diff --git a/CHANGELOG.md b/CHANGELOG.md index 9e812c7b..fb48f72a 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,12 @@ # Agent Workbench 更新记录 +## 1.5.1 · 2026-10-29 + +- 修复大工件结果因请求预算不足而再次被封存、模型无法看到刚读内容的问题:读取页按下一请求的实际JSON编码字节预算选择,返回next_offset、实际字符数及预算缩小标记;完整原工件保留。 +- 历史选择优先保留直接代码及原执行工件依据,历史索引用于按需定位,减少不断读取更新索引的循环。 +- 同Task持久记录不可变工件的已读区间;3次连续重复已覆盖区间或连续追逐6个新历史索引时明确no_progress暂停。新的区间/直接证据与显式续接有各自处理,重启不自动发送请求或重放工具。 +- 手机上可查看重复读取原因并显式继续原Task。完整回归、包与失败样本续接结果分别保存到[验收记录](docs/VALIDATION.md),1.5.0的真实失败记录保留。 + ## 1.5.0 · 2026-10-29 - 长任务按完整工具调用/结果组选取上下文,原参数和历史保留;任务目标、全部追加要求及最后用户消息完整发送,超大历史通过SHA256工件分段读取。 diff --git a/README.md b/README.md index 81b95e59..c33f3423 100644 --- a/README.md +++ b/README.md @@ -1,6 +1,6 @@ # 本地 Agent 工作台 -Agent Workbench 1.5.0 将 Claude Code 与 Cursor 的机制研究整合为独立的本地 AI 编程工作台。工程索引、精确补丁、多文件整批审查、实际检查、本地Git提交及交付报告可以在同一个Task中完成。 +Agent Workbench 1.5.1 将 Claude Code 与 Cursor 的机制研究整合为独立的本地 AI 编程工作台。工程索引、精确补丁、多文件整批审查、实际检查、本地Git提交及交付报告可以在同一个Task中完成。 每次运行保留模型与预算、已完成工具和日志工件。多文件变更集封存同Task的原/新字节、Git基线与审查指纹,应用前核对整批文件;中断后显式核对或回退,保留外部修改。应用改动后可点击「继续检查与交付」,沿用原目标进行实际测试与修复。提交前查看实际文件、diff、署名和检查结果,确认后生成本地提交及可下载的JSON/Markdown报告。版本改动见 [CHANGELOG](CHANGELOG.md),编辑范围见[索引与补丁](docs/CODE-INDEX-AND-PATCHES.md)和[变更集恢复契约](docs/CHANGE-SETS.md)。 @@ -62,7 +62,7 @@ Windows 终端使用 PowerShell 5.1,通过 UTF-8 stdin 传入经确认的完 - [统一机制报告](docs/RESEARCH.md):原版静态证据与本产品实现逐项对照。 - [详细报告与证据索引](docs/research):Claude Code 工具循环、上下文/恢复,以及 Cursor 工具执行/编辑。 - [验收记录](docs/VALIDATION.md):逐版本保留实际结果、失败与部署身份。 -- [本轮验收清单](docs/QA-1.5.md)、[封存四模块工程基准](benchmarks/node-delivery-v1/README.md)与[保留连接的升级说明](docs/UPGRADE.md)。 +- [本轮验收清单](docs/QA-1.5.1.md)、[封存四模块工程基准](benchmarks/node-delivery-v1/README.md)与[保留连接的升级说明](docs/UPGRADE.md)。 - [独立架构](docs/ARCHITECTURE.md)与[界面设计](DESIGN.md)。 原版研究没有执行 Claude Code/Cursor,报告不表示恢复了全部源码或验证了远端服务。完整分析工件保留在原研究工作区;本产品包不包含第三方客户端原二进制、完整提取源码或大体积代码切片。 diff --git a/docs/CONTEXT-PLANS.md b/docs/CONTEXT-PLANS.md index 0f753913..08e532f1 100644 --- a/docs/CONTEXT-PLANS.md +++ b/docs/CONTEXT-PLANS.md @@ -23,3 +23,11 @@ 覆盖单user回合的长Task、近期大小混合结果、Unicode、完整多工具配对、原参数和历史不变、目标和追加要求完整、固定开销/输出余量/Run和Task剩余预算、工件缺失/外部篡改、显式续接及实际进程重启。API/桌面/390px应能读取请求估算与实际用量、历史索引和工件;真实同题样本单独报告用量和人工介入。 本文件描述当前开发契约;实际完成/部署以VALIDATION.md为准。 + +## 1.5.1:读取预算与重复进展 + +1.5.0真实样本暴露较大工件页被整体封存的循环,失败记录保留。1.5.1的read_artifact按下一请求中工具消息的双层JSON字节预算生成实际读取页;默认2000字符,requested_limit与实际范围分别返回,next_offset指向下一段。UTF-16坐标按合法Unicode边界处理;分页不改写历史工具参数,原工件与SHA256保持。 + +请求计划优先选择近期直接代码/原结果组,并提供有界working_evidence定位。目标、追加要求和完整协议仍按原契约保留;不能容纳的原大结果仍整体封存,不宣称所有任意大小工具组都能进入模型窗口。 + +Task保存有界工件覆盖区间、重复次数和最近读取身份。连续3次读取已完整覆盖区间或连续追逐6个更新历史索引时,以no_progress暂停,所有已执行结果先保存并完成工具配对。读取新区间或取得直接证据会结束重复序列;显式续接清除暂停序列,保留已读证据与同Task身份。此检测针对工件读取,尚不是通用任务调度/语义无进展判断。 diff --git a/docs/QA-1.5.1.md b/docs/QA-1.5.1.md new file mode 100644 index 00000000..057b20c1 --- /dev/null +++ b/docs/QA-1.5.1.md @@ -0,0 +1,9 @@ +# 1.5.1 验收清单 + +本轮由[1.5真实失败](context-live-model-failure.json)驱动:较大的工件页在下一请求中被封存,模型反复读取同一片段。修复实际工具分页、直接依据保留和工件重复进展检测。 + +专项4/4读取测试及1项直接代码保留回归通过;原上下文/API/Runtime的25项针对性回归通过。专项包含实际本地HTTP中模型必须收到最新工件结果、完整Unicode重建、重复读取暂停、保存后重新打开不重放,以及显式续接同Task。 + +完整163/163项行为/API/实际退出/升级回归通过,0失败/跳过/取消,见[完整回归](context-fix-tests-validation.json)。基础10、诊断9、任务7、Git9、索引8、批次11及上下文9共63组界面通过,0页面/请求异常,见[界面总表](context-fix-all-ui-validation.json)。新增暂停用例最初误沿用桌面视口,已显式设置并验证390px后复验;手机查看原因、关闭任务抽屉及显式续接保持原Task。 + +便携/源码独立验收、保留原进程升级与同一个真实失败Task的续接随后登记。当前正式服务仍为1.5.0,不将待验收结果写为完成。 diff --git a/docs/VALIDATION.md b/docs/VALIDATION.md index ef733a80..bc22cdbe 100644 --- a/docs/VALIDATION.md +++ b/docs/VALIDATION.md @@ -1,5 +1,13 @@ # Agent Workbench 验收记录 +## 1.5.1 · 实际工件分页与重复读取暂停 + +构建、163/163项完整回归及63组界面全部通过,0失败/跳过/取消/页面或请求异常;见[回归](context-fix-tests-validation.json)、[界面总表](context-fix-all-ui-validation.json)和[QA-1.5.1](QA-1.5.1.md)。工件页按下一请求的双层JSON字节限额生成、返回next_offset;原工件、Unicode与完整协议保留。直接代码依据优先选择,Task已读区间和no_progress暂停持久保存,恢复不自动重放。 + +新增实际HTTP用例证明下一请求包含刚读取的大工件页,分页推进而不重复工具;重复已覆盖区间的实际循环以no_progress暂停,重新打开后保留原Task/已完成结果,显式续接才请求模型。手机用例检查实际390px宽度、可读原因和续接归属;第一次误用桌面截图的验证器已修正并重跑。 + +独立发行和同一真实失败Task续接正在执行,后续追加实际结果。1.5.0首轮失败证据保持,尚无同题多次对照的费用收益结论。 + ## 1.5.0 · 有界请求与完整历史 Windows x64、Node24.22.1、Edge156.0.4314.42。构建与158/158项行为/API/实际进程退出/升级回归通过,0失败/跳过/取消;上下文8、批次11及原有43共62组真实界面全部通过,0页面/请求异常。桌面与390px历史/工件截图已目视核对。 diff --git a/docs/context-fix-all-ui-validation.json b/docs/context-fix-all-ui-validation.json new file mode 100644 index 00000000..68c72286 --- /dev/null +++ b/docs/context-fix-all-ui-validation.json @@ -0,0 +1,38 @@ +{ + "success": true, + "version": "1.5.1", + "testedAt": "2026-10-29T13:47:45.452Z", + "count": 63, + "checks": [ + { + "script": "verify-ui.mjs", + "passed": 10 + }, + { + "script": "verify-diagnostics-ui.mjs", + "passed": 9 + }, + { + "script": "verify-execution-ui.mjs", + "passed": 7 + }, + { + "script": "verify-delivery-ui.mjs", + "passed": 9 + }, + { + "script": "verify-index-ui.mjs", + "passed": 8 + }, + { + "script": "verify-change-sets-ui.mjs", + "passed": 11 + }, + { + "script": "verify-context-ui.mjs", + "passed": 9 + } + ], + "contextUiRecheckedAt": "2026-10-29T13:47:45.452Z", + "mobileReadPauseWidth": 390 +} diff --git a/docs/context-fix-basic-ui-validation.json b/docs/context-fix-basic-ui-validation.json new file mode 100644 index 00000000..bea99ba8 --- /dev/null +++ b/docs/context-fix-basic-ui-validation.json @@ -0,0 +1,27 @@ +{ + "success": true, + "url": "http://127.0.0.1:52218", + "browser": "Microsoft Edge 156.0.4314.42", + "date": "2026-10-29T13:39:52.423Z", + "projectRoot": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\ui-4e108dc4-3f09-4a10-8235-1d0125675aba\\ui-project", + "viewports": [ + 1440, + 780, + 390 + ], + "passed": [ + "project selection and real file preview", + "real command consent, execution and stale snapshot rejection", + "diff review and verified file application", + "manual editor creates a reviewable proposal", + "file responses stay with the current selection", + "denial, session selection, reload, stale response rejection and export", + "project inventory responses cannot replace another project", + "780 and 390 px: project drawer, file, diff and trace controls, no page overflow", + "unified and detailed research reports render", + "model list, ephemeral key, compatible fixture tool loop and return to offline mode" + ], + "pageErrors": [], + "requestFailures": [], + "modelValidation": "controlled local HTTP fixture; no real model" +} diff --git a/docs/context-fix-batch-ui-validation.json b/docs/context-fix-batch-ui-validation.json new file mode 100644 index 00000000..7998afda --- /dev/null +++ b/docs/context-fix-batch-ui-validation.json @@ -0,0 +1,29 @@ +{ + "success": true, + "version": "1.5.1", + "testedAt": "2026-10-29T13:42:38.074Z", + "cases": [ + "Desktop batch scope, per-file diffs, sealed single-write guard and pre-apply bytes", + "Three-file reviewed application preserves exact BOM/CRLF/LF and creates the new file", + "Explicit batch rollback restores exact originals and removes only the reviewed new file", + "A last-file owner conflict blocks the whole batch and shows the real recovery direction", + "Durable partial recovery reads actual hashes and preserves owner content during rollback", + "Changed observation fingerprints reject stale acknowledgement without releasing the project", + "Explicit owner-content acknowledgement ends recovery without claiming successful application", + "Switching projects clears the previous task notification while retaining its durable error and recovery history", + "390px step-limit review, same-task continuation entry, rollback confirmation and keyboard close fit the viewport", + "A delayed preparation response never opens a different project or changes its selected session", + "A failed batch response after its component unmounts cannot display an error in the newly selected project" + ], + "pageErrors": [], + "requestFailures": [], + "screenshots": [ + "docs/screenshots/context-fix-batch-desktop-review.png", + "docs/screenshots/context-fix-batch-desktop-conflict.png", + "docs/screenshots/context-fix-batch-desktop-recovery.png", + "docs/screenshots/context-fix-batch-mobile-review.png", + "docs/screenshots/context-fix-batch-mobile-rollback.png" + ], + "projectRoot": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\batch-ui-261708de-6af0-40ff-ad0d-66e5b444cfa3", + "scope": "Real independent demo host and durable batch states; UI fixtures include actual owner conflict/rollback results. No real model credentials or user project were used." +} diff --git a/docs/context-fix-delivery-ui-validation.json b/docs/context-fix-delivery-ui-validation.json new file mode 100644 index 00000000..11dc8d65 --- /dev/null +++ b/docs/context-fix-delivery-ui-validation.json @@ -0,0 +1,31 @@ +{ + "success": true, + "version": "1.5.1", + "testedAt": "2026-10-29T13:42:22.402Z", + "browser": "156.0.4314.42", + "passed": [ + "actual failing tests and four source proposals retain the original project and task", + "reviewed changes continue the same Task through 14 real tests and model-prepared Git review", + "review shows actual paths, diffs, tests, author and policy; cancelling leaves Git unchanged", + "an index change after confirmation preview refuses the stale job and preserves both staged owners", + "390px confirmation keeps file hashes, diffs and actions inside the modal", + "double confirmation creates one four-file commit, retains owner bytes and publishes report with completion", + "report viewing and Markdown/JSON downloads preserve verified artifact bytes and truthful acceptance limits", + "mobile report preview closes cleanly and keeps the task input fixed", + "unknown Git results block new submission and the explicit reconcile action restores the existing commit" + ], + "pageErrors": [], + "requestFailures": [], + "projectRoot": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\delivery-ui-5b728005-4b5e-4f90-96ad-d21bb97c4efe\\四模块交付工程", + "projectId": "bc9f1b3a-0af3-4919-8a8e-6e736e6ae638", + "sessionId": "e6703858-6abd-40bb-a25f-76267efc3bec", + "taskId": "9d62894b-79d2-429f-a13e-b8e38c0cb036", + "commit": "657c040cdef31e810a75ee3f7668221a84ac16a4", + "recoveryCommit": "34836a8ffdbe75921a28d8523e073043e03d3cd2", + "testCounts": { + "baselineFailed": 13, + "passed": 14 + }, + "model": "controlled local HTTP fixture", + "confirmations": 2 +} diff --git a/docs/context-fix-diagnostics-ui-validation.json b/docs/context-fix-diagnostics-ui-validation.json new file mode 100644 index 00000000..f0bf19cc --- /dev/null +++ b/docs/context-fix-diagnostics-ui-validation.json @@ -0,0 +1,20 @@ +{ + "success": true, + "version": "1.5.1", + "url": "http://127.0.0.1:52218", + "date": "2026-10-29T13:40:38.565Z", + "passed": [ + "model response validation checks the tentative connection before saving", + "a model failure after successful tools shows source, attempts and local runtime diagnosis", + "explicit retry preserves the completed tool result without replaying the task", + "an error question receives the actual session diagnosis and formatted answer", + "review state exposes the review action and prevents application during a running request", + "approved Unicode commands execute in the declared shell and retain Chinese output", + "authentication errors do not retry and error/diagnostic controls fit a 390 px viewport", + "stop cancels pending retry and clears running model state", + "a transient service failure recovers within the bounded retry policy" + ], + "pageErrors": [], + "requestFailures": [], + "modelValidation": "controlled local HTTP fixture; actual configured model is verified separately" +} diff --git a/docs/context-fix-execution-ui-validation.json b/docs/context-fix-execution-ui-validation.json new file mode 100644 index 00000000..f35396ee --- /dev/null +++ b/docs/context-fix-execution-ui-validation.json @@ -0,0 +1,19 @@ +{ + "success": true, + "version": "1.5.1", + "testedAt": "2026-10-29T13:41:30.900Z", + "browser": "156.0.4314.42", + "passed": [ + "model-selected checks wait for approval and display actual counted results", + "verified artifact preview shows the real check log and hash", + "changed source invalidates old checks and manual confirmation runs the new plan", + "390px task drawer, log preview, close controls and fixed composer remain operable", + "reconnection synchronizes missed session changes in other projects", + "budget controls pause before tools and preserve the goal across recorded runs", + "applied proposals continue the original task into actual engineering verification" + ], + "pageErrors": [], + "requestFailures": [], + "projectRoot": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\execution-ui-db8cc267-eb38-49b5-ae45-693fbf088ca5\\project", + "model": "controlled local HTTP fixture" +} diff --git a/docs/context-fix-index-ui-result.json b/docs/context-fix-index-ui-result.json new file mode 100644 index 00000000..32f6099a --- /dev/null +++ b/docs/context-fix-index-ui-result.json @@ -0,0 +1,19 @@ +{ + "checkedAt": "2026-10-29T13:42:27.687Z", + "version": "1.5.1", + "passed": [ + "工程索引可独立于任务使用,实际覆盖范围与损坏文件显示明确", + "真实AST查询打开原文件第181行并仅滚动代码容器,输入区和顶部保持可见", + "查询词保留;过期SHA不高亮旧位置,增量刷新后正确定位第184行", + "切换项目时延迟返回的旧索引不覆盖新项目结果", + "Python候选、字符串误报排除和局部导入文件候选保留真实证据标签", + "两处唯一锚点成为待审核提案,真实界面应用后BOM/CRLF和未选择原字节保留", + "390px索引查询、范围和完整指纹可读且无横向溢出,实际文件行定位通过", + "索引刷新失败保留归属和解释,旧结果禁用并可重试恢复" + ], + "pageErrors": [], + "requestFailures": [], + "root": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\index-ui-982c3cea-8762-4ac7-ab43-3328c8ef105f", + "primaryProjectId": "759b7254-d556-41dd-a529-05154762bc73", + "otherProjectId": "bf558c4e-768f-4ac1-a268-cb53f496283d" +} diff --git a/docs/context-fix-tests-validation.json b/docs/context-fix-tests-validation.json new file mode 100644 index 00000000..401a9235 --- /dev/null +++ b/docs/context-fix-tests-validation.json @@ -0,0 +1,16 @@ +{ + "success": true, + "version": "1.5.1", + "testedAt": "2026-10-29T13:41:56.942Z", + "counts": { + "tests": 163, + "pass": 163, + "fail": 0, + "cancelled": 0, + "skipped": 0, + "todo": 0 + }, + "exitCode": 0, + "logFile": ".qa/context-full-tests.tap", + "testFiles": 19 +} diff --git a/docs/context-fix-ui-validation.json b/docs/context-fix-ui-validation.json new file mode 100644 index 00000000..3112ab45 --- /dev/null +++ b/docs/context-fix-ui-validation.json @@ -0,0 +1,30 @@ +{ + "success": true, + "version": "1.5.1", + "testedAt": "2026-10-29T13:47:45.452Z", + "passed": [ + "desktop request estimate, actual usage, complete goal and history identity are visible", + "paged complete history opens actual tool result artifacts without executing tools", + "historical and resumed Runs retain their own request records", + "390px history, original JSON and sent-body dialogs fit and retain close controls", + "late old history failure stays in its original scope after switching projects", + "late old Run failure stays in its original scope after switching projects", + "late old request-body failure stays in its original scope after switching projects", + "context-limit pause explains the intact requirement and explicit continuation", + "390px repeated-read pause explains the cause and explicit continuation retains the same Task" + ], + "pageErrors": [], + "requestFailures": [], + "screenshots": [ + "docs/screenshots/context-fix-desktop-plan.png", + "docs/screenshots/context-fix-desktop-history.png", + "docs/screenshots/context-fix-mobile-history.png", + "docs/screenshots/context-fix-mobile-original.png", + "docs/screenshots/context-fix-mobile-request-body.png", + "docs/screenshots/context-fix-mobile-read-pause.png" + ], + "root": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\context-ui-17d4e124-33a6-45cc-8004-ff58e3bad461", + "data": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\context-ui-17d4e124-33a6-45cc-8004-ff58e3bad461\\data", + "mode": "saved