test: freeze host-authoritative original Purge validation
This commit is contained in:
@@ -1,5 +1,7 @@
|
||||
# AF:开放模式下原版 Purge AB/BA
|
||||
|
||||
执行状态:AF已在唯一ab-off规划阶段因源码Node/正式Node路径身份不符而停止,0真实请求/0worker,原状态审计和71份原件封存通过。见[AF实际结果](QA-LIVE-PURGE-AF.md)。本计划和原冻结值保留,不续跑AF;后续仅用正式宿主检测API另冻[AG](LIVE-PURGE-AG-PLAN.md)。
|
||||
|
||||
前置AD Node和AE React/Python各自已封存并独立交付通过。原来的有限额度、网络、输出截断、未交付、重复读取与未知缓存/真实预留记录全部保持;不续跑旧样本或用户项目。正式1.25.8/a6e31dec/PID71628保持,本阶段不改产品或重打包。
|
||||
|
||||
冻结后按ab-off→ab-on→ba-on→ba-off串行,各新父项目、新子工作区和新会话。基线复制原Node费用三模块/原九测试,固定Git元数据令初始提交内容完全相同,不复用已有成功源码或对话。模型仅收一次相同业务目标和允许路径,自行读、改、plan_checks/run_checks;不提供实现、人工审批、继续或失败后重试。子工程保持原HEAD/index,父工程只按真实成果自动归集、合并、检查、源码限定Git并finalize,这些自动动作逐项保留,不能冒认模型自己运行了父流程。
|
||||
|
||||
@@ -0,0 +1,17 @@
|
||||
# AG:正式宿主身份下的原版 Purge AB/BA
|
||||
|
||||
AG是新批次,不续跑AF。前置AD Node、AE React/Python真实自主交付与独立核验已通过;AF在规划阶段0真实请求/0worker终止,原失败、审计和71份证据保持。本轮不改产品、不打包或升级正式1.25.8/a6e31dec/PID71628,不运行用户项目。
|
||||
|
||||
唯一实验准备变更是检测权威:冻结和逐组复核使用正式宿主的controlled-availability与purge/source API,不使用源码Node进程推断正式引擎身份。正式controlled SHA为571e9c8cfa30dca12ea5a914cf74a170c91a3bae9975c428fcafb72c8afa662a;Purge源b71679347cac069667d6cc63cb3b7736e69508eda58d50ee72746639bfc40fb0、运行输入e01d158017357cb87e233449f00828b771dad518a4118c420b6865ce15723de1。Node二进制SHA与源码环境相同,绝对路径不同仍必须按实际宿主身份核对,不放宽来源校验。
|
||||
|
||||
其余沿用[AF完整契约](LIVE-PURGE-AF-PLAN.md)和已提交scripts/live-af/contract.mjs:ab-off→ab-on→ba-on→ba-off,四个独立父项目/子工作区/会话、相同初始Git内容与原九测试、相同一次业务目标,DeepSeek/deepseek-flash、Chat/streaming、nativeParameters={}、同一临时require-match连接修订;关闭组仍使用相同Harness,开启组只增加固定可信原版profile/apply/postPrompt因素。60原文件、393运行输入和两个原生模块逐项核对。无实现提示、人工审批、继续、自动重试或切池;自动父归集/合并/检查/Git/finalize单列,不能归成模型亲自执行父流程。
|
||||
|
||||
unrestricted/full/passthrough/preserve与全部产品额度0保持;实际wire无输出上限、完整历史不裁剪。每例10分钟、连续3分钟无新读取/源码/检查事实作为共同外部观察终点,不改变产品默认。失败/unknown不计零成本,旧网络、型号、截断、重复读取和缓存未知不删除。驱动/身份/配置/证据问题停整个批次,先修复再另冻;模型/网络失败保留后可运行下一全新例,不重放失败。
|
||||
|
||||
成功需原九测试/build、仅三源码父Git且当前内容=提交、owner和非目标index/未跟踪保持、子HEAD/index保持、全部原生账本与所属进程清理、每请求身份和实际用量核对。四例同标准全部通过且真实工具schema相同时,才列两对顺序平衡的token/请求/耗时描述差异;样本量2对不作普遍或因果结论,无账单不编造费用。
|
||||
|
||||
先用正式Node绝对路径启动独立端口0/全新数据/自有固定HTTP的AG完整驱动,验证路径不同但SHA相同、四组工程和原状态恢复/封存;不触正式数据/真实Key。初次夹具变量遮蔽在宿主启动前失败保留。相关语法/diff、AG夹具和AF原封存逐SHA通过后本地提交,再唯一冻结.qa/live-purge-ag/current.json。前置夹具结果字段为driverFixture,隔离运行的fixture仅为true。
|
||||
|
||||
冻结记录同期Orven品牌/UI改动:src/client、public、docs/orven、output以及DESIGN.md/README.md/index.html/mobile.html/package.json/src/desktop/main.mjs,另含品牌图标生成器与两个界面验证器。已核对package仅description变化,desktop仅名称/标题/报错标题变化;这些未进入正式已安装的Node引擎和本轮驱动,不覆盖、不代交、不构建或部署。每组核对正式PID/instance/runtimeRoot/版本/配置/连接/来源;其它有效驱动/后端未提交则拒绝冻结。
|
||||
|
||||
结束恢复原高性能池与连接身份策略,单调修订声明;原表每行、全部导出/工件/模型尝试/凭据存在状态独立审计,全部新增父子归属核对、空闲/窗口0/9229关闭,再封存/回顾/本地提交。两处产品缺口(Python检查解释器、终端Git未登记delivery_job)另切修复,不在本次对照中改动。三系统接口反馈只读记录,不改变冻结条件。不委派、不推送,不读取/复制/解密/重存真实Key。
|
||||
@@ -1,5 +1,7 @@
|
||||
# 1.25 计划:任务要求与交付复核
|
||||
|
||||
AF在正式协作规划时因源码Node/正式Node路径参与身份摘要而被正确拒绝,0真实请求/0worker,原状态审计及71份原件封存保持。新AG以正式检测API为权威,完整隔离四组16自有HTTP、原九测试/父Git和恢复审计通过;下一本地提交、唯一冻结AG后执行原版Purge ABBA。产品未改、Purge收益仍unknown。见[AF结果](QA-LIVE-PURGE-AF.md)与[AG计划](LIVE-PURGE-AG-PLAN.md)。下方为历史接续。
|
||||
|
||||
