humanish 0.107.0 → 0.108.0

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

@@ -5,16 +5,16 @@ import { type CuaSubjectProvenanceArg, type ParticipantRunOutcome } from "./type
 /** The human-readable state story appended to the provenance event (and review.md via it). */
 export declare function describeSubjectState(state: RunSubjectProvenance["state"], dryRun: boolean): string;
 /**
- * Feedback candidates derived from what LIVE participants actually reported (#392).
+ * Feedback candidates derived from what live participants actually reported.
  *
- * A live run's feedback draft used to fall through to a dry-run template, because no browser route
- * ever built a candidate. The candidate worth filing is the one the study produced: a participant
+ * A live run's feedback draft comes from these candidates, never from a dry-run template. The
+ * candidate worth filing is the one the study produced: a participant
  * who reported friction on the way (the most valuable thing a run captures), or one who stopped
- * trying. A clean pass files nothing here — feedback exists to carry findings, and a run without
- * any falls back to an honest live summary in the draft layer instead of a template.
+ * trying. A clean pass files nothing here: feedback exists to carry findings, and a run without
+ * any falls back to a live summary in the draft layer instead of a template.
  *
- * Everything quoted is already scrub+redacted — participant messages and `session.reason` pass
- * through redactNarration in the loop — and passes redactText again here as defense-in-depth.
+ * Everything quoted is already scrub+redacted (participant messages and `session.reason` pass
+ * through redactNarration in the loop) and passes redactText again here as defense-in-depth.
  */
 export declare function participantFeedbackCandidates(args: {
     runId: string;
@@ -34,8 +34,8 @@ export declare function participantFeedbackCandidates(args: {
         commsArtifactPath?: string;
     }>;
 }): RunFeedbackCandidate[];
-/** Human-readable provenance line for the single-participant subject.provenance event (invariant 5):
- *  claims "cloned/packed and served" ONLY when it actually happened. */
+/** Human-readable provenance line for the single-participant subject.provenance event:
+ *  claims "cloned/packed and served" only when it actually happened. */
 export declare function subjectProvenanceMessage(provenance: CuaSubjectProvenanceArg, publicAppUrl: string, dryRun: boolean, hasSession: boolean): string;
 export declare function providerResourcesForOutcome(args: {
     outcome: ParticipantRunOutcome | undefined;

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

@@ -14,9 +14,9 @@ export function describeSubjectState(state, dryRun) {
         case "seeded":
             return `seeded (${state.seed?.length ?? 0} step(s): ${(state.seed ?? []).map((record) => record.name).join(", ")})`;
         case "unpinned":
-            return `UNPINNED (external: ${(state.externalEnvNames ?? []).join(", ")})`;
+            return `unpinned (external: ${(state.externalEnvNames ?? []).join(", ")})`;
         case "declared-not-run":
-            return `declared, not run (${dryRun ? "dry-run contract" : "provisioning did not complete"})`;
+            return `declared, not run (${dryRun ? "dry run" : "provisioning did not complete"})`;
         case "undeclared":
             return "undeclared";
         case "external-public":
@@ -24,16 +24,16 @@ export function describeSubjectState(state, dryRun) {
     }
 }
 /**
- * Feedback candidates derived from what LIVE participants actually reported (#392).
+ * Feedback candidates derived from what live participants actually reported.
  *
- * A live run's feedback draft used to fall through to a dry-run template, because no browser route
- * ever built a candidate. The candidate worth filing is the one the study produced: a participant
+ * A live run's feedback draft comes from these candidates, never from a dry-run template. The
+ * candidate worth filing is the one the study produced: a participant
  * who reported friction on the way (the most valuable thing a run captures), or one who stopped
- * trying. A clean pass files nothing here — feedback exists to carry findings, and a run without
- * any falls back to an honest live summary in the draft layer instead of a template.
+ * trying. A clean pass files nothing here: feedback exists to carry findings, and a run without
+ * any falls back to a live summary in the draft layer instead of a template.
  *
- * Everything quoted is already scrub+redacted — participant messages and `session.reason` pass
- * through redactNarration in the loop — and passes redactText again here as defense-in-depth.
+ * Everything quoted is already scrub+redacted (participant messages and `session.reason` pass
+ * through redactNarration in the loop) and passes redactText again here as defense-in-depth.
  */
 export function participantFeedbackCandidates(args) {
     const candidates = [];
@@ -59,8 +59,8 @@ export function participantFeedbackCandidates(args) {
             persona_id: participant.personaId,
             actor: "computer-use",
             substrate: args.substrate,
-            // The participant is reporting on the PRODUCT: friction and abandonment are target-app
-            // findings by the three-roles rule. A harness failure never reaches this builder — it is
+            // The participant is reporting on the product: friction and abandonment are target-app
+            // findings by the three-roles rule. A harness failure never reaches this builder; it is
             // not a participant report.
             failure_owner: "target-app",
             summary,
@@ -109,12 +109,12 @@ export function participantFeedbackCandidates(args) {
     }
     return candidates;
 }
-/** Human-readable provenance line for the single-participant subject.provenance event (invariant 5):
- *  claims "cloned/packed and served" ONLY when it actually happened. */
+/** Human-readable provenance line for the single-participant subject.provenance event:
+ *  claims "cloned/packed and served" only when it actually happened. */
 export function subjectProvenanceMessage(provenance, publicAppUrl, dryRun, hasSession) {
     if (provenance.source === "clone") {
         if (dryRun) {
-            return `Subject declared: clone of ${provenance.repo}, to be served at ${publicAppUrl} in-sandbox (dry-run contract; nothing cloned)`;
+            return `Subject declared: clone of ${provenance.repo}, to be served at ${publicAppUrl} in-sandbox (dry run; nothing cloned)`;
         }
         if (provenance.commit) {
             return hasSession
@@ -124,7 +124,7 @@ export function subjectProvenanceMessage(provenance, publicAppUrl, dryRun, hasSe
         return `Subject clone attempted from ${provenance.repo}; commit unresolved (provisioning failed before resolution)`;
     }
     if (dryRun) {
-        return `Subject declared: local working tree, to be packed and served at ${publicAppUrl} in-sandbox (dry-run contract; nothing packed)`;
+        return `Subject declared: local working tree, to be packed and served at ${publicAppUrl} in-sandbox (dry run; nothing packed)`;
     }
     if (provenance.archiveSha256) {
         const dirtyLabel = provenance.dirty === true