synthetic context evidence; no real model or project commands executed by this UI verifier", + "browser": "156.0.4314.42" +} diff --git a/docs/screenshots/context-fix-batch-desktop-conflict.png b/docs/screenshots/context-fix-batch-desktop-conflict.png new file mode 100644 index 00000000..d0018c5b Binary files /dev/null and b/docs/screenshots/context-fix-batch-desktop-conflict.png differ diff --git a/docs/screenshots/context-fix-batch-desktop-recovery.png b/docs/screenshots/context-fix-batch-desktop-recovery.png new file mode 100644 index 00000000..bc463fa1 Binary files /dev/null and b/docs/screenshots/context-fix-batch-desktop-recovery.png differ diff --git a/docs/screenshots/context-fix-batch-desktop-review.png b/docs/screenshots/context-fix-batch-desktop-review.png new file mode 100644 index 00000000..cd8f166c Binary files /dev/null and b/docs/screenshots/context-fix-batch-desktop-review.png differ diff --git a/docs/screenshots/context-fix-batch-mobile-review.png b/docs/screenshots/context-fix-batch-mobile-review.png new file mode 100644 index 00000000..4d395dfc Binary files /dev/null and b/docs/screenshots/context-fix-batch-mobile-review.png differ diff --git a/docs/screenshots/context-fix-batch-mobile-rollback.png b/docs/screenshots/context-fix-batch-mobile-rollback.png new file mode 100644 index 00000000..de01c85b Binary files /dev/null and b/docs/screenshots/context-fix-batch-mobile-rollback.png differ diff --git a/docs/screenshots/context-fix-desktop-history.png b/docs/screenshots/context-fix-desktop-history.png new file mode 100644 index 00000000..e449a477 Binary files /dev/null and b/docs/screenshots/context-fix-desktop-history.png differ diff --git a/docs/screenshots/context-fix-desktop-plan.png b/docs/screenshots/context-fix-desktop-plan.png new file mode 100644 index 00000000..85792b75 Binary files /dev/null and b/docs/screenshots/context-fix-desktop-plan.png differ diff --git a/docs/screenshots/context-fix-mobile-history.png b/docs/screenshots/context-fix-mobile-history.png new file mode 100644 index 00000000..301788e7 Binary files /dev/null and b/docs/screenshots/context-fix-mobile-history.png differ diff --git a/docs/screenshots/context-fix-mobile-original.png b/docs/screenshots/context-fix-mobile-original.png new file mode 100644 index 00000000..83c4c419 Binary files /dev/null and b/docs/screenshots/context-fix-mobile-original.png differ diff --git a/docs/screenshots/context-fix-mobile-read-pause.png b/docs/screenshots/context-fix-mobile-read-pause.png new file mode 100644 index 00000000..04af1340 Binary files /dev/null and b/docs/screenshots/context-fix-mobile-read-pause.png differ diff --git a/docs/screenshots/context-fix-mobile-request-body.png b/docs/screenshots/context-fix-mobile-request-body.png new file mode 100644 index 00000000..f4846ab1 Binary files /dev/null and b/docs/screenshots/context-fix-mobile-request-body.png differ diff --git a/findings.md b/findings.md index 2e3e75e8..159e03b7 100644 --- a/findings.md +++ b/findings.md @@ -196,3 +196,9 @@ Claude Code 报告已验证工具输入校验、权限决定、队列并发屏 - 独立1.5源码离线安装、6份编译字节对照及158/158项回归全部通过。升级验收器首次读取export未携带既有接口控制令牌,返回HTTP403并在升级前停止,正式进程无变化;补齐bootstrap控制令牌后继续,未读取模型Key。 - 真实1.5模型失败根因:读取历史索引16000字符经JSON双层编码后超过当前硬输入预算,最新完整工具组被整体封存;后续计划始终保留同一组[0,2,12,13,14,15],模型看不到刚读取的片段,重复offset4000/5000。最终服务超时,27请求/249537已报告tokens、无提案。需要工具实际响应按当轮可用字节预算分页,让最新结果能发送;还需按不可变工件的已覆盖区间识别循环,并优先保留直接代码/检查依据。 + +- 1.5.1修复:真实工件页按下一请求中的双层JSON字节限额选择,保持Unicode与next_offset;完整工具参数/原工件保留。Task持久记录已读区间,3次连续重复或6个不断更新的索引触发no_progress暂停,显式续接保留原Task。相关原25项回归通过。首轮新测试3/4通过,HTTP fixture错把/v1/chat/completions当无效路径;诊断显示0实际fixture请求,已改按实际端点后缀核对。一次组合补丁误用Store的数据库import,校验拒绝且未写入,已拆分按实际源码完成。 + +- 冻结1.5.1完整163项回归与63组UI全部通过。目视发现新增mobile-read-pause截图实际为1440px:该用例沿用了迟到错误测试的桌面视口,标签不能作为手机证据。已显式切到390px、验证宽度/溢出,手机查看暂停原因后关闭任务抽屉再续接,重新独立跑9组上下文UI。 + +- 390px最终暂停截图已目视确认原因、运行下拉和关闭控件清晰;完整63组UI0页面/请求异常。一个只读路径误写tests/git-delivery.test.mjs,查询失败且无改动;rg实际定位为tests/delivery.test.mjs和delivery-recovery.test.mjs。 diff --git a/package-lock.json b/package-lock.json index 6234fdbf..ba675d92 100644 --- a/package-lock.json +++ b/package-lock.json @@ -1,12 +1,12 @@ { "name": "local-agent-workbench", - "version": "1.5.0", + "version": "1.5.1", "lockfileVersion": 3, "requires": true, "packages": { "": { "name": "local-agent-workbench", - "version": "1.5.0", + "version": "1.5.1", "dependencies": { "@babel/parser": "7.29.11", "diff": "^8.0.2", diff --git a/package.json b/package.json index 94d13881..8d2e11a5 100644 --- a/package.json +++ b/package.json @@ -1,6 +1,6 @@ { "name": "local-agent-workbench", - "version": "1.5.0", + "version": "1.5.1", "private": true, "type": "module", "description": "A local coding workbench with explicit tool execution, reviewable changes and recoverable sessions.", diff --git a/progress.md b/progress.md index f0196254..0acbf2cd 100644 --- a/progress.md +++ b/progress.md @@ -147,3 +147,7 @@ - 1.5最终158项回归及62组桌面/手机UI通过,原历史证据已隔离保存;桌面/390px上下文截图目视确认。公开说明、能力/路线和QA清单已更新为已实现范围与尚待发行实测的区别。检查diff并保存本地实现,随后不可变打包、独立源码验收及保留原进程升级。 - 1.5独立源码158/158回归、6份构建一致及便携2180/源码343份归档验证通过,原PID23080/连接和7项目9会话完整保留更新。真实模型首轮失败:27请求/249537tokens、358秒服务超时,无提案或源码/Git写入。工件读取偏大导致最新结果不能进入下一请求,连续重复索引/相同片段;原记录已单独封存。下一步优先1.5.1读取预算、当前修复依据和无进展检测,再接续已保存的工作区实现草稿。 + +- 1.5.1工件预算/Unicode分页、直接代码优先保留及Task已读范围/no_progress已实现;4项真实HTTP/重复暂停/重新打开/显式续接专项与1项优先选择验证通过。完整163项回归与63组UI进行中,现有独立UI基础10/诊断9已通过。正式1.5.0保持,等待相关检查后创建新包并续接原失败Task。 + +- 1.5.1冻结163/163完整回归和63组UI通过。390px新增暂停用例已显式复验并目视核对,查看原因/关闭抽屉/续接原Task可用;1.5.0历史证据保持。已准备只复用原失败工程和标准审核/检查/Git循环的续接验收器,不创建新Task、不重跑基线、不修改四源码之外文件;等待独立包验收与升级后执行。 diff --git a/scripts/upgrade-running.mjs b/scripts/upgrade-running.mjs index 93286a3d..7b8c4fdf 100644 --- a/scripts/upgrade-running.mjs +++ b/scripts/upgrade-running.mjs @@ -64,7 +64,7 @@ export async function upgradeRunning({ url, pid, currentRoot, newRoot, handlerRo if (!fs.existsSync(originalScript) || !fs.existsSync(upgradeScript) || !fs.existsSync(path.join(next, 'dist-client', 'index.html'))) throw new Error('Upgrade bundles are incomplete.'); if (!Number.isInteger(pid) || pid < 1 || pid === process.pid) throw new Error('Invalid target process.'); const before = await readUpgradeJson(origin + '/api/bootstrap', 5000, 'Read original Workbench'); - if (!before.token || !['0.1.0', '1.0.0', '1.1.0', '1.2.0', '1.3.0', '1.4.0', '1.5.0'].includes(before.version)) throw new Error('This migration accepts a supported existing Workbench.'); + if (!before.token || !['0.1.0', '1.0.0', '1.1.0', '1.2.0', '1.3.0', '1.4.0', '1.5.0', '1.5.1'].includes(before.version)) throw new Error('This migration accepts a supported existing Workbench.'); if (before.sessions.some(session => ['running', 'awaiting_approval', 'stopping'].includes(session.status))) throw new Error('Wait for existing tasks and approvals before upgrading.'); let protocol, socket, enabledHere = false, paused = false; const breakpoints = []; diff --git a/scripts/verify-change-sets-ui.mjs b/scripts/verify-change-sets-ui.mjs index 62af8c41..755a2926 100644 --- a/scripts/verify-change-sets-ui.mjs +++ b/scripts/verify-change-sets-ui.mjs @@ -33,9 +33,9 @@ child.stdout.on('data', value => { output += value; for (const line of output.sp child.stderr.on('data', value => { errors += value; }); const wait = async predicate => { const deadline = Date.now() + 15000; while (!await predicate()) { if (Date.now() > deadline) throw new Error('UI fixture timed out: ' + errors); await new Promise(resolve => setTimeout(resolve, 40)); } }; await wait(() => url); const bootstrap = await (await fetch(url + '/api/bootstrap')).json(); -assert.ok(['1.4.0', '1.5.0'].includes(bootstrap.version)); assert.equal(bootstrap.config.mode, 'demo'); assert.equal(bootstrap.config.hasKey, false); -const proofPrefix = bootstrap.version === '1.4.0' ? 