AE新React/Python真实交付均已通过:20请求367365报告tokens/0unknown,原九测试/页面与原十pytest/编译、源码限定Git及owner/index独立通过,225份原证据封存;原48100行/116导出/工件保持。内置Python选错环境的失败与终端Git未登记内部交付账本保留,见[AE验收](QA-LIVE-ENGINEERING-AE.md)。下一先核对开放worker/engine契约,再另冻原版Purge AB/BA,收益仍unknown;不改变任何历史样本结论。
|
||||
|
||||
新AD已完成:五线十次真实预检8通过、两次Claude上游400缺必填max_tokens,目录两处网络异常保持;DeepSeek新Node自主11请求、原九测试/build/三源码实际Git交付成功,0人工审批/继续。全批21请求199250报告tokens+2unknown,216份原证据及原状态对账完成,见[AD验收](QA-LIVE-VALIDATION-AD.md)。本例30工具schema保持、7次不同文件读取及122624上游缓存输入tokens不作普遍收益;终端Git未登记内置交付账本的缺口单列。下一另冻AE React/Python,通过后才原版Purge AB/BA;历史失败/真实预留边界/Purge收益未知保持。
|
||||
|
||||
@@ -1,5 +1,13 @@
|
||||
# AF 原版Purge对照:准备与实际结果
|
||||
|
||||
AF已结束并封存,真实模型请求0、worker0。唯一ab-off在规划协作时被正式宿主拒绝:“受控引擎与刚才检测的程序/插件身份不一致”。没有续跑、重试或执行其余三组,不具备收益比较条件。下方准备记录保留原时点,不代表AF仍待启动。
|
||||
|
||||
原因已独立核对:源码驱动使用tools目录的Node,正式宿主使用安装包runtime/node.exe;二进制SHA相同,但绝对路径参与受控身份摘要。源码探测controlled=5f4cb8df、runtime=16e74cf9;正式API探测controlled=571e9c8c、runtime=e01d1580。桥与Purge源码相同,正式身份验证正确拒绝错误快照。解决办法是另冻AG、以正式宿主检测API作为权威;不削弱产品校验、不修改AF原输入。
|
||||
|
||||
原50837行、118导出、全部旧工件/模型尝试/凭据行保持;仅新增1自有父项目、1会话、1事件及outbox。DeepSeek15→17、config116→118为暂选/恢复的声明修订,原高性能池和其余配置恢复,正式PID71628/1.25.8空闲、窗口0、9229关闭。最终审计final-d2482bc7通过,71份30311984字节封存且再次逐SHA通过,见[live-purge-af-evidence.json](live-purge-af-evidence.json)。
|
||||
|
||||
封存后发现公开索引未生成:AF基线的fixture字段保存的是“前置夹具结果对象”,封存器误把它当“当前运行就是夹具”的布尔值。原summary和FINAL-MANIFEST均未改写;逐项离线核对后只补公开索引。新驱动仅把fixture===true作为隔离运行,AG的前置结果另存driverFixture。新AG隔离夹具第一次还因batch变量遮蔽触发TDZ,启动宿主前即退出,原ag-driver-cca2392e失败保留,真实模型请求0。
|
||||
|
||||
契约见[LIVE-PURGE-AF-PLAN.md](LIVE-PURGE-AF-PLAN.md)。AD Node和AE React/Python已各自独立交付封存,本阶段不改产品,正式仍1.25.8/a6e31dec/PID71628。真实对照尚未开始;未观察前收益unknown。
|
||||
|
||||
原生开放契约本机验证open-original-engine-13b6c268通过,两组各四次自有HTTP和12000068合成报告tokens、完整历史与无输出cap、原版入口/apply/两个原生模块、父工程检查/Git与owner/index/进程树退出均过,真实请求0。前两个夹具的驱动函数名错误和提交后误要求整个HEAD/index保持的断言失败保留。
|
||||
|
||||
@@ -0,0 +1,9 @@
|
||||
# AG 原版Purge对照:验证与回顾
|
||||
|
||||
计划见[LIVE-PURGE-AG-PLAN.md](LIVE-PURGE-AG-PLAN.md)。新批次以正式宿主检测API为身份权威,AF的0请求规划失败保持,不续跑。当前进行隔离完整驱动验证,真实AG尚未冻结或发请求,Purge工程收益仍unknown。
|
||||
|
||||
首ag-driver-cca2392e因同块batch变量遮蔽触发TDZ,在启动宿主前失败;已改为batchResult,原结果保留。封存器另修fixture结果对象被误当隔离布尔值的问题,AF原manifest不改,仅补逐SHA核对后的公开索引。后续实际结果另附,不能用夹具响应证明真实推理或收益。
|
||||
|
||||
完整隔离ag-driver-c5ff7602-52a8-440c-a24c-329d4c3257df通过:以正式Node绝对路径启动新端口/新数据宿主,实际检测路径与源码不同但二进制SHA相同;正式API权威贯穿四组,16自有HTTP、每组原九测试/build、父工程限定Git/owner/index、原版入口/apply/两个原生模块及进程清理均通过。原11行/1导出/工件/凭据状态保持,新增8父子项目/8会话/16尝试全部自有,最终审计9c83415c通过;477份19575831字节封存且逐SHA核对。真实模型请求0。
|
||||
|
||||
回顾:AF原因在不同Node绝对路径参与身份摘要;本次夹具复现同样环境差异并以正式检测API通过,不放宽产品核对。有效驱动语法和diff检查通过,AF公开索引离线恢复通过,正式程序未变。准备本地提交后唯一冻结AG,再运行一次四组真实对照。
|
||||
@@ -0,0 +1,27 @@
|
||||
# Workbench 三系统接口反馈
|
||||
|
||||
2026-11-06,只读接收Mnemo协调的三系统共识v1。三产品独立运营、部署、认证和账本;保留本产品全局开放偏好,以可替换接口关联。本轮保持AG原版Purge实验条件,不跨项目写库、切线、运行付费任务或群发消息。当前正式程序1.25.8,Orven品牌/UI在并行开发,不能把未提交界面当正式发行。
|
||||
|
||||
## 当前可提供及证据
|
||||
|
||||
本地HTTP接口已有项目/Session创建、messages启动、stop、Run/事件查询、会话export及工件原字节读取;export的schemaVersion/executionSchemaVersion均为1。Host/Origin与x-workbench-token属于本地访问机制,不等于已交付跨产品认证。检查、页面与Git工件绑定项目/Session/Task/Run和SHA。源码依据src/server/app.mjs;实际AD Node原九测试/build/Git、AE React九测试/类型/构建/1280及390页面和Python十pytest/compileall见[AD](QA-LIVE-VALIDATION-AD.md)、[AE](QA-LIVE-ENGINEERING-AE.md)。这些是既有封存结果,本次未重跑。
|
||||
|
||||
