diff --git a/docs/RELIABILITY-1.25.2.md b/docs/RELIABILITY-1.25.2.md index fed4ec9d..4f021b1b 100644 --- a/docs/RELIABILITY-1.25.2.md +++ b/docs/RELIABILITY-1.25.2.md @@ -47,3 +47,9 @@ W六请求71044 tokens、0提案0交付是事实;末轮新增468字符、还 每个真实工具轮开始/结算记录新返回字符及重复读取数,按当前Run区间并集计算;同轮确有新字符时,不因尾部三次重复而整轮暂停,并在边界清掉重复streak。空页/仅改变历史索引不算新源码依据;全重复仍暂停,未变检查/相同无效提案/连续相同状态的独立保护仍权威。shadow模块与其判断不参与停止决策。 联合43/43通过后补强独立相同状态保护,针对性2/2通过。HTTP夹具忠实复现945→1413新增468及三小模块重复,下一真实本机请求收到完整新页且工具配对成立;全重复和独立保护仍暂停。测试保持16K/2048与原权限门,不证明真实模型一定交付。原证据在reading-contract.tap、reading-mixed-final.tap。 + +### 纯读取批次吞吐 + +readBudgetPolicyVersion=3把多文件/原工件读取作为共享预算批次,按实际请求大小优先完成小来源,释放剩余份额给测试/日志。仅纯读取批次排序,工具调用ID与原参数不变;变更、检查、审批和任意混合动作保持原序。单工件读取沿用原分页分配,v2封存保留原算法;完整下一请求、用量预留、16K/2048/1600保护不放宽。 + +首轮38项33通过5失败:三项断言仍固定新版策略号为2;另外两项揭示把整个混合文件/日志组强留会挤掉已有依据。因此保留原planner纯文件组规则,不将混合原日志组提升到必留;单工件分页保持原策略。重新38/38通过,三协议Unicode/CRLF分页、原片段聚合/检查事实/来源篡改/重开验证均通过。新增W形状HTTP夹具保留三完整小模块和原日志,第一测试页>945字符(原W首轮473);这是一项受控吞吐证据,实际新样本另验。原read-throughput.tap与read-throughput-final.tap保留;完整构建通过。 diff --git a/progress.md b/progress.md index 331a7223..d5e7a611 100644 --- a/progress.md +++ b/progress.md @@ -10,6 +10,7 @@ - 安装文件变化复核12/12通过;10971文件只读ABBA全校验8.15–8.45s→9.96–11.08s、四次身份exact,记录新增检查的真实成本。已补阶段回顾并独立本地提交。 - 新读取HTTP用例首33项32通过1失败;有第3请求但被误设required的检查阻断。debug证实verification_incomplete;补明确“不运行工程检查”等否定及正向对照后,读取/无进展/shadow/task-control联合43/43通过。原首失败保留,task-control先独立提交。 - 混合读取按每工具轮真实新增区间结算;新468字符+3模块重复可进入下一请求,纯重复/相同检查/相同提案/相同状态保护保持。43/43后独立保护补强2/2通过,回顾后本地提交。源码定位曾误用task-contract文件名,改真实task-control;Windows rg文件通配失败改-g过滤。验证依赖中两套现有Python缺pytest,已定位Python310/312供后续核对;浏览器下载进行中。 +- 读取吞吐v3纯读批次共享/小来源先完成,首38项33通过5失败保留;去掉会挤走原源码的混合组强留、保持单工件原分页并更新新版号断言后最终38/38、完整构建通过。补丁一度误删preparedSets,diff检查立即恢复,未进入测试或提交。Python310自有venv已装固定pytest8.4.2;Chromium仍后台下载。 ## 1.25接续 diff --git a/src/core/context/read-budget.mjs b/src/core/context/read-budget.mjs index 3781cc7b..02161b6d 100644 --- a/src/core/context/read-budget.mjs +++ b/src/core/context/read-budget.mjs @@ -5,6 +5,8 @@ export const BROWSER_READ_HINT = '当前Run已经返回此页面结果区间, export const readMetadataReserve = metadata => Object.keys(metadata).length ? Buffer.byteLength(JSON.stringify(JSON.stringify(metadata)), 'utf8') + 8 : 0; const boundary = (text, end) => end > 0 && end < text.length && /[\uD800-\uDBFF]/.test(text[end - 1]) ? end - 1 : end; +export const isPagedRead = call => ['read_file', 'read_artifact'].includes(call.function.name); +const pooledReads = calls => calls?.length && calls.every(isPagedRead) && (calls.length > 1 || calls[0].function.name === 'read_file'); // This is the actual tool page, not a rewrite of a historical result. Count // the escaped JSON tool message which will be present in the next request. @@ -36,8 +38,8 @@ export function readMessageBudget(plan, message) { // The next request also gains bounded read coverage and route observations. // Reserve their growth before returning a page, rather than dropping that // new page or enlarging the request budget afterwards. - const pureReads = plan.readBudgetPolicyVersion === 2 && message.tool_calls?.length - && message.tool_calls.every(call => call.function.name === 'read_file'); + const pureReads = message.tool_calls?.length && (plan.readBudgetPolicyVersion === 3 && pooledReads(message.tool_calls) + || plan.readBudgetPolicyVersion === 2 && message.tool_calls.every(call => call.function.name === 'read_file')); // A pure read batch cannot create a plan, command output or change set. Its // bounded state/coverage growth uses 512 tokens plus 256 for each additional // read. Mixed batches and historical plans retain the original 2560 reserve. @@ -50,9 +52,19 @@ export function readMessageBudget(plan, message) { export function readBatchBudget(plan, message) { const maximum = readMessageBudget(plan, message); - if (plan.readBudgetPolicyVersion !