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

@@ -2,7 +2,7 @@
 // and buildCuaRunBundle sends it to the single-participant or the fan-out builder with that
 // shape's arguments. renderCuaReviewMarkdown writes review.md. shared-world/bundle.ts holds the
 // same for that route.
-import { judgeOneParticipant, judgeParticipants } from "../../run/judge.js";
+import { judgeOneParticipant, judgeParticipants, verdictText, } from "../../run/judge.js";
 import { desktopSpanToMinutes } from "../../run/cost-summary.js";
 import { e2bDesktopTemplate } from "../../substrates/e2b/sandbox.js";
 import { providerResourcesForOutcome } from "./bundle-parts.js";
@@ -157,13 +157,13 @@ export function renderCuaReviewMarkdown(bundle) {
         "",
         `- run: ${bundle.runId}`,
         `- mode: ${bundle.mode}`,
-        `- run gate: ${bundle.review.verdict}`,
+        `- run gate: ${verdictText(bundle.review.verdict, bundle.mode)}`,
         `- summary: ${bundle.review.summary}`,
         ...(provenance ? [`- subject: ${provenance.message}`] : []),
         ...(trace
             ? [
                 `- actor: ${trace.provider} (${trace.lane}/${trace.protocol})`,
-                // Honest count: name the trace's actual screenshot mode ("raw" | "blurred"); say
+                // Name the trace's actual screenshot mode ("raw" | "blurred"); say
                 // nothing when no frames exist ("n/a") rather than claim a redaction that never ran.
                 `- evidence: ${trace.items.length} trace item(s), ${trace.counts.screenshots ?? 0} ${trace.redaction.screenshots === "raw" || trace.redaction.screenshots === "blurred"
                     ? `${trace.redaction.screenshots} screenshot(s)`

package/dist/routes/computer-use/e2b-desktop/comms.d.ts

@@ -19,7 +19,7 @@ export interface ParticipantComms {
 }
 /**
  * On an in-sandbox subject route, redirect the app's email-API sends into an in-sandbox catch
- * (loopback) so its verification mail is CAPTURED, not sent to the internet. Undefined when the
+ * (loopback) so its verification mail is captured instead of sent to the internet. Undefined when the
  * lab declares no fake email, which leaves the participant unchanged.
  */
 export declare function planParticipantComms(comms: LabConfig["comms"], serveUrl: string | undefined, targetUrl: string, inSandboxSubject: boolean): ParticipantComms | undefined;

