page-agent-sdk 3.35.0 → 3.36.0

package/README.md

@@ -163,6 +163,7 @@ CDN zero-config: `<script src="https://unpkg.com/page-agent-sdk"></script>` →
 | 🤖 Agent-driven compression (2.33+) | `capabilities.agentCompression` (opt-in) lets the summary LLM decide per-trigger compression strategy via an `inspect_context` tool loop (keepRounds / windowRatio / summary mode / recall / preserve); `shouldTriggerCompression` gate avoids per-message LLM cost; decide failure/timeout degrades to static; `decisionTimeoutMs` / `decisionMaxTokens` configurable | `capabilities.agentCompression` + `summaryLlm` |
 | 🎯 Cross-session user preference memory | `capabilities.preferences` (**opt-in, default off** — auto-writing the user's browser is behavior-sensitive): the agent captures durable user preferences from conversation — strong signal (explicit commands like "Remember: …", zero LLM) / medium signal (pattern-word prefilter + small-LLM extraction; the core test is **durable taste vs this-round task instruction**) / behavioral inference **not captured** (better to miss than to learn wrong — one false preference would ride along every future session); preferences persist independently (preferenceStore, IndexedDB, same shape as storage/skillStorage; same topic **later statement overrides earlier**, FIFO ≤20); injected as a pin segment into the system prompt each round (survives sessions and compression); manage wrongly-learned entries via `sdk.getPreferences()/removePreference(id)/clearPreferences()`, plus a read-only DebugDrawer "User preferences" section | `capabilities: { preferences: true }` + optional `preferenceStorage` |
 | 🧭 Instruction adherence (3.35+) | **Completion gate**: if the agent tries to close with plain text while todos still have unfinished items, a "two-exit" nudge is injected (mark done via update_todo, or keep executing; ≤2 retries) — fixes "planned 3 tasks, did 1, then stopped" premature interruption. **Question-intent guard**: a 3-tier regex heuristic classifies each user message as a question; on hit a "answer first, don't act" pin segment is injected (survives compression) — fixes long-chat questions being dragged into actions by history (e.g. asking "what is this component" but ending up generating code). Both default on, zero config, prefer-miss-over-false-positive | built-in |
+| 🎨 Subagent model/thinking tiering | `createHtmlSubagent({ llm, thinkingMode })`: code-gen subagents get their own stronger model (main stays light for orchestration) + thinking-depth lock (`'deep'` injects thinking params for quality / `'simple'` strips them to save tokens; top-level `subagent.thinkingMode` as global default). LLMConfig construction path only (pre-built instance → warn + no-op); requires a thinking-capable model (deepseek-thinking / claude); `inspect().subagent.subagents` reflects the effective state | `createHtmlSubagent({ llm, thinkingMode })` |
 | ⚡ host actions (2.20+) | Register save/publish/preview etc; SDK auto-generates named tools, agent triggers page ops directly (no `trigger_action` indirection) | `actions` |
 | 🧩 schema tiered disclosure (2.20+) | Large schema → systemPrompt injects top-level overview only (no constraints/no recursion); deep constraints via `schema_data` on demand; small schema unaffected (full) | `schemaHint` |
 | 📌 cross-compress working memory (2.20+) | Pin recent read/query paths + hashes across compression; no re-fetch, correct optimistic-lock hash | `capabilities.workingMemory` |
@@ -493,8 +494,8 @@ function switchTo(i: number) {
 ## Self-tests
 
 ```bash
-npm test            # 2560 assertions (tsx, source-level; no LLM dependency)
-npm run test:e2e    # 843 integration assertions (node, built dist; covers APIs/options/modules/simple&complex scenes: default systemPrompt(capability overview) / dynamic register + inspect sync / inspect(tools/middleware/subagent/verify/mcp/todos/lastCompression/checkpoints reflect config) / custom tools/middleware/skills/memory injection / runtime dynamic reconfiguration(setTools/addTool/removeTool/setLlm/setMemory/setSubagents reflect) / switchSession(on/off) / shareContext on/off sharing/independent / storage backends + object config / presets(3) / checkpoint / exports complete(39+ fns/components) / util fns usable(isQuotaError/estimateTokens/jpEval/searchJson) / source=builtin / mount boundary / hook multi-listener / llm config / hide/show / error scenes)
+npm test            # 2573 assertions (tsx, source-level; no LLM dependency)
+npm run test:e2e    # 856 integration assertions (node, built dist; covers APIs/options/modules/simple&complex scenes: default systemPrompt(capability overview) / dynamic register + inspect sync / inspect(tools/middleware/subagent/verify/mcp/todos/lastCompression/checkpoints reflect config) / custom tools/middleware/skills/memory injection / runtime dynamic reconfiguration(setTools/addTool/removeTool/setLlm/setMemory/setSubagents reflect) / switchSession(on/off) / shareContext on/off sharing/independent / storage backends + object config / presets(3) / checkpoint / exports complete(39+ fns/components) / util fns usable(isQuotaError/estimateTokens/jpEval/searchJson) / source=builtin / mount boundary / hook multi-listener / llm config / hide/show / error scenes)
 ```
 
