humanish 0.107.0 → 0.108.0

package/dist/routes/computer-use/self-report.js

@@ -1,12 +1,12 @@
 import { selfReportedBlocker } from "../../run/judge.js";
-// "can't" followed by a PERCEPTION verb describes what the screen showed, not an inability to
+// "can't" followed by a perception verb describes what the screen showed, not an inability to
 // proceed: "the canvas truncates it so you can't even read the whole thing", "I can't tell from
 // the screen whether the rename is persisted", "so I could not read its full description". Five
 // of five completed live runs on 2026-09-01 (two on drawDB, three on the planted benchmark app)
 // were refused as "not a credible pass" on exactly these sentences, every one a defect report
-// written AFTER the participant reached the goal. The more precisely a participant describes a
-// display defect, the more likely the scan was to refuse the run — the incentive inversion #453
-// fixed for resolved arcs, back in a new shape. "could not complete", "could not connect",
+// written after the participant reached the goal. The more precisely a participant describes a
+// display defect, the more likely the scan was to refuse the run: the incentive inversion the
+// resolved-arc rule fixed, back in a new shape. "could not complete", "could not connect",
 // "unable to get focus" still count: those name an inability to act.
 const PERCEPTION_AFTER_MODAL = /\b(can'?t|cannot|could ?not|couldn'?t|unable to|wasn'?t able to)\s+(even\s+|quite\s+|really\s+|fully\s+)?(read|see|tell|view|make out|verify|confirm|be sure|be certain|judge|know)\b/g;
 function hasBlockerLanguage(text) {
@@ -15,17 +15,17 @@ function hasBlockerLanguage(text) {
         /\berror[:.]/.test(text) ||
         /what would you like me to do|please tell me|need (the )?(task|credentials|instructions)/.test(text));
 }
-/** The friction scan (inclusive): does the narrative report ANY blocker-shaped language,
+/** The friction scan (inclusive): does the narrative report any blocker-shaped language,
  *  resolved or not? Feeds the participants `reportedFriction` tally and feedback candidates. */
-// Report-shaped language: what a participant writes when it finished AND has something to say.
+// Report-shaped language: what a participant writes when it finished and has something to say.
 // Every one of the day's eleven drawDB reports (2026-09-01) opened a section "What confused me"
 // or "Accessibility defects:"; none contained a blocker word, so none became a feedback candidate
 // and a participant that had just replicated a keyboard-accessibility defect three times drafted "Live
-// study completed without a participant-reported finding". Friction is the INCLUSIVE scan; a
+// study completed without a participant-reported finding". Friction is the inclusive scan; a
 // false positive here adds a candidate a person then reads, which is the cheap direction.
 const REPORTED_DEFECT_LANGUAGE = /\b(defects?|bugs?|accessibilit(y|ies)|inaccessible|not (keyboard|screen.?reader)[- ]?accessible|confus(ed|ing)|hesitat(ed|ion)|unexpected(ly)?|unclear|hard to (find|tell|see|read|reach)|no (visible )?focus|overlap(ped|ping|s)?|truncat(ed|es|ion)|cut off|did nothing|nothing happened|no effect)\b/;
-// The friction scan's own negations (#614). "Nothing was confusing", "no defects", "not unclear"
-// are what a participant writes when it has NOTHING to report, and until 2026-09-03 each of them
+// The friction scan's own negations. "Nothing was confusing", "no defects", "not unclear"
+// are what a participant writes when it has nothing to report, and until 2026-09-03 each of them
 // counted as reported friction and became a feedback candidate whose "actual" was a sentence
 // reporting no problem. Only the report-shaped adjectives are negatable here: "no visible focus",
 // "not keyboard-accessible" and "did nothing" are defects and stay.
@@ -80,25 +80,25 @@ function interimMessageReportsFriction(message) {
     });
 }
 /** The verdict scan (strict): like the friction scan, but resolved-arc segments are stripped
- *  first — failure narration the participant itself reports as overcome is friction on the
- *  road, not a blocker at the destination (#453). */
+ *  first: failure narration the participant itself reports as overcome counts as friction and
+ *  never as a blocker. */
 function completionReasonBlocksVerdict(reason) {
-    // Perception phrases are stripped for the VERDICT only: "I could not read the full description"
+    // Perception phrases are stripped for the verdict only: "I could not read the full description"
     // is friction worth a tally count and a feedback candidate (the friction scan above keeps it),
     // and it is not a reason to refuse the pass.
