humanish 0.106.0 → 0.107.0

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

@@ -3,8 +3,7 @@ import type { SubjectPhaseEvent } from "../../subject/steps.js";
 import type { DesktopBrowserEvidence } from "../../substrates/e2b/desktop-browser.js";
 import type { ActorPersonaRef } from "../../actors/contract.js";
 import type { CuaLoopResult } from "../../actors/computer-use/loop.js";
-import { type RunLabProvenance } from "../../run/status.js";
-import { type RunBundle, type RunFeedbackCandidate, type RunProviderResource } from "../../run/bundle.js";
+import { type BundleRun, type RunBundle, type RunFeedbackCandidate, type RunProviderResource } from "../../run/bundle.js";
 import { type RunDesktopGeometry, type RunStream } from "../../run/streams.js";
 import { type Verdict } from "../../run/judge.js";
 import { type DesktopUsage } from "../../run/cost-summary.js";
@@ -17,8 +16,8 @@ export declare function buildSingleParticipantBundle(args: {
     /** The run's verdict, from the judge. */
     verdict: Verdict;
     realEmail?: boolean;
-    /** Lab provenance for the bundle's own `lab` field (#455). */
-    lab?: RunLabProvenance;
+    /** The run this bundle belongs to; the bundle head reads its id, mode, start and lab. */
+    run: BundleRun;
     actorId: string;
     appUrl: string;
     /** The participant's plan id, saved as the stream's laneId ("lane-01" when absent). */
@@ -26,7 +25,6 @@ export declare function buildSingleParticipantBundle(args: {
     actorType?: string;
     surface?: string;
     caseGroup?: string;
-    createdAt: string;
     dryRun: boolean;
     labId: string;
     labTitle?: string;
@@ -44,7 +42,6 @@ export declare function buildSingleParticipantBundle(args: {
     /** Device-preset touch metadata echoed on the measured stream viewport (a prompt signal on
      *  this route, never a rendered claim); the measured width/height/DPR stay authoritative. */
     isMobile?: boolean;
-    runId: string;
     screenshots: string[];
     /** Relative run-dir path of the digest-only comms-thread evidence artifact (humanish.comms-thread.v1),
      *  when a comms lab captured mail; registered as a "log" stream artifact. */
@@ -59,10 +56,10 @@ export declare function buildSingleParticipantBundle(args: {
     session?: CuaLoopResult;
     sessionError?: string;
     /**
-     * The lane'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 lane
+     * 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
      * (zero engagement, a final message that describes a blocker). The review has to say the same
-     * thing the lane's exit code says, or the durable bundle reports a participant reaching the
+     * 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.
      */
     credibility?: {
@@ -80,7 +77,7 @@ export declare function buildSingleParticipantBundle(args: {
      * declared honestly as caller-provisioned/unpinned with no E2B. Absent: a plain app-url entry.
      */
     entryKind?: "local-app";
-    /** The custom E2B desktop template (image) this lane launched on, when configured (provenance). */
+    /** The custom E2B desktop template (image) this participant launched on, when configured (provenance). */
     desktopTemplate?: string;
     /** The configured browser choice and the command that opened, when explicitly configured. */
     desktopBrowser?: DesktopBrowserEvidence;
@@ -90,7 +87,7 @@ export declare function buildSingleParticipantBundle(args: {
     /** Completed subject-phase records (clone/upload/extract/install/build/ready/state groups)
      *  to fold into bundle.events, so run.json carries real phase timing after the fact. */
     phaseEvents?: SubjectPhaseEvent[];
-    /** Host-side E2B desktop billed span for this lane, in minutes (from ParticipantRunOutcome
+    /** Host-side E2B desktop billed span for this participant, in minutes (from ParticipantRunOutcome
      *  desktopDurationMs). Absent when no sandbox ran (in-process/dry-run) → no desktop cost line. */
     desktopMinutes?: number;
     desktopUsage?: DesktopUsage;