 ## Local npm package test

package/README.zh-CN.md

@@ -165,6 +165,7 @@ CDN 零配置：`<script src="https://unpkg.com/page-agent-sdk"></script>` → `
 | 🌍 UI 定制与国际化 (3.17+~3.22+) | 对话框 UI 免 fork 全定制:`dialog.icons` 逐图标覆盖(纯文本或净化后 HTML 片段)+ 内置深色主题 `dialog.theme:'dark'` + **顶层 `i18n` 配置组(3.22+)**:`locale:'en-US'` 切内置文案包(聊天面 + Debug 抽屉 + Skill 面板 + 代码预览;`formatTime`/autoTitle 跟随,**默认 systemPrompt 切英文** → agent 回复语言与 UI 一致)、`messages` 键级覆盖(如 `statusDone: '<b style="color:#10b981">Done ✓</b>'` —— 富文本渲染位支持行内 HTML 片段,文案白名单净化)——换语言与改个别文案一个配置组;`DialogMessages`(~226 键)+ `MESSAGES_ZH_CN`/`MESSAGES_EN_US`/`resolveDialogMessages` 导出供自建 UI 复用 | `dialog.{icons,theme}` + `i18n.{locale,messages}` |
 | 🎯 跨会话用户偏好记忆 | `capabilities.preferences`(**opt-in 默认关**,自动写用户浏览器属行为敏感项):agent 从对话中捕获用户持久偏好 —— 强信号(「记住:…」显式命令,零 LLM)/ 中信号(模式词初筛 + 小 LLM 提炼,核心判定「持久口味 vs 本轮任务指令」)/ 行为推断**不捕获**(宁漏勿误:学错一条假偏好,之后每个会话都带着跑);偏好独立持久化(preferenceStore,IndexedDB,与 storage/skillStorage 同构;同 topic **后说覆盖前说**,FIFO ≤20);每轮经 pin 段注入 system prompt(跨会话/跨压缩生效);`sdk.getPreferences()/removePreference(id)/clearPreferences()` 管理学错条目,DebugDrawer「用户偏好」只读小节可查 | `capabilities: { preferences: true }` + 可选 `preferenceStorage` |
 | 🧭 指令执行力增强 (3.35+) | **完结门禁**:todos 有未完成项却欲纯文本收尾 → 回灌「双出口」反馈续跑(≤2 次),防「拆 3 项做 1 项就收口」的莫名中断;**问句意图守卫**:正则三档启发式逐消息定性问句,命中注入「先答勿做」pin 段(跨压缩存活),防长对话提问被历史拖着误路由成操作(如问「这是啥组件」却去生成代码)。均默认开、零配置、宁漏勿误 | 内置 |
+| 🎨 子 agent 模型/思考分层 | `createHtmlSubagent({ llm, thinkingMode })`:代码生成子 agent 独立强模型(主保持轻量编排)+ 思考深度锁定(`'deep'` 注入思考参数质量优先 / `'simple'` 剥除省 token;顶层 `subagent.thinkingMode` 全局缺省)。仅 LLMConfig 构造路径生效(预构造实例 warn+no-op);需模型支持思考(deepseek thinking 版/claude);`inspect().subagent.subagents` 反射生效状态 | `createHtmlSubagent({ llm, thinkingMode })` |
 
 能力默认开（`verify`/`approval`/`checkpoint` 默认关；**主动征询 `humanConfirm` 默认开**——AI 遇不确定/多方案主动问你、不猜测），可经 `capabilities` 关掉无用的省 token。
 
@@ -496,8 +497,8 @@ function switchTo(i: number) {
 ## 自测
 
 ```bash
-npm test            # 2560 项断言（tsx 源码级，不依赖 LLM）
-npm run test:e2e    # 843 项集成断言（node 跑构建产物 dist；覆盖各 API/配置项/功能模块/简单与复杂场景：默认 systemPrompt(含能力概述) / 动态注册与 inspect 同步 / inspect(tools/middleware/subagent/verify/mcp/todos/lastCompression/checkpoints 反映配置) / 自定义 tools/middleware/skills/memory 注入 / 运行时动态重配置(setTools/addTool/removeTool/setLlm/setMemory/setSubagents 反映) / switchSession(开/未开) / shareContext 开/关共享独立 / storage 后端+对象配置 / presets 三预设 / checkpoint / 导出项完整(39+ 函数/组件) / 工具函数可用(isQuotaError/estimateTokens/jpEval/searchJson) / source=builtin / mount 边界 / hook 多监听器 / llm 配置 / 错误场景）
+npm test            # 2573 项断言（tsx 源码级，不依赖 LLM）
+npm run test:e2e    # 856 项集成断言（node 跑构建产物 dist；覆盖各 API/配置项/功能模块/简单与复杂场景：默认 systemPrompt(含能力概述) / 动态注册与 inspect 同步 / inspect(tools/middleware/subagent/verify/mcp/todos/lastCompression/checkpoints 反映配置) / 自定义 tools/middleware/skills/memory 注入 / 运行时动态重配置(setTools/addTool/removeTool/setLlm/setMemory/setSubagents 反映) / switchSession(开/未开) / shareContext 开/关共享独立 / storage 后端+对象配置 / presets 三预设 / checkpoint / 导出项完整(39+ 函数/组件) / 工具函数可用(isQuotaError/estimateTokens/jpEval/searchJson) / source=builtin / mount 边界 / hook 多监听器 / llm 配置 / 错误场景）
 ```
 
 ## 本地 npm 包测试