From 5ca65a45fb48d654f9235c409d1ab9ee771089c9 Mon Sep 17 00:00:00 2001 From: Agent Workbench Dev Date: Thu, 12 Nov 2026 01:01:38 +0800 Subject: [PATCH] test(claude): retain supplier restriction and corrected catalogue evidence --- AGENTS.md | 2 + docs/DEVELOPMENT-ROADMAP-OFFICIAL-FIRST.md | 2 + docs/M4-CLAUDE-CORRECTION-PLAN-2026-11-12.md | 2 + docs/PLAN-1.25.md | 2 + docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md | 34 ++ docs/TIANSHU-ADOPTION-PLAN.md | 2 + docs/claude-proxy-correction-2026-11-12.json | 422 ++++++++++++++++++ ...aude-proxy-correction-seal-2026-11-12.json | 28 ++ findings.md | 2 + progress.md | 2 + task_plan.md | 2 + 11 files changed, 500 insertions(+) create mode 100644 docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md create mode 100644 docs/claude-proxy-correction-2026-11-12.json create mode 100644 docs/claude-proxy-correction-seal-2026-11-12.json diff --git a/AGENTS.md b/AGENTS.md index 3a2b701f..80212fea 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -1,5 +1,7 @@ # Agent Workbench项目执行规范 +2026-11-12 修后Claude独立两预检结束:目录完整14型号/选定在内;两新请求均上游403 permission_error“分组仅允许Claude Code官方客户端”,无响应型号/SSE/usage,用量/费用null、工具执行0。见[修后验收](docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md)。原58298行/137导出/418请求/5576工件及当前七连接/池/选择/配置/凭据行SHA一次exact,41成员36284086字节新封存SHA9b8a1cfc通过;原M4的40请求/541清单保持。不追加或重派两请求,不回退人类更正或冒充Claude Code,供应商须另开放标准Messages客户端。当前验收提交/普通推送收尾,产品未改不重发行/重启;M8原版Purge另冻,全部历史unknown保留。 + 2026-11-12 M4收尾提交e132601b和此前六条已普通增量推送,远端HEAD/重新fetch七条逐SHA一致,见[推送回执](docs/gitatlas-m4-live-push.json)。回执与完成接续最后另提交并核远端;原40请求/541封存保持,1.25.13未改。下一仅按人类Claude代理Key更正另冻至多两新预检,旧失败和Purge未验保留;下方“推送收尾中”为原时点。 2026-11-12 M4原批次完成最终审计和封存:40真实请求/629106报告tokens/0unknown,Node、React、Python原检查及限定Git全部通过;541成员/134884464字节清单SHA558c9b54逐成员复核通过,见[最终验收](docs/QA-M4-LIVE-BASELINE.md)。NRO七份公开来源另根只读登记,不代签Mnemo语义或原AD/AG。当前只剩限定收尾提交及既有GitAtlas普通增量推送;1.25.13产品未变,不重发行/安装/重启。封存后人类反馈Claude代理Key已更正,按新公开修订另冻文本/工具至多两请求,不回退人类修改、不改旧两mismatch和目录缺项,不续旧任务或重跑三工程/Purge;凭据不读/复制/重存,工具只观察。Purge另冻、真实预留边界、生产语义/物理手机/长期和普遍收益unknown保持。下方阶段“下一封存”均为原时点。 diff --git a/docs/DEVELOPMENT-ROADMAP-OFFICIAL-FIRST.md b/docs/DEVELOPMENT-ROADMAP-OFFICIAL-FIRST.md index 761b2cf8..96723b10 100644 --- a/docs/DEVELOPMENT-ROADMAP-OFFICIAL-FIRST.md +++ b/docs/DEVELOPMENT-ROADMAP-OFFICIAL-FIRST.md @@ -1,5 +1,7 @@ # 官方优先开发路线(2026-11-06) +2026-11-12 Claude修后范围已执行并单独封存:完整目录14/选定在内,两新请求被上游403分组客户端权限拒绝,usage/身份/费用未知,工具执行0;见[修后验收](QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md)。41原件/当前七连接与全部原状态通过,原M4的40请求/541封存保持;提交推送收尾,供应商须另开放普通Messages客户端,不追加两请求或重跑工程/Purge。M8后续另冻,真实收益与M5–M7边界保持。 + 2026-11-12 M4验收e132601b及前六条已普通增量推送,远端/fetch逐SHA一致,见[回执](gitatlas-m4-live-push.json)。完成回执另提交核最终HEAD;原40请求/541封存和1.25.13程序保持。当前先另冻人类修后Claude两预检,随后M8原版Purge另做可比实验;下方推送收尾为原时点。 2026-11-12 当前:M4新三线六预检与Node/React/Python完整结束,共40真实请求629106报告tokens/0unknown,原检查/Git及状态审计通过,541原件134884464字节封存逐SHA复核;见[最终验收](QA-M4-LIVE-BASELINE.md)。普通GitAtlas增量推送收尾中,1.25.13程序不变。封存后人类更正Claude代理Key,另冻当前修订及新Task/Run最多两文本/工具预检,旧身份失败/目录缺项保留,不重跑工程/Purge或覆盖人类配置。原版Purge的M8对照另冻,M5–M7、Mnemo生产语义/真实手机/长期与各因果收益仍须自身证据。下方“下一M4”按历史时点理解。 diff --git a/docs/M4-CLAUDE-CORRECTION-PLAN-2026-11-12.md b/docs/M4-CLAUDE-CORRECTION-PLAN-2026-11-12.md index f302ce86..4afd584c 100644 --- a/docs/M4-CLAUDE-CORRECTION-PLAN-2026-11-12.md +++ b/docs/M4-CLAUDE-CORRECTION-PLAN-2026-11-12.md @@ -13,3 +13,5 @@ 结束后严格核对冻结前的全部原行/字段/schema、导出和工件。