niceeval 0.12.2-canary.31 → 0.13.1

package/dist/runner/attempt.cjs

@@ -12,6 +12,8 @@ exports.experimentRunInfo = experimentRunInfo;
 const effect_1 = require("effect");
 const runtime_ts_1 = require("../sandbox/runtime.cjs");
 const registry_ts_1 = require("../sandbox/registry.cjs");
+const paths_ts_1 = require("../sandbox/paths.cjs");
+const request_executor_ts_1 = require("../sandbox/request-executor.cjs");
 const cleanup_timeout_ts_1 = require("./cleanup-timeout.cjs");
 const timeout_ts_1 = require("./timeout.cjs");
 const deadline_ts_1 = require("../sandbox/deadline.cjs");
@@ -46,10 +48,10 @@ const backend_ts_1 = require("../sandbox/backend.cjs");
 const redaction_ts_1 = require("../sandbox/redaction.cjs");
 const link_ts_1 = require("../sandbox/link.cjs");
 const config_identity_ts_1 = require("./config-identity.cjs");
-const facts_ts_1 = require("../shared/facts.cjs");
 const contracts_ts_1 = require("../plugin/contracts.cjs");
 const source_loc_ts_1 = require("../source-loc.cjs");
 const source_producer_ts_1 = require("./source-producer.cjs");
+const sandbox_record_producer_ts_1 = require("./sandbox-record-producer.cjs");
 const runner_producer_ts_1 = require("../o11y/record/runner-producer.cjs");
 const failure_class_ts_2 = require("../shared/failure-class.cjs");
 const sink_ts_1 = require("./feedback/sink.cjs");
@@ -75,7 +77,7 @@ function attemptFailureDeclaration(classify, phase, e) {
 function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, parentSignal, invocationSignal, onPhase, concurrencySlot, onFailureClass, onSealedEvaluation, reusedSandbox, }) {
     const config = opts.config;
     const { evalDef, run, attempt } = a;
-    const niceevalRoot = opts.niceevalRoot ?? `${process.cwd()}/.niceeval`;
+    const coordinationRoot = opts.coordinationRoot ?? `${process.cwd()}/.niceeval`;
     const t0 = Date.now();
     // Source bytes are snapshotted at the author action, then held only through
     // an internal capability until the Record producer seals its own Attachments.
@@ -176,6 +178,8 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
     // Effect.timeoutTo 调用点的注释)。
     let timedOut = false;
     let timeoutDiff;
+    let fileChangesCapture;
+    let timeoutFileChangesCapture;
     let timeoutSources;
     const enterPhase = (phase) => {
         lastPhase = phase;
@@ -206,15 +210,11 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
             });
         }
     };
-    // 本 attempt 累计的运行事实(与 verdict/diagnostics 独立):layer prepare/cleanup 与
-    // agent setup·send·teardown 经 ctx.fact() 上报的都落这里(同一 attempt 内后写覆盖先写),
-    // 收尾时并入结果的 facts 字段(见 finally 末尾,与 diagnostics 同一种「累加器 + finally 并入」模式)。
-    const facts = {};
     // 本 attempt 累计的诊断(与 verdict 独立):ScopedFeedback.diagnostic 与 teardown 失败都落这里,
     // 收尾时并入结果;dedupeKey 相同的并发诊断折叠成一条并累计 count。
     const diagnostics = [];
     const dedupeIndex = new Map();
-    // Sandbox 命令证据的累加器(与 diagnostics/facts 同一种「共享容器」模式):
+    // Sandbox 命令证据的累加器(与 diagnostics 同一种「共享容器」模式):
     // withCommandTiming 在 CommandResult 交还调用方之前把每条命令(成功与非零退出都算)push 进来,
     // runAttemptBody 的 finally 与本函数的超时/中断兜底分支都读同一个数组引用
     // (见 docs/feature/record/architecture.md「commandsjson」「证据在 CommandResult 返回
@@ -226,7 +226,7 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
     // Package-internal only: the final EvalResult is weakly associated with this
     // capture after all Runner evidence is sealed. No public result field or
     // legacy artifact is used as a transport.
