niceeval 0.17.1 → 0.19.0

package/dist/runner/run.cjs

@@ -2,7 +2,7 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.RunModeConflictError = void 0;
 exports.runEvals = runEvals;
-// @concord-file ne-runner-dispatch
+// @concord-file
 // @concord-implements docs/feature/experiments/architecture.md
 // @concord-implements docs/feature/run/lifecycle.md
 // 运行器主调度:发现产出的 eval × agent × runs → attempt,有界并发调度。
@@ -33,7 +33,7 @@ const cleanup_timeout_ts_1 = require("./cleanup-timeout.cjs");
 const timeout_ts_1 = require("./timeout.cjs");
 const node_os_1 = require("node:os");
 const root_ts_1 = require("../record/platform/root.cjs");
-const contracts_ts_1 = require("../plugin/contracts.cjs");
+const link_ts_1 = require("../plugin/link.cjs");
 const teardown_registry_ts_1 = require("./teardown-registry.cjs");
 const lock_ts_1 = require("./lock.cjs");
 const shared_state_lease_ts_1 = require("./shared-state-lease.cjs");
@@ -189,7 +189,7 @@ function runEvals(opts) {
         // only this projection and its own frozen view; no historic result facade
         // participates in dispatch or display.
         const targetPlan = yield* (0, fingerprint_ts_1.planProjectTarget)(opts.evals, opts.agentRuns, opts.config.timeoutMs, {
-            configJudge: opts.config.judgeRuntime,
+            configJudge: opts.config.judgeProvider,
             ...(opts.keepSandbox === undefined ? {} : { keepSandbox: opts.keepSandbox }),
         });
         const { preparedPairsByKey, plannedConfigHashes, resolvedJudgesByKey, plannedFingerprints, renameFingerprintsByKey, } = targetPlan;
@@ -418,7 +418,7 @@ function runEvals(opts) {
         // prepare VM is released before any ordinary Attempt can acquire a VM.
         const setupPrefixPreparation = yield* (0, setup_prefix_preparation_ts_1.prepareSetupPrefixes)(attempts, {
             signal: opts.signal,
-            maxConcurrency: opts.maxSetupPrefixConcurrency ?? 2,
+            maxConcurrency: opts.maxBeforeCacheConcurrency ?? 2,
             onActivity: (event) => {
                 (0, sink_ts_1.reportRunActivity)({
                     id: event.id,
@@ -575,42 +575,7 @@ function runEvals(opts) {
             maxConcurrency: opts.maxConcurrency,
             runIds,
         };
-        // eval 级 reporters:实例只观测引用它的 eval(经 scopeReporter 过滤转发)。
-        // 已经挂在全局 reporters 里的同一实例不重复挂;同一实例被多个 eval 引用时合并观测集
-        // (共享一个目的地,如同一个 Braintrust 实验)。本次没有任何被观测 eval 要跑时整个跳过。
-        const scopedSets = new Map();
-        for (const e of opts.evals) {
-            for (const r of e.reporters ?? []) {
-                if (opts.reporters.some((reg) => reg.reporter === r))
-                    continue;
-                let ids = scopedSets.get(r);
-                if (!ids)
-                    scopedSets.set(r, (ids = new Set()));
-                ids.add(e.id);
-            }
-        }
-        const reporters = [...opts.reporters];
-        // EvalDef.reporters 是用户在单个 eval 上挂的补充观测(如「这个 eval 单独也发一份到某个
-        // dashboard」),不是 CLI 显式注册的默认/机器出口——与 Config.reporters 同样默认
-        // best-effort(见 ReporterRegistration 的字段注释:required 只留给 --junit)。name 用「scope 内第几个」编号,足以在诊断里区分「哪一个 eval 级 reporter」,
-        // 不需要用户自己起名字。
-        let evalReporterIndex = 0;
-        for (const [r, ids] of scopedSets) {
-            const scopedRuns = attempts.filter((a) => ids.has(a.evalDef.id)).length;
-            if (scopedRuns === 0)
-                continue;
-            reporters.push({
-                reporter: (0, report_ts_1.scopeReporter)(r, ids, {
-                    evals: [...ids].filter((id) => runningIds.has(id)).length,
-                    configs: effectiveAgentRuns.length,
-                    totalAttempts: scopedRuns,
-                    maxConcurrency: opts.maxConcurrency,
-                    runIds,
-                }),
-                name: `eval-reporter-${evalReporterIndex++}`,
-                required: false,
-            });
-        }
+        const reporters = opts.reporters;
         for (const reg of reporters) {
             // reporter 只是结果消费方:单个 reporter 抛错记 diagnostic,不能让整次调度崩,也不阻断
             // 其它 reporter 的必要收尾(required/best-effort 的判定权重在 runReporter 内部处理)。