仅允许本批一个新项目/两新会话及其请求/事件/工件、明确项目前插位置和两序列增量;当前七连接/池/选择、配置和凭据行SHA保持。如有外部变化,保留失败和真实增量,不通过回退配置使审计过关。新原件单独封存、回顾、提交和普通GitAtlas推送;原 M4 两 mismatch、目录缺项和四十请求/541清单不改签。 历史网络、unknown、输出截断、重复读取、R/W真机预留、Purge/Mnemo语义/物理手机/长期与普遍收益边界仍保留。M8新原版Purge对照留到后续独立冻结。 + +执行结束:驱动97f132d3在真实前10项离线通过,当前公开状态一次冻结。目录14型号完整/选定在内,两新文本/工具请求各HTTP403 permission_error,分组仅允许Claude Code客户端;两次usage/身份未知、工具执行0,未追加请求。结束审计原58298行/137导出/418请求/5576工件及当前七连接/池/选择/配置一次exact;41成员36284086字节单独封存逐SHA通过,原M4的541清单保持。见[修后验收](QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md)。兼容未通过,执行与证据收尾完成;供应商权限修正后也须另有新范围,不重派本次原键。 diff --git a/docs/PLAN-1.25.md b/docs/PLAN-1.25.md index 94afcbac..125fa4f0 100644 --- a/docs/PLAN-1.25.md +++ b/docs/PLAN-1.25.md @@ -1,5 +1,7 @@ # 1.25 计划:任务要求与交付复核 +2026-11-12 Claude修后独立两预检结束,目录完整14/选定在内,两请求上游403 permission_error、分组仅允许Claude Code客户端;usage/费用/响应身份未知,工具执行0。[验收](QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md)与41成员新封存/原状态核对通过;当前七连接/人类更正与原M4的40请求/541清单保持。普通提交推送收尾,不追加请求/重跑工程/Purge或回退配置;供应商权限须另解决,历史unknown不改签。 + 2026-11-12 M4验收e132601b与前六条已普通增量推送且远端/fetch逐SHA一致,见[回执](gitatlas-m4-live-push.json)。原40请求/541封存保持,1.25.13未变;完成回执另提交核最终HEAD。下一人类修后Claude至多两新预检独立登记,不改旧失败,不重跑工程/Purge。 2026-11-12 M4最终审计和封存通过:40请求629106报告tokens/0unknown,三新工程原验收与Git完整,541成员134884464字节逐SHA核对;见[最终验收](QA-M4-LIVE-BASELINE.md)。收尾提交和既有GitAtlas普通推送进行中,产品未变不重打/重启1.25.13。封存后人类修正Claude代理Key,另冻最多两次原文本/工具预检,不回退配置、不改旧mismatch/目录缺项或续旧样本;三工程/Purge不重跑。M8收益实验另冻,全部历史unknown和Mnemo生产语义/手机/长期边界保持。下方“下一封存”按原时点理解。 diff --git a/docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md b/docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md new file mode 100644 index 00000000..4dc75f1b --- /dev/null +++ b/docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md @@ -0,0 +1,34 @@ +# 人类修正 Claude 代理 Key 后的独立兼容复核 + +目录已通过,模型请求仍未通过:新目录完整返回14型号,包含 `claude-opus-5-5`;原文本和工具两项新请求均被上游 HTTP403 拒绝,原文为“该分组仅允许使用 Claude Code 官方客户端。 This group is restricted to Claude Code clients (/v1/messages only)”。两次没有模型响应型号、SSE或usage;用量和费用均未知,未追加请求或执行工具。 + +机器记录:[claude-proxy-correction-2026-11-12.json](claude-proxy-correction-2026-11-12.json);清单核对:[claude-proxy-correction-seal-2026-11-12.json](claude-proxy-correction-seal-2026-11-12.json);[冻结计划](M4-CLAUDE-CORRECTION-PLAN-2026-11-12.md)。原M4[最终验收](QA-M4-LIVE-BASELINE.md)和四十请求/541清单保持,不能将本次两请求并入原封存或改签旧失败。 + +## 实际请求与结果 + +正式仍1.25.13/PID63892、原instance,unrestricted/passthrough/full,产品累计/步数均0。选定连接 `9ae348a5-d0a6-42ed-8664-0da5f9effe34` 为 Messages/claude-opus-5-5,连接revision8→9、credentialRevision1→2;config130保持,registry182→186。另新增Fable连接只记录公开身份,未选择/测试或推断操作者。人类报告此前Key写错的原因没有独立证明。 + +两请求均实际发送 `https://api.apikey.fan/v1/messages`、`anthropic-version: 2023-06-01`、原配置的x-api-key鉴权、`stream:true` 和官方精确型号后备 `max_tokens:128000`。字段及原body SHA已核对,未保存新参数/身份策略或主动本地额度。上游返回完整188字节JSON错误,type=`permission_error`,没有重定向或重试;SSE、工具能力与模型身份未取得证据。目录列出型号不能代替模型调用权限。 + +| 新用例 | Task / Run | 本地物理尝试 | 上游请求ID | +| --- | --- | --- | --- | +| 文本 | 22f44c69 / ba7da2e3 | 81f1bf6d-0cae-4a7a-b482-3b1381f4a606 | 02acb49c-9ca1-4fdc-9a4c-c45c71d4b000 | +| 工具观察 | 5decb3cb / b0468e4b | 9efaf829-859b-482a-a401-1c02531cd84a | 65157030-cf66-414a-bb14-8db6f4b70acc | + +完整project/session/Task/Run、规划requestId和物理wireAttemptId均在机器记录,两个用例各一新请求。