-    const observabilityRuntime = (0, runner_producer_ts_1.createRunnerAttemptObservabilityRuntimeV1)({
+    const observabilityRuntime = (0, runner_producer_ts_1.createRunnerAttemptObservabilityRuntime)({
         providerName: run.agent.name,
         sensitiveValues,
     });
@@ -394,8 +394,8 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
         // 退避重试(runtime.ts → retry.ts)期间临时归还这个名额:被限流的 provider 只是在
         // setTimeout 里睡觉,不该攥着 sandboxSem 的槽位陪跑,不然一批 429 能把整体并发拖成个位数。
         const provisionSlot = {
-            release: () => assertFirst.requestEffect(sandboxSem.release(1)).then(() => { }),
-            reacquire: () => assertFirst.requestEffect(sandboxSem.take(1)).then(() => { }),
+            release: sandboxSem.release(1),
+            reacquire: sandboxSem.take(1),
         };
         // Sample release(receiver close + provider stop)整段计成 sandbox.stop:先加的 finalizer
         // 后跑(LIFO),所以「先加的」在 release 链末尾打终点戳、「后加的」在 release 开始前打起点戳;
@@ -441,7 +441,6 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
                                 signal,
                                 progress: scopedFeedback.progress,
                                 diagnostic: scopedFeedback.diagnostic,
-                                fact: (key, value) => (0, facts_ts_1.recordFact)(facts, key, value),
                             },
                             buildLocators,
                             agent: run.agent.kind === "sandbox" ? run.agent : undefined,
@@ -462,7 +461,7 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
                                     const suspendStart = recorder.offsetNow();
                                     return yield* tryPromiseAsDefect(() => (0, keep_ts_1.suspendSandbox)(sb)).pipe(effect_1.Effect.tap(() => effect_1.Effect.sync(() => {
                                         recorder.record("sandbox.suspend", recorder.offsetNow() - suspendStart);
-                                    })), effect_1.Effect.zipRight((0, keep_registry_ts_1.updateKeptEntryEffect)(niceevalRoot, (0, keep_registry_ts_1.keptEntryId)(providerName, sb.sandboxId), {
+                                    })), effect_1.Effect.zipRight((0, keep_registry_ts_1.updateKeptEntryEffect)(coordinationRoot, (0, keep_registry_ts_1.keptEntryId)(providerName, sb.sandboxId), {
                                         state: "dormant",
                                     }).pipe(effect_1.Effect.ignore)), effect_1.Effect.catchAllCause((cause) => effect_1.Cause.isInterrupted(cause)
                                         ? effect_1.Effect.failCause(cause)
@@ -479,7 +478,11 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
                                 }),
                             },
                         });
-                        return materialized.sandbox;
+                        // fresh Attempt:author facade 在当前 Attempt Scope 内构建。executor 捕获本
+                        // Scope 的 Runtime,Scope 关闭时统一中断在飞请求并拒绝新调用;release closure
+                        // 继续用 materialized.sandbox(资源面 facade),两者生命周期互不借用。
+                        const executor = yield* (0, request_executor_ts_1.makeSandboxRequestExecutor)();
+                        return (0, paths_ts_1.makeSandboxAuthorFacade)(materialized.authorBackend, executor, runtimeCapabilities.provider);
                     }));
                 })
                 : (0, direct_agent_sandbox_ts_1.createDirectAgentSandbox)());
@@ -596,8 +599,49 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
             if (run.agent.kind === "sandbox" && liveLedger) {
                 const startedAt = recorder.offsetNow();
                 const diff = yield* effect_1.Effect.either((0, cleanup_timeout_ts_1.cleanupCallback)(() => liveLedger.exportWindows()));
-                if (effect_1.Either.isRight(diff))
+                if (effect_1.Either.isRight(diff)) {
                     timeoutDiff = diff.right;
+                    let document;