package/dist/routes/computer-use/e2b-desktop/comms.js

@@ -1,4 +1,4 @@
-// Off-app email for one E2B desktop participant (#297): the in-sandbox catch the subject app sends to, the
+// Off-app email for one E2B desktop participant: the in-sandbox catch the subject app sends to, the
 // persona's inbox surface, the real-email receiving surface, and the evidence drained at teardown.
 import { setTimeout as delay } from "node:timers/promises";
 import { buildOriginMap } from "../../../comms/capture-surface.js";
@@ -15,7 +15,7 @@ import { addressedRecipients } from "../../../lab/parse/comms.js";
 const INBOX_SURFACE_CADENCE_MS = 2500;
 /**
  * On an in-sandbox subject route, redirect the app's email-API sends into an in-sandbox catch
- * (loopback) so its verification mail is CAPTURED, not sent to the internet. Undefined when the
+ * (loopback) so its verification mail is captured instead of sent to the internet. Undefined when the
  * lab declares no fake email, which leaves the participant unchanged.
  */
 export function planParticipantComms(comms, serveUrl, targetUrl, inSandboxSubject) {
@@ -42,7 +42,7 @@ export function planParticipantComms(comms, serveUrl, targetUrl, inSandboxSubjec
 export function participantCommsEnv(comms) {
     if (comms === undefined)
         return {};
-    // injectEnv is absent on an adopter-hosted plane (#328): there is no subject env to inject
+    // injectEnv is absent on an adopter-hosted plane: there is no subject env to inject
     // because the operator points their own app at their own catch.
     const env = comms.email.injectEnv === undefined
         ? {}
@@ -75,23 +75,23 @@ export async function startCommsCatch(shell, comms, requestTimeoutMs) {
     if (!deployed.ready) {
         throw new Error(`comms email catch did not become ready on 127.0.0.1:${comms.port} in the subject sandbox`);
     }
-    // Write the EMPTY inbox once up front so the persona's /inbox always resolves to the "No messages
-    // yet." page — never a bare 404 — the instant it navigates there, even before any mail arrives OR if
+    // Write the empty inbox once up front so the persona's /inbox always resolves to the "No messages
+    // yet." page (never a bare 404) the instant it navigates there, even before any mail arrives or if
     // the app sends to an address no declared recipient matches (the loop only re-renders on new mail).
     await writeInboxSurface(shell, deployed.surfaceDir, [], {
         originMap: comms.originMap,
         requestTimeoutMs,
     });
-    // The surface uses its OWN FakeInbox + cursor, independent of the teardown evidence drain (two
-    // readers of the append-only NDJSON — no double-count).
+    // The surface uses its own FakeInbox + cursor, independent of the teardown evidence drain (two
+    // readers of the append-only NDJSON, so no double-count).
     const stop = new AbortController();
     let renderedCount = 0;
     const loop = (async () => {
         // Render-first (so even a short session gets a populated inbox), then refresh on a cadence. The
-        // cadence uses a REAL timer, NOT the injected instant clock: this loop is unbounded, so an instant
+        // cadence uses a real timer instead of the injected instant clock: this loop is unbounded, so an instant
         // sleep would busy-spin and starve the session's own timers. `stop` interrupts the wait
         // (and clears the timer) so teardown never blocks for a full cadence. Each refresh
-        // is a full, idempotent rebuild; `renderedCount` only advances on a SUCCESSFUL render so a
+        // is a full, idempotent rebuild; `renderedCount` only advances on a successful render so a
         // transient failure retries cleanly (no duplicate emails).
         for (;;) {
             try {
@@ -200,19 +200,19 @@ export async function drainCommsEvidence(args) {
             return path;
         }
         if (collected.captured > 0) {
-            // Captured mail that matched no declared recipient must not vanish silently (invariant 6:
-            // honest signals): tell the operator to declare comms.email.recipients[].address to match
+            // Captured mail that matched no declared recipient must not vanish silently: tell the
+            // operator to declare comms.email.recipients[].address to match
             // the address the app actually sends to (e.g. the one the persona surface will sign up with).
-            warnings.push(`Comms catch captured ${collected.captured} email send(s) but none matched a declared recipient inbox — no comms evidence written. Declare comms.email.recipients[].address to match the address the app sends to.`);
+            warnings.push(`The email catch captured ${collected.captured} email send(s), but none matched a declared recipient inbox, so no email evidence was written. Set comms.email.recipients[].address to the address the app sends to.`);
         }
         else {
-            // Zero captures is the silent-broken shape (#351): the app never posted to the catch at
+            // Zero captures is the silent-broken shape: the app never posted to the catch at
             // all, so the personas stared at an empty inbox. Most common cause: the app does not
             // actually read the declared injectEnv var for its email API base URL.
             const transportHint = comms.email.smtp
                 ? `Verify the app reads ${comms.email.smtp.hostEnv}/${comms.email.smtp.portEnv} for its SMTP host and port`
                 : `Verify the app reads ${comms.email.injectEnv} for its email API base URL (an SDK that ignores it sends real mail or throws)`;
-            warnings.push(`Comms catch captured ZERO email sends — the app never delivered mail through the catch. ${transportHint} and that the flow reached an email step.`);
+            warnings.push(`The email catch captured no email sends: the app never delivered mail through it. ${transportHint} and that the flow reached an email step.`);
         }
     }
     catch (error) {