     return hasBlockerLanguage(stripResolvedArcSegments(stripQuotedSpans(stripNegatedNonBlockerPhrases(reason.toLowerCase()))).replace(PERCEPTION_AFTER_MODAL, ""));
 }
 // A failure segment counts as a resolved arc when the recovery is self-reported either in the
-// SAME segment ("the import failed but then went through") or — the common report shape — in the
-// immediately FOLLOWING segment as a retry/alternative that succeeded ("my first import failed
+// same segment ("the import failed but then went through") or, in the common report shape, in the
+// immediately following segment as a retry/alternative that succeeded ("my first import failed
 // with a parser error. A simpler SQL import succeeded."). The lookahead demands the retry flavor
 // on purpose: unrelated praise ("Separately, the search box worked") must never launder an
 // unresolved failure. "Login failed so I gave up" has no recovery anywhere and stays a blocker.
-// (#453 — the run-1 false negative: a defect report after demonstrated success failed the participant,
+// (Without this, a defect report after demonstrated success failed the participant:
 // an incentive inversion against exactly the participant behavior a study wants most.)
 const RESOLUTION_TERMS = /\b(succeed(?:ed|s)?|success(?:ful|fully)?|worked|works around|then worked|now works?|resolved|fixed|recovered|got it working|went through)\b/;
 const RETRY_RESOLUTION = /\b(simpler|simplified|retry(?:ing)?|retried|second (?:attempt|try)|another (?:attempt|try|approach)|different (?:approach|way|route)|instead|then|eventually|after that)\b[^.!?\n]{0,80}\b(succeed(?:ed|s)?|success(?:ful|fully)?|worked|went through|completed|passed)\b/;
-/** Drop sentence/bullet segments whose failure language is part of a self-reported RESOLVED arc. */
+/** Drop sentence/bullet segments whose failure language is part of a self-reported resolved arc. */
 function stripResolvedArcSegments(text) {
     const segments = text.split(/(?<=[.!?])\s+|\n+/);
     return segments
@@ -112,15 +112,15 @@ function stripResolvedArcSegments(text) {
     })
         .join(" ");
 }
-// Negations that DESCRIBE a defect rather than deny one. Kept out of every clause drop below.
+// Negations that describe a defect rather than deny one. Kept out of every clause drop below.
 const DEFECT_SHAPED_NEGATION = /\bno\s+(?:visible\s+)?focus\b|\bno\s+(?:keyboard|screen.?reader)[- ]?(?:access|path|route|way|alternative|equivalent)|\bnot\s+(?:keyboard|screen.?reader)[- ]?accessible\b|\bno\s+(?:effect|feedback|response)\b|\b(?:did|does)\s+nothing\b|\bnothing\s+happened\b/;
 function stripNegatedNonBlockerPhrases(text) {
     return (text
-        // FIRST, before the narrower rules eat the "no blockers" and leave "encountered ... error"
+        // First, before the narrower rules eat the "no blockers" and leave "encountered ... error"
         // behind: "I encountered no blockers or unclear error output." refused a clean passing run on
         // 2026-09-01. A verb of encounter followed by "no" negates the whole clause, so drop the clause.
-        // ... unless the clause names a DEFECT: "the delete control had no visible focus" is a
-        // finding, and the verb it happens to use must not decide whether it counts (#622).
+        // ... unless the clause names a defect: "the delete control had no visible focus" is a
+        // finding, and the verb it happens to use must not decide whether it counts.
         .replace(/\b(?:encountered|hit|saw|found|met|had|got|ran into)\s+no\s+[^.!?\n]*/g, (clause) => (DEFECT_SHAPED_NEGATION.test(clause) ? clause : " "))
         .replace(/\bno\s+(?:real\s+|remaining\s+|actual\s+)?(?:blocker|blockers|blocking issue|blocking issues|error|errors|failure|failures)\s+(?:was\s+|were\s+)?(?:encountered|observed|found|hit|seen|reported|detected)\b/g, "")
         .replace(/\bwithout\s+(?:a\s+|any\s+)?(?:real\s+|remaining\s+|actual\s+)?(?:blocker|blockers|blocking issue|blocking issues|error|errors|failure|failures)\b/g, "")
@@ -134,9 +134,9 @@ function stripNegatedNonBlockerPhrases(text) {
 }
 /**
  * Remove double-quoted spans and markdown blockquote lines before the blocker scan, so a persona
- * that faithfully QUOTES the subject app's own copy (e.g. a banner reading "cannot be undone") is
- * not misread as the actor reporting its OWN blocker. Only double quotes (straight and smart) and
- * `>` blockquotes are stripped — never single quotes, which would mangle contractions like `can't`.