== 2 || !message.tool_calls?.length - || !message.tool_calls.every(call => call.function.name === 'read_file')) return null; - return { remainingBytes: maximum * message.tool_calls.length, remainingCalls: message.tool_calls.length }; + if (!message.tool_calls?.length || !(plan.readBudgetPolicyVersion === 3 && pooledReads(message.tool_calls) + || plan.readBudgetPolicyVersion === 2 && message.tool_calls.every(call => call.function.name === 'read_file'))) return null; + return { policyVersion: plan.readBudgetPolicyVersion, remainingBytes: maximum * message.tool_calls.length, remainingCalls: message.tool_calls.length }; +} + +export function orderReadCalls(batch, calls, estimateBytes) { + if (batch?.policyVersion !== 3) return calls; + // Independent reads finish with their original IDs. Small sources release + // unused space to larger files/logs even if the model listed the latter first. + return calls.map((call, index) => { + let bytes = Infinity; try { const value = estimateBytes(call); if (Number.isFinite(value) && value >= 0) bytes = value; } catch {} + return { call, index, bytes }; + }).sort((a, b) => a.bytes - b.bytes || a.index - b.index).map(value => value.call); } export function spendReadBudget(batch, call, result) { diff --git a/src/core/context/request-plan.mjs b/src/core/context/request-plan.mjs index 493f0601..233b0aab 100644 --- a/src/core/context/request-plan.mjs +++ b/src/core/context/request-plan.mjs @@ -357,7 +357,7 @@ export class RequestPlanner { workingRetentionTokens: Math.max(0, metricsFor(new Set(lastUser ? [lastUser.index] : []), true, retainedSources).estimatedInputTokens - metricsFor(new Set(lastUser ? [lastUser.index] : []), true, []).estimatedInputTokens), workingRetentionPolicyVersion: 1, - readBudgetPolicyVersion: 2, + 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), archivedGroups: omitted.length, retainedGroups: groups.filter(group => selected.has(group.index)).map(group => group.index), historyIndexArtifactId: reference?.id ?? null, historyIndexSha256: reference?.sha256 ?? null, actualUsage: null, diff --git a/src/server/runtime.mjs b/src/server/runtime.mjs index d78d0c10..bfa867b9 100644 --- a/src/server/runtime.mjs +++ b/src/server/runtime.mjs @@ -28,7 +28,7 @@ import { RequestPlanner, verifyPlanEvidence } from '../core/context/request-plan import { ENGINE_SEED_RESERVATION } from '../shared/contracts/engine-context.mjs'; import { InputAttachments } from '../core/context/input-attachments.mjs'; import { BudgetResume } from '../core/context/budget-resume.mjs'; -import { artifactPage, filePage, readMessageBudget, readBatchBudget, spendReadBudget, recordArtifactRead, recordBrowserRead, recordFileRead, recordUnchangedCheck, clearReadStreak, recordSessionInspection, recordUnchangedAction, readCoverageContext, readMetadataReserve, BROWSER_READ_HINT, beginReadTurn, finishReadTurn } from '../core/context/read-budget.mjs'; +import { artifactPage, filePage, readMessageBudget, readBatchBudget, spendReadBudget, recordArtifactRead, recordBrowserRead, recordFileRead, recordUnchangedCheck, clearReadStreak, recordSessionInspection, recordUnchangedAction, readCoverageContext, readMetadataReserve, BROWSER_READ_HINT, beginReadTurn, finishReadTurn, orderReadCalls } from '../core/context/read-budget.mjs'; import { taskDeliveryProgress } from '../core/context/delivery-progress.mjs'; import { ChangeSetManager } from '../core/workspaces/change-sets.mjs'; import { assertChangeSetsResolved } from '../shared/contracts/change-sets.mjs'; @@ -484,6 +484,13 @@ export class Runtime { beginReadTurn(task, run.id, stepRecord.id); run.readPageCharacters = planned.record.budget.inputBudget - planned.record.nextReadFixedInputTokens - (planned.record.workingRetentionTokens ?? 0) < 4096 ? 