humanish 0.120.1 → 0.122.0

package/dist/routes/computer-use/result.js

@@ -1,6 +1,7 @@
 import { adapterScoreFailureMessage, applyBrowserScorer } from "../../study/adapter-extension.js";
 import { redactText } from "../../evidence/redaction.js";
-import { foldScorerFailures, judgeExecution, judgeParticipantRecords, OUTCOME_POLICIES, resultOk, } from "../../run/judge.js";
+import { foldScorerFailures, judgeExecution, judgeParticipantRecords, OUTCOME_POLICIES, participantHarnessFailed, resultOk, } from "../../run/judge.js";
+import { readHostSuspensions, } from "../../run/host-suspension.js";
 import { validatePreparedRunArtifactPaths } from "../../run/paths.js";
 import { participantExecutionFailures, participantOutcomeOk, participantFactsOf, } from "./participant-facts.js";
 import { participantCapWarning } from "./participant-model.js";
@@ -53,13 +54,39 @@ function buildParticipantSummary(outcomes, participantCount, participantPlan, dr
         waves: participantPlan.waves,
     };
 }
+/**
+ * Each participant as the host-suspension reading reads it: a harness error, its session's time
+ * limit or a skip, when it started, and when its runner returned, which a setup failure with no
+ * session also records.
+ */
+function suspendedParticipants(outcomes) {
+    return outcomes.map((outcome) => {
+        const facts = participantFactsOf(outcome);
+        const trace = outcome.session?.trace;
+        const failure = facts.skipped
+            ? "skipped"
+            : participantHarnessFailed(facts)
+                ? "harness"
+                : trace?.stopCause === "time_limit"
+                    ? "time-limit"
+                    : undefined;
+        const startedAtMs = outcome.arrival?.startedAt ?? (trace === undefined ? undefined : Date.parse(trace.startedAt));
+        const endedAtMs = outcome.endedAt ?? (trace === undefined ? undefined : Date.parse(trace.completedAt));
+        return {
+            id: outcome.spec.planned.id,
+            ...(failure === undefined ? {} : { failure }),
+            ...(startedAtMs === undefined ? {} : { startedAtMs }),
+            ...(endedAtMs === undefined ? {} : { endedAtMs }),
+        };
+    });
+}
 /**
  * The computer-use study result for a finished run. The run passes only when the Observer rendered,
  * every participant passed (dry-run participants pass as contracts), and no adapter score or declared scorer
  * verdict failed; otherwise the error names the first reason.
  */
 function cuaStudyResult(args) {
-    const { studyId, cwd, runId, appUrl, dryRun, participantRuns, outcomes, subjects, aggregateSubject, participantPlan, rerunLineage, bundle, observer, receivingWarnings, aggregateWarnings, adapterWarnings, } = args;
+    const { studyId, cwd, runId, appUrl, dryRun, participantRuns, outcomes, subjects, aggregateSubject, participantPlan, rerunLineage, bundle, observer, hostSuspension, receivingWarnings, aggregateWarnings, adapterWarnings, } = args;
     const participantCount = participantRuns.length;
     // Participant-level pass: dry-run participants are contract-ok; live ones need a passed,
     // engaged session.
@@ -77,6 +104,10 @@ function cuaStudyResult(args) {
     const participantResults = participantRuns.map((spec, index) => toParticipantResult(spec, outcomes?.[index], subjects[index], dryRun));
     const summary = buildParticipantSummary(outcomes, participantCount, participantPlan, dryRun);
     const firstOutcome = outcomes?.[0];
+    // A failure the host suspension likely caused names the suspension first.
+    const suspended = hostSuspension !== undefined && hostSuspension.participantIds.length > 0
+        ? hostSuspension.summary
+        : undefined;
     const errorResult = (() => {
         if (ok)
             return undefined;
@@ -87,24 +118,31 @@ function cuaStudyResult(args) {
             const outcome = firstOutcome;
             return {
                 code: outcome?.failureCode ?? "HUMANISH_COMPUTER_USE_FAILED",
-                message: outcome?.sessionError ??
+                message: `${suspended === undefined ? "" : `${suspended} `}${outcome?.sessionError ??
                     outcome?.providerCleanupError ??
                     outcome?.providerPolicyError ??
                     (outcome?.noEngagement || outcome?.selfReportedBlocker || observer.ok
                         ? (participantResults[0]?.error?.message ??
                             judgeParticipantRecords([participantFactsOf(outcome)]).participants[0]
                                 .notPassedMessage)
-                        : (observer.error?.message ?? "Observer failed for the computer-use run.")),
+                        : (observer.error?.message ?? "Observer failed for the computer-use run."))}`,
             };
         }
