@swmansion/argent 0.30.2-next.2 → 0.30.2-next.20

package/dist/mcp-server.mjs

@@ -24566,6 +24566,7 @@ function spawnToolsServer(paths, port, options = {}) {
       stdio: ["ignore", "pipe", logFd],
       env: buildToolsServerEnv(paths, port, process.env, options)
     });
+    fs.closeSync(logFd);
     child.unref();
     const pid = child.pid;
     if (!pid) {
@@ -25313,6 +25314,7 @@ var FAILURE_CODES = {
   REACT_PROFILER_DEVTOOLS_RENDERER_MISSING: "REACT_PROFILER_DEVTOOLS_RENDERER_MISSING",
   REACT_PROFILER_START_FAILED: "REACT_PROFILER_START_FAILED",
   REACT_PROFILER_START_VERIFY_FAILED: "REACT_PROFILER_START_VERIFY_FAILED",
+  REACT_PROFILER_SESSION_TORN_DOWN: "REACT_PROFILER_SESSION_TORN_DOWN",
   REACT_PROFILER_HOOK_ERROR: "REACT_PROFILER_HOOK_ERROR",
   REACT_PROFILER_ANALYZE_NO_DATA: "REACT_PROFILER_ANALYZE_NO_DATA",
   NATIVE_PROFILER_FACTORY_OPTIONS_MISSING: "NATIVE_PROFILER_FACTORY_OPTIONS_MISSING",
@@ -25577,7 +25579,7 @@ async function getResolvedToolsUrl() {
     const link2 = await readLinkConfig();
     const envToken = process.env[AUTH_TOKEN_ENV];
     return {
-      url: envUrl,
+      url: envUrl.replace(/\/+$/, ""),
       source: "env",
       ...envToken ? { token: envToken } : {},
       ...link2 ? { shadowedLink: link2 } : {}
@@ -26485,8 +26487,13 @@ async function consumeToolStream(name, body, onProgress, signal) {
     if (msg.event === "progress") {
       progress++;
       onProgress(msg.data);
-    } else if (msg.event === "result") final = { data: msg.data, note: msg.note };
-    else if (msg.event === "error") {
+    } else if (msg.event === "result") {
+      final = {
+        data: msg.data,
+        note: msg.note,
+        ...msg.profiling === true ? { profiling: true } : {}
+      };
+    } else if (msg.event === "error") {
       throw new ToolInvocationError(msg.error ?? "tool invocation failed", {
         errorCode: msg.error_code,
         errorKind: msg.error_kind
@@ -26533,7 +26540,11 @@ async function consumeToolStream(name, body, onProgress, signal) {
   if (!final) {
     throw brokenStream(name, "the stream ended without a result", progress);
   }
-  return { data: final.data, note: final.note };
+  return {
+    data: final.data,
+    note: final.note,
+    ...final.profiling ? { profiling: true } : {}
+  };
 }
 function errorBodyMessage(body) {
   if (Array.isArray(body.issues) && typeof body.message === "string") return body.message;
@@ -26706,7 +26717,8 @@ function createToolsClient(options = {}) {
       return {
         data: await settle2(streamed.data),
         note: streamed.note,
-        outputHint: meta2?.outputHint
+        outputHint: meta2?.outputHint,
+        ...streamed.profiling ? { profiling: true } : {}
       };
     }
     let json;
@@ -26731,7 +26743,12 @@ function createToolsClient(options = {}) {
       });
     }
     opts?.signal?.throwIfAborted();
-    return { data: await settle2(json.data), note: json.note, outputHint: meta2?.outputHint };
+    return {
+      data: await settle2(json.data),
+      note: json.note,
+      outputHint: meta2?.outputHint,
+      ...json.profiling === true ? { profiling: true } : {}
+    };
   }
   return { fetchTools, fetchTool, callTool, baseUrl };
 }
@@ -28485,7 +28502,9 @@ var AUTO_SCREENSHOT_TOOLS = /* @__PURE__ */ new Set([
   "restart-app",
   "open-url",
   "describe",
-  "run-sequence"
+  "run-sequence",
+  // A wait is almost always followed by a look at the screen it waited for.
