diff --git a/findings.md b/findings.md index f00366d2..9ee49b7c 100644 --- a/findings.md +++ b/findings.md @@ -984,3 +984,13 @@ D原版加载实证:31模块闭包还依赖dsh-llm,60包输入不足以代 G修复开发:完整批次交接旧回归2例失败、边界5例通过;修改后9/9通过。React原件离线以原system/tools/历史/32K/逐请求剩余额度重建,修复前第7–12次估算全部与原件相同,六文件完整保留为6/5/5/4/2/2;稳定小源码优先后6次均6/6,实际输入15714–16993,仍在原硬预算内,16K软目标未当成硬上限。无需新模型请求,原件保持。QA驱动初始化/SQLite/漏传路由工具的原失败保留,最终补实际路由工具后实现精确复现。 G最终回顾:正式升级的多轮verify-current/verify-candidate分别包含full-manifest、directory-inventory和loaded-inputs,在预检和指针发布前重复核对,属于有界完整性流程。正式C盘目录耗时明显,后续性能优化需保留变更检测、发布前基线和防篡改约束。本轮已完成三项确认修复及1.24.1真实安装/原生窗口/原数据核对;真实模型交付与Purge收益不由本机夹具或离线重放证明。 + +### I 接续:原生整批审查交接 + +原专项通过1/1(native-review-test.tap)。完整批次已封存且Task needs_review时,宿主关闭仍有待返回工具的原生会话被错记failed;新增严格reviewReady/原host结果/进程退出核对后保留paused且verified=false,不当作原生完成。正在跑完整engine-worker及实际固定Harness Chat 32K新隔离样本;原native-chat失败原件保持,无真实配置模型调用或正式宿主改动。 + +恢复时批读历史长文再次超过总输出上限,且查询不存在的.agents/.codex目录报错;截断内容不作审查依据,之后仅读已知文件短段。没有重跑已完成产品流程。 + +原生Chat 32K修复轮已走完子工程原9失败/修复/9通过,但验收器仍要求暂停Run包含completed endEvent,读不存在activity报错;原件保留controlled-context-0d12521e-bc48-4245-a878-48ebba40f679。修正为仅完成Run核对completed,最终Run仍必须verified completed;审查暂停只核对原host结果与关闭,不伪造完成事件。顺带审查发现reviewReady异常可能跳过关闭,已让异常进入原引擎failure且始终关闭,加入审查封条变化的进程退出回归。 + +32K原生新轮两项通过后在“必须发生压缩”旧断言失败(ae604b4f);9次原请求估算6276–18714tokens,审查交接提前关闭历史后自然不超过32K,不是丢失原证据。验收器改为如实记录32K实际压缩次数并继续核对所有请求/原日志/预算,另设16K压力轮仍硬性要求发生已核验压缩;原32K失败不覆盖。最终编译身份未变,额外16K仅自有HTTP夹具新隔离工程,不增加真实模型预算。 diff --git a/progress.md b/progress.md index 0dea78a4..c7071826 100644 --- a/progress.md +++ b/progress.md @@ -1505,3 +1505,15 @@ G封存格式补核:Computer Use原图是JPEG,首保存误用.png扩展名 G首次Git原字节核对明确拒绝:普通目录git add被通用.log忽略规则过滤,未包含全部证据。原失败保留.qa/efficiency-fixes-g/git-index.log;改为仅逐manifest文件显式暂存,不扩大到.qa,再执行索引原字节核对。 G最终Git索引201份/43818454字节全部原字节一致,原E/F manifest保持;独立离线200份证据通过。收尾为文档/证据提交,不改已验证的ba07713发行载荷、正式宿主或原任务。 + +### I 接续:原生整批审查交接 + +原专项通过1/1(native-review-test.tap)。完整批次已封存且Task needs_review时,宿主关闭仍有待返回工具的原生会话被错记failed;新增严格reviewReady/原host结果/进程退出核对后保留paused且verified=false,不当作原生完成。正在跑完整engine-worker及实际固定Harness Chat 32K新隔离样本;原native-chat失败原件保持,无真实配置模型调用或正式宿主改动。 + +恢复时批读历史长文再次超过总输出上限,且查询不存在的.agents/.codex目录报错;截断内容不作审查依据,之后仅读已知文件短段。没有重跑已完成产品流程。 + +原生Chat 32K修复轮已走完子工程原9失败/修复/9通过,但验收器仍要求暂停Run包含completed endEvent,读不存在activity报错;原件保留controlled-context-0d12521e-bc48-4245-a878-48ebba40f679。修正为仅完成Run核对completed,最终Run仍必须verified completed;审查暂停只核对原host结果与关闭,不伪造完成事件。顺带审查发现reviewReady异常可能跳过关闭,已让异常进入原引擎failure且始终关闭,加入审查封条变化的进程退出回归。 + +32K原生新轮两项通过后在“必须发生压缩”旧断言失败(ae604b4f);9次原请求估算6276–18714tokens,审查交接提前关闭历史后自然不超过32K,不是丢失原证据。验收器改为如实记录32K实际压缩次数并继续核对所有请求/原日志/预算,另设16K压力轮仍硬性要求发生已核验压缩;原32K失败不覆盖。最终编译身份未变,额外16K仅自有HTTP夹具新隔离工程,不增加真实模型预算。 + +最终编译宿主两协议各3组/7请求通过。实际Harness最终Chat32K与Chat16K各5组,16K发生4次有原来源压缩,32K自然为0;Responses32K也5组通过。界面截图目视发现身份不一致的行标题仍显示“模型连接失败”,正文虽准确仍易误导;只在React的model_identity行使用已有诊断标题,后端/权限/原事件不改写。增加UI真实断言与发行目录参数,复验最终界面和随包目标。早先最终18文件后端回归继续运行,后端输入保持;此末次改动仅前端呈现及QA入口。 