原采集器可选requestKey字段为空,实际请求键从原账本requestId记录,没有补造或重派。原文件后缀按请求流式设置记为 `.sse`,其原字节实际为HTTP403 JSON错误,响应类型/完整性按实际记录,未将后缀当SSE证明。 + +两请求均 `usageStatus:unknown`,实际tokens和费用为null;已报告tokens的空集合合计0仅是账本合计,不能声称零消费或零计费。HTTP错误原文完整接收与模型用量未知分别成立。工具执行0,Node/React/Python与Purge均未重跑。 + +## 保持核对与封存 + +驱动 `97f132d3` 在真实请求前通过10项离线判断(8合成边界、2旧型号不符原回执只读复核),没有把它计为真实兼容。实际结束后 `audit-44ea125a-b46f-4355-bbac-5026b017c2e7` 一次通过:冻结前58298行/全部原字段及schema、137导出、418请求、5576工件及当前七连接/池/选择/配置/凭据行SHA保持。 + +仅新增自有一个项目/两会话/两请求,36events及outbox、16records、7新工件;项目前插+1、事件序列+36/请求序列+2明确核对。没有恢复旧Key或人类更正前注册表,PID/instance与空闲保持,9229关闭;未读、复制、解密或重存Key,产品程序未改,不重发行或重启。 + +新根 `.qa/m4-claude-correction-2026-11-12/` 封存41成员/36284086字节,MANIFEST SHA256 `9b8a1cfcc4a8842e5c751990fdfc4a274831bf4dcafe73c557514c992490f2e5`,全部成员逐bytes/SHA验证通过。原封存回执bytes表示成员总字节,独立公开回执另列manifest文件字节与原回执引用。原M4的541成员也再次只读核对通过,两根不再追加,原件不进Git。 + +## 回顾与下一步 + +当前明确阻塞是上游分组对客户端的权限要求。需要供应商开放该Key分组给标准Messages API客户端,或使用允许此客户端的分组;本轮不冒充Claude Code客户端或尝试更多请求。两次兼容复核已按授权范围结束,模型身份仍未验证,型号匹配本来也不认证权重。 + +历史两次自报GPT和旧目录缺项不改签,新403单列;网络/输出截断/重复读/R-W真机预留边界保持。Purge收益、Mnemo生产绑定与语义、物理手机/长期、其他模型能力仍依各自新实验,下一M8原版Purge另冻,不能以目录成功或已通过三工程替代对照收益。 diff --git a/docs/TIANSHU-ADOPTION-PLAN.md b/docs/TIANSHU-ADOPTION-PLAN.md index 7bd99ebb..e377839e 100644 --- a/docs/TIANSHU-ADOPTION-PLAN.md +++ b/docs/TIANSHU-ADOPTION-PLAN.md @@ -1,5 +1,7 @@ # Tianshu 研究后的 Agent Workbench 采用计划 +2026-11-12 修后Claude两新例另封存:目录14型号/选定在内,两模型请求上游403“仅允许Claude Code客户端”,usage/响应身份未知、费用null、0工具执行;[验收](QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md)。41成员36284086字节/原状态审计通过,当前七连接与人类更正保持,原40请求/541清单不改签。验收普通提交推送收尾,不追加请求或重跑三工程/Purge;M8收益仍须另冻。 + 2026-11-12 M4验收e132601b及此前六条已普通增量推送,远端和重新fetch逐SHA一致,见[回执](gitatlas-m4-live-push.json)。原40请求/541封存不变;下一按人类修正另冻Claude两项预检,Purge收益仍须新对照。完成回执和接续另提交核最终HEAD,下方收尾中按原时点保留。 2026-11-12 M4原批次最终完成:六兼容预检与三独立工程合计40请求629106报告tokens/0unknown,原要求/检查/Git和全部原状态通过;541成员/134884464字节已封存复核,见[最终验收](QA-M4-LIVE-BASELINE.md)。当前收尾提交及普通增量GitAtlas推送,不重发行1.25.13。封存后人类更正Claude代理Key,另用新修订/Task/Run至多两文本与工具预检,旧两mismatch及目录缺项不改签,不重跑三工程或Purge、不覆盖人类配置。M8原版Purge对照另冻;历史异常/真实预留边界/记忆与缓存因果/物理手机/长期未验保持。下方进度按原时点保留。 diff --git a/docs/claude-proxy-correction-2026-11-12.json b/docs/claude-proxy-correction-2026-11-12.json new file mode 100644 index 00000000..5b877db1 --- /dev/null +++ b/docs/claude-proxy-correction-2026-11-12.json @@ -0,0 +1,422 @@ +{ + "schemaVersion": 1, + "at": "2026-11-11T16:58:33.956Z", + "batch": "m4-claude-correction-2026-11-12", + "evidenceRoot": ".qa/m4-claude-correction-2026-11-12", + "driverCommit": "97f132d3cb2611dbbb758cde0a1399485690b7da", + "driverSha256": "52c8a5ba231f7b2caac8bb2bf9b81dcee5903a735c0a56a135c78abe8015b5d8", + "sourceNotice": { + "path": "D:/Ai-project/Mnemo/docs/integration/coordination-2026-11-12-claude-proxy-correction-user.json", + "bytes": 6801, + "sha256": "0a215dcbb4a18e08d25cb7afc5b28822dbc41a7ddc3c0dd2c72872494defe663", + "scope": "Public human correction notice only; keys absent; source not written." + }, + "humanReportedCause": "wrong_proxy_key", + "historicalMismatchCauseIndependentlyProven": false, + "originalM4": { + "requests": 40, + "reportedTotalTokens": 