humanish 0.126.0 → 0.127.0

package/CHANGELOG.md

@@ -8,6 +8,19 @@ The Unreleased section holds the full notes for the next version until it is tag
 
 ## Unreleased
 
+## 0.127.0: Clock-end passes counted by what participants declared, a scaled wind-down, the concurrency cap in the plan (2026-10-13)
+
+humanish 0.127.0 makes long scheduled studies report what happened. A participant that passes at
+its scheduled end or session cap without declaring `REACHED THE GOAL.` is now counted as running
+until that end (`participants.ranUntilEnd`), where 0.126.0 counted it as reaching the goal. The
+default wind-down is a quarter of the window, at most 10 minutes, so a short window is no longer
+told its end from its first turn. `HUMANISH_CUA_MAX_CONCURRENCY` now lowers the plan itself, on
+shared-world studies too, and a participant whose scheduled end passes while it waits for a slot
+is skipped before its desktop is created. The Observer keeps updating past 100,000 items for a
+participant, and the bundle estimate is re-derived for run bundle v3.
+
+[Release notes](https://github.com/danielgwilson/humanish/releases/tag/v0.127.0)
+
 ## 0.126.0: Per-participant time zones, session caps and desktop files, page paths, a lighter live Observer (2026-10-13)
 
 humanish 0.126.0 adds per-participant settings for long, mixed studies and a lighter live Observer.

package/dist/actors/computer-use/loop/ending.d.ts

@@ -23,7 +23,9 @@ export interface Stop {
 /**
  * Read the fixed closing line the prompt asks for: the first non-empty line of the
  * participant's last message or closing report summary, exactly one of three phrases,
- * punctuation and case forgiven. Anything else is absence, never a guess. Exported for tests.
+ * punctuation and case forgiven. Anything else is absence, never a guess. A `>` blockquote line
+ * is skipped: it quotes something else, such as the page, and the blocker scan
+ * (routes/computer-use/self-report.ts) drops it too. Exported for tests.
  */
 export declare function declaredOutcomeFromClosingLine(message: string | undefined): ParticipantDeclaredOutcome | undefined;
 /**

package/dist/actors/computer-use/loop/ending.js

@@ -7,7 +7,9 @@ import { plural } from "../../../run/text.js";
 /**
  * Read the fixed closing line the prompt asks for: the first non-empty line of the
  * participant's last message or closing report summary, exactly one of three phrases,
- * punctuation and case forgiven. Anything else is absence, never a guess. Exported for tests.
+ * punctuation and case forgiven. Anything else is absence, never a guess. A `>` blockquote line
+ * is skipped: it quotes something else, such as the page, and the blocker scan
+ * (routes/computer-use/self-report.ts) drops it too. Exported for tests.
  */
 export function declaredOutcomeFromClosingLine(message) {
     if (message === undefined)
@@ -15,10 +17,10 @@ export function declaredOutcomeFromClosingLine(message) {
     const first = message
         .split(/\r?\n/)
         .map((line) => line.trim())
-        .find((line) => line.length > 0);
+        .find((line) => line.length > 0 && !line.startsWith(">"));
     if (first === undefined)
         return undefined;
-    const normalized = first.replace(/^[*_#>\s-]+|[*_.!\s]+$/g, "").toLowerCase();
+    const normalized = first.replace(/^[*_#\s-]+|[*_.!\s]+$/g, "").toLowerCase();
     if (normalized === "reached the goal")
         return "reached";
     if (normalized === "did not reach the goal")

package/dist/actors/computer-use/openai-provider.js

@@ -512,7 +512,7 @@ export function createOpenAiResponsesProvider(options) {
         // Both are offered in both modes. A threaded request reads the session the server holds; an
         // explicit_context one carries its newest screens whole and its oldest turns summarized past
         // the budget, which the trace's conversation record counts. A participant that worked a long
-        // shift writes its closing account from what it carries, as a person writes one from memory.
+        // session writes its closing account from what it carries, as a person writes one from memory.
         debrief: (req, signal) => requestTurn(req, signal, "report"),
         requestImpressions: (req, signal) => requestTurn(req, signal, "impressions"),
     };

