doc-detective 4.7.0 → 4.8.0

package/dist/core/tests.js

@@ -6,8 +6,9 @@ import kill from "tree-kill";
 // so we don't carry a top-level `import type` whose `typeof` would refer
 // to a non-runtime identifier.
 import { loadHeavyDep, resolveHeavyDepPath } from "../runtime/loader.js";
+import { requiredBrowserAssets, ensureBrowserInstalled, } from "../runtime/browsers.js";
 import os from "node:os";
-import { log, replaceEnvs, selectSpecsForRun, findFreePort } from "./utils.js";
+import { log, replaceEnvs, selectSpecsForRun, findFreePort, runConcurrent, rollUpResults, createAppiumPool, } from "./utils.js";
 import axios from "axios";
 import { instantiateCursor } from "./tests/moveTo.js";
 import { goTo } from "./tests/goTo.js";
@@ -31,13 +32,13 @@ import { spawn } from "node:child_process";
 import { randomUUID } from "node:crypto";
 import { setAppiumHome } from "./appium.js";
 import { resolveExpression } from "./expressions.js";
-import { getEnvironment, getAvailableApps, clearAppCache } from "./config.js";
+import { getEnvironment, getAvailableApps, clearAppCache, resolveConcurrentRunners, } from "./config.js";
 import { uploadChangedFiles } from "./integrations/index.js";
 import http from "node:http";
 import https from "node:https";
 import { fileURLToPath } from "node:url";
 const __dirname = path.dirname(fileURLToPath(import.meta.url));
-export { runSpecs, runViaApi, getRunner, ensureChromeAvailable, getDriverCapabilities, getDefaultBrowser, isSupportedContext, };
+export { runSpecs, runViaApi, getRunner, ensureChromeAvailable, ensureContextBrowserInstalled, combinationKey, warmUpDecision, selectWarmUpTargets, getDriverCapabilities, getDefaultBrowser, isSupportedContext, };
 // exports.appiumStart = appiumStart;
 // exports.appiumIsReady = appiumIsReady;
 // exports.driverStart = driverStart;
@@ -57,6 +58,30 @@ const driverActions = [
 // Browser names getDriverCapabilities knows how to build caps for. `safari` is
 // rewritten to `webkit` during context resolution, so both appear here.
 const KNOWN_BROWSERS = ["firefox", "chrome", "safari", "webkit"];
+/**
+ * Stable identity for a "context combination" — the platform + browser pairing
+ * that determines whether a driver session can be created. The runner memoizes
+ * warm-up outcomes by this key so a combination that fails to start once isn't
+ * re-attempted (with its slow driverStart backoff) for every later context.
+ * headless is intentionally excluded: headed/headless are two attempts at the
+ * same combination (the loop retries headless on failure), not distinct ones.
+ * `webkit` is normalized to `safari` so the key matches getAvailableApps naming.
+ */
+function combinationKey(context) {
+    const rawName = context?.browser?.name;
+    const name = rawName === "webkit" ? "safari" : rawName || "<none>";
+    return `${context?.platform}::${name}`;
+}
+/**
+ * Decide whether a context combination should be attempted or skipped, given
+ * its prior warm-up outcome in this run. Pure so the memoization branching is
+ * unit-testable without spinning up Appium. A previously-failed combination is
+ * skipped outright; everything else is attempted (and its outcome recorded by
+ * the caller).
+ */
+function warmUpDecision(prev) {
+    return prev === "failed" ? "skip" : "attempt";
+}
 // Get Appium driver capabilities and apply options.
 function getDriverCapabilities({ runnerDetails, name, options }) {
     let capabilities = {};
@@ -129,8 +154,15 @@ function getDriverCapabilities({ runnerDetails, name, options }) {
                 args.push(`--auto-select-desktop-capture-source=RECORD_ME`);
                 if (options.headless)
                     args.push("--headless", "--disable-gpu");
-                if (process.platform === "linux")
+                if (process.platform === "linux") {
                     args.push("--no-sandbox");
+                    // Chrome writes shared memory to /dev/shm, which is only ~64MB on
+                    // many Linux/CI hosts. A single browser fits, but several launched
+                    // at once under concurrentRunners exhaust it and ChromeDriver
+                    // "crashed during startup". Redirect that allocation to /tmp so
+                    // parallel browser contexts start reliably.
+                    args.push("--disable-dev-shm-usage");
+                }
                 // Set capabilities
                 capabilities = {
                     platformName: runnerDetails.environment.platform,
@@ -157,24 +189,11 @@ function getDriverCapabilities({ runnerDetails, name, options }) {
     }
     return capabilities;
 }
-// Check if any steps require an Appium driver.
-function isAppiumRequired(specs) {
-    let appiumRequired = false;
-    specs.forEach((spec) => {
-        spec.tests.forEach((test) => {
-            test.contexts.forEach((context) => {
-                // Check if test includes actions that require a driver.
-                if (isDriverRequired({ test: context })) {
-                    appiumRequired = true;
-                }
-            });
-        });
-    });
-    return appiumRequired;
-}
 function isDriverRequired({ test }) {
     let driverRequired = false;
-    test.steps.forEach((step) => {
+    // The resolved shape doesn't guarantee `steps` — treat a stepless test or
+    // context as needing no driver instead of throwing.
+    (test.steps || []).forEach((step) => {
         // Check if test includes actions that require a driver.
         driverActions.forEach((action) => {
             if (typeof step[action] !== "undefined")
@@ -376,9 +395,11 @@ async function runViaApi({ resolvedTests, apiKey, config = {} }) {
 /**
  * Orchestrates execution of resolved test specifications and returns a hierarchical run report.
  *
- * Executes each spec -> test -> context -> step, conditionally starts Appium and browser drivers,
- * applies viewport/window sizing, handles unsafe-step policies and recording, aggregates per-step,
- * per-context, per-test, and per-spec results, and performs resource cleanup.
+ * Flattens every context across all specs and tests into one job list and runs it through a
+ * worker pool sized by config.concurrentRunners (default 1 = sequential). Conditionally starts
+ * Appium and browser drivers, applies viewport/window sizing, handles unsafe-step policies and
+ * recording, then rolls per-step, per-context, per-test, and per-spec results up in a
+ * deterministic post-pass. Report order always matches input order.
  *
  * @param {Object} resolvedTests - Resolved test bundle containing configuration and specs to run.
  * @param {Object} resolvedTests.config - Runner configuration used during execution.
@@ -422,9 +443,18 @@ async function runSpecs({ resolvedTests }) {
     };
     // Set initial shorthand values
     const platform = runnerDetails.environment.platform;
-    const availableApps = runnerDetails.availableApps;
+    // `let`, not `const`: an on-demand browser install during the context loop
+    // re-detects available apps and reassigns this snapshot (see the support
+    // gate below).
+    let availableApps = runnerDetails.availableApps;
     const metaValues = { specs: {} };
-    let appium;
+    // Per-run memoization, shared across the concurrent context pool below.