diff --git a/scripts/verify-controlled-projects.mjs b/scripts/verify-controlled-projects.mjs index c96e02d8..6319f2ad 100644 --- a/scripts/verify-controlled-projects.mjs +++ b/scripts/verify-controlled-projects.mjs @@ -163,7 +163,10 @@ async function runSample(kind, { sampleName = kind, strategy = null, comparison } if (failure) throw failure; const native = [...f.store.core.engine_run.values()].filter(job => job.taskId === child.taskId).at(-1); - assert.equal(native.status, 'completed', native.error); + if(session.status==='needs_review'&&native.reviewHandoff){ + assert.equal(native.status,'paused');assert.equal(native.verified,false);assert.equal(native.requiresVerification,false); + assert.equal(native.reviewHandoff.kind,'change_set');assert.equal(native.native.cleanup.treeConfirmed,true); + }else assert.equal(native.status, 'completed', native.error); } assert.equal(group.status, 'ready_to_collect', group.summary); assert.deepEqual(ownerBytes(f), original); verify(childRoot); const childChecks = [...f.store.core.check_job.values()].filter(job => job.taskId === child.taskId); @@ -171,13 +174,16 @@ async function runSample(kind, { sampleName = kind, strategy = null, comparison assert.equal(childChecks.at(-1).results[0].tests.passed, spec.tests); assert.equal(childChecks.at(-1).status, 'passed'); const engineRuns = [...f.store.core.engine_run.values()].filter(job => job.taskId === child.taskId), childSession = f.store.session(child.sessionId); for (const job of engineRuns) { - originalEngineRun(f.store, childSession, job); assert.equal(job.status, 'completed', job.error); assert.equal(job.native.cleanup.treeConfirmed, true); + originalEngineRun(f.store, childSession, job); + if(job.reviewHandoff){assert.equal(job.status,'paused');assert.equal(job.verified,false);}else assert.equal(job.status, 'completed', job.error); + assert.equal(job.native.cleanup.treeConfirmed, true); if(originalPurgeMode){ const warning='(node:'+job.native.process.pid+') [PERM0001] SecurityWarning: The flag --allow-addons must be used with extreme caution. It could invalidate the permission model.\n(Use `node --trace-warnings ...` to show where the warning was created)\n'; const expected=warning+['package.json','lib/index.js'].map(file=>'AWB_MODULE_REFUSED '+path.join(installation.root,'packages/context/agent-instructions',file)+'\n').join(''); assert.equal(job.native.stderr,expected);assert.equal(job.native.originalPurge.readiness.originalApplyCalled,true); assert.equal(job.native.originalPurge.nativeLoads.length,2);assert.ok(job.native.originalPurge.durabilityReceipts.length>0); - }else assert.equal(job.native.stderr, ''); assert.equal(job.native.activity.endEvents[0].data.reason.kind, 'completed'); + }else assert.equal(job.native.stderr, ''); + if(!job.reviewHandoff)assert.equal(job.native.activity.endEvents[0].data.reason.kind, 'completed'); for (const request of job.requests) { const saved = verifyEngineModelInput(f.store, childSession, job, request); if (strategy) { assertPurgeModelInput(saved.native, job.installation.strategy); @@ -186,6 +192,8 @@ async function runSample(kind, { sampleName = kind, strategy = null, comparison else assert.ok(nativeSystemTexts(saved.native).every(text => !text.includes('[Agent Workbench engineering prompt experiment]'))); } } + assert.equal(engineRuns.at(-1).status,'completed','The final native Run must complete after explicit review