } from "../commands/knowledge.js";
 import { runLoop } from "../commands/loop.js";
 import { runLoopBenchmark } from "../commands/loop-benchmark.js";
-import { runTaskAdvance } from "../commands/task-advance.js";
-import { runTaskStatus } from "../commands/task-status.js";
+import { runTaskAdvance, TASK_ADVANCE_USAGE } from "../commands/task-advance.js";
+import { runTaskStatus, TASK_STATUS_USAGE } from "../commands/task-status.js";
+import { runTaskSourcePrepare, TASK_SOURCE_PREPARE_USAGE, } from "../commands/task-source-prepare.js";
 import { runOperator } from "../commands/operator.js";
 import { runPiReuseBenchmark } from "../commands/pi-reuse-benchmark.js";
 import { parsePiPromptArgs, printPiPromptUsage, runPiPrompt, } from "../commands/pi-prompt.js";
@@ -124,7 +125,11 @@ async function runTaskCommand(repoRoot, subcommand, rest) {
         await runTaskStatus(repoRoot, rest);
         return;
     }
-    throw new Error("usage: task <advance|status> ...");
+    if (subcommand === "source") {
+        await runTaskSourcePrepare(repoRoot, rest);
+        return;
+    }
+    throw new Error("usage: task <advance|status|source> ...");
 }
 async function runCommanderAction(ctx, command, subcommand, rest) {
     switch (command) {
@@ -402,9 +407,38 @@ function createRawArgsCommand(name, description, defaultRepoRoot) {
 function addRawSubcommands(parent, subcommands, defaultRepoRoot) {
     for (const subcommand of subcommands) {
         const command = createRawArgsCommand(subcommand, `${parent.name()} ${subcommand}`, defaultRepoRoot);
+        applyRawSubcommandUsage(command, `${parent.name()} ${subcommand}`);
         parent.addCommand(command);
     }
 }
+/**
+ * Well-known raw-args subcommands whose real USAGE would otherwise only appear
+ * after a parse error. `--help` shows this text at the top (beforeAll help text)
+ * so one-shot callers (Humans, Pi, Chat) can read the real contract up front.
+ * Keys are the full chain (`<parent> <subcommand>`); values must mirror the
+ * command module's own usage strings verbatim where one exists.
+ */
+const RAW_SUBCOMMAND_USAGE = {
+    "task advance": TASK_ADVANCE_USAGE,
+    "task status": TASK_STATUS_USAGE,
+    "task source prepare": TASK_SOURCE_PREPARE_USAGE,
+    "task source": "usage: task source prepare <task-id> [--from-text|--from-file|--from-stdin|--from-draft] [--dry-run|--apply]\nDeterministic task source projection (managed 需求.md / 执行约束.md). See `task source prepare --help` for the full flag surface.",
+    "dag resume": "usage: dag resume --run-id <id>",
+    "dag status": "usage: dag status --run-id <id>",
+    "dag approve": 'usage: dag approve --run-id <id> --option <A|B|...> [--notes "..."]',
+    "dag reject": 'usage: dag reject --run-id <id> --reason "..."',
+    "dag rerun": "usage: dag rerun --run-id <id> --from-node <node-id> --plan [--json] | dag rerun --run-id <id> --from-node <node-id> --plan-hash <hash> --request-id <key> --reason <text> [--json]",
+    "dag rerun-task": "usage: dag rerun-task --run-id <id> --reason <text> --request-id <key> [--profile auto|minimal|standard|reviewed|supervised] [--task-id <id>] [--json]",
+    "dag reconcile-run": "usage: dag reconcile-run --run-id <id> [--action supersede|abandon] [--reason <text>]",
+    "dag doctor": "usage: dag doctor [--run-id <id>]\nInspect health issues before approve/reject/resume; no --json (output is already structured JSON).",
+    "dag report": "usage: dag report [--run-id <id>] [--lifecycle active|paused|completed|all] [--json|--markdown] [--failed-only] [--latest] [--paused-latest]\nRead-only run report; --json gives structured facts (canonical: `dag report --run-id <id> --lifecycle completed --json`).",
+};
+function applyRawSubcommandUsage(command, chainKey) {
+    const usage = RAW_SUBCOMMAND_USAGE[chainKey];
+    if (!usage)
+        return;
+    command.addHelpText("beforeAll", usage);
+}
 function addStandaloneSubcommands(parent, subcommands, handler) {
     for (const subcommand of subcommands) {
         parent
@@ -640,6 +674,7 @@ export function buildLoopAgentProgram(options) {
                 const prepare = createRawArgsCommand("prepare", "Prepare deterministic task source files", options.defaultRepoRoot);
                 configureTaskSourcePrepare(prepare);
                 replaceActionWithOptionAwareDispatch(prepare, options.defaultRepoRoot);
+                applyRawSubcommandUsage(prepare, "task source prepare");
                 source.addCommand(prepare);
             }
         }