+    // installAttempts keeps a browser's on-demand install from being retried for
+    // every context that uses it; warmUpResults keeps a context combination that
+    // can't start a driver from being re-attempted (with its slow driverStart
+    // backoff) for the rest of the run.
+    const installAttempts = new Map();
+    const warmUpResults = new Map();
     const report = {
         summary: {
             specs: {
@@ -454,377 +484,191 @@ async function runSpecs({ resolvedTests }) {
         },
         specs: [],
     };
-    // Determine which apps are required
-    const appiumRequired = isAppiumRequired(specs);
-    // Warm up Appium
-    let appiumPort;
-    if (appiumRequired) {
-        setAppiumHome({ cacheDir: config?.cacheDir });
-        appiumPort = await findFreePort();
-        log(config, "debug", `Starting Appium on port ${appiumPort}`);
-        // Resolve appium's actual JS entrypoint via `require.resolve`
-        // (shim node_modules first, runtime cache second) and invoke it
-        // with `node <entry>`. This sidesteps every shell-injection trap
-        // at once: no `.cmd` shim, so no Windows-requires-shell:true; no
-        // `npx`, so no PATH lookup; no user-controlled paths in a shell-
-        // interpreted string. Works for both `--omit=optional` users
-        // (appium in cache only) and default installs (appium in shim).
-        const appiumEntry = resolveHeavyDepPath("appium", { cacheDir: config?.cacheDir });
-        if (!appiumEntry) {
-            throw new Error("appium is not installed. The runtime pre-flight should have installed it; check DOC_DETECTIVE_CACHE_DIR / config.cacheDir or run `doc-detective install runtime appium`.");
-        }
-        appium = spawn(process.execPath, [appiumEntry, "-a", "127.0.0.1", "-p", String(appiumPort)], {
-            windowsHide: true,
-            cwd: path.join(__dirname, "../.."),
-        });
-        appium.on("error", (err) => {
-            log(config, "warning", `Appium process error: ${err?.stack ?? err?.message ?? String(err)}`);
-        });
-        appium.stdout.on("data", (data) => {
-            // console.log(`stdout: ${data}`);
-        });
-        appium.stderr.on("data", (data) => {
-            // console.error(`stderr: ${data}`);
-        });
-        try {
-            await appiumIsReady(appiumPort);
-        }
-        catch (error) {
-            // appiumIsReady threw or timed out — the spawned child is still
-            // alive and would leak (orphan process, port still bound). Tear
-            // it down before propagating so subsequent runs don't trip on
-            // the stale state.
-            try {
-                if (appium && appium.pid)
-                    kill(appium.pid);
-            }
-            catch {
-                // best-effort cleanup; the parent error is what matters
-            }
-            throw error;
-        }
-        log(config, "debug", "Appium is ready.");
-    }
-    // Iterate specs
+    // Resolve concurrency up front (defensive re-resolve: API callers can hand
+    // runSpecs a config that never went through core setConfig, leaving
+    // concurrentRunners as `true`). Drives both the worker pool and how many
+    // Appium servers to start.
+    const limit = resolveConcurrentRunners(config);
+    // Phase 1: pre-build the report skeleton and a flat list of context jobs
+    // across all specs and tests. Slots are pre-assigned so report order always
+    // matches input order, no matter what order concurrent contexts finish in.
     log(config, "info", "Running test specs.");
+    const jobs = [];
     for (const spec of specs) {
         log(config, "debug", `SPEC: ${spec.specId}`);
-        // Set spec report
-        let specReport = {
+        // Create-if-missing: specIds (and testIds) aren't guaranteed unique
+        // across the run, and all registration now happens up front — an
+        // overwrite here would wipe an earlier spec's registered tests.
+        metaValues.specs[spec.specId] ??= { tests: {} };
+        const specReport = {
             specId: spec.specId,
             description: spec.description,
             contentPath: spec.contentPath,
             tests: [],
         };
-        // Set meta values
-        metaValues.specs[spec.specId] = { tests: {} };
-        // Iterates tests
+        report.specs.push(specReport);
         for (const test of spec.tests) {
             log(config, "debug", `TEST: ${test.testId}`);
-            // Set test report
-            let testReport = {
+            metaValues.specs[spec.specId].tests[test.testId] ??= { contexts: {} };
+            const testReport = {
                 testId: test.testId,
                 description: test.description,
                 contentPath: test.contentPath,
                 detectSteps: test.detectSteps,
-                contexts: [],
+                contexts: new Array(test.contexts.length),
             };
-            // Set meta values
-            metaValues.specs[spec.specId].tests[test.testId] = { contexts: {} };
-            // Iterate contexts
-            // TODO: Support both serial and parallel execution
-            for (const context of test.contexts) {
-                // If "platform" is not defined, set it to the current platform
-                if (!context.platform)
-                    context.platform = runnerDetails.environment.platform;
-                // Attach OpenAPI definitions to context
-                if (config.integrations?.openApi) {
-                    context.openApi = [
-                        ...(context.openApi || []),
-                        ...config.integrations.openApi,
-                    ];
+            specReport.tests.push(testReport);
+            test.contexts.forEach((context, slot) => {
+                context.contextId = context.contextId || randomUUID();
+                jobs.push({ spec, test, context, contexts: testReport.contexts, slot });
+            });
+        }
+    }
+    if (limit > 1 &&
+        jobs.some((job) => job.context.steps?.some((step) => step.record !== undefined))) {
+        // Concurrent headed recordings capture via getDisplayMedia and can grab
+        // the wrong window. Recording filenames in the temp dir can also collide.
+        log(config, "warning", "Tests include record steps while concurrentRunners is greater than 1. Concurrent recordings can capture the wrong window; set concurrentRunners to 1 for recording runs.");
+    }
+    // Start one Appium server per concurrent runner that will actually use a
+    // driver (capped at the number of driver contexts). Each server owns a
+    // distinct port, so parallel contexts never create sessions on the same
+    // server — that contention crashed ChromeDriver when every context shared
+    // one server. Non-driver runs start none.
+    const driverJobCount = jobs.filter((job) => isDriverRequired({ test: job.context })).length;
+    let appiumServers = [];
+    let appiumPool;
+    if (driverJobCount > 0) {
+        setAppiumHome({ cacheDir: config?.cacheDir });
+        // Resolve appium's actual JS entrypoint via `require.resolve` (shim
+        // node_modules first, runtime cache second) and invoke it with
+        // `node <entry>`. This sidesteps every shell-injection trap at once: no
+        // `.cmd` shim, so no Windows-requires-shell:true; no `npx`, so no PATH
+        // lookup; no user-controlled paths in a shell-interpreted string. Works
+        // for both `--omit=optional` users (appium in cache only) and default
+        // installs (appium in shim).
