Execution 0.1.5 — 2026-10-17 performance refresh

2026-10-17 refresh — still 0.1.5 / 同版本更新

Download the refreshed Windows installer / 下载更新安装器 · Checksums / 校验文件 · Exact build source / 确切构建源码

This is build 20261017-01, retaining Runtime 0.1.5 and the optional Amadeus extension 0.0.2. Existing installations require an explicit update. The original tag and four original assets remain unchanged; choose the dated installer above for this refresh.

本次构建为 20261017-01，沿用 Runtime 0.1.5，可选 Amadeus 扩展仍为 0.0.2。已有安装需要明确更新；原标签与原四个附件保留，下载本次更新请使用上方带日期的安装器。

What changed / 改动

• Adaptive observation now also covers default visual Enter submissions. It checks when the image has changed and settled, bounded by the original grace period. Stable pixels select the check time; they do not count as successful execution.
• Enter 提交也采用自适应观察，在原宽限内检查画面变化和稳定状态；稳定只决定送图时机，不代替结果核验。
• Current-result verification overlaps with one next-step read-only grounding request. The next action still waits for persisted predecessor success and all existing action gates. Changed frames, cancellation and handoff invalidate prefetched results.
• 当前结果复核与下一步的只读定位并行；前步成功落账、原动作检查通过后才可执行。换图、取消和接管均使预取失效。
• Includes the earlier ordinary-click wait optimization and verified SSL-context reuse. Grounding models, prompts, image resolution, Decision acceptance rules and action permissions are unchanged. Learning-mode promotion and active development are paused; its source and historical releases remain available.
• 同时包含此前普通点击等待优化及经验证的 SSL 上下文复用。定位模型、提示词、分辨率、Decision 通过规则和动作权限未变；暂时隐藏学习模式宣传入口、暂停开发，保留源码及历史发布。

Actual action intervals / 真实动作间隔

Source A/B comparison, four matched intervals per application/arm. A already includes the earlier ordinary-click/SSL optimizations; B adds Enter observation and overlap. This is not a direct comparison with the original published installer.

源码 A/B 对照，每应用／组四个匹配区间；A 已含普通点击／SSL 优化，B 再加入 Enter 观察和并行。不是旧发布安装器与新包的直接对照。

Application / 应用	Median start-to-start interval A → B / 相邻逻辑动作开始间隔中位数
Notepad / 记事本	5.803 → 5.597 s
7-Zip	7.805 → 7.254 s
Edge	10.786 → 9.280 s

Intervals run from one logical step's first physical mouse-down to the next, including intervening verification and preparation. They are not pure idle time or API latency. The four-second goal has not been reached. All 12 source plans / 36 steps passed on the first attempt; one Edge B plan was slower than A. Small samples and uncontrolled model/network/cache variation limit attribution. Source performance plus separate cancellation/recovery checks used 139 requests / 268,593 reported tokens.

间隔按相邻逻辑步骤首次真实鼠标按下统计，包含中间核验与准备，不是纯等待或 API 响应时间。尚未达到四秒目标。 源码 12 个计划／36 步首次完成，但 Edge 的一轮 B 比 A 慢；小样本及模型、网络、缓存波动限制因果结论。源码性能加单列取消／恢复检查共 139 次请求／268,593 个已报告 token。

New installed-package acceptance / 新安装包验收

Application / 应用	Reviewer / 验收者	Three-step plan 1 / 2 / 三步计划第一轮／第二轮
Notepad / 记事本	Main	20.744 / 18.191 s
Edge	Main	27.043 / 27.430 s
7-Zip	Independent Sol / 独立 Sol	25.918 / 24.324 s

Six plans / 18 steps passed on the first attempt, using the same frozen installed payload. Original screenshots were inspected; Notepad file bytes matched; all owned windows and hosts closed normally. 7-Zip covered folder/archive navigation, not compression/extraction. Edge used a retained independent profile without clearing browser data. Times cover complete plan calls including visual verification, excluding launch, between-round manual inspection and cleanup.

同一冻结安装载荷的六个计划／18 步均首次完成，原图核对、记事本磁盘内容和窗口／宿主收尾均通过。7-Zip 为目录及预制压缩包浏览，不含压缩解压。Edge 沿用保留的独立资料目录，不清缓存。时间为包含视觉核验的完整计划调用，不含启动、轮间人工查看和收尾。

Installed runs used 66 requests: 18 grounding + 46 Decision + 2 ordinary review; 127,178 reported tokens, with no unknown usage. Consumed prefetch copies are deduplicated. Main/worker subscription usage is excluded. These scoped passes do not establish a general accuracy rate; false accepts remain a known model limitation.

