@diegosouzacdv/jev-browser-mcp 0.7.0 → 0.7.2

package/mcp_servers/jev-browser-npm/src/server.mjs

@@ -1,79 +1,43 @@
 import { spawnSync } from "node:child_process";
 import { readlink } from "node:fs/promises";
 import path from "node:path";
-import { fileURLToPath } from "node:url";
-import { McpServer } from "@modelcontextprotocol/server";
+import { fileURLToPath } from "node:url";
+import { McpServer } from "@modelcontextprotocol/server";
 import { serveStdio } from "@modelcontextprotocol/server/stdio";
 import { chromium } from "playwright";
 import * as z from "zod/v4";
 import packageManifest from "../../../package.json" with { type: "json" };
 import { loadSettings, JevBrowserError } from "./config.mjs";
-import { chooseNextAction, createMutationAuthorizationStore, describeBrowserActions, executeBrowserFlow } from "./flow.mjs";
+import { chooseNextAction, collectFlowContractIssues, createMutationAuthorizationStore, describeBrowserActions, executeBrowserFlow } from "./flow.mjs";
+import { createBrowserFlowRunStore } from "./browser-flow-runs.mjs";
+import { executeTestSuite } from "./test-suites.mjs";
+import { cleanupUploadStagingDirectory, prepareUploadStagingDirectory } from "./upload-staging.mjs";
 
 const browserContract = describeBrowserActions();
-const actionNames = Object.keys(browserContract.actions);
 const jsonObject = z.record(z.string(), z.unknown());
-function browserStepFieldSchema(field) {
-  if (["mutating", "screenshot", "expect_download", "expect_popup", "blur", "sensitive", "full_page"].includes(field)) return z.boolean();
-  if (["timeout_ms", "timeout_seconds"].includes(field)) return z.number().positive();
-  if (["index", "context_index", "ms", "duration_ms", "status", "close_code"].includes(field)) return z.number().int();
-  if (field === "file_paths") return z.union([z.string(), z.array(z.string())]);
-  if (["near", "within", "if_visible", "json"].includes(field)) return jsonObject;
-  if (field === "name_match") return z.enum(["exact", "contains", "regex"]);
-  if (field === "condition") return z.enum(["network_idle", "angular_idle", "url", "hidden"]);
-  if (field === "standard") return z.enum(["wcag2a", "wcag2aa", "wcag21a", "wcag21aa", "wcag22aa", "section508"]);
-  return z.string();
-}
-function browserStepVariant(action, contract, aliases = []) {
-  const required = new Set(contract.required);
-  const optional = new Set([...contract.optional, "timeout_ms", "mutating", "screenshot", ...aliases]);
-  if (action === "wait") { required.delete("ms"); optional.add("ms"); }
-  if (action === "wait_for_text") { required.delete("text"); optional.add("text"); }
-  if (action === "assert_text" || action === "assert_value") { required.delete("expected"); optional.add("expected"); }
-  if (action === "upload_file") { required.delete("file_paths"); optional.add("file_paths"); }
-  if (action === "wait") optional.add("duration_ms");
-  if (action === "wait_for_text" || action === "assert_text" || action === "assert_value") optional.add("value");
-  if (action === "upload_file") optional.add("file_path");
-  if (action === "confirm_modal") {
-    required.delete("button");
-    optional.add("button");
-    optional.add("trigger");
-    optional.add("confirm");
-  }
-  const shape = { action: z.literal(action) };
-  for (const key of new Set([...required, ...optional])) {
-    const field = browserStepFieldSchema(key);
-    shape[key] = required.has(key) ? field : field.optional();
-  }
-  return z.object(shape).strict();
-}
-const browserStepVariants = Object.entries(browserContract.actions).map(([action, contract]) => browserStepVariant(action, contract));
-browserStepVariants.push(
-  browserStepVariant("fill", browserContract.actions.type, ["value"]),
-  browserStepVariant("press_key", browserContract.actions.press),
-);
-const browserStepSchema = z.discriminatedUnion("action", browserStepVariants);
-const candidatePlanSchema = z.object({
-  description: z.string(),
-  steps: z.array(browserStepSchema).min(1),
-  comment: z.string().optional(),
-  when: z.object({ visible: z.string() }).strict().optional(),
-}).strict();
 const browserOptionsShape = Object.fromEntries(browserContract.options.map((key) => {
   const value = key === "return_snapshot" ? z.enum(["diff", "full", "none"])
+    : key === "capture_network" ? z.enum(["none", "errors", "all"])
     : key === "reuse_page_match" ? z.enum(["url", "path"])
       : key === "color_scheme" ? z.enum(["light", "dark", "no-preference"])
         : key === "permissions" ? z.array(z.enum(["microphone", "geolocation"]))
           : key === "console_levels" ? z.array(z.enum(["log", "info", "debug", "warn", "error"]))
             : key === "ready" ? z.union([z.string(), z.record(z.string(), z.unknown()), z.array(z.union([z.string(), z.record(z.string(), z.unknown())]))])
               : ["allow_mutations", "block_trackers", "capture_console_errors", "capture_network_error_bodies", "capture_network_errors", "auto_angular_idle", "block_fonts", "continue_from_current_page", "dry_run", "fast_path", "local_only", "ready_network_idle", "reuse_page", "screenshot_on_failure", "screenshot_on_success", "record_video", "mobile", "snapshot_include_hidden", "include_unnamed_controls", "stop_on_expected", "trace_on_failure", "trace_on_success", "fresh_context", "clear_storage"].includes(key) ? z.boolean()
-                : ["max_flow_steps", "ready_stable_ms", "ready_timeout_seconds", "step_timeout_seconds"].includes(key) ? z.number().positive()
+                : key === "max_flow_steps" ? z.number().int().positive()
+                  : ["ready_stable_ms", "ready_timeout_seconds", "step_timeout_seconds"].includes(key) ? z.number().positive()
                   : ["busy_selectors", "login_text", "login_url_contains"].includes(key) ? z.array(z.string())
                     : ["viewport", "geolocation", "wait_for_http", "fake_media"].includes(key) ? z.union([z.string(), jsonObject])
                       : z.string();
   return [key, value.optional()];
 }));
 const browserOptionsSchema = z.object(browserOptionsShape).passthrough();
+const candidatePlanSchema = z.object({
+  description: z.string().optional(),
+  steps: z.array(z.record(z.string(), z.unknown())).optional(),
+  comment: z.string().optional(),
+  when: z.unknown().optional(),
+}).strict();
 const expectedOutcomeSchema = z.union([
   z.string(),
   z.object({
@@ -85,7 +49,7 @@ const expectedOutcomeSchema = z.union([
   }).strict().refine((value) => Boolean(value.text || value.request), "expected_outcome needs text or request"),
 ]);
 
