feat: preserve bounded model requests and complete task history

This commit is contained in:
Agent Workbench Dev
2026-10-29 21:09:50 +08:00
parent 19b738b468
commit 8f32dec6cc
58 changed files with 1269 additions and 108 deletions
+3 -1
View File
@@ -28,7 +28,9 @@ ProjectInspector识别有界工程文件、语言、规则、npm脚本及当前
模型只准备批次,文件应用/回退由独立审查动作执行。运行、检查、Git和旧逐项应用共享项目互斥与未核对批次检查,界面不能绕过服务端规则。报告记录批次身份/指纹/状态并核对快照完整性。
## 上下文
完整会话仍保留在磁盘。模型请求使用有预算的投影:以完整 user/assistant/tool 组为单位保留近期内容,较早内容形成有长度限制的本地提要。计数为估算,服务返回的 usage 另外记录。该机制是本产品的独立提取式实现,不声称与 Claude Code 的模型摘要等价。
1.5.0的RequestContextPlanner固定Task目标、全部追加要求、项目/会话身份和最后用户消息,按原字节选择完整assistant/tool协议组。省略的原组保存为SHA256工件,超大Unicode记录按合法边界分片;有界通知提供实际索引与结果引用。原providerMessages不改写,缺失或损坏的封存证据不会静默重建。
预算以实际请求JSON正文估算,包含系统、工具定义、历史、model/tool_choice/stream及输出预留,工作输入目标最多16000估算tokens。持久保存正文、冻结计划与context_plan记录成功后才联网;Provider核对正文SHA256/字节。实际usage另行保存,重启将未完成计划标interrupted而不重放。API及任务面板可查看历史运行、发送内容和分页原记录;报告深度核对正文、索引、原组、分片和结果归属。详细边界见[上下文契约](CONTEXT-PLANS.md)。
## 模型与凭据
支持 OpenAI 兼容的 Chat Completions 工具契约,服务 URL 和模型由用户配置。API Key 仅在服务内存或专用环境变量中使用,不写入会话、导出或配置文件。默认离线演示明确标记,不连接第三方账户。
+9 -2
View File
@@ -1,6 +1,6 @@
# Agent Workbench 能力目标与验收矩阵
日期:2026-10-29。1.0.0为基础里程碑,1.1.0实现M1,1.2.0增加内置Git交付,1.3.0增加工程索引和精确补丁,1.4.0增加可恢复多文件变更集。下表保留1.0.0缺口与完整产品目标;每项完成范围以实际验证记录为准。
日期:2026-10-29。1.0.0为基础里程碑,1.1.0实现M1,1.2.0增加内置Git交付,1.3.0增加工程索引和精确补丁,1.4.0增加可恢复多文件变更集,1.5.0增加有界完整请求计划和历史追溯。下表保留1.0.0缺口与完整产品目标;每项完成范围以实际验证记录为准。
总体方案见 [PRODUCT-PLAN.md](PRODUCT-PLAN.md),研究问题与来源见 [RESEARCH-ROADMAP.md](RESEARCH-ROADMAP.md)。
@@ -42,6 +42,13 @@
- C04/C12:轮次上限后整批应用可继续原Task检查,其他停止原因保持原要求;切换项目清理临时错误提示、忽略过期批次响应。
- C13/C23:Git报告引用封存身份/指纹/实际状态与快照校验,桌面/390px整批确认及恢复入口。实际回归、包与真实模型结果见[QA-1.4](QA-1.4.md)。完整M2三类工程、M3引擎与M5桌面/配对尚待完成。
### 1.5.0 M2请求上下文与历史追溯切片
- C04:完整Task目标/追加要求/项目身份,完整协议组选择、原参数结果及SHA256封存;实际发送正文、历史运行与分页原记录可查看。
- C15:完整JSON输入估算、输出/安全余量和Run/Task累计预算预检,估算与服务实际用量分别记录;尚无价格表、费用或多worker总预算。
- C02/C13/C23:请求计划联网前持久保存、重启中断而不重放、损坏/缺失证据明确失败;报告深度核对正文、原组、分片及结果归属。
- 158项完整回归与62组界面通过。独立发行与真实模型结果见[QA-1.5](QA-1.5.md),worktree、React/Python完整工程和M2/M6成功率继续实施。
| ID / 能力 | 1.0.0基础与缺口 | 后续产品目标 | 参考问题 / 模块 / 阶段 | 必须看到的验收证据 |
| --- | --- | --- | --- | --- |
| C01 需求与任务规划 | 有对话和有限步数工具循环,没有任务依赖图 | 固定目标、约束和完成标准;拆任务、判断依赖、接收追加要求与重规划 | CX-1、WB-1 / tasks / M1–M2 | 追加要求后仍保留原目标;任务完成由验收条件决定 |
@@ -71,7 +78,7 @@
## 3. 当前实现的源码锚点
- [provider.mjs](../src/server/provider.mjs):固定工具定义、深度3/160项的项目上下文、提取式投影、非流式模型请求。
- [provider.mjs](../src/server/provider.mjs):固定工具定义、有界项目/诊断状态、实际正文指纹核对及非流式模型请求;[request-plan.mjs](../src/core/context/request-plan.mjs)负责完整请求预算、协议组选择、可核验封存和历史追溯。
- [runtime.mjs](../src/server/runtime.mjs):有限循环、内存active/waiters、命令审批与60秒进程生命周期。
- [store.mjs](../src/server/store.mjs):SQLite任务与事件、旧JSON迁移、备份与重启状态处理。