         const failing = (outcomes ?? []).find((outcome) => !participantOk(outcome));
+        const firstFailure = suspended !== undefined
+            ? `; ${suspended.replace(/\.$/, "")}`
+            : failing?.sessionError !== undefined
+                ? `; first failure: ${failing.sessionError}`
+                : failing === undefined && args.outcome.execution.failures[0] !== undefined
+                    ? `; ${args.outcome.execution.failures[0].message}`
+                    : "";
         // A participant whose desktop failed for a known cause gives the run that cause's code.
         const codedFailure = (outcomes ?? []).find((outcome) => outcome.failureCode !== undefined);
         const code = codedFailure?.failureCode ?? "HUMANISH_COMPUTER_USE_FAILED";
         return {
             code,
             message: observer.ok
-                ? `Fan-out run failed: ${summary.passed}/${plural(participantCount, "participant")} passed (${summary.skipped} skipped, ${plural(summary.harnessErrors, "harness error")}, ${summary.hollow} without engagement)${failing?.sessionError !== undefined ? `; first failure: ${failing.sessionError}` : failing === undefined && args.outcome.execution.failures[0] !== undefined ? `; ${args.outcome.execution.failures[0].message}` : ""}.`
+                ? `Fan-out run failed: ${summary.passed}/${plural(participantCount, "participant")} passed (${summary.skipped} skipped, ${plural(summary.harnessErrors, "harness error")}, ${summary.hollow} without engagement)${firstFailure}.`
                 : (observer.error?.message ?? "Observer failed for the computer-use fan-out run."),
         };
     })();
@@ -149,6 +187,7 @@ function cuaStudyResult(args) {
         laneSummary: summary,
         ...(rerunLineage === undefined ? {} : { rerun: rerunLineage }),
         observer,
+        ...(hostSuspension === undefined ? {} : { hostSuspension }),
         warnings,
         ...(errorResult === undefined ? {} : { error: errorResult }),
     };
@@ -162,7 +201,7 @@ export async function finishCuaRun(setup, finish, ran) {
     const appUrl = plannedAppUrl(plan.runner.subject);
     const { runId } = run;
     const participantCount = participantRuns.length;
-    const { outcomes, failFastReason, receiving, receivingWarnings, externalCommsWarnings } = ran;
+    const { outcomes, receiving, receivingWarnings, externalCommsWarnings } = ran;
     // Per-participant subject projections.
     const subjects = projectParticipantSubjects({ ...subjectArgs, outcomes, dryRun });
     const aggregate = aggregateCuaSubject({ subjects, outcomes, participantCount, dryRun });
@@ -175,13 +214,15 @@ export async function finishCuaRun(setup, finish, ran) {
     ];
     // One judgment for the whole run: the bundle's verdict and the result's ok both read it.
     const judgment = judgeComputerUseRun(bundleBase, { dryRun, outcomes });
+    // One reading of the host suspensions: the review, the outcome and the result all name it.
+    const hostSuspension = readHostSuspensions(run.hostSuspensions(), suspendedParticipants(outcomes ?? []), Date.parse(run.createdAt));
     const bundle = buildCuaRunBundle(bundleBase, {
         judgment,
         dryRun,
         outcomes,
         subjects,
         aggregateSubject,
-        ...(failFastReason === undefined ? {} : { failFastReason }),
+        ...(hostSuspension === undefined ? {} : { hostSuspension }),
     });
     const adapterWarnings = [];
     const scorerResult = await applyBrowserScorer({
@@ -206,8 +247,14 @@ export async function finishCuaRun(setup, finish, ran) {
     if (receiving)
         bundle.commsReceiving = receiving.snapshot();
     const policy = OUTCOME_POLICIES["computer-use"];
-    // FinishedRun.renderObserver adds an Observer that did not render.
-    const execution = judgeExecution(participantExecutionFailures(outcomes, runId), policy);
+    // FinishedRun.renderObserver adds an Observer that did not render. A suspension that likely
+    // caused participant failures is the run's first failure, ahead of each one it explains.
+    const execution = judgeExecution([
+        ...(hostSuspension !== undefined && hostSuspension.participantIds.length > 0
+            ? [{ kind: "run", message: hostSuspension.summary }]
+            : []),
+        ...participantExecutionFailures(outcomes, runId),
+    ], policy);
     const finished = await run.finish(bundle, {
         ok: resultOk({ judgment, execution, scorerFailures: scorerResult.failures, policy }),
         execution,
@@ -236,6 +283,7 @@ export async function finishCuaRun(setup, finish, ran) {
         bundle,
         outcome: finished.outcome,
         observer,
+        hostSuspension,
         receivingWarnings,
         aggregateWarnings,
         adapterWarnings,