方向须分开:现有外部Harness桥是Workbench作为调用方,启动Harness0.2.1-rc.2的SDK/所属进程,关联工具和模型桥、报告与停止。输入/报告schemaVersion=1不表示双方有完整协议版本协商;SDK serverInfo和发行包版本独立核对。原版Purge固定来源profile可加载,不等于工程收益已证明。认知引擎作为调用方委托Workbench尚缺公开稳定的版本化/幂等入站适配,不应把现有引擎出站接口反过来冒充已实现的入站服务。
|
||||
|
||||
## 需要两方提供
|
||||
|
||||
认知侧提供有版本和授权范围的subject/goal/job及记忆快照、来源修订/失效事件、稳定idempotencyKey与请求规范摘要、验收要求,以及接收原始证据候选的接口。正式记忆由认知侧核对,不把模型总结或工程笔记自动变为事实。
|
||||
|
||||
智枢提供版本化模型目录/能力与逐路由Chat/Responses/Messages契约、推理参数、请求/实际返回型号、routeRevision、同一物理请求trace及原始usage/unknown、查询/取消结果和上游原错。缓存tokens按上游报告保存;模型或路由名称不能代替兼容与质量验证。三方各留账本,以同一请求关联,不能重复计作费用。
|
||||
|
||||
## 已知差异与阻塞
|
||||
|
||||
Python内置plan_checks/run_checks在AE选中无pytest的Hermes解释器,原失败保持;模型自行以指定Python终端验证成功,不能倒填check_job。后续入站契约需解释器绝对路径/版本/依赖及检查输入身份。终端Git提交未登记内部delivery_job,delivery_report可能仍in_progress;应核对实际commit/parent/文件范围/源码SHA/检查证据,报告缺失不允许重复提交,产品登记衔接另切修复。
|
||||
|
||||
目录网络中断、Claude Messages省略必填max_tokens的原HTTP400保持;完整历史/默认无本地输出cap不能被适配器悄悄改变。Purge收益仍unknown;当前AF在真实调用前因Node路径身份不同正确被拒绝,AG另冻验证,不重写AF。现行本地REST不具备已验收的远程多租户、入站idempotencyKey/事务回放或中央费用桥接,不能宣称三套已连通。
|
||||
|
||||
## 最小接入与验收
|
||||
|
||||
先离线只读消费一份已完成工程的脱敏回执:来源版本、请求摘要、项目/Session/Task/Run、检查/页面/Git工件ID及SHA全部可核对,导入认知侧为候选,0模型请求/0工程写入。先定义engineering-delegation/v1草案(这是建议的新契约,尚未实现),再用全新合成工程和自有固定HTTP验证入站启动、原游标查询和停止。
|
||||
|
||||
验收覆盖同key同参返回同任务、同key异参拒绝、响应丢失先查询不重发、重启恢复、外部源码漂移、失败/unknown、取消接收与所属进程确已退出分开、已有Git但内部报告缺失。一次委托只产生一次允许写入/一次提交,实际测试/页面与当前commit绑定。委托若要求额外范围或费用条件而通道不能兑现,应在启动前说明;不修改用户全局权限或截断历史来伪称已满足。以上通过后再另冻真实调用与质量样本,三方原质量失败和Purge unknown不因接口通过而关闭。
|
||||
@@ -1,5 +1,7 @@
|
||||
# Tianshu 研究后的 Agent Workbench 采用计划
|
||||
|
||||
AF在正式协作规划时因源码Node/正式Node路径参与身份摘要而被正确拒绝,0真实请求/0worker,原状态审计及71份原件封存保持。新AG以正式检测API为权威,完整隔离四组16自有HTTP、原九测试/父Git和恢复审计通过;下一本地提交、唯一冻结AG后执行原版Purge ABBA。产品未改、Purge收益仍unknown。见[AF结果](QA-LIVE-PURGE-AF.md)与[AG计划](LIVE-PURGE-AG-PLAN.md)。下方为历史接续。
|
||||
|
||||
AE新React/Python真实交付均已通过:20请求367365报告tokens/0unknown,原九测试/页面与原十pytest/编译、源码限定Git及owner/index独立通过,225份原证据封存;原48100行/116导出/工件保持。内置Python选错环境的失败与终端Git未登记内部交付账本保留,见[AE验收](QA-LIVE-ENGINEERING-AE.md)。下一先核对开放worker/engine契约,再另冻原版Purge AB/BA,收益仍unknown;不改变任何历史样本结论。
|
||||
|
||||
新AD已完成:五线十次真实预检8通过、两次Claude上游400缺必填max_tokens,目录两处网络异常保持;DeepSeek新Node自主11请求、原九测试/build/三源码实际Git交付成功,0人工审批/继续。全批21请求199250报告tokens+2unknown,216份原证据及原状态对账完成,见[AD验收](QA-LIVE-VALIDATION-AD.md)。本例30工具schema保持、7次不同文件读取及122624上游缓存输入tokens不作普遍收益;终端Git未登记内置交付账本的缺口单列。下一另冻AE React/Python,通过后才原版Purge AB/BA;历史失败/真实预留边界/Purge收益未知保持。
|
||||
|
||||
@@ -0,0 +1,110 @@
|
||||
{
|
||||
"summary": {
|
||||
"schema": 1,
|
||||
"version": "1.25.8",
|
||||
"driverCommit": "01e64f3fcfb062eed92cb6e4260672b16e531146",
|
||||
"samples": [
|
||||
{
|
||||
"sample": "ab-off",
|
||||
"enabled": false,
|
||||
"success": false,
|
||||
"failure": {
|
||||
"error": "Error: /sessions/b5e47756-36dd-4ac6-9a37-efec00c9ea15/collaborations: 受控引擎与刚才检测的程序/插件身份不一致,请重新检测后制定新计划。\n at api (file:///D:/Ai-project/REA%E9%80%86%E5%90%91/Agent-Workbench/scripts/live-ad/common.mjs:64:43)\n at process.processTicksAndRejections (node:internal/process/task_queues:104:5)\n at async file:///D:/Ai-project/REA%E9%80%86%E5%90%91/Agent-Workbench/scripts/live-af/run.mjs:100:23",
|
||||
"retry": false
|
||||
},
|
||||
"parentSessionId": "b5e47756-36dd-4ac6-9a37-efec00c9ea15",