- [git-delivery.mjs](../src/core/git-delivery.mjs):当前Task已审核文件的提交计划、独立索引/引用事务、结果核对和恢复。
+25
View File
@@ -0,0 +1,25 @@
# 有界模型请求计划
1.5.0实现目标:长Task通过同一目标、追加要求、已保存结果和工件继续工程任务。完整providerMessages保持不变,请求只选择完整协议组;不改写工具参数、不截断工具结果或把省略记录表示成一次成功执行。
## 请求与预算
每次Chat Completions请求先构造完整JSON正文,包含model、系统/项目/Task状态、所选历史、工具定义、tool_choice、max_tokens和stream。UTF-8字节数、正文SHA256和估算方法随Step/Run持久保存;Authorization不属于正文,也不进入计划工件。
输入估算采用UTF-8字节数除以3,外加消息/工具协议余量;这是兼容服务没有公开分词器时的本地估算,不保证等于服务实际tokens。上下文窗口扣除配置的输出预留与安全余量,再与Run/Task剩余预算比较;工作上下文目标最多16000估算输入tokens。服务返回的实际用量另行记录,不能把字节减少直接换算成费用收益。
当前Task目标、全部追加要求/约束/验收要求、项目/会话身份和最后一条用户消息必须完整保留。如果这些必需内容已不能容纳,暂停并明确context_limit或对应累计预算;不截掉用户要求,也不自动新建任务。可保留最新完整工具组到硬预算,以免因工作目标较小就省略正在处理的结果。
## 完整工具组与封存
一个assistant的全部tool_calls及对应tool结果是不可分割组,按原顺序、原字节字符串选择。孤立/重复/缺少结果的协议在联网前拒绝;已知中断由既有恢复账本核对,不假造结果。过大的整组只能整体封存,模型通过实际工件分段读取。
省略组写入SHA256工件;超过单工件上限时按合法Unicode边界分片,清单保存原组SHA256、分片顺序及完整范围。每次请求的历史索引保存所有省略组身份、工具调用/结果引用与所选范围,正文只携带有界索引和最近省略组的精确状态/身份。模型可用read_artifact读取原结果或完整消息组,不重新执行旧工具。
工件先fsync、SQLite归属和request-plan记录保存成功,才发出模型请求。重启只将未完成计划标成interrupted,不发送请求、不重放工具。重用封存工件必须通过SHA256和归属核对;缺失或损坏时停止该投影,保留原记录。
## 验收
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。
+3 -3
View File
@@ -1,6 +1,6 @@
# Agent Workbench 后续产品与执行开发方案
日期:2026-10-29。本文保留完整产品规划;1.1.0实现M1,1.2.0增加Task本地Git交付,1.3.0增加索引与精确补丁,1.4.0增加可恢复多文件变更集。实际实现、验收与部署分别见[能力矩阵](CAPABILITIES.md)及[验证记录](VALIDATION.md),完整阶段目标继续实施。
日期:2026-10-29。本文保留完整产品规划;1.1.0实现M1,1.2.0增加Task本地Git交付,1.3.0增加索引与精确补丁,1.4.0增加可恢复多文件变更集,1.5.0增加有界完整请求与历史追溯。实际实现、验收与部署分别见[能力矩阵](CAPABILITIES.md)及[验证记录](VALIDATION.md),完整阶段目标继续实施。
## 1. 产品目标
@@ -144,8 +144,8 @@ Purge的包声明Apache-2.0,另有商业许可联系说明;Harness声明MIT
## 9. 下一轮最先执行的三个开发切片
1. 1.4.0已完成141项回归、54组UI、独立包与真实模型整批精确交付;最终独立时间戳包继续逐字节核对并保留原进程更新。HTTP520恢复和预算暂停分别保留,不以文字“完成”替代检查/Git证据。
2. 下一批先修复同一任务长历史无法充分投影的问题:请求预算包含系统/工具/消息,旧工具结果以可校验工件引用保留,实际同题核对质量与用量;随后实现独立Git/worktree,再完成React/TypeScript和Python检查及封存任务。当前单Node样本不计为三类工程完成。
1. 1.4.0最终不可变包和真实模型批次交付已完成;HTTP520恢复、385579已报告tokens及模型预算暂停保留。1.5.0请求计划/协议组封存/历史追溯已通过158项回归和62组UI,继续独立包、原进程更新与真实模型验收。
2. 下一批实现封存HEAD的独立Git/worktree、稳定来源/工作区身份和持久恢复记录,再完成React/TypeScript和Python检查及封存任务。原HEAD/index及未提交修改必须保留;当前单Node样本不计为三类工程完成。
3. 在写入/取消/事件归属边界稳定后按M3接入REA及Codex/DeepSeek EngineAdapter的契约探针,分别固定Harness/Purge/profile版本,再比较真实同题工程收益。桌面Electron与手机配对网关复用同一执行宿主,按M5推进。
这些切片优先复用已验证的文件范围、基线备份、UTF-8、取消与诊断能力。研究结论每次至少绑定一个实现模块、一个契约检查和一个用户场景,按实际失败分类持续调整顺序。
+22
View File
@@ -0,0 +1,22 @@
# 1.5.0 验收清单
本轮交付有界完整模型请求、原历史追溯与请求证据;范围见[上下文契约](CONTEXT-PLANS.md)。完整M2/M6、隔离工作区、外部引擎及桌面/手机配对继续实施。
| 检查 | 结果 | 证据 |
| --- | --- | --- |
| 构建及行为/API/实际退出/升级 | 158/158通过,0失败/跳过/取消 | [完整回归](context-tests-validation.json) |