+                    try {
+                        document = (0, workspace_diff_ts_1.createAgentWorkspaceDiff)({
+                            windows: diff.right,
+                            policy: liveLedger.attribution,
+                        });
+                    }
+                    catch {
+                        timeoutFileChangesCapture = (0, diff_ts_2.createEmptyFileChangesCapture)({
+                            policy: liveLedger.attribution,
+                            limitations: [Object.freeze({
+                                    code: "capture-failed",
+                                    stage: "normalize",
+                                    atWindowId: null,
+                                })],
+                        });
+                    }
+                    if (document !== undefined) {
+                        timeoutFileChangesCapture = (0, diff_ts_2.createFileChangesCapture)({
+                            document,
+                            limitations: [Object.freeze({
+                                    code: "capture-interrupted",
+                                    stage: "finalizer-export",
+                                    atWindowId: null,
+                                })],
+                        });
+                    }
+                }
+                else {
+                    timeoutFileChangesCapture = (0, diff_ts_2.createEmptyFileChangesCapture)({
+                        policy: liveLedger.attribution,
+                        limitations: [Object.freeze({
+                                code: "capture-interrupted",
+                                stage: "finalizer-export",
+                                atWindowId: null,
+                            })],
+                    });
+                }
+                if (timeoutFileChangesCapture !== undefined) {
+                    (0, diff_ts_2.assertFileChangesCaptureRecordable)(timeoutFileChangesCapture);
+                }
                 // 沙箱不可用 / 导出挂起到点:如实缺失,不阻塞收尾。
                 recorder.record("workspace.diff", recorder.offsetNow() - startedAt);
             }
@@ -661,10 +705,9 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
                         if (node)
                             recorder.pushParent(node);
                         try {
-                            const cleanupSignal = AbortSignal.timeout(cleanup_timeout_ts_1.CLEANUP_TIMEOUT_MS);
-                            await assertFirst.requestEffect((0, cleanup_timeout_ts_1.cleanupCallback)(() => cleanup.command(commandTarget, {
+                            await assertFirst.requestEffect((0, cleanup_timeout_ts_1.cleanupCallback)((signal) => cleanup.command(commandTarget, {
                                 ...cleanup.context,
-                                signal: cleanupSignal,
+                                signal,
                             })));
                         }
                         catch (error) {
@@ -709,7 +752,6 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
             ...(deadlineAt !== undefined ? { deadlineAt } : {}),
             feedback: scopedFeedback,
             diagnostics,
-            facts,
             commands,
             sensitiveValues,
             observabilityRuntime,
@@ -725,6 +767,9 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
             registerLedger: (ledger) => {
                 liveLedger = ledger;
             },
+            retainFileChangesCapture: (capture) => {
+                fileChangesCapture = capture;
+            },
             assertFirst,
             registerAssertions: (runtime, state) => {
                 if (liveAssertions !== undefined && liveAssertions !== runtime) {
@@ -815,7 +860,7 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
                 const enter = (0, keep_ts_1.nativeEnterCommand)(providerName, sandbox.sandboxId);