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

@@ -1,6 +1,5 @@
 import { participantAssignment } from "../../lab/participant-assignment.js";
 import { redactText } from "../../evidence/redaction.js";
-import {} from "../../run/status.js";
 import { REVIEW_SCHEMA, bundleArtifacts, bundleHead, } from "../../run/bundle.js";
 import {} from "../../run/streams.js";
 import { aggregateTaskFunnels, tallyParticipantOutcomes, withCuaReviewProvenance, } from "../../run/outcomes.js";
@@ -94,8 +93,8 @@ function singleSimulation(args, view) {
                 : 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}.`,
-        startedAt: args.createdAt,
-        updatedAt: args.createdAt,
+        startedAt: args.run.createdAt,
+        updatedAt: args.run.createdAt,
     });
 }
 function singleStream(args, view) {
@@ -111,7 +110,7 @@ function singleStream(args, view) {
         label: `CUA browser — ${args.labId}`,
         status,
         transport: "snapshot",
-        updatedAt: args.createdAt,
+        updatedAt: args.run.createdAt,
         embed: lastScreenshot
             ? { kind: "screenshot", url: lastScreenshot, title: `CUA desktop (${screenshotMode})` }
             : { kind: "placeholder", title: "CUA desktop" },
@@ -162,7 +161,7 @@ function singleEvents(args, view) {
     const events = [
         {
             id: "event-000-created",
-            at: args.createdAt,
+            at: args.run.createdAt,
             level: "info",
             type: "cua-lab.run.created",
             message: `Created computer-use lab run for ${args.labId} (actor ${args.actorId}).`,
@@ -170,7 +169,7 @@ function singleEvents(args, view) {
         args.subjectProvenance
             ? participantEvent(SINGLE, {
                 id: "event-001-subject",
-                at: args.createdAt,
+                at: args.run.createdAt,
                 level: "info",
                 type: "cua-lab.subject.provenance",
                 // HONEST WORDING: claim "cloned/packed and served" only when it actually happened.
@@ -178,7 +177,7 @@ function singleEvents(args, view) {
             })
             : participantEvent(SINGLE, {
                 id: "event-001-subject",
-                at: args.createdAt,
+                at: args.run.createdAt,
                 level: "info",
                 type: "cua-lab.subject.declared",
                 // Invariant 5: declare what the subject WAS, including the ABSENCE of a pin. A
@@ -195,7 +194,7 @@ function singleEvents(args, view) {
         args.session
             ? participantEvent(SINGLE, {
                 id: "event-002-session",
-                at: args.createdAt,
+                at: args.run.createdAt,
                 level: args.session.status === "passed" ? "info" : "warn",
                 type: `cua-lab.session.${args.session.completionReason}`,
                 message: `${args.session.status}: ${args.session.reason}`,
@@ -203,7 +202,7 @@ function singleEvents(args, view) {
             : args.inProgress === true
                 ? participantEvent(SINGLE, {
                     id: "event-002-running",
-                    at: args.createdAt,
+                    at: args.run.createdAt,
                     level: "info",
                     type: "cua-lab.session.running",
                     message: "Live computer-use session is running; terminal evidence has not been written yet.",
@@ -211,14 +210,14 @@ function singleEvents(args, view) {
                 : args.sessionError !== undefined
                     ? participantEvent(SINGLE, {
                         id: "event-002-session",
-                        at: args.createdAt,
+                        at: args.run.createdAt,
                         level: "error",
                         type: "cua-lab.session.error",
                         message: args.sessionError,
                     })
                     : participantEvent(SINGLE, {
                         id: "event-002-contract",
-                        at: args.createdAt,
+                        at: args.run.createdAt,
                         level: "info",
                         type: "cua-lab.contract.ready",
                         message: "Dry-run contract bundle ready; switch scenario.mode to live for a real desktop session.",
@@ -242,7 +241,7 @@ function singleEvents(args, view) {
     for (const warning of desktopGeometry?.warnings ?? []) {
         record({
             id: `event-${String(phaseEventSeq++).padStart(3, "0")}-geometry-warning`,
-            at: args.createdAt,
+            at: args.run.createdAt,
             level: "warn",
             type: "cua-lab.geometry.warning",
             message: warning,
@@ -254,12 +253,12 @@ 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
     // declared protocol (or a dry run) means no `tasks` field, never an empty one.
-    const singleStudyTasks = args.inProgress !== true && args.session?.trace.taskFunnel !== undefined
+    const singleRunTasks = args.inProgress !== true && args.session?.trace.taskFunnel !== undefined
         ? aggregateTaskFunnels([args.session.trace.taskFunnel])
         : undefined;
-    // What happened to the participant, as the LANE judged it — the same rule the fan-out roll-up
+    // 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 lane refused as "not a credible pass" was written up as verdict pass, 1/1 reached
+    // 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".
@@ -274,14 +273,14 @@ function singleReview(args, view, stream) {
     const review = withCuaReviewProvenance({
         schema: REVIEW_SCHEMA,
         verdict: args.verdict,
-        // One lane is still a study with a denominator of one, and saying so keeps a single-lane
-        // result from being read as though it generalized.
+        // One participant is still a study with a denominator of one, and saying so keeps a
+        // single-participant result from being read as though it generalized.
         ...(participantStatus !== undefined && args.inProgress !== true
             ? {
                 participants: tallyParticipantOutcomes([participantStatus], [args.credibility?.reportedFriction === true]),
             }
             : {}),
-        ...(singleStudyTasks === undefined ? {} : { tasks: singleStudyTasks }),
+        ...(singleRunTasks === undefined ? {} : { tasks: singleRunTasks }),
         summary: credibilityNote === undefined ? reason : `${credibilityNote} ${reason}`,
         gaps: args.session || args.sessionError !== undefined
             ? []
@@ -305,15 +304,11 @@ export function buildSingleParticipantBundle(args) {
     const events = singleEvents(args, view);
     const review = singleReview(args, view, stream);
     return {
-        ...bundleHead({
+        ...bundleHead(args.run, {
             ...(args.realEmail && !args.dryRun
                 ? { publication: { restrictions: ["real-communications"] } }
                 : {}),
-            runId: args.runId,
-            mode: args.dryRun ? "dry-run" : "live",
             participants: 1,
-            createdAt: args.createdAt,
-            ...(args.lab === undefined ? {} : { lab: args.lab }),
             source: args.source,
         }),
         persona: {
@@ -331,7 +326,7 @@ export function buildSingleParticipantBundle(args) {
         },
         lifecycle: [
             {
-                at: args.createdAt,
+                at: args.run.createdAt,
                 event: "cua-lab.run.created",
                 message: `Created computer-use lab run with one participant in a desktop browser (actor ${args.actorId}).`,
             },
@@ -356,7 +351,7 @@ export function buildSingleParticipantBundle(args) {
         feedbackCandidates: args.dryRun || args.inProgress === true
             ? []
             : participantFeedbackCandidates({
-                runId: args.runId,
+                runId: args.run.runId,
                 scenarioId: `cua-${args.labId}`,
                 adapterId: args.labId,
                 goal: redactText(args.mission),