package/dist/routes/computer-use/e2b-desktop/desktop.js

@@ -15,7 +15,7 @@ export function createE2BParticipantDesktop(spec, deps, warnings) {
         warnings,
         targetUrl,
         desktopCliRoute: deps.subject.kind === "desktop-cli",
-        // Off-app comms (#297): gated entirely on config.comms; no comms declared, no change.
+        // Off-app comms: gated entirely on config.comms; no comms declared, no change.
         comms: planParticipantComms(deps.residual.comms, participantServeUrl(deps.subject), targetUrl, deps.subject.kind === "clone" || deps.subject.kind === "local-tree"),
         // The default or injected sink sees every event, started and completed alike, so an operator
         // watching stderr sees both halves of each phase; the state keeps only completed ones.
@@ -23,9 +23,9 @@ export function createE2BParticipantDesktop(spec, deps, warnings) {
             if (event.ok !== undefined)
                 state.phaseRecords.push(event);
             deps.reportSubjectPhase(event, {
-                laneId: spec.planned.id,
-                laneIndex: spec.planned.index,
-                laneCount: deps.participantCount,
+                id: spec.planned.id,
+                index: spec.planned.index,
+                count: deps.participantCount,
             });
         },
     };

package/dist/routes/computer-use/e2b-desktop/fidelity.d.ts

@@ -19,7 +19,8 @@ export declare function mobileLaunchFlags(residual: Readonly<ResidualConfig>, sp
  * observation. A request that cannot be applied fails the participant closed with the reason. Only
  * participants on a mobile preset are emulated: a run-wide flag must not hand a desktop or tablet
  * participant an iPhone user agent (the first live proof did exactly that to the desktop newcomer
- * beside the phone participant). Those participants carry no fidelity block, which is honest.
+ * beside the phone participant). Those participants carry no fidelity block, since nothing was
+ * emulated.
  */
 export declare function applyParticipantMobileFidelity(args: {
     desktop: E2BDesktopSandbox;
@@ -44,8 +45,8 @@ export declare function participantBrowserStateObserver(args: {
 }): ReturnType<typeof makeChromeBrowserStateObserver>;
 /**
  * The participant's final browser geometry, measured while the sandbox is alive. A final capture that
- * measured EITHER field wins whole, so a partial final capture omits fields the launch-time capture
- * had (honest omission); only a final capture that measured NOTHING falls back to the launch-time
+ * measured either field wins whole, so a partial final capture omits fields the launch-time capture
+ * had; only a final capture that measured nothing falls back to the launch-time
  * capture. Also reads the emulation holder's log into the fidelity block.
  */
 export declare function finalParticipantGeometry(args: {

package/dist/routes/computer-use/e2b-desktop/fidelity.js

@@ -1,5 +1,5 @@
-// Device fidelity for one E2B desktop participant: mobile emulation beyond viewport size (#221), the
-// browser-state observer that watches it for drift (#623), and the final browser geometry.
+// Device fidelity for one E2B desktop participant: mobile emulation beyond viewport size, the
+// browser-state observer that watches it for drift, and the final browser geometry.
 import { redactText, toErrorMessage } from "../../../evidence/redaction.js";
 import { readDetachedLog } from "../../../substrates/detached.js";
 import { applyMobileEmulation, DEFAULT_MOBILE_USER_AGENT, makeChromeBrowserStateObserver, } from "../../../substrates/e2b/desktop-cdp.js";
@@ -28,7 +28,8 @@ function cdpEndpoint(identity, targetUrl) {
  * observation. A request that cannot be applied fails the participant closed with the reason. Only
  * participants on a mobile preset are emulated: a run-wide flag must not hand a desktop or tablet
  * participant an iPhone user agent (the first live proof did exactly that to the desktop newcomer
- * beside the phone participant). Those participants carry no fidelity block, which is honest.
+ * beside the phone participant). Those participants carry no fidelity block, since nothing was
+ * emulated.
  */
 export async function applyParticipantMobileFidelity(args) {
     const { spec, deps } = args;
@@ -63,10 +64,10 @@ export function participantBrowserStateObserver(args) {
     return makeChromeBrowserStateObserver(args.desktop, deps.requestTimeoutMs, cdpEndpoint(args.launchIdentity, args.targetUrl), {
         targetId: args.targetId,
         // Once per participant: a dark observation channel is a gap in the instrument, and the
-        // funnel's NEVER MEASURED count needs this line to explain itself (#514).
+        // funnel's "never measured" count needs this line to explain itself.
         onUnavailable: (reason) => {
             warnings.push(`Browser-state observer unavailable for participant ${spec.planned.id} (${redactText(deps.scrubKnownValues(reason))}); ` +
-                "urlIncludes/urlPathEquals/textIncludes stop conditions and task criteria are NOT being measured this session.");
+                "urlIncludes, urlPathEquals and textIncludes stop conditions and task criteria are not measured in this session.");
         },
         drift: fidelity.emulatedTargetId === undefined
             ? undefined
@@ -75,11 +76,11 @@ export function participantBrowserStateObserver(args) {
                 expectedWidth: spec.planned.device.preset.width,
                 expectTouch: fidelity.applied?.requested.touch === true,
                 onDrift: (reason) => {
-                    warnings.push(`Mobile emulation drift for participant ${spec.planned.id}: ${reason} (#623).`);
+                    warnings.push(`Mobile emulation drift for participant ${spec.planned.id}: ${reason}.`);
                 },
                 onCovered: (coveredTargetId, read) => {
                     // A later tab the page itself reported at the phone width: evidence that
-                    // the emulation followed the participant (#623), kept on the bundle.
+                    // the emulation followed the participant, kept on the bundle.
                     if (fidelity.applied === undefined)
                         return;
                     fidelity.applied = {
@@ -95,8 +96,8 @@ export function participantBrowserStateObserver(args) {
 }
 /**
  * The participant's final browser geometry, measured while the sandbox is alive. A final capture that
- * measured EITHER field wins whole, so a partial final capture omits fields the launch-time capture
- * had (honest omission); only a final capture that measured NOTHING falls back to the launch-time
+ * measured either field wins whole, so a partial final capture omits fields the launch-time capture
+ * had; only a final capture that measured nothing falls back to the launch-time
  * capture. Also reads the emulation holder's log into the fidelity block.
  */
 export async function finalParticipantGeometry(args) {
@@ -126,7 +127,7 @@ export async function finalParticipantGeometry(args) {
     ];
     warnings.push(...geometryWarnings);
     // The emulation holder's own log, after its announce line: which later targets it
-    // attached to, what it sent, and any reply that came back as an error (#623). Read while
+    // attached to, what it sent, and any reply that came back as an error. Read while
     // the sandbox is alive; the first live proof had no way to say what the holder did.
     if (fidelity.applied !== undefined && fidelity.holderName !== undefined) {
         const holderLog = await readDetachedLog(e2bShell(desktop), fidelity.holderName, deps.requestTimeoutMs).catch(() => "");