'm2-batch' : 'context-batch'; -const screenshotPrefix = bootstrap.version === '1.4.0' ? '1.4-batch' : 'context-batch'; +assert.ok(['1.4.0', '1.5.0', '1.5.1'].includes(bootstrap.version)); assert.equal(bootstrap.config.mode, 'demo'); assert.equal(bootstrap.config.hasKey, false); +const proofPrefix = bootstrap.version === '1.4.0' ? 'm2-batch' : bootstrap.version === '1.5.0' ? 'context-batch' : 'context-fix-batch'; +const screenshotPrefix = bootstrap.version === '1.4.0' ? '1.4-batch' : proofPrefix; const api = async (route, input) => { const response = await fetch(url + '/api' + route, { method: input === undefined ? 'GET' : 'POST', headers: { 'Content-Type': 'application/json', 'X-Workbench-Token': bootstrap.token, Connection: 'close' }, ...(input === undefined ? {} : { body: JSON.stringify(input) }) }); const value = await response.json(); assert.ok(response.ok, value.error); return value; }; const browser = await chromium.launch({ channel: 'msedge', headless: true }), page = await browser.newPage({ viewport: { width: 1440, height: 1000 } }); diff --git a/scripts/verify-context-ui.mjs b/scripts/verify-context-ui.mjs index 72b88423..e5c9e45d 100644 --- a/scripts/verify-context-ui.mjs +++ b/scripts/verify-context-ui.mjs @@ -12,16 +12,17 @@ import { call } from '../tests/helpers.mjs'; const root = path.resolve('.qa', 'context-ui-' + randomUUID()), data = path.join(root, 'data'); fs.mkdirSync(root, { recursive: true }); fs.mkdirSync('.qa/screenshots', { recursive: true }); const store = new Store(data), cases = {}; -for (const name of ['long', 'other', 'paused']) { +for (const name of ['long', 'other', 'paused', 'paused-repeat']) { const projectRoot = path.join(root, 'context-' + name); fs.mkdirSync(projectRoot); fs.writeFileSync(path.join(projectRoot, 'README.md'), '# Owned context UI fixture\n'); const project = store.addProject(projectRoot), session = store.createSession(project.id, '上下文验收 ' + name); const goal = '这是独立界面验收样本。完整任务要求与实际原记录保留,读取工件不得重新执行工具。'; store.message(session, 'user', goal); session.providerMessages.push({ role: 'user', content: goal }); const run = store.beginRun(session, goal, false, store.config), planner = new RequestPlanner(store); - if (name === 'paused') { - session.status = 'paused'; store.message(session, 'system', '完整任务要求超过请求上下文预算,请核对配置后继续。'); - store.finishRun(session, run, 'context_limit'); cases[name] = { project, session, projectRoot, runs: [run.id] }; continue; + if (name === 'paused' || name === 'paused-repeat') { + session.status = 'paused'; store.message(session, 'system', name === 'paused' ? '完整任务要求超过请求上下文预算,请核对配置后继续。' : '模型重复读取已覆盖工件区间,原任务已暂停,请核对后继续。'); + if (name === 'paused-repeat') store.core.task.get(run.taskId).readProgress = { schemaVersion: 1, reads: [], repeated: 3, changingIndexes: [], warning: '重复读取已覆盖区间', pause: true }; + store.finishRun(session, run, name === 'paused' ? 'context_limit' : 'no_progress'); cases[name] = { project, session, projectRoot, runs: [run.id] }; continue; } if (name === 'long') for (let index = 0; index < 18; index++) { const tool = call('ui-original-' + index, 'read_file', { path: 'README.md', fixture: 'Preserve exact historical JSON.' }); @@ -59,8 +60,9 @@ child.stderr.on('data', bytes => { errors += bytes.toString('utf8'); }); const until = async predicate => { const deadline = Date.now() + 20000; while (!await predicate()) { if (Date.now() > deadline) throw new Error('Context UI fixture timeout: ' + errors); await new Promise(resolve => setTimeout(resolve, 40)); } }; await until(() => url || child.exitCode !== null); assert.ok(url, errors); const bootstrap = await (await fetch(url + '/api/bootstrap', { headers: { Connection: 'close' } })).json(); -assert.equal(bootstrap.version, '1.5.0'); assert.equal(bootstrap.config.mode, 'demo'); assert.equal(bootstrap.config.hasKey, false); +assert.ok(['1.5.0', '1.5.1'].includes(bootstrap.version)); assert.equal(bootstrap.config.mode, 'demo'); assert.equal(bootstrap.config.hasKey, false); const api = async route => { const response = await fetch(url + '/api' + route, { headers: { 'X-Workbench-Token': bootstrap.token, Connection: 'close' } }); const value = await response.json(); assert.ok(response.ok, value.error); return value; }; +const proofPrefix = bootstrap.version === '1.5.0' ? 'context' : 'context-fix'; const browser = await chromium.launch({ channel: 'msedge', headless: true }), page = await browser.newPage({ viewport: { width: 1440, height: 1000 } }); const passed = [], pageErrors = [], requestFailures = [], screenshots = []; page.on('pageerror', error => pageErrors.push(error.message)); @@ -68,7 +70,7 @@ page.on('requestfailed', request => { if (request.resourceType() !== 'eventsourc const mark = value => { passed.push(value); console.log('PASS ' + value); }; const panel = () => page.getByLabel('请求上下文', { exact: true }), history = () => page.getByRole('dialog', { name: '完整历史记录', exact: true }); async function screenshot(name) { - const file = '.qa/screenshots/context-' + name + '.png'; await page.screenshot({ path: file, fullPage: true }); screenshots.push(file); + const file = '.qa/screenshots/' + proofPrefix + '-' + name + '.png'; await page.screenshot({ path: file, fullPage: true }); screenshots.push(file); } async function choose(name) { const select = page.getByLabel('当前项目', { exact: true }); @@ -132,13 +134,26 @@ try { await lateFailure('**/artifacts/' + currentPlan.requestArtifactId + '?*', () => panel().getByRole('button', { name: '查看发送内容', exact: true }).click(), 'late old request-body failure'); await choose('paused'); await expect(page.locator('.run-stop')).toContainText('完整任务要求超过请求上下文预算'); await expect(page.getByRole('button', { name: '继续任务', exact: true })).toBeVisible(); mark('context-limit pause explains the intact requirement and explicit continuation'); + await page.setViewportSize({ width: 390, height: 844 }); + await choose('paused-repeat'); await expect(page.locator('.run-stop')).toContainText('重复读取未推进任务'); + await page.locator('.run-stop').scrollIntoViewIfNeeded(); + assert.equal(await page.evaluate(() => window.innerWidth), 390); + assert.ok(await page.evaluate(() => document.documentElement.scrollWidth <= window.innerWidth + 2)); + await screenshot('mobile-read-pause'); + await page.keyboard.press('Escape'); await expect(page.getByRole('button', { name: '继续任务', exact: true })).toBeVisible(); + const pausedTaskId = (await api('/sessions/' + cases['paused-repeat'].session.id)).execution.task.id; + await page.getByRole('button', { name: '继续任务', exact: true }).click(); + await until(async () => (await api('/sessions/' + cases['paused-repeat'].session.id)).execution.runs.length === 2); + const resumed = await api('/sessions/' + cases['paused-repeat'].session.id); + assert.equal(resumed.execution.task.id, pausedTaskId); assert.equal(resumed.execution.run.resumeOf, cases['paused-repeat'].runs[0]); + mark('390px repeated-read pause explains the cause and explicit continuation retains the same Task'); assert.deepEqual(pageErrors, []); assert.deepEqual(requestFailures, []); const result = { success: true, version: bootstrap.version, testedAt: new Date().toISOString(), passed, pageErrors, requestFailures, screenshots, root, data, mode: 'saved synthetic context evidence; no real model or project commands executed by this UI verifier', browser: browser.version() }; - fs.writeFileSync('.qa/context-ui-validation.json', JSON.stringify(result, null, 2) + '\n'); console.log(JSON.stringify(result, null, 2)); + fs.writeFileSync('.qa/' + proofPrefix + '-ui-validation.json', JSON.stringify(result, null, 2) + '\n'); console.log(JSON.stringify(result, null, 2)); } catch (error) { await screenshot('failure').catch(() => {}); - fs.writeFileSync('.qa/context-ui-validation.json', JSON.stringify({ success: false, version: bootstrap.version, passed, pageErrors, requestFailures, error: error.message, screenshots }, null, 2) + '\n'); + fs.writeFileSync('.qa/' + proofPrefix + '-ui-validation.json', JSON.stringify({ success: false, version: bootstrap.version, passed, pageErrors, requestFailures, error: error.message, screenshots }, null, 2) + '\n'); throw error; } finally { await browser.close(); diff --git a/src/client/TaskPanel.tsx b/src/client/TaskPanel.tsx index ec352d8e..804e36eb 100644 --- a/src/client/TaskPanel.tsx +++ b/src/client/TaskPanel.tsx @@ -7,7 +7,7 @@ import { DeliveryPanel } from './DeliveryPanel'; import { RequestContextPanel } from './RequestContextPanel'; import './TaskPanel.css'; -const stopReasons: Record = { context_limit: '完整任务要求超过请求上下文预算', time_limit: '到达运行时间预算', token_limit: '到达本次 token 预算', task_token_limit: '到达任务累计预算', step_limit: '到达工具轮次上限', cancelled: '用户停止', model_error: '模型请求失败', interrupted: '进程中断' }; +const stopReasons: Record = { no_progress: '重复读取未推进任务,请核对后继续', context_limit: '完整任务要求超过请求上下文预算', time_limit: '到达运行时间预算', token_limit: '到达本次 token 预算', task_token_limit: '到达任务累计预算', step_limit: '到达工具轮次上限', cancelled: '用户停止', model_error: '模型请求失败', interrupted: '进程中断' }; const number = (value: number | undefined) => (value ?? 