-const browserPools = new Set();
+const browserPools = new Map();
 
 class BrowserPool {
   #settings;
@@ -97,8 +61,9 @@ class BrowserPool {
   #contextBrowsers = new Map();
   #page;
   #startedAt;
-  #everStarted = false;
-  #navigationError;
+  #everStarted = false;
+  #navigationError;
+  #unrestoredPage = false;
   #flowQueue = Promise.resolve();
   #configurationKey;
   #lastConfiguration = {};
@@ -107,12 +72,13 @@ class BrowserPool {
     this.#settings = settings;
   }
 
-  getNavigationError() {
-    return this.#navigationError;
-  }
-
-  setNavigationError(error) {
-    this.#navigationError = error || undefined;
+  getNavigationError() {
+    return this.#unrestoredPage ? "BROWSER_PAGE_NOT_RESTORED" : this.#navigationError;
+  }
+
+  setNavigationError(error) {
+    this.#navigationError = error || undefined;
+    this.#unrestoredPage = error === "BROWSER_PAGE_NOT_RESTORED";
   }
 
   async runFlow(work) {
@@ -125,9 +91,10 @@ class BrowserPool {
     } finally {
       release();
     }
-  }
-
+  }
+
   async getPage(configuration) {
+    const hadStartedBeforeRequest = this.#everStarted;
     if (configuration === undefined) configuration = this.#lastConfiguration;
     else this.#lastConfiguration = configuration;
     const configurationKey = JSON.stringify(configuration);
@@ -156,7 +123,12 @@ class BrowserPool {
         throw error;
       }
     }
-    return { page: this.#page, context: this.#context, sessionMs: isOpening ? performance.now() - this.#startedAt : 0 };
+    return {
+      page: this.#page,
+      context: this.#context,
+      sessionMs: isOpening ? performance.now() - this.#startedAt : 0,
+      recovered: isOpening && hadStartedBeforeRequest,
+    };
   }
 