package/dist/routes/computer-use/e2b-desktop/prepare.js

@@ -24,7 +24,7 @@ export async function acquireParticipantDesktop(ctx, state) {
     const { residual, env } = deps;
     const subjectEnvNames = participantSubjectEnv(deps.subject);
     const subjectEnvValues = residual.subject.envValues ?? {};
-    // Off-app comms (#297): the base-URL env is injected at sandbox create, so the app reads it at
+    // Off-app comms: the base-URL env is injected at sandbox create, so the app reads it at
     // boot; the catch starts right after create.
     const commsEnv = participantCommsEnv(ctx.comms);
     const desktopModule = await (deps.desktopModule ?? loadE2BDesktopModule)();
@@ -39,10 +39,10 @@ export async function acquireParticipantDesktop(ctx, state) {
             metadata: {
                 ...CUA_ACTOR_LAB_PROVIDER_METADATA,
                 labId: deps.labId,
-                simId: spec.recordId,
-                laneId: spec.planned.id,
-                laneIndex: String(spec.planned.index),
-                laneCount: String(deps.participantCount),
+                recordId: spec.recordId,
+                participantId: spec.planned.id,
+                participantIndex: String(spec.planned.index),
+                participantCount: String(deps.participantCount),
             },
             // The participant's model key never enters the sandbox (the model drives from outside).
             // The subject's declared env names are provisioned here on the clone route.
@@ -129,7 +129,7 @@ export async function verifyParticipantScreen(ctx, state, desktop) {
         state.desktopGeometry = { ...state.desktopGeometry, warnings: [screenGeometry.warning] };
     }
     if (screenGeometry.error && deps.screenMismatchPolicy !== "record-evidence") {
-        state.failureCode = "HUMANISH_CUA_LAB_DEVICE_GEOMETRY";
+        state.failureCode = "HUMANISH_COMPUTER_USE_DEVICE_GEOMETRY";
         throw new Error(screenGeometry.error);
     }
     if (screenGeometry.error && screenGeometry.verified) {
@@ -149,7 +149,7 @@ export async function provisionParticipantSubject(ctx, state, desktop) {
     const { residual, subject } = deps;
     const shell = e2bShell(desktop);
     if (subject.kind === "desktop-cli") {
-        // Prepare the runtime and any declared product install, UNKEYED. With install omitted,
+        // Prepare the runtime and any declared product install, with no keys. With install omitted,
         // the participant discovers and installs the product from its public surfaces.
         await provisionDesktopCli(shell, {
             product: subject.product.name,