0).toLocaleString('zh-CN'); const time = (value: string | null | undefined) => value ? new Date(value).toLocaleString('zh-CN', { month: '2-digit', day: '2-digit', hour: '2-digit', minute: '2-digit', second: '2-digit' }) : '执行中'; diff --git a/src/core/context/read-budget.mjs b/src/core/context/read-budget.mjs new file mode 100644 index 00000000..60d3f9f6 --- /dev/null +++ b/src/core/context/read-budget.mjs @@ -0,0 +1,64 @@ +import { createHash } from 'node:crypto'; + +const messageBytes = (result, callId) => Buffer.byteLength(JSON.stringify({ role: 'tool', tool_call_id: callId, content: JSON.stringify(result) }), 'utf8'); +const boundary = (text, end) => end > 0 && end < text.length && /[\uD800-\uDBFF]/.test(text[end - 1]) ? end - 1 : end; + +// This is the actual tool page, not a rewrite of a historical result. Count +// the escaped JSON tool message which will be present in the next request. +export function artifactPage(artifact, text, args, { maximumMessageBytes = 16000, callId = '' } = {}) { + const requestedOffset = args.offset ?? 0, requestedLimit = args.limit ?? 2000; + let offset = Math.min(requestedOffset, text.length); + if (offset > 0 && /[\uDC00-\uDFFF]/.test(text[offset]) && /[\uD800-\uDBFF]/.test(text[offset - 1])) offset--; + let maximumEnd = boundary(text, Math.min(offset + requestedLimit, text.length)); + if (maximumEnd === offset && offset < text.length) maximumEnd = boundary(text, Math.min(offset + 2, text.length)); + const make = end => ({ artifact_id: artifact.id, kind: artifact.kind, sha256: artifact.sha256, + content: text.slice(offset, end), offset, next_offset: end, returned_characters: end - offset, + total_characters: text.length, truncated: end < text.length, requested_offset: requestedOffset, requested_limit: requestedLimit, + budget_limited: end < maximumEnd, maximum_message_bytes: maximumMessageBytes }); + let low = offset, high = maximumEnd; + if (messageBytes(make(low), callId) > maximumMessageBytes) throw new Error('当前请求预算无法容纳工件读取元数据,请先核对任务预算。'); + while (low < high) { + const middle = Math.ceil((low + high) / 2), end = boundary(text, middle); + if (messageBytes(make(end), callId) <= maximumMessageBytes) low = middle; else high = middle - 1; + } + const end = boundary(text, low), result = make(end); + if (offset < maximumEnd && end <= offset) throw new Error('当前请求预算无法容纳下一个完整字符,请先核对任务预算。'); + return result; +} + +export function readMessageBudget(plan, message) { + const calls = message.tool_calls?.length ?? 1; + const assistantTokens = Math.ceil(Buffer.byteLength(JSON.stringify(message), 'utf8') / 3) + 8; + // Reserve space for the current call's framing and the next state/notice. + const tokens = plan.budget.inputBudget - plan.mandatoryInputTokens - assistantTokens - calls * 8 - 512; + return Math.max(1024, Math.floor(tokens * 3 / calls)); +} + +export function recordArtifactRead(task, artifact, result, invocationId) { + const progress = task.readProgress ??= { schemaVersion: 1, reads: [], repeated: 0, changingIndexes: [], warning: null, pause: false }; + const identity = artifact.id + ':' + artifact.sha256, start = result.offset, end = result.next_offset; + let read = progress.reads.find(value => value.identity === identity); + if (!read) { read = { identity, ranges: [] }; progress.reads.push(read); if (progress.reads.length > 128) progress.reads.shift(); } + const covered = read.ranges.some(range => start >= range[0] && end <= range[1]); + progress.repeated = covered ? progress.repeated + 1 : 0; + if (!covered) { + const ranges = [...read.ranges, [start, end]].sort((a, b) => a[0] - b[0]), merged = []; + for (const range of ranges) { const previous = merged.at(-1); if (previous && range[0] <= previous[1]) previous[1] = Math.max(previous[1], range[1]); else merged.push(range); } + read.ranges = merged.slice(-64); + } + if (artifact.kind === 'context-history-index') { + if (!progress.changingIndexes.includes(identity)) progress.changingIndexes.push(identity); + progress.changingIndexes = progress.changingIndexes.slice(-8); + } else progress.changingIndexes = []; + progress.warning = progress.repeated >= 1 ? '刚才读取的工件区间已完整读取,请使用已返回内容,或按next_offset读取新的区间。' + : progress.changingIndexes.length >= 3 ? '连续读取更新的历史索引没有推进工程;请直接读取实际结果或项目文件,依据已有信息继续任务。' : null; + progress.pause = progress.repeated >= 3 || progress.changingIndexes.length >= 6; + progress.last = { artifactId: artifact.id, sha256: artifact.sha256, offset: start, nextOffset: end, invocationId, + contentSha256: createHash('sha256').update(result.content, 'utf8').digest('hex') }; + return progress; +} + +export function clearReadStreak(task) { + if (!task.readProgress) return; + task.readProgress.repeated = 0; task.readProgress.changingIndexes = []; task.readProgress.warning = null; task.readProgress.pause = false; +} diff --git a/src/core/context/request-plan.mjs b/src/core/context/request-plan.mjs index abbac457..bb02fa6c 100644 --- a/src/core/context/request-plan.mjs +++ b/src/core/context/request-plan.mjs @@ -74,7 +74,10 @@ function describe(group, ledger, artifacts) { const assistant = group.messages[0], tools = (assistant.tool_calls ?? []).map(call => { const result = group.messages.find(message => message.role === 'tool' && message.tool_call_id === call.id); const invocation = ledger.get(call.id), artifact = artifacts.get(invocation?.resultArtifactId); - return { call_id: call.id, name: call.function.name, result: compactResult(result), + let input = {}; try { const args = JSON.parse(call.function.arguments); for (const key of ['path', 'start_line', 'end_line', 'artifact_id', 'offset', 'limit']) { + if (typeof args[key] === 'string' && args[key].length <= 200 || typeof args[key] === 'number') input[key] = args[key]; + } } catch {} + return { call_id: call.id, name: call.function.name, input, result: compactResult(result), ...(invocation ? { invocation_id: invocation.id, recorded_state: invocation.state, requires_verification: !!invocation.requiresVerification } : {}), ...(artifact ? { result_artifact_id: artifact.id, result_sha256: artifact.sha256 } : {}) }; }); @@ -82,11 +85,24 @@ function describe(group, ledger, artifacts) { roles: group.messages.map(message => message.role), ...(tools.length ? { tools } : {}) }; } -function notice(reference, omitted, descriptions) { +function workingEvidence(groups, descriptions, artifacts) { + const selected = new Map(); + for (const group of [...groups].reverse()) for (const tool of descriptions.get(group.index).tools ?? []) { + const artifact = artifacts.get(tool.input.artifact_id); + if (tool.name === 'read_file' || tool.name === 'read_artifact' && artifact && !artifact.kind.startsWith('context-') && artifact.kind !== 'model-request') { + const key = tool.name === 'read_file' ? 