+ * that faithfully quotes the subject app's own copy (e.g. a banner reading "cannot be undone") is
+ * not misread as the actor reporting its own blocker. Only double quotes (straight and smart) and
+ * `>` blockquotes are stripped, never single quotes, which would mangle contractions like `can't`.
  */
 function stripQuotedSpans(text) {
     return text
@@ -148,15 +148,15 @@ function traceHasStopWhenMatch(session) {
     return session.trace.items.some((item) => item.kind === "notice" &&
         item.status === "matched" &&
         // A dwell window that ended the session (then: stop) is the same class of harness-owned,
-        // structured completion as a matched stopWhen (#510).
+        // structured completion as a matched stopWhen.
         (item.title.startsWith("stopWhen matched") || item.title === "dwell window complete"));
 }
 /**
- * A goal_satisfied participant counts as a self-reported blocker ONLY when its final narrative contradicts
- * the goal AND the run's own stop predicate did NOT fire. A matched stopWhen is independent,
- * structured completion evidence, so it overrides a text scan of the free-form narrative — which can
- * otherwise trip on the subject app's OWN quoted copy (e.g. a relayed "cannot be undone" banner).
- * Resolved-arc segments never block the verdict (#453). Returns the offending reason, or undefined
+ * A goal_satisfied participant counts as a self-reported blocker only when its final narrative contradicts
+ * the goal and the run's own stop predicate did not fire. A matched stopWhen is independent,
+ * structured completion evidence, so it overrides a text scan of the free-form narrative, which can
+ * otherwise trip on the subject app's own quoted copy (e.g. a relayed "cannot be undone" banner).
+ * Resolved-arc segments never block the verdict. Returns the offending reason, or undefined
  * when the participant is a clean pass. Exported for testing.
  */
 export function resolveSelfReportedBlocker(session) {
@@ -187,10 +187,10 @@ export function sessionEnding(session) {
     };
 }
 /**
- * Friction is independent of how a completed session ended (#657). Read the participant's
+ * Friction is independent of how a completed session ended. Read the participant's
  * redacted messages, including earlier reports, rather than the harness-owned reason that
  * stopWhen/dwell writes. Reasoning, observations, and notices are not participant reports.
- * Resolved arcs still count (#453); quoted copy and negated reports still do not. This read
+ * Resolved arcs still count; quoted copy and negated reports still do not. This read
  * never changes the verdict. Exported for testing.
  */
 export function resolveSelfReportedFriction(session) {

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

@@ -106,7 +106,7 @@ export declare function admitCuaRun(plan: ComputerUsePlan, input: ComputerUseRun
             result: CuaActorLabResult;
         };
         descriptor: CuaActorDescriptor;
-        runSession: (options: import("../../actors/computer-use/actor.js").CuaActorSessionOptions) => Promise<import("../../index.js").CuaLoopResult>;
+        runSession: (options: import("../../actors/computer-use/actor.js").CuaActorSessionOptions) => Promise<import("../../index.js").ComputerUseLoopResult>;
         externalCommsConfig: LabCommsExternal | undefined;
         externalCommsEmail: LabCommsEmail | undefined;
         participantRuns: DesktopParticipantRun[];

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

@@ -17,13 +17,14 @@ import { e2bRequestTimeoutMs } from "../../substrates/e2b/lifetime.js";
 import { liveCuaRejection } from "./preflight.js";
 import { cuaDescriptorOf, declaredAppUrl, plannedAppUrl } from "./plan.js";
 import { trackRuntimeStreams } from "./live-flush.js";
-import { participantOf, phaseEvent, planEvent } from "../../lab/run-lab-events.js";
+import { phaseEvent, planEvent } from "../../lab/run-lab-events.js";
 import { defaultSubjectPhaseSink } from "../../subject/steps.js";
 import {} from "./bundle.js";
 import { packRunLocalTree } from "./local-tree-pack.js";
 import { projectParticipantSubjects, subjectProvenanceArg } from "./subject-projection.js";
-import { CUA_ACTOR_LAB_SCHEMA, participantSubjectEnv, } from "./types.js";
+import { participantSubjectEnv, } from "./types.js";
 import { labPersonaIds } from "../../lab/persona-resolve.js";
+import { studyResultIdentity } from "../../run/study-result.js";
 /** The physical project, bound before any caller hook runs. */
 async function bindProject(cwd) {
     // Capture the physical project before reading or invoking any caller hook. A supported
@@ -43,10 +44,9 @@ export async function refuseCuaLab(options, refusal) {
     if (refusal.stage === "after-personas")
         await compileParticipantPersonas(projectRoot, labPersonaIds(config));
     return {
-        schema: CUA_ACTOR_LAB_SCHEMA,