   #invalidate(reason) {
@@ -290,17 +262,47 @@ class BrowserPool {
     this.#contextPromise = undefined;
     this.#navigationError = undefined;
     this.#configurationKey = undefined;
+    this.#unrestoredPage = false;
   }
 }
 
-function toolError(error, settings) {
-  let message = error instanceof JevBrowserError ? error.message : error?.message || `Jev browser flow failed (${error?.name || "Error"})`;
-  const credential = settings.env?.[settings.jev.credentialEnv];
-  if (credential) message = message.split(credential).join("[redacted]");
-  message = message.replace(/\bBearer\s+\S+/gi, "Bearer [redacted]")
-    .replace(/([?&](?:token|api[_-]?key|password|secret)=)[^&\s]+/gi, "$1[redacted]");
-  return { isError: true, content: [{ type: "text", text: `[jev-browser ${packageManifest.version}] ${message}` }] };
-}
+function safeErrorMessage(error, settings) {
+  let message = error instanceof JevBrowserError ? error.message : error?.message || `Jev browser flow failed (${error?.name || "Error"})`;
+  const credential = settings.env?.[settings.jev.credentialEnv];
+  if (credential) message = message.split(credential).join("[redacted]");
+  message = message.replace(/\bBearer\s+\S+/gi, "Bearer [redacted]")
+    .replace(/([?&](?:token|api[_-]?key|password|secret)=)[^&\s]+/gi, "$1[redacted]");
+  return message;
+}
+
+function toolError(error, settings) {
+  const message = safeErrorMessage(error, settings);
+  return { isError: true, content: [{ type: "text", text: `[jev-browser ${packageManifest.version}] ${message}` }] };
+}
+
+function isBrowserDisconnectError(error) {
+  return /(?:target page, context or browser has been closed|browser has been closed|browser disconnected|context has been closed|connection closed|connection is closed)/i
+    .test(String(error?.message || error?.name || error || ""));
+}
+
+function isBrowserStartupError(error) {
+  return /(?:could not start the configured .* browser|browser profile is in use by PID \d+)/i
+    .test(String(error?.message || error?.name || error || ""));
+}
+
+function browserEnvironmentError(reason, navigationError, extra = {}) {
+  return {
+    server_version: packageManifest.version,
+    status: "environment_error",
+    navigation_error: navigationError,
+    reason,
+    acceptance_criteria_met: false,
+    expected_outcome_visible: false,
+    steps_executed: 0,
+    steps: [],
+    ...extra,
+  };
+}
 
 export function fitToolResponse(value, maxBytes) {
   const clone = structuredClone(value && typeof value === "object" ? value : { result: value });
@@ -389,46 +391,66 @@ async function activeProfileOwnerPid(profileDir) {
   }
 }
 
-export function createJevBrowserServer({ env = process.env, configPath, fetchImpl = fetch } = {}) {
-  const settings = loadSettings({ env, ...(configPath ? { configPath } : {}) });
-  const browserPool = new BrowserPool(settings);
-  const mutationAuthorizationStore = createMutationAuthorizationStore();
-  browserPools.add(browserPool);
+export function createJevBrowserServer({ env = process.env, configPath, fetchImpl = fetch } = {}) {
+  const settings = loadSettings({ env, ...(configPath ? { configPath } : {}) });
+  const browserPool = new BrowserPool(settings);
+  const mutationAuthorizationStore = createMutationAuthorizationStore();
+  const flowRunStore = createBrowserFlowRunStore({
+    maxEntries: settings.browser.maxBackgroundFlowRuns,
+    retentionMs: settings.browser.backgroundFlowResultTtlMs,
+  });
+  browserPools.set(browserPool, settings);
   const server = new McpServer({
     name: "jev-browser",
     version: packageManifest.version,
     description: "Bounded Playwright screen flows selected by Jev through the OpenRouter Decisions API.",
   });
   const closeServer = server.close.bind(server);
-  server.close = async () => {
-    try {
-      await closeServer();
-    } finally {
-      await browserPool.close();
-      browserPools.delete(browserPool);
-    }
-  };
+  server.close = async () => {
+    try {
+      await closeServer();
+    } finally {
+      try {
+        await browserPool.close();
+      } finally {
+        browserPools.delete(browserPool);
+        await cleanupUploadStagingDirectory(settings);
+      }
+    }
+  };
 