629106, + "unknownRequests": 0, + "sealedFiles": 541, + "manifestSha256": "558c9b548559798279d7f8e61deb3b7344920753a53dbf4e84eec374c6f3a3df", + "mismatchResponses": 2, + "configuredModelPreviouslyListed": false, + "oldResultsRewritten": false + }, + "formal": { + "version": "1.25.13", + "pid": 63892, + "instanceId": "75209495-f90a-4552-9a90-81a680d50c7c", + "configRevision": 130, + "registryRevision": 186, + "selected": { + "type": "pool", + "id": "35b99d94-2df1-4f25-afc3-25d063b6ff84" + } + }, + "currentConnection": { + "id": "9ae348a5-d0a6-42ed-8664-0da5f9effe34", + "name": "apikey-Claude", + "protocol": "anthropic-messages", + "model": "claude-opus-5-5", + "revision": 9, + "credentialRevision": 2, + "hasKey": true, + "nativeParameters": {} + }, + "postSealPublicChanges": { + "configRevision": [ + 130, + 130 + ], + "registryRevision": [ + 182, + 186 + ], + "selectedConnectionRevision": [ + 8, + 9 + ], + "credentialRevision": [ + 1, + 2 + ], + "definitionChangedFields": [ + "revision", + "updatedAt", + "credentialBindingRevision", + "upstreamModel", + "credentialRevision", + "modelSpecification", + "capabilities" + ], + "addedConnections": [ + { + "id": "02f67692-e3ce-4add-bde1-977c2162e0af", + "name": "apikey-Claude · claude-fable-5-2", + "model": "claude-fable-5-2", + "revision": 1 + } + ], + "observationRegistryRevision": 186, + "actorOfAdditionalChanges": "unestablished; preserved as observed; no connection restored" + }, + "catalog": { + "complete": true, + "models": 14, + "selectedListed": true, + "source": { + "path": "catalog.json", + "bytes": 24517, + "sha256": "78986dce5418965b1fff60c5e3a3b13a17601dd847bc7a3c73ab293c9ca71d50" + }, + "evidenceScope": "Public product catalogue with supplied source metadata; no invented raw HTTP capture." + }, + "preflights": [ + { + "scope": "text", + "projectId": "6afd88f9-0e28-41e0-ae2f-48dbf6e9f6c6", + "sessionId": "aa93b8ae-31b0-4cfc-8b28-b6db44d2e2b1", + "taskId": "22f44c69-5b5e-4d23-9298-9207e7550e34", + "runId": "ba7da2e3-a51a-491c-a0fe-c5092843e7f7", + "requestId": "b714df32-711a-4307-b7ec-1b68e1000950", + "attemptId": "81f1bf6d-0cae-4a7a-b482-3b1381f4a606", + "wireAttemptId": "a0c136b9-74f3-4648-9a40-6fa77272dad6", + "upstreamRequestId": "02acb49c-9ca1-4fdc-9a4c-c45c71d4b000", + "connectionId": "9ae348a5-d0a6-42ed-8664-0da5f9effe34", + "connectionRevision": 9, + "credentialRevision": 2, + "requestedModel": "claude-opus-5-5", + "responseModel": null, + "modelIdentityStatus": "incomplete", + "requestUrl": "https://api.apikey.fan/v1/messages", + "responseUrl": "https://api.apikey.fan/v1/messages", + "httpStatus": 403, + "responseFormat": "json_http_error_body", + "responseCaptureComplete": true, + "responseBytes": 188, + "originalUpstreamError": { + "error": { + "message": "该分组仅允许使用 Claude Code 官方客户端。 