安装态共 66 次请求：18 次定位＋46 次 Decision＋2 次普通复核；127,178 个已报告 token，无缺失用量，预取消费副本已去重，不含 Main／工作 Agent 套餐用量。限定场景通过不代表通用准确率；模型误放仍是已知限制。

Package and checks / 安装包与验证

• 395 affected checks passed; after race fixes, 82 overlapping checks passed. Counts are not summed as unique tests. Cancellation/recovery and delayed/forged/stale evidence have separate coverage.
• 395 项相关检查通过；竞态修复后 82 项重叠检查通过，不相加为独立测试数；取消恢复、迟到变化、伪造及旧证据另有覆盖。
• All 659 payload file hashes matched before/after installed acceptance. Check-mode same-version upgrade, fresh install/uninstall, preserved external data and injected-failure rollback passed. A new isolated Python 3.11 environment installed the 45 locked API-route dependencies and passed functional-entrypoint checks.
• 实测前后 659 个载荷文件哈希一致；检查模式的同版本升级、新装卸载、保留外部数据及注入失败回滚通过；新建隔离 Python 3.11 环境安装 45 个锁定 API 路线依赖，真实功能入口验证通过。
• Installer remains unsigned and requires Python/dependency setup. This refresh does not add normal OS registry/shortcut or clean-machine acceptance. Existing Main review/continuation remains supported; the Amadeus wheel is unchanged and was not requalified in a new Amadeus pipeline in this refresh.
• 安装器仍未签名，需要准备 Python 和依赖；本次未新增正常注册表／快捷方式集成及全新机器验收。保留既有 Main 审核续行能力；Amadeus wheel 未改动，本次未重新跑 Amadeus 整链。

Installer SHA-256: fb24a66fee62d68876d344d0cc09b00df592c2f853f3ca041b810f1621cd5875

Source commit / 源码提交: f3a680405f4d221ab2ab080d671c15afb4b89f39. The dated source ZIP is the exact build input; this commit has identical runtime code with subsequent documentation-status updates. The original instant-v0.1.5 tag is retained, so GitHub's automatic tag-source archives still describe the original build; use the dated ZIP for this refresh.

日期源码 ZIP 是确切构建输入；上述提交的运行时代码相同，补充了后续文档状态。原 instant-v0.1.5 标签保留，因此 GitHub 自动生成的标签源码包仍对应原构建；复现本次更新请使用带日期的源码 ZIP。

Full method, limitations and per-application evidence / 完整方法、限制与多应用证据 · English setup · 中文安装指南

Original 0.1.5 build — historical evidence / 原构建历史记录

Execution 0.1.5 adds a way for the original planning Agent (Main) to review settled text input and continue the same live command without typing it again. It also delivers the optional Amadeus MCP + Pi Skill extension as a local wheel.

执行版 0.1.5 新增：原规划 Agent（Main）可以审核已完成的文字输入，并明确继续同一条存活命令，不重复输入。同时交付可选的 Amadeus MCP＋Pi Skill 本地扩展包。

What changes / 本次变化

• After an authenticated Decision + ordinary-model nonpass on settled text input, Runtime preserves the original job, bound screenshot and review packet. Main records its actual judgment, then explicitly continues the original command. Remaining submit actions still require fresh target checks and result verification. Authentication/protocol failures, unknown input outcome, cancellation and process restart cannot be approved away through this route.
• 已结算输入遇到经过认证的 Decision＋普通模型双重非通过后，Runtime 保留原任务、绑定截图和审核包。Main 记录实际判断后，明确续行原命令；剩余提交动作仍需新鲜目标检查及结果核验。此入口不能覆盖认证／协议错误、输入结果未知、取消或进程重启。
• Extension 0.0.2 contributes a stdio MCP connection and a native Pi Skill, launching an explicitly configured external Runtime with its existing Python 3.11 environment. It contains no second executor, Runtime payload, model weights or credentials. Install through Amadeus's local plugin manager; a marketplace is not required.
• 扩展 0.0.2 提供 stdio MCP 连接与原生 Pi Skill，启动明确配置的外部 Runtime 及已有 Python 3.11 环境。不包含第二套执行器、Runtime 载荷、模型权重或密钥。通过 Amadeus 本地插件管理器安装，无需插件商店。