+        ...studyResultIdentity("computer-use", config.id),
         ok: false,
         cwd: projectRoot.physicalPath,
-        labId: config.id,
         actor: refusal.actor ?? config.actors[0]?.type ?? "",
         appUrl: declaredAppUrl(config),
         dryRun,
@@ -74,10 +74,9 @@ export async function admitCuaRun(plan, input, config) {
     const appUrl = plannedAppUrl(subject);
     const subjectEnvNames = [...participantSubjectEnv(subject)];
     const fail = (code, message, actorLabel) => ({
-        schema: CUA_ACTOR_LAB_SCHEMA,
+        ...studyResultIdentity("computer-use", plan.labId),
         ok: false,
         cwd,
-        labId: plan.labId,
         actor: actorLabel ?? plan.actor,
         appUrl,
         dryRun,
@@ -96,7 +95,7 @@ export async function admitCuaRun(plan, input, config) {
     if (config.subject.source === "app-url" &&
         config.execution?.target === "local" &&
         input.localVm === undefined)
-        return refuse("HUMANISH_CUA_LAB_LOCAL_DESKTOP_MISSING", "An app-url lab with execution.target: local needs a local desktop. runLab starts one; a direct route call or an in-process executor does not.");
+        return refuse("HUMANISH_COMPUTER_USE_LOCAL_DESKTOP_MISSING", "An app-url lab with execution.target: local needs a local desktop. runLab starts one; a direct route call or an in-process executor does not.");
     const descriptor = cuaDescriptorOf(plan.actor);
     const baseRunSession = seams.runSession ?? descriptor.runSession;
     // A local VM study's sessions abort on its signal.
@@ -104,11 +103,11 @@ export async function admitCuaRun(plan, input, config) {
     const runSession = localVmSignal === undefined
         ? baseRunSession
         : (options) => baseRunSession({ ...options, signal: localVmSignal });
-    // Adopter-hosted comms plane on the app-url route (#380): humanish provisions no subject here,
-    // so it cannot host a catch — the OPERATOR runs one, and humanish still does every other part
+    // Adopter-hosted comms plane on the app-url route: humanish provisions no subject here,
+    // so it cannot host a catch. The operator runs one, and humanish still does every other part
     // of the funnel: tells each persona its address and inbox URL, drains the catch over HTTP after
     // the participants, and writes the same digest-only evidence. Declaring `external` previously did
-    // nothing on this route (and, per #387, on every other) while its docs said otherwise.
+    // nothing on this route (or on any other) while its docs said otherwise.
     const comms = plan.residual.comms;
     const externalCommsConfig = subject.kind !== "clone" && subject.kind !== "local-tree" && !inProcess
         ? comms?.email?.external
@@ -120,7 +119,7 @@ export async function admitCuaRun(plan, input, config) {
     }
     const { participantRuns, participantPlan, rerunLineage } = participants;
     const participantCount = participantRuns.length;
-    // Pre-flight plan: BEFORE any sandbox or provider call (dry-run AND live). onEvent gets it for
+    // Pre-flight plan: before any sandbox or provider call (dry-run and live). onEvent gets it for
     // every N; the stderr table prints for fan-out (N>1) so single-participant runs stay as quiet as they
     // always were.
     if (participantCount > 1) {
@@ -141,7 +140,7 @@ export async function admitCuaRun(plan, input, config) {
     sanitizeParticipantRuns(participantRuns, scrubKnownValues);
     const redactRepoLabel = plan.residual.policies?.redactRepos ?? subjectEnvNames.includes("GITHUB_TOKEN");
     const publicRepo = subject.kind === "clone" ? (redactRepoLabel ? "repo-01" : subject.repo) : undefined;
-    // Key-gating is route-aware: the in-process route uses the caller's OWN model + executor, and
+    // Key-gating is route-aware: the in-process route uses the caller's own model + executor, and
     // the local-agent route uses a CLI the operator has already signed in to.
     if (!dryRun && !inProcess) {
         const rejection = await liveCuaRejection({
@@ -154,8 +153,8 @@ export async function admitCuaRun(plan, input, config) {
         if (rejection)
             return refuse(rejection.code, rejection.message, descriptor.id);
     }
-    // Pack the working tree ONCE per run, on the host, BEFORE any sandbox or provider call: every
-    // fan-out participant below uploads this SAME archive, so one archiveSha256 describes every
+    // Pack the working tree once per run, on the host, before any sandbox or provider call: every
+    // fan-out participant below uploads this same archive, so one archiveSha256 describes every
     // participant's digest. Dry-run packs nothing (no fs side effects; the contract bundle carries no
     // archiveSha256). A packing failure fails the run closed here, before any sandbox is
     // created and before the run directory exists.