package/dist/actors/goal-source.d.ts

@@ -1,5 +1,35 @@
-/** Presentation provenance for computer-use completion; never reclassifies the stored outcome. */
-export type CuaGoalSource = "participant_report" | "condition_matched" | "unavailable";
+/**
+ * The endpoint a computer-use completion came from, as analysis provenance records it
+ * (cuaEndpointSource). Saved analyses hash it, so a new value would make them invalid on read.
+ */
+export declare const CUA_ENDPOINT_SOURCES: readonly ["participant_report", "condition_matched", "unavailable"];
+export type CuaEndpointSource = (typeof CUA_ENDPOINT_SOURCES)[number];
+/**
+ * Presentation provenance for computer-use completion; never reclassifies the stored outcome.
+ * `scheduled_end` and `session_cap` name a session that passed at that clock end without the
+ * participant declaring it reached the goal (ranUntilClockEnd).
+ */
+export type CuaGoalSource = CuaEndpointSource | ClockEnd;
+/** The two clock ends a computer-use session passes at: the stop causes of the same names. */
+export type ClockEnd = "scheduled_end" | "session_cap";
+export declare function isClockEnd(source: CuaGoalSource | undefined): source is ClockEnd;
+/**
+ * The clock end a session ran until, when the participant did not declare that it reached the
+ * goal: its scheduled end or its session cap. Such a session ends as passed, because ending at
+ * the clock is the plan, and its closing report says `DID NOT REACH THE GOAL.` or declares
+ * nothing. The tally and every completion label read this, so none counts it as reaching the
+ * goal. Undefined for every other ending, including a declared `reached` at a clock end.
+ */
+export declare function ranUntilClockEnd(trace: {
+    readonly stopCause?: unknown;
+    readonly declaredOutcome?: unknown;
+}): ClockEnd | undefined;
 export declare function isCuaTrace(value: unknown): boolean;
 export declare function cuaGoalSource(value: unknown, recordedStatus?: unknown): CuaGoalSource | undefined;
+/**
+ * cuaGoalSource for analysis provenance. A clock end reads as it did before it had a source of
+ * its own, so an analysis saved then still matches its run when re-read. The provenance's
+ * `stopCause` and `declaredOutcome` name the clock end.
+ */
+export declare function cuaEndpointSource(value: unknown, recordedStatus?: unknown): CuaEndpointSource | undefined;
 export declare const CUA_COMPLETION_NOTE = "Participant reports alone do not establish task success. A matched stop condition establishes only its declared condition. Run gate and share-safety results are separate.";

package/dist/actors/goal-source.js

@@ -1,3 +1,27 @@
+/**
+ * The endpoint a computer-use completion came from, as analysis provenance records it
+ * (cuaEndpointSource). Saved analyses hash it, so a new value would make them invalid on read.
+ */
+export const CUA_ENDPOINT_SOURCES = [
+    "participant_report",
+    "condition_matched",
+    "unavailable",
+];
+export function isClockEnd(source) {
+    return source === "scheduled_end" || source === "session_cap";
+}
+/**
+ * The clock end a session ran until, when the participant did not declare that it reached the
+ * goal: its scheduled end or its session cap. Such a session ends as passed, because ending at
+ * the clock is the plan, and its closing report says `DID NOT REACH THE GOAL.` or declares
+ * nothing. The tally and every completion label read this, so none counts it as reaching the
+ * goal. Undefined for every other ending, including a declared `reached` at a clock end.
+ */
+export function ranUntilClockEnd(trace) {
+    if (trace.stopCause !== "scheduled_end" && trace.stopCause !== "session_cap")
+        return undefined;
+    return trace.declaredOutcome === "reached" ? undefined : trace.stopCause;
+}
 export function isCuaTrace(value) {
     if (typeof value !== "object" || value === null)
         return false;
@@ -32,7 +56,22 @@ export function cuaGoalSource(value, recordedStatus) {
         (item.title.startsWith("stopWhen matched:") || item.title === "dwell window complete"))) {
         return "condition_matched";
     }
-    // A legacy trace without enough detail cannot establish which endpoint ended the loop.
-    return trace.items.length > 0 ? "participant_report" : "unavailable";
+    return ranUntilClockEnd(trace) ?? endpointOfItems(trace.items);
+}
+// A legacy trace without enough detail cannot establish which endpoint ended the loop.
+function endpointOfItems(items) {
+    return items.length > 0 ? "participant_report" : "unavailable";
+}
+/**
+ * cuaGoalSource for analysis provenance. A clock end reads as it did before it had a source of
+ * its own, so an analysis saved then still matches its run when re-read. The provenance's
+ * `stopCause` and `declaredOutcome` name the clock end.
+ */
+export function cuaEndpointSource(value, recordedStatus) {
+    const source = cuaGoalSource(value, recordedStatus);
+    // cuaGoalSource returns a clock end only after it has checked that `items` is an array.
+    return isClockEnd(source)
+        ? endpointOfItems(value.items)
+        : source;
 }
 export const CUA_COMPLETION_NOTE = "Participant reports alone do not establish task success. A matched stop condition establishes only its declared condition. Run gate and share-safety results are separate.";