package/dist/routes/computer-use/setup.d.ts

@@ -38,7 +38,7 @@ export interface CuaRunSetup {
 /** What only runStudyParticipants reads. */
 export interface CuaParticipantsSetup {
     env: Record<string, string | undefined>;
-    deps: Omit<CuaParticipantDeps, "signalProvisioned">;
+    deps: Omit<CuaParticipantDeps, "signalReady">;
     /** Filled by runStudyParticipants on a live run; deps.onTrace reads it. */
     liveTrace: {
         flush?: LiveTraceFlush["flush"];

package/dist/routes/computer-use/single-bundle.d.ts

@@ -7,6 +7,7 @@ import { type BundleRun, type RunBundle, type RunFeedbackCandidate, type RunProv
 import { type RunDesktopGeometry, type RunStream } from "../../run/streams.js";
 import { type Verdict } from "../../run/judge.js";
 import { type DesktopUsage } from "../../run/cost-summary.js";
+import type { HostSuspensionReading } from "../../run/host-suspension.js";
 /**
  * The bundle of a single-participant run: one participant and no rerun. buildCuaRunBundle
  * (bundle.ts) maps the run's base and state into these arguments.
@@ -92,4 +93,6 @@ export declare function buildSingleParticipantBundle(args: {
      *  desktopDurationMs). Absent when no sandbox ran (in-process/dry-run) → no desktop cost line. */
     desktopMinutes?: number;
     desktopUsage?: DesktopUsage;
+    /** The run's host suspensions and the failure they likely caused. */
+    hostSuspension?: HostSuspensionReading;
 }): RunBundle;

package/dist/routes/computer-use/single-bundle.js

@@ -332,10 +332,11 @@ function singleReview(args, view, stream) {
             }
             : {}),
         ...(singleRunTasks === undefined ? {} : { tasks: singleRunTasks }),
-        summary: singleSummary(args, view, {
+        // A host suspension comes first: it may explain everything after it.
+        summary: `${args.hostSuspension === undefined ? "" : `${args.hostSuspension.summary} `}${singleSummary(args, view, {
             ...(singleRunTasks === undefined ? {} : { tasks: singleRunTasks }),
             ...(credibilityNote === undefined ? {} : { credibilityNote }),
-        }),
+        })}`,
         gaps: args.session || args.sessionError !== undefined
             ? []
             : args.inProgress === true