continuation.'); + assert.equal(engineRuns.at(-1).verified,true); assert.equal(f.store.core.task.get(child.taskId).usage.requests, modelRequests.length); assert.equal(f.store.core.task.get(child.taskId).usage.totalTokens, modelRequests.length * 168); if (responsesMode) { @@ -197,13 +205,13 @@ async function runSample(kind, { sampleName = kind, strategy = null, comparison mark(label + ': actual Harness uses owned model/tool bridges, retains failing tests, and completes reviewed repairs with one accounting entry per HTTP request'); if(contextMode){ const bounded=engineRuns.flatMap(job=>job.requests.map(request=>verifyEngineModelInput(f.store,childSession,job,request))).filter(input=>input.view.mode==='bounded'); - assert.ok(bounded.length>0,'The original long failure/log history must exercise the bounded bridge.'); + if(contextLimit===16000)assert.ok(bounded.length>0,'The tighter original context profile must exercise the bounded bridge.'); assert.ok(bounded.every(input=>input.view.beforeInputTokens>input.view.inputBudget&&input.plan.estimatedInputTokens<=input.view.inputBudget)); - const sent=JSON.stringify(bounded.map(input=>input.body));assert.ok(sent.includes('logArtifactId'));assert.ok(sent.includes('failed')); + const sent=JSON.stringify(engineRuns.flatMap(job=>job.requests.map(request=>verifyEngineModelInput(f.store,childSession,job,request).body)));assert.ok(sent.includes('logArtifactId'));assert.ok(sent.includes('failed')); const failed=childChecks.find(job=>job.status==='failed'),log=f.store.core.artifact.get(failed.results[0].logArtifactId); assert.ok(f.store.artifactStore.read(log).includes('fail 9'));assert.ok(engineRuns.some(job=>job.calls.some(call=>call.name==='read_artifact'))); assert.ok(engineRuns.some(job=>job.calls.some(call=>call.name==='prepare_change_set')));assert.ok(engineRuns.every(job=>!job.inputRejections?.length)); - mark(contextLimit/1000+'K: original nine failures and paged log survive verified native compaction; all three proposals form one reviewed batch without replay'); + mark(contextLimit/1000+'K: original nine failures and paged log retained; '+bounded.length+' native requests used verified compaction; all three proposals form one reviewed batch without replay'); } await (await manager.collect(f.session, group.id, group.fingerprint)).promise; assert.equal(group.status, 'collected', group.summary); const collection = await manager.readCollection(f.session, group); assert.deepEqual(collection.conflicts, []); assert.deepEqual(collection.paths, writable); diff --git a/src/core/engines/engine-runner.mjs b/src/core/engines/engine-runner.mjs index fe149877..31a9472b 100644 --- a/src/core/engines/engine-runner.mjs +++ b/src/core/engines/engine-runner.mjs @@ -13,6 +13,7 @@ import { workerForSession } from '../../shared/contracts/workers.mjs'; import { selectPurgeStrategy, assertPurgeModelInput } from './purge-strategy.mjs'; import { engineHistoryGroups,engineContextEvidence,engineContextIndex,engineContextView,engineSeedPrompt } from '../../shared/contracts/engine-context.mjs'; import { sanitizeModelText } from '../models/upstream-error.mjs'; +import { ChangeSetManager } from '../workspaces/change-sets.mjs'; const at = () => new Date().toISOString(); const deferred = () => { @@ -51,6 +52,15 @@ export class EngineRunner { } async