+  "await-ui-element"
 ]);
 var AUTO_SCREENSHOT_DELAY_MS_BY_TOOL = {
   "launch-app": 3e3,
@@ -28506,9 +28525,13 @@ var AUTO_SCREENSHOT_DELAY_MS_BY_TOOL = {
   "fold": 1500,
   "keyboard": 300,
   "paste": 300,
-  "describe": 100
+  "describe": 100,
+  // The wait returns on the first match, often mid-transition; the cap must
+  // leave `await-screen-idle` room to see two identical reads.
+  "await-ui-element": 1500
 };
 var DEFAULT_DELAY_MS = 1400;
+var CAPTURE_TIMEOUT_MS = 15e3;
 function autoScreenshotEnabled(options) {
   return !isFlagEnabled("disable-auto-screenshot", options);
 }
@@ -28528,11 +28551,13 @@ var AUTO_DESCRIBE_TOOLS = /* @__PURE__ */ new Set([
   "launch-app",
   "restart-app",
   "open-url",
-  "run-sequence"
+  "run-sequence",
+  "await-ui-element"
 ]);
 function autoDescribeEnabled(options) {
   return !isFlagEnabled("disable-auto-describe", options);
 }
+var PROFILING_CAPTURE_SKIPPED_NOTE = "Auto-screenshot and element tree skipped while a profiler is recording on this device. Call screenshot or describe if you need to see the screen.";
 var AUTO_DESCRIBE_HEADER = "--- Elements after action (describe) ---";
 function shouldAutoDescribe(toolName) {
   return AUTO_DESCRIBE_TOOLS.has(normalizeToolName(toolName));
@@ -28588,6 +28613,7 @@ function toMcpTool(t) {
     name: t.name,
     description: t.description,
     inputSchema: { type: "object", ...withoutRetiredParams(t.inputSchema) },
+    ...t.readOnly ? { annotations: { readOnlyHint: true } } : {},
     ...Object.keys(meta2).length > 0 ? { _meta: meta2 } : {}
   };
 }
@@ -28623,22 +28649,50 @@ async function fetchWithReconnect(getUrl, reconnect, config2) {
   for (let attempt = 0; attempt <= maxRetries; attempt++) {
     const controller = new AbortController();
     const timer = fetchTimeoutMs !== null ? setTimeout(() => controller.abort(), fetchTimeoutMs) : void 0;
+    let callerSignal;
     try {
       const attemptInit = typeof init === "function" ? init() : init;
-      return await fetch(getUrl(), { ...attemptInit, signal: controller.signal });
+      callerSignal = attemptInit?.signal;
+      const signal = callerSignal ? AbortSignal.any([controller.signal, callerSignal]) : controller.signal;
+      return await fetch(getUrl(), { ...attemptInit, signal });
     } catch (err) {
       lastError = err;
-      if (attempt === maxRetries) break;
-      if (attempt === 0) {
-        await reconnect();
+      if (callerSignal?.aborted || attempt === maxRetries) break;
+      const pause = (async () => {
+        if (attempt === 0) {
+          await reconnect();
+        }
+        await new Promise((r) => setTimeout(r, expBackoffBase * Math.pow(2, attempt)));
+      })();
+      await untilAborted(pause, callerSignal);
+      if (callerSignal?.aborted) {
+        lastError = callerSignal.reason;
+        break;
       }
-      await new Promise((r) => setTimeout(r, expBackoffBase * Math.pow(2, attempt)));
     } finally {
       clearTimeout(timer);
     }
   }
   throw lastError;
 }
+async function untilAborted(work, signal) {
+  if (!signal) {
+    await work;
+    return;
+  }
+  let onAbort = () => {
+  };
+  const aborted2 = new Promise((resolve10) => {
+    onAbort = resolve10;
+    if (signal.aborted) resolve10();
+    else signal.addEventListener("abort", onAbort, { once: true });
+  });
+  try {
+    await Promise.race([work, aborted2]);
+  } finally {
+    signal.removeEventListener("abort", onAbort);
+  }
+}
 function parseUrl(url) {
   let parsed;
   try {
@@ -28683,9 +28737,9 @@ function createToolCaller(deps) {
   });
   return {
     fetchTools: client.fetchTools,
-    async callTool(name, args) {
-      const { data, note, outputHint } = await client.callTool(name, args);
-      return { result: data, outputHint, note };
+    async callTool(name, args, opts) {
