skyloom 1.16.1 → 1.17.0

package/dist/tools/spawn.d.ts

@@ -0,0 +1,23 @@
+/**
+ * spawn_agent tool — launch a general-purpose, isolated-context subagent to
+ * handle one focused, self-contained task and return only its final report.
+ *
+ * This is sky's analogue of Claude Code's Task tool. Subagent types are
+ * discovered from built-ins plus `.claude/agents` / `.sky/agents` definition
+ * files (see core/subagent). Subagents never receive spawn_agent themselves,
+ * so there is no recursive fan-out.
+ */
+import type { ToolDefinition } from '../core/tool';
+import type { ToolRegistry } from '../core/tool';
+import type { SkillRegistry } from '../core/skill';
+import type { LLMClient } from '../core/llm';
+import type { MessageBus } from '../core/bus';
+export declare function createSpawnAgentTool(opts: {
+    config: any;
+    llm: LLMClient;
+    bus: MessageBus;
+    baseToolRegistry: ToolRegistry;
+    baseSkillRegistry: SkillRegistry;
+    cwd?: string;
+}): ToolDefinition;
+//# sourceMappingURL=spawn.d.ts.map

package/dist/tools/spawn.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"spawn.d.ts","sourceRoot":"","sources":["../../src/tools/spawn.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,cAAc,CAAC;AACnD,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,cAAc,CAAC;AACjD,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AACnD,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,aAAa,CAAC;AAC7C,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAG9C,wBAAgB,oBAAoB,CAAC,IAAI,EAAE;IACzC,MAAM,EAAE,GAAG,CAAC;IACZ,GAAG,EAAE,SAAS,CAAC;IACf,GAAG,EAAE,UAAU,CAAC;IAChB,gBAAgB,EAAE,YAAY,CAAC;IAC/B,iBAAiB,EAAE,aAAa,CAAC;IACjC,GAAG,CAAC,EAAE,MAAM,CAAC;CACd,GAAG,cAAc,CAmEjB"}

package/dist/tools/spawn.js

@@ -0,0 +1,77 @@
+"use strict";
+/**
+ * spawn_agent tool — launch a general-purpose, isolated-context subagent to
+ * handle one focused, self-contained task and return only its final report.
+ *
+ * This is sky's analogue of Claude Code's Task tool. Subagent types are
+ * discovered from built-ins plus `.claude/agents` / `.sky/agents` definition
+ * files (see core/subagent). Subagents never receive spawn_agent themselves,
+ * so there is no recursive fan-out.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.createSpawnAgentTool = createSpawnAgentTool;
+const subagent_1 = require("../core/subagent");
+function createSpawnAgentTool(opts) {
+    const cwd = opts.cwd || process.cwd();
+    const buildDescription = () => {
+        const defs = (0, subagent_1.loadSubagentDefinitions)(cwd);
+        const lines = [...defs.values()].map((d) => `  - ${d.name}: ${d.description}`);
+        return ('派生一个隔离上下文的子智能体来独立完成一个聚焦、自洽的任务,只返回它的最终报告。' +
+            '当任务需要大量搜索/调研、或你想把一段独立工作从主上下文里隔离出去时使用。' +
+            '子智能体看不到你的对话历史,所以 task 必须自带全部所需上下文(目标、相关文件、约束)。' +
+            '它无法反问你,会一次性完成并汇报。可并行派生多个互不依赖的子智能体。\n\n可用子智能体类型:\n' +
+            lines.join('\n'));
+    };
+    return {
+        name: 'spawn_agent',
+        description: buildDescription(),
+        parameters: [
+            {
+                name: 'agent_type',
+                type: 'string',
+                description: '子智能体类型(见工具描述中的可用类型,如 general-purpose / explore 或自定义)',
+                required: true,
+            },
+            {
+                name: 'task',
+                type: 'string',
+                description: '完整、自洽的任务描述。子智能体看不到对话历史,必须在此写清目标、相关文件路径与约束。',
+                required: true,
+            },
+            {
+                name: 'description',
+                type: 'string',
+                description: '可选:对该任务的简短(3-5 词)标签,用于展示。',
+                required: false,
+            },
+        ],
+        // Long-running by nature (full nested agent loop); give it generous headroom.
+        timeout: 600000,
+        handler: async (params) => {
+            const agentType = String(params.agent_type || '').trim();
+            const task = String(params.task || '').trim();
+            if (!agentType)
+                return '[spawn_agent error] agent_type is required.';
+            if (!task)
+                return '[spawn_agent error] task is required.';
+            const defs = (0, subagent_1.loadSubagentDefinitions)(cwd);
+            const def = defs.get(agentType);
+            if (!def) {
+                const available = [...defs.keys()].join(', ');
+                return `[spawn_agent error] unknown agent_type '${agentType}'. Available: ${available}`;
+            }
+            const report = await (0, subagent_1.runSubagent)({
+                def,
+                task,
+                config: opts.config,
+                llm: opts.llm,
+                bus: opts.bus,
+                baseToolRegistry: opts.baseToolRegistry,
+                baseSkillRegistry: opts.baseSkillRegistry,
+            });
+            const header = `[subagent ${def.name} 完成]`;
+            return `${header}\n${report}`;
+        },
+    };
+}
+//# sourceMappingURL=spawn.js.map