finish(session, run) { const driver = this.drivers.get(run.id); if (!driver) return; + const review = run.record.review; + try { + if (!driver.failure && !driver.modelActive && !run.controller.signal.aborted && session.status === 'needs_review' + && run.stopReason === 'change_set_review' && review?.kind === 'change_set' + && await new ChangeSetManager(this.store).reviewReady(session, review.changeSetId, run.record)) { + if (this.store.core.change_set.get(review.changeSetId)?.fingerprint !== review.fingerprint) throw engineError('整批审查指纹已变化,不能记录为已核对交接。'); + driver.prepareReviewHandoff(review); + } + } catch (error) { driver.fail(error); } try { await driver.close(); } finally { this.drivers.delete(run.id); } if (driver.job.requiresVerification) { session.status = 'paused'; run.stopReason = 'engine_verification'; @@ -317,6 +327,18 @@ class EngineRunDriver { if (this.closePromise) return this.closePromise; return this.closePromise = this.finishClose(); } + prepareReviewHandoff(review) { + if (this.queue.length || !this.pendingCalls.size) return; + const tools = []; + for (const [callId, pending] of this.pendingCalls) { + const tool = [...this.store.core.tool.values()].find(value => value.runId === this.run.id && value.callId === callId); + const result = this.store.savedToolResult(this.session.id, callId); + if (!pending.delivered || !tool || tool.state !== 'completed' || tool.requiresVerification || !result || !tool.resultArtifactId) return; + tools.push({ callId, invocationId: tool.id, resultArtifactId: tool.resultArtifactId, resultSha256: controlledDigest(result) }); + } + this.job.reviewHandoff = { schemaVersion: 1, kind: 'change_set', changeSetId: review.changeSetId, fingerprint: review.fingerprint, + scope: 'host_results_saved_native_session_closed_for_review_not_completed', tools }; + } async finishClose() { this.job.status = 'closing'; this.save('closing'); for (const pending of this.pendingCalls.values()) pending.result.reject(new Error('所属引擎已结束,未返回结果保留原状态。')); @@ -328,10 +350,13 @@ class EngineRunDriver { const launched = !!this.native?.client || !!native?.process?.pid || !!this.job.pid; const unknown = launched && (!native?.cleanup?.rootExited || native?.cleanup?.treeConfirmed === false || native?.status === 'unknown'); this.job.requiresVerification = unknown; - this.job.status = unknown ? 'unknown' : this.finished && !this.failure && native?.status === 'closed' ? 'completed' + const reviewPaused = !unknown && this.job.reviewHandoff && !this.run.controller.signal.aborted + && ['closed', 'cancelled'].includes(native?.status); + this.job.status = unknown ? 'unknown' : reviewPaused ? 'paused' : this.finished && !this.failure && native?.status === 'closed' ? 'completed' : this.run.controller.signal.aborted ? 