   server.registerTool(
     "browser_health",
     {
       description: "Check the managed browser connection; set restart=true to close and restart only the browser session owned by this MCP process.",
       inputSchema: { restart: z.boolean().optional() },
-    },
+    },
     async ({ restart = false } = {}) => {
       const health = restart ? await browserPool.restart() : await browserPool.health();
+      const uploadStagingDirectory = settings.browser.managedUploadStaging
+        ? await prepareUploadStagingDirectory(settings)
+        : undefined;
       return toolText({
         ...health,
         server_version: packageManifest.version,
         browser_mode: settings.browser.mode,
         session_id: settings.browser.sessionId,
+        ...(uploadStagingDirectory ? { upload_staging_directory: uploadStagingDirectory } : {}),
         capabilities: {
           actions: Object.keys(describeBrowserActions().actions),
           options: describeBrowserActions().options,
+          limits: describeBrowserActions(settings).limits,
+          environment_profiles: Object.keys(settings.browser.environmentProfiles.profiles),
+          profile_environments: Object.fromEntries(Object.entries(settings.browser.environmentProfiles.profiles).map(([name, profile]) => [name, profile.environment])),
+          test_suite_directory: ".jev-browser/suites",
+          suite_report_format: "Markdown",
+          database_engines: ["postgresql", "oracle"],
+          remote_file_protocols: ["ftp", "ftps", "sftp"],
           media_permissions: ["microphone", "geolocation"],
           fake_media: "WAV audio input",
           storage_state: "named account snapshots",
           contexts: ["current", "new isolated context"],
+          file_upload_staging: settings.browser.managedUploadStaging ? "private project-local fixtures and uploads; managed files are cleaned after flows and suites" : "managed project-local staging unavailable",
         },
       }, settings.browser.maxToolResponseBytes);
     },
@@ -440,7 +462,7 @@ export function createJevBrowserServer({ env = process.env, configPath, fetchImp
       description: "Return the live JSON contract for every supported Jev Browser MCP action, option, alias, timeout and placeholder rule.",
       inputSchema: {},
     },
-    async () => toolText({ server_version: packageManifest.version, ...describeBrowserActions() }, settings.browser.maxToolResponseBytes),
+    async () => toolText({ server_version: packageManifest.version, ...describeBrowserActions(settings) }, settings.browser.maxToolResponseBytes),
   );
 