This group is restricted to Claude Code clients (/v1/messages only)", + "type": "permission_error" + }, + "type": "error" + }, + "requestOutputFields": { + "max_tokens": 128000 + }, + "outputFieldSource": "official_model_spec", + "streamRequested": true, + "sseObserved": false, + "sseMessageStopObserved": false, + "usageStatus": "unknown", + "usage": null, + "actualTotalTokens": null, + "actualCost": null, + "startedAt": "2026-11-11T16:54:45.390Z", + "finishedAt": "2026-11-11T16:54:45.684Z", + "elapsedMs": 294, + "actualToolsExecuted": 0, + "requestSource": { + "attemptId": "81f1bf6d-0cae-4a7a-b482-3b1381f4a606", + "artifactId": "4d9cfdd4-a322-43cf-bac7-7364b5f2f869", + "kind": "request", + "path": "text/captures/81f1bf6d-0cae-4a7a-b482-3b1381f4a606-request.json", + "bytes": 313, + "sha256": "b1f387575f4cdb2025ce242cb207f1bf3f160623df84bb2b2285c869aae9b285" + }, + "responseSource": { + "attemptId": "81f1bf6d-0cae-4a7a-b482-3b1381f4a606", + "artifactId": "a9f7983c-9b5d-4f2d-b1b1-390f54ae95d9", + "kind": "response", + "path": "text/captures/81f1bf6d-0cae-4a7a-b482-3b1381f4a606-response.sse", + "bytes": 188, + "sha256": "a9a64c798135516fcbfa0f036deeb21e38954e8d442a0dbc063fbc2905018ef8" + }, + "artifactSuffixNote": "Original collector used configured streaming to select .sse suffix. These original bytes are a complete JSON HTTP 403 error body, not SSE." + }, + { + "scope": "tools", + "projectId": "6afd88f9-0e28-41e0-ae2f-48dbf6e9f6c6", + "sessionId": "d19148aa-c9fe-4b8b-ae8c-34d83944c119", + "taskId": "5decb3cb-ea08-4d14-aa28-2f697c37ebb0", + "runId": "b0468e4b-2841-4111-8915-fbf6f0688f8a", + "requestId": "30510069-50fd-4e88-a35e-5a835799f921", + "attemptId": "9efaf829-859b-482a-a401-1c02531cd84a", + "wireAttemptId": "cec8475b-59cc-449f-b7d3-c2253eabc0e5", + "upstreamRequestId": "65157030-cf66-414a-bb14-8db6f4b70acc", + "connectionId": "9ae348a5-d0a6-42ed-8664-0da5f9effe34", + "connectionRevision": 9, + "credentialRevision": 2, + "requestedModel": "claude-opus-5-5", + "responseModel": null, + "modelIdentityStatus": "incomplete", + "requestUrl": "https://api.apikey.fan/v1/messages", + "responseUrl": "https://api.apikey.fan/v1/messages", + "httpStatus": 403, + "responseFormat": "json_http_error_body", + "responseCaptureComplete": true, + "responseBytes": 188, + "originalUpstreamError": { + "error": { + "message": "该分组仅允许使用 Claude Code 官方客户端。 This group is restricted to Claude Code clients (/v1/messages only)", + "type": "permission_error" + }, + "type": "error" + }, + "requestOutputFields": { + "max_tokens": 128000 + }, + "outputFieldSource": "official_model_spec", + "streamRequested": true, + "sseObserved": false, + "sseMessageStopObserved": false, + "usageStatus": "unknown", + "usage": null, + "actualTotalTokens": null, + "actualCost": null, + "startedAt": "2026-11-11T16:54:46.069Z", + "finishedAt": "2026-11-11T16:54:46.473Z", + "elapsedMs": 404, + "actualToolsExecuted": 0, + "requestSource": { + "attemptId": "9efaf829-859b-482a-a401-1c02531cd84a", + "artifactId": "fb534e24-b3a7-4f54-8316-55bc61082bce", + "kind": "request", + "path": "tools/captures/9efaf829-859b-482a-a401-1c02531cd84a-request.json", + "bytes": 648, + "sha256": "74fa50f9b8d49d9acbe83f7548602644a4a5fb0c5a55cc8d1084e96105e2b439" + }, + "responseSource": { + "attemptId": "9efaf829-859b-482a-a401-1c02531cd84a", + "artifactId": "d760a4de-d24f-4ef9-a870-619cc15f3871", + "kind": "response", + "path": "tools/captures/9efaf829-859b-482a-a401-1c02531cd84a-response.sse", + "bytes": 188, + "sha256": "a9a64c798135516fcbfa0f036deeb21e38954e8d442a0dbc063fbc2905018ef8" + }, + "artifactSuffixNote": "Original collector used configured streaming to select .sse