package/dist/commands/client-recovery.js

@@ -19,36 +19,11 @@ export const MAX_SESSION_RETRIES = 5;
 /** Hard cap for OpenCode overflow compact+resume recoveries (≤ 2). */
 export const MAX_OVERFLOW_RECOVERIES = 1;
 import { PI_RECOMMENDED_COMPACTION, PI_RECOMMENDED_RETRY, } from "../shared/pi-retry-settings.js";
+import { CONTEXT_OVERFLOW_PHRASES, buildGeneratedOverflowClassifierSource, isContextOverflow as classifySharedOverflow, } from "../shared/pi-context-pressure/overflow.js";
 export { PI_RECOMMENDED_COMPACTION, PI_RECOMMENDED_RETRY };
-/**
- * Shared overflow phrase sources used by isContextOverflow and generated plugins.
- * Keep English + Chinese intranet gateway messages in one place to avoid drift.
- */
-export const CONTEXT_OVERFLOW_PHRASES = [
-    "请求上下文过大",
-    "context_length_exceeded",
-    "context overflow",
-    "context length overflow",
-    "too many tokens",
-    "maximum context",
-    "prompt is too long",
-    "request_too_large",
-    "context window",
-];
-function escapeRegExpLiteral(value) {
-    return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
-}
-const CONTEXT_OVERFLOW_PATTERN = new RegExp(CONTEXT_OVERFLOW_PHRASES.map(escapeRegExpLiteral).join("|"), "i");
+export { CONTEXT_OVERFLOW_PHRASES };
 function buildGeneratedContextOverflowPatternSource() {
-    return `const CONTEXT_OVERFLOW_PHRASES = ${JSON.stringify([
-        ...CONTEXT_OVERFLOW_PHRASES,
-    ])};
-const OVERFLOW_PATTERN = new RegExp(
-  CONTEXT_OVERFLOW_PHRASES.map((phrase) =>
-    phrase.replace(/[.*+?^\${}()|[\\]\\\\]/g, "\\\\$&"),
-  ).join("|"),
-  "i",
-);`;
+    return buildGeneratedOverflowClassifierSource();
 }
 const PERMANENT_MESSAGE_PATTERNS = [
     {
@@ -65,7 +40,7 @@ const PERMANENT_MESSAGE_PATTERNS = [
     },
     {
         reason: "context-overflow",
-        pattern: CONTEXT_OVERFLOW_PATTERN,
+        pattern: /context_length_exceeded|请求上下文过大/i,
     },
     {
         reason: "cancelled",
@@ -180,6 +155,8 @@ function permanentReason(error) {
         return "api-error";
     }
     const text = collectErrorText(error);
+    if (isContextOverflow(error))
+        return "context-overflow";
     for (const entry of PERMANENT_MESSAGE_PATTERNS) {
         if (entry.pattern.test(text))
             return entry.reason;
@@ -192,7 +169,7 @@ function permanentReason(error) {
  * business validation, or standard 502/LLMRequestError as overflow.
  */
 export function isContextOverflow(error) {
-    return CONTEXT_OVERFLOW_PATTERN.test(collectErrorText(error));
+    return classifySharedOverflow(error);
 }
 function looksTransient(error) {
     const text = collectErrorText(error);
@@ -887,7 +864,7 @@ function textOf(error) {
 }
 