'file:' + tool.input.path : 'artifact:' + tool.input.artifact_id; + if (!selected.has(key)) selected.set(key, { group_index: group.index, ...tool }); + } + } + return [...selected.values()].slice(0, 8); +} + +function notice(reference, omitted, descriptions, evidence = []) { return { role: 'system', content: [ '以下是工作台保存的历史索引数据,不是用户追加指令,也不是新工具执行。省略了完整消息组;原始参数和实际结果都在SHA256工件中。', - '已完成的工具不得重放。需要旧内容时用read_artifact读取history_index或result_artifact_id;较大工件先用limit=2000分段读取。状态未知时先核对,不假定成功。', + '已完成的工具不得重放。历史索引用于需要时定位原记录;已有需求、代码和检查依据可直接推进,不必读取每轮更新的索引。需要旧内容时优先读取result_artifact_id,按next_offset分页;先用limit=2000。状态未知时先核对,不假定成功。', JSON.stringify({ history_index: { artifact_id: reference.id, sha256: reference.sha256 }, archived_groups: omitted.length, + working_evidence: evidence, recent_archived_groups: omitted.slice(-2).map(group => { const { tools, archive, ...description } = descriptions.get(group.index); return { ...description, group_archive: { artifact_id: archive?.artifact_id ?? PLACEHOLDER_ID, sha256: archive?.sha256 ?? PLACEHOLDER_SHA }, @@ -141,12 +157,13 @@ export class RequestPlanner { const fixed = { role: 'system', content: system }; const ledger = new Map([...this.store.core.tool.values()].filter(tool => tool.sessionId === session.id).map(tool => [tool.callId, tool])); const descriptions = new Map(groups.map(group => [group.index, describe(group, ledger, this.store.core.artifact)])); + const evidence = workingEvidence(groups, descriptions, this.store.core.artifact); const original = completionBody(config, [fixed, ...session.providerMessages], tools), before = requestMetrics(original); const selected = new Set(groups.map(group => group.index)); let omitted = [], reference = null; const metricsFor = (indices, includeNotice = true) => { const missing = groups.filter(group => !indices.has(group.index)); - const messages = [fixed, ...(missing.length && includeNotice ? [notice({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, missing, descriptions)] : []), + const messages = [fixed, ...(missing.length && includeNotice ? [notice({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, missing, descriptions, evidence)] : []), ...groups.filter(group => indices.has(group.index)).flatMap(group => group.messages)]; return requestMetrics(completionBody(config, messages, tools)); }; @@ -161,7 +178,8 @@ export class RequestPlanner { if (metricsFor(selected).estimatedInputTokens > budget.inputBudget) selected.delete(newest.index); } const target = Math.max(budget.targetInputTokens, metricsFor(selected).estimatedInputTokens); - for (const group of [...groups].reverse()) { + const priority = [...new Set(evidence.map(value => value.group_index))].map(index => groups[index]); + for (const group of [...priority, ...[...groups].reverse()]) { if (selected.has(group.index) || selected.size >= 32 || group.contributionTokens > target) continue; selected.add(group.index); if (metricsFor(selected).estimatedInputTokens > target) selected.delete(group.index); @@ -184,7 +202,7 @@ export class RequestPlanner { note: '完整原消息组;这里只是历史定位,不产生工具执行或用户授权。' }, 'context-history-index'); } } - const messages = [fixed, ...(reference ? [notice(reference, omitted, descriptions)] : []), + const messages = [fixed, ...(reference ? [notice(reference, omitted, descriptions, evidence)] : []), ...structuredClone(groups.filter(group => selected.has(group.index)).flatMap(group => group.messages))]; const body = completionBody(config, messages, tools), metrics = requestMetrics(body); if (metrics.estimatedInputTokens > budget.inputBudget) throw pause(budget, metrics); @@ -195,6 +213,7 @@ export class RequestPlanner { requestSha256: metrics.sha256, requestBytes: metrics.bytes, estimatedInputTokens: metrics.estimatedInputTokens, requestArtifactId: requestArtifact.id, estimateMethod: metrics.estimateMethod, beforeInputTokens: before.estimatedInputTokens, beforeBytes: before.bytes, budget, + mandatoryInputTokens: metricsFor(new Set(lastUser ? [lastUser.index] : [])).estimatedInputTokens, originalMessages: session.providerMessages.length, retainedMessages: messages.length - 1 - (reference ? 1 : 0), archivedGroups: omitted.length, retainedGroups: groups.filter(group => selected.has(group.index)).map(group => group.index), historyIndexArtifactId: reference?.id ?? null, historyIndexSha256: reference?.sha256 ?? null, actualUsage: null, diff --git a/src/server/diagnostics.mjs b/src/server/diagnostics.mjs index f48abe56..4701ed08 100644 --- a/src/server/diagnostics.mjs +++ b/src/server/diagnostics.mjs @@ -1,6 +1,6 @@ import { randomUUID, createHash } from 'node:crypto'; -export const APP_VERSION = '1.5.0'; +export const APP_VERSION = '1.5.1'; export function executionEnvironment() { const windows = process.platform === 'win32'; diff --git a/src/server/provider.mjs b/src/server/provider.mjs index 3d65de07..6cf7f53b 100644 --- a/src/server/provider.mjs +++ b/src/server/provider.mjs @@ -11,7 +11,7 @@ export const toolDefinitions = [ { type: 'function', function: { name: 'propose_patch', description: '按读取时的原文件SHA256,用唯一old_text锚点提出精确替换。所有替换基于同一原文且不得重叠;自动保留UTF-8 BOM和统一换行格式。只保存待审查提案,不写盘;哈希变化或锚点不唯一拒绝。', parameters: { type: 'object', properties: { path: { type: 'string' }, base_sha256: { type: 'string', pattern: '^[a-fA-F0-9]{64}$' }, reason: { type: 'string' }, edits: { type: 'array', minItems: 1, maxItems: 32, items: { type: 'object', properties: { old_text: { type: 'string', minLength: 1 }, new_text: { type: 'string' } }, required: ['old_text', 'new_text'], additionalProperties: false } } }, required: ['path', 'base_sha256', 'reason', 'edits'] } } }, { type: 'function', function: { name: 'plan_delivery', description: '为当前Task已经审核应用且实际检查通过的文件准备只读Git提交预览,保留其他暂存与工作区修改。只返回计划、路径、diff工件和署名,不提交;用户须在任务记录界面核对并确认。', parameters: { type: 'object', properties: { message: { type: 'string', maxLength: 1600, description: '本次本地提交说明;可省略并采用任务目标。' } } } } }, { type: 'function', function: { name: 'delivery_report', description: '从当前Task的源码指纹、实际检查、提交和Run用量生成JSON/Markdown证据报告工件,不执行项目命令。报告注明未检查/待审核/结果未知和业务需求未验收;运行中的报告为阶段快照。', parameters: { type: 'object', properties: {} } } }, - { type: 'function', function: { name: 'read_artifact', description: '读取当前会话保存的实际执行工件或检查日志,按artifact_id校验归属和SHA256。可按offset/limit分段读取;每次最多16000字符。依据失败日志修复时使用,不重新执行检查。', parameters: { type: 'object', properties: { artifact_id: { type: 'string' }, offset: { type: 'integer', minimum: 0 }, limit: { type: 'integer', minimum: 1, maximum: 16000 } }, required: ['artifact_id'] } } }, + { type: 'function', function: { name: 'read_artifact', description: '读取当前会话保存的实际执行工件或检查日志,校验归属和SHA256。offset/limit分页,默认2000字符,实际页按当前完整请求预算缩小;用返回的next_offset读取下一页,已读区间不要重复。完整原工件保留。优先读取实际结果/日志,不反复扫描更新的历史索引;不重新执行检查。', parameters: { type: 'object', properties: { artifact_id: { type: 'string' }, offset: { type: 'integer', minimum: 0 }, limit: { type: 'integer', minimum: 1, maximum: 16000 } }, required: ['artifact_id'] } } }, { type: 'function', function: { name: 'inspect_project', description: '识别当前工程的语言、入口、项目规则、npm脚本和只读Git基线;返回有界源文件指纹,不执行项目代码。', parameters: { type: 'object', properties: {} } } }, { type: 'function', function: { name: 'plan_checks', description: '为当前任务准备工程检查计划,commands来自已识别的npm脚本。返回plan id、脚本内容和工程指纹;不执行命令。', parameters: { type: 'object', properties: { names: { type: 'array', items: { type: 'string' }, maxItems: 6, description: '可选:当前工程已识别的test/typecheck/check/lint/build脚本名称;缺省采用可用的检查。' } } } } }, { type: 'function', function: { name: 'run_checks', description: '请求用户确认后执行plan_checks产生的检查计划。