test(providers): reproduce agent creation pinned to a failed catalog

packages/app/e2e/mobile/workspace-memory/README.md

@@ -1,21 +0,0 @@
-# Android workspace memory diagnostic
-
-Supplemental reproduction for [#4408](https://github.com/getpaseo/paseo/issues/4408), using the production 0.11.2 APK on an Android 16 emulator. This tests an adapted failure: opening the sidebar with 800 projects, each containing one workspace, produced a `LOW_MEMORY` exit after 66–90 seconds. It does not reproduce the reported 18-second cold-start ANR or establish its cause.
-
-Before running:
-
-- Use a dedicated emulator with the production APK installed and notifications permission settled. Bind the agent-device session to that emulator before running.
-- Pair the app with an isolated daemon containing 800 active workspaces, each in its own project. Use Project grouping and leave projects expanded. The observed fixture used existing directories inside a Git checkout, with workspace titles permuted by `(index * 317) % 800`, no agents, and no terminals.
-- Record the daemon revision, emulator RAM and Java heap settings, and the project/workspace distribution. Device memory pressure controls whether Android kills the app.
-
-Run from the repository root with the agent-device session, Android serial, and an artifact directory:
-
-```sh
-node packages/app/e2e/mobile/workspace-memory/android.mjs workspace-memory emulator-5556 .dev/workspace-memory
-```
-
-The test cold-starts the paired app, opens its sidebar, samples memory and the process ID every five seconds for two minutes, and verifies that closing the sidebar responds. A process exit or restart fails the assertion. Memory, exit records, and logs stay in the supplied artifact directory.
-
-Control: repeat with 800 workspaces grouped into one project. That distribution survived the two-minute automated check and responded to closing the sidebar, with total PSS near 300 MB. The project-per-workspace fixture reached about 2.5 GB PSS before Android recorded `LOW_MEMORY`; its Java heap rose and fell rather than showing the reporter's full-heap ANR. Neither result proves that shared collators would fix the report.
-
-This diagnostic is deliberately outside CI and contains no production change.

packages/app/e2e/mobile/workspace-memory/android.mjs

@@ -1,65 +0,0 @@
-// Requires a paired production app and the workspace fixture described in README.md.
-import assert from "node:assert/strict";
-import { execFileSync } from "node:child_process";
-import { mkdirSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
-import { setTimeout as delay } from "node:timers/promises";
-
-const [, , session, serial, artifacts] = process.argv;
-assert(session && serial && artifacts, "Pass session, Android serial, and artifact directory");
-mkdirSync(artifacts, { recursive: true });
-
-function ad(...args) {
-  return execFileSync("agent-device", [...args, "--session", session], {
-    encoding: "utf8",
-    timeout: 60_000,
-  });
-}
-function adb(...args) {
-  return execFileSync("adb", ["-s", serial, ...args], {
-    encoding: "utf8",
-    timeout: 15_000,
-  });
-}
-function pid() {
-  try {
-    return adb("shell", "pidof", "sh.paseo").trim();
-  } catch {
-    return "";
-  }
-}
-
-ad("open", "sh.paseo", "--relaunch", "--platform", "android");
-ad("wait", "text", "Open menu", "15000");
-assert.equal(
-  JSON.parse(ad("perf", "memory", "sample", "--json")).data.deviceId,
-  serial,
-  "Session and adb serial must address the same device",
-);
-const initialPid = pid();
-assert(initialPid, "App must have a process before opening the sidebar");
-ad("press", 'label="Open menu"', "--settle");
-const startedAt = Date.now();
-const samples = [];
-try {
-  while (Date.now() - startedAt < 120_000) {
-    const currentPid = pid();
-    samples.push({
-      elapsedMs: Date.now() - startedAt,
-      pid: currentPid,
-      meminfo: currentPid ? adb("shell", "dumpsys", "meminfo", "sh.paseo") : "",
-    });
-    writeFileSync(join(artifacts, "samples.json"), JSON.stringify(samples, null, 2));
-    assert.equal(currentPid, initialPid, "Opening the workspace sidebar must keep the app alive");
-    await delay(5_000);
-  }
-  ad("press", 'label="Close sidebar"', "--settle");
-  ad("wait", "text", "Open menu", "10000");
-  console.log("PASS: app survived two minutes with the sidebar open and responded to Close");
-} finally {
-  writeFileSync(
-    join(artifacts, "exit-info.txt"),
-    adb("shell", "dumpsys", "activity", "exit-info", "sh.paseo"),
-  );
-  writeFileSync(join(artifacts, "logcat.txt"), adb("logcat", "-d", "-t", "4000"));
-}

