From 72115647d9c1826286733cc85e04493baa60a2eb Mon Sep 17 00:00:00 2001 From: Agent Workbench Dev Date: Fri, 6 Nov 2026 01:18:37 +0800 Subject: [PATCH] fix(runtime): preserve model capacity history and cache evidence with resumable file proposals --- docs/QA-PROJECT-GENERATION-REPAIR.md | 24 ++++- findings.md | 2 + progress.md | 4 + scripts/verify-model-capacity-ui.mjs | 104 +++++++++++++++++++ src/client/App.tsx | 4 +- src/client/BudgetPausePanel.tsx | 2 +- src/client/ModelBudgetSettings.tsx | 12 ++- src/client/ModelCenter.tsx | 10 +- src/client/RequestContextPanel.tsx | 4 +- src/client/types.ts | 6 +- src/core/context/budget-resume.mjs | 5 +- src/core/context/prefix-audit.mjs | 23 ++++ src/core/context/request-plan.mjs | 61 ++++++++--- src/core/git-delivery.mjs | 1 + src/core/models/config-import.mjs | 2 + src/core/models/model-catalog.mjs | 5 +- src/core/models/model-hub.mjs | 5 +- src/core/models/protocols.mjs | 20 ++-- src/core/models/runtime-settings.mjs | 10 +- src/core/storage/legacy-import.mjs | 6 +- src/core/workspaces/file-proposal-drafts.mjs | 70 +++++++++++++ src/server/app.mjs | 4 +- src/server/files.mjs | 4 +- src/server/provider.mjs | 9 +- src/server/runtime.mjs | 27 ++++- src/shared/contracts/execution.mjs | 2 + src/shared/contracts/file-draft-tools.mjs | 14 +++ src/shared/contracts/model-capacity.d.mts | 4 + src/shared/contracts/model-capacity.mjs | 24 +++++ src/shared/contracts/model-request.mjs | 5 +- src/shared/contracts/task-control.mjs | 10 +- src/shared/contracts/workers.mjs | 4 +- task_plan.md | 8 +- tests/file-proposal-drafts.test.mjs | 96 +++++++++++++++++ tests/helpers.mjs | 9 +- tests/model-capacity.test.mjs | 82 +++++++++++++++ tests/model-catalog.test.mjs | 2 +- tests/prefix-stability.test.mjs | 81 +++++++++++++++ tests/task-control.test.mjs | 9 +- 39 files changed, 699 insertions(+), 75 deletions(-) create mode 100644 scripts/verify-model-capacity-ui.mjs create mode 100644 src/core/context/prefix-audit.mjs create mode 100644 src/core/workspaces/file-proposal-drafts.mjs create mode 100644 src/shared/contracts/file-draft-tools.mjs create mode 100644 src/shared/contracts/model-capacity.d.mts create mode 100644 src/shared/contracts/model-capacity.mjs create mode 100644 tests/file-proposal-drafts.test.mjs create mode 100644 tests/model-capacity.test.mjs create mode 100644 tests/prefix-stability.test.mjs diff --git a/docs/QA-PROJECT-GENERATION-REPAIR.md b/docs/QA-PROJECT-GENERATION-REPAIR.md index a384087e..98abec22 100644 --- a/docs/QA-PROJECT-GENERATION-REPAIR.md +++ b/docs/QA-PROJECT-GENERATION-REPAIR.md @@ -19,7 +19,29 @@ 原全局、新连接、上游目录导入均默认2048,运行时再取各层最小值;旧1.25.4恢复提示仅要求每轮一个文件,而单个index.html本身已超过额度,反复从头生成仍截断。用户明确要求正常开发不套用该实验额度。 -本切片实施:新开发默认采用较充分的可调容量;模型容量模式消除全局重复裁低;上游明确的容量直接进入连接,未知容量保留为用户可核对的配置;手动模式和旧Run冻结额度保持。长文件须能分段组装、完整后才产生可审查提案,不把不完整JSON当文件执行。完成后补实际请求及界面验证。 +已实现: + +- 新安装默认按连接容量,未知容量的可编辑初始值128000/16384明确不是已验证能力。上游声明直接带入新连接;已有连接的覆盖值保持,编辑时可明确采用上游容量。去掉本地16000输出/200000上下文的通用clamp;安全整数、真实容量和累计预算仍核对。 +- 模型模式消除全局第二层裁低,自动池按实际成员发送各自输出额度,不再把大模型降到最小成员。池输入仍采用共同窗口及最大输出预留,是保守预算规划;容量不足暂停,不删原文。手动模式保留显式取小策略。旧配置保持manual/compact,旧Run、旧实验和旧用量不迁改。 +- `propose_file_chunk`保存所属Task的不可变片段,核对基线SHA、原草稿SHA和偏移,重启后可续接;最后完整组装才走原文件提案/审查流程。半截JSON、重复/乱序、源文件变化、异Task、Unicode损坏、保存失败均不能发布半完成提案。未完成草稿不能完成任务或Git交付;放弃也保留原片段。原512KB文件提案边界保持,不等于token截断。 + +相关回归102/102通过;新容量/前缀/草稿边界14项首轮13通过,新增池测试误用不存在的upsertPool,改用原pool接口后原断言精确补验1/1通过。其中长文件保存失败与重启5/5,实际HTTP完整30000+字符答复与原响应保持、三协议131072输出字段、池4096/32768各自发送已核对。测试夹具显式保持旧16K/2048/manual/compact,避免把既有有限预算/压缩回归改测成新默认;新安装行为另由modernDefaults独立覆盖。 + +## 切片三:完整历史与缓存事实 + +新增明确的历史策略。preserve保持原历史文字及顺序,装不下时暂停,不把原消息静默换成摘要;compact保留既有可检索整理策略。完整模式把固定工程规则和worker权限契约留在系统前缀,把目录、当前状态、记忆及动态能力目录放在历史后;协议转换、凭据脱敏和实际权限校验保持。 + +每次封存请求额外记录实际工具名称、schema SHA/字节数,以及同Task最近实际发送请求的有序prompt共同前缀字节。跨Run的两次“继续”仍是两个原消息实例;不将诊断元数据注入模型请求。不相同路线/型号不做伪对照,变更工具surface有明确原因,历史计划可按原请求重算核对。 + +Chat/Responses/DeepSeek及Messages的上游缓存用量按实际回执保留原字段与规范化计数;缺失不补为命中0,异常计数不估算。界面分开显示真实cached tokens与共同前缀B;后者不是命中或费用节约。未添加供应商不明示支持的cache参数,未实现Claude显式cache_control开关,也不保证供应商缓存不过期。三次官方OpenAI文档抓取均403,原响应保留,不把403页当文档依据。 + +## 界面中发现的附带阻断 + +原句`Inspect the directory once and answer only. Do not modify files or run checks.`被本地规则误认成必须检查,只读回答后verification_incomplete。原网页/Electron失败与会话保存;共享否定判断现覆盖并列or/nor,并用于首次要求与追加要求,明确重新要求检查仍可恢复required。原检查/交付门未放宽;task-control 10/10通过。 + +新增验收器覆盖实际网页与Electron、1440/390、策略保存刷新、旧2048显示及采用上游32768、旧修订拒绝但不丢编辑、实际HTTP请求与缓存事实。首轮误取alert角色、夹具缺少工具声明、移动断点转换未等待React提交/重新打开面板的验收器失败均保留。最终入口结果在发行段独立记录;源码阶段完成不等于正式部署。 + +源码最终网页3组/实际Electron3组各6图通过,结果分别在`capacity-web-83090bed-9f25-4643-84e2-504f0beb9aa2`和`capacity-electron-5da38c19-a972-4b4a-8f16-ebc2a847e0e2`;每套2次自有HTTP模型请求、1次自有目录GET、真实供应商0。截图回看后把策略控件并入原表单网格,模型模式不再显示未启用的旧2048,切回手动保留原值;此最终改动已复验。旧超预算/未知用量/续接预览入口两视口2组通过(`.qa/model-budget-ui-36ef7526-88a9-4425-af34-b596acc314dd`),未重放原Task或派发续接请求。 ## 结论边界 diff --git a/findings.md b/findings.md index 4caae1b4..db418e28 100644 --- a/findings.md +++ b/findings.md @@ -1,5 +1,7 @@ # 汇总发现 +用户补充“透传/不截流/高缓存”:不只是调大默认值,还要保留实际请求与完整响应、用户设置不被静默clamp。发现request-plan对大窗口仍以16000输入target提前压缩,system含动态项目清单且runtime每轮重建任务/诊断段,项目记忆/Skills/MCP目录附带新工件身份插在历史前,会打断前缀。缓存上游报告现有Claude计数已保留;需要把Chat/Responses/DeepSeek实际报告正规化,不将共同前缀或本地文件索引缓存当供应商命中。容量设置会有model/manual策略,旧配置保持manual,目录明确容量不再被源连接2048裁低。两次apply_patch整批因猜测标题/占位import校验失败,已按实际文本重新补,失败操作无产品文件变化。 + “引航演示”起步实证:正式实际1.25.4/原PID36552,项目d2a23086(D:/Ai-project/引航演示)只有一个失败会话be9397b1。原目标为从零开发数字人课程制作产品。原inspect_project/list_files都返回空工程,项目rg匹配无文件(exit1是空列表,不是工程构建错误)。两Run共9模型请求,4个实际工具事件:inspect/list成功、plan_task首次工具枚举校验拒绝后修正成功;0提案/检查/写入。首Run池在DeepSeek→GPT→Claude间选线,有两次length恢复后Messages初始块格式失败;用户续接后GPT连续三次length耗尽2次恢复。最后诊断output_truncated。需要核对原SSE确切结构、单文件输出长度与池路由,不能笼统归网络或认为1.25.4恢复已解决单大文件问题。只读原记录已存.qa/user-project-errors-20261106,尚未发送模型/操作Key/改用户工程。 AB另冻DeepSeek低强度后真实工程7请求65440tokens失败,末次SSE的DONE与length同时存在;一个完整propose_file后第二个未完成,整条未执行。不能归“始终只读”,也无证据说2048全推理。最后请求三源码完整/测试2583非末换行字符/当前检查job均在,八次read无同SHA同范围重复;原九失败/build通过、0交付,Purge前置未满足。AA+AB14次66404报告tokens+1unknown原预留2235;恢复/全原行/导出/工件/尝试核对通过。现83/99/6875/14761,4016工件,原正式1.25.3/PID36552/Provider/凭据保持。原网络间歇/预留边界/普遍重复读取与Purge收益未知继续保留。 diff --git a/progress.md b/progress.md index 21000ae2..3186525b 100644 --- a/progress.md +++ b/progress.md @@ -2,6 +2,10 @@ ## 用户项目“引航演示”错误排查 +- 源码最终网页/Electron各3组6图、旧预算两视口2组均通过并回看。策略控件沿用原表单网格,模型模式隐藏未启用旧2048,手动原值保持;各自有HTTP32768输出和实际回执700缓存tokens/共同前缀B分开核对。阶段diff回顾后本地提交,再独立发行1.25.5;正式仍1.25.4/原PID36552,没有真实模型请求。 + +- 容量/完整历史/前缀与长文件已实施;102/102相关回归,最终边界14项13通过+池测试接口名修正后1/1;任务否定语义10/10,完整构建通过。自动池现各发成员额度,原始输出/用量/缓存缺失保持可核对。真实模型请求0,源码尚未部署。实际网页/Electron复核中发现并列“不要修改或检查”误判已修;alert角色、夹具能力和响应式等待等验收器首失败均保留。误运行不存在build.mjs、tsc -b生成的自有tsbuildinfo及文档403亦如实记录;生成文件已移除,不作构建成功依据。 + - 切片一隔离相关62项首轮61通过,新增测试调用参数错误及过窄断流类别断言已修,最终精确补验3/3通过,原首失败日志均保留。回顾见docs/QA-PROJECT-GENERATION-REPAIR.md;检查diff后阶段本地提交,正式仍1.25.4,没有真实模型调用。 - 原9请求wire/响应SHA已核对:75,666报告tokens、5次length、0有效提案/文件/检查。合法Messages无起始signature与plan_task缺少只读enum已用真实HTTP回归复现,首轮2/2失败保存在parser-schema-before.txt;修复不放宽最终签名/流终结/工具完整性。用户明确反对正常使用2048限制,当前追踪全局、连接、目录和池的叠加裁剪,旧实验/旧Run不改。首次规划补丁因progress标题不匹配被拒,前两份记录已更新,按实际标题补齐本记录。 diff --git a/scripts/verify-model-capacity-ui.mjs b/scripts/verify-model-capacity-ui.mjs new file mode 100644 index 00000000..31cb2cb8 --- /dev/null +++ b/scripts/verify-model-capacity-ui.mjs @@ -0,0 +1,104 @@ +import fs from 'node:fs'; +import path from 'node:path'; +import net from 'node:net'; +import assert from 'node:assert/strict'; +import { randomUUID } from 'node:crypto'; +import { chromium, _electron } from 'playwright'; +import { expect } from 'playwright/test'; +import { fixture, compatibleFixture, completion, call, launchWorkbench, until } from '../tests/helpers.mjs'; +import { Provider } from '../src/server/provider.mjs'; +import { HostController } from '../src/core/desktop/host-controller.mjs'; +import { desktopExecutableIdentity } from '../src/core/desktop/app-installation.mjs'; +import { projectEnvironment } from '../src/core/process-environment.mjs'; + +if (process.argv.includes('--help')) { console.log('node scripts/verify-model-capacity-ui.mjs [--desktop] — isolated actual capacity/history save, upstream adoption, stale-edit recovery and reported-cache UI. Optional AGENT_WORKBENCH_QA_APP_ROOT selects the immutable app. No real-model keys.'); process.exit(0); } +assert.ok(process.argv.slice(2).every(value => value === '--desktop')); +const desktop = process.argv.includes('--desktop'), appRoot = path.resolve(process.env.AGENT_WORKBENCH_QA_APP_ROOT || '.'); +const directory = path.resolve('.qa/user-project-errors-20261106', (desktop ? 'capacity-electron-' : 'capacity-web-') + randomUUID()); +fs.mkdirSync(directory, { recursive: true }); console.log('EVIDENCE ' + directory); +const disposers = [], t = { after: fn => disposers.push(fn) }, f = fixture(t), checks = [], screenshots = [], errors = []; +let host, controller, browser, application, page, window, result; +const dialog = () => page.getByRole('dialog', { name: '模型中心', exact: true }); +const mark = name => { checks.push(name); console.log('PASS ' + name); }; +async function open() { if (!await dialog().count()) await page.getByRole('button', { name: '打开模型中心', exact: true }).click(); } +async function resize(width) { + if (window) { await window.evaluate((win, size) => win.setContentSize(size.width, size.height), { width, height: 1080 }); await until(() => page.evaluate(w => innerWidth === w, width)); } + else await page.setViewportSize({ width, height: 1080 }); + const compactControl = page.locator('.task-toolbar').getByRole('button', { name: '查看文件', exact: true }); + if (width <= 940) await expect(compactControl).toBeVisible(); else await expect(compactControl).toHaveCount(0); +} +async function shot(name) { + const file = path.join(directory, name + '.png'); await page.screenshot({ path: file, fullPage: true }); screenshots.push(file); + fs.writeFileSync(path.join(directory, name + '.txt'), await page.locator('body').innerText()); +} +try { + const model = await compatibleFixture(t, (url, body) => url.endsWith('/models') ? { data: [{ id: 'owned-capacity', context_length: 131072, max_output_tokens: 32768, capabilities: { text: true, tools: true } }] } + : { ...completion(body.messages.some(row => row.role === 'tool') ? [] : [call('owned-list', 'list_files', { path: '' })], 'Owned reference answer.'), + usage: { prompt_tokens: 1000, completion_tokens: 20, prompt_tokens_details: { cached_tokens: 700 } } }); + const hub = new Provider().modelHub(f.store), source = hub.upsert({ name: 'Owned source', model: 'source', baseUrl: model.baseUrl, requiresKey: false }); + const preview = await hub.modelCatalog().fetch({ id: source.id, revision: source.revision }); + const applied = hub.modelCatalog().apply({ previewId: preview.id, models: ['owned-capacity'], routing: 'manual', defaultModel: 'owned-capacity' }); + const connection = hub.upsert({ ...hub.connection(applied.added[0]), contextLimit: 16000, maxOutputTokens: 2048 }); + assert.equal(connection.upstreamModel.maxOutputTokens, 32768); f.store.close(); + if (desktop) { + const socket = net.createServer(); await new Promise(resolve => socket.listen(0, '127.0.0.1', resolve)); const port = socket.address().port; await new Promise(resolve => socket.close(resolve)); assert.notEqual(port, 4318); + controller = new HostController({ appRoot, dataRoot: f.data, port, nodePath: process.execPath, timeoutMs: 60000 }); await controller.start(); + host = { api: (route, input) => controller.api(route, input ? { method: 'POST', input } : {}), bootstrap: await controller.api('/bootstrap') }; + const executable = desktopExecutableIdentity(appRoot).path; + application = await _electron.launch({ executablePath: executable, args: [...(executable.toLowerCase().endsWith('agentworkbench.exe') ? [] : [path.join(appRoot, 'dist-desktop/main.cjs')]), + '--app-root', appRoot, '--data-root', f.data, '--node', process.execPath, '--port', String(port)], env: projectEnvironment({ AGENT_WORKBENCH_QA: '1', AGENT_WORKBENCH_NODE: process.execPath }) }); + page = await application.firstWindow(); window = await application.browserWindow(page); await page.waitForLoadState(); + } else { + host = await launchWorkbench(t, f.data, {}, { appRoot }); assert.notEqual(new URL(host.url).port, '4318'); + browser = await chromium.launch({ channel: 'msedge', headless: true }); page = await browser.newPage(); await page.goto(host.url); + } + page.on('pageerror', error => errors.push(error.message)); await page.waitForLoadState('networkidle'); await resize(1440); await open(); + await dialog().getByRole('button', { name: '运行预算', exact: true }).click(); + const budgets = () => dialog().getByRole('form', { name: '运行预算设置', exact: true }); + await expect(budgets().getByLabel('容量策略', { exact: true })).toHaveValue('manual'); + await budgets().getByLabel('容量策略', { exact: true }).selectOption('model'); await budgets().getByLabel('历史传递', { exact: true }).selectOption('preserve'); + await expect(budgets().getByLabel('输出上限', { exact: true })).toHaveCount(0); + await budgets().getByLabel('容量策略', { exact: true }).selectOption('manual'); await expect(budgets().getByLabel('输出上限', { exact: true })).toHaveValue('2048'); + await budgets().getByLabel('容量策略', { exact: true }).selectOption('model'); + await budgets().getByRole('button', { name: '保存运行预算', exact: true }).click(); await expect(dialog().getByRole('status')).toContainText('运行预算已保存'); + await page.reload(); await open(); await dialog().getByRole('button', { name: '运行预算', exact: true }).click(); + await expect(budgets().getByLabel('容量策略', { exact: true })).toHaveValue('model'); await expect(budgets().getByLabel('历史传递', { exact: true })).toHaveValue('preserve'); + await shot('capacity-history-1440'); mark('capacity and complete-history choices save and survive reload without model or credential changes'); + await dialog().getByRole('button', { name: '模型连接', exact: true }).click(); + const card = () => dialog().locator('.mc-connection').filter({ has: page.getByRole('heading', { name: connection.name, exact: true }) }); + await card().getByRole('button', { name: '编辑', exact: true }).click(); + const editor = () => dialog().getByRole('form', { name: '编辑模型连接', exact: true }); + await expect(editor().getByLabel('输出上限', { exact: true })).toHaveValue('2048'); + await editor().getByRole('button', { name: '采用上游容量', exact: true }).click(); await expect(editor().getByLabel('输出上限', { exact: true })).toHaveValue('32768'); + const center = await host.api('/model-center'), current = center.connections.find(row => row.id === connection.id); + await host.api('/model-center/connections', { ...current, maxOutputTokens: 8192 }); + await editor().getByRole('button', { name: '保存模型连接', exact: true }).click(); await expect(dialog().getByRole('status')).toContainText('刷新'); + await expect(editor().getByLabel('输出上限', { exact: true })).toHaveValue('32768'); await shot('stale-edit-retained'); + assert.equal((await host.api('/model-center')).connections.find(row => row.id === connection.id).maxOutputTokens, 8192); + await editor().getByRole('button', { name: '取消', exact: true }).click(); await page.reload(); await open(); + await card().getByRole('button', { name: '编辑', exact: true }).click(); await editor().getByRole('button', { name: '采用上游容量', exact: true }).click(); + await editor().getByRole('button', { name: '保存模型连接', exact: true }).click(); await expect(editor()).toHaveCount(0); + assert.equal((await host.api('/model-center')).connections.find(row => row.id === connection.id).maxOutputTokens, 32768); + await page.reload(); await open(); await card().getByRole('button', { name: '编辑', exact: true }).click(); await expect(editor().getByLabel('输出上限', { exact: true })).toHaveValue('32768'); + for (const width of [1440, 390]) { await resize(width); await editor().getByLabel('输出上限', { exact: true }).scrollIntoViewIfNeeded(); assert.ok(await page.evaluate(() => document.documentElement.scrollWidth <= innerWidth + 1)); await shot('upstream-capacity-' + width); } + mark('old 2048 stays visible until upstream adoption; 32768 saves and persists; stale revision rejects while retaining the edit; both viewport widths fit'); + await dialog().getByRole('button', { name: '关闭窗口', exact: true }).click(); + await host.api('/sessions/' + f.session.id + '/messages', { content: 'Inspect the directory once and answer only. Do not modify files or run checks.' }); + let state; await until(async () => { state = await host.api('/sessions/' + f.session.id); return !['running', 'awaiting_approval', 'stopping'].includes(state.status); }, 20000); + assert.equal(state.status, 'completed'); assert.equal(model.requests.filter(row => row.input).length, 2); + assert.ok(model.requests.filter(row => row.input).every(row => row.input.max_tokens === 32768)); + await page.reload(); + const project = page.getByLabel('当前项目', { exact: true }); if (!await project.isVisible()) await page.getByRole('button', { name: '展开项目栏', exact: true }).click(); await project.selectOption(f.project.id); + await page.getByRole('button', { name: '查看任务记录', exact: true }).click(); + const context = page.getByLabel('请求上下文', { exact: true }); await expect(context).toContainText('700 tokens'); await expect(context).toContainText('共同前缀'); await expect(context).toContainText('不等于上游命中'); + for (const width of [1440, 390]) { await resize(width); if (!await context.isVisible()) await page.getByRole('button', { name: '查看任务记录', exact: true }).click(); await expect(context).toBeVisible(); await context.scrollIntoViewIfNeeded(); assert.ok(await page.evaluate(() => document.documentElement.scrollWidth <= innerWidth + 1)); await shot('reported-cache-' + width); } + assert.deepEqual(errors, []); fs.writeFileSync(path.join(directory, 'session.json'), JSON.stringify(state, null, 2)); + mark('actual owned HTTP sends 32768; reported 700 cached tokens and prefix bytes remain separate and persist in the request record'); + result = { success: true, version: host.bootstrap.version, appRoot, actualElectron: desktop, checks, screenshots, errors, realModelRequests: 0, ownedModelRequests: 2, directoryGets: 1, officialServiceTouched: false }; +} catch (error) { if (page) await shot('failure').catch(() => {}); if (host) fs.writeFileSync(path.join(directory, 'failed-session.json'), JSON.stringify(await host.api('/sessions/' + f.session.id).catch(cause => ({ error: cause.message })), null, 2)); result = { success: false, appRoot, actualElectron: desktop, checks, screenshots, errors, error: error.stack, realModelRequests: 0 }; } +finally { + if (application) await application.close(); if (browser) await browser.close(); if (controller && await controller.health()) await controller.stop(); + await host?.close?.(); for (const dispose of disposers.reverse()) await dispose(); + fs.writeFileSync(path.join(directory, 'result.json'), JSON.stringify(result, null, 2)); +} +console.log(JSON.stringify({ directory, ...result })); if (!result.success) process.exitCode = 1; diff --git a/src/client/App.tsx b/src/client/App.tsx index a1c8d282..5289f3e6 100644 --- a/src/client/App.tsx +++ b/src/client/App.tsx @@ -28,7 +28,7 @@ import { BudgetPausePanel } from './BudgetPausePanel'; const activeStates = ['running', 'awaiting_approval', 'stopping']; const activeMcpStates = ['queued', 'connecting', 'discovering', 'calling', 'closing', 'stopping']; -const defaultConfig: Config = { mode: 'demo', baseUrl: 'http://127.0.0.1:11434/v1', model: '', contextLimit: 16000, maxSteps: 12, maxOutputTokens: 2048, hasKey: false, maxRunDurationMs: 600000, maxRunTokens: 200000, maxTaskTokens: 1000000, modelResponseTimeoutMs: 300000, modelStreamIdleTimeoutMs: 120000 }; +const defaultConfig: Config = { mode: 'demo', baseUrl: 'http://127.0.0.1:11434/v1', model: '', contextLimit: 128000, maxSteps: 12, maxOutputTokens: 16384, capacityMode: 'model', hasKey: false, maxRunDurationMs: 600000, maxRunTokens: 200000, maxTaskTokens: 1000000, modelResponseTimeoutMs: 300000, modelStreamIdleTimeoutMs: 120000 }; export default function App() { const [token, setAccess] = useState(''); @@ -646,7 +646,7 @@ export default function App() { {desktopOpen && desktop && setDesktopOpen(false)}> { if (selectedProjectRef.current === scope.projectId && selectedSessionRef.current === scope.sessionId && selectedTaskRef.current === scope.taskId) { setDesktopOpen(false); showInspector('task'); } }} />} {settingsOpen && setSettingsOpen(false)}>
{ event.preventDefault(); saveSettings(); }}>
{settings.mode === 'compatible' && <>
Chat Completions 兼容接口
只在本次服务内存中使用,不写入会话与配置。{config.hasKey && }
填写服务提供的接口根地址。
{modelCheck &&
{modelCheck.message}
}} -
上下文与任务预算