package/dist/routes/computer-use/types.d.ts

@@ -1,4 +1,5 @@
 import type { Verdict } from "../../run/judge.js";
+import type { HostSuspensionReading } from "../../run/host-suspension.js";
 import type { LocalCapacity } from "../../substrates/local/capacity.js";
 import type { RunDesktopRecording } from "../../evidence/desktop-recording-types.js";
 import type { CuaExecutor, CuaLiveMetadata, CuaLoopResult, CuaProvider } from "../../actors/computer-use/loop.js";
@@ -111,6 +112,8 @@ export interface CuaParticipantPlanEntry {
     targetDigest?: string;
     /** The participant's declared start, in ms after the run starts its participants. */
     startAfterMs?: number;
+    /** The participant's declared end, on the same clock. */
+    endAfterMs?: number;
 }
 /** The pre-flight spend and participant plan (pure; printed to stderr + recorded as a bundle event before
  *  any sandbox or provider call; identical in dry-run, marked $0). */
@@ -137,7 +140,8 @@ export interface CuaParticipantPlan {
     lanes: CuaParticipantPlanEntry[];
 }
 /** One participant's outcome in the result projection. Always present in `result.lanes` (length
- *  1 at N=1). A `blocked` participant is one the pipeline-gate / fail-fast skipped before it ran. */
+ *  1 at N=1). A `blocked` participant is one the pipeline gate or the study budget skipped before
+ *  it ran. */
 export interface CuaParticipantResult {
     id: string;
     actorType?: string;
@@ -148,7 +152,7 @@ export interface CuaParticipantResult {
     device: string;
     /** Requested E2B/X screen resolution. See the run stream's desktopGeometry for measurements. */
     resolution: [number, number];
-    /** Terminal participant status; "blocked" = skipped (gate/fail-fast); "contract_proof_only" = dry-run. */
+    /** Terminal participant status; "blocked" = skipped (gate or study budget); "contract_proof_only" = dry-run. */
     status: ActorStatus | "blocked" | "contract_proof_only";
     ok: boolean;
     session?: {
@@ -182,7 +186,7 @@ export interface CuaParticipantSummary {
     total: number;
     /** Participants whose own verdict is ok (terminal, engaged, no harness error). */
     passed: number;
-    /** Participants skipped by the pipeline gate / fail-fast. */
+    /** Participants skipped by the pipeline gate or the study budget. */
     skipped: number;
     /** Participants that ended in a harness error. */
     harnessErrors: number;
@@ -233,6 +237,11 @@ export interface CuaActorStudyResult extends AutomaticAnalysisResult, StudyResul
     rerun?: RunRerunLineage;
     observer?: ObserverResult;
     diagnostics?: CuaDiagnostics;
+    /**
+     * Present when the host was suspended during the run (a laptop that slept): when, and which
+     * participant failures it likely caused. Each participant's own result is unchanged.
+     */
+    hostSuspension?: HostSuspensionReading;
     warnings: string[];
     error?: {
         code: CuaActorStudyErrorCode;
@@ -357,6 +366,16 @@ export interface ParticipantModelDeps {
     env: Record<string, string | undefined>;
     openaiApiKey: string;
     timeoutMs: number;
+    /**
+     * The participant's scheduled end on `now`'s clock, when it declares participants[].endAfterMs
+     * (scheduledParticipantDeps). Its session ends there with stop cause `scheduled_end`.
+     */
+    scheduledEnd?: {
+        readonly atMs: number;
+        readonly windDownMs: number;
+    };
+    /** The run's clock. A participant with a scheduled end runs its session on it. */
+    now: () => number;
     participantCount: number;
     artifactRoot: PreparedOutputRoot;
     redactScreenshots: boolean;
@@ -391,8 +410,8 @@ export type CuaParticipantDeps = E2BDesktopDeps & ParticipantModelDeps & {
     /** The caller's inProcess executor with the run's config bound. Read by the in-process desktop. */
     inProcessExecutor?: (appUrl: string) => Promise<CuaExecutor>;
     runSession: (options: CuaActorSessionOptions) => Promise<CuaLoopResult>;
-    /** Lane-0 only: signal the pipeline gate after provisioning succeeds (true) or fails (false). */
-    signalProvisioned?: (ok: boolean) => void;
+    /** Called once, when the participant's desktop and model are ready and its session starts. */
+    signalReady?: () => void;
 };
 /** Why a sandbox was not confirmed released. */
 export interface SandboxReleaseFact {
@@ -435,10 +454,16 @@ export interface ParticipantRunOutcome {
      *  folded into bundle.events at build time. Empty on the in-process route (no provisioning). */
     phaseRecords: SubjectPhaseEvent[];
     warnings: string[];
-    /** Set when the participant was skipped by the pipeline gate / fail-fast (a pinned reason). */
+    /** Set when the pipeline gate or the study budget skipped the participant (a pinned reason). */
     skippedReason?: string;
     /** When a fan-out participant was due and when its desktop was requested. */
     arrival?: ParticipantArrival;
+    /**
+     * When the participant's runner returned this outcome, on the deps clock (epoch ms): after its
+     * session, or after a setup failure that left it none. Absent on a skipped participant, which
+     * never ran, and on outcomes built outside the runner.
+     */
+    endedAt?: number;
     noEngagement: boolean;
     selfReportedBlocker: boolean;
     /** The inclusive friction read: blocker-shaped narration incl. self-resolved arcs.
@@ -469,7 +494,8 @@ export interface CuaFanoutBundleArgs {
     source: RunBundle["source"];
     participantPlan: CuaParticipantPlan;
     rerun?: RunRerunLineage;
-    failFastReason?: string;
+    /** The run's host suspensions and the participant failures they likely caused. */
+    hostSuspension?: HostSuspensionReading;
     publicRepo?: string;
     inProgress?: boolean;
 }