+        const appiumEntry = resolveHeavyDepPath("appium", {
+            cacheDir: config?.cacheDir,
+        });
+        if (!appiumEntry) {
+            throw new Error("appium is not installed. The runtime pre-flight should have installed it; check DOC_DETECTIVE_CACHE_DIR / config.cacheDir or run `doc-detective install runtime appium`.");
+        }
+        const serverCount = Math.min(limit, driverJobCount);
+        log(config, "debug", `Starting ${serverCount} Appium server(s).`);
+        // Start servers one at a time rather than all at once: concurrent
+        // findFreePort() calls share a close-to-rebind window (two could hand out
+        // the same port), and spawning every Appium at once spikes CPU during
+        // startup. Sequential startup is a one-time per-run cost (serverCount <= 4)
+        // that removes the port race and fails fast on the first server that can't
+        // come up, tearing down any already started so they don't leak.
+        try {
+            for (let i = 0; i < serverCount; i++) {
+                appiumServers.push(await startAppiumServer(appiumEntry, config));
+            }
+        }
+        catch (error) {
+            for (const server of appiumServers) {
+                try {
+                    kill(server.process.pid);
                 }
-                // If "browser" isn't defined but is required by the test, set it to the first available browser in the sequence of Firefox, Chrome, Safari
-                if (!context.browser && isDriverRequired({ test: context })) {
-                    context.browser = getDefaultBrowser({ runnerDetails });
+                catch {
+                    // best-effort
                 }
-                // Set context report
-                let contextReport = {
-                    contextId: context.contextId || randomUUID(),
-                    platform: context.platform,
-                    browser: context.browser,
+            }
+            throw error;
+        }
+        appiumPool = createAppiumPool(appiumServers.map((s) => s.port));
+    }
+    // Everything that uses the Appium servers runs inside this try so the
+    // shutdown in `finally` always reaches them — otherwise a throw in
+    // warmUpContexts (e.g. getAvailableApps failing during the re-detect) would
+    // leak the started servers, leaving orphaned processes bound to their ports.
+    try {
+        // For concurrent runs, resolve missing browser dependencies and warm up
+        // each unique driver combination serially *before* the pool. Two contexts
+        // can't then race on an on-demand install (which mutates the shared app
+        // cache), and a combination that can't start a driver is recorded once here
+        // so every parallel context sharing it skips instantly instead of re-paying
+        // driverStart's backoff. This pre-populates installAttempts /
+        // warmUpResults / runnerDetails.availableApps, so runContext's own gates
+        // below collapse to fast cache hits. Sequential runs (limit 1) keep #338's
+        // natural first-context-warms-up behavior in runContext — no pre-pass, no
+        // extra driver start, byte-identical to before.
+        if (limit > 1 && appiumPool) {
+            await warmUpContexts({
+                jobs,
+                config,
+                runnerDetails,
+                appiumPool,
+                installAttempts,
+                warmUpResults,
+            });
+        }
+        // Phase 2: run every context job through one flat worker pool. A limit of 1
+        // (the default) is strictly sequential in input order.
+        await runConcurrent(jobs, limit, async (job) => {
+            try {
+                job.contexts[job.slot] = await runContext({
+                    config,
+                    spec: job.spec,
+                    test: job.test,
+                    context: job.context,
+                    runnerDetails,
+                    appiumPool,
+                    metaValues,
+                    installAttempts,
+                    warmUpResults,
+                    logPrefix: limit > 1 ? `[${job.test.testId}/${job.context.contextId}]` : "",
+                });
+            }
+            catch (error) {
+                // Error isolation: one crashing context must not abort sibling jobs.
+                // Guard against non-Error throws (a thrown string/object has no
+                // .message) so the real failure detail survives in logs and report.
+                const detail = error?.message ?? String(error);
+                log(config, "error", `Context '${job.context.contextId}' crashed: ${detail}`);
+                job.contexts[job.slot] = {
+                    contextId: job.context.contextId,
+                    platform: job.context.platform,
+                    browser: job.context.browser,
+                    result: "FAIL",
+                    resultDescription: `Unexpected error: ${detail}`,
                     steps: [],
                 };
-                // Set meta values
-                metaValues.specs[spec.specId].tests[test.testId].contexts[context.contextId] = { steps: {} };
-                // If a driver is required but no browser could be resolved (e.g.
-                // getDefaultBrowser found nothing installed, or the context supplied a
-                // browser object with no name), skip with an explicit reason instead of
-                // letting it fail later as "Failed to start context 'undefined'".
-                if (isDriverRequired({ test: context }) && !context.browser?.name) {
-                    const errorMessage = `Skipping context on '${context.platform}': no supported browser is available in the current environment.`;
-                    log(config, "warning", errorMessage);
-                    contextReport = {
-                        ...contextReport,
-                        result: "SKIPPED",
-                        resultDescription: errorMessage,
-                    };
-                    report.summary.contexts.skipped++;
-                    testReport.contexts.push(contextReport);
-                    continue;
-                }
-                // Check if current environment supports given contexts
-                const supportedContext = isSupportedContext({
-                    context: context,
-                    apps: availableApps,
-                    platform: platform,
-                });
-                // If context isn't supported, skip it
-                if (!supportedContext) {
-                    log(config, "info", `Skipping context. The current system doesn't support this context: {"platform": "${context.platform}", "apps": ${JSON.stringify(context.apps)}}`);
-                    contextReport = { result: "SKIPPED", ...contextReport };
-                    report.summary.contexts.skipped++;
-                    testReport.contexts.push(contextReport);
-                    continue;
-                }
-                log(config, "debug", `CONTEXT:\n${JSON.stringify(context, null, 2)}`);
-                let driver;
-                // Ensure context contains a 'steps' property
-                if (!context.steps) {
-                    context.steps = [];
-                }
-                const driverRequired = isDriverRequired({ test: context });
-                if (driverRequired) {
-                    // Define driver capabilities
-                    // TODO: Support custom apps
-                    let caps = getDriverCapabilities({
-                        runnerDetails: runnerDetails,
-                        name: context.browser.name,
-                        options: {
-                            width: context.browser?.window?.width || 1200,
-                            height: context.browser?.window?.height || 800,
-                            headless: context.browser?.headless !== false,
-                        },
-                    });
-                    log(config, "debug", "CAPABILITIES:");
-                    log(config, "debug", caps);
-                    if (appiumPort === undefined) {
-                        throw new Error("Driver requested but Appium was not started. " +
-                            "isAppiumRequired(specs) and isDriverRequired(context) disagreed; this is a bug.");
-                    }
-                    // Instantiate driver
-                    try {
-                        driver = await driverStart(caps, appiumPort, 4, { cacheDir: config?.cacheDir });
-                    }
-                    catch (error) {
-                        try {
-                            // If driver fails to start, try again as headless
-                            log(config, "warning", `Failed to start context '${context.browser?.name}' on '${platform}'. Retrying as headless.`);
-                            context.browser.headless = true;
-                            caps = getDriverCapabilities({
-                                runnerDetails: runnerDetails,
-                                name: context.browser.name,
-                                options: {
-                                    width: context.browser?.window?.width || 1200,
-                                    height: context.browser?.window?.height || 800,
-                                    headless: context.browser?.headless !== false,
-                                },
-                            });
-                            driver = await driverStart(caps, appiumPort, 4, { cacheDir: config?.cacheDir });
-                        }
-                        catch (error) {
-                            let errorMessage = `Failed to start context '${context.browser?.name}' on '${platform}'.`;
-                            if (context.browser?.name === "safari")
-                                errorMessage =
-                                    errorMessage +
-                                        " Make sure you've run `safaridriver --enable` in a terminal and enabled 'Allow Remote Automation' in Safari's Develop menu.";
-                            log(config, "error", errorMessage);
-                            contextReport = {
-                                result: "SKIPPED",
