humanish 0.87.0 → 0.88.1

package/dist/run.d.ts

@@ -1,5 +1,6 @@
 import { type CodexAppServerTrace } from "./codex-app-server.js";
 import { type ActorStatus, type ActorTrace, type ActorTraceItem } from "./actor-contract.js";
+import { type CuaGoalSource } from "./actor-goal-source.js";
 import type { TaskFunnel } from "./tasks.js";
 import { type CapturedGitState } from "./core/git-state.js";
 import type { E2BDesktopModule } from "./e2b-desktop-launch.js";
@@ -844,7 +845,7 @@ export interface RunCostLine {
     modelId?: string;
     /** null = NOT MEASURED / no rate; never coerced to 0. */
     estimatedCostUsd: number | null;
-    reason?: "no_rate_for_model" | "no_rate_for_desktop" | "no_token_usage" | "no_duration" | "closing_usage_unreported" | "no_desktop_resources" | "desktop_lifetime_incomplete";
+    reason?: "no_rate_for_model" | "no_rate_for_desktop" | "no_token_usage" | "no_duration" | "closing_usage_unreported" | "interaction_usage_unreported" | "no_desktop_resources" | "desktop_lifetime_incomplete";
     /** Pricing provenance date; non-null iff estimatedCostUsd is non-null. */
     ratesAsOf: string | null;
     source?: string;
@@ -933,7 +934,7 @@ export interface RunRerunLineage {
 export interface ParticipantOutcomes {
     /** Participants whose sessions reached a terminal state — the denominator for every count below. */
     total: number;
-    /** Reached the goal. */
+    /** Recorded successful sessions; completion provenance depends on the actor and its evidence. */
     reachedGoal: number;
     /** Stopped trying. A finding about the product. */
     abandoned: number;
@@ -1011,11 +1012,23 @@ export interface StudyTaskFunnel {
 export declare function aggregateTaskFunnels(funnels: readonly TaskFunnel[]): StudyTaskFunnel | undefined;
 /** The funnel as one line, denominator on every number: `signup 2/2 · verify-email 1/2`. */
 export declare function formatStudyTaskFunnel(funnel: StudyTaskFunnel): string;
-/** One line a stakeholder can read, with the denominator attached to every number. */
-export declare function formatParticipantOutcomes(outcomes: ParticipantOutcomes, terminalCauses?: readonly {
+type ParticipantOutcomeDetail = {
     status: ActorStatus;
     label?: string;
-}[]): string;
+    goalSource?: CuaGoalSource;
+};
+/** One line a stakeholder can read, with the denominator attached to every number. */
+export declare function formatParticipantOutcomes(outcomes: ParticipantOutcomes, terminalCauses?: readonly ParticipantOutcomeDetail[]): string;
+/** Presentation details tolerate optional legacy actor payloads without changing their tallies. */
+export declare function participantOutcomeDetails(streams: readonly {
+    actor?: unknown;
+    status?: unknown;
+}[]): ParticipantOutcomeDetail[];
+/** Refresh a CUA completion claim from recorded traces, leaving original evidence and enums intact. */
+export declare function withCuaReviewProvenance(review: ReviewSummary, streams: readonly {
+    actor?: unknown;
+    status?: unknown;
+}[]): ReviewSummary;
 export declare function buildRunSource(args: {
     cwd: string;
     capturedAt?: Date | string;