@@ -741,7 +706,7 @@ function runEvals(opts) {
                 pool = new sandbox_pool_ts_1.ReusableSandboxPool(a.plan, capacity, {
                     progress: setupContext.progress,
                     diagnostic: setupContext.diagnostic,
-                }, setupContext, runtime_ts_1.liveSandboxRuntimeServices, agent, runTiming, materializationOwnerId, () => nextReuseSandboxNumber(a.run), a.evalDef.evalGroup === undefined ? Object.freeze([]) : (0, contracts_ts_1.linkPluginLifecycles)(a.evalDef.evalGroup.plugins ?? [], "group"), groupPluginContext);
+                }, setupContext, runtime_ts_1.liveSandboxRuntimeServices, agent, runTiming, materializationOwnerId, () => nextReuseSandboxNumber(a.run), (0, link_ts_1.groupPluginLifecycles)(a.evalDef, a.run), groupPluginContext);
                 bySpec.set(key, pool);
                 return pool.managed().pipe(effect_1.Effect.map((managed) => ({ _tag: "Reuse", pool: managed })));
             }
@@ -961,12 +926,12 @@ function runEvals(opts) {
         /**
          * Owner tokens are an explicit inspection capability, not ordinary Runner
          * feedback. Keep both the live feedback event and the durable Run diagnostic
-         * limited to the key and recovery obligation; `exp --teardown` is the only
+         * limited to the key and recovery obligation; `exp --after` is the only
          * public surface that renders immutable owner evidence.
          */
         const reportSharedStateRecoveryRequired = (input) => {
             const key = input.run.sharedState?.key ?? "";
-            const message = `sharedState key ${key} requires explicit recovery before another Invocation can use it. Inspect immutable owner evidence with \`niceeval exp <selector> --teardown --recover-shared-state <key>\`.
+            const message = `sharedState key ${key} requires explicit recovery before another Invocation can use it. Inspect immutable owner evidence with \`niceeval exp <selector> --after --recover-shared-state <key>\`.
 `.trimEnd();
             (0, sink_ts_1.reportDiagnostic)({
                 key: `state-lease-recovery-required:${key || input.experimentId}`,
@@ -1391,7 +1356,7 @@ function runEvals(opts) {
                                 host: currentHost,
                                 startedAt: new Date().toISOString(),
                             }).pipe(effect_1.Effect.catch((error) => effect_1.Effect.sync(() => {
-                                const message = `writing the crash-recovery teardown registration for experiment ${experimentId} failed: ${error instanceof Error ? error.message : String(error)}. The run continues normally, but a SIGKILL during this run cannot be recovered via \`niceeval exp --teardown\` or the startup self-heal — check the canonical ProjectDatabase write failure.
+                                const message = `writing the crash-recovery teardown registration for experiment ${experimentId} failed: ${error instanceof Error ? error.message : String(error)}. The run continues normally, but a SIGKILL during this run cannot be recovered via \`niceeval exp --after\` or the startup self-heal — check the canonical ProjectDatabase write failure.
 `.trimEnd();
                                 (0, sink_ts_1.reportDiagnostic)({
                                     key: `teardown-registration-write-failed:${experimentId}`,
@@ -2264,7 +2229,7 @@ function runEvals(opts) {
                     const leaseError = leased._tag === "Failed" ? (0, attempt_ts_1.errorFromThrown)(leased.error, "sandbox.create") : undefined;
                     if (leased._tag === "Failed") {
                         cancelReuseAttempt(a, true);
-                        const declaration = (0, attempt_ts_1.attemptFailureDeclaration)(a.run.classifyFailure, "sandbox.create", leased.error);
+                        const declaration = (0, attempt_ts_1.attemptFailureDeclaration)("sandbox.create", leased.error);
                         if (declaration)
                             closeHaltGate(a, declaration);
                         // This Attempt did start its physical acquisition and is therefore recorded below