-                                resultDescription: errorMessage,
-                                ...contextReport,
-                            };
-                            report.summary.contexts.skipped++;
-                            testReport.contexts.push(contextReport);
-                            continue;
-                        }
-                    }
-                    if (context.browser?.viewport?.width ||
-                        context.browser?.viewport?.height) {
-                        // Set driver viewport size
-                        await setViewportSize(context, driver);
-                    }
-                    else if (context.browser?.window?.width ||
-                        context.browser?.window?.height) {
-                        // Get driver window size
-                        const windowSize = await driver.getWindowSize();
-                        // Resize window if necessary
-                        await driver.setWindowSize(context.browser?.window?.width || windowSize.width, context.browser?.window?.height || windowSize.height);
-                    }
-                }
-                // Iterates steps
-                let stepExecutionFailed = false;
-                for (let step of context.steps) {
-                    // Set step id if not defined
-                    if (!step.stepId)
-                        step.stepId = randomUUID();
-                    log(config, "debug", `STEP:\n${JSON.stringify(step, null, 2)}`);
-                    if (step.unsafe && runnerDetails.allowUnsafeSteps === false) {
-                        log(config, "warning", `Skipping unsafe step: ${step.description} in test ${test.testId} context ${context.contextId}`);
-                        // Mark as skipped
-                        const stepReport = {
-                            ...step,
-                            result: "SKIPPED",
-                            resultDescription: "Skipped because unsafe steps aren't allowed.",
-                        };
-                        contextReport.steps.push(stepReport);
-                        report.summary.steps.skipped++;
-                        continue;
-                    }
-                    if (stepExecutionFailed) {
-                        // Mark as skipped
-                        const stepReport = {
-                            ...step,
-                            result: "SKIPPED",
-                            resultDescription: "Skipped due to previous failure in context.",
-                        };
-                        contextReport.steps.push(stepReport);
-                        report.summary.steps.skipped++;
+            }
+        });
+        // Phase 3: roll results up the tree and count the summary in one
+        // deterministic pass after all contexts have finished.
+        for (const specReport of report.specs) {
+            for (const testReport of specReport.tests) {
+                for (const contextReport of testReport.contexts) {
+                    // Every slot is assigned by the pool callback (even on crash), so
+                    // this guard should never fire — it documents the invariant and
+                    // keeps a future gap from surfacing as a cryptic undefined read.
+                    if (!contextReport)
                         continue;
+                    for (const stepReport of contextReport.steps) {
+                        report.summary.steps[stepReport.result.toLowerCase()]++;
                     }
-                    // Set meta values
-                    metaValues.specs[spec.specId].tests[test.testId].contexts[context.contextId].steps[step.stepId] = {};
-                    // Run step
-                    const stepResult = await runStep({
-                        config: config,
-                        context: context,
-                        step: step,
-                        driver: driver,
-                        metaValues: metaValues,
-                        options: {
-                            openApiDefinitions: context.openApi || [],
-                        },
-                    });
-                    log(config, "debug", `RESULT: ${stepResult.status}\n${JSON.stringify(stepResult, null, 2)}`);
-                    stepResult.result = stepResult.status;
-                    stepResult.resultDescription = stepResult.description;
-                    delete stepResult.status;
-                    delete stepResult.description;
-                    // Add step result to report
-                    const stepReport = {
-                        ...step,
-                        ...stepResult,
-                    };
-                    contextReport.steps.push(stepReport);
-                    report.summary.steps[stepReport.result.toLowerCase()]++;
-                    // If this step failed, set flag to skip remaining steps
-                    if (stepReport.result === "FAIL") {
-                        stepExecutionFailed = true;
-                    }
-                }
-                // If recording, stop recording
-                if (config.recording) {
-                    const stopRecordStep = {
-                        stopRecord: true,
-                        description: "Stopping recording",
-                        stepId: randomUUID(),
-                    };
-                    const stepResult = await runStep({
-                        config: config,
-                        context: context,
-                        step: stopRecordStep,
-                        driver: driver,
-                        options: {
-                            openApiDefinitions: context.openApi || [],
-                        },
-                    });
-                    stepResult.result = stepResult.status;
-                    stepResult.resultDescription = stepResult.description;
-                    delete stepResult.status;
-                    delete stepResult.description;
-                    // Add step result to report
-                    const stepReport = {
-                        ...stopRecordStep,
-                        ...stepResult,
-                    };
-                    contextReport.steps.push(stepReport);
-                    report.summary.steps[stepReport.result.toLowerCase()]++;
-                }
-                // Parse step results to calc context result
-                // If any step fails, context fails
-                let contextResult;
-                if (contextReport.steps.find((step) => step.result === "FAIL"))
-                    contextResult = "FAIL";
-                // If any step warns, context warns
-                else if (contextReport.steps.find((step) => step.result === "WARNING"))
-                    contextResult = "WARNING";
-                // If all steps skipped, context skipped
-                else if (contextReport.steps.length ===
-                    contextReport.steps.filter((step) => step.result === "SKIPPED").length)
-                    contextResult = "SKIPPED";
-                // If all steps pass, context passes
-                else
-                    contextResult = "PASS";
-                contextReport = { result: contextResult, ...contextReport };
-                testReport.contexts.push(contextReport);
-                report.summary.contexts[contextResult.toLowerCase()]++;
-                if (driverRequired) {
-                    // Close driver
-                    try {
-                        await driver.deleteSession();
-                    }
-                    catch (error) {
-                        log(config, "error", `Failed to delete driver session: ${error.message}`);
-                    }
+                    report.summary.contexts[contextReport.result.toLowerCase()]++;
                 }
+                testReport.result = rollUpResults(testReport.contexts.filter(Boolean));
+                report.summary.tests[testReport.result.toLowerCase()]++;
             }
-            // Parse context results to calc test result
-            // If any context fails, test fails
-            let testResult;
-            if (testReport.contexts.find((context) => context.result === "FAIL"))
-                testResult = "FAIL";
-            // If any context warns, test warns
-            else if (testReport.contexts.find((context) => context.result === "WARNING"))
-                testResult = "WARNING";
-            // If all contexts skipped, test skipped
-            else if (testReport.contexts.length ===
-                testReport.contexts.filter((context) => context.result === "SKIPPED")
-                    .length)
-                testResult = "SKIPPED";
-            // If all contexts pass, test passes
-            else
-                testResult = "PASS";
-            testReport = { result: testResult, ...testReport };
-            specReport.tests.push(testReport);
-            report.summary.tests[testResult.toLowerCase()]++;
+            specReport.result = rollUpResults(specReport.tests);
+            report.summary.specs[specReport.result.toLowerCase()]++;
         }
-        // Parse test results to calc spec result
-        // If any context fails, test fails
-        let specResult;
-        if (specReport.tests.find((test) => test.result === "FAIL"))
-            specResult = "FAIL";
-        // If any test warns, spec warns
-        else if (specReport.tests.find((test) => test.result === "WARNING"))
-            specResult = "WARNING";
-        // If all tests skipped, spec skipped
-        else if (specReport.tests.length ===
-            specReport.tests.filter((test) => test.result === "SKIPPED").length)
-            specResult = "SKIPPED";
-        // If all contexts pass, test passes
-        else
-            specResult = "PASS";
-        specReport = { result: specResult, ...specReport };
-        report.specs.push(specReport);
-        report.summary.specs[specResult.toLowerCase()]++;
     }
-    // Close appium server
-    if (appium) {
-        log(config, "debug", "Closing Appium server");
-        try {
-            kill(appium.pid);
-        }
-        catch {
-            // Process may already be terminated
+    finally {
+        // Close every Appium server we started.