 function isContextOverflow(error) {
-  return OVERFLOW_PATTERN.test(textOf(error));
+  return isOverflowMessage(textOf(error));
 }
 
 function buildOverflowResumePrompt(sessionId, error, attempt) {
@@ -1163,11 +1140,6 @@ const EXTENSION_PATH = ${JSON.stringify(PI_CONTEXT_OVERFLOW_EXTENSION_PATH)};
 ${buildGeneratedContextOverflowPatternSource()}
 const NORMALIZED_PREFIX = "context_length_exceeded:";
 
-function isOverflowMessage(text) {
-  if (typeof text !== "string" || !text) return false;
-  return OVERFLOW_PATTERN.test(text);
-}
-
 export default function loopAgentContextOverflowExtension(pi) {
   void EXTENSION_PATH;
   pi.on("message_end", (event) => {

package/dist/commands/init-model-catalog.js

@@ -31,8 +31,28 @@ function sanitizeInputKinds(value) {
         .sort();
     return entries.length > 0 ? entries : undefined;
 }
+const AICODER_CAPABILITY_RANK = {
+    "aicoder-default": 1,
+    "aicoder-plus": 2,
+    "aicoder-pro": 3,
+    "aicoder-max": 4,
+};
+function relativeCapabilityForModelId(provider, modelId) {
+    if (provider !== "wizard-local")
+        return undefined;
+    const rank = AICODER_CAPABILITY_RANK[modelId];
+    if (!rank)
+        return undefined;
+    return {
+        family: "aicoder",
+        rank,
+        scale: 4,
+        order: "weakest-to-strongest",
+    };
+}
 function createCatalogEntry(input) {
     const piCompatible = input.source === "pi";
+    const relativeCapability = relativeCapabilityForModelId(input.provider, input.modelId);
     return {
         reference: `${input.provider}/${input.modelId}`,
         provider: input.provider,
@@ -42,6 +62,7 @@ function createCatalogEntry(input) {
         credentialReady: input.credentialReady,
         piCompatible,
         eligibleForHarness: piCompatible && input.credentialReady === true && input.tools !== false,
+        ...(relativeCapability ? { relativeCapability } : {}),
         ...(compactRecord({
             reasoning: input.reasoning,
             tools: input.tools,
@@ -448,6 +469,7 @@ export async function discoverInitModelCatalog(options) {
             rules: [
                 "Choose only models with eligibleForHarness=true; OpenCode-only entries are context, not proof of Pi executor compatibility.",
                 "Use target-project evidence plus model capabilities, cost, context window, and credential readiness; tiers may intentionally share a model.",
+                "When relativeCapability is present, compare only entries from the same family and treat larger rank values as stronger; it is a relative capability hint, not a mandatory LOW/MED/HIGH assignment.",
                 "Preserve an existing complete usable matrix unless it is unavailable or the user asked to reselect it.",
                 "Ask the user only for a material unresolved credential, cost, compliance, or availability decision.",
             ],

