humanish 0.30.0 → 0.32.0

package/README.md

@@ -262,6 +262,18 @@ you own (a Vercel preview, staging), use an `app-url` subject with
 only declared subject env names do. `humanish init` scaffolds an example at
 `humanish/labs/cua-browser.yaml`.
 
+**Off-app email/SMS verification (`comms`).** When a flow is gated behind an email
+or SMS the app itself sends — a signup verification link, a one-time code, a magic
+link — add a `comms:` block to the lab. Humanish redirects the app's email-API sends
+(via one adopter-named env var, e.g. `RESEND_API_URL`) into a catch **inside** the
+sandbox, so nothing leaves the machine; the persona then opens a synthetic inbox to
+read and click through, and the run bundle gets a digest-only
+`humanish.comms-thread.v1` artifact (from/to/subject/link digests + an OTP count —
+no raw address, link, or code persists). It works on the single-lane clone/local-tree
+route and the concurrent shared-world route, and is vendor-neutral (Resend/SendGrid
+shaped, or a custom profile). See `docs/contracts/schemas.md` for the full `comms:`
+shape.
+
 **Screenshots are full-fidelity by default.** Run bundles live in gitignored
 `.humanish/`, so the Observer shows exactly what the persona saw — the point of
 simming your own app. Set `policies.redactScreenshots: true` to persist blurred

package/dist/comms-types.d.ts

@@ -4,7 +4,7 @@ export interface CommsAddress {
     channel: CommsChannelKind;
     /** Which lane owns this inbox. */
     actorId: string;
-    /** Runtime-only raw address, e.g. patient-07@example.test | +15550137. */
+    /** Runtime-only raw address, e.g. user-07@example.test | +15550137. */
     value: string;
     /** sha256-short(value) — the only form persisted (redaction.digestText). */
     digest: string;

package/dist/cua-actor-lab.d.ts

@@ -612,6 +612,14 @@ export declare function captureDesktopBrowserGeometry(args: {
     viewport?: RunDesktopGeometry["viewport"];
     warnings: string[];
 }>;
+/**
+ * A goal_satisfied lane 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).
+ * Returns the offending reason, or undefined when the lane is a clean pass. Exported for testing.
+ */
+export declare function resolveSelfReportedBlocker(session: CuaLoopResult | undefined): string | undefined;
 /**
  * Run ONE E2B desktop lane end-to-end: create the sandbox (per-lane metadata + the lane's device
  * resolution), prepareDesktop, verify geometry, (clone+serve+seed the subject per lane), open the

package/dist/cua-actor-lab.js

@@ -981,7 +981,7 @@ async function startDesktopStream(desktop, browserWindowId) {
     }
 }
 function completionReasonContradictsGoal(reason) {
-    const text = stripNegatedNonBlockerPhrases(reason.toLowerCase());
+    const text = stripQuotedSpans(stripNegatedNonBlockerPhrases(reason.toLowerCase()));
     return /\b(can'?t|cannot|could not|unable|blocked|blocker|failed|invalid|not set)\b/.test(text)
         || /\b(shows|showing|hit|encountered|returned|got)\b.{0,80}\berror\b/.test(text)
         || /\berror[:.]/.test(text)
@@ -993,11 +993,37 @@ function stripNegatedNonBlockerPhrases(text) {
         .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, "")
         .replace(/\bnot\s+(?:blocked|a blocker|an error|failed)\b/g, "");
 }
+/**
+ * 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`.
+ */
+function stripQuotedSpans(text) {
+    return text
+        .replace(/"[^"]*"/g, " ")
+        .replace(/“[^”]*”/g, " ")
+        .replace(/^\s*>.*$/gm, " ");
+}
 function traceHasStopWhenMatch(session) {
     return session.trace.items.some((item) => item.kind === "notice"
         && item.status === "matched"
         && item.title.startsWith("stopWhen matched"));
 }
+/**
+ * A goal_satisfied lane 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).
+ * Returns the offending reason, or undefined when the lane is a clean pass. Exported for testing.
+ */
+export function resolveSelfReportedBlocker(session) {
+    return session?.completionReason === "goal_satisfied"
+        && completionReasonContradictsGoal(session.reason)
+        && !traceHasStopWhenMatch(session)
+        ? session.reason
+        : undefined;
+}
 /**
  * Run ONE E2B desktop lane end-to-end: create the sandbox (per-lane metadata + the lane's device
  * resolution), prepareDesktop, verify geometry, (clone+serve+seed the subject per lane), open the
@@ -1468,9 +1494,7 @@ export async function runCuaLane(spec, deps) {
     if (noEngagement) {
         warnings.push("Actor returned goal_satisfied with ZERO actions and ZERO messages — it likely saw a blank or still-loading screen and stopped without engaging. NOT counted as a pass. Check the screenshot; raise execution.timeoutMs or confirm the subject painted before the first turn.");
     }
-    const blockerReason = session?.completionReason === "goal_satisfied" && completionReasonContradictsGoal(session.reason)
-        ? session.reason
-        : undefined;
+    const blockerReason = resolveSelfReportedBlocker(session);
     const selfReportedBlocker = blockerReason !== undefined;
     if (selfReportedBlocker) {
         warnings.push(`Actor returned goal_satisfied while its final message describes a blocker or asks for missing instructions — NOT counted as a pass: ${redactText(deps.scrubKnownValues(blockerReason))}`);
@@ -1538,9 +1562,7 @@ async function runInProcessLane(spec, deps) {
     if (noEngagement) {
         warnings.push("Actor returned goal_satisfied with ZERO actions and ZERO messages — it likely saw a blank or still-loading screen and stopped without engaging. NOT counted as a pass. Check the screenshot; raise execution.timeoutMs or confirm the subject painted before the first turn.");
     }
-    const blockerReason = session?.completionReason === "goal_satisfied" && completionReasonContradictsGoal(session.reason)
-        ? session.reason
-        : undefined;
+    const blockerReason = resolveSelfReportedBlocker(session);
     const selfReportedBlocker = blockerReason !== undefined;
     if (selfReportedBlocker) {
         warnings.push(`Actor returned goal_satisfied while its final message describes a blocker or asks for missing instructions — NOT counted as a pass: ${blockerReason}`);