   server.registerTool(
@@ -470,53 +492,186 @@ export function createJevBrowserServer({ env = process.env, configPath, fetchImp
   server.registerTool(
     "run_browser_flow",
     {
-      description: "Run a bounded screen flow in an isolated per-server browser profile. Continues the current SPA page by default. Supports media fakes, tabs and contexts, read-only page inspection, conditional plans, HTTP/network/WebSocket assertions, named storage state, guarded mutation confirmation, per-step timeouts and evidence. expected_outcome accepts descriptive text or structured text/request assertions. Use describe_actions for required and optional fields and the exact options contract. Jev chooses among eligible supplied plans; a single plan skips the decision call when fast_path is enabled.",
+      description: "Run a bounded screen flow in an isolated per-server browser profile. Set background=true to return a run_id immediately and poll get_browser_flow_result instead of holding a long MCP call open. Continues the current SPA page by default. Supports media fakes, tabs and contexts, read-only page inspection, conditional plans, HTTP/network/WebSocket assertions, named storage state, guarded mutation confirmation, per-step timeouts and evidence. expected_outcome accepts descriptive text or structured text/request assertions. Use describe_actions for required and optional fields and the exact options contract. Jev chooses among eligible supplied plans; a single plan skips the decision call when fast_path is enabled.",
       inputSchema: {
-        flow: z.string(),
+        flow: z.string().optional(),
         initial_url: z.string().optional(),
         expected_outcome: expectedOutcomeSchema.optional(),
-        candidate_plans: z.record(z.string(), candidatePlanSchema),
+        candidate_plans: z.record(z.string(), candidatePlanSchema).optional(),
         params: z.record(z.string(), z.unknown()).optional(),
         options: browserOptionsSchema.optional(),
+        background: z.boolean().optional(),
       },
-    },
-    async ({ flow, initial_url: initialUrl, expected_outcome: expectedOutcome, candidate_plans: candidatePlans, params, options }) => {
-      try {
-        return toolText(await browserPool.runFlow(async () => {
-          const args = {
-          flow,
-          initialUrl,
-          expectedOutcome,
-          candidatePlans,
-          params,
-          options,
-          settings,
-          browserPool,
-          fetchImpl,
-          serverVersion: packageManifest.version,
-          mutationAuthorizationStore,
-          };
-          const first = await executeBrowserFlow(args);
-          if (first.status !== "environment_error" || first.navigation_error !== "BROWSER_DISCONNECTED"
-            || first.mutating_steps?.length || options?.confirmation_token) return first;
-          const recovery = await browserPool.restart();
-          if (!recovery.connected) return { ...first, browser_restart: recovery, retry_count: 0 };
-          const retry = await executeBrowserFlow({ ...args, options: { ...options, reuse_page: false } });
-          return { ...retry, browser_restart: recovery, retry_count: 1 };
-        }), settings.browser.maxToolResponseBytes);
-      } catch (error) {
-        return toolError(error, settings);
-      }
-    },
-  );
-
-  return server;
-}
+    },
+    async ({ flow, initial_url: initialUrl, expected_outcome: expectedOutcome, candidate_plans: candidatePlans, params, options, background = false }) => {
+      try {
+        const contractIssues = collectFlowContractIssues({ flow, candidatePlans });
+        if (contractIssues.length) throw new JevBrowserError(`flow contract validation failed:\n- ${contractIssues.join("\n- ")}`);
+        const execute = () => browserPool.runFlow(async () => {
+          const args = {
+            flow,
+            initialUrl,
+            expectedOutcome,
+            candidatePlans,
+            params,
+            options,
+            settings,
+            browserPool,
+            fetchImpl,
+            serverVersion: packageManifest.version,
+            mutationAuthorizationStore,
+          };
+          const retryAfterRestart = async (first) => {
+            const recovery = await browserPool.restart();
+            if (recovery.reason) recovery.reason = safeErrorMessage(new Error(recovery.reason), settings);
+            if (!recovery.connected) {
+              if (!initialUrl) browserPool.setNavigationError("BROWSER_PAGE_NOT_RESTORED");
+              return {
+                ...first,
+                ...(initialUrl ? {} : { navigation_error: "BROWSER_PAGE_NOT_RESTORED" }),
+                browser_restart: recovery,
+                retry_count: 0,
+              };
+            }
+            if (!initialUrl) {
+              browserPool.setNavigationError("BROWSER_PAGE_NOT_RESTORED");
+              return {
+                ...first,
+                navigation_error: "BROWSER_PAGE_NOT_RESTORED",
+                reason: "The browser reconnected, but the previous page was not restored. Provide initial_url to start a new navigation.",
+                browser_restart: recovery,
+                retry_count: 0,
+              };
+            }
+            try {
+              const retry = await executeBrowserFlow({ ...args, options: { ...options, reuse_page: false } });
+              return { ...retry, browser_restart: recovery, retry_count: 1 };
+            } catch (error) {