package/dist/run.js

@@ -13,6 +13,8 @@ import { CODEX_APP_SERVER_TRACE_SCHEMA, runCodexAppServerSessionInPreparedRoot }
 import { getActor } from "./actor-registry.js";
 import { artifactReferenceIfWritten, hasWrittenScreenshot } from "./artifact-reference.js";
 import { ACTOR_TRACE_SCHEMA } from "./actor-contract.js";
+import { cuaGoalSource, isCuaTrace, CUA_COMPLETION_NOTE } from "./actor-goal-source.js";
+import { actorEnding } from "./actor-stop-cause.js";
 import { captureGitState, GIT_STATE_SCHEMA } from "./core/git-state.js";
 import { inspectVerifiedGitWorkspace } from "./core/git-workspace.js";
 import { mapWithConcurrency } from "./concurrency.js";
@@ -137,11 +139,34 @@ export function formatStudyTaskFunnel(funnel) {
     })
         .join(" · ");
 }
+function participantCompletionLine(outcomes, terminalCauses) {
+    const completions = terminalCauses.filter((entry) => entry.status === "passed");
+    const reported = completions.filter((entry) => entry.goalSource === "participant_report").length;
+    const matched = completions.filter((entry) => entry.goalSource === "condition_matched").length;
+    let goalLine = outcomes.reachedGoal === 0
+        ? `0/${outcomes.total} recorded completions`
+        : `${outcomes.reachedGoal}/${outcomes.total} reached the goal`;
+    if (outcomes.reachedGoal > 0 && terminalCauses.some((entry) => entry.goalSource !== undefined)) {
+        const count = `${outcomes.reachedGoal}/${outcomes.total}`;
+        goalLine = completions.length !== outcomes.reachedGoal
+            ? `${count} recorded completions (completion source unavailable)`
+            : reported === outcomes.reachedGoal
+                ? `${count} reported reaching the goal`
+                : matched === outcomes.reachedGoal
+                    ? `${count} met a recorded completion condition`
+                    : `${count} recorded completions (${[
+                        reported > 0 ? `${reported} participant-reported` : undefined,
+                        matched > 0 ? `${matched} condition-matched` : undefined,
+                        outcomes.reachedGoal - reported - matched > 0 ? `${outcomes.reachedGoal - reported - matched} other or unavailable source` : undefined
+                    ].filter(Boolean).join(", ")})`;
+    }
+    return goalLine;
+}
 /** One line a stakeholder can read, with the denominator attached to every number. */
 export function formatParticipantOutcomes(outcomes, terminalCauses = []) {
     if (outcomes.total === 0)
         return "no participants reached a terminal state";
-    const parts = [`${outcomes.reachedGoal}/${outcomes.total} reached the goal`];
+    const parts = [participantCompletionLine(outcomes, terminalCauses)];
     // Detail may explain a recorded outcome, but must never change its count or invent a match
     // between a tally and an incomplete set of traces.
     const append = (count, statuses, fallback) => {
@@ -174,6 +199,46 @@ export function formatParticipantOutcomes(outcomes, terminalCauses = []) {
         parts.push(`${outcomes.reportedFriction} reported friction`);
     return parts.join(", ");
 }
+/** Presentation details tolerate optional legacy actor payloads without changing their tallies. */
+export function participantOutcomeDetails(streams) {
+    return streams.flatMap((stream) => {
+        if (!isRecord(stream.actor))
+            return [];
+        const actor = stream.actor;
+        const goalSource = cuaGoalSource(actor, stream.status);
+        if (!["passed", "abandoned", "incomplete", "blocked", "timed_out", "failed"].includes(String(actor.status))) {
+            return goalSource === "unavailable" ? [{ status: "passed", goalSource }] : [];
+        }
+        const ending = Array.isArray(actor.items) && actor.items.every(isRecord) ? actorEnding(actor) : undefined;
+        return [{ status: actor.status,
+                ...(ending === undefined ? {} : { label: ending.label }),
+                ...(goalSource === undefined ? {} : { goalSource }) }];
+    });
+}
+/** Refresh a CUA completion claim from recorded traces, leaving original evidence and enums intact. */
+export function withCuaReviewProvenance(review, streams) {
+    const details = participantOutcomeDetails(streams);
+    if (!isRecord(review.participants)
+        || !["total", "reachedGoal", "abandoned", "ranOut", "blocked", "harnessFailed", "reportedFriction"].every((key) => isNonNegativeSafeInteger(review.participants[key]))
+        || (review.participants.reachedGoal === 0
+            ? !streams.some((stream) => isCuaTrace(stream.actor))
+            : !details.some((entry) => entry.goalSource !== undefined)))
+        return review;
+    const outcomes = formatParticipantOutcomes(review.participants, details);
+    // Preserve rerun context, participant narration and adapter-specific findings. Refreshing a
+    // historical summary qualifies its old tally instead of silently discarding that context.
+    const header = `Run gate: ${review.verdict}. Participants: ${outcomes}.${review.tasks ? ` Tasks: ${formatStudyTaskFunnel(review.tasks)}.` : ""}`;
+    const prefix = `${header} Recorded summary: `;
+    const recorded = review.summary.startsWith(prefix) ? review.summary.slice(prefix.length) : review.summary;
+    const oldGoal = `${review.participants.reachedGoal}/${review.participants.total} reached the goal`;
+    const qualified = recorded.split(oldGoal).join(participantCompletionLine(review.participants, details));
+    return {
+        ...review,
+        summary: `${prefix}${qualified}`,
+        gaps: review.participants.reachedGoal === 0 ? review.gaps
+            : [...review.gaps.filter((gap) => gap !== CUA_COMPLETION_NOTE), CUA_COMPLETION_NOTE]
+    };
+}
 export async function buildRunSource(args) {
     const gitOptions = args.capturedAt === undefined ? {} : { capturedAt: args.capturedAt };
     return {
@@ -2988,8 +3053,11 @@ export async function readReview(cwdInput, runInput) {
             }
         };
     }
+    const bundle = runPaths ? await readRunJsonIfExists(runPaths, "run.json") : null;
+    const projected = isRecord(bundle) && Array.isArray(bundle.streams)
+        ? withCuaReviewProvenance(review, bundle.streams.filter(isRecord)) : review;
     return {
-        ...review,
+        ...projected,
         path: path.relative(cwd, path.join(runPaths.absoluteRunRoot, "review.json")),
         runId: path.basename(runPaths.absoluteRunRoot)
     };