package/dist/tools/spawn.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"spawn.js","sourceRoot":"","sources":["../../src/tools/spawn.ts"],"names":[],"mappings":";AAAA;;;;;;;;GAQG;;AASH,oDA0EC;AA5ED,+CAAwE;AAExE,SAAgB,oBAAoB,CAAC,IAOpC;IACC,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,IAAI,OAAO,CAAC,GAAG,EAAE,CAAC;IAEtC,MAAM,gBAAgB,GAAG,GAAW,EAAE;QACpC,MAAM,IAAI,GAAG,IAAA,kCAAuB,EAAC,GAAG,CAAC,CAAC;QAC1C,MAAM,KAAK,GAAG,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC;QAC/E,OAAO,CACL,0CAA0C;YAC1C,uCAAuC;YACvC,gDAAgD;YAChD,mDAAmD;YACnD,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CACjB,CAAC;IACJ,CAAC,CAAC;IAEF,OAAO;QACL,IAAI,EAAE,aAAa;QACnB,WAAW,EAAE,gBAAgB,EAAE;QAC/B,UAAU,EAAE;YACV;gBACE,IAAI,EAAE,YAAY;gBAClB,IAAI,EAAE,QAAQ;gBACd,WAAW,EAAE,sDAAsD;gBACnE,QAAQ,EAAE,IAAI;aACf;YACD;gBACE,IAAI,EAAE,MAAM;gBACZ,IAAI,EAAE,QAAQ;gBACd,WAAW,EAAE,4CAA4C;gBACzD,QAAQ,EAAE,IAAI;aACf;YACD;gBACE,IAAI,EAAE,aAAa;gBACnB,IAAI,EAAE,QAAQ;gBACd,WAAW,EAAE,2BAA2B;gBACxC,QAAQ,EAAE,KAAK;aAChB;SACF;QACD,8EAA8E;QAC9E,OAAO,EAAE,MAAM;QACf,OAAO,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE;YACxB,MAAM,SAAS,GAAG,MAAM,CAAC,MAAM,CAAC,UAAU,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;YACzD,MAAM,IAAI,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;YAC9C,IAAI,CAAC,SAAS;gBAAE,OAAO,6CAA6C,CAAC;YACrE,IAAI,CAAC,IAAI;gBAAE,OAAO,uCAAuC,CAAC;YAE1D,MAAM,IAAI,GAAG,IAAA,kCAAuB,EAAC,GAAG,CAAC,CAAC;YAC1C,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;YAChC,IAAI,CAAC,GAAG,EAAE,CAAC;gBACT,MAAM,SAAS,GAAG,CAAC,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAC9C,OAAO,2CAA2C,SAAS,iBAAiB,SAAS,EAAE,CAAC;YAC1F,CAAC;YAED,MAAM,MAAM,GAAG,MAAM,IAAA,sBAAW,EAAC;gBAC/B,GAAG;gBACH,IAAI;gBACJ,MAAM,EAAE,IAAI,CAAC,MAAM;gBACnB,GAAG,EAAE,IAAI,CAAC,GAAG;gBACb,GAAG,EAAE,IAAI,CAAC,GAAG;gBACb,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;gBACvC,iBAAiB,EAAE,IAAI,CAAC,iBAAiB;aAC1C,CAAC,CAAC;YAEH,MAAM,MAAM,GAAG,aAAa,GAAG,CAAC,IAAI,MAAM,CAAC;YAC3C,OAAO,GAAG,MAAM,KAAK,MAAM,EAAE,CAAC;QAChC,CAAC;KACF,CAAC;AACJ,CAAC"}