                 const keptAt = new Date().toISOString();
                 const expiresAt = (0, keep_ts_1.computeExpiresAt)(providerName, keptAt);
-                const registered = yield* effect_1.Effect.either((0, keep_registry_ts_1.writeKeptEntryEffect)(niceevalRoot, {
+                const registered = yield* effect_1.Effect.either((0, keep_registry_ts_1.writeKeptEntryEffect)(coordinationRoot, {
                     sandboxId: sandbox.sandboxId,
                     provider: providerName,
                     evalId: evalDef.id,
@@ -902,7 +947,6 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
         const withEvidence = {
             ...r,
             ...(diagnostics.length > 0 ? { diagnostics: [...diagnostics] } : {}),
-            ...(Object.keys(facts).length > 0 ? { facts: { ...facts } } : {}),
             ...(commands.length > 0 ? { commands: [...commands] } : {}),
         };
         const phases = recorder.finalize();
@@ -925,7 +969,13 @@ function runAttemptEffect(a, opts, sandboxSem, { buildLocators, runTiming, paren
             }
             : withSandbox;
         const finalResult = (0, source_producer_ts_1.retainRunnerAttemptSourceCapture)(redactEvalResultEvidence(completed, sensitiveValues), sourceCapture.snapshot());
-        (0, runner_producer_ts_1.bindRunnerAttemptObservabilityCaptureV1)(finalResult, observabilityRuntime);
+        const capture = timedOut
+            ? timeoutFileChangesCapture ?? fileChangesCapture
+            : fileChangesCapture;
+        if (capture !== undefined) {
+            (0, sandbox_record_producer_ts_1.retainRunnerAttemptFileChangesCapture)(finalResult, capture);
+        }
+        (0, runner_producer_ts_1.bindRunnerAttemptObservabilityCapture)(finalResult, observabilityRuntime);
         return finalResult;
     }));
 }
@@ -1115,10 +1165,7 @@ function assertDeadlineFitsPlan(capabilities, deadlineMs) {
 // 资源已由 runAttemptEffect 的 Sample 持有;这里只在 finally 跑 agent 自己的 cleanup/teardown。
 async function runAttemptBody(a, config, t0, base, res, parentSignal) {
     const { evalDef, run, attempt } = a;
-    const { sandbox: rawSandbox, sourceRegistry, sourceCapture, receiver, telemetry, otel, signal, log, enterPhase, getPhase, setSendActive, recorder, feedback, diagnostics, facts, commands, sensitiveValues, observabilityRuntime, concurrencySlot, declareFailure, isDeadlineTimedOut, layerCleanups, registerEvidence, registerLedger, assertFirst, registerAssertions, prepareCoordinator, } = res;
-    // ctx.fact() 闭包:校验后写进 res.facts(与 runAttemptEffect 共享的同一个 Record 引用),
-    // finally 把它原样挂到即将返回的结果上(与 diagnostics 同一种「共享容器 + finally 并入」模式)。
-    const fact = (key, value) => (0, facts_ts_1.recordFact)(facts, key, value);
+    const { sandbox: rawSandbox, sourceRegistry, sourceCapture, receiver, telemetry, otel, signal, log, enterPhase, getPhase, setSendActive, recorder, feedback, diagnostics, commands, sensitiveValues, observabilityRuntime, concurrencySlot, declareFailure, isDeadlineTimedOut, layerCleanups, registerEvidence, registerLedger, retainFileChangesCapture, assertFirst, registerAssertions, prepareCoordinator, } = res;
     const usesSandbox = run.agent.kind === "sandbox";
     // 命令时间树:所有经这个包装 sandbox 发出的 runCommand/runShell 都挂成当前阶段(或当前 hook
     // 节点)下的 command 子节点。包装只在最外层公开调用记录一次——provider 内部转调不经过它。
@@ -1142,7 +1189,6 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
         telemetry,
         progress: feedback.progress,
         diagnostic: feedback.diagnostic,
-        fact,
         // log 是 progress({ message }) 的别名,不是第二条通道(见 AgentContext.log 注释)。
         log,
     };
@@ -1160,7 +1206,6 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
         signal,
         progress: feedback.progress,
         diagnostic: feedback.diagnostic,
-        fact,
     };
     // Only the Assert-first runtime is registered with the outer Scope. The
     // body requests its single seal after gathering non-authoring evidence.