package/dist/routes/computer-use/e2b-desktop/start.js

@@ -7,6 +7,7 @@ import { captureDesktopBrowserGeometry } from "../../../substrates/e2b/desktop-g
 import { prepareDesktopMedia, startE2BDesktopMedia, } from "../../../substrates/e2b/desktop-media.js";
 import { startE2BDesktopRecording } from "../../../substrates/e2b/desktop-recording.js";
 import { applyParticipantMobileFidelity, mobileLaunchFlags } from "./fidelity.js";
+import { streamEvent } from "../../../lab/run-lab-homes.js";
 /** Start the declared speech worker and screen recording. A recording that cannot start is a warning. */
 export async function startParticipantMedia(ctx, state, desktop) {
     const { spec, deps, warnings } = ctx;
@@ -47,13 +48,13 @@ export async function openParticipantSurface(ctx, state, desktop) {
     if (subject.kind === "desktop-cli") {
         // A terminal window, opened the way the browser is opened on every other route: the
         // participant arrives at a desktop with the thing they were asked to use already in front
-        // of them. They can still open another from the dock — that is the point of a desktop.
+        // of them. They can still open another from the dock; that is the point of a desktop.
         await openDesktopTerminal(desktop, deps.requestTimeoutMs, subject.product.workdir);
         await desktop.wait(DESKTOP_SETTLE_MS).catch(() => undefined);
         return;
     }
     const requestedMedia = residual.execution?.desktop?.media;
-    // A declared camera (#509) is in place before the browser starts: the feed is generated or
+    // A declared camera is in place before the browser starts: the feed is generated or
     // uploaded first, and a feed that cannot be produced fails the participant closed here.
     const mediaEvidence = requestedMedia === undefined
         ? undefined
@@ -76,7 +77,7 @@ export async function openParticipantSurface(ctx, state, desktop) {
     state.browserLaunched = true;
     noteDevToolsReadiness(ctx, browserLaunch.devTools, emulationFlags.length > 0);
     await desktop.wait(DESKTOP_SETTLE_MS).catch(() => undefined);
-    // Mobile fidelity (#221) is applied outside startParticipantStream's best-effort catch, so a request
+    // Mobile fidelity is applied outside startParticipantStream's best-effort catch, so a request
     // that cannot be applied fails the participant closed.
     state.fidelity = await applyParticipantMobileFidelity({
         desktop,
@@ -91,7 +92,7 @@ export async function openParticipantSurface(ctx, state, desktop) {
 }
 /**
  * DevTools answering later than this after launch is recorded as a warning on the participant. Six live
- * launches answered in 4.8-7.7 s; the reads this wait now guards used to give up at about 11 s.
+ * launches answered in 4.8-7.7 s; without this wait the reads it guards give up at about 11 s.
  */
 const SLOW_DEVTOOLS_MS = 10_000;
 /**
@@ -168,14 +169,14 @@ export async function startParticipantStream(ctx, state, desktop) {
         });
         if (typeof candidateStreamUrl === "string" && candidateStreamUrl.trim().length > 0) {
             state.streamUrl = candidateStreamUrl;
-            await deps.onStream({
+            await deps.onStream(streamEvent({
                 type: "ready",
                 participantId: spec.planned.id,
                 sandboxId: desktop.sandboxId,
-                simId: spec.recordId,
+                recordId: spec.recordId,
                 streamId: spec.streamId,
                 url: candidateStreamUrl,
-            });
+            }));
         }
         else {
             warnings.push("Live desktop stream started but did not return a usable watch URL; Observer will fall back to screenshots.");
@@ -188,7 +189,7 @@ export async function startParticipantStream(ctx, state, desktop) {
     // never a participant finding about missing controls. Both computer-use and shared-world participants
     // use this route.
     if (state.initialBrowserGeometry?.unusable !== undefined) {
-        state.failureCode = "HUMANISH_CUA_LAB_DEVICE_GEOMETRY";
+        state.failureCode = "HUMANISH_COMPUTER_USE_DEVICE_GEOMETRY";
         throw new Error(`${state.failureCode}: ${state.initialBrowserGeometry.unusable} Participant actions were not started.`);
     }
 }

package/dist/routes/computer-use/e2b-desktop/teardown.js

@@ -7,6 +7,7 @@ import { readE2BRelease } from "../../../substrates/e2b/sandbox.js";
 import { e2bShell } from "../../../substrates/e2b/shell.js";
 import { drainCommsEvidence } from "./comms.js";
 import { finalParticipantGeometry } from "./fidelity.js";
+import { streamEvent } from "../../../lab/run-lab-homes.js";
 /**
  * Each route's own keep flag gates its own participants only: a clone.keep can never leak into a
  * local-tree participant's teardown decision, and vice versa.
@@ -145,16 +146,16 @@ export async function finishE2BDesktop(ctx, state, failed) {
         state.sandboxTornDownAtMs = deps.now();
         // The participant's live stream is now a dead page whichever teardown path ran (released, kept, or
         // release-failed-awaiting-TTL); tell the watch overlay so the tile falls back to recorded
-        // evidence instead of "sandbox not found" (#357). Guarded: a viewer callback must never
+        // evidence instead of "sandbox not found". Guarded: a viewer callback must never
         // break teardown.
         if (state.streamUrl !== undefined) {
             try {
-                await deps.onStream({
+                await deps.onStream(streamEvent({
                     type: "ended",
                     participantId: spec.planned.id,
-                    simId: spec.recordId,
+                    recordId: spec.recordId,
                     streamId: spec.streamId,
-                });
+                }));
             }
             catch {
                 // viewer-side only; nothing to record