| 基础界面 | 10/10通过 | [基础证据](context-basic-ui-validation.json) |
| 诊断/错误/Unicode命令/恢复 | 9/9通过 | [诊断证据](context-diagnostics-ui-validation.json) |
| Task/工件/预算/审核后继续 | 7/7通过 | [运行证据](context-execution-ui-validation.json) |
| Git审查/提交/报告/恢复 | 9/9通过 | [交付证据](context-delivery-ui-validation.json) |
| 索引/文件定位/归属 | 8/8通过 | [索引证据](context-index-ui-result.json) |
| 多文件批次/恢复/归属 | 11/11通过 | [批次证据](context-batch-ui-validation.json) |
| 上下文/历史运行/分页/工件/迟到错误/预算暂停 | 8/8通过 | [上下文证据](context-ui-validation.json) |
独立端口和自有数据目录运行全部62组界面。上下文UI使用已保存合成记录,不请求真实模型、不执行工程命令;其他验证器使用明确标记的离线或本地HTTP fixture。真实大模型的用量和质量另行记录。
完整Task要求和原参数保留;整体工具组选择、7MB以上Unicode分片、冻结正文/计划、归属/SHA256深度核对、工件缺失或篡改、原Task续接及计划保存后的真实进程退出均有行为检查。重启不发送请求、不重放完成工具。估算包含系统/工具/历史,服务返回的实际usage单独保存。
已目视核对[桌面请求记录](screenshots/context-desktop-plan.png)和[390px完整历史](screenshots/context-mobile-history.png)。原记录与实际发送JSON通过滚动弹窗查看,Escape关闭弹窗后仍可操作任务抽屉。
便携/源码独立验收、原PID升级、真实配置模型与最终发行身份随后写入[验证记录](VALIDATION.md)。开发测试通过不替代这些结果,也不宣称费用收益。
+14
View File
@@ -1,5 +1,19 @@
# Agent Workbench 验收记录
## 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。
| 验证 | 结果 | 证据 |
| --- | --- | --- |
| 完整请求/预算/协议组/Unicode分片/归属/缺失损坏/恢复与原有契约 | 158/158通过 | [完整回归](context-tests-validation.json) |
| 请求计划、历史运行、分页工件、手机关闭及迟到错误归属 | 8/8通过 | [上下文界面](context-ui-validation.json) |
| 批次审查/恢复及原有基础/诊断/任务/Git/索引 | 11/10/9/7/9/8通过 | [QA-1.5](QA-1.5.md) |
首轮完整回归误丢诊断中的历史applied状态,157/158通过;修正为保留最近32项所有状态后完整复验通过。手机Escape问题按真实document/window事件顺序复现并修复;三个旧作用域迟到失败用例分别通过。旧1.4批次证据恢复为19b738b内容,本轮结果单独保存context-*。
本段记录开发与界面验收;1.5独立便携/源码包、正式升级和真实配置模型结果随后追加,未以合成用量声称准确token或成本下降。完整M2的隔离worktree、React/Python工程及M3/M5继续开发。
## 1.4.0 · 多文件审查与恢复
Windows x64、Node24.22.1。构建与141/141项行为/API/实际进程退出/升级回归通过,0失败/跳过/取消;批次11及原有43共54组真实界面流程全部通过,0页面/请求异常。桌面及390px审查/回退截图已目视核对。
+27
View File
@@ -0,0 +1,27 @@
{
"success": true,
"url": "http://127.0.0.1:63763",
"browser": "Microsoft Edge 156.0.4314.42",
"date": "2026-10-29T12:55:47.238Z",
"projectRoot": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\ui-ac1ba709-7631-4715-b46d-435f0d4b463d\\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"
}
+29
View File
@@ -0,0 +1,29 @@
{
"success": true,
"version": "1.5.0",
"testedAt": "2026-10-29T12:55:39.594Z",
"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-batch-desktop-review.png",
"docs/screenshots/context-batch-desktop-conflict.png",
"docs/screenshots/context-batch-desktop-recovery.png",
"docs/screenshots/context-batch-mobile-review.png",
"docs/screenshots/context-batch-mobile-rollback.png"
],
"projectRoot": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\batch-ui-d0ea8d59-4cec-4d70-9fc0-fd5e2d152ed7",
"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."