上下文使用本地估算;实际用量与未知用量分开保存。达到预算会暂停,已完成的工具结果保留。

{anyRunning &&

有任务在执行,请先停止任务再改变连接。

}
} +
上下文与任务预算

上下文使用本地估算;实际用量与未知用量分开保存。达到预算会暂停,已完成的工具结果保留。

{anyRunning &&

有任务在执行,请先停止任务再改变连接。

}
} {diagnosticsOpen && setDiagnosticsOpen(false)}> openDiagnostics()} />} {reaOpen && session && selectedProject && setReaOpen(false)}>} {engineOpen && session && selectedProject && setEngineOpen(false)}> { setEngineOpen(false); await navigateWorkspace(projectId, sessionId); }} />} diff --git a/src/client/BudgetPausePanel.tsx b/src/client/BudgetPausePanel.tsx index 6c8a5860..78c3aa76 100644 --- a/src/client/BudgetPausePanel.tsx +++ b/src/client/BudgetPausePanel.tsx @@ -32,6 +32,6 @@ export function BudgetPausePanel({session,config,onSession,onRecovery}:{session: {pause?.tokenReservation&&

{calibrationSources?<>依据同一模型线路的 {calibrationSources} 次已知响应校准,:'按本地估算,'}下一次请求需预留 {pause.tokenReservation.reservedTokens.toLocaleString()} tokens{pause.remainingTokens!==undefined?<>,剩余 {pause.remainingTokens.toLocaleString()}:null}。已在发送前暂停;这项预留用于后续预算,原始用量记录保持。

} {pause?.tokenAccounting&&<>
本次运行上游报告用量
{pause.tokenAccounting.run.reportedTokens.toLocaleString()}
用量未知请求 / 本地预留
{pause.tokenAccounting.run.unknownRequests.toLocaleString()} / {pause.tokenAccounting.run.estimatedUnknownTokens.toLocaleString()}
本次运行预算 / 记账超出
{pause.tokenAccounting.run.maximumTokens>0?pause.tokenAccounting.run.maximumTokens.toLocaleString():'不限'} / {pause.tokenAccounting.run.exceededTokens.toLocaleString()}
任务累计预算 / 记账超出
{pause.tokenAccounting.task.maximumTokens>0?pause.tokenAccounting.task.maximumTokens.toLocaleString():'不限'} / {pause.tokenAccounting.task.exceededTokens.toLocaleString()}

单个在途响应可能超过预估。超额后停止后续请求;用量未知的请求仍按本地预留控制预算,不补造实际计费。

} {controlled?<>

受限子任务在隔离协作中处理。当前协作沿用已封存容量与累计预算;若配置不足,先保留现场,再为新目标选择合适配置。已完成工具和原失败不会自动重放。

{onRecovery&&}:} - {open&&{if(!busy)setOpen(false);}} wide>

Task {session.execution?.task?.id}
原 Run {run?.id}

{([['contextLimit','上下文窗口',1000,200000],['maxOutputTokens','响应预留',256,16000],['maxRunTokens','新Run累计tokens',1000,2000000],['maxTaskTokens','原任务累计tokens',1000,10000000],['maxSteps','新Run工具轮次',1,30],['maxRunDurationMs','新Run时间(分钟)',1,60]] as const).map(([key,label,min,max])=>)}

容量须符合服务实际限制;池按成员最小容量核对。新增预算只用于这次明确续接,不修改全局设置。

{error&&

{error}

}
{plan&&
将创建原Task的新Run

可用模型 {plan.eligibleConnections.length?plan.eligibleConnections.map(row=>row.name+' / '+row.model).join('、'):plan.config.model} · 上下文 {plan.budget.contextLimit.toLocaleString()} · 可用输入 {plan.budget.inputBudget.toLocaleString()} tokens

保留 {plan.completedToolIds.length} 项完成工具、{plan.attachmentIds.length} 张附件;{plan.unknownToolIds.length} 项未知结果仍须核对。

{plan.note}

预览 SHA-256 {plan.sha256}

}
} + {open&&{if(!busy)setOpen(false);}} wide>

Task {session.execution?.task?.id}
原 Run {run?.id}

{([['contextLimit','上下文窗口',1000,Number.MAX_SAFE_INTEGER],['maxOutputTokens','响应预留',256,Number.MAX_SAFE_INTEGER],['maxRunTokens','新Run累计tokens',1000,2000000],['maxTaskTokens','原任务累计tokens',1000,10000000],['maxSteps','新Run工具轮次',1,30],['maxRunDurationMs','新Run时间(分钟)',1,60]] as const).map(([key,label,min,max])=>)}

容量须符合服务实际限制;池按成员最小容量核对。新增预算只用于这次明确续接,不修改全局设置。

{error&&

{error}

}
{plan&&
将创建原Task的新Run

可用模型 {plan.eligibleConnections.length?plan.eligibleConnections.map(row=>row.name+' / '+row.model).join('、'):plan.config.model} · 上下文 {plan.budget.contextLimit.toLocaleString()} · 可用输入 {plan.budget.inputBudget.toLocaleString()} tokens

保留 {plan.completedToolIds.length} 项完成工具、{plan.attachmentIds.length} 张附件;{plan.unknownToolIds.length} 项未知结果仍须核对。

{plan.note}

预览 SHA-256 {plan.sha256}

}
} ; } diff --git a/src/client/ModelBudgetSettings.tsx b/src/client/ModelBudgetSettings.tsx index 250ff266..e321afb7 100644 --- a/src/client/ModelBudgetSettings.tsx +++ b/src/client/ModelBudgetSettings.tsx @@ -2,16 +2,20 @@ import { useState } from 'react'; import type { Config } from './types'; export function ModelBudgetSettings({ value, disabled, onSave }: { value: Config; disabled: boolean; onSave: (input: Partial) => void }) { - const [draft, setDraft] = useState(() => ({ contextLimit: value.contextLimit, maxOutputTokens: value.maxOutputTokens, + const [draft, setDraft] = useState(() => ({ capacityMode: value.capacityMode ?? 'manual', historyPolicy: value.historyPolicy ?? 'compact', contextLimit: value.contextLimit, maxOutputTokens: value.maxOutputTokens, maxSteps: value.maxSteps, maxRunDurationMs: value.maxRunDurationMs, maxRunTokens: value.maxRunTokens, maxTaskTokens: value.maxTaskTokens })); const fields = [ - ['contextLimit', '上下文预算', 1000, 200000], ['maxOutputTokens', '输出上限', 256, 16000], + ['contextLimit', '上下文预算', 1000, Number.MAX_SAFE_INTEGER], ['maxOutputTokens', '输出上限', 256, Number.MAX_SAFE_INTEGER], ['maxSteps', '最大工具轮次', 1, 30], ['maxRunDurationMs', '运行时间(秒)', 1, 3600], ['maxRunTokens', '单次运行 tokens', 1000, 2000000], ['maxTaskTokens', '任务累计 tokens', 1000, 10000000], ] as const; return
{ event.preventDefault(); onSave(draft); }}> -

运行预算

保存仅调整后续运行的预算,保留已选连接或模型池。实际上下文与输出上限还受所选连接限制;旧运行的预算与用量保持原记录。

-
{fields.map(([key, label, min, max]) =>
diff --git a/src/client/RequestContextPanel.tsx b/src/client/RequestContextPanel.tsx index 080c506c..deeab97f 100644 --- a/src/client/RequestContextPanel.tsx +++ b/src/client/RequestContextPanel.tsx @@ -38,12 +38,14 @@ export function RequestContextPanel({ session, runId, openArtifact, onError }: {
请求输入估算
{number(plan.estimatedInputTokens)} tokens
输出 / 安全余量
{number(plan.budget.outputReservation)} / {number(plan.budget.safetyTokens)} tokens
模型上下文窗口
{number(plan.budget.contextLimit)} tokens
已报告输入
{plan.actualUsage?.prompt_tokens === undefined ? '服务未返回' : number(plan.actualUsage.prompt_tokens) + ' tokens'}
历史记录
保留 {plan.retainedMessages} 条 · 保存 {plan.archivedGroups} 组原记录
+
上游缓存命中
{plan.actualUsage?.prompt_tokens_details?.cached_tokens === undefined ? '上游未报告' : number(plan.actualUsage.prompt_tokens_details.cached_tokens) + ' tokens'}
+ {plan.prefixAudit && <>
缓存前缀核对
{plan.prefixAudit.variants.map((variant, index) =>
{variant.model} · {variant.previousPromptBytes === null ? '首次请求或新路线' : '共同前缀 ' + number(variant.commonPrefixBytes) + ' B'}{variant.reason === 'tool_surface_changed' ? ' · 可用工具已变化' : ''}
)}共同前缀用于核对稳定性,不等于上游命中。
} {!!plan.workingSetEntries && <>
当前任务依据
{plan.workingSetEntries} 项代码与执行结果
} {!!plan.projectMemoryEntries && <>
项目记忆
{plan.projectMemoryEntries} 条来源已核对的记录
} {!!plan.skillCatalogEntries && <>
技能目录
{plan.skillCatalogEntries} 项本次目录记录
}
发送内容
{number(plan.requestBytes)} B{plan.requestSha256.slice(0, 16)}
-

估算包含系统提示、工具定义和历史。任务要求完整保留;省略的原记录可从本机工件检索。实际用量以模型服务为准。

+

估算包含系统提示、工具定义和历史。{plan.historyPolicy === 'preserve' ? '本次完整传递历史,容量不足时暂停。' : '任务要求完整保留;整理过的原记录可从本机工件检索。'}实际用量与缓存命中以模型服务为准。