package/docs/OPTIMIZATION_PLAN.md

@@ -120,13 +120,13 @@
 
 - [x] **P4.1 Session 恢复 + 持久化修复**：新增 CLI `/sessions`（编号列表、标记当前）、`/resume <序号|id>`、`/new`。**并修复了一个严重 bug**：`persistDb()`（唯一写盘的代码）**从未被调用** → sql.js 纯内存 → 会话/长期记忆/工作记忆**重启全丢**（"启动自动恢复最新会话"形同虚设）。改为所有写经 `dbRun` 触发**防抖落盘** + `close()` 同步保存。新增跨实例持久化回归测试。
 - [x] **P4.2 自动压缩（catalog 感知触发）**：`shouldAutoCompact()`/`contextUsage()` 不再硬编码 128K —— 改为按当前模型在 catalog 里的真实 `context` 窗口判断（留 20% 余量给回复）。修复了小窗口模型（deepseek-reasoner 64K、mixtral 32K）长聊时**先于压缩就溢出**的隐患；状态栏 % 与模型名也正确了。压缩本身（摘要 + 保留近 N 条 + 指令保真）已存在并已接线。后续：结构化 checkpoint + 溢出后重试（对标 opencode）。
-- [ ] **P4.3 上下文快照**：环境信息、日期、工作区、激活技能合成一份"系统上下文"，模型可见但与历史分离（轻量版 Context Epoch）。
+- [x] **P4.3 环境上下文快照**（对标 Claude Code `<env>` 块）：[`envcontext.ts`](../src/core/envcontext.ts) 的 `buildEnvBlock` 合成「运行环境」块(工作目录、平台、Node、git 仓库/分支、日期),`gitInfo` 用纯文件读取识别仓库与分支(兼容 worktree 的 `.git` 文件)。`BaseAgent.injectEnvironment` 在 init/reinitLanguage 把它注入系统提示,与对话历史分离。新增 [`tests/envcontext.test.ts`](../tests/envcontext.test.ts) 6 用例。
 
 ### Phase 5 — 工具与插件健壮性｜~1 天
 
-- [ ] **P5.1 工具 IO 校验**：工具定义带 input/output schema，非法输入不执行、非法输出不算成功（对标 opencode `tools.md`）。
-- [ ] **P5.2 权限内聚**：危险操作的 `permission.assert` 由工具自身发起，security 模块只评估策略 + 管理审批，统一 `ask/allow/deny` 语义。
-- [ ] **P5.3 插件 hook**：目录加载器升级为有序 hook（`init` / `tool.register` / `provider.update`），插件在自己的 scope 注册，卸载即移除。
+- [x] **P5.1 工具输入校验 + 强制 coercion**（对标 opencode `tools.md`）：`ToolRegistry.validateAndCoerce` 在 `execute` 里于缓存/执行**之前**校验并归一化入参 —— 必填项缺失即拒、按声明类型 coerce（`"5"`→5、`"true"`→true、JSON 串→array/object）、enum 成员校验，handler 收到的是干净的类型化值，非法输入返回**可操作的**错误供模型重试。修了 `parseInt` 截断浮点(`"3.5"`→3)的隐患。见 [`tests/tool.test.ts`](../tests/tool.test.ts)（+7 用例）。输出校验暂未做。
+- [x] **P5.2 权限模式内聚**（对标 Claude Code 权限模式）：把审批逻辑收敛为一个纯函数 `decideApproval(level, mode, tool) → allow/ask/deny`（[security.ts](../src/core/security.ts)），`checkApproval` 只做红线门禁 + 调用它 + 回调。新增模式 `acceptEdits`（自动放行文件编辑类工具,其余照 default 询问）与 `bypass`（除红线外全放行）;保留 `auto/interactive/strict` 行为不变。`/perm <default|auto|accept|strict|bypass>` 运行时切换(linear + loom),`config.cli.approvalMode` 启动时生效。新增 [`tests/security.test.ts`](../tests/security.test.ts) 11 用例覆盖决策矩阵。后续:让工具自身发起 `permission.assert`(细粒度 scope)。
+- [x] **P5.3 插件 hook 生命周期**（对标 opencode 插件）：[`plugins/loader.ts`](../src/plugins/loader.ts) 从扁平加载器升级为有序 hook 系统。插件导出 `activate(ctx)`,通过 `ctx.registerTool` / `ctx.on(hook, fn)` 在**自己的 scope** 注册;`unload(name)` 精确移除其工具与 hook handler(并调 `deactivate`)。核心 hook:`init`(加载后、agent 起来前由 `SystemContext.initAll` 触发)、`tool.register`、`provider.update`。handler 按注册顺序触发、单个抛错隔离。**保留 legacy `register(registry)`**(diff 注册表追踪其工具以支持卸载)。新增 [`tests/plugins.test.ts`](../tests/plugins.test.ts) 7 用例。后续:`provider.update` 在 model_config 变更处触发。
 
