page-agent-sdk 4.9.3 → 4.11.0

package/README.md

@@ -159,6 +159,9 @@ CDN zero-config: `<script src="https://unpkg.com/page-agent-sdk"></script>` →
 | 🛡️ compression-safe | Live data snapshot + preserved tool results in summary; write returns hint available paths; `systemPromptHelpers.reliableWriteRules` | built-in |
 | 💰 Context economy (3.10/3.11+) | Compression cost cap `promptSoftCapTokens` (defaults to 160K when window ≥320K — huge-window models no longer burn hundreds of thousands of tokens before compressing; reflected via `inspect().compression`) + agent budget self-awareness (70%-rounds / half-cap token hint, consecutive write-failure reminder, per-invocation `roundTokenBudget` friendly wrap-up) + tool-description slimming (-40% prompt) | `contextOptions.promptSoftCapTokens`, `roundTokenBudget` |
 | 💾 persistence | IndexedDB multi-session + quota eviction + switch | `storage` |
+| ⚡ quick actions / session transfer (4.10+) | `dialog.quickActions` chip row above the input sends high-frequency prompts in one click (no pre-fill; queueing/gate semantics inherited); `sdk.exportSession`/`importSession` fully restorable session JSON (copy semantics, never overwrites; `dialog.sessionTransfer:true` adds history-panel entries); `dialog.onDropElement` callback when a host element is dragged into the input (focus entry, mapping stays host-side); `approval.preview:true` renders the write-approval bar as a structured old→new diff (validation failures visible before approving) | `dialog.{quickActions,sessionTransfer,onDropElement}` + `approval.preview` |
+| 🖥 server-side (node) execution (smoke-backed 4.10+) | the same SDK runs in node: `page-agent-sdk/headless` subpath + `ui:false`, backed by a real-LLM dual-protocol smoke test (full read→write→restore_data tool loop); browser-domain APIs explicitly unavailable (get_dom returns a friendly pointer to data tools; canvas render checks auto-degrade); unattended combo (approval auto-reject + `conflictPolicy` + `batch`) for scheduled/webhook long tasks | `ui:false` + headless subpath (`npm run test:node-real` is the in-repo smoke) |
+| 🧪 regression toolkit eval-toolkit | the SDK's own real-LLM regression core, exported for integrators to run **their own** pre-upgrade scenario regressions: `createEvalHarness({sdk})` (waitForIdle dual-condition idle detection + collectReport snapshots) + `diffReport` (token ±15% AND ±2000 / toolCount ±3 thresholds, variance-proof); report shape matches the SDK's own — bug reports parse for free; shipped from both main and headless entries | `import { createEvalHarness, diffReport } from 'page-agent-sdk'` |
 | 👁 DOM inspect (2.20+) | `get_dom` structure read + `dom_search` (selector/text) + `dom_info` (content/computed styles/event bindings from inline/Vue props/listener recorder) — lazy-injected via the `dom-inspect` skill so they don't occupy standing tool context | `capabilities.domInspect` |
 | 📊 Context inspector | Snapshot actual-LLM-message composition (total / occupancy / category ratio); DebugDrawer `📊 上下文` tab + `inspectContext()`; zero LLM cost, default on | `capabilities.contextInspector` |
 | 🤖 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` |

package/README.zh-CN.md

@@ -159,6 +159,9 @@ CDN 零配置：`<script src="https://unpkg.com/page-agent-sdk"></script>` → `
 | 🛡️ 压缩不丢信息 | 摘要内嵌当前 data 快照 + 保留指定工具结果；写返回附可操作 path；`systemPromptHelpers.reliableWriteRules` | 内置 |
 | 💰 上下文经济性 (3.10/3.11+) | 压缩触发成本上限 `promptSoftCapTokens`(窗口 ≥320K 默认 160K,大窗口模型不再烧几十万 token 才压缩;`inspect().compression` 反射)+ agent 预算自感知(轮次 70%/token 半程注入提示、连续写失败提醒、单轮预算 `roundTokenBudget` 友好收口)+ 工具描述瘦身(-40% prompt) | `contextOptions.promptSoftCapTokens`、`roundTokenBudget` |
 | 💾 持久化 | IndexedDB 多会话 + 配额淘汰 + 切换 | `storage` |
+| ⚡ 快捷指令 / 会话转移 (4.10+) | `dialog.quickActions` 输入区 chip 行一键发高频 prompt(不预填输入框,排队/门禁语义自动继承);`sdk.exportSession`/`importSession` 会话导出导入(可复全 JSON,副本语义不覆盖;`dialog.sessionTransfer:true` 历史面板入口);`dialog.onDropElement` 拖宿主元素进输入框回调(聚焦入口,映射归宿主);`approval.preview:true` write 审批确认条渲染结构化 old→new diff(校验失败批准前可见) | `dialog.{quickActions,sessionTransfer,onDropElement}` + `approval.preview` |
+| 🖥 服务端(node)运行 (4.10+ 冒烟背书) | 同一套 SDK 在 node 跑:`page-agent-sdk/headless` 子路径 + `ui:false`,真 LLM 双协议冒烟(read→write→restore_data 完整工具循环)背书;浏览器域 API 明示不可用(get_dom 友好回退/canvas 渲染检查自动降级);无人值守组合(approval 自动拒 + `conflictPolicy` + `batch`)跑定时/webhook 长任务 | `ui:false` + headless 子路径(`npm run test:node-real` 为仓库内冒烟) |
+| 🧪 回归工具 eval-toolkit | SDK 自用真 LLM 回归判定核公开导出,集成方为**自己的场景**跑升级前回归:`createEvalHarness({sdk})`(waitForIdle 双条件 idle 判定 + collectReport 报告快照)+ `diffReport`(token ±15% 且 ±2000 / toolCount ±3 阈值判定,防方差误报);与 SDK 自用报告同构,报问题格式互通;主包/headless 双入口同带 | `import { createEvalHarness, diffReport } from 'page-agent-sdk'` |
 | 🤖 无人值守自动化 (2.20+) | 资源预算闸（`tokenBudget`/`timeBudgetMs`）+ 致命错误自动恢复（`maxAutoRetries`：回退 checkpoint + 重试）+ 刷新续跑 + `sdk.batch(tasks)` 批处理 | `capabilities.automation` |
 | 📐 上下文健壮性 (2.30+) | 硬地板 `contextWindow ≥200K`(启动拒绝 <200K 模型如老款 `deepseek`/`gpt-4o`/`glm-4.5`);三道闸(压缩/trim/offload)阈值在 `setLlm` 后跟随实时窗口;遇 `context_length_exceeded` 反应性重试(激进 trim → 重试一次,不裸失败);vfs 大结果引用受保护免 LRU 淘汰 + OOM 1.5× 兜底;系统段预算(25% 窗口,丢弃非 pin 段保 base/mission/workingMemory) | 内置 |
 | 🎯 focus 自动切换 (2.31+) | AI 自动判断任务范围 → `set_focus`(局部任务)/ `clear_focus`(全局/完成);focus 跨刷新/切会话持久化(restore 经 `getSchemaAtPath` 校验 path,失效丢弃);子 agent 继承主焦点(三层收敛;主未聚焦 → 子无 focus 中间件,零回归) | `capabilities.focus` |