+              if (!isBrowserDisconnectError(error) && !isBrowserStartupError(error)) throw error;
+              return {
+                ...browserEnvironmentError(
+                  `The browser reconnected, but the flow could not restart: ${safeErrorMessage(error, settings)}`,
+                  "BROWSER_RESTART_FAILED",
+                ),
+                browser_restart: recovery,
+                retry_count: 1,
+              };
+            }
+          };
+
+          let first;
+          try {
+            first = await executeBrowserFlow(args);
+          } catch (error) {
+            if (isBrowserStartupError(error)) {
+              const profileInUse = /browser profile is in use by PID \d+/i.test(String(error.message || ""));
+              return browserEnvironmentError(
+                safeErrorMessage(error, settings),
+                profileInUse ? "BROWSER_PROFILE_IN_USE" : "BROWSER_START_FAILED",
+                { retry_count: 0 },
+              );
+            }
+            if (!isBrowserDisconnectError(error)) throw error;
+            const failedBeforeSteps = browserEnvironmentError(
+              "The browser disconnected before the flow could obtain its page.",
+              "BROWSER_DISCONNECTED",
+            );
+            return retryAfterRestart(failedBeforeSteps);
+          }
+          const disconnected = /^(?:BROWSER_DISCONNECTED|CONNECTION_CLOSED)$/.test(String(first.navigation_error || ""));
+          if (first.status !== "environment_error" || !disconnected
+            || first.steps_executed > 0 || first.mutating_steps?.length || options?.confirmation_token) return first;
+          return retryAfterRestart(first);
+        });
+        if (background) {
+          const started = flowRunStore.start(async () => {
+            try {
+              return fitToolResponse(await execute(), Math.floor(settings.browser.maxToolResponseBytes * 0.8));
+            } catch (error) {
+              throw new Error(safeErrorMessage(error, settings));
+            }
+          });
+          return toolText(started, settings.browser.maxToolResponseBytes);
+        }
+        return toolText(await execute(), settings.browser.maxToolResponseBytes);
+      } catch (error) {
+        return toolError(error, settings);
+      }
+    },
+  );
+
+  server.registerTool(
+    "run_test_suite",
+    {
+      description: "Run a validated JSON suite from .jev-browser/suites/<suite_name>.json sequentially, retaining the browser and integration session across setup, acceptance criteria and always-attempted teardown. Produces a Markdown evidence report and cleans project-local fixtures after teardown. Set background=true to poll the long run with get_browser_flow_result.",
+      inputSchema: {
+        suite_name: z.string().min(1).max(80),
+        params: z.record(z.string(), z.unknown()).optional(),
+        environment_profile: z.string().min(1).max(48).optional(),
+        report_path: z.string().min(1).max(2048).optional(),
+        background: z.boolean().optional(),
+      },
+    },
+    async ({ suite_name: suiteName, params, environment_profile: environmentProfile, report_path: reportPath, background = false }) => {
+      try {
+        const execute = () => browserPool.runFlow(() => executeTestSuite({
+          suiteName,
+          params,
+          environmentProfile,
+          reportPath,
+          settings,
+          browserPool,
+          fetchImpl,
+          serverVersion: packageManifest.version,
+          mutationAuthorizationStore,
+          executeFlow: executeBrowserFlow,
+        }));
+        if (background) {
+          const started = flowRunStore.start(async () => {
+            try { return fitToolResponse(await execute(), Math.floor(settings.browser.maxToolResponseBytes * 0.8)); }
+            catch (error) { throw new Error(safeErrorMessage(error, settings)); }
+          });
+          return toolText(started, settings.browser.maxToolResponseBytes);
+        }
+        return toolText(await execute(), settings.browser.maxToolResponseBytes);
+      } catch (error) {
+        return toolError(error, settings);
+      }
+    },
+  );
+
+  server.registerTool(
+    "get_browser_flow_result",
+    {
+      description: "Poll a background run_browser_flow call. Set wait_ms up to 30000 to long-poll; completed results are retained for 30 minutes in this MCP process.",
+      inputSchema: {
+        run_id: z.string().min(1),
+        wait_ms: z.number().int().min(0).max(30000).optional(),
+      },
+    },
+    async ({ run_id: runId, wait_ms: waitMs = 0 }) => {
+      const result = await flowRunStore.get(runId, waitMs);
+      if (!result) return toolError(new JevBrowserError("background flow run was not found or its result expired"), settings);
+      return toolText(result, settings.browser.maxToolResponseBytes);
+    },
+  );
+
+  return server;
+}
 
-export async function closeBrowserSessions() {
-  await Promise.all([...browserPools].map((pool) => pool.close()));
-  browserPools.clear();
-}
+export async function closeBrowserSessions() {
+  try {
+    await Promise.all([...browserPools].map(async ([pool, settings]) => {
+      try {
+        await pool.close();
+      } finally {
+        await cleanupUploadStagingDirectory(settings);
+      }
+    }));
+  } finally {
+    browserPools.clear();
+  }
+}
 
 export function installConfiguredBrowser({ env = process.env } = {}) {
   const settings = loadSettings({ env });