 ### Phase 6 — 测试与质量门禁｜~1.5 天
 
@@ -176,3 +176,20 @@ M5 打磨成品     Phase 7        → 美学工程化 + 品牌资产 + 发布
 | 流式与工具执行交错复杂 | 复用已存在的 `chatStreamImpl` 事件模型，只接线不重写 |
 | Catalog 迁移漏模型 | 以 `config/models.yaml` 为真值源迁移，向导/README 派生 |
 | 美学改动破坏现有观感 | 设计 token 抽取为"提取"非"重画"，先快照对比 |
+
+---
+
+## 6. Phase 8 — Claude Code / opencode 能力对齐 ✅ 已完成
+
+四项核心能力补齐,全部 `tsc --noEmit` 绿 + 单测覆盖(351 → 387 用例,+36):
+
+- [x] **P8.1 通用可定义子智能体**(对标 Claude Code `Task` 工具)
+  [`src/core/subagent.ts`](../src/core/subagent.ts) + [`src/tools/spawn.ts`](../src/tools/spawn.ts):`spawn_agent` 工具派生**隔离上下文**的子智能体(独立临时记忆,用完即删),只回传最终报告。内置 `general-purpose` / `explore`;支持 `.sky/agents/*.md` 与 `.claude/agents/*.md` 自定义(frontmatter 兼容 Claude Code,工具名自动映射)。子智能体永不持有 `spawn_agent`(无递归)。`/agents` 列举。新增 [`tests/subagent.test.ts`](../tests/subagent.test.ts) 13 用例。
+- [x] **P8.2 精确编辑 + diff**(对标 Claude Code `Edit`)
+  [`src/core/diff.ts`](../src/core/diff.ts) + 升级 `edit_file`:强制**唯一匹配**(歧义即拒)、`replace_all`、no-op 检测、返回统一 diff(+/- 统计)。并修复旧实现 `String.replace` 把 `$&`/`$1` 当替换模式的隐患。新增 [`tests/edit_diff.test.ts`](../tests/edit_diff.test.ts) 10 用例。
+- [x] **P8.3 后台任务 / 长进程**(对标 Claude Code Bash `run_in_background`)
+  [`src/core/bgproc.ts`](../src/core/bgproc.ts):`run_bash` 加 `background=true` 派后台子进程,`bash_output`(增量读)/`list_bash`/`kill_bash` 控制。复用沙箱红线预检;滚动输出上限;会话退出 `killAll`(无孤儿)。新增 [`tests/bgproc.test.ts`](../tests/bgproc.test.ts) 5 用例。
+- [x] **P8.4 诊断(LSP 关键能力)**(对标 opencode LSP)
+  [`src/core/diagnostics.ts`](../src/core/diagnostics.ts) + `get_diagnostics` 工具:TS/JS 经**工作区** TypeScript 编译器 API 取真实语义诊断(行:列 + TS 码),零额外安装;其他语言走 `config.diagnostics` 配置的外部 checker(解析 `file:line:col` 输出)。新增 [`tests/diagnostics.test.ts`](../tests/diagnostics.test.ts) 8 用例。
+
+> 这是把"完整的多 Agent 终端"推进到**与 Claude Code / opencode 同代能力面**的一步:可派生子智能体、精确可审计的编辑、后台进程、按文件诊断闭环。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "skyloom",
-  "version": "1.16.1",
+  "version": "1.17.0",
   "description": "天空织机 Skyloom — 6 weather-themed AI agents: Fog, Rain, Frost, Snow, Dew, Fair",
   "preferGlobal": true,
   "type": "commonjs",