package/dist/analysis/evidence-sources.js

@@ -1,6 +1,6 @@
 import { receivingAnalysisContext } from "../comms/receiving-evidence.js";
 import { eventRecordIdOf, recordIdOf, streamParticipantIdOf } from "../run/participant-records.js";
-import { cuaGoalSource } from "../actors/goal-source.js";
+import { cuaEndpointSource } from "../actors/goal-source.js";
 import { isRecord } from "../run/type-guards.js";
 import { ACTION_CAPTURE_VERSION, } from "./types.js";
 import { decodesToPlainRelativePath, isEvidencePath } from "../run/evidence-files.js";
@@ -82,7 +82,7 @@ export function participantSource(stream, captureVersion) {
             actorStatus: actor?.status ?? null,
             completionReason: actor?.completionReason ?? null,
             stopCause: actor?.stopCause ?? null,
-            goalSource: cuaGoalSource(actor, stream.status) ?? null,
+            goalSource: cuaEndpointSource(actor, stream.status) ?? null,
             declaredOutcome: actor?.declaredOutcome ?? null,
             taskOutcomes: actor?.taskFunnel === undefined
                 ? null

package/dist/analysis/types.d.ts

@@ -1,5 +1,5 @@
 import type { ActorStatus, ActorCompletionReason, ActorStopCause, ParticipantDeclaredOutcome } from "../actors/contract.js";
-import type { CuaGoalSource } from "../actors/goal-source.js";
+import type { CuaEndpointSource } from "../actors/goal-source.js";
 import type { AutomaticAnalysisView } from "./job.js";
 import type { RunAnalysisCost } from "../run/run-cost.js";
 /** Independent interpretation of retained evidence; never a participant or harness verdict. */
@@ -55,7 +55,7 @@ export interface AnalysisParticipantInput {
         actorStatus: ActorStatus | null;
         completionReason: ActorCompletionReason | null;
         stopCause: ActorStopCause | null;
-        goalSource: CuaGoalSource | null;
+        goalSource: CuaEndpointSource | null;
         declaredOutcome: ParticipantDeclaredOutcome | null;
         taskOutcomes: Array<{
             taskId: string;

package/dist/analysis/validation.js

@@ -2,6 +2,7 @@ import { validStoredCodexAnalysisConfig } from "./codex-config.js";
 import { createHash } from "node:crypto";
 import { z } from "zod";
 import { ACTOR_STATUSES, ACTOR_STOP_CAUSES } from "../actors/contract.js";
+import { CUA_ENDPOINT_SOURCES } from "../actors/goal-source.js";
 import { ANALYSIS_LIMITS, EVIDENCE_LIMITS, MAX_ANALYSIS_COHORTS, hasControlCharacter, hasRevisionFields, } from "./analysis-limits.js";
 import { SHA256_HEX_PATTERN, ACTION_CAPTURE_VERSION, ANALYSIS_SCHEMA, ANALYSIS_CORRECTION_SCHEMA, } from "./types.js";
 const text = (max) => z
@@ -178,7 +179,7 @@ const analysisParticipantProvenanceSchema = z
     ])
         .nullable(),
     stopCause: z.enum(ACTOR_STOP_CAUSES).nullable(),
-    goalSource: z.enum(["participant_report", "condition_matched", "unavailable"]).nullable(),
+    goalSource: z.enum(CUA_ENDPOINT_SOURCES).nullable(),
     declaredOutcome: z.enum(["reached", "not_reached", "blocked"]).nullable(),
     taskOutcomes: z
         .array(z

package/dist/observer/data.js

@@ -156,6 +156,13 @@ export function observerDataFromSaved(value) {
         ? null
         : joinObserverData(saved, null);
 }
+const GOAL_SOURCE_LABELS = {
+    participant_report: "Reported complete",
+    condition_matched: "Condition matched",
+    scheduled_end: "Ran until scheduled end",
+    session_cap: "Ran until session cap",
+    unavailable: "Completion source unavailable",
+};
 /** Refresh presentation from recorded actor evidence, including older exported snapshots. */
 export function withObserverEndings(data) {
     const streams = (data.streams ?? []).map(({ ending: _previousEnding, ...stream }) => {
@@ -166,11 +173,7 @@ export function withObserverEndings(data) {
             ...(ending === undefined ? {} : { ending }),
             ...((stream.status === "passed" || stream.status === "complete") && source !== undefined
                 ? {
-                    statusLabel: source === "participant_report"
-                        ? "Reported complete"
-                        : source === "condition_matched"
-                            ? "Condition matched"
-                            : "Completion source unavailable",
+                    statusLabel: GOAL_SOURCE_LABELS[source],
                 }
                 : {}),
             ...(stream.status === "incomplete" || (ending !== undefined && stream.status === "abandoned")