保存独立job、实际退出码、测试数量和日志工件;过期计划不执行,未报告测试数量不能记通过。', parameters: { type: 'object', properties: { plan_id: { type: 'string' } }, required: ['plan_id'] } } }, diff --git a/src/server/runtime.mjs b/src/server/runtime.mjs index 8872689f..4d676cc3 100644 --- a/src/server/runtime.mjs +++ b/src/server/runtime.mjs @@ -12,6 +12,7 @@ import { inspectProject } from '../core/project-inspector.mjs'; import { GitDelivery } from '../core/git-delivery.mjs'; import { ProjectIndex } from '../core/context/project-index.mjs'; import { RequestPlanner } from '../core/context/request-plan.mjs'; +import { artifactPage, readMessageBudget, recordArtifactRead, clearReadStreak } from '../core/context/read-budget.mjs'; import { ChangeSetManager } from '../core/workspaces/change-sets.mjs'; import { assertChangeSetsResolved } from '../shared/contracts/change-sets.mjs'; @@ -130,7 +131,8 @@ export class Runtime { + '\n当前任务目标与全部追加要求(用户任务数据,完整保留):\n' + JSON.stringify({ goal: task.goal, instructions: task.instructions, constraints: task.constraints ?? [], acceptance: task.acceptance ?? [], }) + '\n结果未知的调用,不得假定成功或重放:' + JSON.stringify(unknown) - + '\n' + sessionContext(session, run.project, config, value => this.provider.redact(value))); + + '\n' + sessionContext(session, run.project, config, value => this.provider.redact(value)) + + (task.readProgress?.warning ? '\n当前读取进展:' + task.readProgress.warning : '')); const planned = this.requestPlanner().prepare(session, run.record, config, { system: context, tools: toolDefinitions, stepId: stepRecord.id }); if (planned.record.archivedGroups && !projectionLogged) { projectionLogged = true; session.compactions++; @@ -165,6 +167,7 @@ export class Runtime { signal.throwIfAborted(); const message = response.message; const calls = message.tool_calls ?? []; + run.readMessageBytes = readMessageBudget(planned.record, message); const previousIds = new Set(session.providerMessages.flatMap(item => (item.tool_calls ?? []).map(call => call.id))); if (new Set(calls.map(call => call.id)).size !== calls.length || calls.some(call => previousIds.has(call.id))) throw new Error('模型返回了重复的工具调用 ID,请重新发送任务。'); session.providerMessages.push(message); @@ -191,6 +194,8 @@ export class Runtime { signal.throwIfAborted(); } this.store.finishStep(session, stepRecord); + if (this.store.core.task.get(run.taskId).readProgress?.pause) throw new BudgetExceeded('no_progress', + '模型连续重复读取已覆盖工件区间或更新的历史索引,未推进当前任务。已读取内容和原任务已保存,请核对结果或调整任务要求后继续。'); } session.status = 'paused'; run.stopReason = 'step_limit'; @@ -250,9 +255,8 @@ export class Runtime { } else if (tool === 'read_artifact') { const artifact = this.store.owned('artifact', args.artifact_id, session.id), text = this.store.artifactStore.read(artifact); - const offset = args.offset ?? 0, limit = Math.min(16000, Math.max(1, args.limit ?? 8000)); - result = { artifact_id: artifact.id, kind: artifact.kind, sha256: artifact.sha256, content: text.slice(offset, offset + limit), offset, - total_characters: text.length, truncated: offset + limit < text.length }; + result = artifactPage(artifact, text, args, { callId: call.id, maximumMessageBytes: run.readMessageBytes ?? 16000 }); + recordArtifactRead(this.store.core.task.get(run.taskId), artifact, result, event.invocationId); } else if (tool === 'prepare_change_set') { const set = await this.changeSetManager().prepare(session, args.change_ids, { source: 'agent' }); @@ -294,6 +298,7 @@ export class Runtime { else if (tool === 'run_command') result = await this.requestCommand(session, event, args, run); else throw new Error('当前能力表没有工具:' + tool); event.phase = ['denied', 'interrupted', 'unknown', 'unverified', 'stale'].includes(result.status) ? result.status : result.status === 'failed' ? 'failed' : 'completed'; + if (tool !== 'read_artifact' && !['error', 'interrupted', 'unknown', 'denied'].includes(result.status)) clearReadStreak(this.store.core.task.get(run.taskId)); if (typeof result.stdout === 'string') { result.stdout = this.provider.redact(result.stdout); result.stderr = this.provider.redact(result.stderr ?? ''); } event.output = this.provider.redact(typeof result.stdout === 'string' ? result.stdout + (result.stderr ? '\n' + result.stderr : '') : JSON.stringify(result).slice(0, 16000)); if (result.exit_code !== undefined) event.exitCode = result.exit_code; diff --git a/src/server/store.mjs b/src/server/store.mjs index a42f4e99..e679b20f 100644 --- a/src/server/store.mjs +++ b/src/server/store.mjs @@ -337,6 +337,9 @@ export class Store extends EventEmitter { const goal = content || session.messages.find(message => message.role === 'user')?.content || session.title; task = createTask(session, goal, config, previous); this.core.task.set(task.id, task); } else if (content && (!resume || instructionSource)) task.instructions.push({ id: randomUUID(), content, createdAt: now(), ...(instructionSource ? { source: instructionSource } : {}) }); + if (task.readProgress) { + task.readProgress.repeated = 0; task.readProgress.changingIndexes = []; task.readProgress.warning = null; task.readProgress.pause = false; + } const run = createRun(session, task, config, resume ? previousRunId : null); this.core.run.set(run.id, run); task.activeRunId = run.id; session.taskId = task.id; session.runId = run.id; session.status = 'running'; diff --git a/task_plan.md b/task_plan.md index 5469ed6e..57471301 100644 --- a/task_plan.md +++ b/task_plan.md @@ -178,3 +178,10 @@ 2. [in_progress] 单长Task、多语言/大结果、完整协议、续接/实际退出及归属验证通过,158项完整回归和62组桌面/手机界面通过;请求/原历史/深度校验报告已接入。继续独立不可变1.5包、原PID升级和真实配置模型验证。 3. [pending] 封存已提交HEAD、独立命名分支/worktree、持久准备/核对记录和稳定项目身份;保留原HEAD/index/用户改动,不隐式执行hooks或内容过滤器。 4. [pending] 继续React/TypeScript与Python工程检查,以及证据绑定的REA/Codex/DeepSeek适配和有界Purge实验;只将实测通过的切片标已实现。 + +## 真实工件循环修复(1.5.1,当前优先) + +1. [complete] 1.5独立包/源码158项回归与保留原PID升级完成;真实首轮27请求/249537tokens、358秒超时、无提案的失败已封存,定位最新大页未进入请求的循环。 +2. [complete] 工件页按双层JSON预算/Unicode边界/next_offset返回,直接代码依据优先选择;Task持久已读区间及no_progress暂停,专项4项和1项优先选择回归通过。 +3. [in_progress] 完整163项回归、63组UI、独立发行与原PID升级;继续原真实失败Task,核对新请求及全部历史/源码/检查/Git证据。 +4. [pending] 接续.qa中已保存的独立Git/worktree创建/核对草稿,按实际原工作区保护、退出恢复与桌面/手机流程完成内置能力。 diff --git a/tests/read-budget.test.mjs b/tests/read-budget.test.mjs new file mode 100644 index 00000000..0d820032 --- /dev/null +++ b/tests/read-budget.test.mjs @@ -0,0 +1,92 @@ +import test from 'node:test'; +import assert from 'node:assert/strict'; +import { randomUUID, createHash } from 'node:crypto'; +import { artifactPage, recordArtifactRead, clearReadStreak } from '../src/core/context/read-budget.mjs'; +import { Runtime } from '../src/server/runtime.mjs'; +import { Provider } from '../src/server/provider.mjs'; +import { Store } from '../src/server/store.mjs'; +import { fixture, compatibleFixture, call, completion, until, assertToolPairs } from './helpers.mjs'; + +const sha = text => createHash('sha256').update(text, 'utf8').digest('hex'); +const bytes = (page, id) => Buffer.byteLength(JSON.stringify({ role: 'tool', tool_call_id: id, content: JSON.stringify(page) }), 'utf8'); + +test('budgeted artifact pages count escaped protocol bytes and reconstruct original multilingual JSON without losing characters', () => { + const text = JSON.stringify({ source: '原文🙂\r\n"\\'.repeat(4000) }), artifact = { id: randomUUID(), sha256: sha(text), kind: 'tool-result' }; + let offset = 0, reconstructed = '', pages = 0; + while (offset < text.length) { + const page = artifactPage(artifact, text, { offset, limit: 16000 }, { maximumMessageBytes: 7000, callId: 'literal-read' }); + assert.ok(bytes(page, 'literal-read') <= 7000); assert.ok(page.next_offset > offset); assert.equal(page.offset, offset); + assert.equal(page.content, text.slice(offset, page.next_offset)); assert.equal(page.truncated, page.next_offset < text.length); + assert.ok(!