|
||||
"requests": 0,
|
||||
"usage": {
|
||||
"requests": 0,
|
||||
"inputTokens": 0,
|
||||
"outputTokens": 0,
|
||||
"totalTokens": 0,
|
||||
"unknownRequests": 0,
|
||||
"estimatedUnknownTokens": 0
|
||||
},
|
||||
"dispatchStage": "before-worker-and-real-model",
|
||||
"originalPurge": []
|
||||
}
|
||||
],
|
||||
"requests": 0,
|
||||
"reportedRequests": 0,
|
||||
"reportedTokens": 0,
|
||||
"unknownRequests": 0,
|
||||
"unknownLocalEstimate": 0,
|
||||
"actualToolSha256": [],
|
||||
"sameTools": false,
|
||||
"comparison": {
|
||||
"eligible": false,
|
||||
"reason": "Four equal-success samples with stable tools are required; early failure is not token savings."
|
||||
},
|
||||
"originalStateAudit": "audits/final-d2482bc7-d9a9-4b32-a0c6-f4b4ee07b724/result.json",
|
||||
"originalStateAuditResult": {
|
||||
"success": true,
|
||||
"version": "1.25.8",
|
||||
"pid": 71628,
|
||||
"originalProjects": 93,
|
||||
"originalSessions": 118,
|
||||
"addedProjects": 1,
|
||||
"addedSessions": 1,
|
||||
"checkedOriginalRows": 50837,
|
||||
"addedRows": {
|
||||
"events": 1,
|
||||
"metadata": 0,
|
||||
"outbox": 1,
|
||||
"projects": 1,
|
||||
"records": 0,
|
||||
"sessions": 1,
|
||||
"sqlite_sequence": 0
|
||||
},
|
||||
"newArtifactCount": 0,
|
||||
"originalExports": "exact",
|
||||
"originalArtifactBytes": "exact",
|
||||
"originalCredentialRow": "exact hash; no decryption or copy",
|
||||
"allOldModelAttempts": "exact",
|
||||
"newModelAttempts": 0,
|
||||
"configComparison": {
|
||||
"original": 116,
|
||||
"current": 118,
|
||||
"allOtherFields": "exact"
|
||||
},
|
||||
"connectionChanges": [
|
||||
{
|
||||
"id": "1a7a55a9-29f2-4b9f-8ea2-5a55afd7eacd",
|
||||
"originalRevision": 15,
|
||||
"currentRevision": 17,
|
||||
"originalUpdatedAt": "2026-11-06T05:05:04.685Z",
|
||||
"currentUpdatedAt": "2026-11-06T05:42:01.952Z"
|
||||
}
|
||||
],
|
||||
"connectionOrder": {
|
||||
"before": [
|
||||
"5b4daed9-b2f7-4de0-b949-573eba70c7b3",
|
||||
"93d0cf6d-4f32-4918-9da8-6f125aa8a6e5",
|
||||
"09b0293b-faad-45c3-832d-7c8b615813aa",
|
||||
"9ae348a5-d0a6-42ed-8664-0da5f9effe34",
|
||||
"1a7a55a9-29f2-4b9f-8ea2-5a55afd7eacd"
|
||||
],
|
||||
"after": [
|
||||
"5b4daed9-b2f7-4de0-b949-573eba70c7b3",
|
||||
"93d0cf6d-4f32-4918-9da8-6f125aa8a6e5",
|
||||
"09b0293b-faad-45c3-832d-7c8b615813aa",
|
||||
"9ae348a5-d0a6-42ed-8664-0da5f9effe34",
|
||||
"1a7a55a9-29f2-4b9f-8ea2-5a55afd7eacd"
|
||||
]
|
||||
},
|
||||
"originalRows": "exact except declared project positions, sequence and validated public metadata",
|
||||
"keysReadOrResaved": false,
|
||||
"inspectorClosed": true
|
||||
},
|
||||
"keysReadOrResaved": false,
|
||||
"historicalSamplesReplayed": false,
|
||||
"priorFailedExperimentsRemainUnchanged": true
|
||||
},
|
||||
"evidence": {
|
||||
"root": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\live-purge-af",
|
||||
"files": 71,
|
||||
"bytes": 30311984,
|
||||
"manifestSha256": "9ad1ba3365e4927ab0c35422452e8387aafb4338a14b64eba1ad86baa1cceec5"
|
||||
},
|
||||
"indexRecovery": "Offline index only: the immutable AF baseline used fixture for the prior local validation result, not the isolated-fixture boolean; its original summary and manifest are unchanged."