+      const { data, note, outputHint, profiling } = await client.callTool(name, args, opts);
+      return { result: data, outputHint, note, ...profiling ? { profiling } : {} };
     }
   };
 }
@@ -28699,6 +28753,7 @@ async function startMcpServer(options) {
   }
   const autoScreenshotOn = autoScreenshotEnabled();
   const autoDescribeOn = autoDescribeEnabled();
+  const captureTimeoutMs = options.captureTimeoutMs ?? CAPTURE_TIMEOUT_MS;
   let TOOLS_URL;
   let AUTH_TOKEN;
   const resolved = await getResolvedToolsUrl();
@@ -28758,11 +28813,75 @@ async function startMcpServer(options) {
     reconnect,
     extraHeaders: aiClientHeaders
   });
+  async function awaitScreenSettled(udid, toolName) {
+    const maxWaitMs = getAutoScreenshotDelayMs(toolName);
+    if (maxWaitMs === 0) return true;
+    const started = Date.now();
+    try {
+      const idle = await callTool(
+        "await-screen-idle",
+        { udid, timeoutMs: maxWaitMs },
+        // The tool bounds its poll to maxWaitMs; the rest covers its setup.
+        { signal: AbortSignal.timeout(maxWaitMs + captureTimeoutMs) }
+      );
+      await spyLog({
+        ts: (/* @__PURE__ */ new Date()).toISOString(),
+        event: "auto_screenshot_readiness",
+        name: toolName,
+        maxWaitMs,
+        ...idle.result
+      });
+      return idle.result?.settled === true;
+    } catch {
+      await new Promise((r) => setTimeout(r, Math.max(0, maxWaitMs - (Date.now() - started))));
+      return false;
+    }
+  }
+  async function captureScreenshotBlocks(udid, signal) {
+    try {
+      const { result } = await callTool("screenshot", { udid }, { signal });
+      const blocks = await toMcpContent(result, "image", {
+        toolsUrl: TOOLS_URL,
+        authToken: AUTH_TOKEN,
+        deviceId: udid
+      });
+      if (!blocks.some((b) => b.type === "image")) return [];
+      return [{ type: "text", text: "--- Screen after action ---" }, ...blocks];
+    } catch {
+      return [];
+    }
+  }
+  async function captureTreeBlocks(udid, toolName, signal) {
+    const t1 = Date.now();
+    try {
+      const d = await callTool("describe", { udid }, { signal });
+      const desc = d.result?.description;
+      await spyLog({
+        ts: (/* @__PURE__ */ new Date()).toISOString(),
+        event: "auto_describe",
+        name: toolName,
+        durationMs: Date.now() - t1,
+        chars: typeof desc === "string" ? desc.length : 0
+      });
+      if (typeof desc !== "string" || desc.length === 0) return [];
+      return [{ type: "text", text: `${AUTO_DESCRIBE_HEADER}
+${desc}` }];
+    } catch (e) {
+      await spyLog({
+        ts: (/* @__PURE__ */ new Date()).toISOString(),
+        event: "auto_describe",
+        name: toolName,
+        durationMs: Date.now() - t1,
+        error: String(e instanceof Error ? e.message : e)
+      });
+      return [];
+    }
+  }
   const server = new Server(
     { name: "argent", version: getInstalledVersion() },
     {
       capabilities: { tools: {} },
-      instructions: "Argent \u2014 iOS Simulator, Android Emulator, and Chromium app control for interacting, testing, profiling and debugging mobile and Chromium applications. Interaction tools return the screen after the action: a screenshot plus the accessibility element tree with normalized tap frames. Take coordinates from that tree; call describe (or debugger-component-tree) only when no fresh tree is available \u2014 never guess coordinates from pixels. On session end: call stop-all-simulator-servers with devices: [...] naming the devices this session used, and perform any necessary cleanup. One tool-server is shared by every agent using this argent install, so an unscoped call tears down their devices too \u2014 reserve it for a deliberate machine-wide cleanup. Full guidance is in the argent rule loaded from .claude/rules/argent.md."