Verified scope / 已验证范围

Check / 检查	Result / 结果
Affected-source regression / 受影响源码回归	135 passed, 0 failed, 0 skipped. Overlapping earlier groups are not added. / 135 项通过，零失败、零跳过，与先前重叠分组不累加。
Final installer / 最终安装包	Isolated entrypoint/dependency checks and all 654 installed file hashes passed. / 隔离入口与依赖检查、654 个安装文件哈希核对通过。
Independent final-package desktop run / 最终包独立实机	One three-step Notepad edit → File → Save plan passed on the first attempt in 27.70161 s, including visual verification. 10 HTTP requests: 7 Decision + 3 grounding; no ordinary review, Main override or rerun. Exact 55-byte disk result and owned-window/Runtime cleanup passed. / 一次三步记事本「编辑→文件→保存」计划首次通过，27.70161 秒，包含视觉核验；10 次 HTTP：7 次 Decision＋3 次定位，无普通复核、Main 接管或重跑；55 字节磁盘内容精确匹配，窗口及 Runtime 收尾通过。
Normal Amadeus pipeline / 正常 Amadeus 链路	Chat → Pi → installed extension → installed Runtime completed two tasks, six steps and cleanup. The same Chat used distinct Pi Work and Runtime sessions. One ordinary review was used; no Main semantic override. The second Work corrected two outer errors before dispatch, so this is not a zero-error first-attempt claim. / 完成两项任务、六步与收尾；同一 Chat，分别使用 Pi Work 与 Runtime 会话。发生一次普通复核，无 Main 语义接管；第二项在下发动作前修正了两个外层错误，不宣称零错误首次通过。
Extension lifecycle / 扩展生命周期	Install/configure/enable/restart, disable/restart, remove/restart passed; all 149 retained data files kept their original hashes after removal with data retention. / 安装、配置、启用并重启、禁用并重启、移除并重启通过；选择保留数据后，149 个文件哈希均未改变。

These are scoped functional results, not a general success-rate or speedup benchmark. The independent run's total token count is unknown because one request did not expose usage; the visible partial count is not presented as a total. Earlier 0.1.4 demos and timing comparisons keep their original version labels.

以上是限定场景的功能结果，不代表通用成功率或提速 benchmark。独立实测中一条请求未公开 usage，故总 token 数未知，不把可见部分当作总量。旧 0.1.4 演示和计时继续保留原版本标注。

The earlier live04 controlled TEST dual-refusal run proved that original Pi received the bound image and continued the original command without retyping. It used an earlier unreleased payload; it is not a new natural model-error/accuracy test on this final installer. Role speech playback was observed in software logs during normal integration; microphone recognition and acoustic quality were not evaluated.

先前 live04 显式 TEST 双拒绝实测验证了原 Pi 收到绑定原图、续行原命令且不重复输入；使用的是此前未发布载荷，不能记为本次最终安装包的新自然误拒／准确率测试。正常集成中有角色语音播放的软件日志证据，未评估麦克风识别或听感质量。

Install and compatibility / 安装与兼容

• Download AgentGUIRuntimeExecution-Setup.exe for execution 0.1.5, and verify SHA256SUMS.txt. The installer is unsigned and does not bundle a prepared Python environment. This acceptance reused an existing Python 3.11 environment; first-time setup on a clean machine and normal registry/shortcut integration were not newly verified.
• 下载执行版 AgentGUIRuntimeExecution-Setup.exe，按 SHA256SUMS.txt 校验。安装器未签名，不内置已准备好的 Python 环境。本轮使用已有 Python 3.11；未新增验证全新机器依赖配置及常规注册表／快捷方式集成。
• For Amadeus, also download amadeus_agent_gui_runtime-0.0.2-py3-none-any.whl and follow the English guide / 中文指南. The exact frozen wheel was accepted; its embedded description retains build-time notes, while these companion guides provide current delivery instructions.
• Amadeus 还需下载上述 0.0.2 wheel 并按指南安装。本次交付验收过的原字节 wheel；其内嵌说明保留构建时记录，当前交付方法以仓库配套指南为准。
• Tested Host base: a9a488ae, with Amadeus PR #225 and PR #226 applied. These upstream PRs remain separate from this Runtime release. Do not assume every published Amadeus build is compatible. Main visual review requires an image-capable execution model and preservation of MCP image content.
• 已测 Host 基线为 a9a488ae，应用 PR #225／#226；上游 PR 与本 Runtime 发布分开，不保证任意已发布 Amadeus 均兼容。Main 视觉审核需要支持图像的执行模型及保留 MCP 图像的适配链路。

Wallpaper follow-up / 壁纸问题

The observed reporting failure was traced to Lively 2.2.1.0 mapping a null display name into a non-null protobuf field, producing a false empty wallpaper list. Original/fixed-copy A/B reproduced the difference. A backed-up local third-party DLL repair then passed two normal installed-path cold start → mount → stop/restore cycles. It is not an official Lively update and is not included in either download. An affected Lively installation needs separate repair; installing this extension alone does not resolve that defect.

已定位 Lively 2.2.1.0 将空显示器名称写入非空 protobuf 字段，导致错误地返回空壁纸列表；原版／修正版副本 A/B 已复现差异。本机备份原件后进行第三方 DLL 本地修复，正常安装路径连续两次冷启动、挂载与退出恢复通过。此修复不是 Lively 官方更新，也不包含在两个下载文件中；受影响的 Lively 需单独修复，安装扩展本身不能解决它。

Amadeus's separate uncertain-mount recovery fix is PR #228: retain the original error, owned host and recovery record until an explicit reliable recovery. It does not silently treat an empty query as a successful restore. See detailed evidence and limitations.

Amadeus 对「挂载结果未确认」的恢复修复见 PR #228：保留原错误、自有宿主及恢复记录，等待明确且可信的恢复，不再把空查询静默记为恢复成功。详见上述验收文档。