|
||||
}
|
||||
@@ -1306,3 +1306,5 @@ Node新样本已真实交付:11请求180554tokens,7次读取分别属于不
|
||||
AE两工程均真实自主交付通过,但Python内置检查选中了Hermes的无pytest环境,模型自主使用目标中现成解释器的终端验证才通过。目标文字没有成为检查环境的结构化选择,不能把终端绿测补写成原check_job通过。React/Python均未登记内部delivery_job;报告必须区分真实Git与内部登记的证据。这两处需独立产品修复,不能在Purge对照中更换实现。
|
||||
|
||||
原版Harness开放模式由collaborationInput将四项budget规范化为0;worker configFor跳过有限额度计算,engine credit以Infinity处理,capacityMode=passthrough不投影裁剪也不发输出cap。本机原生off/on各四请求和12000068合成tokens后继续完成父检查/Git,入口/apply、两个原生模块与进程树退出通过。夹具证明执行契约,不能代替真实自主交付/Purge收益。
|
||||
|
||||
AF正式规划零请求即拒绝,原因是Node绝对路径进入controlled/runtime身份:源码与正式二进制SHA相同也不代表身份相同。新AG必须以正式检测API为权威,隔离宿主用正式Node路径的四组夹具已通过。AF封存器把前置fixture结果对象误作隔离运行布尔值,原manifest已完整而公开索引缺失;离线逐SHA后仅补索引,新版仅fixture===true跳过公开索引。
|
||||
|
||||
@@ -1944,3 +1944,5 @@ AD现已完成并封存216份60197251字节。真实兼容8/10,jojocode/Claude
|
||||
AE已封存225份52216677字节:React/Python均10请求、合计367365tokens/0unknown,自主原检查/实际页面及限定Git/owner/index独立通过。Python原内置检查误选无pytest的Hermes环境,失败保留,模型自主用目标提供的现成解释器验证成功;两例内部交付job缺失保持。最终审计原48100行/116导出/工件精确保留,新增2项目2会话均归属AE,原池与配置恢复、正式空闲/9229关闭。下一先本机核对开放原版引擎,再另冻PurgeAB/BA。
|
||||
|
||||
AF准备完成:原生开放off/on四请求12000068合成tokens仍继续完成父交付;完整af-driver-c6ca1f0f四组16本机HTTP、原九测试/build/父Git/owner/index和原版入口/apply/模块/进程树全部通过,原状态审计及477份证据封存通过。旧函数名、Git断言、route-plan取证、null元数据和省略身份策略恢复失败全部保留。下一本地提交并唯一冻结AF真实ABBA;本机固定响应不算真实自主交付/Purge收益,不改产品和默认额度。
|
||||
|
||||
AF正式协作规划因源码Node/正式Node路径身份不同而拒绝,0真实请求/0worker;原50837行/118导出/全部工件和旧尝试保持,新增1自有父项目/会话/事件,池与配置恢复。71份30311984字节封存逐SHA通过,原manifest不改、公开索引离线补齐。AG另冻不续跑AF:首夹具TDZ失败保留,完整c5ff7602四组16自有HTTP、原九测试/build/父Git/原生与原状态审计通过,477份19575831字节封存,覆盖正式Node路径不同但二进制SHA相同。三系统接口只读反馈已记录,未扩大运行授权。下一本地提交后唯一冻结AG,真实收益仍unknown。
|
||||
|
||||
@@ -8,10 +8,10 @@ import { Agent, fetch } from 'undici';
|
||||
export let root = path.resolve('.qa/live-validation-ad'), origin = 'http://127.0.0.1:4318';
|
||||
export function engineeringEvidence() { root = path.resolve('.qa/live-engineering-ae'); }
|
||||
let purgeFixture;
|
||||
export function purgeEvidence() { root = purgeFixture?.root ?? path.resolve('.qa/live-purge-af'); if (purgeFixture) origin = purgeFixture.origin; }
|
||||
export function purgeEvidence() { root = purgeFixture?.root ?? path.resolve(process.argv.includes('--batch=ag') ? '.qa/live-purge-ag' : '.qa/live-purge-af'); if (purgeFixture) origin = purgeFixture.origin; }
|
||||
export function configurePurgeFixture(options) {
|
||||
const target = path.resolve(options.root), url = new URL(options.origin);
|
||||
assert.equal(path.dirname(target), path.resolve('.qa')); assert.match(path.basename(target), /^af-driver-[a-f0-9-]{36}$/);
|
||||
assert.equal(path.dirname(target), path.resolve('.qa')); assert.match(path.basename(target), /^a[fg]-driver-[a-f0-9-]{36}$/);
|
||||
assert.equal(url.hostname, '127.0.0.1'); assert.notEqual(url.port, '4318'); assert.ok(url.port);
|
||||
purgeFixture = { root: target, origin: url.origin }; purgeEvidence();
|
||||
}
|
||||
|
||||
@@ -80,8 +80,12 @@ try {
|
||||
const store = (file, value) => save(prefix + '/' + file, value), startedAt = Date.now();
|
||||
let own, session, worker, group, workerState, failure, result, route;
|
||||
try {
|
||||
const installation = resolveControlledInstallation(process.cwd(), frozen.installation.root);
|
||||
const source = resolveOriginalPurgeSource(process.cwd(), frozen.source.root, installation);
|
||||
const installation = frozen.availabilityAuthority === 'host-api'
|
||||
? await api('/engines/deepseek/controlled-availability?root=' + encodeURIComponent(frozen.installation.root))
|
||||
: resolveControlledInstallation(process.cwd(), frozen.installation.root);
|
||||
const source = frozen.availabilityAuthority === 'host-api'
|
||||
? await api('/engines/purge/source?kind=' + ORIGINAL_PURGE_KIND + '&root=' + encodeURIComponent(frozen.source.root) + '&harnessRoot=' + encodeURIComponent(installation.root))