@@ -1205,7 +1250,6 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
                         signal,
                         progress: feedback.progress,
                         diagnostic: feedback.diagnostic,
-                        facts: fact,
                     };
                     const context = {
                         ...cleanupContext,
@@ -1249,7 +1293,6 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
                 if (effect_1.Either.isLeft(verified))
                     throw verified.left;
                 const ensureEffect = (0, provisioner_ts_1.runAgentEnsure)(run.agent.ensure, run.agent.installers, sandbox, {
-                    fact,
                     coordinator: effect_1.Option.fromNullable(prepareCoordinator),
                     targetPlatform: a.plan.providerPlan.target.platform,
                     signal,
@@ -1327,6 +1370,8 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
                 } }),
             model: run.model,
             reasoningEffort: run.reasoningEffort,
+            // maxCost 断言的估算价目表;与 observed usage.costUSD 互不兜底。
+            pricing: config.pricing,
             flags: run.flags,
             experimentId: run.experimentId,
             judge: a.judge,
@@ -1339,7 +1384,6 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
             telemetry,
             otel,
             feedback,
-            fact,
             concurrencySlot,
             // 实验分类器:send 链上排在 adapter 的 classifySendFailure 之前(见
             // docs/feature/error-classification/architecture.md「分类链」)。与本文件 declareFailure
@@ -1365,9 +1409,14 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
             // 才带(show `--execution`/`--timing` 的 turn 头行读 TimingNode.usage,见 docs/feature/
             // results/architecture.md「result.json」TimingNode.usage)。
             onTurn: (info) => {
-                sourceCapture.onTurn(info);
+                const turnId = (0, runner_producer_ts_1.captureRunnerPhysicalConversationTurn)({
+                    runtime: observabilityRuntime,
+                    outcome: info.outcome,
+                    events: info.events,
+                });
+                sourceCapture.onTurn({ ...info, ...(turnId === undefined ? {} : { turnId }) });
                 if (info.usage !== undefined) {
-                    (0, runner_producer_ts_1.captureRunnerTurnUsageV1)(observabilityRuntime, info.usage);
+                    (0, runner_producer_ts_1.captureRunnerTurnUsage)(observabilityRuntime, info.usage);
                 }
                 return recorder.child((0, timing_ts_1.turnActivity)({
                     label: info.label,
@@ -1376,6 +1425,7 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
                     ...(info.failed ? { failed: true } : {}),
                     sessionIndex: info.sessionIndex,
                     turnIndex: info.turnIndex,
+                    ...(turnId === undefined ? {} : { turnId }),
                     ...(info.traceId !== undefined ? { traceId: info.traceId } : {}),
                     ...(info.traceAttribution !== undefined ? { traceAttribution: info.traceAttribution } : {}),
                     ...(info.usage !== undefined ? { usage: info.usage } : {}),
@@ -1417,6 +1467,7 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
             enterPhase("workspace.diff");
         let diffWindows = [];
         let diffArtifactAvailable = true;
+        let diffExportAvailable = true;
         const captureFailure = state.manager.ledgerCaptureFailure;
         if (captureFailure !== undefined) {
             // `afterSend` is intentionally non-fatal to the completed agent Turn,
@@ -1430,7 +1481,7 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
                 data: evidenceDiagnosticData("capture", captureFailure.error, a.plan._tag === "Sandbox" ? a.plan.providerPlan.provider : undefined),
             });
         }
-        if (!skipReason && usesSandbox && ledger && captureFailure === undefined) {
+        if (usesSandbox && ledger) {
             const startedAt = recorder.offsetNow();
             const operation = recorder.child((0, timing_ts_1.workspaceDiffExportActivity)({
                 label: "export workspace diff",
@@ -1452,6 +1503,7 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
                 if (isBodyInterrupted(signal, parentSignal, isDeadlineTimedOut()))
                     throw diffError;
                 diffArtifactAvailable = false;
+                diffExportAvailable = false;
                 feedback.diagnostic({
                     code: "workspace-diff-unavailable",
                     level: "warning",
@@ -1468,39 +1520,71 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
         }
         const diff = (0, diff_ts_1.deriveDiffData)(diffWindows);
         let workspaceDiff;
-        if (usesSandbox && ledger && diffArtifactAvailable && !skipReason) {
-            // This is the one export/freeze point. Post-run Assertion closures and
-            // the later Evaluation Record adapter share this exact object; neither
-            // side rebuilds a diff.