package/src/cli/loom_chat.ts

@@ -424,6 +424,17 @@ export async function loomChat(ctx: any, startAgent: any, deps: LoomChatDeps): P
         dim(`模式 → ${mode.current}${mode.current === "default" ? "" : " · Shift+Tab 或 /default 切回"}`);
         continue;
       }
+      if (cmdL === "/perm" || cmdL.startsWith("/perm ")) {
+        const { getSecurity, PERMISSION_MODE_ALIASES } = require("../core/security");
+        const sec = getSecurity();
+        const arg = inp.split(/\s+/)[1]?.toLowerCase();
+        if (!arg) { dim(`权限模式: ${sec.approvalMode} · 可选 default | auto | accept | strict | bypass`); continue; }
+        const m = PERMISSION_MODE_ALIASES[arg];
+        if (!m) { dim(`未知权限模式 '${arg}' · 可选 default | auto | accept | strict | bypass`); continue; }
+        sec.setMode(m);
+        dim(`✓ 权限模式 → ${m}`);
+        continue;
+      }
       if (cmdL === "/context") {
         try {
           const d = agent.contextDetail();

package/src/cli/main.ts

@@ -303,8 +303,14 @@ async function chat(agentName: string, modelOverride?: string, classic?: boolean
 
   // Wire up security approval — prompt user for HIGH/CRITICAL operations
   try {
-    const { getSecurity, DangerLevel } = require("../core/security");
+    const { getSecurity, DangerLevel, PERMISSION_MODE_ALIASES } = require("../core/security");
     const sec = getSecurity();
+    // Honor a configured permission mode (config.yaml cli.approvalMode), mapped
+    // through the same aliases as /perm.
+    const cfgMode = (ctx as any).config?.cli?.approvalMode || (ctx as any).config?.cli?.approval_mode;
+    if (cfgMode && PERMISSION_MODE_ALIASES[String(cfgMode).toLowerCase()]) {
+      sec.setMode(PERMISSION_MODE_ALIASES[String(cfgMode).toLowerCase()]);
+    }
     sec.setApprovalCallback(async (tool: string, args: Record<string, any>, level: number) => {
       process.stdout.write(chalk.yellow(`\n  ⚠ ${tool} ( danger level ${level} )\n`));
       process.stdout.write(chalk.dim(`     args: ${JSON.stringify(args).slice(0, 80)}\n`));
@@ -415,6 +421,17 @@ async function chat(agentName: string, modelOverride?: string, classic?: boolean
       process.stdout.write(chalk.dim(`  模式 → ${MODE.current} · ${MODE.describe()}\n`));
       continue;
     }
+    if (cmdL === "/perm" || cmdL.startsWith("/perm ")) {
+      const { getSecurity, PERMISSION_MODE_ALIASES } = require("../core/security");
+      const sec = getSecurity();
+      const arg = inp.split(/\s+/)[1]?.toLowerCase();
+      if (!arg) { process.stdout.write(chalk.dim(`  权限模式: ${sec.approvalMode} · 可选 default | auto | accept | strict | bypass\n`)); continue; }
+      const m = PERMISSION_MODE_ALIASES[arg];
+      if (!m) { process.stdout.write(chalk.yellow(`  未知权限模式 '${arg}' · 可选 default | auto | accept | strict | bypass\n`)); continue; }
+      sec.setMode(m);
+      process.stdout.write(chalk.green(`  ✓ 权限模式 → ${m}\n`));
+      continue;
+    }
     if (cmdL === "/context") {
       try {
         const d = currentAgent.contextDetail();
@@ -438,6 +455,19 @@ async function chat(agentName: string, modelOverride?: string, classic?: boolean
       process.stdout.write("\n");
       continue;
     }
+    if (cmdL === "/agents") {
+      const { loadSubagentDefinitions } = require("../core/subagent");
+      const defs = loadSubagentDefinitions();
+      process.stdout.write(chalk.bold(`\n  可派生子智能体 · spawn_agent\n`));
+      for (const d of defs.values()) {
+        const scope = d.source === "builtin" ? "内置" : "自定义";
+        const tools = d.tools === null ? "全部工具" : `${d.tools.length} 个工具`;
+        process.stdout.write(chalk.dim(`  ◇ ${String(d.name).padEnd(18)} ${d.description}\n`));
+        process.stdout.write(chalk.dim(`    └ ${scope} · ${tools}${d.model ? ` · ${d.model}` : ""}\n`));
+      }
+      process.stdout.write(chalk.dim(`\n  自定义: 在 .sky/agents/ 或 .claude/agents/ 放 <name>.md (frontmatter: description/tools/model)\n\n`));
+      continue;
+    }
     if (cmdL === "/trace") {
       const trace = (currentAgent as any).getLastTrace?.();
       if (!trace || !trace.spans?.length) { process.stdout.write(chalk.dim("  本会话还没有可追踪的运行\n")); continue; }