@@ -168,7 +167,7 @@ export async function admitCuaRun(plan, input, config) {
             localTreeArchiveBuffer = packed.buffer;
         }
         catch (error) {
-            return refuse("HUMANISH_CUA_LAB_SUBJECT_INVALID", `local-tree packing failed: ${redactText(scrubKnownValues(toErrorMessage(error)))}`, descriptor.id);
+            return refuse("HUMANISH_COMPUTER_USE_SUBJECT_INVALID", `local-tree packing failed: ${redactText(scrubKnownValues(toErrorMessage(error)))}`, descriptor.id);
         }
     }
     return {
@@ -229,7 +228,7 @@ export async function startCuaRun(plan, input, admitted, scope) {
         humanishSource: "present",
         packageName: "humanish",
     });
-    // Live-trace flush seam (#441): runLabParticipants fills it when a live run has an in-progress bundle
+    // Live-trace flush seam: runLabParticipants fills it when a live run has an in-progress bundle
     // to grow; participants call it through deps.onTrace. It exists before deps so deps can reference it as
     // a stable indirection.
     const liveTrace = {};
@@ -334,7 +333,7 @@ function cuaParticipantDeps(plan, input, admitted, run) {
         redactScreenshots,
         scrubKnownValues,
         runSession,
-        // The study-level ledger exists once per RUN, shared by every participant (#299). Dry runs never
+        // The study-level ledger exists once per run, shared by every participant. Dry runs never
         // spend, so they carry none.
         ...(dryRun || plan.caps.maxTotalUsd === undefined
             ? {}
@@ -374,7 +373,7 @@ function subjectPhaseReporter(input) {
     const sink = input.deps?.subjectPhaseSink ?? defaultSubjectPhaseSink;
     return (event, participant) => {
         sink(event, participant);
-        input.emit?.(phaseEvent(event, { kind: "participant", participant: participantOf(participant) }));
+        input.emit?.(phaseEvent(event, { kind: "participant", participant }));
     };
 }
 function makeCuaRunId() {

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

@@ -49,15 +49,15 @@ export declare function buildSingleParticipantBundle(args: {
     /**
      * Capture-time screenshot policy ("blurred" when policies.redactScreenshots, else "raw").
      * When a session ran, its trace's `redaction.screenshots` is the evidence-of-record and
-     * wins; this fallback keeps labels honest for frames written before a mid-session failure
-     * (no trace exists to testify then). Defaults to "raw" — the engine default.
+     * wins; this fallback labels the frames written before a mid-session failure
+     * (no trace exists to testify then). Defaults to "raw", the engine default.
      */
     captureRedaction?: "raw" | "blurred";
     session?: CuaLoopResult;
     sessionError?: string;
     /**
-     * The route's own credibility read of a goal_satisfied session (#476). The actor's status is
-     * evidence of what it CLAIMED; whether the harness counts the claim is decided by the route
+     * The route's own credibility read of a goal_satisfied session. The actor's status is
+     * evidence of what it claimed; whether the harness counts the claim is decided by the route
      * (zero engagement, a final message that describes a blocker). The review has to say the same
      * thing the participant's exit code says, or the durable bundle reports a participant reaching the
      * goal on a run the harness refused to count.
@@ -72,9 +72,9 @@ export declare function buildSingleParticipantBundle(args: {
      * + digests only, never values or command text), including the subject's state story. */
     subjectProvenance?: CuaSubjectProvenanceArg;
     /**
-     * Entry kind for the non-clone subject.declared event (invariant 5 — declare what the subject
-     * WAS). "local-app": an already-running LOCAL dev server driven in-process, un-pinnable —
-     * declared honestly as caller-provisioned/unpinned with no E2B. Absent: a plain app-url entry.
+     * Entry kind for the non-clone subject.declared event, which declares what the subject
+     * was. "local-app": an already-running local dev server driven in-process, un-pinnable, and
+     * declared as caller-provisioned/unpinned with no E2B. Absent: a plain app-url entry.