{plan.error &&

{plan.error}

}
diff --git a/src/client/types.ts b/src/client/types.ts index 8a1b4452..505e8969 100644 --- a/src/client/types.ts +++ b/src/client/types.ts @@ -1,6 +1,8 @@ export interface Config { mode: 'demo' | 'compatible'; baseUrl: string; model: string; contextLimit: number; maxSteps: number; maxOutputTokens: number; hasKey: boolean; + capacityMode?: 'model' | 'manual'; + historyPolicy?: 'preserve' | 'compact'; maxRunDurationMs: number; maxRunTokens: number; maxTaskTokens: number; modelResponseTimeoutMs?: number; modelStreamIdleTimeoutMs?: number; } @@ -179,7 +181,9 @@ export interface ContextPlan { latestToolsArtifactId?: string | null; latestToolsSha256?: string | null; latestToolsEntries?: number; projectMemoryArtifactId?: string | null; projectMemorySha256?: string | null; projectMemoryEntries?: number; skillCatalogArtifactId?: string | null; skillCatalogSha256?: string | null; skillCatalogEntries?: number; - actualUsage: { prompt_tokens?: number; completion_tokens?: number; total_tokens?: number } | null; requestAttempts: number; error?: string; + actualUsage: { prompt_tokens?: number; completion_tokens?: number; total_tokens?: number; prompt_tokens_details?: { cached_tokens?: number; cache_creation_tokens?: number } } | null; requestAttempts: number; error?: string; + historyPolicy?: 'preserve' | 'compact'; contextLayout?: string; + prefixAudit?: { variants: { connectionId: string | null; model: string; commonPrefixBytes: number; previousPromptBytes: number | null; reason: string; toolNames: string[]; toolsSha256: string }[] }; budget: { contextLimit: number; outputReservation: number; safetyTokens: number; inputBudget: number; targetInputTokens: number; runRemainingTokens: number | null; taskRemainingTokens: number | null }; } export interface ContextHistoryGroup { diff --git a/src/core/context/budget-resume.mjs b/src/core/context/budget-resume.mjs index 57e8379a..88b11844 100644 --- a/src/core/context/budget-resume.mjs +++ b/src/core/context/budget-resume.mjs @@ -2,10 +2,11 @@ import { randomUUID,createHash } from 'node:crypto'; import { budgetFor } from './request-plan.mjs'; import { accountedTokens,emptyUsage } from '../../shared/contracts/execution.mjs'; import { capabilityRejections } from '../../shared/contracts/model-capabilities.mjs'; +import { CAPACITY_RANGES } from '../../shared/contracts/model-capacity.mjs'; const fail=(message,status=409)=>Object.assign(new Error(message),{status}); const hash=value=>createHash('sha256').update(JSON.stringify(value)).digest('hex'); -const ranges={contextLimit:[1000,200000],maxOutputTokens:[256,16000],maxSteps:[1,30],maxRunTokens:[1000,2000000],maxTaskTokens:[1000,10000000],maxRunDurationMs:[1000,3600000]}; +const ranges={...CAPACITY_RANGES,maxSteps:[1,30],maxRunTokens:[1000,2000000],maxTaskTokens:[1000,10000000],maxRunDurationMs:[1000,3600000]}; export class BudgetResume { constructor(store,provider,runtime){this.store=store;this.provider=provider;this.runtime=runtime;} @@ -18,7 +19,7 @@ export class BudgetResume { prepare(session,input){ const {task,run}=this.current(session); if(!input||Object.keys(input).some(key=>![...Object.keys(ranges),'selection'].includes(key)))throw fail('续接预览字段无效。',400); - const config={...this.store.config,maxTaskTokens:task.budget.maxTokens}; + const config={...this.store.config,capacityMode:'manual',maxTaskTokens:task.budget.maxTokens}; for(const [key,[minimum,maximum]] of Object.entries(ranges))if(input[key]!==undefined){ if(!Number.isSafeInteger(input[key])||input[key]maximum)throw fail('续接预算 '+key+' 超出允许范围。',400); config[key]=input[key]; diff --git a/src/core/context/prefix-audit.mjs b/src/core/context/prefix-audit.mjs new file mode 100644 index 00000000..3fb5a0ae --- /dev/null +++ b/src/core/context/prefix-audit.mjs @@ -0,0 +1,23 @@ +import { createHash } from 'node:crypto'; + +const hash = value => createHash('sha256').update(value).digest('hex'); +const variants = body => body.kind === 'model-route-plan' ? body.variants : [{ connectionId: null, protocol: body.input ? 'responses' : Array.isArray(body.system) ? 'anthropic-messages' : 'chat-completions', body }]; +const prompt = body => Buffer.from(JSON.stringify({ model: body.model, tools: body.tools ?? [], system: body.system ?? null, messages: body.messages ?? body.input ?? [] })); + +// This measures ordered prompt fields, not a supplier cache hit or saved +// tokens. Metadata stays outside the request and cannot invalidate its prefix. +export function prefixAudit(body, previous = null) { + const old = previous ? variants(previous) : []; + return { schemaVersion: 1, method: 'ordered-wire-prompt-fields-utf8', upstreamCache: 'unknown', variants: variants(body).map(current => { + const tools = JSON.stringify(current.body.tools ?? []), bytes = prompt(current.body); + const prior = old.find(value => value.connectionId === current.connectionId && value.protocol === current.protocol && value.body.model === current.body.model); + const previousBytes = prior ? prompt(prior.body) : null; + let common = 0; + if (previousBytes) while (common < Math.min(previousBytes.length, bytes.length) && previousBytes[common] === bytes[common]) common++; + const toolsChanged = !!prior && JSON.stringify(prior.body.tools ?? []) !== tools; + return { connectionId: current.connectionId, protocol: current.protocol, model: current.body.model, + toolsSha256: hash(tools), toolBytes: Buffer.byteLength(tools), toolNames: (current.body.tools ?? []).map(tool => tool.function?.name ?? tool.name), + promptBytes: bytes.length, previousPromptBytes: previousBytes?.length ?? null, commonPrefixBytes: common, + reason: !prior ? 'new_route_or_first_request' : toolsChanged ? 'tool_surface_changed' : bytes.equals(previousBytes) ? 'same_prompt' : 'prompt_append_or_change' }; + }) }; +} diff --git a/src/core/context/request-plan.mjs b/src/core/context/request-plan.mjs index 233b0aab..c2e40d89 100644 --- a/src/core/context/request-plan.mjs +++ b/src/core/context/request-plan.mjs @@ -8,6 +8,7 @@ import { skillCatalogMessage, SKILL_LIMITS } from '../skills/skill-registry.mjs' import { mcpCatalogMessage, MCP_LIMITS } from '../tools/mcp-contract.mjs'; import { modelReservation,taskConsumption,verifyModelReservation } from '../../shared/contracts/model-reservation.mjs'; import { fileContinuation } from './read-continuation.mjs'; +import { prefixAudit } from './prefix-audit.mjs'; const hash = value => createHash('sha256').update(typeof value === 'string' ? value : JSON.stringify(value), 'utf8').digest('hex'); const PLACEHOLDER_ID = '00000000-0000-0000-0000-000000000000'; @@ -61,7 +62,7 @@ export function budgetFor(config, run, task, transportReservationTokens = 0) { if (taskRemainingTokens !== null && taskRemainingTokens < limit) { limit = taskRemainingTokens; reason = 'task_token_limit'; } if(!Number.isSafeInteger(transportReservationTokens)||transportReservationTokens<0)throw new Error('请求封装预留无效。'); const inputBudget = limit - outputReservation - safetyTokens - transportReservationTokens; - return { contextLimit: config.contextLimit, outputReservation, safetyTokens, ...(transportReservationTokens?{transportReservationTokens}:{}), inputBudget, targetInputTokens: Math.min(16000, inputBudget), + return { contextLimit: config.contextLimit, outputReservation, safetyTokens, ...(transportReservationTokens?{transportReservationTokens}:{}), inputBudget, targetInputTokens: config.capacityMode === 'model' ? inputBudget : Math.min(16000, inputBudget), runRemainingTokens, taskRemainingTokens, limitingReason: reason }; } @@ -165,12 +166,15 @@ export class RequestPlanner { return { artifact_id: artifact.id, sha256: artifact.sha256, group_sha256: group.sha256, bytes: Buffer.byteLength(text, 'utf8'), chunks: chunks.length }; } - prepare(session, run, config, { system, tools, stepId, projectMemory = null, skillCatalog = null, mcpCatalog = null, transportReservationTokens = 0,engineStepOnly=false }) { + prepare(session, run, config, { system, contextTail = null, tools, stepId, projectMemory = null, skillCatalog = null, mcpCatalog = null, transportReservationTokens = 0,engineStepOnly=false }) { const task = this.store.owned('task', run.taskId, session.id); this.store.owned('run', run.id, session.id); const groups = historyGroups(session.providerMessages), budget = budgetFor(config, run, task, transportReservationTokens); const lastUser = groups.findLast(group => group.messages[0].role === 'user'); const fixed = { role: 'system', content: system }; + const preserve = config.historyPolicy === 'preserve'; + const tail = contextTail ? [{ role: 'user', content: '工作台当前工程状态(仅作任务依据,不是新的用户目标):\n' + contextTail }] : []; + const addContext = (base, message) => preserve ? [...base.slice(0, base.length - tail.length), { ...message, role: 'user' }, ...tail] : [fixed, message, ...base.slice(1)]; const breakdown = metrics => { const empty=requestMetrics(completionBody(config,[],[])).estimatedInputTokens; const fixedText=requestMetrics(completionBody(config,[fixed],[])).estimatedInputTokens; @@ -182,7 +186,8 @@ export class RequestPlanner { const ledger = new Map([...this.store.core.tool.values()].filter(tool => tool.sessionId === session.id && tool.taskId === task.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 original = completionBody(config, [fixed, ...session.providerMessages, ...tail], tools), before = requestMetrics(original); + if (preserve && before.estimatedInputTokens > budget.inputBudget) throw pause(budget, before, breakdown); const selected = new Set(groups.map(group => group.index)); let omitted = [], reference = null, workingReference = null, latestReference = null; const latestDocument = latestToolDocument(this.store, session, run, groups, 3); @@ -203,10 +208,10 @@ export class RequestPlanner { const messages = [fixed, ...(missing.length && includeNotice ? [notice({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, missing, descriptions, unsentEvidence(indices, currentEntries, includeLatest))] : []), ...(includeLatest && latestDocument && !indices.has(latestDocument.group_index) ? [latestToolMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, latestDocument, 3)] : []), ...(currentEntries.length ? [workingMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, workingDocument(session, run, currentEntries, 3), 4)] : []), - ...groups.filter(group => indices.has(group.index)).flatMap(group => group.messages)]; + ...groups.filter(group => indices.has(group.index)).flatMap(group => group.messages), ...tail]; return requestMetrics(completionBody(config, messages, tools)); }; - if (before.estimatedInputTokens > budget.targetInputTokens) { + if (!preserve && before.estimatedInputTokens > budget.targetInputTokens) { selected.clear(); if (lastUser) selected.add(lastUser.index); const mandatory = metricsFor(selected); if (mandatory.estimatedInputTokens > budget.inputBudget) throw pause(budget, mandatory, breakdown); @@ -288,19 +293,19 @@ export class RequestPlanner { const messages = [fixed, ...(reference ? [notice(reference, omitted, descriptions, unsentEvidence(selected, entries, true))] : []), ...(latestReference ? [latestToolMessage(latestReference, latestDocument, 3)] : []), ...(workingReference ? [workingMessage(workingReference, workingDocument(session, run, entries, 3), 4)] : []), - ...structuredClone(groups.filter(group => selected.has(group.index)).flatMap(group => group.messages))]; + ...structuredClone(groups.filter(group => selected.has(group.index)).flatMap(group => group.messages)), ...tail]; let mcpReference = null, mcpDocument = null; if (mcpCatalog?.projectId === session.projectId && mcpCatalog.entries?.length) { mcpDocument = { ...mcpCatalog, sessionId: session.id, taskId: run.taskId, runId: run.id, entries: [] }; for (const entry of mcpCatalog.entries.slice(0, MCP_LIMITS.allowedTools)) { mcpDocument.entries.push(entry); - const candidate = [fixed, mcpCatalogMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, mcpDocument), ...messages.slice(1)]; + const candidate = addContext(messages, mcpCatalogMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, mcpDocument)); if (Buffer.byteLength(JSON.stringify(mcpDocument), 'utf8') > 24000 || requestMetrics(completionBody(config, candidate, tools)).estimatedInputTokens > Math.min(budget.inputBudget, budget.targetInputTokens)) mcpDocument.entries.pop(); } if (mcpDocument.entries.length) { mcpReference = this.artifact(session, run, mcpDocument, 'context-mcp-catalog'); - messages.splice(1, 0, mcpCatalogMessage(mcpReference, mcpDocument)); + messages.splice(0, messages.length, ...addContext(messages, mcpCatalogMessage(mcpReference, mcpDocument))); } } let skillReference = null, skillDocument = null; @@ -308,13 +313,13 @@ export class RequestPlanner { skillDocument = { ...skillCatalog, sessionId: session.id, taskId: run.taskId, runId: run.id, entries: [] }; for (const entry of skillCatalog.entries.slice(0, SKILL_LIMITS.enabled)) { skillDocument.entries.push(entry); - const candidate = [fixed, skillCatalogMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, skillDocument), ...messages.slice(1)]; + const candidate = addContext(messages, skillCatalogMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, skillDocument)); if (Buffer.byteLength(JSON.stringify(skillDocument), 'utf8') > SKILL_LIMITS.catalogBytes || requestMetrics(completionBody(config, candidate, tools)).estimatedInputTokens > Math.min(budget.inputBudget, budget.targetInputTokens)) skillDocument.entries.pop(); } if (skillDocument.entries.length) { skillReference = this.artifact(session, run, skillDocument, 'context-skill-catalog'); - messages.splice(1, 0, skillCatalogMessage(skillReference, skillDocument)); + messages.splice(0, messages.length, ...addContext(messages, skillCatalogMessage(skillReference, skillDocument))); } } // Optional catalogs and memory use only space left after goals, paired history and current code evidence. @@ -325,13 +330,13 @@ export class RequestPlanner { for (const entry of projectMemory.entries.slice(0, MEMORY_CONTEXT_LIMITS.entries)) { if (entry.sourceFingerprint !