package/src/core/agent.ts

@@ -166,6 +166,17 @@ export class BaseAgent {
     }
   }
 
+  /** Consolidated environment snapshot (cwd/platform/git/date) — see envcontext. */
+  protected injectEnvironment(prompt: string): string {
+    try {
+      const { buildEnvBlock } = require('./envcontext');
+      const lang = (this.config as any).llm?.language || 'zh';
+      return prompt + '\n\n' + buildEnvBlock({ lang });
+    } catch {
+      return prompt;
+    }
+  }
+
   /** Always return the live current time — never stale. */
   protected currentTimeTag(): string {
     const date = new Date();
@@ -229,6 +240,7 @@ export class BaseAgent {
     this._baseSystemPrompt = '';
     this._baseSystemPrompt = this.resolveSystemPrompt();
     this._baseSystemPrompt = this.injectWorkspaceInfo(this._baseSystemPrompt);
+    this._baseSystemPrompt = this.injectEnvironment(this._baseSystemPrompt);
     this._baseSystemPrompt = this.injectBehaviorRules(this._baseSystemPrompt);
     this._baseSystemPrompt = this.injectProgrammingWisdom(this._baseSystemPrompt);
     this._baseSystemPrompt = this.injectProjectMemory(this._baseSystemPrompt);
@@ -254,6 +266,7 @@ export class BaseAgent {
 
     this._baseSystemPrompt = this.resolveSystemPrompt();
     this._baseSystemPrompt = this.injectWorkspaceInfo(this._baseSystemPrompt);
+    this._baseSystemPrompt = this.injectEnvironment(this._baseSystemPrompt);
     this._baseSystemPrompt = this.injectBehaviorRules(this._baseSystemPrompt);
     this._baseSystemPrompt = this.injectProgrammingWisdom(this._baseSystemPrompt);
     this._baseSystemPrompt = this.injectProjectMemory(this._baseSystemPrompt);
@@ -609,8 +622,12 @@ export class BaseAgent {
 
     for (let i = 0; i < parsed.length; i++) {
       const p = parsed[i];
-      // Dedup: only for cacheable, non-dangerous tools with identical args
-      if (options?.dedupCacheable && p.toolArgs && p.tool && (p.tool as ToolDefinition).cacheable && !(p.tool as ToolDefinition).dangerous) {
+      // Dedup identical calls within this round for read-only tools (idempotent
+      // or cacheable, never dangerous): the model often emits the same
+      // read_file / web_search twice in one parallel batch — run it once and
+      // share the result. Safe because one round observes one world state.
+      const td = p.tool as ToolDefinition | undefined;
+      if (options?.dedupCacheable && p.toolArgs && td && (td.idempotent || td.cacheable) && !td.dangerous) {
         const key = `${p.toolName}:${JSON.stringify(p.toolArgs, Object.keys(p.toolArgs).sort())}`;
         if (seenDedupKeys.has(key)) {
           execPlan.push({ idx: i, prep: p, isDuplicate: true });
@@ -629,6 +646,11 @@ export class BaseAgent {
     const results = new Array<{ tc: ToolCall; result: string; success: boolean; toolName: string } | null>(parsed.length).fill(null);
     const uniquePlan = execPlan.filter(e => !e.isDuplicate);
     const concurrency = resolveConcurrency((this.config as any)?.llm?.tool_concurrency);
+    // Enter ACTING once for the whole batch rather than per-tool (each parallel
+    // worker re-setting the same state was a redundant async no-op).
+    if (uniquePlan.some(e => e.prep.tool && !e.prep.parseError && !e.prep.denied)) {
+      await this.setState(AgentState.ACTING);
+    }
     const completed = await mapBounded(
       uniquePlan,
       async ({ idx, prep }, _i, aborted) => {
@@ -651,7 +673,6 @@ export class BaseAgent {
         }
 