'cancelled' : 'failed'; this.job.verified = this.job.status === 'completed'; this.job.finishedAt = at(); - this.job.error = this.failure ? this.provider.redact(this.failure.message) : native?.error ?? null; + if (reviewPaused && this.failure) this.job.reviewHandoff.closeDiagnostic = this.provider.redact(this.failure.message); + this.job.error = reviewPaused ? null : this.failure ? this.provider.redact(this.failure.message) : native?.error ?? null; this.job.originalOutcome = { status: this.job.status, requiresVerification: this.job.requiresVerification, verified: this.job.verified, error: this.job.error }; try { verifyEngineInput(this.store, this.session, this.job); diff --git a/src/server/runtime.mjs b/src/server/runtime.mjs index c2793213..13c7de91 100644 --- a/src/server/runtime.mjs +++ b/src/server/runtime.mjs @@ -556,6 +556,7 @@ export class Runtime { } signal.throwIfAborted(); if (!ready) continue; + run.record.review = { kind: 'change_set', changeSetId: id, fingerprint: this.store.core.change_set.get(id).fingerprint, status: 'pending' }; session.status = 'needs_review'; run.stopReason = 'change_set_review'; this.store.message(session, 'system', '当前任务的完整变更批次已封存,同轮工具结果已保存。请审查并应用或撤销批次;等待审查期间不继续请求模型。应用后可显式续接原任务进行检查。'); this.store.save(session, { type: 'change_set.awaiting_review', detail: { changeSetId: id } }); diff --git a/src/shared/contracts/engine-execution.mjs b/src/shared/contracts/engine-execution.mjs index 1252674f..cf8a1249 100644 --- a/src/shared/contracts/engine-execution.mjs +++ b/src/shared/contracts/engine-execution.mjs @@ -23,6 +23,7 @@ export function engineRunReport(job) { return { schemaVersion: 1, owner: { sessionId: job.sessionId, projectId: job.projectId, taskId: job.taskId, runId: job.runId, jobId: job.id }, inputSha256: job.inputSha256, installationSha256: job.installation.controlledSha256, sdkSessionId: job.sdkSessionId, outcome: job.originalOutcome, native: job.native, requests: job.requests, calls: job.calls, + ...(job.reviewHandoff ? { reviewHandoff: job.reviewHandoff } : {}), ...(job.inputRejections?{inputRejections:job.inputRejections}:{}) }; } export function publicEngineRun(job) { diff --git a/task_plan.md b/task_plan.md index c78124ff..e4576869 100644 --- a/task_plan.md +++ b/task_plan.md @@ -7,7 +7,7 @@ 1. [complete] I-A 请求/返回型号分离、JSON/SSE来源绑定及严格暂停已实现;76相关/16补充、构建和编译宿主3组UI通过,正式未部署;更广引擎/交付回归持续运行。新实测先严格首请求身份预检。 2. [complete] I-B 当前SHA/检查失效/下一可验证动作已接入;2专项、构建/10组UI及编译宿主Chat原3失败→应用→审批复测3通过→Git完整链路通过。Responses及实际Harness补最终验证。 3. [complete] I-C 同SHA区间并集、有限已读摘要、相同提案/复用证明不清进展已修复;原分页回归及新空转10/10。首9/10的原失败保留,新增状态开销预留后10/10,未扩大预算或删除原日志。 -4. [pending] 最终来源冻结、相关回归/实际界面与工程流、固定发行和保留状态部署;小步本地Git及阶段回顾。真实请求仅在完成部署及身份预检合同后有界执行,遇不一致立即停,不跑旧Task。 +4. [in_progress] 实际Harness验证发现整批审查关闭被错记失败,专项已通过,正在核对完整引擎回归及Chat/Responses真实Harness;随后最终来源冻结、相关回归/实际界面与工程流、固定发行和保留状态部署。真实请求仅在完成部署及身份预检合同后有界执行,遇不一致立即停,不跑旧Task。 ## 当前优先 H:1.24.1 真实交付与原版 Purge 重新实测 diff --git a/tests/engine-worker.test.mjs b/tests/engine-worker.test.mjs index 0db1604b..d6e24aff 100644 --- a/tests/engine-worker.test.mjs +++ b/tests/engine-worker.test.mjs @@ -18,6 +18,47 @@ async function prepare(f, context) { } const jobs = (f, group) => [...f.store.core.engine_run.values()].filter(job => job.collaborationId === group.id); +test('a sealed whole-batch review pauses the owned native session without certifying engine completion or losing the saved host result',async t=>{ + const f=fixture(t);let requests=0; + const context=controlledWorker(f,{respond:({project})=>{ + requests++;const session=[...f.store.sessions.values()].find(value=>value.projectId===project.id); + const calls=requests===1?