suffix. These original bytes are a complete JSON HTTP 403 error body, not SSE." + } + ], + "requests": 2, + "maximumPhysicalRequests": 2, + "requestsWithKnownUsage": 0, + "unknownRequests": 2, + "reportedTokenSum": 0, + "actualTotalTokens": null, + "actualTotalCost": null, + "costScope": "Complete HTTP403 errors provide no model usage or cost. Zero sum of available reports is not zero upstream use or billing.", + "qualification": { + "catalog": true, + "text": false, + "tools": false, + "streaming": false, + "responseModelIdentity": false, + "requestContractFieldsVerified": true, + "modelWeightsAuthenticated": false + }, + "blocker": { + "origin": "upstream", + "httpStatus": 403, + "type": "permission_error", + "message": "该分组仅允许使用 Claude Code 官方客户端。 This group is restricted to Claude Code clients (/v1/messages only)", + "scope": "Provider group accepts Claude Code clients only; standard Workbench Messages requests were rejected.", + "remedy": "Supplier must grant ordinary Messages API client permission or provide a group that accepts such clients. No client impersonation or extra model requests were attempted." + }, + "driverValidation": { + "at": "2026-11-11T16:52:49.879Z", + "checks": [ + { + "name": "declared complete synthetic Messages receipt", + "passed": true + }, + { + "name": "missing mandatory output field", + "passed": true + }, + { + "name": "old local 2048 is not the declared maximum", + "passed": true + }, + { + "name": "missing message_stop", + "passed": true + }, + { + "name": "invalid SSE JSON", + "passed": true + }, + { + "name": "physical request count exceeds case scope", + "passed": true + }, + { + "name": "observed tool cannot have executed", + "passed": true + }, + { + "name": "unknown usage cannot qualify", + "passed": true + }, + { + "name": "original sealed text mismatch remains failed", + "passed": true + }, + { + "name": "original sealed tools mismatch remains failed", + "passed": true + } + ], + "syntheticCases": 8, + "originalMismatchReceiptsReadOnly": 2, + "realModelRequests": 0, + "oldRootsWritten": false, + "keysReadOrResaved": false + }, + "audit": { + "success": true, + "at": "2026-11-11T16:56:12.964Z", + "pid": 63892, + "instanceId": "75209495-f90a-4552-9a90-81a680d50c7c", + "version": "1.25.13", + "originalSessions": 137, + "originalModelRequests": 418, + "originalArtifacts": 5576, + "checkedOriginalRows": 58298, + "addedRows": { + "events": 36, + "metadata": 0, + "model_attempts": 2, + "outbox": 36, + "projects": 1, + "records": 16, + "sessions": 2, + "sqlite_sequence": 0 + }, + "sequenceChanges": [ + { + "name": "events", + "before": 24175, + "after": 24211, + "increment": 36 + }, + { + "name": "model_attempts", + "before": 418, + "after": 420, + "increment": 2 + } + ], + "newArtifacts": 7, + "currentSevenConnectionDefinitionsPoolsSelection": "exact", + "originalExportsAndArtifactBytes": "exact", + "originalMetadataAndCredentialRow": "exact SHA; no credential decrypted or copied", + "originalProjects": "exact after declared single front insertion", + "requests": 2, + "inspectorClosed": true, + "keysReadOrResaved": false, + "noRestorationOfOldHumanSettings": true, + "source": { + "path": "audit-44ea125a-b46f-4355-bbac-5026b017c2e7/result.json", + "bytes": 