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

@@ -2,16 +2,14 @@ import type { ComputerUsePlan } from "../../lab/plan-types.js";
 import type { LabSubjectState } from "../../lab/types.js";
 import { type RunSubjectProvenance, type RunSubjectStateStepRecord } from "../../run/bundle.js";
 import { type LocalTreeArchive } from "../../subject/local-tree-archive.js";
-import { type CuaSubjectRoute } from "./plan.js";
 import { type DesktopParticipantRun, type CuaSubjectProjection, type CuaSubjectProvenanceArg, type ParticipantRunOutcome } from "./types.js";
 /**
- * Each lane's subject projection. A lane's outcome adds the subject commit it resolved and the
+ * Each participant's subject projection. A participant's outcome adds the subject commit it resolved and the
  * state steps it executed; without outcomes (dry run, or a run still in progress) the declared
  * state is projected as not yet run.
  */
 export declare function projectParticipantSubjects(args: {
     plan: ComputerUsePlan;
-    subjectRoute: CuaSubjectRoute;
     publicRepo?: string;
     localTreeArchive?: LocalTreeArchive;
     runs: readonly DesktopParticipantRun[];
@@ -35,10 +33,10 @@ export declare function resolveSubjectState(args: {
  *  default branch that builder already handles). */
 export declare function subjectProvenanceArg(subject: CuaSubjectProjection, publicRepo: string | undefined, subjectEnvNames: string[]): CuaSubjectProvenanceArg | undefined;
 /**
- * The run-level subject for the top level and the bundle. Local-tree lanes all pack from the same
- * once-per-run archive, so every lane already carries the identical archiveSha256/commit/dirty and
- * the first lane's projection is the aggregate. Clone lanes each resolve their own commit; the
- * aggregate carries it only when every lane agrees, and warns when they diverge.
+ * The run-level subject for the top level and the bundle. Local-tree participants all pack from the
+ * same once-per-run archive, so every participant already carries the identical archiveSha256/commit/dirty and
+ * the first participant's projection is the aggregate. Clone participants each resolve their own
+ * commit; the aggregate carries it only when every participant agrees, and warns when they diverge.
  */
 export declare function aggregateCuaSubject(args: {
     subjects: readonly CuaSubjectProjection[];