+        for (const server of appiumServers) {
+            log(config, "debug", `Closing Appium server on port ${server.port}`);
+            try {
+                kill(server.process.pid);
+            }
+            catch {
+                // Process may already be terminated
+            }
         }
     }
     // Upload changed files back to source integrations (best-effort)
@@ -853,6 +697,468 @@ async function runSpecs({ resolvedTests }) {
     }
     return report;
 }
+/**
+ * Pick which contexts warmUpContexts should warm up: one representative per
+ * unique platform::browser combination among the driver-required jobs. Applies
+ * the same platform default and default-browser resolution runContext uses, so
+ * the combination keys it produces match the ones runContext looks up in the
+ * pool. Non-driver and browserless contexts are excluded. Mutates
+ * context.platform / context.browser in place — idempotent, since runContext
+ * applies the identical defaults. Pure (no I/O) so the selection + de-dup +
+ * normalization logic is unit-testable without Appium.
+ */
+function selectWarmUpTargets(jobs, runnerDetails) {
+    const platform = runnerDetails.environment.platform;
+    const seen = new Set();
+    const targets = [];
+    for (const job of jobs) {
+        const context = job.context;
+        if (!context.steps)
+            context.steps = [];
+        // Default platform to the runner's, matching runContext. Without this a
+        // resolved context of `{}` (no runOn — the common case) keys as
+        // `undefined::<browser>`, fails the support check, and is skipped — which
+        // would defeat the warm-up/install de-racing the pre-pass exists for.
+        if (!context.platform)
+            context.platform = platform;
+        if (!context.browser && isDriverRequired({ test: context })) {
+            context.browser = getDefaultBrowser({ runnerDetails });
+        }
+        if (!isDriverRequired({ test: context }))
+            continue;
+        // No resolvable browser — runContext skips these per-context with its own
+        // message; nothing to warm up.
+        if (!context.browser?.name)
+            continue;
+        const combo = combinationKey(context);
+        if (seen.has(combo))
+            continue;
+        seen.add(combo);
+        targets.push({ context, combo });
+    }
+    return targets;
+}
+/**
+ * Serial pre-pass for concurrent runs. For each unique driver combination
+ * (platform::browser) among the jobs, resolves a missing browser dependency on
+ * demand and then warms up a driver once, recording the outcome. Runs before
+ * the worker pool so:
+ *   - on-demand installs never race (they mutate the shared app cache), and
+ *   - a combination that can't start a driver is recorded once, so every
+ *     parallel context sharing it is skipped instantly by runContext's warm-up
+ *     gate instead of each re-paying driverStart's retry/backoff.
+ * Mirrors the install + driver-start logic in runContext so the memoization
+ * state (installAttempts / warmUpResults / runnerDetails.availableApps) is
+ * identical to what the first same-combo context would have produced serially.
+ */
+async function warmUpContexts({ jobs, config, runnerDetails, appiumPool, installAttempts, warmUpResults, }) {
+    const platform = runnerDetails.environment.platform;
+    // Which unique combinations to warm up (with the same platform/browser
+    // normalization runContext applies) is extracted into selectWarmUpTargets so
+    // it can be unit-tested without spinning up Appium.
+    for (const { context } of selectWarmUpTargets(jobs, runnerDetails)) {
+        const combo = combinationKey(context);
+        // On-demand install + re-detect (serial), mirroring runContext's gate.
+        let supported = isSupportedContext({
+            context,
+            apps: runnerDetails.availableApps,
+            platform,
+        });
+        if (!supported &&
+            context.platform === platform &&
+            Array.isArray(context?.steps) &&
+            requiredBrowserAssets(context.browser?.name).length > 0) {
+            const firstAttempt = !installAttempts.has((context.browser?.name ?? "<none>").toLowerCase());
+            const outcome = await ensureContextBrowserInstalled({
+                browserName: context.browser?.name,
+                config,
+                installAttempts,
+                deps: {
+                    ensureBrowser: (asset, options) => ensureBrowserInstalled(asset, options),
+                    log,
+                },
+            });
+            if (firstAttempt && (outcome === "installed" || outcome === "failed")) {
+                clearAppCache(config);
+                runnerDetails.availableApps = await getAvailableApps({ config });
+                supported = isSupportedContext({
+                    context,
+                    apps: runnerDetails.availableApps,
+                    platform,
+                });
+            }
+        }
+        // Unsupported combinations are left unmarked; runContext skips each with the
+        // appropriate per-context reason (install-but-undetected vs unsupported).
+        if (!supported)
+            continue;
+        // Warm-up probe: start a driver once to prove the combination works.
+        // driverStart's own transient retry absorbs concurrent-launch flakiness;
+        // a headless fallback (on a throwaway caps object, so the real contexts
+        // keep their configured headedness) matches runContext so a headed-only
+        // failure on a headless-capable box doesn't poison the combination.
+        const port = await appiumPool.acquire();
+        let warmDriver;
+        try {
+            const options = {
+                width: context.browser?.window?.width || 1200,
+                height: context.browser?.window?.height || 800,
+                headless: context.browser?.headless !== false,
+            };
+            try {
+                warmDriver = await driverStart(getDriverCapabilities({
+                    runnerDetails,
+                    name: context.browser.name,
+                    options,
+                }), port, 4, { cacheDir: config?.cacheDir });
+            }
+            catch {
+                log(config, "warning", `Warm-up for ${combo} failed headed; retrying headless.`);
+                warmDriver = await driverStart(getDriverCapabilities({
+                    runnerDetails,
+                    name: context.browser.name,
+                    options: { ...options, headless: true },
+                }), port, 4, { cacheDir: config?.cacheDir });
+            }
+            warmUpResults.set(combo, "ok");
+            log(config, "debug", `Warm-up succeeded for ${combo}.`);
+        }
+        catch (error) {
+            warmUpResults.set(combo, "failed");
+            log(config, "warning", `Warm-up failed for ${combo}; contexts using it will be skipped: ${error?.message ?? String(error)}`);
+        }
+        finally {
+            if (warmDriver) {
+                try {
+                    await warmDriver.deleteSession();
+                }
+                catch {
+                    // best-effort teardown of the warm-up session
+                }
+            }
+            appiumPool.release(port);
+        }
+    }
+}
+/**
+ * Runs a single resolved context to completion and returns its finished
+ * contextReport (steps array + rolled-up result). Never touches the shared
+ * report or summary counters — the caller owns aggregation, which keeps this
+ * function safe to run concurrently with sibling contexts.