|
||||
: resolveOriginalPurgeSource(process.cwd(), frozen.source.root, installation);
|
||||
assert.equal(installation.controlledSha256, frozen.installation.controlledSha256);
|
||||
assert.equal(source.sourceSha256, frozen.source.sourceSha256); assert.equal(source.runtimeSha256, frozen.source.runtimeSha256);
|
||||
const current = await api('/model-center');
|
||||
|
||||
@@ -8,6 +8,8 @@ const auditPath = process.argv[2]; assert.ok(auditPath, 'Explicit completed audi
|
||||
const audit = read(auditPath); assert.equal(audit.success, true);
|
||||
const frozen = read('current.json'), result = read('result.json'), samples = [], allAttempts = [], tools = new Set();
|
||||
for (const sample of result.samples) {
|
||||
if (!sample.workerSessionId) { samples.push({ ...sample, requests: 0, usage: { requests: 0, inputTokens: 0, outputTokens: 0, totalTokens: 0,
|
||||
unknownRequests: 0, estimatedUnknownTokens: 0 }, dispatchStage: 'before-worker-and-real-model', originalPurge: [] }); continue; }
|
||||
const prefix = sample.sample, state = read(prefix + '/worker/session.json'), attempts = read(prefix + '/worker/attempts.json');
|
||||
const jobs = read(prefix + '/worker/engine-jobs.json'), proof = read(prefix + '/worker/engine-proof.json');
|
||||
assert.equal(attempts.length, state.execution.task.usage.requests);
|
||||
@@ -68,9 +70,9 @@ function walk(directory) {
|
||||
}
|
||||
}
|
||||
walk(root); files.sort((a, b) => a.path.localeCompare(b.path)); const bytes = files.reduce((sum, row) => sum + row.bytes, 0);
|
||||
save('FINAL-MANIFEST.json', { schema: 1, files, bytes, scope: 'All AF raw requests/responses, original engine inputs/reports, source/Git, checks and audits. Historical evidence is not repackaged.' });
|
||||
save('FINAL-MANIFEST.json', { schema: 1, files, bytes, scope: 'All ' + (frozen.batch ?? 'af').toUpperCase() + ' raw requests/responses, original engine inputs/reports, source/Git, checks and audits. Historical evidence is not repackaged.' });
|
||||
for (const row of files) { const value = fs.readFileSync(path.join(root, row.path)); assert.equal(value.length, row.bytes); assert.equal(sha(value), row.sha256); }
|
||||
if (!frozen.fixture) fs.writeFileSync('docs/live-purge-af-evidence.json', JSON.stringify({ summary, evidence: { root, files: files.length, bytes,
|
||||
if (frozen.fixture !== true) fs.writeFileSync('docs/live-purge-' + (frozen.batch ?? 'af') + '-evidence.json', JSON.stringify({ summary, evidence: { root, files: files.length, bytes,
|
||||
manifestSha256: sha(fs.readFileSync(path.join(root, 'FINAL-MANIFEST.json'))) } }, null, 2) + '\n', { flag: 'wx' });
|
||||
console.log(JSON.stringify({ success: true, realRequests: summary.requests, reportedTokens: summary.reportedTokens, unknown: summary.unknownRequests,
|
||||
allFourSuccessful: fourSuccessful, sameTools, comparison, files: files.length, bytes }));
|
||||
|
||||
@@ -12,7 +12,8 @@ import { resolveControlledInstallation } from '../../src/core/engines/controlled
|
||||
import { resolveOriginalPurgeSource } from '../../src/core/engines/purge-strategy.mjs';
|
||||
import { allowed } from './contract.mjs';
|
||||
|
||||
const output = path.resolve('.qa/af-driver-' + randomUUID()), disposers = [], t = { after: fn => disposers.push(fn) };
|
||||
const batch = process.argv.includes('--batch=ag') ? 'ag' : 'af';
|
||||
const output = path.resolve('.qa/' + batch + '-driver-' + randomUUID()), disposers = [], t = { after: fn => disposers.push(fn) };
|
||||
fs.mkdirSync(output); const f = fixture(t, { modernDefaults: true });
|
||||
f.store.saveConfig({ approvalMode: 'full', mode: 'compatible' });
|
||||
// Model the existing production metadata shape before freezing this fresh fixture.