== projectMemory.sourceFingerprint) continue; memoryDocument.entries.push(entry); - const candidate = [fixed, projectMemoryMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, memoryDocument), ...messages.slice(1)]; + const candidate = addContext(messages, projectMemoryMessage({ id: PLACEHOLDER_ID, sha256: PLACEHOLDER_SHA }, memoryDocument)); if (Buffer.byteLength(JSON.stringify(memoryDocument), 'utf8') > MEMORY_CONTEXT_LIMITS.bytes || requestMetrics(completionBody(config, candidate, tools)).estimatedInputTokens > memoryBudget) memoryDocument.entries.pop(); } if (memoryDocument.entries.length) { memoryReference = this.artifact(session, run, memoryDocument, 'context-project-memory'); - messages.splice(1, 0, projectMemoryMessage(memoryReference, memoryDocument)); + messages.splice(0, messages.length, ...addContext(messages, projectMemoryMessage(memoryReference, memoryDocument))); } } const body = completionBody(config, messages, tools), metrics = requestMetrics(body); @@ -341,11 +346,18 @@ export class RequestPlanner { if(tokenReservation&&tokenReservation.reservedTokens>credit)throw pause({...budget,limitingReason:runCredit<=taskCredit?'token_limit':'task_token_limit', tokenReservation,remainingTokens:credit,pauseKind:'request_reservation'},metrics,breakdown); const requestArtifact = this.artifact(session, run, body, 'model-request'); + const tailArtifact = tail.length ? this.artifact(session, run, tail[0], 'context-state-tail') : null; + const previousPlan = [...this.store.core.context_plan.values()].filter(plan => plan.sessionId === session.id && plan.taskId === task.id && plan.requestAttempts > 0) + .sort((a, b) => a.createdAt.localeCompare(b.createdAt) || a.requestIndex - b.requestIndex).at(-1) ?? null; + const previousBody = previousPlan ? JSON.parse(this.store.artifactStore.read(this.store.owned('artifact', previousPlan.requestArtifactId, session.id))) : null; + const prefix = { ...prefixAudit(body, previousBody), previousPlanId: previousPlan?.id ?? null, previousRequestSha256: previousPlan?.requestSha256 ?? null }; const record = { schemaVersion: 1, id: randomUUID(), sessionId: session.id, projectId: session.projectId, taskId: run.taskId, runId: run.id, stepId, requestIndex: [...this.store.core.context_plan.values()].filter(plan => plan.runId === run.id).length + 1, status: 'prepared', createdAt: new Date().toISOString(), finishedAt: null, requestSha256: metrics.sha256, requestBytes: metrics.bytes, estimatedInputTokens: metrics.estimatedInputTokens, requestArtifactId: requestArtifact.id, + prefixAudit: prefix, + ...(preserve ? { contextLayout: 'stable-prefix-1', historyPolicy: 'preserve', contextTailArtifactId: tailArtifact?.id ?? null, contextTailSha256: tailArtifact?.sha256 ?? null } : {}), ...(tokenReservation?{tokenReservation,maxBudgetAttempts:Math.min(8,Math.floor(credit/tokenReservation.reservedTokens))}:{engineStepOnly:true}), estimateMethod: metrics.estimateMethod, beforeInputTokens: before.estimatedInputTokens, beforeBytes: before.bytes, budget, ...(metrics.imageCount?{imageCount:metrics.imageCount,imageReservationTokens:metrics.imageReservationTokens,imageEstimate:metrics.imageEstimate, @@ -358,7 +370,7 @@ export class RequestPlanner { - metricsFor(new Set(lastUser ? [lastUser.index] : []), true, []).estimatedInputTokens), workingRetentionPolicyVersion: 1, readBudgetPolicyVersion: 3, - originalMessages: session.providerMessages.length, retainedMessages: messages.length - 1 - (reference ? 1 : 0) - (workingReference ? 1 : 0) - (latestReference ? 1 : 0) - (memoryReference ? 1 : 0) - (skillReference ? 1 : 0) - (mcpReference ? 1 : 0), + originalMessages: session.providerMessages.length, retainedMessages: messages.length - 1 - tail.length - (reference ? 1 : 0) - (workingReference ? 1 : 0) - (latestReference ? 1 : 0) - (memoryReference ? 1 : 0) - (skillReference ? 1 : 0) - (mcpReference ? 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, workingSetArtifactId: workingReference?.id ?? null, workingSetSha256: workingReference?.sha256 ?? null, workingSetEntries: entries.length, @@ -404,6 +416,23 @@ export function verifyPlanEvidence(store, session, plan, cache = new Map()) { historyGroups(requestHistory(body)); for (const variant of requestVariants(body)) historyGroups(requestHistory(variant)); const frozen = JSON.parse(store.artifactStore.read(get(plan.artifactId, 'context-plan'))); + for (const key of ['contextLayout', 'historyPolicy', 'contextTailArtifactId', 'contextTailSha256', 'prefixAudit']) { + if (JSON.stringify(frozen[key] ?? null) !== JSON.stringify(plan[key] ?? null)) throw new Error('前缀或历史策略与发送前封存不匹配。'); + } + if (plan.prefixAudit) { + const previous = plan.prefixAudit.previousPlanId ? store.owned('context_plan', plan.prefixAudit.previousPlanId, session.id) : null; + if (previous && (previous.taskId !== plan.taskId || previous.id === plan.id || previous.createdAt > plan.createdAt || previous.requestSha256 !== plan.prefixAudit.previousRequestSha256)) throw new Error('前缀对照来源不匹配。'); + const previousBody = previous ? JSON.parse(store.artifactStore.read(get(previous.requestArtifactId, 'model-request', previous.requestSha256))) : null; + const expected = { ...prefixAudit(body, previousBody), previousPlanId: previous?.id ?? null, previousRequestSha256: previous?.requestSha256 ?? null }; + if (JSON.stringify(expected) !== JSON.stringify(plan.prefixAudit)) throw new Error('前缀观测与原请求字节不匹配。'); + } + const carriesContext = content => requestHistory(body).filter(value => typeof value.content === 'string' + && (value.role === 'system' && value.content === content || plan.contextLayout === 'stable-prefix-1' && value.role === 'user' && value.content.includes(content))).length === 1; + if (plan.contextTailArtifactId) { + const artifact = get(plan.contextTailArtifactId, 'context-state-tail', plan.contextTailSha256); + const tail = JSON.parse(store.artifactStore.read(artifact)); + if (artifact.taskId !== plan.taskId || artifact.runId !== plan.runId || tail.role !== 'user' || !carriesContext(tail.content)) throw new Error('动态状态与实际发送内容不匹配。'); + } for(const key of ['tokenReservation','maxBudgetAttempts'])if(JSON.stringify(frozen[key]??null)!==JSON.stringify(plan[key]??null))throw new Error('模型用量预留与发送前封存不一致。'); for(const key of ['imageCount','imageReservationTokens','imageEstimate','inputImages']) { if(JSON.stringify(frozen[key]??null)!==JSON.stringify(plan[key]??null))throw new Error('图片身份或预算与发送前封存不匹配。'); @@ -451,7 +480,7 @@ export function verifyPlanEvidence(store, session, plan, cache = new Map()) { || document.entries.length > MEMORY_CONTEXT_LIMITS.entries || Buffer.byteLength(JSON.stringify(document), 'utf8') > MEMORY_CONTEXT_LIMITS.bytes || document.entries.some(entry => entry.sourceFingerprint !== document.sourceFingerprint)) throw new Error('项目记忆工件归属或范围不匹配。'); const message = projectMemoryMessage(artifact, document); - if (requestHistory(body).filter(value => value.role === 'system' && value.content === message.content).length !== 1) throw new Error('项目记忆与实际发送内容不匹配。'); + if (!carriesContext(message.content)) throw new Error('项目记忆与实际发送内容不匹配。'); } if ((frozen.skillCatalogArtifactId ?? null) !== (plan.skillCatalogArtifactId ?? null) || (frozen.skillCatalogSha256 ?? null) !== (plan.skillCatalogSha256 ?? null) @@ -465,7 +494,7 @@ export function verifyPlanEvidence(store, session, plan, cache = new Map()) { || document.entries.length > SKILL_LIMITS.enabled || Buffer.byteLength(JSON.stringify(document), 'utf8') > SKILL_LIMITS.catalogBytes || new Set(document.entries.map(entry => entry.id)).size !== document.entries.length) throw new Error('技能目录归属或范围不匹配。'); const message = skillCatalogMessage(artifact, document); - if (requestHistory(body).filter(value => value.role === 'system' && value.content === message.content).length !== 1) throw new Error('技能目录与实际发送内容不匹配。'); + if (!carriesContext(message.content)) throw new Error('技能目录与实际发送内容不匹配。'); } if ((frozen.mcpCatalogArtifactId ?? null) !== (plan.mcpCatalogArtifactId ?? null) || (frozen.mcpCatalogSha256 ?? null) !== (plan.mcpCatalogSha256 ?? null) @@ -478,7 +507,7 @@ export function verifyPlanEvidence(store, session, plan, cache = new Map()) { || document.kind !== 'registered-mcp-catalog' || !Array.isArray(document.entries) || document.entries.length !== plan.mcpCatalogEntries || document.entries.length > MCP_LIMITS.allowedTools || Buffer.byteLength(JSON.stringify(document), 'utf8') > 24000) throw new Error('MCP目录归属或范围不匹配。'); const message = mcpCatalogMessage(artifact, document); - if (requestHistory(body).filter(value => value.role === 'system' && value.content === message.content).length !== 1) throw new Error('MCP目录与实际发送内容不匹配。'); + if (!carriesContext(message.content)) throw new Error('MCP目录与实际发送内容不匹配。'); } if (plan.latestToolsArtifactId) { const artifact = get(plan.latestToolsArtifactId, 'context-latest-tools', plan.latestToolsSha256); diff --git a/src/core/git-delivery.mjs b/src/core/git-delivery.mjs index 4e968f54..213e3cb5 100644 --- a/src/core/git-delivery.mjs +++ b/src/core/git-delivery.mjs @@ -204,6 +204,7 @@ export class GitDelivery { } async plan(session, options = {}, { source = 'user' } = {}) { + if (session.fileDrafts?.some(row => row.taskId === session.taskId && row.status === 'staging')) throw conflict('分段文件草稿尚未完成,请完成提案或显式放弃后再交付。'); assertWorkersResolved(this.store, session.projectId); if (workerForProject(this.store, session.projectId)) throw conflict('受限子任务不能独立Git提交;请在父任务审查合并后交付。'); const idleOptions = { exceptSessionId: source === 'agent' ? session.id : null }; diff --git a/src/core/models/config-import.mjs b/src/core/models/config-import.mjs index f7960f27..4b46bad0 100644 --- a/src/core/models/config-import.mjs +++ b/src/core/models/config-import.mjs @@ -35,6 +35,8 @@ export class ConfigImporter { ...(item.outputTokenField !== undefined ? { outputTokenField: item.outputTokenField } : {}), ...(item.parameterProfile !== undefined ? { parameterProfile: item.parameterProfile } : {}), ...(item.protocol === 'anthropic-messages' ? { messagesAuth: item.messagesAuth ?? 'api-key' } : {}), + ...(item.contextLimit !== undefined ? { contextLimit: item.contextLimit } : {}), + ...(item.maxOutputTokens !== undefined ? { maxOutputTokens: item.maxOutputTokens } : {}), baseUrl: item.baseUrl, apiKey: item.apiKey, requiresKey: item.requiresKey !== false, brand: item.brand }); } } else if (source === 'ccswitch') { diff --git a/src/core/models/model-catalog.mjs b/src/core/models/model-catalog.mjs index 119d2f72..d2826d5a 100644 --- a/src/core/models/model-catalog.mjs +++ b/src/core/models/model-catalog.mjs @@ -2,6 +2,7 @@ import { randomUUID, createHash } from 'node:crypto'; import { catalogPage, catalogRoute } from '../../shared/contracts/model-catalog.mjs'; import { messagesAuthSettings, validateProtocolEndpoint } from '../../shared/contracts/model-endpoint.mjs'; import { modelProtocol, capabilityDeclarations } from '../../shared/contracts/model-capabilities.mjs'; +import { DEFAULT_MODEL_CAPACITY } from '../../shared/contracts/model-capacity.mjs'; const fail = (message, status = 400) => Object.assign(new Error(message), { status }); const stamp = () => new Date().toISOString(); @@ -128,8 +129,8 @@ export class ModelCatalog { protocol: value.connection.protocol, baseUrl: value.connection.baseUrl, brand: value.connection.brand, messagesAuth: value.connection.messagesAuth, requiresKey: value.connection.requiresKey, concurrency: value.connection.concurrency, parameterProfile: value.connection.parameterProfile ?? 'auto', capabilityDeclarations: Object.fromEntries(Object.entries(model.capabilityDeclarations).map(([key, state]) => [key, state === 'unknown' ? declarations[key] : state])), streaming: false, - contextLimit: Math.min(value.connection.contextLimit ?? 16000, model.contextLimit ?? 200000), - maxOutputTokens: Math.min(value.connection.maxOutputTokens ?? 2048, model.maxOutputTokens ?? 16000), + contextLimit: model.contextLimit ?? value.connection.contextLimit ?? DEFAULT_MODEL_CAPACITY.contextLimit, + maxOutputTokens: model.maxOutputTokens ?? value.connection.maxOutputTokens ?? DEFAULT_MODEL_CAPACITY.maxOutputTokens, ...(!owner && value.explicitKey ? { apiKey: value.explicitKey, persistKey: input.persistKey === true } : {}) }); const connection = hub.connection(profile.id); connection.upstreamModel = upstreamModel; connection.source = 'upstream-catalog'; if (!owner) owner = connection; diff --git a/src/core/models/model-hub.mjs b/src/core/models/model-hub.mjs index 010db4c1..63d6d590 100644 --- a/src/core/models/model-hub.mjs +++ b/src/core/models/model-hub.mjs @@ -13,6 +13,7 @@ import { assertModelParameterSupport } from '../../shared/contracts/model-parame import { modelIdentitySettings } from '../../shared/contracts/model-identity.mjs'; import { NetworkDiagnostics } from './network-diagnostics.mjs'; import { ModelCatalog, credentialScope } from './model-catalog.mjs'; +import { capacitySettings, routedCapacity } from '../../shared/contracts/model-capacity.mjs'; const legacyId = 'legacy-default'; const stamp = () => new Date().toISOString(); @@ -217,7 +218,7 @@ export class ModelHub { capabilitySchema: 1, capabilityDeclarations: declarations, streaming, ...(native ?? {}), ...modelIdentitySettings(input, previous ?? {}), ...messagesAuthSettings(protocol, input.messagesAuth ?? previous?.messagesAuth), baseUrl: validateProtocolEndpoint(protocol, modelEndpoint(input.baseUrl)), enabled: input.enabled !== false, requiresKey: input.requiresKey !== false, - contextLimit: bounded(input.contextLimit, 16000, 1000, 200000), maxOutputTokens: bounded(input.maxOutputTokens, 2048, 256, 16000), + ...capacitySettings(input, previous ?? undefined), concurrency: bounded(input.concurrency, 1, 1, 8), revision: (previous?.revision ?? 0) + 1, source: previous?.source ?? 'manual', ...(previous?.importOrigin ? { importOrigin: previous.importOrigin, importRevision: previous.importRevision } : {}), createdAt: previous?.createdAt ?? stamp(), updatedAt: stamp() }; @@ -291,7 +292,7 @@ export class ModelHub { return { ...config, ...messagesAuthSettings(members[0].protocol, members[0].messagesAuth), baseUrl: members[0].baseUrl, model: members[0].model, protocol: modelProtocol(members[0].protocol), streaming: members[0].streaming === true, nativeParameters: members[0].nativeParameters ?? {}, outputTokenField: members[0].outputTokenField, parameterProfile: members[0].parameterProfile, modelConnectionId: members[0].id, modelCredentialRevision: members[0].credentialRevision, - contextLimit: Math.min(config.contextLimit, ...members.map(value => value.contextLimit)), maxOutputTokens: Math.min(config.maxOutputTokens, ...members.map(value => value.maxOutputTokens)), + ...routedCapacity(config, members), modelRouting: { type: selection.type, id: selection.id, revision: pool?.revision ?? members[0].revision, strategy: pool?.strategy ?? 'round-robin', sticky: pool?.sticky ?? true, maxAttempts: pool?.maxAttempts ?? 1, queueTimeoutMs: pool?.queueTimeoutMs ?? 15000, members } }; } diff --git a/src/core/models/protocols.mjs b/src/core/models/protocols.mjs index fcd73dd6..86e4b865 100644 --- a/src/core/models/protocols.mjs +++ b/src/core/models/protocols.mjs @@ -6,7 +6,7 @@ export const protocolFailure = (message, kind = 'protocol', extra = null) => new // even a syntactically complete leading tool call is dropped, because any later call in // the same turn may be cut mid-argument. The diagnostic must tell the model what happened // and how to proceed, instead of a generic incompleteness message it cannot act on. -export const OUTPUT_TRUNCATED_HINT = '模型输出达到单次 output token 上限(finish_reason=length),本次响应整体未执行任何工具(即使其中一个工具参数已完整也不执行,避免应用被截断的半成品)。请减少单次响应的工具数量:把多文件提案拆成每次一个文件的 propose_file/propose_patch,或分多轮 prepare_change_set,不要在一个响应里堆叠多个大参数工具。'; +export const OUTPUT_TRUNCATED_HINT = '上游模型在本次配置的输出额度处停止(finish_reason=length),本次响应整体未执行工具,原响应和用量已保留。请拆分输出:已有文件使用propose_patch;单个长文件用propose_file_chunk分多轮追加较短的完整片段,最后complete=true生成完整提案。已有成功片段使用原draft_id/next_offset/draft_sha256继续,不从头重复。不要在一个响应中堆叠多个大工具参数。原目标和已读依据保留。'