+ */
+async function runContext({ config, spec, test, context, runnerDetails, appiumPool, metaValues, installAttempts, warmUpResults, logPrefix = "", }) {
+    const platform = runnerDetails.environment.platform;
+    // `let`, not `const`: an on-demand browser install below re-detects available
+    // apps and reassigns this snapshot.
+    let availableApps = runnerDetails.availableApps;
+    // Context-scoped log: prefixed only when contexts run concurrently, so
+    // sequential output stays unchanged.
+    const clog = (level, message) => log(config, level, logPrefix && typeof message === "string"
+        ? `${logPrefix} ${message}`
+        : message);
+    // Ensure context contains a 'steps' property before anything walks it —
+    // isDriverRequired iterates context.steps and the resolved shape doesn't
+    // guarantee the field.
+    if (!context.steps) {
+        context.steps = [];
+    }
+    // If "platform" is not defined, set it to the current platform
+    if (!context.platform)
+        context.platform = runnerDetails.environment.platform;
+    // Attach OpenAPI definitions to context
+    if (config.integrations?.openApi) {
+        context.openApi = [
+            ...(context.openApi || []),
+            ...config.integrations.openApi,
+        ];
+    }
+    // If "browser" isn't defined but is required by the test, set it to the first available browser in the sequence of Firefox, Chrome, Safari
+    if (!context.browser && isDriverRequired({ test: context })) {
+        context.browser = getDefaultBrowser({ runnerDetails });
+    }
+    // Set context report
+    const contextReport = {
+        contextId: context.contextId,
+        platform: context.platform,
+        browser: context.browser,
+        steps: [],
+    };
+    // Set meta values (create-if-missing — ids aren't guaranteed unique)
+    metaValues.specs[spec.specId].tests[test.testId].contexts[context.contextId] ??= { steps: {} };
+    // If a driver is required but no browser could be resolved (e.g.
+    // getDefaultBrowser found nothing installed, or the context supplied a
+    // browser object with no name), skip with an explicit reason instead of
+    // letting it fail later as "Failed to start context 'undefined'".
+    if (isDriverRequired({ test: context }) && !context.browser?.name) {
+        const errorMessage = `Skipping context on '${context.platform}': no supported browser is available in the current environment.`;
+        clog("warning", errorMessage);
+        contextReport.result = "SKIPPED";
+        contextReport.resultDescription = errorMessage;
+        return contextReport;
+    }
+    // Check if current environment supports given contexts
+    let supportedContext = isSupportedContext({
+        context: context,
+        apps: availableApps,
+        platform: platform,
+    });
+    // If the context needs a browser that isn't available yet, try to resolve
+    // the missing dependency on demand before giving up — e.g. Firefox declared
+    // but geckodriver absent because the pre-flight was skipped or its install
+    // failed. Memoized per browser (installAttempts) so a failed/no-op install
+    // isn't retried for every later context. The install + re-detect mutate the
+    // shared runnerDetails.availableApps; under concurrency that's racy, but it
+    // only fires for a genuinely-missing browser (rare) and the app list only
+    // grows, so a sibling reading a slightly stale snapshot still re-detects.
+    let freshInstallRedetected = false;
+    if (!supportedContext &&
+        context.platform === platform &&
+        // Mirror isSupportedContext's own guard: isDriverRequired iterates
+        // context.steps, so a malformed context without a steps array would
+        // otherwise crash here instead of skipping cleanly.
+        Array.isArray(context?.steps) &&
+        isDriverRequired({ test: context }) &&
+        requiredBrowserAssets(context.browser?.name).length > 0) {
+        // Whether this browser was already attempted earlier this run; a cached
+        // outcome installed nothing new, so there's no point paying for a re-detect.
+        const firstAttempt = !installAttempts.has((context.browser?.name ?? "<none>").toLowerCase());
+        const outcome = await ensureContextBrowserInstalled({
+            browserName: context.browser?.name,
+            config,
+            installAttempts,
+            deps: {
+                ensureBrowser: (asset, options) => ensureBrowserInstalled(asset, options),
+                log,
+            },
+        });
+        // Re-detect after a real attempt regardless of outcome: a "failed" install
+        // can still have materialized assets before it threw, so a stale snapshot
+        // could wrongly skip a now-usable browser.
+        if (firstAttempt && (outcome === "installed" || outcome === "failed")) {
+            freshInstallRedetected = true;
+            clearAppCache(config);
+            availableApps = await getAvailableApps({ config });
+            runnerDetails.availableApps = availableApps;
+            supportedContext = isSupportedContext({
+                context: context,
+                apps: availableApps,
+                platform: platform,
+            });
+        }
+    }
+    // If context isn't supported, skip it
+    if (!supportedContext) {
+        // Distinguish "we installed the dependency but still can't see it" from a
+        // plain unsupported context, so the skip reason points at the real problem.
+        const errorMessage = freshInstallRedetected
+            ? `Skipping context '${context.browser?.name}' on '${context.platform}': the missing browser dependency was installed but still could not be detected.`
+            : `Skipping context. The current system doesn't support this context: {"platform": "${context.platform}", "apps": ${JSON.stringify(context.apps)}}`;
+        clog(freshInstallRedetected ? "warning" : "info", errorMessage);
+        contextReport.result = "SKIPPED";
+        contextReport.resultDescription = errorMessage;
+        return contextReport;
+    }
+    clog("debug", `CONTEXT:\n${JSON.stringify(context, null, 2)}`);
+    let driver;
+    let appiumPort;
+    const driverRequired = isDriverRequired({ test: context });
+    if (driverRequired && !appiumPool) {
+        throw new Error("Driver requested but no Appium server pool was created; " +
+            "driverJobCount and isDriverRequired(context) disagreed; this is a bug.");
+    }
+    // Warm-up memoization. The first context of each combination acts as the
+    // warm-up; if that combination already failed to start a driver earlier in
+    // this run, skip it outright instead of paying driverStart's retry/backoff
+    // again. Under concurrency this is a best-effort speedup, not correctness —
+    // same-combo contexts may start before one records a result.