|
||||
@@ -37,17 +38,23 @@ try {
|
||||
assert.ok(requests <= 16); console.log(JSON.stringify({ fixtureRequest: requests, step }));
|
||||
return { ...completion(calls, calls.length ? '' : 'Owned fixture done.'), model: body.model };
|
||||
});
|
||||
const host = await launchWorkbench(t, f.data), originalCenter = await host.api('/model-center');
|
||||
const hostNode = batch === 'ag' ? JSON.parse(fs.readFileSync('.qa/live-purge-af/host-availability.json')).installation.node.path : process.execPath;
|
||||
const host = await launchWorkbench(t, f.data, {}, { nodePath: hostNode }), originalCenter = await host.api('/model-center');
|
||||
const connection = await host.api('/model-center/connections', { name: 'Owned AF local fixture', brand: 'Owned fixture', model: 'owned-af', protocol: 'chat-completions',
|
||||
baseUrl: model.baseUrl, apiKey: 'owned-local-fixture-only', streaming: false, nativeParameters: {} });
|
||||
const boot = await host.api('/bootstrap'), center = await host.api('/model-center'), health = await host.api('/host/status');
|
||||
configurePurgeFixture({ root: output, origin: host.url });
|
||||
const installation = resolveControlledInstallation(process.cwd()), source = resolveOriginalPurgeSource(process.cwd(), undefined, installation);
|
||||
const localInstallation = resolveControlledInstallation(process.cwd());
|
||||
const installation = batch === 'ag' ? await host.api('/engines/deepseek/controlled-availability?root=' + encodeURIComponent(localInstallation.root)) : localInstallation;
|
||||
const source = batch === 'ag' ? await host.api('/engines/purge/source?kind=purge-original-restricted-v1&harnessRoot=' + encodeURIComponent(installation.root))
|
||||
: resolveOriginalPurgeSource(process.cwd(), undefined, installation);
|
||||
if (batch === 'ag') { assert.notEqual(installation.node.path, localInstallation.node.path); assert.equal(installation.node.sha256, localInstallation.node.sha256); }
|
||||
save('current.json', { host: health, config: boot.config, center, connection, originalProjects: boot.projects,
|
||||
originalPublicMetadata: publicMetadata(f.data), baseline: await publicSnapshot(host.url, f.data), installation, source, fixture: true });
|
||||
await import('./run.mjs'); const batch = JSON.parse(fs.readFileSync(path.join(output, 'result.json')));
|
||||
assert.equal(batch.success, true, JSON.stringify(batch.samples.map(row => ({ sample: row.sample, failure: row.failure, cleanupErrors: row.cleanupErrors }))));
|
||||
assert.equal(requests, 16); assert.equal(batch.samples.length, 4); assert.ok(batch.samples.every(row => row.final.test.stats.pass === 9));
|
||||
originalPublicMetadata: publicMetadata(f.data), baseline: await publicSnapshot(host.url, f.data), installation, source, fixture: true,
|
||||
batch, ...(batch === 'ag' ? { availabilityAuthority: 'host-api', localInstallation } : {}) });
|
||||
await import('./run.mjs'); const batchResult = JSON.parse(fs.readFileSync(path.join(output, 'result.json')));
|
||||
assert.equal(batchResult.success, true, JSON.stringify(batchResult.samples.map(row => ({ sample: row.sample, failure: row.failure, cleanupErrors: row.cleanupErrors }))));
|
||||
assert.equal(requests, 16); assert.equal(batchResult.samples.length, 4); assert.ok(batchResult.samples.every(row => row.final.test.stats.pass === 9));
|
||||
await import('./finish.mjs');
|
||||
const audits = fs.readdirSync(path.join(output, 'audits')); assert.equal(audits.length, 1);
|
||||
const auditPath = 'audits/' + audits[0] + '/result.json';
|
||||
@@ -56,7 +63,8 @@ try {
|
||||
try { await import('./seal.mjs'); } finally { process.argv[2] = oldArgument; }
|
||||
const after = await host.api('/model-center'); assert.deepEqual(after.selected, originalCenter.selected);
|
||||
result = { success: true, realModelRequests: 0, fixedHttpRequests: requests, actualBatchDriverExecuted: true,
|
||||
originalTestsPerSample: 9, completeSamples: 4, originalStateAuditPassed: true, implementationSuppliedToLocalFixtureOnly: true };
|
||||
originalTestsPerSample: 9, completeSamples: 4, originalStateAuditPassed: true, implementationSuppliedToLocalFixtureOnly: true,
|
||||
distinctHostExecutablePathVerified: batch === 'ag', availabilityAuthority: batch === 'ag' ? 'host-api' : 'source' };
|
||||
} catch (error) { result = { success: false, error: error.stack, realModelRequests: 0 }; process.exitCode = 1; }
|
||||
finally { for (const dispose of disposers.reverse()) await dispose(); fs.writeFileSync(path.join(output, 'driver-result.json'), JSON.stringify(result, null, 2)); }
|
||||
console.log(JSON.stringify({ output, ...result }));
|
||||
|
||||
@@ -0,0 +1,43 @@
|
||||
import fs from 'node:fs';
|
||||
import assert from 'node:assert/strict';
|
||||
import { execFileSync } from 'node:child_process';
|
||||
import { purgeEvidence, root, save, client, publicMetadata } from '../live-ad/common.mjs';
|
||||
import { publicSnapshot } from '../retained-host-state.mjs';
|
||||
import { inspectorClosed } from '../live-ad/inspector.mjs';
|
||||
import { order, allowed, goal, purgePrompt, observations } from '../live-af/contract.mjs';
|
||||
|
||||
assert.ok(process.argv.includes('--batch=ag'));
|
||||
const af = JSON.parse(fs.readFileSync('.qa/live-purge-af/summary.json'));
|
||||
assert.equal(af.requests, 0); assert.equal(af.originalStateAuditResult.success, true);
|
||||