package/dist/routes/computer-use/external-comms.d.ts

@@ -2,7 +2,7 @@ import type { LabCommsEmail, LabCommsExternal } from "../../lab/types.js";
 import type { PreparedRunArtifactPaths } from "../../run/paths.js";
 import type { DesktopParticipantRun, ParticipantRunOutcome } from "./types.js";
 /**
- * Adopter-hosted drain (#380): once per RUN, after every participant finished — the catch is one
+ * Adopter-hosted drain: once per run, after every participant finished, because the catch is one
  * shared external endpoint, not a per-sandbox file. Same routing and digest-only artifact as
  * the in-sandbox drain; the artifact is registered on every participant that declared a recipient
  * address, since the thread carries each inbox's mail. A drain failure never fails the run.

package/dist/routes/computer-use/external-comms.js

@@ -1,7 +1,7 @@
 import { collectExternalCommsEvidence } from "../../comms/external-evidence.js";
 import { participantHasInboxRecipient } from "./participant-desktop.js";
 /**
- * Adopter-hosted drain (#380): once per RUN, after every participant finished — the catch is one
+ * Adopter-hosted drain: once per run, after every participant finished, because the catch is one
  * shared external endpoint, not a per-sandbox file. Same routing and digest-only artifact as
  * the in-sandbox drain; the artifact is registered on every participant that declared a recipient
  * address, since the thread carries each inbox's mail. A drain failure never fails the run.