+      instructions: "Argent \u2014 iOS Simulator, Android Emulator, and Chromium app control for interacting, testing, profiling and debugging mobile and Chromium applications. Interaction tools return the screen after the action: a screenshot plus the accessibility element tree with normalized tap frames. Take coordinates from that tree; call describe (or debugger-component-tree) only when no fresh tree is available \u2014 never guess coordinates from pixels. Do not stop simulator servers when a task finishes: the next task reuses them. Call stop-all-simulator-servers when the user asks to clean up or ends the session, or when a skill tells you to recycle a device's services, always with devices: [...] naming the devices this session used. Never stop a Metro the user started unless they ask. One tool-server is shared by every agent using this argent install, so an unscoped call tears down their devices too \u2014 reserve it for a deliberate machine-wide cleanup. Full guidance is in the argent rule loaded from .claude/rules/argent.md."
     }
   );
   server.setRequestHandler(ListToolsRequestSchema, async () => {
@@ -28791,7 +28910,7 @@ async function startMcpServer(options) {
       args: params.arguments
     });
     try {
-      const { result, outputHint, note } = await callTool(params.name, params.arguments);
+      const { result, outputHint, note, profiling } = await callTool(params.name, params.arguments);
       await spyLog({
         ts: (/* @__PURE__ */ new Date()).toISOString(),
         event: "tool_result",
@@ -28824,73 +28943,18 @@ async function startMcpServer(options) {
             text: "Auto-screenshot and element tree skipped: the input contains a {{secret:\u2026}} placeholder, and a capture of this screen could reveal the typed secret. The secret is already typed \u2014 do not send the typing step again, or the field will hold two copies of it. Submit or navigate away, then verify the resulting screen as usual. Only this call is covered: the next call is captured normally, and shows the secret if the field is still on screen. To cover the submit as well, put the typing and the submit in ONE `run-sequence` the next time you type a secret."
           }
         ];
-      } else if (udid && (wantScreenshot || wantTree)) {
-        const maxWaitMs = getAutoScreenshotDelayMs(params.name);
-        if (maxWaitMs > 0) {
-          try {
-            const idle = await callTool("await-screen-idle", { udid, timeoutMs: maxWaitMs });
-            await spyLog({
-              ts: (/* @__PURE__ */ new Date()).toISOString(),
-              event: "auto_screenshot_readiness",
-              name: params.name,
-              maxWaitMs,
-              ...idle.result
-            });
-          } catch {
-            await new Promise((r) => setTimeout(r, maxWaitMs));
-          }
-        }
-        if (wantScreenshot) {
-          try {
-            const screenshotResult = await callTool("screenshot", { udid });
-            const screenshotContent = await toMcpContent(screenshotResult.result, "image", {
-              toolsUrl: TOOLS_URL,
-              authToken: AUTH_TOKEN,
-              deviceId: udid
-            });
-            const hasImage = screenshotContent.some((b) => b.type === "image");
-            if (hasImage) {
-              content = [
-                ...content,
-                {
-                  type: "text",
-                  text: "--- Screen after action ---"
-                },
-                ...screenshotContent
-              ];
-            }
-          } catch {
-          }
-        }
-        if (wantTree) {
-          const t1 = Date.now();
-          try {
-            const d = await callTool("describe", { udid });
-            const desc = d.result?.description;
-            if (typeof desc === "string" && desc.length > 0) {
-              content = [
-                ...content,