package/dist/routes/shared-world/provisioned.js

@@ -383,8 +383,8 @@ async function publishInProgress(plane, ctx, live, setup) {
     startParticipantFlush(ctx, live, inProgressBundle);
 }
 // Launch N actor sandboxes independently on the study's schedule (runCuaParticipant +
-// runOnSchedule; runCuaParticipants would add a pipeline gate and fail-fast). Each actor's window is
-// measured on the one orchestrator clock.
+// runOnSchedule; runCuaParticipants would add a pipeline gate). Each actor's window is measured on
+// the one orchestrator clock.
 function runParticipants(plane, ctx, live, setup) {
     const { now } = ctx;
     const participants = ctx.plan.plane.participants;

package/dist/run/host-suspension.d.ts

@@ -0,0 +1,56 @@
+import type { RunEvent } from "./bundle.js";
+/** The run's clock and heartbeat timer. Tests pass one whose wall clock they move. */
+export interface HostClock {
+    /** Wall-clock time in epoch milliseconds. */
+    now(): number;
+    /** Call `tick` every `intervalMs` until the returned function runs. Never keeps the process alive. */
+    every(intervalMs: number, tick: () => void): () => void;
+}
+export declare const systemHostClock: HostClock;
+/** A time the run's heartbeat did not run: from its last tick before the gap to its first after. */
+export interface HostSuspension {
+    startedAt: string;
+    endedAt: string;
+    durationMs: number;
+}
+export interface Heartbeat {
+    /** The suspensions so far. A read also checks the time since the last tick. */
+    suspensions(): readonly HostSuspension[];
+    /** Stop the timer. Later reads return the suspensions recorded until then. */
+    stop(): void;
+}
+/** Run `onTick` every `intervalMs` on `clock`, and record each gap of 30 s or more. */
+export declare function startHeartbeat(clock: HostClock, intervalMs: number, onTick: () => void): Heartbeat;
+/** One warn event per suspension, at its start, naming its start, end, length and offset. */
+export declare function hostSuspensionEvents(suspensions: readonly HostSuspension[], runStartMs: number): RunEvent[];
+/** One participant as the attribution reads it. */
+export interface SuspendedParticipant {
+    id: string;
+    /**
+     * How it did not pass, when a suspension can explain that: an error in the harness, its
+     * session's time limit, or a skip after another participant's harness error. Absent when it
+     * passed or its session ended on its own report.
+     */
+    failure?: "harness" | "time-limit" | "skipped";
+    /** When it started; absent when no record says. */
+    startedAtMs?: number;
+    /** When it stopped, after its session or after a setup failure; absent when no record says. */
+    endedAtMs?: number;
+}
+/** What a run says about its host suspensions. */
+export interface HostSuspensionReading {
+    /** One sentence for the run outcome, the CLI and the review. */
+    summary: string;
+    /** The participants whose failure is likely a suspension's effect, in the order given. */
+    participantIds: string[];
+    suspensions: HostSuspension[];
+}
+/**
+ * Which failures the run's suspensions likely caused, and the sentence that says so; undefined
+ * without a suspension. A frozen loop's failures surface at the wake (macOS fires every overdue
+ * timer at once) or up to a turn timeout later (Linux timers do not count suspended time), so the
+ * rule reads who was in flight across the suspension. A skip follows another participant's
+ * harness error (the pipeline gate), so skips count only when the run had harness errors and
+ * every one was in flight across a suspension.
+ */
+export declare function readHostSuspensions(suspensions: readonly HostSuspension[], participants: readonly SuspendedParticipant[], runStartMs: number): HostSuspensionReading | undefined;