-            try {
-                const document = (0, workspace_diff_ts_1.createAgentWorkspaceDiff)({
-                    windows: diffWindows,
+        if (usesSandbox && ledger) {
+            let document;
+            if (diffExportAvailable) {
+                try {
+                    document = (0, workspace_diff_ts_1.createAgentWorkspaceDiff)({
+                        windows: diffWindows,
+                        policy: ledger.attribution,
+                    });
+                }
+                catch (diffError) {
+                    diffArtifactAvailable = false;
+                    feedback.diagnostic({
+                        code: "workspace-diff-unavailable",
+                        level: "warning",
+                        message: `workspace diff freeze failed; continuing with command/event evidence: ${(0, util_ts_1.firstLine)((0, util_ts_1.formatThrown)(diffError))}`,
+                        data: evidenceDiagnosticData("freeze", diffError, a.plan._tag === "Sandbox" ? a.plan.providerPlan.provider : undefined),
+                    });
+                }
+            }
+            const limitations = [
+                ...(captureFailure === undefined ? [] : [Object.freeze({
+                        code: "capture-failed",
+                        stage: "checkpoint",
+                        atWindowId: captureFailure.label,
+                    })]),
+                ...(!diffExportAvailable ? [Object.freeze({
+                        code: "capture-failed",
+                        stage: "export",
+                        atWindowId: null,
+                    })] : []),
+            ];
+            if (document === undefined) {
+                const capture = (0, diff_ts_2.createEmptyFileChangesCapture)({
                     policy: ledger.attribution,
+                    limitations: [
+                        ...limitations,
+                        ...(diffExportAvailable ? [Object.freeze({
+                                code: "capture-failed",
+                                stage: "normalize",
+                                atWindowId: null,
+                            })] : []),
+                    ],
                 });
-                // Diff is supplemental unless an Assertion declared it required. Check
-                // the exact durable v1 representation here, while the shared late
-                // state can still become unavailable. Discovering the Record JSON
-                // limit only during publication would abort the whole Run after all
-                // Attempts had already settled.
-                (0, diff_ts_2.assertAgentWorkspaceDiffRecordableV1)(document);
+                (0, diff_ts_2.assertFileChangesCaptureRecordable)(capture);
+                retainFileChangesCapture(capture);
+            }
+            else {
+                const capture = (0, diff_ts_2.createFileChangesCapture)({ document, limitations });
+                (0, diff_ts_2.assertFileChangesCaptureRecordable)(capture);
+                retainFileChangesCapture(capture);
+            }
+            // A malformed family payload or its blob closure is a producer defect,
+            // not a collector limitation. Keep this outside the collector catches.
+            if (!skipReason && diffArtifactAvailable && captureFailure === undefined && document !== undefined) {
                 state.late.diff = Object.freeze({
                     state: "available",
                     document,
                 });
                 workspaceDiff = document;
             }
-            catch (diffError) {
-                diffArtifactAvailable = false;
+            else {
                 state.late.diff = Object.freeze({
                     state: "unavailable",
                     reason: "producer-failed",
                 });
-                feedback.diagnostic({
-                    code: "workspace-diff-unavailable",
-                    level: "warning",
-                    message: `workspace diff freeze failed; continuing with command/event evidence: ${(0, util_ts_1.firstLine)((0, util_ts_1.formatThrown)(diffError))}`,
-                    data: evidenceDiagnosticData("freeze", diffError, a.plan._tag === "Sandbox" ? a.plan.providerPlan.provider : undefined),
-                });
             }
         }
         else {
@@ -1600,9 +1684,11 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
         recorder.closeCurrent();
         const durationMs = recorder.offsetNow();
         const o11y = (0, derive_ts_1.buildO11ySummary)(events);
-        // 实测成本(网关带回)优先,缺则按 model + 用量查价格表估算(见 o11y/cost.ts)。
+        // 单向字段契约:estimatedCostUSD 恒为 estimateCost(run.model, usage, config.pricing)
+        // 的 runtime/config price table 估算,永远独立计算;observed 成本(网关/adapter 显式回报)
+        // 只留在 usage.costUSD,两者互不覆盖、互不兜底——即使 observed 存在也照常估算。
         // 权威唯一在 result.json 的 estimatedCostUSD;o11y.json 只留行为计数(见 docs/feature/record/architecture.md「o11y.json」)。
-        const cost = usage.costUSD ?? (0, cost_ts_1.estimateCost)(run.model, usage, config.pricing);
+        const estimatedCostUSD = (0, cost_ts_1.estimateCost)(run.model, usage, config.pricing);
         // Assert-first entries carry no legacy source graph. Session events and the
         // captured entry module still provide the stable source artifact surface.