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

@@ -1,32 +1,33 @@
-// Each lane's subject projection and the subject-state marker, from the declared subject and what
-// the lanes ran; the run-level subject aggregated from them; and the provenance argument the
+// Each participant's subject projection and the subject-state marker, from the declared subject
+// and what the participants ran; the run-level subject aggregated from them; and the provenance argument the
 // bundle builders take.
 import { commandDigestOf } from "../../subject/state.js";
 import { cuaDeclaredState } from "./plan.js";
 import {} from "../../run/bundle.js";
 import {} from "../../subject/local-tree-archive.js";
-import {} from "./plan.js";
-import {} from "./types.js";
+import { participantSubjectEnv, } from "./types.js";
 /**
- * Each lane's subject projection. A lane's outcome adds the subject commit it resolved and the
+ * Each participant's subject projection. A participant's outcome adds the subject commit it resolved and the
  * state steps it executed; without outcomes (dry run, or a run still in progress) the declared
  * state is projected as not yet run.
  */
 export function projectParticipantSubjects(args) {
-    const { subjectRoute, publicRepo, localTreeArchive } = args;
+    const { publicRepo, localTreeArchive } = args;
+    const { subject } = args.plan.runner;
+    const subjectEnvNames = [...participantSubjectEnv(subject)];
+    // Only a provisioned subject declares state; cuaDeclaredState is undefined for the others.
     const declaredState = cuaDeclaredState(args.plan);
     return args.runs.map((_spec, index) => {
         const outcome = args.outcomes?.[index];
         const subjectState = resolveSubjectState({
-            declared: subjectRoute.provisionedRoute ? declaredState : undefined,
+            declared: declaredState,
             dryRun: args.dryRun,
             executed: outcome?.stateStepRecords ?? [],
         });
         return participantSubjectProjection({
-            cloneRoute: subjectRoute.cloneRoute,
-            localTreeRoute: subjectRoute.localTreeRoute,
+            kind: subject.kind,
             ...(publicRepo === undefined ? {} : { publicRepo }),
-            subjectEnvNames: subjectRoute.subjectEnvNames,
+            subjectEnvNames,
             ...(outcome?.subjectCommit === undefined ? {} : { subjectCommit: outcome.subjectCommit }),
             ...(localTreeArchive === undefined ? {} : { localTreeArchive }),
             subjectState,
@@ -73,12 +74,12 @@ export function resolveSubjectState(args) {
         ...(external.length > 0 ? { externalEnvNames: external } : {}),
     };
 }
-/** Build the per-lane subject projection (invariant 5). Local-tree lanes all share ONE
- *  host-packed archive, so every lane's projection carries the identical archiveSha256/
- *  commit/dirty (no per-lane divergence is possible, unlike the clone route's per-lane
+/** 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/
+ *  commit/dirty (no divergence is possible, unlike the clone route's per-participant
  *  in-sandbox commit). */
 function participantSubjectProjection(args) {
-    if (args.cloneRoute && args.publicRepo) {
+    if (args.kind === "clone" && args.publicRepo) {
         return {
             source: "clone",
             repo: args.publicRepo,
@@ -87,7 +88,7 @@ function participantSubjectProjection(args) {
             state: args.subjectState,
         };
     }
-    if (args.localTreeRoute) {
+    if (args.kind === "local-tree") {
         const archive = args.localTreeArchive;
         return {
             source: "local-tree",
@@ -127,10 +128,10 @@ export function subjectProvenanceArg(subject, publicRepo, subjectEnvNames) {
     return undefined;
 }
 /**
- * The run-level subject for the top level and the bundle. Local-tree lanes all pack from the same
- * once-per-run archive, so every lane already carries the identical archiveSha256/commit/dirty and
- * the first lane's projection is the aggregate. Clone lanes each resolve their own commit; the
- * aggregate carries it only when every lane agrees, and warns when they diverge.
+ * The run-level subject for the top level and the bundle. Local-tree participants all pack from the
+ * same once-per-run archive, so every participant already carries the identical archiveSha256/commit/dirty and
+ * the first participant's projection is the aggregate. Clone participants each resolve their own
+ * commit; the aggregate carries it only when every participant agrees, and warns when they diverge.
  */
 export function aggregateCuaSubject(args) {
     const { subjects, outcomes, participantCount, dryRun } = args;