packages/server/src/server/agent/provider-snapshot-manager.test.ts

@@ -142,6 +142,46 @@ async function runTestCatalogActivities(
 }
 
 describe("ProviderSnapshotManager public surface", () => {
+  test("agent creation retries a failed directory catalog without restarting the manager", async () => {
+    const cwd = resolveSnapshotCwd("/tmp/cursor-retry");
+    let attempts = 0;
+    const manager = new ProviderSnapshotManager({
+      logger: createTestLogger(),
+      providerOverrides: {
+        cursor: { label: "Cursor", extends: "acp", command: ["cursor-agent", "acp"] },
+      },
+      extraClients: {
+        cursor: createExtraClient("cursor", {
+          isAvailable: async () => true,
+          async fetchCatalog() {
+            attempts++;
+            if (attempts === 1) throw new Error("Failed to initialize session services");
+            return { models: [{ provider: "cursor", id: "model", label: "Model" }], modes: [] };
+          },
+        }),
+      },
+    });
+    const create = () =>
+      manager.resolveCreateConfig({
+        cwd,
+        provider: "cursor",
+        requestedMode: undefined,
+        featureValues: undefined,
+        parent: null,
+        unattended: false,
+      });
+    try {
+      await expect(create()).rejects.toThrow("Failed to initialize session services");
+      expect(attempts).toBe(1);
+      await expect(create()).resolves.toBeDefined();
+      expect(attempts).toBe(2);
+      await create();
+      expect(attempts).toBe(2);
+    } finally {
+      manager.destroy();
+    }
+  });
+
   test("publishes one model per provider identity and retains the first definition", async () => {
     const first = { provider: "codex", id: "shared", label: "First", isDefault: true };
     const manager = new ProviderSnapshotManager({

packages/server/src/server/agent/providers/pi/agent.test.ts

@@ -1943,48 +1943,6 @@ describe("PiRpcAgentSession", () => {
     expect(session.describePersistence()?.metadata?.thinkingOptionId).toBe("off");
   });
 
-  test.each(["high", "off", "max"] as const)(
-    "publishes an extension's thinking change to %s before refreshing state",
-    async (level) => {
-      const { pi, session } = await createSession();
-      onTestFinished(() => session.close());
-      const changes: AgentStreamEvent[] = [];
-      session.subscribe((event) => {
-        if (event.type === "thinking_option_changed") {
-          expect(session.describePersistence()?.metadata?.thinkingOptionId).toBe(level);
-          changes.push(event);
-        }
-      });
-      const runtime = pi.latestSession();
-      runtime.state = { ...runtime.state, thinkingLevel: level };
-      runtime.emit({ type: "thinking_level_changed", level });
-
-      expect(changes).toEqual([
-        { type: "thinking_option_changed", provider: "pi", thinkingOptionId: level },
-      ]);
-      expect(session.describePersistence()?.metadata?.thinkingOptionId).toBe(level);
-      await expect(session.getRuntimeInfo()).resolves.toMatchObject({ thinkingOptionId: level });
-    },
-  );
-
-  test.each([undefined, null, "invalid", 123])(
-    "ignores a thinking change with invalid level %s",
-    async (level) => {
-      const { pi, session } = await createSession();
-      onTestFinished(() => session.close());
-      const initialThinking = session.describePersistence()?.metadata?.thinkingOptionId;
-      const changes: AgentStreamEvent[] = [];
-      session.subscribe((event) => {
-        if (event.type === "thinking_option_changed") changes.push(event);
-      });
-
-      pi.latestSession().emit({ type: "thinking_level_changed", level });
-
-      expect(changes).toEqual([]);
-      expect(session.describePersistence()?.metadata?.thinkingOptionId).toBe(initialThinking);
-    },
-  );
-
   test("surfaces state refresh failures after a Pi thinking update", async () => {
     const { pi, session } = await createSession();
     onTestFinished(() => session.close());

packages/server/src/server/agent/providers/pi/agent.ts

@@ -1005,8 +1005,6 @@ function isProcessExitEvent(
 
 function isPiAgentSessionEvent(event: PiRuntimeEvent): event is PiAgentSessionEvent {
   switch (event.type) {
-    case "thinking_level_changed":
-      return typeof event.level === "string" && normalizePiThinkingOption(event.level) !== null;
     case "agent_start":
     case "turn_start":
     case "message_start":
@@ -2138,15 +2136,6 @@ export class PiRpcAgentSession implements AgentSession {
     }
 
     switch (event.type) {
-      case "thinking_level_changed":
-        this.state = { ...this.state, thinkingLevel: event.level };
-        this.config.thinkingOptionId = event.level;
-        this.emit({
-          type: "thinking_option_changed",
-          provider: this.provider,
-          thinkingOptionId: event.level,
-        });
-        return;
       case "agent_start":
         this.activeTurnStarted = true;
         this.clearNoTurnBuffers();

packages/server/src/server/agent/providers/pi/rpc-types.ts

@@ -160,7 +160,6 @@ export type PiAssistantMessageEvent =
   | { type: "start" | "text_start" | "text_end" | "thinking_start" | "thinking_end" | "done" };
 
 export type PiAgentSessionEvent =
-  | { type: "thinking_level_changed"; level: PiThinkingLevel }
   | { type: "agent_start" }
   | { type: "turn_start" }
   | { type: "message_start"; message: PiAgentMessage }