         const sources = await collectSources(evalDef.source, sourceRegistry);
@@ -1625,7 +1711,7 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
             evaluationKind: evalDef.evaluationKind ?? "pass",
             ...(state.manager.retryAttempts.length > 0 ? { retryAttempts: state.manager.retryAttempts } : {}),
             usage,
-            estimatedCostUSD: cost,
+            estimatedCostUSD,
             error,
             skipReason,
             events,
@@ -1698,13 +1784,21 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
                     // AgentTeardown | DirectAgentTeardown 混成无法调用的签名。
                     if (run.agent.kind === "sandbox") {
                         const teardown = run.agent.teardown;
-                        if (teardown)
-                            await assertFirst.requestEffect((0, cleanup_timeout_ts_1.cleanupCallback)(() => teardown(sandbox, sandboxAttemptCtx)));
+                        if (teardown) {
+                            await assertFirst.requestEffect((0, cleanup_timeout_ts_1.cleanupCallback)((signal) => teardown(sandbox, {
+                                ...sandboxAttemptCtx,
+                                signal,
+                            })));
+                        }
                     }
                     else {
                         const teardown = run.agent.teardown;
-                        if (teardown)
-                            await assertFirst.requestEffect((0, cleanup_timeout_ts_1.cleanupCallback)(() => teardown(attemptCtx)));
+                        if (teardown) {
+                            await assertFirst.requestEffect((0, cleanup_timeout_ts_1.cleanupCallback)((signal) => teardown({
+                                ...attemptCtx,
+                                signal,
+                            })));
+                        }
                     }
                 }
                 catch (e) {
@@ -1719,9 +1813,9 @@ async function runAttemptBody(a, config, t0, base, res, parentSignal) {
             if (lifecycle.teardown === undefined)
                 continue;
             try {
-                await effect_1.Effect.runPromise((0, cleanup_timeout_ts_1.cleanupCallback)(() => lifecycle.teardown({
+                await assertFirst.requestEffect((0, cleanup_timeout_ts_1.cleanupCallback)((signal) => lifecycle.teardown({
                     ...evalPluginContext,
-                    signal: AbortSignal.timeout(cleanup_timeout_ts_1.CLEANUP_TIMEOUT_MS),
+                    signal,
                 })));
             }
             catch (error) {
@@ -1832,7 +1926,6 @@ const EVAL_RESULT_REDACTED_EVIDENCE_KEYS = [
     "retryAttempts",
     "error",
     "diagnostics",
-    "facts",
     "phases",
     "skipReason",
     "events",
@@ -1875,8 +1968,16 @@ function redactEvalResultEvidence(result, sensitiveValues) {
  * (docs/feature/record/architecture.md「commandsjson」);登记不改变 `runCommand` 的
  * 返回/抛错语义,调用方可以处理非零退出并继续,证据仍保留。checked 方法抛出的
  * `SandboxCommandExitError` 自带 `CommandResult`,同样登记；没有结果的 transport 错误只落
- * 时间树 `failed` 标记。
+ * 时间树 `failed` 标记，同时保留 manifest 和准确的非正常终态。
  */
+function isCommandSpawnFailure(error) {
+    if (typeof error !== "object" || error === null)
+        return false;
+    const candidate = error;
+    return ((candidate.code === "ENOENT" || candidate.code === "EACCES") &&
+        typeof candidate.syscall === "string" &&
+        candidate.syscall.startsWith("spawn"));
+}
 function withCommandTiming(sandbox, recorder, getPhase, commands, sensitiveValues, deadlineAt, observabilityRuntime, attemptSignal) {
     const recordCommandExit = (node, display, result, checked) => {
         if (!node)
@@ -1896,7 +1997,7 @@ function withCommandTiming(sandbox, recorder, getPhase, commands, sensitiveValue
         const display = (0, timing_ts_1.commandDisplay)(cmd, args, sensitiveValues);
         // Register the command manifest capability before process launch. The
         // later result can only be attached through this opaque handle.
