Execution 0.1.6 + Amadeus extension 0.0.4

Runtime 0.1.6 and Amadeus extension 0.0.4 ship together. A model refusal before a click can now return its bound screenshot to the original Main agent for explicit review and separate continuation of the same task.

English

• Native Streamable HTTP MCP, restricted to authenticated loopback access, alongside existing stdio. The extension supplies HTTP MCP + Skill; install and start Runtime separately.
• Pre-action double nonpass → Main reads the before image → explicit review → separate continue of the original job/candidate. Settled text-input review still uses after; completed actions are not replayed.
• Temporary queue-lock contention no longer becomes permanent task-binding failure. Input, window, freshness and cancellation gates remain. Decision/Luna thresholds, prompts and defaults are unchanged.
• New API/Agent environments default to Python 3.12, retaining supported 3.11 environments. Local-model setup remains 3.11; Amadeus and Runtime use separate environments. API vision needs no local GPU or model weights.
• Includes execution-only maintenance history: learning entry points are archived; shared execution dependencies remain.

Install: Runtime installer → prepare/start Runtime → import the 0.0.4 wheel into a compatible Amadeus → configure the local URL/token → bind MCP + Skill to Pi → enable, restart backend and create a new Work. Provider API keys stay in Runtime. Complete English instructions.

Evidence: candidate05 completed a direct five-step and native Amadeus/Pi four-step 7-Zip navigation workflow after an explicit TEST refusal pair and actual Main image review. These are assisted functional recovery demonstrations. The final installer adds documentation/SVG updates: 609 production/configuration files, including 582 Python modules, are byte-identical to the accepted candidate. All 676 final payload hashes, ZIP/embedded bytes, native check-install, isolated installed entrypoints and check-uninstall passed. The latest affected regression passed 188 tests on Python 3.11 and 188 on Python 3.12, including ten contention cases; overlapping suites are not added. Full delivery scope · Accepted recovery evidence · Final packaging report.

Limits: the installer is unsigned; the packaging-only rebuild is not a fresh desktop/model run or clean-machine/OS-integration qualification. Earlier failed attempts remain failures. Pi saw a verified native full-frame resized derivative, not original PNG bytes. Full Pi receipts remain token-heavy; no new speed, token-saving, cost or general accuracy improvement is claimed. The wheel/source ZIP preserve their tested build-time bytes; use the current setup guide. Store approval remains separate from artifact publication; manual wheel import is available.

简体中文

• 新增仅绑定本机、带鉴权的 Streamable HTTP MCP，保留 stdio。扩展提供 HTTP MCP＋Skill，Runtime 需单独安装、启动。
• 点击前双拒 → 原 Main 查看 before 绑定图 → 明确审核 → 单独继续原任务／候选。已结算文字输入仍用 after 图审核，已完成动作不重放。
• 修复临时队列锁竞争被误记为永久任务绑定失效；输入、窗口、新鲜度与取消门禁保留。Decision／Luna 阈值、提示词和默认方案未改。
• 新建 API／Agent 环境默认 Python 3.12，保留受支持的 3.11 环境；本地模型路线仍用 3.11。Amadeus 与 Runtime 的环境分开，外部视觉 API 不需本地 GPU 或模型权重。
• 同步执行模式维护历史：学习入口已封存，执行共享依赖继续保留。

安装顺序： Runtime 安装器 → 配置并启动 Runtime → 在兼容 Amadeus 中导入 0.0.4 wheel → 填写本机地址／连接 token → 将 MCP＋Skill 绑定 Pi → 启用、重启后端、新建 Work。模型 API Key 只保存在 Runtime 环境。完整中文安装说明。

验证范围： candidate05 的本体五步、原生 Amadeus／Pi 四步 7-Zip 导航流程均在显式 TEST 双拒及实际 Main 看图后完成，属于辅助恢复功能验收。最终安装器仅更新文档和 SVG；609 项生产／配置文件（含 582 个 Python 模块）与实测候选逐字节一致。最终 676 项载荷、ZIP／内嵌字节、原生 check-install、隔离安装态入口和 check-uninstall 均通过。最新相关回归在 Python 3.11 和 3.12 下各 188 项通过，已包含十项锁竞争用例，重叠分组不累加。验收、最终包身份及保留失败见上方完整记录。

限制： 安装器未签名，文档重封装不代表新增桌面／模型实测或全新机器／OS 集成验收。Pi 看到的是经核对的原生全帧缩放派生图，不是原 PNG 字节。完整回执仍消耗较多 token，本版不宣称新的提速、token 节省、费用或通用准确率提升。wheel／源码 ZIP 保留已测构建时字节，最新安装方法以上方指南为准。公开发布不等于商店审核通过；可先手动导入 wheel。

Downloads and verification / 下载与校验

• SHA256SUMS.txt
• Runtime source payload ZIP · payload manifest
• Exact extension source ZIP · author preflight · bounded public acceptance

Runtime installer SHA-256: 3c5594fa23d1522559f42de88ed17c841814b74c6aac61310527313c0a133287

Extension wheel SHA-256: 9acc7f5d1d667d1d44e3750cdf0f351abefeb5e5136b7fad91d50d77a6d29b31