; export const outputTruncatedFailure = (kind = 'protocol') => protocolFailure(OUTPUT_TRUNCATED_HINT, kind, { reason: 'output_truncated' }); export { protocolRoute } from '../../shared/contracts/model-endpoint.mjs'; @@ -20,14 +20,22 @@ export function reportedUsage(config, value) { const input = raw.input_tokens + cached + created, total = input + raw.output_tokens; if (!Number.isSafeInteger(total)) return null; return { prompt_tokens: input, completion_tokens: raw.output_tokens, total_tokens: total, - prompt_tokens_details: { cached_tokens: cached, cache_creation_tokens: created }, raw: structuredClone(raw), protocol: 'anthropic-messages' }; + ...(raw.cache_read_input_tokens !== undefined || raw.cache_creation_input_tokens !== undefined ? { prompt_tokens_details: { + ...(raw.cache_read_input_tokens !== undefined ? { cached_tokens: cached } : {}), + ...(raw.cache_creation_input_tokens !== undefined ? { cache_creation_tokens: created } : {}) } } : {}), + raw: structuredClone(raw), protocol: 'anthropic-messages' }; } const input = config.protocol === 'responses' ? raw.input_tokens : raw.prompt_tokens; const output = config.protocol === 'responses' ? raw.output_tokens : raw.completion_tokens; const valid = number => Number.isSafeInteger(number) && number >= 0; if (!valid(raw.total_tokens) && !(valid(input) && valid(output))) return null; + const cached = config.protocol === 'responses' ? raw.input_tokens_details?.cached_tokens : raw.prompt_tokens_details?.cached_tokens ?? raw.prompt_cache_hit_tokens; + const cacheDetails = valid(cached) && valid(input) && cached <= input ? { cached_tokens: cached } : null; + const reasoning = config.protocol === 'responses' ? raw.output_tokens_details?.reasoning_tokens : raw.completion_tokens_details?.reasoning_tokens; return { ...(valid(input) ? { prompt_tokens: input } : {}), ...(valid(output) ? { completion_tokens: output } : {}), - total_tokens: valid(raw.total_tokens) ? raw.total_tokens : input + output }; + total_tokens: valid(raw.total_tokens) ? raw.total_tokens : input + output, + ...(cacheDetails ? { prompt_tokens_details: cacheDetails } : {}), ...(valid(reasoning) && valid(output) && reasoning <= output ? { completion_tokens_details: { reasoning_tokens: reasoning } } : {}), + raw: structuredClone(raw), protocol: config.protocol ?? 'chat-completions' }; } function acceptsEmptyObject(tools, name) { @@ -84,9 +92,10 @@ export function normalizeProtocolResponse(config, value) { } if (config.protocol !== 'responses') { const received = value?.choices?.[0]?.message; - if (received?.reasoning_content === undefined || received.reasoning_content === null) return value; + if (received?.reasoning_content === undefined || received.reasoning_content === null) return { ...value, usage: reportedUsage(config, value) }; if (typeof received.reasoning_content !== 'string') throw protocolFailure('Chat reasoning_content必须为字符串,原响应未用于执行工具。'); const normalized = structuredClone(value), target = normalized.choices[0].message; + normalized.usage = reportedUsage(config, value); delete target.reasoning_content; target.providerState = { protocol: 'chat-completions', endpoint: config.baseUrl, model: config.model, connectionId: config.modelConnectionId ?? null, credentialRevision: config.modelCredentialRevision ?? null, reasoningContent: received.reasoning_content }; @@ -109,8 +118,7 @@ export function normalizeProtocolResponse(config, value) { } } else if (item.type !== 'reasoning') throw protocolFailure('Responses包含当前尚未接入的服务端工具或输出类型:' + String(item.type ?? '未标注').slice(0, 100) + '。'); } - const raw = value.usage, usage = raw ? { prompt_tokens: raw.input_tokens, completion_tokens: raw.output_tokens, - ...(raw.total_tokens !== undefined ? { total_tokens: raw.total_tokens } : {}), raw, protocol: 'responses' } : null; + const usage = reportedUsage(config, value); return { choices: [{ message: { role: 'assistant', content: texts.join(''), ...(calls.length ? { tool_calls: calls } : {}), providerState: { protocol: 'responses', endpoint: config.baseUrl, model: config.model, responseId: value.id ?? null, items: value.output } }, finish_reason: calls.length ? 'tool_calls' : 'stop' }], usage }; } diff --git a/src/core/models/runtime-settings.mjs b/src/core/models/runtime-settings.mjs index 6c1a0f41..9275ddb3 100644 --- a/src/core/models/runtime-settings.mjs +++ b/src/core/models/runtime-settings.mjs @@ -1,15 +1,19 @@ +import { CAPACITY_RANGES, capacityMode } from '../../shared/contracts/model-capacity.mjs'; + const ranges = { - contextLimit: [1000, 200000], maxOutputTokens: [256, 16000], maxSteps: [1, 30], + ...CAPACITY_RANGES, maxSteps: [1, 30], maxRunDurationMs: [1000, 3600000], maxRunTokens: [1000, 2000000], maxTaskTokens: [1000, 10000000], }; // Budget-only edits cannot change the selected route, endpoint or credential. export function validateRuntimeBudgets(input) { if (!input || typeof input !== 'object' || Array.isArray(input) || !Object.keys(input).length - || Object.keys(input).some(key => !Object.hasOwn(ranges, key))) throw new Error('只接受运行预算字段;模型连接请在模型选择中切换。'); + || Object.keys(input).some(key => !['capacityMode', 'historyPolicy'].includes(key) && !Object.hasOwn(ranges, key))) throw new Error('只接受运行预算字段;模型连接请在模型选择中切换。'); for (const [key, value] of Object.entries(input)) { + if (key === 'capacityMode') { capacityMode(value); continue; } + if (key === 'historyPolicy') { if (!['preserve', 'compact'].includes(value)) throw new Error('请选择完整传递或允许整理历史。'); continue; } const [min, max] = ranges[key]; if (!Number.isSafeInteger(value) || value < min || value > max) throw new Error(key + ' 需填写范围内的整数。'); } - return { ...input }; + return { ...input, ...(!Object.hasOwn(input, 'capacityMode') && ['contextLimit', 'maxOutputTokens'].some(key => Object.hasOwn(input, key)) ? { capacityMode: 'manual' } : {}) }; } diff --git a/src/core/storage/legacy-import.mjs b/src/core/storage/legacy-import.mjs index e7d965a1..168281bc 100644 --- a/src/core/storage/legacy-import.mjs +++ b/src/core/storage/legacy-import.mjs @@ -2,9 +2,10 @@ import fs from 'node:fs'; import path from 'node:path'; import { createHash, randomUUID } from 'node:crypto'; import { isId } from '../../shared/contracts/execution.mjs'; +import { DEFAULT_MODEL_CAPACITY } from '../../shared/contracts/model-capacity.mjs'; export const DEFAULT_CONFIG = { - mode: 'demo', baseUrl: 'http://127.0.0.1:11434/v1', model: '', contextLimit: 16000, maxSteps: 12, maxOutputTokens: 2048, + mode: 'demo', baseUrl: 'http://127.0.0.1:11434/v1', model: '', ...DEFAULT_MODEL_CAPACITY, capacityMode: 'model', historyPolicy: 'preserve', maxSteps: 12, maxRunDurationMs: 600000, maxRunTokens: 200000, maxTaskTokens: 1000000, configRevision: 1, modelResponseTimeoutMs: 300000, modelStreamIdleTimeoutMs: 120000, }; @@ -12,6 +13,9 @@ export const DEFAULT_CONFIG = { export function cleanConfig(value = {}) { const result = { ...DEFAULT_CONFIG }; for (const key of Object.keys(result)) if (Object.hasOwn(value, key)) result[key] = value[key]; + // Old settings stay manual. Every saved Run keeps its original snapshot. + if (!Object.hasOwn(value, 'capacityMode') && (Object.hasOwn(value, 'contextLimit') || Object.hasOwn(value, 'maxOutputTokens'))) result.capacityMode = 'manual'; + if (!Object.hasOwn(value, 'historyPolicy') && (Object.hasOwn(value, 'contextLimit') || Object.hasOwn(value, 'maxOutputTokens'))) result.historyPolicy = 'compact'; for (const key of ['nativeParameters', 'outputTokenField', 'parameterProfile']) if (Object.hasOwn(value, key)) result[key] = structuredClone(value[key]); return result; } diff --git a/src/core/workspaces/file-proposal-drafts.mjs b/src/core/workspaces/file-proposal-drafts.mjs new file mode 100644 index 00000000..4c136f1c --- /dev/null +++ b/src/core/workspaces/file-proposal-drafts.mjs @@ -0,0 +1,70 @@ +import fs from 'node:fs'; +import { randomUUID } from 'node:crypto'; +import { safePath, readText, hash, proposeFile } from '../../server/files.mjs'; +import { assertWorkersResolved, assertWorkerWriteScope } from '../../shared/contracts/workers.mjs'; +import { assertMcpResolved } from '../tools/mcp-contract.mjs'; +import { assertChangeSetsResolved } from '../../shared/contracts/change-sets.mjs'; +import { assertWorkspacesResolved } from '../../shared/contracts/workspaces.mjs'; + +const fail = message => { throw Object.assign(new Error(message), { status: 409 }); }; +const emptySha = hash(Buffer.alloc(0)); +export const pendingFileDrafts = session => (session.fileDrafts ?? []).filter(row => row.taskId === session.taskId && row.status === 'staging'); + +// Fragments are immutable artifacts. A partial draft is never a file proposal, +// never a change-set member, and never written into the user's project. +export function proposeFileChunk(store, session, args) { + if (Object.keys(args).some(key => !['path', 'content', 'reason', 'offset', 'complete', 'draft_id', 'expected_sha256', 'base_sha256'].includes(key)) + || typeof args.path !== 'string' || typeof args.content !== 'string' || typeof args.reason !== 'string' || !args.reason.trim() + || !Number.isSafeInteger(args.offset) || args.offset < 0 || typeof args.complete !== 'boolean' + || args.content.includes('\0') || !args.content.isWellFormed() || !args.complete && !args.content.length) fail('分段提案需要有效路径、文本、说明、offset和complete;未完成片段不能为空。'); + const run = store.owned('run', session.runId, session.id); + if (run.taskId !== session.taskId) fail('分段提案缺少当前Task/Run归属。'); + assertWorkersResolved(store, session.projectId); assertWorkerWriteScope(store, session, args.path); + assertMcpResolved(store, session.projectId); assertChangeSetsResolved(store, session.projectId); + assertWorkspacesResolved(store, session.projectId, { includePrepared: true }); + const project = store.project(session.projectId), target = safePath(project, args.path, true); + const base = fs.existsSync(target) ? readText(project, args.path).sha256 : null; + let draft = args.draft_id ? (session.fileDrafts ?? []).find(row => row.id === args.draft_id) : null; + if (args.draft_id && (!draft || draft.status !== 'staging' || draft.taskId !== session.taskId || draft.projectId !== session.projectId || draft.path !== args.path)) fail('草稿不属于当前Task/路径或已结束,请核对原回执。'); + if (!draft) { + if (args.offset !== 0 || args.expected_sha256 !== undefined && args.expected_sha256 !== emptySha) fail('新草稿必须从offset 0开始。'); + if (pendingFileDrafts(session).some(row => row.path === args.path) || pendingFileDrafts(session).length >= 16) fail('该路径已有未完成草稿或当前Task已达16份草稿,请继续原草稿或显式放弃。'); + if (Object.hasOwn(args, 'base_sha256') && args.base_sha256 !== base) fail('原文件SHA已变化,请重新核对。'); + draft = { id: randomUUID(), projectId: session.projectId, taskId: session.taskId, path: args.path, status: 'staging', baseSha256: base, + sha256: emptySha, characters: 0, bytes: 0, chunks: [], createdAt: new Date().toISOString() }; + } else if (args.offset !== draft.characters || args.expected_sha256 !== draft.sha256) fail('片段offset或SHA与原草稿不一致,未追加;请使用上一次成功回执,不能从头重复。'); + if (draft.baseSha256 !== base || Object.hasOwn(args, 'base_sha256') && args.base_sha256 !== draft.baseSha256) fail('分段期间原文件已变化,草稿不能覆盖新内容;请放弃后按新SHA准备。'); + const previous = draft.chunks.map(part => { + const artifact = store.owned('artifact', part.id, session.id); + if (artifact.kind !== 'file-proposal-fragment' || artifact.taskId !== session.taskId || artifact.projectId !== session.projectId || artifact.sha256 !== part.sha256) fail('原片段身份不匹配。'); + return store.artifactStore.read(artifact); + }).join(''); + if (hash(Buffer.from(previous)) !== draft.sha256 || previous.length !== draft.characters || Buffer.byteLength(previous) !== draft.bytes) fail('原草稿校验失败,未采用任何片段。'); + const content = previous + args.content, bytes = Buffer.byteLength(content); + if (bytes > 512000 || draft.chunks.length >= 256) fail('完整文本超过512KB或256段;请按工程模块拆分文件。'); + const artifact = store.artifactStore.write(session, run, args.content, { kind: 'file-proposal-fragment', mediaType: 'text/plain' }); + store.core.artifact.set(artifact.id, artifact); + const next = { ...draft, sha256: hash(Buffer.from(content)), characters: content.length, bytes, chunks: [...draft.chunks, { id: artifact.id, sha256: artifact.sha256 }] }; + const original = session.fileDrafts ?? []; + const originalChanges = [...session.changes], originalStatus = session.status; + session.fileDrafts = [...original.filter(row => row.id !== draft.id), next]; + let result; + try { + if (args.complete) { + next.status = 'finalized'; + result = proposeFile(store, session, { path: args.path, reason: args.reason, content, base_sha256: next.baseSha256 }, { persist: false }); + next.changeId = result.change_id ?? null; + } + store.save(session); + } catch (cause) { session.fileDrafts = original; session.changes = originalChanges; session.status = originalStatus; throw cause; } + return { ...(result ?? { status: 'staged', message: '片段已保存,尚无可应用提案。继续使用draft_id、next_offset和draft_sha256;最后一段complete=true才创建完整审查提案。' }), + draft_id: next.id, path: next.path, next_offset: next.characters, draft_sha256: next.sha256, bytes: next.bytes, complete: args.complete }; +} + +export function discardFileDraft(store, session, args) { + if (Object.keys(args).some(key => !['draft_id', 'expected_sha256', 'reason'].includes(key)) || typeof args.reason !== 'string' || !args.reason.trim()) fail('放弃草稿需要原身份、SHA与说明。'); + const draft = pendingFileDrafts(session).find(row => row.id === args.draft_id); + if (!draft || draft.sha256 !== args.expected_sha256) fail('草稿归属、状态或SHA不匹配。'); + draft.status = 'discarded'; draft.discardReason = args.reason.slice(0, 600); store.save(session); + return { status: 'discarded', draft_id: draft.id, path: draft.path, message: '草稿已放弃,原片段保留,项目文件没有改变。' }; +} diff --git a/src/server/app.mjs b/src/server/app.mjs index 9c44c69e..4359b0fb 100644 --- a/src/server/app.mjs +++ b/src/server/app.mjs @@ -1,4 +1,5 @@ import { validateRuntimeBudgets } from '../core/models/runtime-settings.mjs'; +import { capacitySettings, capacityMode } from '../shared/contracts/model-capacity.mjs'; import fs from 'node:fs'; import { createHash } from 'node:crypto'; import path from 'node:path'; @@ -864,9 +865,8 @@ return async (request, response) => { if (!['demo', 'compatible'].includes(input.mode)) throw new Error('请选择有效运行模式。'); const config = { mode: input.mode, baseUrl: String(input.baseUrl || '').trim(), model: String(input.model || '').trim(), - contextLimit: Math.trunc(Math.max(1000, Math.min(200000, Number(input.contextLimit) || 16000))), + ...capacitySettings(input, store.config), capacityMode: capacityMode(input.capacityMode), maxSteps: Math.trunc(Math.max(1, Math.min(30, Number(input.maxSteps) || 12))), - maxOutputTokens: Math.trunc(Math.max(256, Math.min(16000, Number(input.maxOutputTokens) || 2048))), ...validateModelWaitSettings(input, store.config), maxRunDurationMs: Math.trunc(Math.max(1000, Math.min(3600000, Number(input.maxRunDurationMs) || store.config.maxRunDurationMs))), maxRunTokens: Math.trunc(Math.max(1000, Math.min(2000000, Number(input.maxRunTokens) || store.config.maxRunTokens))), diff --git a/src/server/files.mjs b/src/server/files.mjs index ef8e277f..a48452a2 100644 --- a/src/server/files.mjs +++ b/src/server/files.mjs @@ -129,7 +129,7 @@ export function searchFiles(project, args) { scope: '有数量/体积上限的文本搜索;无结果不表示所有文件都已搜索。' }; } -export function proposeFile(store, session, args, { patch = null, sourceBinding = null } = {}) { +export function proposeFile(store, session, args, { patch = null, sourceBinding = null, persist = true } = {}) { assertWorkersResolved(store, session.projectId); assertWorkerWriteScope(store, session, args.path); assertMcpResolved(store, session.projectId); @@ -159,7 +159,7 @@ export function proposeFile(store, session, args, { patch = null, sourceBinding }; session.changes.push(change); if (!