-        const observabilityCommand = (0, runner_producer_ts_1.captureRunnerCommandStartV1)({
+        const observabilityCommand = (0, runner_producer_ts_1.captureRunnerCommandStart)({
             runtime: observabilityRuntime,
             phase: getPhase() ?? "eval.run",
             invocationKind,
@@ -1935,7 +2036,7 @@ function withCommandTiming(sandbox, recorder, getPhase, commands, sensitiveValue
                     exitCode,
                     ...result,
                 }, checked);
-                (0, runner_producer_ts_1.captureRunnerCommandResultV1)({
+                (0, runner_producer_ts_1.captureRunnerCommandResult)({
                     handle: observabilityCommand,
                     exitCode,
                     stdout: result.stdout,
@@ -1943,7 +2044,7 @@ function withCommandTiming(sandbox, recorder, getPhase, commands, sensitiveValue
                 });
             }
             else {
-                (0, runner_producer_ts_1.captureRunnerCommandCaptureFailedV1)(observabilityCommand);
+                (0, runner_producer_ts_1.captureRunnerCommandCaptureFailed)(observabilityCommand);
             }
             // CommandResult.command:最外层公开调用恰好是「eval 实际跑了什么」的定义点,摘要
             // 与时间树节点同一份。记录边界是权威来源，provider 即使给过原始 command 也必须覆盖，
@@ -1980,7 +2081,7 @@ function withCommandTiming(sandbox, recorder, getPhase, commands, sensitiveValue
             }));
             if (commandResult !== undefined) {
                 recordCommandExit(node, display, commandResult, checked);
-                (0, runner_producer_ts_1.captureRunnerCommandResultV1)({
+                (0, runner_producer_ts_1.captureRunnerCommandResult)({
                     handle: observabilityCommand,
                     exitCode: commandResult.exitCode,
                     stdout: commandResult.stdout,
@@ -1988,16 +2089,16 @@ function withCommandTiming(sandbox, recorder, getPhase, commands, sensitiveValue
                 });
             }
             else if (e instanceof deadline_ts_1.SandboxCommandTimeoutError) {
-                (0, runner_producer_ts_1.captureRunnerCommandTimeoutV1)(observabilityCommand);
+                (0, runner_producer_ts_1.captureRunnerCommandTimeout)(observabilityCommand);
             }
             else if (attemptSignal.aborted ||
                 (opts?.signal?.aborted === true)) {
-                // Cancellation has no observed terminal result; preserve the partial
-                // capture state rather than inventing one.
-                (0, runner_producer_ts_1.captureRunnerCommandInterruptedV1)(observabilityCommand);
+                // The provider call was already entered, so cancellation terminates
+                // this registered command rather than claiming it never started.
+                (0, runner_producer_ts_1.captureRunnerCommandInterrupted)(observabilityCommand);
             }
             else {
-                (0, runner_producer_ts_1.captureRunnerCommandCaptureFailedV1)(observabilityCommand);
+                (0, runner_producer_ts_1.captureRunnerCommandCaptureFailed)(observabilityCommand, isCommandSpawnFailure(e) ? "spawn-failed" : "transport-lost");
             }
             throw e;
         }