2000 : Infinity; + const executionCalls = orderReadCalls(run.readBatchBudget, calls, call => { + const args = JSON.parse(call.function.arguments); + validateToolArgs(call.function.name, args); + const size = call.function.name === 'read_file' ? fs.statSync(safePath(run.project, args.path)).size + : this.store.owned('artifact', args.artifact_id, session.id).bytes; + return Math.max(0, Math.min(size - (args.offset ?? 0), args.limit ?? (call.function.name === 'read_artifact' ? 2000 : Infinity))); + }); const previousIds = new Set(session.providerMessages.flatMap(item => (item.tool_calls ?? []).map(call => call.id))); if (new Set(calls.map(call => call.id)).size !== calls.length || calls.some(call => previousIds.has(call.id))) throw new Error('模型返回了重复的工具调用 ID,请重新发送任务。'); session.providerMessages.push(message); @@ -539,7 +546,7 @@ export class Runtime { } let deliveryReview = null, mcpReview = null; const preparedSets = new Set(); - for (const call of calls) { + for (const call of executionCalls) { let result; if (deliveryReview || mcpReview) { const event = session.events.find(value => value.toolCallId === call.id && value.runId === run.id); diff --git a/task_plan.md b/task_plan.md index 3b37d306..a626466b 100644 --- a/task_plan.md +++ b/task_plan.md @@ -6,8 +6,8 @@ 1. [complete] 预留v3绑定原账本消费前缀,拒绝重排/重复/跨Task;保留v1/v2算法和原封存核验,输入异常不再丢弃已报告输出上界。专项13/13、上下文/预算/引擎相关24/24通过。 2. [complete] inventory恢复每文件treeFile检查,扫描后新增硬链接被拒,原身份/字节一致;12/12安装通过,关联11项通过与升级3/3复验。10971文件只读ABBA全校验旧8.15–8.45s、新9.96–11.08s,未恢复逐文件祖先遍历。 -3. [in_progress] 修复混合进展读取停止,保留真正重复/相同检查/相同提案的有界停止;核对实际请求中读页与检查事实并提高可用分页吞吐。 -4. [pending] 完整相关/全量/构建与实际界面、不可变发行及隔离安装验收;保留状态升级1.25.2,阶段回顾和本地提交。 +3. [complete] 混合进展按同轮实际新增区间结算,独立无变化保护保持;新v3读取批次释放小来源剩余预算给文件/日志、变更操作原序不动。联合43/43、补强2/2及吞吐最终38/38与构建通过,首失败保留。 +4. [in_progress] 完整相关/全量/构建与实际界面、不可变发行及隔离安装验收;保留状态升级1.25.2,阶段回顾和本地提交。 5. [pending] 新冻结单次诊断/身份预检/有界Node;通过后才React/Python及Purge AB/BA,分别报告失败、成本和因果边界,不循环重试求通过。 细节和阶段回顾见docs/RELIABILITY-1.25.2.md。开始时session-catchup无待同步输出;计划文件首次补丁因findings标题上下文不符被拒且无文件变化,已按原标题修正。以下按历史时点理解。 diff --git a/tests/reading-reliability.test.mjs b/tests/reading-reliability.test.mjs index 1552e65f..7daaba1a 100644 --- a/tests/reading-reliability.test.mjs +++ b/tests/reading-reliability.test.mjs @@ -4,7 +4,7 @@ import fs from 'node:fs'; import path from 'node:path'; import { createHash } from 'node:crypto'; import { fileContinuation } from '../src/core/context/read-continuation.mjs'; -import { filePage, readMessageBudget, readBatchBudget, spendReadBudget } from '../src/core/context/read-budget.mjs'; +import { filePage, readMessageBudget, readBatchBudget, spendReadBudget, orderReadCalls } from '../src/core/context/read-budget.mjs'; import { readLines } from '../src/server/files.mjs'; import { Runtime, validateToolArgs } from '../src/server/runtime.mjs'; import { Provider } from '../src/server/provider.mjs'; @@ -47,6 +47,49 @@ test('pure-file batch returns unused small-page space while mixed and historical assert.equal(readBatchBudget(plan, { ...message, tool_calls: [call('artifact', 'read_artifact', { artifact_id: 'raw' })] }), null); }); +test('v3 pools mixed file/log reads, serves small sources first and leaves historical/mutating batches unchanged', () => { + const plan={readBudgetPolicyVersion:3,budget:{inputBudget:12352},nextReadFixedInputTokens:9000,workingRetentionTokens:600}; + const calls=[call('large','read_file',{path:'large'}),call('log','read_artifact',{artifact_id:'log'}),call('small','read_file',{path:'small'})]; + const