1097, + "sha256": "17ff401656a6353808d2a0163e58d7d0f39fe6ad00751fca8d53804605041950" + }, + "afterSource": { + "path": "audit-44ea125a-b46f-4355-bbac-5026b017c2e7/after.json", + "bytes": 16447360, + "sha256": "9298d9a572e99499dc0254afc0a60bbb238023c07c73d62f1de6abe9473db743" + } + }, + "sources": [ + { + "path": "observation.json", + "bytes": 1048438, + "sha256": "0c0ec165c3dfcc776cc3106f9b1b930cd655104302d5a41330aed274a7f17621" + }, + { + "path": "frozen.json", + "bytes": 18119119, + "sha256": "2710d091eb84aa7f32e1be6872796f51f428082026ca940e392ad131e36ebb9b" + }, + { + "path": "claim.json", + "bytes": 181, + "sha256": "7761a2644549cfd140ca32f9d5d6e3af23339c4912a1855706884a1d95b05ea2" + }, + { + "path": "catalog.json", + "bytes": 24517, + "sha256": "78986dce5418965b1fff60c5e3a3b13a17601dd847bc7a3c73ab293c9ca71d50" + }, + { + "path": "catalog-cancel.json", + "bytes": 24, + "sha256": "5e5626d3e3b68f8f5ba416eb1c04e6f1de60ff931a236cf7bdb48199ff750561" + }, + { + "path": "result.json", + "bytes": 18687, + "sha256": "95caf7e791299218eedbbe24bcdf66cc116f243534bce39d82317084df91987c" + }, + { + "path": "canonical-request-identities.json", + "bytes": 5799, + "sha256": "ece30dca26b2b404860144fe3c76f1caf3d7d9684f4fbf514fd427c1214336f8" + }, + { + "path": "driver-validation.json", + "bytes": 1069, + "sha256": "da2cb65067aafc18218f8573f3a2191a30a72d9a208bdee207041ee08eaccd7e" + }, + { + "path": "driver-source.mjs", + "bytes": 24449, + "sha256": "52c8a5ba231f7b2caac8bb2bf9b81dcee5903a735c0a56a135c78abe8015b5d8" + }, + { + "path": "human-correction-source.json", + "bytes": 6801, + "sha256": "0a215dcbb4a18e08d25cb7afc5b28822dbc41a7ddc3c0dd2c72872494defe663" + } + ], + "boundaries": { + "keysReadOrResaved": false, + "oldM4RootWritten": false, + "modelPreferencesRestored": false, + "currentSevenConnectionsPreserved": true, + "toolsExecuted": 0, + "engineeringOrPurgeReplayed": false, + "productSourceChanged": false, + "releaseOrRestart": false, + "messagesSentToSource": 0, + "authorizationExpanded": false + }, + "unverified": [ + "Corrected Claude text/tool/streaming/response identity", + "Real upstream model weights", + "Purge and cache/experience/role causal benefits", + "Mnemo production binding and semantic benefit", + "Physical iPhone and long term stability", + "Historical R/W live outlier boundary" + ] +} diff --git a/docs/claude-proxy-correction-seal-2026-11-12.json b/docs/claude-proxy-correction-seal-2026-11-12.json new file mode 100644 index 00000000..7ed733bd --- /dev/null +++ b/docs/claude-proxy-correction-seal-2026-11-12.json @@ -0,0 +1,28 @@ +{ + "schemaVersion": 1, + "verifiedAt": "2026-11-11T17:00:06.064Z", + "evidenceRoot": ".qa/m4-claude-correction-2026-11-12", + "files": 41, + "bytes": 36284086, + "manifest": { + "path": "MANIFEST.json", + "bytes": 6867, + "sha256": "9b8a1cfcc4a8842e5c751990fdfc4a274831bf4dcafe73c557514c992490f2e5" + }, + "originalSealReceipt": { + "path": "seal.json", + "bytes": 346, + "sha256": "aacca201e3b5822ffcda363bc51a55951279f2e55ffb337d4aa8751933f96b73" + }, + "originalReceiptBytesMeaning": "Total member bytes; the manifest file length is explicitly separate above.", + "originalM4ManifestSha256": "558c9b548559798279d7f8e61deb3b7344920753a53dbf4e84eec374c6f3a3df", + "requests": 