+    const combo = combinationKey(context);
+    try {
+        if (driverRequired) {
+            if (warmUpDecision(warmUpResults.get(combo)) === "skip") {
+                const errorMessage = `Skipping context '${context.browser?.name}' on '${context.platform}': this context combination could not start a driver earlier in this run.`;
+                clog("warning", errorMessage);
+                contextReport.result = "SKIPPED";
+                contextReport.resultDescription = errorMessage;
+                return contextReport;
+            }
+            // Check out a server for this context's lifetime — released in the
+            // finally so the next queued context can reuse it.
+            appiumPort = await appiumPool.acquire();
+            // Define driver capabilities
+            // TODO: Support custom apps
+            let caps = getDriverCapabilities({
+                runnerDetails: runnerDetails,
+                name: context.browser.name,
+                options: {
+                    width: context.browser?.window?.width || 1200,
+                    height: context.browser?.window?.height || 800,
+                    headless: context.browser?.headless !== false,
+                },
+            });
+            clog("debug", "CAPABILITIES:");
+            clog("debug", caps);
+            // Instantiate driver
+            try {
+                driver = await driverStart(caps, appiumPort, 4, { cacheDir: config?.cacheDir });
+            }
+            catch (error) {
+                try {
+                    // If driver fails to start, try again as headless
+                    clog("warning", `Failed to start context '${context.browser?.name}' on '${platform}'. Retrying as headless.`);
+                    context.browser.headless = true;
+                    caps = getDriverCapabilities({
+                        runnerDetails: runnerDetails,
+                        name: context.browser.name,
+                        options: {
+                            width: context.browser?.window?.width || 1200,
+                            height: context.browser?.window?.height || 800,
+                            headless: context.browser?.headless !== false,
+                        },
+                    });
+                    driver = await driverStart(caps, appiumPort, 4, { cacheDir: config?.cacheDir });
+                }
+                catch (error) {
+                    let errorMessage = `Failed to start context '${context.browser?.name}' on '${platform}'.`;
+                    // `safari` is normalized to `webkit` during context resolution, so
+                    // match both or this Safari-specific hint never fires on real runs.
+                    if (context.browser?.name === "safari" ||
+                        context.browser?.name === "webkit")
+                        errorMessage =
+                            errorMessage +
+                                " Make sure you've run `safaridriver --enable` in a terminal and enabled 'Allow Remote Automation' in Safari's Develop menu.";
+                    clog("error", errorMessage);
+                    // Record the combination as failed so every later context that shares
+                    // it is skipped instantly (see the warm-up check above).
+                    if (!warmUpResults.has(combo))
+                        warmUpResults.set(combo, "failed");
+                    contextReport.result = "SKIPPED";
+                    contextReport.resultDescription = errorMessage;
+                    return contextReport;
+                }
+            }
+            // Driver started (first attempt or headless retry) — mark this
+            // combination as known-good for the rest of the run.
+            if (!warmUpResults.has(combo))
+                warmUpResults.set(combo, "ok");
+            if (context.browser?.viewport?.width ||
+                context.browser?.viewport?.height) {
+                // Set driver viewport size
+                await setViewportSize(context, driver);
+            }
+            else if (context.browser?.window?.width ||
+                context.browser?.window?.height) {
+                // Get driver window size
+                const windowSize = await driver.getWindowSize();
+                // Resize window if necessary
+                await driver.setWindowSize(context.browser?.window?.width || windowSize.width, context.browser?.window?.height || windowSize.height);
+            }
+        }
+        // Iterates steps
+        let stepExecutionFailed = false;
+        for (let step of context.steps) {
+            // Set step id if not defined
+            if (!step.stepId)
+                step.stepId = randomUUID();
+            clog("debug", `STEP:\n${JSON.stringify(step, null, 2)}`);
+            if (step.unsafe && runnerDetails.allowUnsafeSteps === false) {
+                clog("warning", `Skipping unsafe step: ${step.description} in test ${test.testId} context ${context.contextId}`);
+                // Mark as skipped
+                const stepReport = {
+                    ...step,
+                    result: "SKIPPED",
+                    resultDescription: "Skipped because unsafe steps aren't allowed.",
+                };
+                contextReport.steps.push(stepReport);
+                continue;
+            }
+            if (stepExecutionFailed) {
+                // Mark as skipped
+                const stepReport = {
+                    ...step,
+                    result: "SKIPPED",
+                    resultDescription: "Skipped due to previous failure in context.",
+                };
+                contextReport.steps.push(stepReport);
+                continue;
+            }
+            // Set meta values
+            metaValues.specs[spec.specId].tests[test.testId].contexts[context.contextId].steps[step.stepId] = {};
+            // Run step
+            const stepResult = await runStep({
+                config: config,
+                context: context,
+                step: step,
+                driver: driver,
+                metaValues: metaValues,
+                options: {
+                    openApiDefinitions: context.openApi || [],
+                },
+            });
+            clog("debug", `RESULT: ${stepResult.status}\n${JSON.stringify(stepResult, null, 2)}`);
+            stepResult.result = stepResult.status;
+            stepResult.resultDescription = stepResult.description;
+            delete stepResult.status;
+            delete stepResult.description;
+            // Add step result to report
+            const stepReport = {
+                ...step,
+                ...stepResult,
+            };
+            contextReport.steps.push(stepReport);
+            // If this step failed, set flag to skip remaining steps
+            if (stepReport.result === "FAIL") {
+                stepExecutionFailed = true;
+            }
+        }
+        // If recording, stop recording
+        if (driver?.state?.recording) {
+            const stopRecordStep = {
+                stopRecord: true,
+                description: "Stopping recording",
+                stepId: randomUUID(),
+            };
+            const stepResult = await runStep({
+                config: config,
+                context: context,
+                step: stopRecordStep,
+                driver: driver,
+                options: {
+                    openApiDefinitions: context.openApi || [],
+                },
+            });
+            stepResult.result = stepResult.status;
+            stepResult.resultDescription = stepResult.description;
+            delete stepResult.status;
+            delete stepResult.description;
+            // Add step result to report
+            const stepReport = {
+                ...stopRecordStep,
+                ...stepResult,
+            };
+            contextReport.steps.push(stepReport);
+        }
+    }
+    finally {
+        // Close driver. In a finally so an unexpected throw can't leak a session
+        // while sibling contexts keep running.
+        if (driver) {
+            try {
+                await driver.deleteSession();
+            }
+            catch (error) {
+                clog("error", `Failed to delete driver session: ${error.message}`);
+            }
+        }
+        // Return the Appium server to the pool for the next queued context. Always
+        // runs (even on the driver-start-failure early return) so a port can't
+        // leak out of the pool and starve later contexts.