/[\uD800-\uDBFF]$/.test(page.content)); reconstructed += page.content; offset = page.next_offset; pages++; + } + assert.ok(pages > 5); assert.equal(reconstructed, text); assert.equal(sha(reconstructed), artifact.sha256); + const unicode = artifactPage({ ...artifact, sha256: sha('🙂雪') }, '🙂雪', { offset: 1, limit: 1 }, { maximumMessageBytes: 1200 }); + assert.equal(unicode.offset, 0); assert.equal(unicode.content, '🙂'); assert.equal(unicode.next_offset, 2); +}); + +test('known artifact coverage detects semantic repeats despite changing limits, and advancing pages or direct evidence restore progress', () => { + const task = {}, text = 'Immutable engineering evidence\n'.repeat(30), artifact = { id: randomUUID(), kind: 'check-log', sha256: sha(text) }; + const page = artifactPage(artifact, text, { offset: 0, limit: 200 }, { maximumMessageBytes: 2000 }); + recordArtifactRead(task, artifact, page, 'first'); + for (const limit of [100, 150, 180]) recordArtifactRead(task, artifact, + artifactPage(artifact, text, { offset: 0, limit }, { maximumMessageBytes: 2000 }), 'repeat-' + limit); + assert.equal(task.readProgress.pause, true); assert.equal(task.readProgress.repeated, 3); + recordArtifactRead(task, artifact, artifactPage(artifact, text, { offset: 200, limit: 200 }, { maximumMessageBytes: 2000 }), 'next'); + assert.equal(task.readProgress.pause, false); assert.deepEqual(task.readProgress.reads[0].ranges, [[0, 400]]); + for (let index = 0; index < 6; index++) recordArtifactRead(task, { ...artifact, id: randomUUID(), kind: 'context-history-index' }, page, 'index-' + index); + assert.equal(task.readProgress.pause, true); clearReadStreak(task); + assert.equal(task.readProgress.pause, false); assert.equal(task.readProgress.reads[0].identity, artifact.id + ':' + artifact.sha256); +}); + +test('actual large artifact reads are present in the next bounded HTTP request and next_offset progresses without redoing tools', async t => { + const f = fixture(t), original = JSON.stringify({ log: '真实证据🙂\n"\\'.repeat(2200) }); + const artifact = f.store.artifactStore.write(f.session, null, original, { kind: 'check-log' }); + f.store.core.artifact.set(artifact.id, artifact); f.store.save(f.session); + let requests = 0, observedPages = []; + const model = await compatibleFixture(t, (route, body) => { + if (!route.endsWith('/chat/completions')) return { body: {} }; + requests++; + const received = body.messages.filter(value => value.role === 'tool').map(value => JSON.parse(value.content)).filter(value => value.artifact_id === artifact.id); + if (requests > 1) { + const page = received.at(-1); assert.ok(page, 'Newest actual result must be sent, not archived before the model sees it'); + assert.ok(page.budget_limited); assert.ok(page.next_offset > page.offset); assert.equal(page.content, original.slice(page.offset, page.next_offset)); + observedPages.push(page); + if (requests === 4) return { body: completion([], '已根据三段实际结果完成读取验收。') }; + return { body: completion([call('page-' + requests, 'read_artifact', { artifact_id: artifact.id, offset: page.next_offset, limit: 16000 })]) }; + } + return { body: completion([call('page-first', 'read_artifact', { artifact_id: artifact.id, limit: 16000 })]) }; + }); + f.store.saveConfig({ mode: 'compatible', baseUrl: model.baseUrl, model: 'page-budget-fixture', maxSteps: 8 }); + const runtime = new Runtime(f.store, new Provider()); runtime.start(f.session.id, '读取保存的原工件,分段取得真实依据,不执行命令或修改文件。'); + await until(() => !runtime.isActive(f.session.id), 20000); + assert.equal(f.session.status, 'completed', JSON.stringify({ failure: f.session.failure, requests, pages: observedPages.map(value => ({ offset: value.offset, next: value.next_offset, budget: value.maximum_message_bytes })) })); assert.equal(requests, 4); assert.equal(observedPages.length, 3); + assert.equal(observedPages[1].offset, observedPages[0].next_offset); assert.equal(observedPages[2].offset, observedPages[1].next_offset); + assert.equal(f.store.artifactStore.read(artifact), original); assertToolPairs(f.session.providerMessages); + const plans = [...f.store.core.context_plan.values()]; + for (const page of observedPages) assert.ok(plans.some(plan => plan.status === 'completed' && plan.retainedGroups.length > 0)); +}); + +test('a real repeated artifact loop pauses the same Task after recorded results, survives reload and only resumes explicitly', async t => { + const f = fixture(t), artifact = f.store.artifactStore.write(f.session, null, 'Known complete evidence', { kind: 'check-log' }); + f.store.core.artifact.set(artifact.id, artifact); f.store.save(f.session); + let requests = 0, resume = false; + const model = await compatibleFixture(t, () => { + requests++; + return { body: resume ? completion([], '使用已保存内容完成核对。') + : completion([call('repeat-read-' + requests, 'read_artifact', { artifact_id: artifact.id, limit: 16000 })]) }; + }); + f.store.saveConfig({ mode: 'compatible', baseUrl: model.baseUrl, model: 'repeat-loop-fixture', maxSteps: 12 }); + const runtime = new Runtime(f.store, new Provider()); runtime.start(f.session.id, '使用不可变证据完成当前任务。'); + await until(() => !runtime.isActive(f.session.id), 20000); + const taskId = f.session.taskId; + assert.equal(f.session.status, 'paused'); assert.equal(f.store.core.run.get(f.session.runId).stopReason, 'no_progress'); + assert.equal(requests, 4); assert.equal([...f.store.core.tool.values()].filter(value => value.state === 'completed').length, 4); + assertToolPairs(f.session.providerMessages); f.store.close(); + const reopened = new Store(f.data); f.onCleanup(() => reopened.close()); + const saved = reopened.session(f.session.id); assert.equal(saved.taskId, taskId); assert.equal(requests, 4); + assert.equal(reopened.core.task.get(taskId).readProgress.pause, true); assertToolPairs(saved.providerMessages); + resume = true; const resumed = new Runtime(reopened, new Provider()); resumed.start(saved.id, '', true); + await until(() => !resumed.isActive(saved.id), 20000); + assert.equal(saved.status, 'completed'); assert.equal(saved.taskId, taskId); assert.equal(requests, 5); + assert.equal([...reopened.core.tool.values()].length, 4); assert.equal(reopened.core.task.get(taskId).readProgress.pause, false); +}); diff --git a/tests/request-plan.test.mjs b/tests/request-plan.test.mjs index 5dc4674f..feaf8326 100644 --- a/tests/request-plan.test.mjs +++ b/tests/request-plan.test.mjs @@ -37,6 +37,20 @@ function assertBounded(plan) { assertToolPairs(plan.messages); } +test('direct code evidence remains selected while recent historical-index browsing is archived within the complete request budget', t => { + const f = setup(t, { contextLimit: 10000, maxOutputTokens: 2048 }); + for (const name of ['normalize', 'summary', 'render', 'cli']) append(f, name, 'Actual implementation ' + name + '\n'.repeat(150), { path: 'src/' + name + '.mjs', start_line: 421, end_line: 450 }); + const artifact = f.store.artifactStore.write(f.session, f.run, 'Index fixture', { kind: 'context-history-index' }); + f.store.core.artifact.set(artifact.id, artifact); + for (let index = 0; index < 20; index++) append(f, 'index-' + index, 'metadata'.repeat(1500), { artifact_id: artifact.id, offset: index * 2000, limit: 2000 }, 'read_artifact'); + append(f, 'latest-index', 'Recent bounded index page', { artifact_id: artifact.id, offset: 40000, limit: 2000 }, 'read_artifact'); + const plan = f.prepare(); assertBounded(plan); + for (const name of ['normalize', 'summary', 'render', 'cli']) assert.ok(plan.messages.some(message => message.tool_calls?.some(value => value.id === 'history-' + name)), name + ' actual code group must remain'); + assert.ok(plan.messages.some(message => message.tool_calls?.some(value => value.id === 'history-latest-index'))); + assert.ok(plan.record.archivedGroups > 10); assert.ok(plan.record.mandatoryInputTokens < plan.record.budget.inputBudget); + assert.ok(plan.messages.some(message => message.role === 'system' && message.content.includes('working_evidence'))); +}); + test('one long user turn becomes bounded complete groups with exact UTF-8 history and recent JSON arguments', t => { const f = setup(t); for (let index = 0; index < 60; index++) append(f, index, ('行' + index + '🙂\r\n').repeat(1200), { path: 'README.md', payload: '原参数🙂'.repeat(800) });