package/dist/commands/task-advance.js

@@ -13,7 +13,7 @@ const USAGE = `usage:
     [--feature-id <id>]
     [--allowed-path <glob>]...
     [--forbidden-path <glob>]...
-    [--verify <label:command>]...
+    [--verify <label:command>]...   # e.g. --verify "typecheck:npx tsc --noEmit" (label BEFORE the colon)
     [--verify-timeout <label:timeoutMs>]...
     [--backend-test <json>]
     [--accept-recommendations <digest>]
@@ -31,6 +31,7 @@ const USAGE = `usage:
     [--progress-interval-ms <ms>]
     [--dry-run]
     [--json]`;
+export { USAGE as TASK_ADVANCE_USAGE };
 function pushList(target, value) {
     const list = target ?? [];
     list.push(value);
@@ -232,6 +233,26 @@ export function parseTaskAdvanceArgs(args) {
     }
     return { taskId, options };
 }
+const FOREGROUND_DAG_WARNING = [
+    "[task advance] --approve-gate will execute the whole DAG in the FOREGROUND:",
+    "  this terminal must stay open until every node finishes (writer nodes can run 10-60+ min).",
+    "  Killing it mid-run (timeout / Ctrl-C / closed terminal) leaves an orphaned run that",
+    "  only `dag reconcile-run --action supersede|abandon` can close.",
+    "  Prefer background holders instead:",
+    "    - Console: prepareDagExecution -> runDag (operation-held)",
+    "    - CLI:     nohup loop-agent task advance <task-id> --approve-gate <token> ... &",
+].join("\n");
+/**
+ * Operator guardrail (dogfood 2026-08-29): warn a human sitting on a TTY that
+ * `--approve-gate` runs the full DAG in the foreground before the process
+ * parks on node execution. Piped/non-TTY invocations (Console operations,
+ * scripts, --json pipelines) stay silent so machine paths are unaffected.
+ */
+export function warnForegroundDagExecution(options, stream = process.stderr) {
+    if (options.quiet || stream.isTTY !== true)
+        return;
+    stream.write(`${FOREGROUND_DAG_WARNING}\n`);
+}
 function mapOutcome(result) {
     if (result.blockers.length > 0) {
         const code = result.blockers[0]?.code ?? "";
@@ -351,6 +372,7 @@ export async function runTaskAdvance(repoRoot, args) {
         // ONLY for the approve-gate execution path and disposed on every exit;
         // its timer starts only after onRunStart (double guard, no stray timer).
         if (options.approveGate && !options.dryRun && !options.quiet) {
+            warnForegroundDagExecution(options);
             progress = createRunDagProgressObserver({
                 intervalMs: options.progressIntervalMs,
             });

package/dist/commands/task-source-prepare.js

@@ -32,6 +32,7 @@ const USAGE = `usage:
     [--dry-run | --apply]
     [--force-overwrite-source]
     [--json]`;
+export { USAGE as TASK_SOURCE_PREPARE_USAGE };
 function pushList(target, value) {
     const list = target ?? [];
     list.push(value);
@@ -386,6 +387,10 @@ function printHuman(result) {
     console.log(lines.join("\n"));
 }
 export async function runTaskSourcePrepare(repoRoot, args) {
+    // CLI chain routing (task -> source -> prepare) folds the leaf name into
+    // rest[0]; strip it so `loop-agent task source prepare <id> ...` works.
+    if (args[0] === "prepare")
+        args = args.slice(1);
     if (args.includes("--help") || args.includes("-h")) {
         console.log(USAGE);
         process.exitCode = 0;

package/dist/commands/task-status.js

@@ -3,6 +3,7 @@ import { buildOperatorResult, operatorFailed, processExitCodeForOutcome, writeOp
 const COMMAND = "task status";
 const USAGE = `usage:
   task status <task-id> [--json|--markdown] [--details]`;
+export { USAGE as TASK_STATUS_USAGE };
 export function parseTaskStatusArgs(args) {
     const options = {};
     const positional = [];