+        if (appiumPort !== undefined && appiumPool) {
+            appiumPool.release(appiumPort);
+        }
+    }
+    contextReport.result = rollUpResults(contextReport.steps);
+    return contextReport;
+}
 // Run a specific step
 async function runStep({ config = {}, context = {}, step, driver, metaValues = {}, options = {}, }) {
     let actionResult;
@@ -915,7 +1221,7 @@ async function runStep({ config = {}, context = {}, step, driver, metaValues = {
             step: step,
             driver: driver,
         });
-        config.recording = actionResult.recording;
+        driver.state.recording = actionResult.recording ?? null;
     }
     else if (typeof step.runCode !== "undefined") {
         actionResult = await runCode({ config: config, step: step });
@@ -947,7 +1253,7 @@ async function runStep({ config = {}, context = {}, step, driver, metaValues = {
         };
     }
     // If recording, wait until browser is loaded, then instantiate cursor
-    if (config?.recording) {
+    if (driver?.state?.recording) {
         const currentUrl = await driver.getUrl();
         if (currentUrl !== driver.state.url) {
             driver.state.url = currentUrl;
@@ -971,6 +1277,40 @@ async function runStep({ config = {}, context = {}, step, driver, metaValues = {
     }
     return actionResult;
 }
+// Start one Appium server on a free port and resolve once it answers /status.
+// Each concurrent runner gets its own server (own port) so parallel contexts
+// never create sessions on the same Appium instance.
+async function startAppiumServer(appiumEntry, config) {
+    const port = await findFreePort();
+    log(config, "debug", `Starting Appium on port ${port}`);
+    const proc = spawn(process.execPath, [appiumEntry, "-a", "127.0.0.1", "-p", String(port)], {
+        windowsHide: true,
+        cwd: path.join(__dirname, "../.."),
+    });
+    proc.on("error", (err) => {
+        log(config, "warning", `Appium process error: ${err?.stack ?? err?.message ?? String(err)}`);
+    });
+    proc.stdout.on("data", () => { });
+    proc.stderr.on("data", () => { });
+    try {
+        await appiumIsReady(port);
+    }
+    catch (error) {
+        // appiumIsReady threw or timed out — the spawned child is still alive and
+        // would leak (orphan process, port still bound). Tear it down before
+        // propagating so subsequent runs don't trip on the stale state.
+        try {
+            if (proc && proc.pid)
+                kill(proc.pid);
+        }
+        catch {
+            // best-effort cleanup; the parent error is what matters
+        }
+        throw error;
+    }
+    log(config, "debug", `Appium is ready on port ${port}.`);
+    return { port, process: proc };
+}
 // Delay execution until Appium server is available.
 async function appiumIsReady(port, timeoutMs = 120000) {
     let isReady = false;
@@ -991,10 +1331,20 @@ async function appiumIsReady(port, timeoutMs = 120000) {
 }
 // Start the Appium driver specified in `capabilities`.
 async function driverStart(capabilities, port, maxAttempts = 4, ctx = {}) {
-    // POST /session can race a just-spawned-or-still-dying Appium on Windows:
-    // /status may already return 200 from the outgoing process while /session
-    // is no longer accepting. Retry with linear backoff ONLY on ECONNREFUSED --
-    // any other error is a real session-creation failure and propagates.
+    // Two families of transient, retryable session-creation failures, both worse
+    // under concurrency (the TRANSIENT regex below enumerates the specific
+    // patterns):
+    //   1. POST /session races a just-spawned-or-still-dying Appium (Windows):
+    //      /status returns 200 from the outgoing process while /session no longer
+    //      accepts, or Appium's proxy to chromedriver drops the socket ->
+    //      ECONNREFUSED / ECONNRESET / "socket hang up" / "could not proxy command".
+    //   2. Several Chromes launching at once briefly starve resources and
+    //      ChromeDriver "crashed during startup" / "cannot connect to" /
+    //      "DevToolsActivePort" / "session not created". A staggered retry lets
+    //      the contention clear; it recovers on the next attempt in practice.
+    // Retry these with linear backoff; any other error is a real session-
+    // creation failure and propagates immediately.
+    const TRANSIENT = /ECONNREFUSED|ECONNRESET|socket hang up|could not proxy command|crashed during startup|cannot connect to|DevToolsActivePort|session not created/i;
     const wdio = await loadHeavyDep("webdriverio", { ctx });
     let lastError;
     for (let attempt = 1; attempt <= maxAttempts; attempt++) {
@@ -1009,12 +1359,14 @@ async function driverStart(capabilities, port, maxAttempts = 4, ctx = {}) {
                 connectionRetryTimeout: 120000, // 2 minutes
                 waitforTimeout: 120000, // 2 minutes
             });
-            driver.state = { url: "", x: null, y: null };
+            // Per-context mutable state. `recording` lives here (not on config)
+            // so concurrent contexts can't clobber each other's recordings.
+            driver.state = { url: "", x: null, y: null, recording: null };
             return driver;
         }
         catch (err) {
             lastError = err;
-            if (!/ECONNREFUSED/.test(String(err && err.message)))
+            if (!TRANSIENT.test(String(err && err.message)))
                 throw err;
             if (attempt < maxAttempts) {
                 await new Promise((r) => setTimeout(r, 500 * attempt));
@@ -1106,6 +1458,46 @@ async function ensureChromeAvailable(config, deps) {
     }
     return availableApps;
 }
+/**
+ * On-demand, per-context browser/driver install used by the runner when a
+ * context's browser isn't yet available (e.g. Firefox declared but geckodriver
+ * missing). Attempts to install every asset the browser needs, memoizing the
+ * outcome in `installAttempts` so a failed (or no-op) attempt isn't repeated
+ * for every later context that shares the browser. Like ensureChromeAvailable,
+ * this self-provisions regardless of DOC_DETECTIVE_AUTOINSTALL (that env var
+ * only governs the eager postinstall). Deps are injected for testing.
+ *
+ * @returns "installed" when all assets installed, "failed" when an install
+ *   threw, or "notInstallable" for browsers with no installable asset (safari).
+ */
+async function ensureContextBrowserInstalled({ browserName, config, installAttempts, deps, }) {
+    const key = (browserName ?? "<none>").toLowerCase();
+    const cached = installAttempts.get(key);
+    if (cached)
+        return cached;
+    const assets = requiredBrowserAssets(browserName);
+    if (assets.length === 0) {
+        installAttempts.set(key, "notInstallable");
+        return "notInstallable";
+    }
+    const ctx = { cacheDir: config?.cacheDir };
+    // Bridge runtime modules' (msg, level) logger to core/utils.ts#log, mapping
+    // "warn" → "warning" the same way provisionChromeRuntime does.
+    const logger = (msg, level = "info") => deps.log?.(config, level === "warn" ? "warning" : level, msg);
+    try {
+        deps.log?.(config, "info", `Browser '${browserName}' is not available; attempting on-demand install of: ${assets.join(", ")}.`);
+        for (const asset of assets) {
+            await deps.ensureBrowser(asset, { ctx, deps: { logger } });
+        }
+        installAttempts.set(key, "installed");
+        return "installed";
+    }
+    catch (err) {
+        deps.log?.(config, "warning", `On-demand install for '${browserName}' failed: ${err?.message ?? err}`);
+        installAttempts.set(key, "failed");
+        return "failed";
+    }
+}
 async function getRunner(options = {}) {
     const environment = getEnvironment();
     const config = { ...options.config, environment };