         if (onStatus) onStatus(p.label);
-        await this.setState(AgentState.ACTING);
 
         // File checkpoint: snapshot the target before any mutating file tool
         // runs, so /rewind can restore the pre-turn state.
@@ -1393,7 +1414,7 @@ export class BaseAgent {
         });
 
         // ── Execute all tools via shared pipeline ──
-        await this.executeToolCalls(response.toolCalls, { onStatus: onStatus ?? undefined, ephemeral });
+        await this.executeToolCalls(response.toolCalls, { dedupCacheable: true, onStatus: onStatus ?? undefined, ephemeral });
         await this.setState(AgentState.THINKING);
       }
 

package/src/core/bgproc.ts

@@ -0,0 +1,153 @@
+/**
+ * Background process manager — long-running shells that don't block the agent.
+ *
+ * `run_bash` with background=true spawns a child here and returns immediately
+ * with a job id. The agent later pulls incremental output (bash_output), lists
+ * jobs (list_bash), or terminates one (kill_bash). Children are NOT detached,
+ * so they die with the sky process — no orphan management needed across runs.
+ */
+
+import { spawn, type ChildProcess, execSync } from 'child_process';
+import { getLogger } from './logger';
+import { preflightCheck } from './sandbox';
+
+const log = getLogger('bgproc');
+
+/** Per-job rolling output cap — keep the tail, drop the oldest. */
+const MAX_LOG_BYTES = 512 * 1024;
+
+export type BgStatus = 'running' | 'exited' | 'killed' | 'error';
+
+export interface BgJobView {
+  id: string;
+  command: string;
+  pid: number | null;
+  status: BgStatus;
+  exitCode: number | null;
+  startedAt: number;
+  endedAt: number | null;
+  /** Total bytes ever produced (before any rolling trim). */
+  totalBytes: number;
+}
+
+interface BgJob extends BgJobView {
+  child: ChildProcess | null;
+  log: string;       // combined stdout+stderr in arrival order
+  readOffset: number; // cursor for incremental reads
+  trimmed: number;    // bytes dropped from the front by the rolling cap
+}
+
+class BackgroundManager {
+  private jobs = new Map<string, BgJob>();
+  private seq = 0;
+
+  private append(job: BgJob, text: string): void {
+    job.log += text;
+    job.totalBytes += Buffer.byteLength(text, 'utf8');
+    if (job.log.length > MAX_LOG_BYTES) {
+      const drop = job.log.length - MAX_LOG_BYTES;
+      job.log = job.log.slice(drop);
+      job.trimmed += drop;
+      job.readOffset = Math.max(0, job.readOffset - drop);
+    }
+  }
+
+  /** Start a background command. Returns the job id, or an error string. */
+  start(command: string, opts?: { cwd?: string; env?: Record<string, string> }): { id?: string; error?: string } {
+    const check = preflightCheck(command);
+    if (check) return { error: `[BLOCKED] ${check}` };
+
+    const id = `bg_${(++this.seq).toString(36)}_${Date.now().toString(36)}`;
+    let child: ChildProcess;
+    try {
+      child = spawn(command, {
+        shell: true,
+        cwd: opts?.cwd || process.cwd(),
+        env: { ...process.env, ...(opts?.env || {}) },
+        windowsHide: true,
+      });
+    } catch (e: any) {
+      return { error: `Failed to start background command: ${e.message || e}` };
+    }
+
+    const job: BgJob = {
+      id, command, pid: child.pid ?? null, status: 'running', exitCode: null,
+      startedAt: Date.now(), endedAt: null, totalBytes: 0,
+      child, log: '', readOffset: 0, trimmed: 0,
+    };
+    this.jobs.set(id, job);
+
+    child.stdout?.on('data', (d) => this.append(job, d.toString()));