[call('proposal','propose_file',{path:'src/a.mjs',content:'export const value=1;\n',reason:'owned review'})] + :[call('review','prepare_change_set',{change_ids:session.changes.filter(value=>value.status==='pending').map(value=>value.id)})]; + return {message:{role:'assistant',content:'',tool_calls:calls},usage}; + }}),group=await prepare(f,context); + await(await context.manager.start(f.session,group.id,group.fingerprint)).promise; + const session=f.store.session(group.children[0].sessionId),job=jobs(f,group)[0]; + assert.equal(requests,2);assert.equal(session.status,'needs_review');assert.equal(group.status,'awaiting_review'); + assert.equal(job.status,'paused');assert.equal(job.verified,false);assert.equal(job.requiresVerification,false);assert.equal(job.error,null); + assert.equal(job.native.cleanup.rootExited,true);assert.equal(job.reviewHandoff.tools.length,1); + const result=job.reviewHandoff.tools[0];assert.equal(result.resultArtifactId,f.store.core.tool.get(result.invocationId).resultArtifactId); + assert.equal(f.store.savedToolResult(session.id,result.callId).change_set_id,job.reviewHandoff.changeSetId); + assert.ok(job.calls.some(value=>value.status!=='returned'),'saved host result is not claimed as delivered to the native engine'); + originalEngineRun(f.store,session,job); +}); + +test('a changed review seal still closes the owned native process and cannot claim a verified review handoff',async t=>{ + const f=fixture(t);let requests=0,changed=false; + const save=f.store.save.bind(f.store);f.store.save=(session,event)=>{ + const result=save(session,event); + if(event?.type==='change_set.awaiting_review'){ + const set=f.store.core.change_set.get(event.detail.changeSetId);set.fingerprint='changed-after-review-ready';changed=true; + } + return result; + }; + const context=controlledWorker(f,{respond:({project})=>{ + requests++;const session=[...f.store.sessions.values()].find(value=>value.projectId===project.id); + const calls=requests===1?[call('proposal','propose_file',{path:'src/a.mjs',content:'export const value=1;\n',reason:'owned review'})] + :[call('review','prepare_change_set',{change_ids:session.changes.filter(value=>value.status==='pending').map(value=>value.id)})]; + return {message:{role:'assistant',content:'',tool_calls:calls},usage}; + }}),group=await prepare(f,context); + await(await context.manager.start(f.session,group.id,group.fingerprint)).promise; + const session=f.store.session(group.children[0].sessionId),job=jobs(f,group)[0]; + assert.equal(changed,true);assert.equal(requests,2);assert.equal(job.status,'failed');assert.equal(job.verified,false); + assert.equal(job.reviewHandoff,undefined);assert.equal(job.native.cleanup.rootExited,true); + assert.equal(context.runtime.controlledEngines.isActive(session.id),false);originalEngineRun(f.store,session,job); +}); + test('native text and call IDs preserve original result pairing while unsupported content fails closed', () => { const input = { system: 'fixed', messages: [{ role: 'assistant', content: [{ type: 'reasoning', text: 'private' }, { type: 'tool-call', id: 'native', name: 'read_file', arguments: '{"path":"README.md"}' }] }, { role: 'tool', toolCallId: 'native', content: [{ type: 'text', text: 'result' }] }] };