message={role:'assistant',content:'',tool_calls:calls},batch=readBatchBudget(plan,message); + assert.ok(batch); assert.equal(batch.policyVersion,3); + assert.deepEqual(orderReadCalls(batch,calls,c=>({large:2584,small:51,log:6000})[c.id]).map(c=>c.id),['small','large','log']); + assert.deepEqual(calls.map(c=>c.id),['large','log','small']); + assert.ok(readMessageBudget(plan,message)>readMessageBudget({...plan,readBudgetPolicyVersion:2},message)); + assert.equal(readBatchBudget({...plan,readBudgetPolicyVersion:2},message),null); + assert.equal(orderReadCalls(null,calls,()=>0),calls); + assert.equal(readBatchBudget(plan,{...message,tool_calls:[...calls,call('write','propose_file',{path:'out',content:'x'})]}),null); + assert.deepEqual(orderReadCalls(batch,calls,c=>{if(c.id==='large')throw Error('unavailable');return 10;}).map(c=>c.id),['log','small','large']); +}); + +test('v3 W-shaped mixed reading returns complete small modules and substantially larger first test pages in the next bounded request', async t => { + const f=fixture(t),text=fs.readFileSync(new URL('../benchmarks/live-delivery-v1/node/test/expense.test.mjs',import.meta.url),'utf8'); + fs.writeFileSync(path.join(f.projectRoot,'expense.test.mjs'),text); + for(const [name,content] of [['amount','export const amount = 0;\n'],['summary','export const summary = [];\n'],['report','export const report = "";\n']])fs.writeFileSync(path.join(f.projectRoot,name+'.mjs'),content); + const artifact=f.store.artifactStore.write(f.session,null,'Original failure trace\n'.repeat(500),{kind:'check-log'}); + f.store.core.artifact.set(artifact.id,artifact);f.store.save(f.session); + let requests=0,page; + const model=await compatibleFixture(t,(route,body)=>{ + requests++; + if(requests===1)return{body:completion([call('large-first','read_file',{path:'expense.test.mjs',limit:8000}), + ...['amount','summary','report'].map(name=>call(name,'read_file',{path:name+'.mjs',limit:4000})), + call('log','read_artifact',{artifact_id:artifact.id,limit:8000})])}; + const results=body.messages.filter(m=>m.role==='tool').map(m=>JSON.parse(m.content));page=results.find(v=>v.path==='expense.test.mjs'); + assert.ok(page);assert.ok(page.next_offset>945,JSON.stringify(page));assert.equal(page.content,text.slice(0,page.next_offset)); + for(const name of ['amount','summary','report']){ + const source=results.find(v=>v.path===name+'.mjs');assert.ok(source);assert.equal(source.next_offset,source.total_characters); + } + assert.ok(results.find(v=>v.artifact_id===artifact.id));assertToolPairs(body.messages); + assert.ok(requestMetrics(body).estimatedInputTokens<=12352);return{body:completion([],'完成指定只读核对。')}; + }); + f.store.saveConfig({mode:'compatible',model:'mixed-reading-fixture',baseUrl:model.baseUrl,contextLimit:16000,maxOutputTokens:2048,maxSteps:8}); + const runtime=new Runtime(f.store,new Provider());f.onCleanup(()=>runtime.stopAll()); + runtime.start(f.session.id,'只读取指定测试、三模块和原检查日志,核对真实依据;无需测试,不修改文件。'); + await runtime.active.get(f.session.id).promise; + assert.equal(requests,2,JSON.stringify(f.session.failure));assert.equal(f.session.status,'completed'); + assert.ok(page.next_offset>945);assertToolPairs(f.session.providerMessages); + for(const plan of f.store.core.context_plan.values())assert.doesNotThrow(()=>verifyPlanEvidence(f.store,f.session,plan)); +}); + test('observed upstream high input pauses before a new request, preserves reported usage and distinguishes the limiting ledger', t => { const f = fixture(t), config = { ...f.store.config, model: 'owned', contextLimit: 16000, maxOutputTokens: 2048, maxRunTokens: 90000, maxTaskTokens: 90000 }; @@ -108,7 +151,7 @@ for (const protocol of ['chat-completions', 'responses', 'anthropic-messages']) verifyPlanEvidence(f.store, f.session, plan); const body = JSON.parse(f.store.artifactStore.read(f.store.core.artifact.get(plan.requestArtifactId))); assert.ok(requestMetrics(body).estimatedInputTokens <= 12352); assert.equal(plan.budget.safetyTokens, 1600); - assert.equal(plan.readBudgetPolicyVersion, 2); + assert.equal(plan.readBudgetPolicyVersion, 3); } assert.equal(fs.readFileSync(path.join(f.projectRoot, 'long.txt'), 'utf8'), text); });