2, + "requestsWithKnownUsage": 0, + "unknownRequests": 2, + "actualTotalTokens": null, + "actualTotalCost": null, + "memberBytesAndSha256Verified": true, + "currentSevenConnectionsPreserved": true, + "keysReadOrResaved": false, + "engineeringOrPurgeReplayed": false +} diff --git a/findings.md b/findings.md index aac720b5..c7d088e5 100644 --- a/findings.md +++ b/findings.md @@ -1,5 +1,7 @@ # 汇总发现 +2026-11-12 修后Claude的目录已完整14型号且含选定,但两模型请求均403 permission_error,原文明确分组仅允许Claude Code客户端。实际/v1/messages、x-api-key占位、stream及max_tokens128000字段完整,响应188字节JSON完整,不能因.sse后缀当流式通过;无型号/usage,实际tokens/费用null。规划requestId/物理wireAttemptId从原账本登记,可选requestKey空字段保持,不重派。原状态与当前七连接一审通过,41新封存及旧541保持;见docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md。需要供应商普通API客户端权限,不以目录成功认证权重或收益。 + 2026-11-12 M4普通推送e132601b成功,前六条连同验收七条逐SHA与远端/fetch一致,37待推送文件无私钥/真实Key模式或归档/依赖路径,见docs/gitatlas-m4-live-push.json。源码包和证据原件未入Git,runner结果仍unknown;修后Claude另记录,原M4事实不改签。 2026-11-12 M4最终原状态审计通过,原53527行/127导出/5244工件/377请求与本批九捕获导出保持;三工程共34请求,含预检总40请求629106报告tokens/0unknown。三工程同SHA同范围重复读取0、440960cacheRead仅为观察,cacheWrite未知不补0。541成员134884464字节封存SHA558c9b54及十三驱动已复核;见[最终验收](docs/QA-M4-LIVE-BASELINE.md)。人类在封存后报告Claude代理Key此前写错并更正,原因尚无独立证明;旧两次自报GPT/目录缺项完整保持,修后另冻至多两新预检,原审计不能用于回退人类配置。NRO与AD/另选AG分别只读登记,不代签生产语义或Purge。 diff --git a/progress.md b/progress.md index 96764388..bde7afa3 100644 --- a/progress.md +++ b/progress.md @@ -1,5 +1,7 @@ # 统一执行进度 +2026-11-12 Claude独立修后范围结束:97f132d3驱动离线10项通过后一次冻结,目录14完整/选定在内,两新例各一请求均上游403“仅Claude Code客户端”,2unknown、费用null、0工具执行。audit-44ea125a一次保持原58298行/137导出/418请求/5576工件及当前七连接/池/选择/配置/凭据SHA;41成员36284086字节SHA9b8a1cfc封存复核,原M4 541清单保持。docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md记录原错误和完整身份,当前限定验收提交/普通GitAtlas推送收尾,不追加或重跑工程/Purge/重启。 + 2026-11-12 M4验收e132601b及前六条普通推送成功,七条与远端和重新fetch逐SHA一致;docs/gitatlas-m4-live-push.json回执及完成接续另提交后核最终HEAD。541原件/40请求保持,产品不重发行。下一另冻修后Claude目录及两项预检,原失败和Purge收益unknown保留。 2026-11-12 M4阶段6原批次最终审计、541成员134884464字节封存和逐SHA复核通过;40请求629106报告tokens/0unknown,三新工程原验收与Git已实际完成。本次仅收尾文档/驱动提交与普通增量GitAtlas推送,不重复产品测试、发行或重启。封存后人类更正Claude代理Key,接着另冻当前修订、原文本/工具两预检最多两请求;旧两mismatch/目录缺项及全部历史unknown保留,不回退用户配置、不读Key、不重跑三工程/Purge。NRO七公开副本/原索引只读核对,来源不回信。最终QA在docs/QA-M4-LIVE-BASELINE.md。 diff --git a/task_plan.md b/task_plan.md index a66fbb21..0bde00e4 100644 --- a/task_plan.md +++ b/task_plan.md @@ -1,5 +1,7 @@ # Agent Workbench:核心代理能力研究与产品计划 +2026-11-12 修后Claude执行和证据完成:目录完整14/选定在内,两新请求HTTP403 permission_error“仅Claude Code客户端”,0执行工具、2unknown,实际tokens/费用null。docs/QA-CLAUDE-PROXY-CORRECTION-2026-11-12.md与机器记录/41成员封存SHA9b8a1cfc通过;一次核原58298行/137导出/418请求/5576工件与当前七连接/池/选择/配置。当前限定验收提交及普通推送;原40请求/541清单保持,供应商权限须另解决,不追加请求或重跑工程/Purge,下一M8另冻。 + 2026-11-12 M4阶段6完成推送:e132601b与前六条远端/fetch逐SHA一致,docs/gitatlas-m4-live-push.json保存回执。回执与完成接续最后限定提交核最终HEAD;原40请求和541封存不再变动。下一人类修后Claude新目录及文本/工具至多两预检,不重跑工程/Purge、不改旧异常或用户凭据。 2026-11-12 当前M4原40请求已结束,最终审计和541成员/134884464字节封存复核通过,三新工程全部成功,629106报告tokens/0unknown;见[最终验收](docs/QA-M4-LIVE-BASELINE.md)。原十三驱动及NRO七公开副本核对,限定提交/普通GitAtlas推送收尾。1.25.13未改不重发行。封存后收到人类Claude代理Key更正,后续另冻公开修订和新Task/Run、原文本/工具两例至多两请求;不读Key、不回退人类变更,不续旧失败/unknown、不重跑三工程/Purge,不向来源回信或委派。旧两mismatch/目录缺项、历史网络/重复读/真实预留边界、Mnemo生产语义/长期/手机/Purge收益unknown保持;M8另冻在后。