-                { type: "text", text: `${AUTO_DESCRIBE_HEADER}
-${desc}` }
-              ];
-            }
-            await spyLog({
-              ts: (/* @__PURE__ */ new Date()).toISOString(),
-              event: "auto_describe",
-              name: params.name,
-              durationMs: Date.now() - t1,
-              chars: typeof desc === "string" ? desc.length : 0
-            });
-          } catch (e) {
-            await spyLog({
-              ts: (/* @__PURE__ */ new Date()).toISOString(),
-              event: "auto_describe",
-              name: params.name,
-              durationMs: Date.now() - t1,
-              error: String(e instanceof Error ? e.message : e)
-            });
-          }
+      } else if (udid && (wantScreenshot || wantTree) && profiling) {
+        if (params.name !== "describe") {
+          await awaitScreenSettled(udid, params.name);
+          content = [...content, { type: "text", text: PROFILING_CAPTURE_SKIPPED_NOTE }];
         }
+      } else if (udid && (wantScreenshot || wantTree)) {
+        const screenSettled = await awaitScreenSettled(udid, params.name);
+        const signal = AbortSignal.timeout(captureTimeoutMs);
+        const screenshot = wantScreenshot ? captureScreenshotBlocks(udid, signal) : Promise.resolve([]);
+        if (!screenSettled) await screenshot;
+        const tree = wantTree ? captureTreeBlocks(udid, params.name, signal) : Promise.resolve([]);
+        content = [...content, ...await screenshot, ...await tree];
       }
       if (note) {
         content = [{ type: "text", text: note }, ...content];

package/dist/tool-args.generated.d.ts

@@ -53,6 +53,8 @@ export interface ArgentToolArgs {
         force?: boolean;
         /** Android only: boot the emulator with audio output enabled. Defaults to false: argent boots emulators MUTED so several agent-driven devices don't all play sound on the host machine; pass `true` when the task involves playing, hearing, or testing audio. Takes effect at boot: if the emulator is already running muted, add `force: true` to reboot it with sound. A boot snapshot saved in the other audio mode can't be reused, so the first boot after toggling is a slower cold boot. The `boot-sound` argent flag flips this default to true. Ignored on iOS/Vega/HarmonyOS/Electron, which argent never mutes. */
         sound?: boolean;
+        /** Android on a macOS host only: boot the emulator with guest Vulkan. Defaults to false: guest Vulkan crashes Chrome's GPU process, which breaks Chrome and Custom Tab sign-in flows. Without it apps find no Vulkan device, although the guest still lists the Vulkan features, so a Vulkan-only renderer such as WebGPU (react-native-wgpu) draws nothing; pass `true` to test one. Takes effect at boot: an emulator already running in the other mode is reused with a `note`, and `force: true` reboots it. A boot snapshot saved in the other mode can't be reused, so the first boot after toggling is a slower cold boot. The `ARGENT_EMULATOR_GUEST_VULKAN` env var flips this default to true. Ignored on other hosts and platforms. */
+        guestVulkan?: boolean;
         /** iOS only: boot the simulator core WITHOUT opening its GUI window (Simulator.app, or Device Hub on Xcode 27+). The device still streams via simulator-server; used by Argent Lens. Set the `ARGENT_SIMULATOR_NO_WINDOW` env var (1/true/yes) to force this host-wide without passing the flag per call (the iOS analog of `ARGENT_EMULATOR_NO_WINDOW`). Ignored on Android/Vega/HarmonyOS/Electron, which have no equivalent GUI step. */
         headless?: boolean;
         /** Electron: path to the Electron app to launch. Either a packaged .app bundle / executable, or a project directory whose package.json points the Electron binary at the entry script. Provide exactly one of `udid`, `avdName`, `vvdImage`, `harmonyInstance`, or `electronAppPath`. */