package/dist/run/host-suspension.js

@@ -0,0 +1,106 @@
+// Host suspensions: when the machine running a study sleeps, every participant loop on it stops,
+// and on wake its timers and connections fail together. This module notices the gap and says
+// which participant failures it likely caused, so the run names the suspension as their cause.
+//
+// Detection reads the wall clock across a heartbeat. Date.now counts suspended time on macOS and
+// Linux. performance.now and Node's timers run on libuv's clock, which counts sleep on macOS
+// (mach_continuous_time), so the tick after a sleep fires at wake, and stops during a Linux
+// suspend (CLOCK_MONOTONIC), so the tick fires within one interval of resume. Either way the wall
+// clock between two ticks shows the gap.
+import { formatDuration } from "./projection.js";
+import { plural } from "./text.js";
+export const systemHostClock = {
+    now: () => Date.now(),
+    every(intervalMs, tick) {
+        const timer = setInterval(tick, intervalMs);
+        timer.unref?.();
+        return () => clearInterval(timer);
+    },
+};
+/**
+ * A wall-clock gap between two heartbeats this long or longer is a suspension. A shorter one is
+ * most likely a busy event loop.
+ */
+const HOST_SUSPENSION_MIN_MS = 30_000;
+/** Run `onTick` every `intervalMs` on `clock`, and record each gap of 30 s or more. */
+export function startHeartbeat(clock, intervalMs, onTick) {
+    const found = [];
+    let lastMs = clock.now();
+    let stopped = false;
+    const observe = () => {
+        if (stopped)
+            return;
+        const nowMs = clock.now();
+        if (nowMs - lastMs >= HOST_SUSPENSION_MIN_MS)
+            found.push({
+                startedAt: new Date(lastMs).toISOString(),
+                endedAt: new Date(nowMs).toISOString(),
+                durationMs: nowMs - lastMs,
+            });
+        lastMs = nowMs;
+    };
+    const stopTimer = clock.every(intervalMs, () => {
+        observe();
+        onTick();
+    });
+    return {
+        suspensions() {
+            observe();
+            return [...found];
+        },
+        stop() {
+            stopped = true;
+            stopTimer();
+        },
+    };
+}
+/** `6m 5s at +1m`: how long, and how far into the run it started. */
+function span(suspension, runStartMs) {
+    const offset = Date.parse(suspension.startedAt) - runStartMs;
+    return `${formatDuration(suspension.durationMs)} at +${formatDuration(offset)}`;
+}
+/** One warn event per suspension, at its start, naming its start, end, length and offset. */
+export function hostSuspensionEvents(suspensions, runStartMs) {
+    return suspensions.map((suspension, index) => ({
+        id: `event-host-suspended-${String(index + 1).padStart(3, "0")}`,
+        at: suspension.startedAt,
+        level: "warn",
+        type: "host.suspended",
+        message: `The host was suspended for ${formatDuration(suspension.durationMs)}, from ${suspension.startedAt} to ${suspension.endedAt} (+${formatDuration(Date.parse(suspension.startedAt) - runStartMs)} into the run): the run's heartbeat did not run in that time. A process that was stopped, or a clock set forward, reads the same way.`,
+    }));
+}
+/** The participant was running at some moment of the suspension. */
+function inFlight(participant, suspension) {
+    return ((participant.startedAtMs ?? -Infinity) <= Date.parse(suspension.endedAt) &&
+        (participant.endedAtMs ?? Infinity) >= Date.parse(suspension.startedAt));
+}
+/**
+ * Which failures the run's suspensions likely caused, and the sentence that says so; undefined
+ * without a suspension. A frozen loop's failures surface at the wake (macOS fires every overdue
+ * timer at once) or up to a turn timeout later (Linux timers do not count suspended time), so the
+ * rule reads who was in flight across the suspension. A skip follows another participant's
+ * harness error (the pipeline gate), so skips count only when the run had harness errors and
+ * every one was in flight across a suspension.
+ */
+export function readHostSuspensions(suspensions, participants, runStartMs) {
+    if (suspensions.length === 0)
+        return undefined;
+    const explained = new Set(participants
+        .filter((participant) => (participant.failure === "harness" || participant.failure === "time-limit") &&
+        suspensions.some((suspension) => inFlight(participant, suspension)))
+        .map((participant) => participant.id));
+    const harness = participants.filter((participant) => participant.failure === "harness");
+    const skipsFollow = harness.length > 0 && harness.every((participant) => explained.has(participant.id));
+    const participantIds = participants
+        .filter((participant) => explained.has(participant.id) || (participant.failure === "skipped" && skipsFollow))
+        .map((participant) => participant.id);
+    const one = suspensions.length === 1;
+    const what = one
+        ? `The host was suspended for ${span(suspensions[0], runStartMs)}`
+        : `The host was suspended ${suspensions.length} times (${suspensions.map((suspension) => span(suspension, runStartMs)).join(", ")})`;
+    const count = participantIds.length;
+    const effect = count === 0
+        ? "no participant failure followed"
+        : `the ${plural(count, "participant failure")} after ${one ? "it" : "them"} ${count === 1 ? "is" : "are"} likely ${one ? "its" : "their"} effect`;
+    return { summary: `${what}; ${effect}.`, participantIds, suspensions: [...suspensions] };
+}