+    child.stderr?.on('data', (d) => this.append(job, d.toString()));
+    child.on('error', (e) => {
+      job.status = 'error';
+      job.endedAt = Date.now();
+      this.append(job, `\n[spawn error] ${e.message}\n`);
+      log.warn('bg_error', { id, error: e.message });
+    });
+    child.on('exit', (code, signal) => {
+      if (job.status === 'running') job.status = signal ? 'killed' : 'exited';
+      job.exitCode = code;
+      job.endedAt = Date.now();
+      job.child = null;
+    });
+
+    return { id };
+  }
+
+  get(id: string): BgJob | undefined { return this.jobs.get(id); }
+
+  list(): BgJobView[] {
+    return [...this.jobs.values()].map(toView).sort((a, b) => b.startedAt - a.startedAt);
+  }
+
+  /** Read output produced since the last read; advances the cursor. */
+  read(id: string): { ok: boolean; text?: string; status?: BgStatus; exitCode?: number | null; error?: string } {
+    const job = this.jobs.get(id);
+    if (!job) return { ok: false, error: `No background job '${id}'.` };
+    const chunk = job.log.slice(job.readOffset);
+    job.readOffset = job.log.length;
+    return { ok: true, text: chunk, status: job.status, exitCode: job.exitCode };
+  }
+
+  kill(id: string): { ok: boolean; error?: string } {
+    const job = this.jobs.get(id);
+    if (!job) return { ok: false, error: `No background job '${id}'.` };
+    if (job.status !== 'running' || !job.child) return { ok: false, error: `Job '${id}' is not running (${job.status}).` };
+    const pid = job.child.pid;
+    try {
+      if (process.platform === 'win32' && pid) {
+        // child.kill() doesn't reap the cmd.exe process tree on Windows.
+        try { execSync(`taskkill /pid ${pid} /T /F`, { stdio: 'ignore' }); }
+        catch { job.child.kill(); }
+      } else {
+        job.child.kill('SIGTERM');
+      }
+      job.status = 'killed';
+      job.endedAt = Date.now();
+      return { ok: true };
+    } catch (e: any) {
+      return { ok: false, error: `Failed to kill '${id}': ${e.message || e}` };
+    }
+  }
+
+  /** Terminate all running jobs (called on session shutdown). */
+  killAll(): void {
+    for (const job of this.jobs.values()) {
+      if (job.status === 'running') this.kill(job.id);
+    }
+  }
+}
+
+function toView(j: BgJob): BgJobView {
+  return {
+    id: j.id, command: j.command, pid: j.pid, status: j.status,
+    exitCode: j.exitCode, startedAt: j.startedAt, endedAt: j.endedAt, totalBytes: j.totalBytes,
+  };
+}
+
+let _mgr: BackgroundManager | null = null;
+export function getBackgroundManager(): BackgroundManager {
+  if (!_mgr) _mgr = new BackgroundManager();
+  return _mgr;
+}

package/src/core/commands.ts

@@ -313,6 +313,16 @@ export const BUILTIN_COMMANDS: CommandInfo[] = [
     takesArgs: false,
     source: 'builtin',
   },
+  {
+    name: 'agents',
+    aliases: [],
+    description: 'List spawnable subagents',
+    label: '可派生子智能体',
+    category: 'context',
+    hints: [],
+    takesArgs: false,
+    source: 'builtin',
+  },
   {
     name: 'workspace',
     aliases: [],
@@ -428,6 +438,16 @@ export const BUILTIN_COMMANDS: CommandInfo[] = [
     takesArgs: false,
     source: 'builtin',
   },
+  {
+    name: 'perm',
+    aliases: [],
+    description: 'Set permission mode (default|auto|accept|strict|bypass)',
+    label: '权限模式（default/auto/accept/strict/bypass）',
+    category: 'workflow',
+    hints: ['default', 'auto', 'accept', 'strict', 'bypass'],
+    takesArgs: true,
+    source: 'builtin',
+  },
 
   // ── File & Checkpoint ──
   {