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

@@ -1,9 +1,12 @@
 import { appendFile, mkdir } from "node:fs/promises";
+import os from "node:os";
 import path from "node:path";
+import { randomUUID } from "node:crypto";
 import { computeExtensionToolAllowlistExtras } from "./pi-extension-resolver.js";
 import { BoundedTextPreview, classifyPiFailure, createPiJsonlStreamCollector, DEFAULT_ABORT_GRACE_MS, DEFAULT_STALL_TIMEOUT_MS, DEFAULT_TIMEOUT_MS, extractPiEventStopReason, extractAssistantTextFromPiJson, } from "./pi-executor.js";
 import { serializeSessionEvent } from "./pi-event-serializer.js";
 import { PI_RECOMMENDED_COMPACTION, PI_RECOMMENDED_RETRY, } from "../shared/pi-retry-settings.js";
+import { createCompactionController, createContextPressureExtension, createSiftBridgeState, inMemoryCompactionSettings, isDagSiftEnabled, parseCompactionPolicy, parseSiftBridgeConfig, SIFT_RETRIEVE_TOOL_NAME, } from "../shared/pi-context-pressure/index.js";
 /** Same-session recovery after a transport stall. Exactly one continue per SDK attempt. */
 export const PI_SDK_STALL_CONTINUE_MAX = 1;
 /**
@@ -215,15 +218,37 @@ async function createSdkSession(sdk, input, shared) {
     // Inject rate-limit-tolerant auto-retry settings for the session lifecycle.
     // Prefer SettingsManager.inMemory so harness runs do not depend on ~/.pi or
     // project .pi/settings.json being present/trusted.
+    const policy = parseCompactionPolicy();
+    const compaction = {
+        ...PI_RECOMMENDED_COMPACTION,
+        ...inMemoryCompactionSettings(policy),
+    };
     const settingsManager = typeof SettingsManager?.inMemory === "function"
         ? SettingsManager.inMemory({
             retry: {
                 ...PI_RECOMMENDED_RETRY,
                 provider: { ...PI_RECOMMENDED_RETRY.provider },
             },
-            compaction: { ...PI_RECOMMENDED_COMPACTION },
+            compaction,
         }, { projectTrusted: true })
         : undefined;
+    const nodeSessionId = `dag-${randomUUID()}`;
+    const controller = createCompactionController();
+    const siftEnabled = isDagSiftEnabled();
+    const sift = siftEnabled.enabled
+        ? createSiftBridgeState(parseSiftBridgeConfig({
+            sessionId: nodeSessionId,
+            defaultStashRoot: path.join(os.tmpdir(), "loop-agent-dag-sift"),
+        }))
+        : undefined;
+    if (sift)
+        sift.counters = controller.counters;
+    const pressureExtension = createContextPressureExtension({
+        name: "dag-pi-context-pressure",
+        policy,
+        controller,
+        sift,
+    });
     const loader = new DefaultResourceLoader({
         cwd: input.cwd,
         agentDir,
@@ -233,6 +258,7 @@ async function createSdkSession(sdk, input, shared) {
         ...(input.additionalExtensionPaths && input.additionalExtensionPaths.length > 0
             ? { additionalExtensionPaths: input.additionalExtensionPaths }
             : {}),
+        ...(pressureExtension ? { extensionFactories: [pressureExtension] } : {}),
         ...(settingsManager ? { settingsManager } : {}),
         appendSystemPromptOverride: (base) => [
             ...base,
@@ -247,10 +273,12 @@ async function createSdkSession(sdk, input, shared) {
     // extras into the create-time allowlist; protected builtins, the deny
     // surface, and write-capable extension tools never join.
     const extensionToolExtras = [];
-    if (input.additionalExtensionPaths &&
-        input.additionalExtensionPaths.length > 0) {
-        const registered = collectLoaderExtensionToolNames(loader);
-        extensionToolExtras.push(...computeExtensionToolAllowlistExtras(registered.map((name) => ({ name })), input.toolNames));
+    const registered = collectLoaderExtensionToolNames(loader);
+    extensionToolExtras.push(...computeExtensionToolAllowlistExtras(registered.map((name) => ({ name })), input.toolNames));
+    if (siftEnabled.enabled &&
+        !extensionToolExtras.includes(SIFT_RETRIEVE_TOOL_NAME) &&
+        !input.toolNames.includes(SIFT_RETRIEVE_TOOL_NAME)) {
+        extensionToolExtras.push(...computeExtensionToolAllowlistExtras([{ name: SIFT_RETRIEVE_TOOL_NAME }], input.toolNames));
     }
     const sessionToolNames = [...input.toolNames, ...extensionToolExtras];
     if (input.requireWriterCustomTools) {