      */
     entryKind?: "local-app";
     /** The custom E2B desktop template (image) this participant launched on, when configured (provenance). */

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

@@ -24,7 +24,7 @@ function browserPlace(args) {
             return "in a hosted desktop browser";
     }
 }
-/** Run-level cost ESTIMATE (advisory; omitted when nothing was priced and no sandbox ran). */
+/** Run-level cost estimate (advisory; omitted when nothing was priced and no sandbox ran). */
 function runCost(args) {
     return buildRunCostSummary({
         participants: args.session
@@ -52,15 +52,15 @@ function participantView(args, publicAppUrl) {
         ? "Live computer-use session is running; stream auth URL is available only through the attached Observer server."
         : (args.session?.reason ??
             args.sessionError ??
-            "Contract bundle only: dry-run produced the evidence shape without launching a desktop or spending provider tokens.");
+            "Dry run: the evidence shape was written without launching a desktop or spending provider tokens.");
     const lastScreenshot = args.screenshots[args.screenshots.length - 1];
     const desktopGeometry = args.desktopRoute === false
         ? undefined
         : (args.desktopGeometry ?? {
             screen: { requested: { width: args.resolution[0], height: args.resolution[1] } },
         });
-    // Honest labels (invariant 6: claims match mechanism): every screenshot label names the
-    // run's ACTUAL mode. The session trace is the evidence-of-record; the capture policy covers
+    // Claims match mechanism: every screenshot label names the
+    // run's actual mode. The session trace is the evidence-of-record; the capture policy covers
     // frames written before a mid-session failure produced a trace.
     const traceScreenshotMode = args.session?.trace.redaction.screenshots;
     const screenshotMode = traceScreenshotMode === "raw" || traceScreenshotMode === "blurred"
@@ -92,7 +92,7 @@ function singleSimulation(args, view) {
                 ? `Computer-use actor (${args.actorId}) is driving the subject app ${browserPlace(args)}.`
                 : args.sessionError !== undefined
                     ? `Computer-use lab failed before a terminal session verdict: ${args.sessionError}`
-                    : `Contract participant for the computer-use actor (${args.actorId}) against ${publicAppUrl}.`,
+                    : `Computer-use actor (${args.actorId}) against ${publicAppUrl}; no session ran.`,
         startedAt: args.run.createdAt,
         updatedAt: args.run.createdAt,
     });
@@ -107,13 +107,13 @@ function singleStream(args, view) {
         ...(args.surface === undefined ? {} : { surface: args.surface }),
         ...(args.caseGroup === undefined ? {} : { caseGroup: args.caseGroup }),
         kind: "browser",
-        label: `CUA browser — ${args.labId}`,
+        label: `${args.participantId ?? "lane-01"} · browser`,
         status,
         transport: "snapshot",
         updatedAt: args.run.createdAt,
         embed: lastScreenshot
-            ? { kind: "screenshot", url: lastScreenshot, title: `CUA desktop (${screenshotMode})` }
-            : { kind: "placeholder", title: "CUA desktop" },
+            ? { kind: "screenshot", url: lastScreenshot, title: `Desktop (${screenshotMode})` }
+            : { kind: "placeholder", title: "Desktop" },
         ...(desktopGeometry?.viewport === undefined
             ? {}
             : {
@@ -172,7 +172,7 @@ function singleEvents(args, view) {
                 at: args.run.createdAt,
                 level: "info",
                 type: "cua-lab.subject.provenance",
-                // HONEST WORDING: claim "cloned/packed and served" only when it actually happened.
+                // Claim "cloned/packed and served" only when it actually happened.
                 message: `${subjectProvenanceMessage(args.subjectProvenance, publicAppUrl, args.dryRun, args.session !== undefined)} (subject env names: ${args.subjectProvenance.envNames.length > 0 ? args.subjectProvenance.envNames.join(", ") : "none"}; values never persisted); state: ${describeSubjectState(args.subjectProvenance.state, args.dryRun)}.`,
             })
             : participantEvent(SINGLE, {
@@ -180,13 +180,13 @@ function singleEvents(args, view) {
                 at: args.run.createdAt,
                 level: "info",
                 type: "cua-lab.subject.declared",
-                // Invariant 5: declare what the subject WAS, including the ABSENCE of a pin. A
-                // local-app / in-process subject is an already-running LOCAL dev server the caller
-                // provisioned; it cannot be commit-pinned, so its provenance is honestly UNPINNED and
+                // Declare what the subject was, including the absence of a pin. A
+                // local-app / in-process subject is an already-running local dev server the caller
+                // provisioned; it cannot be commit-pinned, so its provenance is unpinned and
                 // no E2B desktop was created. A plain app-url entry runs inside the desktop sandbox, or,
                 // on the local VM, reaches the host's loopback from the VM's browser.