const ae = JSON.parse(fs.readFileSync('.qa/live-engineering-ae/summary.json'));
|
||||
assert.ok(ae.samples.length === 2 && ae.samples.every(row => row.success));
|
||||
const fixturePath = process.argv[2], fixture = JSON.parse(fs.readFileSync(fixturePath));
|
||||
for (const key of ['success', 'originalStateAuditPassed', 'actualBatchDriverExecuted', 'distinctHostExecutablePathVerified']) assert.equal(fixture[key], true);
|
||||
assert.equal(fixture.realModelRequests, 0); assert.equal(fixture.completeSamples, 4); assert.equal(fixture.availabilityAuthority, 'host-api');
|
||||
purgeEvidence(); assert.equal(fs.existsSync(root), false, 'Never overwrite or continue an observed AG batch.');
|
||||
const git = args => execFileSync('git', args, { encoding: 'utf8', windowsHide: true });
|
||||
const workingChanges = { tracked: git(['diff', '--name-only', 'HEAD', '-z']).split('\0').filter(Boolean), untracked: git(['ls-files', '--others', '--exclude-standard', '-z']).split('\0').filter(Boolean) };
|
||||
const parallelBrandingPaths = ['DESIGN.md', 'README.md', 'index.html', 'mobile.html', 'package.json', 'src/desktop/main.mjs',
|
||||
'scripts/generate-orven-icons.mjs', 'scripts/verify-orven-ui.mjs', 'scripts/verify-model-center-ui.mjs'];
|
||||
for (const name of [...workingChanges.tracked, ...workingChanges.untracked]) assert.ok(parallelBrandingPaths.includes(name)
|
||||
|| ['src/client/', 'public/', 'docs/orven/', 'output/'].some(prefix => name.startsWith(prefix)), 'Effective uncommitted source: ' + name);
|
||||
const { api, boot, close } = await client();
|
||||
try {
|
||||
await inspectorClosed(); const host = await api('/host/status'), center = await api('/model-center');
|
||||
assert.equal(boot.version, '1.25.8'); assert.equal(host.activity.active, false); assert.equal(host.desktopWindows.count, 0);
|
||||
for (const [key, value] of Object.entries({ executionMode: 'unrestricted', approvalMode: 'full', capacityMode: 'passthrough', historyPolicy: 'preserve',
|
||||
maxSteps: 0, maxRunTokens: 0, maxTaskTokens: 0, maxRunDurationMs: 0 })) assert.equal(boot.config[key], value);
|
||||
const connection = center.connections.find(row => row.id === '1a7a55a9-29f2-4b9f-8ea2-5a55afd7eacd');
|
||||
assert.equal(connection.enabled, true); assert.equal(connection.hasKey, true); assert.equal(connection.model, 'deepseek-flash');
|
||||
assert.equal(connection.credentialRevision, 1); assert.deepEqual(connection.nativeParameters, {});
|
||||
const installation = await api('/engines/deepseek/controlled-availability'), source = await api('/engines/purge/source?kind=purge-original-restricted-v1&harnessRoot=' + encodeURIComponent(installation.root));
|
||||
assert.equal(installation.controlledStatus, 'available'); assert.equal(source.status, 'available');
|
||||
const prior = JSON.parse(fs.readFileSync('.qa/live-purge-af/host-availability.json'));
|
||||
assert.equal(installation.controlledSha256, prior.installation.controlledSha256); assert.equal(source.sourceSha256, prior.source.sourceSha256); assert.equal(source.runtimeSha256, prior.source.runtimeSha256);
|
||||
const baseline = await publicSnapshot('http://127.0.0.1:4318', host.identity.dataRoot);
|
||||
save('current.json', { batch: 'ag', availabilityAuthority: 'host-api', host, config: boot.config, center, connection, originalProjects: boot.projects,
|
||||
originalPublicMetadata: publicMetadata(host.identity.dataRoot), baseline, installation, source, order, allowed, goal, purgePrompt, observations,
|
||||
fixturePath, driverFixture: fixture, workingChanges, previousBatch: { id: 'AF', requests: 0, outcome: 'Host rejected source-runtime identity before model dispatch; sealed unchanged.' }, driverCommit: git(['rev-parse', 'HEAD']).trim() });
|
||||
console.log(JSON.stringify({ frozen: true, batch: 'AG', requests: 0, version: boot.version, pid: host.identity.pid, projects: baseline.projects,
|
||||
sessions: baseline.database.sessionIds.length, connectionRevision: connection.revision, controlledSha256: installation.controlledSha256, runtimeSha256: source.runtimeSha256 }));
|
||||
} finally { await close(); }
|
||||
+4
-1
@@ -1,5 +1,7 @@
|
||||
# Agent Workbench:核心代理能力研究与产品计划
|
||||
|
||||
当前阶段AG:AD Node与AE React/Python真实自主交付及原证据封存已完成;AF在正式规划时因源码/宿主Node路径身份不符而0请求终止,71份原件及原状态审计封存。AG完整隔离驱动c5ff7602四组16自有HTTP/原九测试/父Git与恢复审计通过,正式Node路径差异已覆盖。下一提交驱动和计划后唯一冻结AG、一次执行原版Purge ABBA并审计封存;收益仍unknown。不改正式产品、不重放AF/旧样本,不代交并行品牌/UI改动。见docs/QA-LIVE-PURGE-AG.md。
|
||||
|
||||
## 当前接续:C 工具可见性、前缀证据与新真实验证(进行中)
|
||||
|
||||
基线 b14e26e6,feat/task-agreement-1.25,起步工作树干净。正式 1.25.7 / PID71628;用户当前选择「高性能」池 revision8,保留此新配置,不沿用旧 DeepSeek 单线默认。用户授权开始 C → 新供应商预检 → 新 Node → 有条件 React/Python/Purge;不续跑历史样本或用户项目、不委派、不推送、不操作真实 Key。
|
||||
@@ -1017,4 +1019,5 @@ F收尾:57d075e修正独立编译QA路径,6d09ac8封存差异;202610302320
|
||||
3. [complete] 唯一正式安装319241ms,保留原PID/Provider/连接、85项目105会话;正式网页/Electron回看后原状态再次exact,83份证据索引与发行身份封存。
|
||||
4. [complete] AD驱动/真实兼容与独立Node已封存:21请求199250报告tokens+2unknown,Node11请求9/9/build/Git88bb25cb及owner/index通过。
|
||||
5. [complete] AE React/Python均自主交付,原九/十测试、类型/构建/页面、限定Git与owner/index通过;20请求367365tokens/0unknown,225份证据和原数据审计封存。
|
||||
6. [in_progress] 两工程前置已通过;先验证开放worker/engine的0不限契约,再另冻原版Purge AB/BA,真实收益仍unknown。
|
||||
6. [complete] 开放worker/engine的0不限契约与完整四组本机驱动已通过;AF正式规划失败0请求,原失败、71份封存与审计保持。
|
||||
7. [in_progress] 新AG正式宿主API身份完整夹具通过;提交并唯一冻结后执行原版Purge AB/BA,真实收益仍unknown。
|
||||
|
||||
Reference in New Issue
Block a user