['running', 'awaiting_approval', 'stopping'].includes(session.status)) session.status = 'needs_review'; - store.save(session); + if (persist) store.save(session); return { status: 'awaiting_review', change_id: change.id, path: change.path, additions: change.additions, deletions: change.deletions, message: '提案已保存,项目文件尚未改变。等待用户查看差异并应用。' }; } diff --git a/src/server/provider.mjs b/src/server/provider.mjs index 65c39fb9..952e91d4 100644 --- a/src/server/provider.mjs +++ b/src/server/provider.mjs @@ -12,6 +12,7 @@ import { ModelTransport } from '../core/models/transport.mjs'; import { localModelEstimate } from '../shared/contracts/model-reservation.mjs'; import { ERROR_CAPTURE_LIMIT, readErrorResponse, upstreamCapture } from '../core/models/upstream-error.mjs'; import { messagesAuthSettings } from '../shared/contracts/model-endpoint.mjs'; +import { fileDraftToolDefinitions } from '../shared/contracts/file-draft-tools.mjs'; export const engineToolDefinitions = [ { type: 'function', function: { name: 'inspect_engine_checks', description: '读取当前Task已经保存的外部引擎连接核验:真实安装版本、启动完成信号、SDK会话、进程关闭、原报告工件和能力限制。不会启动引擎或发送prompt;完整原记录通过已有read_artifact读取。', parameters: { type: 'object', properties: {}, additionalProperties: false } } }, @@ -50,6 +51,7 @@ export const toolDefinitions = [ offset: { type: 'integer', minimum: 0 }, limit: { type: 'integer', minimum: 1, maximum: 16000 }, expected_sha256: { type: 'string', pattern: '^[a-fA-F0-9]{64}$' } }, required: ['path'] } } }, { type: 'function', function: { name: 'search_files', description: '在当前项目的文本文件中按字面内容检索。结果声明已搜索文件数与数量上限。', parameters: { type: 'object', properties: { query: { type: 'string' }, case_sensitive: { type: 'boolean' } }, required: ['query'] } } }, + ...fileDraftToolDefinitions, { type: 'function', function: { name: 'propose_file', description: '提出一个文本文件的完整新内容。只产生待审查提案,不写入项目;由用户查看差异后应用。', parameters: { type: 'object', properties: { path: { type: 'string' }, content: { type: 'string' }, reason: { type: 'string' }, base_sha256: { type: ['string', 'null'], description: '如已读取文件,可传入该文件的 SHA256 检查读取后变化;新文件为 null。' } }, required: ['path', 'content', 'reason'] } } }, { type: 'function', function: { name: 'run_command', description: '请求在当前项目目录运行终端命令。命令会显示给用户,用户确认后才运行。命令有时间与输出限制。', @@ -62,8 +64,7 @@ export const researchToolDefinitions = [ { type: 'function', function: { name: 'read_research_evidence', description: '读取当前会话原完整REA Evidence的实际字段,校验工件、Bundle及Task来源。pointer使用JSON Pointer,默认/normalized_result;按返回available_fields选择下一级字段,offset/limit和next_offset读取预算分页。静态关系是候选,不能当成运行观测;不会再次调用REA。', parameters: { type: 'object', properties: { research_job_id: { type: 'string' }, pointer: { type: 'string', maxLength: 400 }, offset: { type: 'integer', minimum: 0 }, limit: { type: 'integer', minimum: 1, maximum: 16000 } }, required: ['research_job_id'] } } }, ]; -export function projectContext(project) { - const inventory = listFiles(project, '', 3, 160); +export function projectContext(project, { inventory = true } = {}) { return [ '你是这个本地编程工作台的编程助手。用用户的语言沟通,以实际读取和工具结果为依据。', '当前项目:' + project.name + '。所有文件路径相对于这个项目。', @@ -72,6 +73,7 @@ export function projectContext(project) { '工程任务先用inspect_project识别语言、脚本、规则与Git基线。plan_checks准备已识别npm或Python检查计划,run_checks请求用户确认后执行。Python支持python:test和python:compile,绑定unittest/pytest、实际解释器和源码;混合项目默认保留npm范围,明确选择Python名称才能执行。只凭回答不能声称检查通过。', '修改较大的工程时,先用index_project/query_code定位符号、导入和文件,再用read_file按实际行读取必要内容。索引是静态证据,部分或无结果不代表完整工程;Python声明候选需读取核对。', '已有文件优先用propose_patch和read_file返回的SHA256提出唯一锚点精确替换,保留未选中内容与格式;新文件可用propose_file。补丁锚点重叠/重复或文件变化时重新读取,不猜测写入位置。', + '单个文件大于本次输出额度时,用propose_file_chunk分多轮追加完整片段,核对draft_id/next_offset/draft_sha256;不要从头重复大文件。最后complete=true才生成可审查完整提案。未完成片段不能应用或交付,放弃用discard_file_draft保留原因。', '已有文本文件满足当前目标时,完整分页读取后用verify_existing_files登记当前SHA和原read_call_ids,说明采用原因;跨Task须显式选择source_result_id。无需重写相同文件,不把旧成果冒充新写入;明确修改要求、测试和页面验收仍需各自真实证据。', '跨文件修改的提案准备完整后,用prepare_change_set将本Task的change_id封存成整批审查。准备不表示应用;用户核对整个批次并确认,结果未知时先用工作台恢复入口核对,不重放文件写入。', '复杂工程任务先了解实际范围,再用plan_task保存有限依赖步骤及真实通过条件;verify_task_step引用原工具调用、已应用批次或当前检查/页面job核验。不能删改原条件掩盖失败;inspect_task_plan核对未完成项,模型文字不代替步骤核验。简单问答无需新建工程计划。', @@ -87,8 +89,7 @@ export function projectContext(project) { '当前终端:' + executionEnvironment().shell + '。' + executionEnvironment().shellSyntax + ' 输入和输出使用 UTF-8。', '任务中断后先核对当前状态,不假定缺少结果的工具已经成功。已有提案可等待用户审核,不要重复提出相同改动。', '完成必要步骤后直接给出结果。不要无限继续调用工具。', - '当前文件清单(有数量上限):', - inventory.files.filter(file => file.kind === 'file').map(file => file.path).join('\n'), + ...(inventory ? ['当前文件清单(有数量上限):', listFiles(project, '', 3, 160).files.filter(file => file.kind === 'file').map(file => file.path).join('\n')] : []), ].join('\n'); } diff --git a/src/server/runtime.mjs b/src/server/runtime.mjs index bccf7514..41cf2a7d 100644 --- a/src/server/runtime.mjs +++ b/src/server/runtime.mjs @@ -9,9 +9,11 @@ import { spawn, execFile } from 'node:child_process'; import path from 'node:path'; import { randomUUID } from 'node:crypto'; import { sanitizeModelText } from '../core/models/upstream-error.mjs'; +import { OUTPUT_TRUNCATED_HINT } from '../core/models/protocols.mjs'; import { StringDecoder } from 'node:string_decoder'; import { setImmediate as yieldExecution } from 'node:timers/promises'; import { safePath, listFiles, readLines, searchFiles, proposeFile, proposePatch } from './files.mjs'; +import { proposeFileChunk, discardFileDraft, pendingFileDrafts } from '../core/workspaces/file-proposal-drafts.mjs'; import { projectContext, toolDefinitions, researchToolDefinitions, engineToolDefinitions, browserToolDefinitions } from './provider.mjs'; import { now, repairToolPairs } from './store.mjs'; import { executionEnvironment, sessionContext, sessionSnapshot } from './diagnostics.mjs'; @@ -82,6 +84,8 @@ labels.plan_task = '保存任务依赖计划'; labels.verify_task_step = '核验任务步骤原证据'; labels.inspect_task_plan = '核对任务计划与未完成项'; labels.verify_existing_files = '核对当前已有成果'; +labels.propose_file_chunk = '组装长文件提案'; +labels.discard_file_draft = '放弃未完成文件草稿'; labels.read_registered_skill = '读取已登记技能修订'; labels.call_registered_mcp = '请求调用已登记MCP工具'; @@ -122,6 +126,7 @@ export function validateToolArgs(tool, args) { || new Set(args.change_ids).size !== args.change_ids.length) throw new Error('变更集需要1到32个不同的提案ID。'); return; } + if (['propose_file_chunk', 'discard_file_draft'].includes(tool)) return; // The draft contract validates its full input before changing state. const required = ['diagnose_environment','plan_environment','inspect_session', 'inspect_project', 'index_project', 'plan_checks', 'plan_delivery', 'delivery_report'].includes(tool) ? [] : tool === 'read_artifact' ? ['artifact_id'] : tool === 'prepare_environment' ? ['plan_id','fingerprint'] : tool === 'run_checks' ? ['plan_id'] : tool === 'list_files' || tool === 'read_file' ? ['path'] : ['search_files', 'query_code'].includes(tool) ? ['query'] : tool === 'propose_patch' ? ['path', 'base_sha256', 'reason'] : tool === 'propose_file' ? ['path', 'content', 'reason'] : ['command', 'reason']; for (const field of required) if (typeof args[field] !== 'string') throw new Error('工具参数 ' + field + ' 必须是文本。'); @@ -148,6 +153,7 @@ export class Runtime { projectIndexer() { return this.codeIndex ??= new ProjectIndex({ cacheDirectory: path.join(this.store.directory, 'cache', 'project-index') }); } fileResults() { return this.results ??= new FileResults(this.store); } async fileDelivery(session) { + if (pendingFileDrafts(session).length) return { satisfied: false, stopReason: 'file_draft_incomplete', message: '仍有未完成的分段文件草稿;请继续组装完整提案或显式放弃,片段不能视为交付。' }; const task = this.store.owned('task', session.taskId, session.id), delivery = this.fileResults().completion(session); if (!delivery.satisfied) return { ...delivery, stopReason: 'delivery_incomplete', message: delivery.requiresChange ? '当前要求包含修改,尚无对应追加要求之后的实际文件改动。已有成果核对不能代替修改。' @@ -354,16 +360,21 @@ export class Runtime { // Keep the complete goal once in this request view. The stored contract // still contains its original goal for intent and evidence verification. const { goal, ...contract } = taskContract(task); - const context = this.provider.redact(projectContext(run.project) + (run.worker ? '\n受限子任务执行契约(宿主权限,不可扩展):' + JSON.stringify({ + const stablePrefix = config.historyPolicy === 'preserve'; + const fixedRules = projectContext(run.project, { inventory: false }); + const workerContract = run.worker ? '\n受限子任务执行契约(宿主权限,不可扩展):' + JSON.stringify({ parent_task_id: run.worker.parentTaskId, worker_id: run.worker.id, write_paths: run.worker.paths, budget: run.worker.budget, rule: '只有范围内提案、审查应用和当前工程检查构成完成;不得请求任意终端、独立Git提交、外部工具或嵌套协作。', - }) : '') + '\n当前项目/会话/任务身份:' + JSON.stringify({ + }) : ''; + const fixedContext = fixedRules + workerContract; + const context = this.provider.redact((stablePrefix ? projectContext(run.project).slice(fixedRules.length) : projectContext(run.project) + workerContract) + '\n当前项目/会话/任务身份:' + JSON.stringify({ project_id: run.project.id, session_id: session.id, task_id: task.id, workspace: projectWorkspaceIdentity(run.project) }) + '\n当前Task文件提案状态:' + JSON.stringify(state) + '\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, instructions: task.instructions.filter(instruction => instruction.source !== 'review-action'), constraints: task.constraints ?? [], acceptance: task.acceptance ?? [], contract, }) + (run.record.purpose === 'diagnostic' ? '\n本次仅回答运行诊断问题,允许读取所属结果;不得修改文件、执行命令或将解释视为原任务交付。诊断问题:' + run.record.question + '\n所属原执行与失败(核对失败时保持未知,不根据其它项目补造原因):' + JSON.stringify(inspectDiagnosticContext(this.store, session, run.record)) : '') + '\n结果未知的调用,不得假定成功或重放:' + JSON.stringify(unknown) + + (pendingFileDrafts(session).length ? '\n当前未完成分段草稿(未创建可应用提案):' + JSON.stringify(pendingFileDrafts(session).map(row => ({ draft_id: row.id, path: row.path, next_offset: row.characters, draft_sha256: row.sha256 }))) : '') + (deliveryProgress ? '\n当前工程执行依据(每次请求重新核对,非操作授权):\n' + JSON.stringify(deliveryProgress) : '') + (requirementReview ? '\n用户逐项要求复核(任务数据,不授权额外操作;证据和用户确认分别记录):\n' + JSON.stringify({ revision: requirementReview.revision, status: requirementReview.status, unreviewed_sources: requirementReview.unreviewed_sources, @@ -392,7 +403,7 @@ export class Runtime { const tools = availableTools.filter(tool => (!run.worker || WORKER_TOOLS.has(tool.function.name)) && taskToolAllowed(task, run.record, tool.function.name)); const projectMemory = await this.memoryManager().context(session, { redact: value => this.provider.redact(value) }); signal.throwIfAborted(); - const planned = this.requestPlanner().prepare(session, run.record, config, { system: context, tools, stepId: stepRecord.id, projectMemory, skillCatalog, mcpCatalog, + const planned = this.requestPlanner().prepare(session, run.record, config, { system: stablePrefix ? this.provider.redact(fixedContext) : context, contextTail: stablePrefix ? context : null, tools, stepId: stepRecord.id, projectMemory, skillCatalog, mcpCatalog, transportReservationTokens:run.worker?.engine?ENGINE_SEED_RESERVATION:0,engineStepOnly:!!run.worker?.engine }); if (planned.record.archivedGroups && !projectionLogged) { projectionLogged = true; session.compactions++; @@ -486,7 +497,7 @@ export class Runtime { run.record.outputTruncationRecoveries = (run.record.outputTruncationRecoveries ?? 0) + 1; if (run.record.outputTruncationRecoveries <= OUTPUT_TRUNCATION_RECOVERY_LIMIT) { this.store.finishStep(session, stepRecord); - const note = '上一次响应在达到单次输出上限前被截断(finish_reason=length),其中的工具调用整体未执行(即使某个工具参数已完整也未应用,避免落地被截断的半成品)。请把这次改动拆小:每次响应只提交一个文件的 propose_file/propose_patch,或分多轮 prepare_change_set,不要在一个响应里堆叠多个大参数工具。原目标和已读依据都保留。'; + const note = OUTPUT_TRUNCATED_HINT; // The truncated assistant turn was never appended, so the conversation // stays valid. Feed the instruction to the model as a standalone turn // in the provider history, and mirror it to the visible chat log. @@ -533,6 +544,10 @@ export class Runtime { this.store.finishStep(session, stepRecord); session.status = session.changes.some(change => ['pending', 'conflict'].includes(change.status)) ? 'needs_review' : 'completed'; if (run.record.purpose === 'diagnostic') { session.status = 'completed'; run.stopReason = 'diagnostic_answered'; } + else if (pendingFileDrafts(session).length) { + session.status = 'paused'; run.stopReason = 'file_draft_incomplete'; + this.store.message(session, 'system', '仍有未完成的分段文件草稿,片段不能作为完整交付;原片段和续接位置已保存。'); + } else if (session.status === 'completed' && taskContract(task).kind === 'file') { const delivery = await this.fileDelivery(session); if (!delivery.satisfied) { @@ -851,13 +866,15 @@ export class Runtime { } else if (tool === 'search_files') result = searchFiles(run.project, args); else if (tool === 'propose_file') result = proposeFile(this.store, session, args); + else if (tool === 'propose_file_chunk') result = proposeFileChunk(this.store, session, args); + else if (tool === 'discard_file_draft') result = discardFileDraft(this.store, session, args); else if (tool === 'propose_patch') result = proposePatch(this.store, session, args); 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); if (proofHint) result.tool_invocation_id = event.invocationId; event.phase = ['denied', 'interrupted', 'unknown', 'unverified', 'stale'].includes(result.status) ? result.status : ['failed', 'error'].includes(result.status) ? 'failed' : 'completed'; if (tool === 'inspect_session') recordSessionInspection(this.store.core.task.get(run.taskId), result); - else if (['propose_file', 'propose_patch', 'verify_task_step'].includes(tool)) { + else if (['propose_file', 'propose_patch', 'propose_file_chunk', 'discard_file_draft', 'verify_task_step'].includes(tool)) { const task=this.store.core.task.get(run.taskId); if(result.status==='unchanged'||result.reused===true){ result.unchanged_action_count=recordUnchangedAction(task,tool,result,run.id); diff --git a/src/shared/contracts/execution.mjs b/src/shared/contracts/execution.mjs index 6c3323d5..195e7ec3 100644 --- a/src/shared/contracts/execution.mjs +++ b/src/shared/contracts/execution.mjs @@ -19,6 +19,8 @@ export function modelSnapshot(config) { return { mode: config.mode, baseUrl: config.baseUrl, model: config.model, contextLimit: config.contextLimit, maxOutputTokens: config.maxOutputTokens, + ...(config.capacityMode ? { capacityMode: config.capacityMode } : {}), + ...(config.historyPolicy ? { historyPolicy: config.historyPolicy } : {}), configRevision: config.configRevision ?? 1, ...modelWaitSettings(config), ...modelIdentitySettings(config), diff --git a/src/shared/contracts/file-draft-tools.mjs b/src/shared/contracts/file-draft-tools.mjs new file mode 100644 index 00000000..3fa43606 --- /dev/null +++ b/src/shared/contracts/file-draft-tools.mjs @@ -0,0 +1,14 @@ +export const fileDraftToolDefinitions = [ + { type: 'function', function: { name: 'propose_file_chunk', description: '长文件不能一次完整输出时,分多轮追加完整文本片段。