package/dist/run/judge.d.ts

@@ -34,7 +34,7 @@ export interface ParticipantFacts {
     completionReason?: ActorCompletionReason;
     /** The harness failed before the session reached a terminal status. */
     sessionError?: string;
-    /** The pipeline gate or fail-fast skipped the participant. */
+    /** The pipeline gate or the study budget skipped the participant. */
     skipped: boolean;
     /** hollowCompletion held for its session. */
     noEngagement: boolean;

package/dist/run/outcomes.js

@@ -193,7 +193,9 @@ export function formatParticipantOutcomes(outcomes, terminalCauses = []) {
             const description = entry.label === undefined ? fallback : `interrupted (${entry.label})`;
             counts.set(description, (counts.get(description) ?? 0) + 1);
         }
-        for (const [description, n] of counts)
+        // Participants finish in any order, so the causes are sorted: most first, then by name.
+        const sorted = [...counts].sort(([a, m], [b, n]) => n - m || a.localeCompare(b));
+        for (const [description, n] of sorted)
             parts.push(`${n} ${description}`);
     };
     append(outcomes.abandoned, ["abandoned"], "gave up");

package/dist/run/run.d.ts

@@ -3,6 +3,7 @@ import { type RunBundle } from "./bundle.js";
 import { type ExecutionFailure, type ExecutionOutcome, type OutcomePolicy } from "./judge.js";
 import { type PreparedRunArtifactPaths, type RunIdInUse } from "./paths.js";
 import { RunSecrets } from "./secrets.js";
+import { type HostClock, type HostSuspension } from "./host-suspension.js";
 import type { RunStudyProvenance } from "./study-provenance.js";
 interface StartRunOptions {
     /** Project directory, in the form the route resolved it. */
@@ -21,8 +22,10 @@ interface StartRunOptions {
         open: boolean;
         render?: typeof renderObserver | undefined;
     };
-    /** Clock for `createdAt` and the latest pointer. */
+    /** Clock for `createdAt` and the latest pointer. Defaults to the clock's `now`. */
     now?: (() => number) | undefined;
+    /** The run's wall clock and heartbeat timer. Defaults to systemHostClock. */
+    clock?: HostClock | undefined;
     /** Warnings about the study's own fields. Every bundle write records each as a warn event. */
     warnings?: readonly string[] | undefined;
     /** The route's known values, created before the run so a refusal before it is scrubbed too. */
@@ -42,9 +45,13 @@ interface StartRunInput {
     readonly runId?: string | undefined;
     /** Open the Observer page once it renders. */
     readonly open?: boolean | undefined;
-    /** `renderObserver` is the seam FinishedRun.renderObserver calls. */
+    /**
+     * `renderObserver` is the seam FinishedRun.renderObserver calls; `hostClock` is the run's
+     * clock and heartbeat timer.
+     */
     readonly deps?: {
         readonly renderObserver?: typeof renderObserver | undefined;
+        readonly hostClock?: HostClock | undefined;
     } | undefined;
 }
 /** What a recorded route passes to startRun besides its plan and input. */
@@ -100,6 +107,12 @@ interface Run {
      * and a run that failed before any session (a refused E2B key) does not.
      */
     participantStarted(): void;
+    /**
+     * The times this process did not run while the run was alive, most likely because the host
+     * slept, as the run's heartbeat saw them up to this call. Every bundle write records each as a
+     * `host.suspended` event; a route reads them to say which participant failures they caused.
+     */
+    hostSuspensions(): readonly HostSuspension[];
     /**
      * The one final publication: run.json with `outcome` set from `outcome`, then the status
      * outcome copied from it, then review.json, review.md, events.ndjson,

