@data-fair/lib-agents-sim 0.5.0 → 0.6.0

package/chat-driver.d.ts

@@ -24,11 +24,23 @@ export declare function chatDriverStrings(locale: ChatDriverLocale): {
 };
 export declare const TURN_TIMEOUT_MS: number;
 export declare const SEND_TIMEOUT_MS = 15000;
+/**
+ * How a turn ended. `waiting` means the assistant declared
+ * `wait_for_user_action` and is holding the turn open for the person — the
+ * caller's cue to let them act, then wait again for the turn it resumes.
+ */
+export type TurnOutcome = 'ended' | 'waiting';
+/**
+ * Matched on the activity's kind, not its label: the label is the model's own
+ * words interpolated into a translated string, so any text match would be both
+ * locale-dependent and at the mercy of what the assistant wrote.
+ */
+export declare const WAITING_SELECTOR = "[data-testid=\"chat-activity\"][data-activity=\"waiting\"]";
 export declare function createChatDriver(root: ChatRoot, opts?: {
     locale?: ChatDriverLocale;
 }): {
     sendMessage(text: string): Promise<void>;
-    waitForTurn(timeoutMs?: number): Promise<void>;
+    waitForTurn(timeoutMs?: number): Promise<TurnOutcome>;
     readConversation(): Promise<{
         role: "user" | "assistant";
         text: string;

package/chat-driver.js

@@ -46,6 +46,12 @@ export const TURN_TIMEOUT_MS = 10 * 60 * 1000;
 // later with no diagnosis. Unrelated to TURN_TIMEOUT_MS, which bounds a
 // legitimately long model turn once the message has actually been sent.
 export const SEND_TIMEOUT_MS = 15000;
+/**
+ * Matched on the activity's kind, not its label: the label is the model's own
+ * words interpolated into a translated string, so any text match would be both
+ * locale-dependent and at the mercy of what the assistant wrote.
+ */
+export const WAITING_SELECTOR = '[data-testid="chat-activity"][data-activity="waiting"]';
 export function createChatDriver(root, opts = {}) {
     const strings = chatDriverStrings(opts.locale ?? 'en');
     return {
@@ -77,10 +83,23 @@ export function createChatDriver(root, opts = {}) {
         },
         async waitForTurn(timeoutMs = TURN_TIMEOUT_MS) {
             const stop = root.getByRole('button', { name: strings.stop });
+            const waiting = root.locator(WAITING_SELECTOR);
             // The turn may already be finished by the time we look, so a missing Stop
             // button is not an error — only one that never goes away is.
             await stop.waitFor({ state: 'visible', timeout: 15000 }).catch(() => { });
-            await expect(stop).toHaveCount(0, { timeout: timeoutMs });
+            // A turn can finish two ways, and only one of them is the assistant being
+            // done. `wait_for_user_action` holds the turn open on purpose, having handed
+            // control back to the person — and a simulated person only acts between
+            // turns, so a harness that waited for the Stop button alone could never let
+            // them act on it. Every declared wait then ran its whole window and was
+            // recorded as a wedged turn; at a wait window as long as the harness's own
+            // ceiling, that is every run.
+            const ended = expect(stop).toHaveCount(0, { timeout: timeoutMs }).then(() => 'ended');
+            const armed = expect(waiting).toHaveCount(1, { timeout: timeoutMs }).then(() => 'waiting', 
+            // Never rejects: a wait that is simply not what this turn did must not be
+            // the error a caller sees. The Stop arm owns the timeout message.
+            () => new Promise(() => { }));
+            return await Promise.race([ended, armed]);
         },
         async readConversation() {
             // evaluateAll, not page.evaluate: FrameLocator has no evaluate, and this

package/index.d.ts

@@ -6,6 +6,6 @@ export { writeEvidence, evidenceDir } from './transcript.ts';
 export { computeMetrics, type RunMetrics } from './metrics.ts';
 export { selectCases } from './cases.ts';
 export { reportCases } from './report.ts';
-export { createChatDriver, chatDriverStrings, type ChatRoot, type ChatDriverLocale, TURN_TIMEOUT_MS, SEND_TIMEOUT_MS } from './chat-driver.ts';
+export { createChatDriver, chatDriverStrings, type ChatRoot, type ChatDriverLocale, type TurnOutcome, TURN_TIMEOUT_MS, SEND_TIMEOUT_MS, WAITING_SELECTOR } from './chat-driver.ts';
 export { createPagePerception, truncate, SNAPSHOT_CAP, ACTION_TIMEOUT_MS, MCP_SERVER_NAME as PAGE_MCP_SERVER_NAME } from './page-perception.ts';
 export type { PerceptionRoot, Observation, PagePerception } from './page-perception.ts';

package/index.js

@@ -5,5 +5,5 @@ export { writeEvidence, evidenceDir } from "./transcript.js";
 export { computeMetrics } from "./metrics.js";
 export { selectCases } from "./cases.js";
 export { reportCases } from "./report.js";
-export { createChatDriver, chatDriverStrings, TURN_TIMEOUT_MS, SEND_TIMEOUT_MS } from "./chat-driver.js";
+export { createChatDriver, chatDriverStrings, TURN_TIMEOUT_MS, SEND_TIMEOUT_MS, WAITING_SELECTOR } from "./chat-driver.js";
 export { createPagePerception, truncate, SNAPSHOT_CAP, ACTION_TIMEOUT_MS, MCP_SERVER_NAME as PAGE_MCP_SERVER_NAME } from "./page-perception.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@data-fair/lib-agents-sim",
-  "version": "0.5.0",
+  "version": "0.6.0",
   "description": "Primitives for judged browser simulations of the data-fair agents chat, plus a Claude Code bridge exposing the Agent SDK as an OpenAI-compatible provider.",
   "main": "index.js",
   "type": "module",