首次offset=0;后续必须使用回执draft_id/next_offset/draft_sha256作为draft_id/offset/expected_sha256。每段按本次输出额度拆小,最后一段complete=true才创建一个完整待审提案;此前片段不能应用或交付。不修改项目文件,不接收截断JSON。', parameters: { + type: 'object', additionalProperties: false, required: ['path', 'content', 'reason', 'offset', 'complete'], properties: { + path: { type: 'string' }, content: { type: 'string' }, reason: { type: 'string' }, offset: { type: 'integer', minimum: 0 }, complete: { type: 'boolean' }, + draft_id: { type: 'string' }, expected_sha256: { type: 'string', pattern: '^[a-f0-9]{64}$' }, + base_sha256: { type: ['string', 'null'], description: '首次核对原文件SHA;新文件为null。分段过程中基线不变。' }, + }, + } } }, + { type: 'function', function: { name: 'discard_file_draft', description: '明确放弃当前Task的未完成分段文件草稿。需要原draft_id/draft_sha256与原因;原片段和失败保留,不更改项目文件。', parameters: { + type: 'object', additionalProperties: false, required: ['draft_id', 'expected_sha256', 'reason'], properties: { + draft_id: { type: 'string' }, expected_sha256: { type: 'string', pattern: '^[a-f0-9]{64}$' }, reason: { type: 'string' }, + }, + } } }, +]; diff --git a/src/shared/contracts/model-capacity.d.mts b/src/shared/contracts/model-capacity.d.mts new file mode 100644 index 00000000..7e9f00ef --- /dev/null +++ b/src/shared/contracts/model-capacity.d.mts @@ -0,0 +1,4 @@ +export const DEFAULT_MODEL_CAPACITY: Readonly<{ contextLimit: number; maxOutputTokens: number }>; +export const CAPACITY_RANGES: Readonly>; +export function capacitySettings(input?: { contextLimit?: number; maxOutputTokens?: number }, previous?: { contextLimit?: number; maxOutputTokens?: number }): { contextLimit: number; maxOutputTokens: number }; +export function capacityMode(value?: string): 'model' | 'manual'; diff --git a/src/shared/contracts/model-capacity.mjs b/src/shared/contracts/model-capacity.mjs new file mode 100644 index 00000000..4ab6b73a --- /dev/null +++ b/src/shared/contracts/model-capacity.mjs @@ -0,0 +1,24 @@ +// Editable fallbacks for upstreams that do not publish capacity. These values +// are not verified model capabilities or ceilings on larger models. +export const DEFAULT_MODEL_CAPACITY = Object.freeze({ contextLimit: 128000, maxOutputTokens: 16384 }); +export const CAPACITY_RANGES = Object.freeze({ contextLimit: [1000, Number.MAX_SAFE_INTEGER], maxOutputTokens: [256, Number.MAX_SAFE_INTEGER] }); + +export function capacitySettings(input = {}, previous = DEFAULT_MODEL_CAPACITY) { + return Object.fromEntries(Object.entries(CAPACITY_RANGES).map(([key, [min, max]]) => { + const value = input[key] ?? previous[key] ?? DEFAULT_MODEL_CAPACITY[key]; + if (!Number.isSafeInteger(value) || value < min || value > max) throw Object.assign(new Error(key + ' 需要不小于 ' + min + ' 的安全整数,请核对模型实际容量。'), { status: 400 }); + return [key, value]; + })); +} + +export function capacityMode(value = 'manual') { + if (!['model', 'manual'].includes(value)) throw Object.assign(new Error('请选择按模型容量或手动限制。'), { status: 400 }); + return value; +} + +export function routedCapacity(config, members) { + const mode = capacityMode(config.capacityMode); + return Object.fromEntries(Object.keys(DEFAULT_MODEL_CAPACITY).map(key => [key, + mode === 'model' && key === 'maxOutputTokens' ? Math.max(...members.map(member => member[key])) + : Math.min(...members.map(member => member[key]), ...(mode === 'manual' ? [config[key]] : []))])); +} diff --git a/src/shared/contracts/model-request.mjs b/src/shared/contracts/model-request.mjs index 6207ecf1..07557371 100644 --- a/src/shared/contracts/model-request.mjs +++ b/src/shared/contracts/model-request.mjs @@ -34,10 +34,11 @@ function wireBody(config, messages, tools) { export function completionBody(config, messages, tools) { const members = config.modelRouting?.members; - if (members?.some(member => member.protocol && member.protocol !== 'chat-completions' || member.streaming === true || member.nativeParameters || member.outputTokenField)) { + if (members?.length && (config.capacityMode === 'model' || members.some(member => member.protocol && member.protocol !== 'chat-completions' || member.streaming === true || member.nativeParameters || member.outputTokenField))) { return { schema: 1, kind: 'model-route-plan', messages, variants: members.map(member => ({ connectionId: member.id, protocol: modelProtocol(member.protocol), body: wireBody({ ...config, ...member, nativeParameters: member.nativeParameters ?? {}, outputTokenField: member.outputTokenField, - modelConnectionId: member.id, modelCredentialRevision: member.credentialRevision, maxOutputTokens: config.maxOutputTokens }, messages, tools) })) }; + modelConnectionId: member.id, modelCredentialRevision: member.credentialRevision, + maxOutputTokens: config.capacityMode === 'model' ? member.maxOutputTokens : config.maxOutputTokens }, messages, tools) })) }; } return wireBody(config, messages, tools); } diff --git a/src/shared/contracts/task-control.mjs b/src/shared/contracts/task-control.mjs index b65410d2..31f36dcb 100644 --- a/src/shared/contracts/task-control.mjs +++ b/src/shared/contracts/task-control.mjs @@ -1,6 +1,8 @@ // Task intent is retained independently of short follow-ups and system actions. import { changeRequirement, requestedFilePaths, negatedFileAction } from './file-intent.mjs'; export const isContinuation = content => /^(?:请)?(?:继续(?:完成|执行|开发|原任务|这个任务)?|接着(?:完成|执行)?|continue(?:\s+(?:the\s+)?(?:task|work))?)[。.!!\s]*$/i.test(String(content).trim()); +const skipChecks = /无需(?:任何)?(?:测试|工程检查)|不(?:需要|用|要)?(?:进行|做|运行|执行)?(?:任何)?(?:测试|工程检查)|\b(?:no (?:tests?|checks?)(?: needed|required)?|without (?:running )?(?:tests?|checks?)|(?:do not|don't|never) (?:[^.,;!?\n]{1,60}\s+(?:or|nor)\s+)?(?:run|execute) (?:tests?|checks?))\b/i; +const requireChecks = /(?:运行|执行|通过|修复|补充|添加).{0,12}(?:测试|工程检查)|\b(?:run|pass|fix|add)\b.{0,24}\b(?:tests?|checks?)\b/i; export function isDiagnosticQuestion(content) { const text = String(content).trim(); @@ -19,8 +21,8 @@ export function taskContract(task, lookup = () => null) { original = previous; } const goal = original.contract?.goal ?? original.goal; - const skip = /无需(?:任何)?(?:测试|工程检查)|不(?:需要|用|要)?(?:进行|做|运行|执行)?(?:任何)?(?:测试|工程检查)|\b(?:no (?:tests?|checks?)(?: needed|required)?|without (?:running )?(?:tests?|checks?)|(?:do not|don't|never) (?:run|execute) (?:tests?|checks?))\b/i.test(goal); - const required = !skip && /(?:运行|执行|通过|修复|补充|添加).{0,12}(?:测试|工程检查)|\b(?:run|pass|fix|add)\b.{0,24}\b(?:tests?|checks?)\b/i.test(goal); + const skip = skipChecks.test(goal); + const required = !skip && requireChecks.test(goal); const file = hasFileDeliveryRequest(goal); return { schemaVersion: 1, goal, kind: file ? 'file' : 'general', checks: skip ? 'skip' : required ? 'required' : 'unspecified', ...(original.id !== task.id ? { inheritedFromTaskId: original.id } : {}) }; @@ -41,8 +43,8 @@ export function updateTaskContract(task, instruction) { const current = taskContract(task); if (hasFileDeliveryRequest(instruction)) current.kind = 'file'; // Only an explicit new user instruction changes the check policy. - if (/无需(?:任何)?测试|不(?:需要|用|要)(?:进行|做|运行)?测试|\b(?:no tests?|without (?:running )?tests?)\b/i.test(instruction)) current.checks = 'skip'; - else if (/(?:运行|执行|通过|修复|补充|添加).{0,12}(?:测试|工程检查)|\b(?:run|pass|fix|add)\b.{0,24}\b(?:tests?|checks?)\b/i.test(instruction)) current.checks = 'required'; + if (skipChecks.test(instruction)) current.checks = 'skip'; + else if (requireChecks.test(instruction)) current.checks = 'required'; task.contract = current; task.constraints = (task.constraints ?? []).filter(value => value !== '用户明确要求无需测试;不得自动扩大为工程测试任务。'); if (current.checks === 'skip') task.constraints.push('用户明确要求无需测试;不得自动扩大为工程测试任务。'); diff --git a/src/shared/contracts/workers.mjs b/src/shared/contracts/workers.mjs index 996d4ba3..576ad15a 100644 --- a/src/shared/contracts/workers.mjs +++ b/src/shared/contracts/workers.mjs @@ -3,7 +3,7 @@ import { isId, accountedTokens, emptyUsage } from './execution.mjs'; import { selectedEngineInput } from './engine-execution.mjs'; export const WORKER_TOOLS = new Set(['diagnose_environment','plan_environment','prepare_environment','inspect_session', 'inspect_project', 'index_project', 'query_code', 'read_artifact', - 'list_files', 'read_file', 'search_files', 'propose_file', 'propose_patch', 'prepare_change_set', 'plan_checks', 'run_checks']); + 'list_files', 'read_file', 'search_files', 'propose_file', 'propose_patch', 'propose_file_chunk', 'discard_file_draft', 'prepare_change_set', 'plan_checks', 'run_checks']); export const COLLABORATION_ACTIVE = new Set(['preparing', 'starting', 'running', 'stopping', 'collecting', 'merge_preparing', 'reconciling']); export const COLLABORATION_CLOSED = new Set(['completed', 'discarded', 'resolved_external']); export const workerConflict = message => Object.assign(new Error(message), { status: 409 }); @@ -109,7 +109,7 @@ export function assertWorkerTool(store, session, tool, args) { const worker = workerForSession(store, session); if (!worker) return; if (!WORKER_TOOLS.has(tool)) throw workerConflict('受限子任务不开放这个工具:' + tool); - if (tool === 'propose_file' || tool === 'propose_patch') assertWorkerWriteScope(store, session, args.path); + if (tool === 'propose_file' || tool === 'propose_patch' || tool === 'propose_file_chunk') assertWorkerWriteScope(store, session, args.path); } export function assertWorkersResolved(store, projectId, { recoveryTaskId = null } = {}) { const group = [...(store.core.collaboration?.values() ?? [])].find(value => value.requiresVerification diff --git a/task_plan.md b/task_plan.md index b29e4c3f..61528c6b 100644 --- a/task_plan.md +++ b/task_plan.md @@ -4,10 +4,10 @@ 2026-11-06用户要求查看已加入项目的大量错误,解释与Codex/Claude Code/DeepSeek Harness的差距,并处理能确认的工作台问题。接续干净ada7d364,已有他轮1.25.4修复/AC样本;实际服务版本与来源重新核对,不把浏览器旧release查询串当当前版本。不读取/复制/解密/重存Key、不自动续跑用户任务,不改用户工程原源码/测试/暂存或历史错误。 -1. [in_progress] 只读定位项目/全部所属会话、原用户目标/真实工具/模型错误及当前运行身份,保存可核对证据。 -2. [pending] 分类工作台缺陷、模型/供应商和工程环境;用原输入做隔离重现,确定最小修复。 -3. [pending] 修复并验证相关路径,保留原失败与差异;如产品变更按现有独立发行纪律交付。 -4. [pending] 汇总真实原因、已修与仍未知、采用技术与实际接入范围,阶段回顾并本地提交,不推送。 +1. [complete] 原项目/两Run/9请求与原响应SHA、实际1.25.4运行身份已核对,调查原件保留。 +2. [complete] Claude合法thinking起始误拒、plan_task缺少工具枚举已复现、修复并提交2b178211;相关62项61通过+精确补验3/3,测试驱动首失败保留。 +3. [complete] 容量/完整历史/前缀审计/上游缓存计数/长文件分段已实现;102相关及新增边界、保存失败/重启、三协议与池各自输出通过。最终网页/Electron各3组6图、旧预算两视口2组与构建通过;并列否定误识别必查已修并10/10。未知能力、原失败、旧Run与原实验不改。 +4. [pending] 入口验证后审查diff、回顾并本地提交,再冻结1.25.5独立构建/归档/安装和保留原Provider升级;源码完成不冒称已部署,不推送。真实自主交付/网络/重复读取/真实预留/Purge未知保持。 ## 当前:AB 已验证低强度模式的新工程验证 diff --git a/tests/file-proposal-drafts.test.mjs b/tests/file-proposal-drafts.test.mjs new file mode 100644 index 00000000..c5910338 --- /dev/null +++ b/tests/file-proposal-drafts.test.mjs @@ -0,0 +1,96 @@ +import test from 'node:test'; +import assert from 'node:assert/strict'; +import fs from 'node:fs'; +import path from 'node:path'; +import { fixture, compatibleFixture, completion, call, until } from './helpers.mjs'; +import { proposeFileChunk, discardFileDraft, pendingFileDrafts } from '../src/core/workspaces/file-proposal-drafts.mjs'; +import { hash, applyChange } from '../src/server/files.mjs'; +import { Store } from '../src/server/store.mjs'; +import { Provider } from '../src/server/provider.mjs'; +import { Runtime } from '../src/server/runtime.mjs'; + +const start = t => { const f = fixture(t); f.store.beginRun(f.session, 'Create a complete file.', false, f.store.config); return f; }; +const input = (content, extra = {}) => ({ path: 'long.html', content, reason: 'Owned long file', offset: 0, complete: false, base_sha256: null, ...extra }); +const next = (result, content, complete = false) => input(content, { draft_id: result.draft_id, offset: result.next_offset, expected_sha256: result.draft_sha256, complete }); + +test('fragments persist across restart, reject duplicate/out-of-order appends, and create exactly one whole review proposal', t => { + const f = start(t), first = proposeFileChunk(f.store, f.session, input('中文😀\r\n')); + assert.equal(f.session.changes.length, 0); assert.equal(fs.existsSync(path.join(f.projectRoot, 'long.html')), false); + assert.throws(() => applyChange(f.store, f.session, first.draft_id)); + assert.throws(() => proposeFileChunk(f.store, f.session, { ...next(first, 'duplicate'), offset: 0 }), /offset或SHA/); + f.store.close(); + const reopened = new Store(f.data); f.onCleanup(() => reopened.close()); + const session = reopened.sessions.get(f.session.id); + const last = proposeFileChunk(reopened, session, next(first, '\n', true)); + assert.equal(session.changes.length, 1); assert.equal(session.changes[0].content, '中文😀\r\n\n'); + assert.equal(session.changes[0].sha256, hash(Buffer.from(session.changes[0].content))); + assert.equal(last.change_id, session.changes[0].id); assert.equal(pendingFileDrafts(session).length, 0); + assert.throws(() => proposeFileChunk(reopened, session, next(first, '', true)), /已结束/); + assert.equal(session.changes.length, 1); assert.equal(fs.existsSync(path.join(f.projectRoot, 'long.html')), false); +}); + +test('source edits, foreign tasks, malformed input, oversize text and changed fragment identities cannot finalize', t => { + const f = start(t), target = path.join(f.projectRoot, 'README.md'), base = hash(fs.readFileSync(target)); + const first = proposeFileChunk(f.store, f.session, input('New start', { path: 'README.md', base_sha256: base })); + fs.appendFileSync(target, 'External change\n'); + assert.throws(() => proposeFileChunk(f.store, f.session, { ...next(first, 'tail', true), path: 'README.md', base_sha256: base }), /原文件已变化/); + assert.match(fs.readFileSync(target, 'utf8'), /External change/); assert.equal(f.session.changes.length, 0); + const draft = proposeFileChunk(f.store, f.session, input('Original draft', { path: 'new.txt' })); + for (const malformed of [input('x', { path: '../outside.txt' }), input('\ud800'), input('\0'), input('x'.repeat(512001)), input('x', { complete: 'true' })]) assert.throws(() => proposeFileChunk(f.store, f.session, malformed)); + const fragment = f.store.core.artifact.get(pendingFileDrafts(f.session).find(row => row.id === draft.draft_id).chunks[0].id), originalSha = fragment.sha256; + fragment.sha256 = '0'.repeat(64); + assert.throws(() => proposeFileChunk(f.store, f.session, { ...next(draft, 'tail', true), path: 'new.txt' }), /片段身份/); fragment.sha256 = originalSha; + const oldTask = f.session.taskId; f.session.taskId = '00000000-0000-0000-0000-000000000000'; + assert.throws(() => proposeFileChunk(f.store, f.session, { ...next(draft, 'tail', true), path: 'new.txt' })); f.session.taskId = oldTask; + discardFileDraft(f.store, f.session, { draft_id: draft.draft_id, expected_sha256: draft.draft_sha256, reason: 'Replan the file' }); + assert.equal(f.session.fileDrafts.find(row => row.id === draft.draft_id).status, 'discarded'); +}); + +test('a truncated continuation cannot append half JSON; recovery continues the original draft and produces one complete file', async t => { + const f = fixture(t), pieces = ['\n' + '中文😀\n'.repeat(200), '