                 message: args.entryKind === "local-app"
-                    ? `Subject app declared at ${publicAppUrl} (already-running LOCAL dev server driven in-process; NO clone, NO E2B desktop). Provenance: caller-provisioned and UNPINNED — a running dev server cannot be commit-pinned.`
+                    ? `Subject app declared at ${publicAppUrl}: a running local dev server driven in this process, with no clone and no E2B desktop. Provenance: provided by the caller and unpinned, because a running dev server has no commit to pin.`
                     : runnerSubstrate(args) === "local-desktop"
                         ? `Subject app declared at ${publicAppUrl} (the host's loopback, opened from a browser on a local VM).`
                         : `Subject app declared at ${publicAppUrl} (loopback inside the desktop sandbox).`,
@@ -220,12 +220,12 @@ function singleEvents(args, view) {
                         at: args.run.createdAt,
                         level: "info",
                         type: "cua-lab.contract.ready",
-                        message: "Dry-run contract bundle ready; switch scenario.mode to live for a real desktop session.",
+                        message: "Dry-run bundle ready; switch scenario.mode to live for a real desktop session.",
                     }),
     ];
     const record = (event) => events.push(participantEvent(SINGLE, event));
     // Persisted phase trail (real boot timing, not just a coarse provenance sentence): one
-    // RunEvent per COMPLETED phase boundary (started events never persist here; they carry no
+    // RunEvent per completed phase boundary (started events never persist here; they carry no
     // durationMs). ok:false phases warn rather than error, since the failing phase's own thrown
     // error already becomes the terminal cua-lab.session.error event above.
     let phaseEventSeq = 3;
@@ -251,17 +251,15 @@ function singleEvents(args, view) {
 }
 function singleReview(args, view, stream) {
     const { reason } = view;
-    // A funnel with a denominator of one is still the funnel — and its absence stays honest: no
+    // A funnel with a denominator of one is still the funnel, and its absence stays recorded: no
     // declared protocol (or a dry run) means no `tasks` field, never an empty one.
     const singleRunTasks = args.inProgress !== true && args.session?.trace.taskFunnel !== undefined
         ? aggregateTaskFunnels([args.session.trace.taskFunnel])
         : undefined;
-    // What happened to the participant, as the route judged it — the same rule the fan-out roll-up
-    // applies (participantStatusForOutcome). Before #476 this read the actor's own status, so a
-    // run the route refused as "not a credible pass" was written up as verdict pass, 1/1 reached
-    // the goal, and every projection of the bundle (Observer tally, `runs`, the status index)
-    // repeated it. Found on a real drawDB run whose participant wrote "Blocked after partial
-    // completion".
+    // What happened to the participant, as the route judged it: the same rule the fan-out roll-up
+    // applies (participantStatus in src/run/judge.ts). Reading the actor's own status instead would write up
+    // a run the route refused as "not a credible pass" as verdict pass, 1/1 reached the goal, and
+    // every projection of the bundle (Observer tally, `runs`, the status index) would repeat it.
     const participantStatus = args.session === undefined
         ? undefined
         : participantStatusFor(args.session.status, args.credibility);
@@ -286,7 +284,7 @@ function singleReview(args, view, stream) {
             ? []
             : args.inProgress === true
                 ? ["Live desktop session is still running."]
-                : ["Live desktop session not yet run (dry-run contract only)."],
+                : ["Live desktop session not yet run (dry run only)."],
     }, [stream]);
     return review;
 }
@@ -314,14 +312,14 @@ export function buildSingleParticipantBundle(args) {
         persona: {
             id: args.persona.id,
             name: `Computer-use operator (${args.persona.id})`,
-            source: `lab:${args.labId}`,
+            source: `study:${args.labId}`,
             sourceDigest: args.persona.promptDigest,
         },
         scenario: {
             id: `cua-${args.labId}`,
             title: args.labTitle ?? `Computer-use lab: ${args.labId}`,
             goal: redactText(args.mission),
-            source: `lab:${args.labId}`,
+            source: `study:${args.labId}`,
             sourceDigest: args.persona.promptDigest,
         },
         lifecycle: [
@@ -337,7 +335,7 @@ export function buildSingleParticipantBundle(args) {
         redaction: {
             status: "passed",
             notes: traceScreenshotMode === "raw"
-                ? "Typed text recorded as length only and reasoning/messages pass through text redaction. Screenshots are FULL-FIDELITY (raw), retained for local use — 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. Screenshots are unblurred and kept for local use; they are not redacted for publishing. Set policies.redactScreenshots: true to blur screenshots in a bundle you plan to share."