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

@@ -35,26 +35,26 @@ function fanoutReview(args, streams) {
     // The judge's verdict: live fan-out must prove every participant (judgeParticipants).
     const verdict = args.verdict;
     const passedParticipants = (outcomes ?? []).filter((outcome) => participantPassed(participantFactsOf(outcome))).length;
-    // What happened to the PARTICIPANTS, with the denominator attached. The verdict above has to
+    // What happened to the participants, with the denominator attached. The verdict above has to
     // collapse the run to one word; this does not (docs/principles/three-roles.md).
     const terminalOutcomes = (outcomes ?? []).filter((outcome) => outcome?.session?.status !== undefined);
     const participants = terminalOutcomes.length > 0
         ? tallyParticipantOutcomes(
-        // A NO-ENGAGEMENT participant is not one who reached the goal. It said "done" having
+        // A no-engagement participant is not one who reached the goal. It said "done" having
         // taken zero actions and said nothing, and `passedParticipants` above already refuses to
-        // count it — but `reachedGoal` was reading the trace status directly, so one run could be both
-        // "not a passed participant" AND "1/1 reached the goal". The headline number a researcher reads
+        // count it, but `reachedGoal` was reading the trace status directly, so one run could be both
+        // "not a passed participant" and "1/1 reached the goal". The headline number a researcher reads
         // first was the dishonest one. Found by a provider bug that ended a study on turn one.
         terminalOutcomes.map((outcome) => participantStatus(outcome.session.status, {
             noEngagement: outcome.noEngagement === true,
             selfReportedBlocker: outcome.selfReportedBlocker === true,
         })), 
-        // A participant who reached the goal AND told you the road there was broken is the most
+        // A participant who reached the goal and told you the road there was broken is the most
         // useful result a study produces; reporting only the outcome would bury it.
         terminalOutcomes.map((outcome) => outcome.reportedFriction === true))
         : undefined;
     // The study funnel: per-task completion rates across every session that measured one. This is
-    // "where did people get stuck" as data, next to WHO got stuck (participants) above.
+    // "where did people get stuck" as data, next to who got stuck (participants) above.
     const participantFunnels = (outcomes ?? [])
         .map((outcome) => outcome?.session?.trace.taskFunnel)
         .filter((funnel) => funnel !== undefined);
@@ -74,12 +74,12 @@ function fanoutReview(args, streams) {
         summary: args.inProgress === true
             ? `Live computer-use fan-out is running (${specs.length} participants, one world each); terminal participant evidence has not been written yet.`
             : args.dryRun
-                ? `${args.rerun ? `Rerun contract from ${args.rerun.sourceRunId}: ` : ""}Dry-run fan-out contract: ${specs.length} participants composed for ${args.descriptor.id} against ${args.appUrl}, one world each; no desktops launched, $0 spend.`
+                ? `${args.rerun ? `Rerun from ${args.rerun.sourceRunId}: ` : ""}Dry-run fan-out: ${specs.length} participants composed for ${args.descriptor.id} against ${args.appUrl}, one world each; no desktops launched, $0 spend.`
                 : `${args.rerun ? `Rerun from ${args.rerun.sourceRunId}: ` : ""}Computer-use fan-out (${specs.length} participants, one world each): ${passedParticipants}/${specs.length} participant(s) reached a terminal, engaged verdict${participants ? `: ${formatParticipantOutcomes(participants, participantEndings)}` : ""}${runTasks ? `; tasks: ${formatRunTaskFunnel(runTasks)}` : ""}.`,
         gaps: args.inProgress === true
             ? ["Live fan-out session is still running."]
             : args.dryRun
-                ? ["Live fan-out session not yet run (dry-run contract only)."]
+                ? ["Live fan-out session not yet run (dry run only)."]
                 : specs
                     .map((spec, index) => ({ spec, outcome: outcomes?.[index] }))
                     .filter(({ outcome }) => outcome === undefined ||
@@ -112,7 +112,7 @@ function fanoutDesktopBrowser(args) {
 }
 function fanoutCost(args) {
     const { specs, outcomes } = args;
-    // Run-level cost ESTIMATE: one model-token line per participant that ran a session (from its persisted
+    // Run-level cost estimate: one model-token line per participant that ran a session (from its persisted
     // trace.estimatedCost) + a desktop line per owned allocation, priced at its observed resources.
     // `per-lane-worlds` participants have no shared provisioning to double-count. Omitted on a pure dry-run.
     const costTraces = specs
@@ -218,22 +218,22 @@ export function buildCuaFanoutBundle(args) {
         persona: {
             id: specs[0].persona.id,
             name: `Computer-use fan-out (${specs.length} participants)`,
-            source: `lab:${plan.labId}`,
+            source: `study:${plan.labId}`,
             sourceDigest: specs[0].persona.promptDigest,
         },
         scenario: {
             id: `cua-${plan.labId}`,
             title: plan.title ?? `Computer-use fan-out: ${plan.labId}`,
-            // Redacted at WRITE time, like every other raw-text surface in the bundle. Participant records are
-            // digest-only by design, but scenario.goal keeps one participant's composed instructions verbatim —
+            // Redacted at write time, like every other raw-text surface in the bundle. Participant records are
+            // digest-only by design, but scenario.goal keeps one participant's composed instructions verbatim,
             // and an adopter whose authored participant text must name a runtime world URL (an inbox on a route
             // where the harness does not inject one) put an *.e2b.app address in it. That landed raw here
             // and in observer-data.json, the sensitive-text scanner matched it, and verify failed a bundle
-            // this writer produced. The only adopter-side workaround was scanner evasion (#412).
+            // this writer produced. The only adopter-side workaround was scanner evasion.
             //
             // The instructions the model actually receives are untouched; only the persisted copy changes.
             goal: redactText(specs[0].evidenceInstructions ?? specs[0].instructions),
-            source: `lab:${plan.labId}`,
+            source: `study:${plan.labId}`,
             sourceDigest: specs[0].persona.promptDigest,
         },
         lifecycle: [
@@ -251,14 +251,14 @@ export function buildCuaFanoutBundle(args) {
             status: "passed",
             notes: ranLive
                 ? anyRaw
-                    ? "Typed text recorded as length only and reasoning/messages pass through text redaction. Some participants captured FULL-FIDELITY (raw) screenshots, retained for local use and NOT redacted for publishing; set policies.redactScreenshots: true to blur a share-as-is bundle."
+                    ? "Typed text is recorded as its length only, and reasoning and messages pass through text redaction. Some participants captured unblurred screenshots, kept for local use and not redacted for publishing. Set policies.redactScreenshots: true to blur screenshots in a bundle you plan to share."
                     : "Typed text recorded as length only and reasoning/messages pass through text redaction. Screenshots are blurred at capture (policies.redactScreenshots: true) for a share-as-is bundle."
-                : "Dry-run fan-out contract bundle: no desktops launched and no screenshots captured. Typed text is recorded as length only and reasoning/messages pass through text redaction whenever a session runs.",
+                : "Dry-run fan-out bundle: no desktops launched and no screenshots captured. Typed text is recorded as length only and reasoning/messages pass through text redaction whenever a session runs.",
         },
         artifacts: bundleArtifacts(),
         review,
-        // What the participants reported, when any reported anything (#392). Dry-run and in-progress
-        // bundles carry none — there is no participant yet to quote.
+        // What the participants reported, when any reported anything. Dry-run and in-progress
+        // bundles carry none: there is no participant yet to quote.
         feedbackCandidates: fanoutFeedbackCandidates(args),
         // Selected hosted image, including the optional speech default; omitted on the stock desktop.
         ...(desktopTemplate === undefined ? {} : { desktopTemplate }),