Body

\n', '\n']; + let receipt, count = 0; + const server = await compatibleFixture(t, (_url, body) => { + count++; + const tools = body.messages.filter(row => row.role === 'tool').map(row => JSON.parse(row.content)); + receipt = tools.findLast(row => row.draft_id) ?? receipt; + if (count === 1) return completion([call('chunk-a', 'propose_file_chunk', input(pieces[0]))]); + if (count === 2) return { ...completion([call('never-append', 'propose_file_chunk', '{"path":"long.html","content":"PARTIAL')]), choices: [{ message: { role: 'assistant', content: '', tool_calls: [call('never-append', 'propose_file_chunk', '{"path":"long.html","content":"PARTIAL')] }, finish_reason: 'length' }] }; + if (count === 3) return completion([call('chunk-final', 'propose_file_chunk', next(receipt, pieces[1] + pieces[2], true))]); + return completion([], 'The complete proposal is ready for review.'); + }); + const provider = new Provider(), hub = provider.modelHub(f.store), c = hub.upsert({ name: 'Owned small output', model: 'owned', baseUrl: server.baseUrl, requiresKey: false, maxOutputTokens: 2048 }); + hub.select({ type: 'connection', id: c.id }); const runtime = new Runtime(f.store, provider); f.onCleanup(() => runtime.stopAll()); + runtime.start(f.session.id, 'Create long.html using reviewable proposals, no tests requested.'); await until(() => !runtime.isActive(f.session.id), 20000); + assert.equal(f.session.status, 'needs_review', JSON.stringify(f.session.events.slice(-6))); assert.equal(count, 4); + assert.equal(f.session.changes.length, 1); assert.equal(f.session.changes[0].content, pieces.join('')); + assert.equal(f.session.fileDrafts[0].chunks.length, 2); assert.equal(f.session.fileDrafts[0].status, 'finalized'); + assert.equal(f.session.events.filter(event => event.toolCallId === 'never-append').length, 0); + assert.ok(f.session.providerMessages.some(row => row.role === 'user' && row.content.includes('propose_file_chunk'))); + assert.equal(fs.existsSync(path.join(f.projectRoot, 'long.html')), false); +}); + +test('an unfinished draft cannot be described as a completed task', async t => { + const f = fixture(t), server = await compatibleFixture(t, (_url, body) => body.messages.some(row => row.role === 'tool') ? completion([], 'Done.') : completion([call('partial', 'propose_file_chunk', input('Only the beginning'))])); + const provider = new Provider(), hub = provider.modelHub(f.store), c = hub.upsert({ name: 'Owned partial', model: 'owned', baseUrl: server.baseUrl, requiresKey: false }); + hub.select({ type: 'connection', id: c.id }); const runtime = new Runtime(f.store, provider); f.onCleanup(() => runtime.stopAll()); + runtime.start(f.session.id, 'Prepare the project.'); await until(() => !runtime.isActive(f.session.id), 15000); + assert.equal(f.session.status, 'paused', JSON.stringify(f.session.events.slice(-6))); assert.equal(f.store.core.run.get(f.session.runId).stopReason, 'file_draft_incomplete'); + assert.equal(f.session.changes.length, 0); assert.equal(pendingFileDrafts(f.session).length, 1); +}); + +test('a failed final save preserves the last successful receipt and cannot publish half a finalized proposal', t => { + const f = start(t), first = proposeFileChunk(f.store, f.session, input('Original fragment\n')); + const save = f.store.storage.save; + f.store.storage.save = () => { throw new Error('Owned storage failure'); }; + assert.throws(() => proposeFileChunk(f.store, f.session, next(first, 'Final fragment\n', true)), /Owned storage failure/); + f.store.storage.save = save; + assert.equal(f.session.changes.length, 0); + assert.equal(pendingFileDrafts(f.session)[0].sha256, first.draft_sha256); + assert.equal(pendingFileDrafts(f.session)[0].chunks.length, 1); + const final = proposeFileChunk(f.store, f.session, next(first, 'Final fragment\n', true)); + assert.equal(f.session.changes.length, 1); assert.equal(final.change_id, f.session.changes[0].id); + f.store.close(); const reopened = new Store(f.data); f.onCleanup(() => reopened.close()); + const session = reopened.session(f.session.id); + assert.equal(session.changes.length, 1); assert.equal(session.changes[0].content, 'Original fragment\nFinal fragment\n'); + assert.equal(pendingFileDrafts(session).length, 0); assert.equal(fs.existsSync(path.join(f.projectRoot, 'long.html')), false); +}); diff --git a/tests/helpers.mjs b/tests/helpers.mjs index bb9bcd24..2b956edd 100644 --- a/tests/helpers.mjs +++ b/tests/helpers.mjs @@ -5,18 +5,25 @@ import http from 'node:http'; import { spawn } from 'node:child_process'; import { fileURLToPath } from 'node:url'; import { Store } from '../src/server/store.mjs'; +import { DEFAULT_CONFIG } from '../src/core/storage/legacy-import.mjs'; export const workspace = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); const fixtureRoot = path.join(workspace, '.qa'); const disposalLists = new WeakMap(); -export function fixture(t) { +export function fixture(t, { modernDefaults = false } = {}) { fs.mkdirSync(fixtureRoot, { recursive: true }); const root = fs.mkdtempSync(path.join(fixtureRoot, 'test-')); const projectRoot = path.join(root, 'project'); fs.mkdirSync(projectRoot); fs.writeFileSync(path.join(projectRoot, 'README.md'), '# Fixture project\n\nA local test project with TODO work.\n'); const data = path.join(root, 'data'); + // Existing regression scenarios intentionally exercise the original finite + // budget/compaction contract. New-default coverage opts into a fresh install. + if (!modernDefaults) { + fs.mkdirSync(data); + fs.writeFileSync(path.join(data, 'config.json'), JSON.stringify({ ...DEFAULT_CONFIG, contextLimit: 16000, maxOutputTokens: 2048, capacityMode: 'manual', historyPolicy: 'compact' })); + } const store = new Store(data); const project = store.addProject(projectRoot); const session = store.createSession(project.id); diff --git a/tests/model-capacity.test.mjs b/tests/model-capacity.test.mjs new file mode 100644 index 00000000..68174eea --- /dev/null +++ b/tests/model-capacity.test.mjs @@ -0,0 +1,82 @@ +import test from 'node:test'; +import assert from 'node:assert/strict'; +import { fixture, compatibleFixture, completion, until } from './helpers.mjs'; +import { Provider } from '../src/server/provider.mjs'; +import { Runtime } from '../src/server/runtime.mjs'; +import { DEFAULT_CONFIG, cleanConfig } from '../src/core/storage/legacy-import.mjs'; +import { validateRuntimeBudgets } from '../src/core/models/runtime-settings.mjs'; +import { capacitySettings } from '../src/shared/contracts/model-capacity.mjs'; +import { completionBody } from '../src/shared/contracts/model-request.mjs'; +import { budgetFor } from '../src/core/context/request-plan.mjs'; + +test('fresh capacity defaults are independent from legacy frozen budgets; valid large capacities are not clamped', () => { + assert.equal(DEFAULT_CONFIG.capacityMode, 'model'); + assert.ok(DEFAULT_CONFIG.maxOutputTokens >= 16384); + const old = cleanConfig({ contextLimit: 16000, maxOutputTokens: 2048 }); + assert.equal(old.capacityMode, 'manual'); assert.equal(old.maxOutputTokens, 2048); + assert.deepEqual(capacitySettings({ contextLimit: 1000000, maxOutputTokens: 131072 }), { contextLimit: 1000000, maxOutputTokens: 131072 }); + for (const value of [-1, 1.5, NaN, Infinity, Number.MAX_SAFE_INTEGER + 1, '64000']) assert.throws(() => capacitySettings({ maxOutputTokens: value })); + assert.deepEqual(validateRuntimeBudgets({ capacityMode: 'model' }), { capacityMode: 'model' }); + assert.equal(validateRuntimeBudgets({ maxOutputTokens: 65536 }).capacityMode, 'manual'); + assert.throws(() => validateRuntimeBudgets({ capacityMode: 'unlimited' })); +}); + +test('model mode removes the second global cap; manual mode and connection edits retain explicit limits', t => { + const f = fixture(t), hub = new Provider().modelHub(f.store); + const c = hub.upsert({ name: 'Large', model: 'owned', baseUrl: 'http://127.0.0.1:1', requiresKey: false, contextLimit: 256000, maxOutputTokens: 65536 }); + hub.select({ type: 'connection', id: c.id }); + const low = { ...f.store.config, contextLimit: 16000, maxOutputTokens: 2048 }; + assert.equal(hub.prepare({ ...low, capacityMode: 'model' }).maxOutputTokens, 65536); + assert.equal(hub.prepare({ ...low, capacityMode: 'manual' }).maxOutputTokens, 2048); + const changed = hub.upsert({ ...c, contextLimit: undefined, maxOutputTokens: undefined, name: 'Renamed' }); + assert.equal(changed.maxOutputTokens, 65536); assert.equal(changed.contextLimit, 256000); + const b = budgetFor({ ...low, capacityMode: 'model', contextLimit: 256000, maxOutputTokens: 65536 }); + assert.ok(b.targetInputTokens > 16000); assert.equal(b.targetInputTokens, b.inputBudget); +}); + +test('upstream capacities survive a 2048 source connection and are preserved in every protocol wire field', async t => { + const f = fixture(t), server = await compatibleFixture(t, () => ({ data: [{ id: 'owned-large', context_length: 1000000, max_output_tokens: 131072 }] })); + const hub = new Provider().modelHub(f.store), source = hub.upsert({ name: 'Small source', model: 'owned-source', baseUrl: server.baseUrl, requiresKey: false, contextLimit: 16000, maxOutputTokens: 2048 }); + const preview = await hub.modelCatalog().fetch({ id: source.id, revision: source.revision }); + const result = hub.modelCatalog().apply({ previewId: preview.id, models: ['owned-large'], routing: 'manual', defaultModel: 'owned-large' }); + const target = hub.connection(result.added[0]); + assert.equal(target.contextLimit, 1000000); assert.equal(target.maxOutputTokens, 131072); + for (const [protocol, field] of [['chat-completions', 'max_tokens'], ['responses', 'max_output_tokens'], ['anthropic-messages', 'max_tokens']]) { + const body = completionBody({ ...target, protocol, outputTokenField: undefined, parameterProfile: 'auto' }, [{ role: 'user', content: 'Keep the configured capacity.' }], []); + assert.equal(body[field], 131072); + } + assert.equal(server.requests.length, 1, 'only one local catalog request, no model calls'); +}); + +test('actual HTTP sends the selected output capacity and preserves a long complete answer and raw response', async t => { + const f = fixture(t, { modernDefaults: true }), text = 'A'.repeat(30000) + 'END-OF-COMPLETE-ANSWER'; + const server = await compatibleFixture(t, () => completion([], text)); + const provider = new Provider(), hub = provider.modelHub(f.store); + const connection = hub.upsert({ name: 'Owned large answer', model: 'owned', baseUrl: server.baseUrl, requiresKey: false, maxOutputTokens: 32768 }); + hub.select({ type: 'connection', id: connection.id }); + const runtime = new Runtime(f.store, provider); f.onCleanup(() => runtime.stopAll()); + runtime.start(f.session.id, 'Reply with reference text only; do not create or modify files.'); + await until(() => !runtime.isActive(f.session.id), 15000); + assert.equal(f.session.status, 'completed'); assert.equal(server.requests.length, 1); + assert.equal(server.requests[0].input.max_tokens, 32768); + assert.equal(f.session.providerMessages.at(-1).content, text); + const attempt = hub.attempts.at(-1), artifact = f.store.core.artifact.get(attempt.responseArtifactId); + assert.equal(JSON.parse(f.store.artifactStore.read(artifact)).choices[0].message.content, text); +}); + +test('automatic routing sends each member own output capacity rather than the smallest pool member allowance', async t => { + const f = fixture(t, { modernDefaults: true }); let count = 0; + const server = await compatibleFixture(t, () => ++count === 1 ? { status: 503, body: { error: { message: 'Owned retry boundary' }, usage: { prompt_tokens: 20, completion_tokens: 0, total_tokens: 20 } } } : completion([], 'Complete.')); + const provider = new Provider(), hub = provider.modelHub(f.store); + const members = [4096, 32768].map((maxOutputTokens, index) => hub.upsert({ name: 'Owned route ' + index, model: 'owned-' + index, baseUrl: server.baseUrl, requiresKey: false, contextLimit: 128000, maxOutputTokens })); + const pool = hub.pool({ name: 'Owned unequal capacity', members: members.map(member => ({ connectionId: member.id, weight: 1 })), strategy: 'round-robin', sticky: false, maxAttempts: 2, queueTimeoutMs: 15000 }); + hub.select({ type: 'pool', id: pool.id }); + f.store.config.maxRunTokens = 400000; f.store.config.maxTaskTokens = 800000; + const runtime = new Runtime(f.store, provider); f.onCleanup(() => runtime.stopAll()); + runtime.start(f.session.id, 'Reply only, do not modify files.'); await until(() => !runtime.isActive(f.session.id), 15000); + assert.equal(f.session.status, 'completed', JSON.stringify(f.session.events.slice(-4))); + assert.equal(server.requests.length, 2); + for (const request of server.requests) assert.equal(request.input.max_tokens, members.find(member => member.model === request.input.model).maxOutputTokens); + assert.equal(new Set(server.requests.map(row => row.input.max_tokens)).size, 2); + assert.equal([...f.store.core.context_plan.values()][0].budget.outputReservation, 32768); +}); diff --git a/tests/model-catalog.test.mjs b/tests/model-catalog.test.mjs index 82870a26..6dfa02b4 100644 --- a/tests/model-catalog.test.mjs +++ b/tests/model-catalog.test.mjs @@ -36,7 +36,7 @@ test('new upstream retrieves more than 100 models without a manual model; choosi const [source, child] = result.added.map(id => hub.connection(id)); assert.equal(provider.credentials.size, 1); assert.equal(hub.credential(source), hub.credential(child)); assert.equal(child.credentialSource.connectionId, source.id); assert.equal(child.upstreamModel.id, 'owned-7'); - assert.equal(child.contextLimit, 16000); assert.equal(child.maxOutputTokens, 2048); + assert.equal(child.contextLimit, 200000); assert.equal(child.maxOutputTokens, 8192); assert.equal(hub.status(child).capabilities.vision, 'protocol_declared'); assert.equal(hub.status(child).capabilityEvidence, undefined); assert.equal(f.store.config.mode, 'compatible'); assert.equal(hub.state.selected.id, child.id); const prepared = hub.prepare(f.store.config), snapshot = modelSnapshot(prepared); diff --git a/tests/prefix-stability.test.mjs b/tests/prefix-stability.test.mjs new file mode 100644 index 00000000..5336afb1 --- /dev/null +++ b/tests/prefix-stability.test.mjs @@ -0,0 +1,81 @@ +import test from 'node:test'; +import assert from 'node:assert/strict'; +import { fixture, compatibleFixture, completion, until, call } from './helpers.mjs'; +import { Provider } from '../src/server/provider.mjs'; +import { Runtime } from '../src/server/runtime.mjs'; +import { RequestPlanner, verifyPlanEvidence } from '../src/core/context/request-plan.mjs'; +import { completionBody } from '../src/shared/contracts/model-request.mjs'; +import { reportedUsage } from '../src/core/models/protocols.mjs'; +import { prefixAudit } from '../src/core/context/prefix-audit.mjs'; + +test('complete history above 16K is retained verbatim; an oversized history pauses without a compacted replacement', t => { + const f = fixture(t, { modernDefaults: true }), config = { ...f.store.config, contextLimit: 128000 }; + const run = f.store.beginRun(f.session, 'Keep the complete original text.', false, config); + const history = [{ role: 'user', content: 'Keep the original request.' }, { role: 'assistant', content: 'Owned original reference\n'.repeat(3500) }]; + f.session.providerMessages.push(...history); + const planner = new RequestPlanner(f.store), plan = planner.prepare(f.session, run, config, { system: 'Stable rules', tools: [], contextTail: 'State A' }); + assert.equal(plan.record.archivedGroups, 0); assert.ok(plan.record.estimatedInputTokens > 16000); + assert.deepEqual(plan.body.messages.slice(1, 3), history); assert.equal(plan.record.retainedMessages, history.length); + assert.doesNotThrow(() => verifyPlanEvidence(f.store, f.session, plan.record)); + const before = f.store.core.context_plan.size; + assert.throws(() => planner.prepare(f.session, run, { ...config, contextLimit: 18000 }, { system: 'Stable rules', tools: [] }), cause => cause.reason === 'context_limit'); + assert.equal(f.store.core.context_plan.size, before); assert.deepEqual(f.session.providerMessages, history); +}); + +test('runtime keeps fixed rules and original user messages in front of changing state; cache details remain upstream facts', async t => { + const f = fixture(t, { modernDefaults: true }); + const server = await compatibleFixture(t, (_url, body) => ({ ...completion(body.messages.some(row => row.role === 'tool') ? [] : [call('owned-read', 'list_files', { path: '' })], 'Read once.'), + usage: { prompt_tokens: 1000, completion_tokens: 20, prompt_tokens_details: { cached_tokens: 700 } } })); + const provider = new Provider(), hub = provider.modelHub(f.store), c = hub.upsert({ name: 'Owned stable', model: 'owned', baseUrl: server.baseUrl, requiresKey: false }); + hub.select({ type: 'connection', id: c.id }); + const runtime = new Runtime(f.store, provider); f.onCleanup(() => runtime.stopAll()); + runtime.start(f.session.id, 'Inspect once, answer only; do not change files.'); await until(() => !runtime.isActive(f.session.id), 15000); + assert.equal(f.session.status, 'completed'); assert.equal(server.requests.length, 2); + const [first, second] = server.requests.map(row => row.input); + assert.deepEqual(first.tools, second.tools); assert.deepEqual(first.messages[0], second.messages[0]); + assert.deepEqual(first.messages[1], second.messages[1]); assert.equal(first.messages.at(-1).role, 'user'); + assert.ok(!first.messages[0].content.includes('当前Task文件提案状态')); + const plans = [...f.store.core.context_plan.values()]; + for (const plan of plans) assert.doesNotThrow(() => verifyPlanEvidence(f.store, f.session, plan)); + assert.ok(plans[1].prefixAudit.variants[0].commonPrefixBytes > Buffer.byteLength(first.messages[0].content)); + assert.equal(plans[1].prefixAudit.upstreamCache, 'unknown'); + assert.equal(plans[1].actualUsage.prompt_tokens_details.cached_tokens, 700); + const taskId = f.session.taskId; + for (let index = 0; index < 2; index++) { + runtime.start(f.session.id, '继续'); await until(() => !runtime.isActive(f.session.id), 15000); + assert.equal(f.session.taskId, taskId); + } + const continued = [...f.store.core.context_plan.values()]; + assert.equal(server.requests.length, 4); + assert.equal(f.session.providerMessages.filter(row => row.role === 'user' && row.content === '继续').length, 2); + for (let index = 2; index < continued.length; index++) { + assert.equal(continued[index].prefixAudit.previousPlanId, continued[index - 1].id); + assert.doesNotThrow(() => verifyPlanEvidence(f.store, f.session, continued[index])); + } +}); + +test('Messages keeps changing state out of native system blocks and detects tool or route changes', () => { + const cfg = { protocol: 'anthropic-messages', model: 'owned', maxOutputTokens: 32768, baseUrl: 'https://owned.test' }; + const history = [{ role: 'system', content: 'Fixed native rules' }, { role: 'user', content: 'Original request' }]; + const one = completionBody(cfg, [...history, { role: 'user', content: 'State A' }], []); + const two = completionBody(cfg, [...history, { role: 'assistant', content: 'Observed.' }, { role: 'user', content: 'State B' }], []); + assert.deepEqual(one.system, two.system); assert.equal(two.system.length, 1); + assert.ok(prefixAudit(two, one).variants[0].commonPrefixBytes > 40); + assert.equal(prefixAudit({ ...two, tools: [{ name: 'new-tool' }] }, two).variants[0].reason, 'tool_surface_changed'); + assert.equal(prefixAudit({ ...two, model: 'different' }, two).variants[0].commonPrefixBytes, 0); +}); + +test('Chat, Responses and DeepSeek cache receipts preserve raw counters; missing or impossible counters stay unknown', () => { + for (const [config, raw, expected] of [ + [{ protocol: 'chat-completions' }, { prompt_tokens: 1000, completion_tokens: 50, prompt_tokens_details: { cached_tokens: 750 } }, 750], + [{ protocol: 'responses' }, { input_tokens: 1000, output_tokens: 50, input_tokens_details: { cached_tokens: 800 } }, 800], + [{ protocol: 'chat-completions' }, { prompt_tokens: 1000, completion_tokens: 50, prompt_cache_hit_tokens: 900, prompt_cache_miss_tokens: 100 }, 900], + ]) { + const usage = reportedUsage(config, { usage: raw }); + assert.equal(usage.prompt_tokens_details.cached_tokens, expected); assert.deepEqual(usage.raw, raw); assert.equal(usage.total_tokens, 1050); + } + for (const cached of [undefined, -1, 1001, '900']) assert.equal(reportedUsage({}, { usage: { prompt_tokens: 1000, completion_tokens: 50, prompt_cache_hit_tokens: cached } }).prompt_tokens_details, undefined); + const messages = { protocol: 'anthropic-messages' }; + assert.equal(reportedUsage(messages, { usage: { input_tokens: 1000, output_tokens: 50 } }).prompt_tokens_details, undefined); + assert.equal(reportedUsage(messages, { usage: { input_tokens: 1000, output_tokens: 50, cache_read_input_tokens: 0 } }).prompt_tokens_details.cached_tokens, 0); +}); diff --git a/tests/task-control.test.mjs b/tests/task-control.test.mjs index 79cd2c51..f46febff 100644 --- a/tests/task-control.test.mjs +++ b/tests/task-control.test.mjs @@ -5,7 +5,7 @@ import path from 'node:path'; import { Runtime } from '../src/server/runtime.mjs'; import { Provider } from '../src/server/provider.mjs'; import { applyChange } from '../src/server/files.mjs'; -import { isDiagnosticQuestion, taskContract } from '../src/shared/contracts/task-control.mjs'; +import { isDiagnosticQuestion, taskContract, updateTaskContract, taskToolAllowed } from '../src/shared/contracts/task-control.mjs'; import { fixture, call, assertToolPairs } from './helpers.mjs'; const answer = content => ({ message: { role: 'assistant', content }, usage: { prompt_tokens: 11, completion_tokens: 3 } }); @@ -17,8 +17,13 @@ test('operational questions are diagnostic while explicit repair requests remain test('an explicit prohibition on running engineering checks is not interpreted as a requirement to run them', () => { for (const goal of ['只读核对指定文件范围,不修改文件,不运行工程检查。', '不执行工程检查,只解释原结果', - '不需要做任何测试', '无需工程检查', 'Read the evidence; do not run checks.', "Don't run tests; explain the original failure."]) { + '不需要做任何测试', '无需工程检查', 'Read the evidence; do not run checks.', "Don't run tests; explain the original failure.", + 'Inspect the directory once and answer only. Do not modify files or run checks.']) { assert.equal(taskContract({goal}).checks, 'skip', goal); + const task = { goal: 'Run checks.', instructions: [] }; + updateTaskContract(task, goal); assert.equal(task.contract.checks, 'skip', goal); + assert.equal(taskToolAllowed(task, { purpose: 'execution' }, 'run_checks'), false); + updateTaskContract(task, 'Now run checks.'); assert.equal(task.contract.checks, 'required'); } for (const goal of ['修复并运行工程检查', '运行并通过原测试', 'Fix the source and run checks.']) assert.equal(taskContract({goal}).checks, 'required', goal); });