                 : traceScreenshotMode === "blurred"
                     ? "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."
                     : args.screenshots.length > 0
@@ -346,8 +344,8 @@ export function buildSingleParticipantBundle(args) {
         },
         artifacts: bundleArtifacts(),
         review,
-        // What the participant reported, when it reported anything (#392). Dry-run and in-progress
-        // bundles carry none — there is no participant yet to quote.
+        // What the participant reported, when it reported anything. Dry-run and in-progress
+        // bundles carry none: there is no participant yet to quote.
         feedbackCandidates: args.dryRun || args.inProgress === true
             ? []
             : participantFeedbackCandidates({
@@ -378,8 +376,8 @@ export function buildSingleParticipantBundle(args) {
         ...(args.providerResources === undefined || args.providerResources.length === 0
             ? {}
             : { providerResources: args.providerResources }),
-        // Structured subject provenance (invariant 5): code pin + state story. Uniform and
-        // honest on app-url bundles too — the caller minted the URL, its state is the caller's.
+        // Structured subject provenance: code pin + state story. Uniform, and
+        // stated on app-url bundles too: the caller minted the URL, its state is the caller's.
         // CuaSubjectProvenanceArg's two variants (clone, local-tree) are already RunSubjectProvenance-
         // shaped, so no reconstruction is needed beyond the app-url fallback.
         subject: args.subjectProvenance ?? { source: "app-url", state: { provenance: "undeclared" } },

package/dist/routes/computer-use/subject-projection.d.ts

@@ -18,7 +18,7 @@ export declare function projectParticipantSubjects(args: {
 }): CuaSubjectProjection[];
 /**
  * Resolve the bundle's state marker from the declaration and what actually ran.
- * Precedence: external declared → "unpinned" (seed records, if any, stay attached — a
+ * Precedence: external declared → "unpinned" (seed records, if any, stay attached; a
  * migrated external DB is still unpinned overall); else seed declared → "seeded" only when
  * every declared step executed ok on a live run, otherwise "declared-not-run" (dry-run
  * contract bundles and failed live provisioning); no declaration → "undeclared".

package/dist/routes/computer-use/subject-projection.js

@@ -36,7 +36,7 @@ export function projectParticipantSubjects(args) {
 }
 /**
  * Resolve the bundle's state marker from the declaration and what actually ran.
- * Precedence: external declared → "unpinned" (seed records, if any, stay attached — a
+ * Precedence: external declared → "unpinned" (seed records, if any, stay attached; a
  * migrated external DB is still unpinned overall); else seed declared → "seeded" only when
  * every declared step executed ok on a live run, otherwise "declared-not-run" (dry-run
  * contract bundles and failed live provisioning); no declaration → "undeclared".
@@ -48,8 +48,8 @@ export function resolveSubjectState(args) {
     }
     const declaredSeed = declared.seed ?? [];
     const external = declared.external ?? [];
-    // Dry-run: nothing executes (no sandbox) — record the DECLARED recipe: name, phase, and
-    // command digest only, with NO execution fields.
+    // Dry-run: nothing executes (no sandbox), so record the declared recipe: name, phase, and
+    // command digest only, with no execution fields.
     const seed = args.dryRun
         ? declaredSeed.map((step) => ({
             name: step.name,
@@ -74,8 +74,8 @@ export function resolveSubjectState(args) {
         ...(external.length > 0 ? { externalEnvNames: external } : {}),
     };
 }
-/** Build the per-participant subject projection (invariant 5). Local-tree participants all share
- *  ONE host-packed archive, so every participant's projection carries the identical archiveSha256/
+/** Build the per-participant subject projection. Local-tree participants all share
+ *  one host-packed archive, so every participant's projection carries the identical archiveSha256/
  *  commit/dirty (no divergence is possible, unlike the clone route's per-participant
  *  in-sandbox commit). */
 function participantSubjectProjection(args) {
@@ -144,7 +144,7 @@ export function aggregateCuaSubject(args) {
     const unanimous = !dryRun && commits.length === participantCount && new Set(commits).size === 1;
     const warnings = !dryRun && participantCount > 1 && new Set(commits).size > 1
         ? [
-            "Fan-out participants resolved DIVERGENT subject commits, so the top-level subject.commit is omitted; see each participant's provenance in result.lanes for its pinned commit.",
+            "Participants resolved different subject commits, so the top-level subject.commit is omitted; each participant's provenance in result.lanes has its pinned commit.",
         ]
         : [];
     return {