package/dist/run/run.js

@@ -16,6 +16,7 @@ import { scrubRunSandboxIds, withPublicSandboxIds } from "./sandbox-ids.js";
 import { RunSecrets } from "./secrets.js";
 import { RUN_STATUS_FILE, beginRunStatus, runStatusOutcome, } from "./status.js";
 import { readRunJsonIfExists, runJsonValue } from "./locate.js";
+import { hostSuspensionEvents, systemHostClock, } from "./host-suspension.js";
 /** An entry point called in a state that does not admit it. */
 class RunLifecycleError extends Error {
     name = "RunLifecycleError";
@@ -127,10 +128,14 @@ function runPublisher(args) {
         type: "study.warning",
         message: redactText(message),
     }));
-    const withStudyWarnings = (events) => [
-        ...events.filter((event) => !studyWarnings.some((warning) => warning.id === event.id)),
-        ...studyWarnings,
-    ];
+    // So do the host suspensions the heartbeat has seen by the time of the write.
+    const withRunEvents = (events) => {
+        const owned = [
+            ...studyWarnings,
+            ...hostSuspensionEvents(runStatus.suspensions(), Date.parse(createdAt)),
+        ];
+        return [...events.filter((event) => !owned.some((mine) => mine.id === event.id)), ...owned];
+    };
     let pointerWritten = false;
     // The bundle of the last write asked for, and the outcome every later write carries once a
     // signal stopped the run, so no write after the interrupt can drop it.
@@ -162,7 +167,7 @@ function runPublisher(args) {
         const recorded = interrupted ?? outcome;
         const publicBundle = await withPublicSandboxIds(paths, {
             ...evidence,
-            events: withStudyWarnings(evidence.events),
+            events: withRunEvents(evidence.events),
             cwd: PUBLIC_TARGET_CWD,
             ...(recorded === undefined ? {} : { outcome: recorded }),
         });
@@ -259,8 +264,8 @@ export async function runScope(fn) {
             const created = await createRunArtifactPaths(options.cwd, runId);
             if (!created.ok)
                 return created;
-            // beginRunStatus reads only the mode, the study provenance and the sandboxes field from
-            // the run's options.
+            // beginRunStatus reads only the mode, the study provenance, the sandboxes field and the
+            // clock from the run's options.
             const runStatus = beginRunStatus(created.paths, { ...options, runId });
             status = runStatus;
             let interruptBundle;
@@ -290,7 +295,7 @@ export async function runScope(fn) {
         return args.length === 1 ? startRunWith(args[0]) : startRecordedRun(startRunWith, ...args);
     }
     const openRun = (options, runId, paths, runStatus) => {
-        const now = options.now ?? Date.now;
+        const now = options.now ?? (options.clock ?? systemHostClock).now;
         const createdAt = new Date(now()).toISOString();
         let participantsRan = false;
         const publisher = runPublisher({
@@ -317,6 +322,7 @@ export async function runScope(fn) {
             participantStarted() {
                 participantsRan = true;
             },
+            hostSuspensions: () => runStatus.suspensions(),
             writeSnapshot(bundle) {
                 if (closed)
                     return refuse("The run scope has closed.");
@@ -374,6 +380,7 @@ async function startRecordedRun(start, plan, input, route) {
         renderReview: route.renderReview,
         observer: { open: input.open === true, render: input.deps?.renderObserver },
         now: route.now,
+        clock: input.deps?.hostClock,
         secrets: route.secrets,
     });
     if (!started.ok)