}
+31
View File
@@ -0,0 +1,31 @@
{
"success": true,
"version": "1.5.0",
"testedAt": "2026-10-29T12:57:05.420Z",
"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-1900107a-cfb2-4efb-b9ca-6d5376752302\\四模块交付工程",
"projectId": "ec7bbbb9-583d-46c2-bc46-ac840288e88f",
"sessionId": "abd1c71e-f264-4d83-82ea-ecfb2fe62562",
"taskId": "846567cd-b931-42fa-be96-2c14cf505484",
"commit": "9e1d571256f7d076cc9e5e3ee84fa94fa9cf4864",
"recoveryCommit": "7341988dc9b95b85923a694910481e35ff3a5475",
"testCounts": {
"baselineFailed": 13,
"passed": 14
},
"model": "controlled local HTTP fixture",
"confirmations": 2
}
@@ -0,0 +1,20 @@
{
"success": true,
"version": "1.5.0",
"url": "http://127.0.0.1:63763",
"date": "2026-10-29T12:56:05.455Z",
"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"
}
+19
View File
@@ -0,0 +1,19 @@
{
"success": true,
"version": "1.5.0",
"testedAt": "2026-10-29T12:56:29.423Z",
"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-6954843b-e9de-40b5-b98f-696285ff05fe\\project",
"model": "controlled local HTTP fixture"
}
+19
View File
@@ -0,0 +1,19 @@
{
"checkedAt": "2026-10-29T12:57:10.797Z",
"version": "1.5.0",
"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-c71ccf52-fc1b-4e34-ba2a-c81387d5b0e5",
"primaryProjectId": "266716ca-0ad7-47da-9308-bf542bc04818",
"otherProjectId": "a153af59-8397-4503-afbe-b7f845b46573"
}
+16
View File
@@ -0,0 +1,16 @@
{
"success": true,
"version": "1.5.0",
"testedAt": "2026-10-29T12:44:08.334Z",
"counts": {
"tests": 158,
"pass": 158,
"fail": 0,
"cancelled": 0,
"skipped": 0,
"todo": 0
},
"exitCode": 0,
"logFile": ".qa/context-full-tests.tap",
"testFiles": 18
}
+28
View File
@@ -0,0 +1,28 @@
{
"success": true,
"version": "1.5.0",
"testedAt": "2026-10-29T12:51:02.478Z",
"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"
],
"pageErrors": [],
"requestFailures": [],
"screenshots": [
"docs/screenshots/context-desktop-plan.png",
"docs/screenshots/context-desktop-history.png",
"docs/screenshots/context-mobile-history.png",
"docs/screenshots/context-mobile-original.png",
"docs/screenshots/context-mobile-request-body.png"
],
"root": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\context-ui-1214ad45-63df-4f72-bd75-d99d36b46e9c",
"data": "D:\\Ai-project\\REA逆向\\Agent-Workbench\\.qa\\context-ui-1214ad45-63df-4f72-bd75-d99d36b46e9c\\data",
"mode": "saved synthetic context evidence; no real model or project commands executed by this UI verifier",
"browser": "156.0.4314.42"
}
Binary file not shown.

After

Width:  |  Height:  |  Size: 160 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 154 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 144 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 67 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 72 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 154 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 133 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 64 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 77 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 94 KiB