@codeyam-editor/codeyam-editor 0.1.1 → 0.1.3

package/npm/preflight-guidance.js

@@ -0,0 +1,77 @@
+"use strict";
+
+/**
+ * Launcher-side mirror of `start_preflight` in
+ * `crates/codeyam-editor/src/commands/start.rs`. The Rust side is the source
+ * of truth for the guidance strings; if it changes, update this file in the
+ * same commit and keep the `crates/codeyam-editor/tests/start_preflight_exit_code.rs`
+ * test in sync.
+ *
+ * Mirrors the heuristic in `migration_status::is_legacy_config`: a
+ * `.codeyam/config.json` whose JSON has a top-level `webapps` key marks a
+ * legacy project. Anything else with no `editor.json` is greenfield.
+ *
+ * Rust-side behavior note (improve39): `start_preflight` no longer bails on an
+ * EMPTY greenfield folder — it auto-inits in place and boots the editor's wired
+ * scaffold flow. It still bails (exit 2) for an existing-source `none` folder
+ * (the `greenfieldGuidance` text below) and for legacy. This module is consumed
+ * by the launcher's bail/guidance path only; `classifyProject` does not itself
+ * decide empty-vs-non-empty (the launcher's `planPreflightAction` /
+ * `postLaunchInitBanner` in editor-dev.js own that split and intentionally
+ * auto-init all `none` projects as a local-dev convenience).
+ */
+
+const fs = require("node:fs");
+const path = require("node:path");
+
+function greenfieldGuidance(projectDir) {
+  return (
+    `No codeyam-editor project detected in ${projectDir}.\n\n` +
+    "Run `codeyam-editor init` to install the agent bundle, then\n" +
+    "open Claude Code and invoke `/codeyam-onboard` to set up this project."
+  );
+}
+
+function legacyGuidance(projectDir) {
+  return (
+    `Detected legacy codeyam project in ${projectDir}.\n\n` +
+    "Run `codeyam-editor init` to install the agent bundle (it will skip\n" +
+    "editor.json creation so the legacy classification is preserved), then\n" +
+    "open Claude Code and invoke `/codeyam-onboard` to migrate."
+  );
+}
+
+function isLegacyConfig(configPath) {
+  let raw;
+  try {
+    raw = fs.readFileSync(configPath, "utf8");
+  } catch {
+    return false;
+  }
+  let parsed;
+  try {
+    parsed = JSON.parse(raw);
+  } catch {
+    return false;
+  }
+  return parsed && typeof parsed === "object" && "webapps" in parsed;
+}
+
+/**
+ * Classify a project directory as `modern`, `legacy`, or `none`, mirroring
+ * the Rust `classify` / `start_preflight` pair. `guidance` is the exact
+ * stderr text to print before exiting with code 2.
+ */
+function classifyProject(projectDir) {
+  const editorJson = path.join(projectDir, ".codeyam", "editor.json");
+  if (fs.existsSync(editorJson)) {
+    return { kind: "modern" };
+  }
+  const legacyConfig = path.join(projectDir, ".codeyam", "config.json");
+  if (isLegacyConfig(legacyConfig)) {
+    return { kind: "legacy", guidance: legacyGuidance(projectDir) };
+  }
+  return { kind: "none", guidance: greenfieldGuidance(projectDir) };
+}
+
+module.exports = { classifyProject };

package/npm/scenario-check.js

@@ -11,8 +11,51 @@
 
 const fs = require("fs");
 const path = require("path");
+const { execSync } = require("child_process");
 const { chromium } = require("playwright");
 
+// Substring Playwright has emitted across every 1.x release when the
+// browser cache is empty. Match the substring (not the full string)
+// because Playwright includes the offending path inline.
+const PLAYWRIGHT_MISSING_BROWSER_PATTERN = "Executable doesn't exist";
+const PLAYWRIGHT_INSTALL_COMMAND = "npx playwright install chromium";
+
+// One-shot self-heal around `chromium.launch()`. If the first launch
+// throws the "missing browser" error, run `npx playwright install
+// chromium` synchronously (with `stdio: "inherit"` so the user sees
+// progress) and retry the launch exactly once. If the install or the
+// retry fails, rethrow the ORIGINAL Playwright error so the existing
+// `Scenario check failed: <stderr>` path keeps showing the actionable
+// message — looping would hide a real ops failure under a slow timeout.
+async function launchChromiumWithSelfHeal({
+  launch = () => chromium.launch(),
+  install = () => execSync(PLAYWRIGHT_INSTALL_COMMAND, { stdio: "inherit" }),
+  stderr = process.stderr,
+} = {}) {
+  try {
+    return await launch();
+  } catch (error) {
+    const isMissingBrowser =
+      error &&
+      typeof error.message === "string" &&
+      error.message.includes(PLAYWRIGHT_MISSING_BROWSER_PATTERN);
+    if (!isMissingBrowser) throw error;
+    stderr.write(
+      "Playwright's Chromium browser is missing — installing it now (one-time ~150 MB download). Subsequent runs will be instant.\n",
+    );
+    try {
+      install();
+    } catch (_installError) {
+      throw error;
+    }
+    try {
+      return await launch();
+    } catch (_retryError) {
+      throw error;
+    }
+  }
+}
+
 const {
   findErrorPattern,
   buildErrorContextSnippet,
@@ -31,9 +74,14 @@ const {
 } = require("./scenario-mocks");
 
 const {
+  assertAppPortReachable,
   loadScenarioInIframe,
+  loadScenarioTopLevel,
   waitForStablePage,
+  createNetworkTracker,
+  waitForNetworkQuiet,
   collectContentState,
+  performInteraction,
 } = require("./scenario-playwright");
 
 const {
@@ -43,12 +91,196 @@ const {
   handleRequestFailed,
 } = require("./scenario-handlers");
 
-const BLANK_RETRY_DELAY_MS = 500;
+const {
+  probeInteractivity,
+} = require("./scenario-interactivity");
+
+// Read project-specific loading markers from `.codeyam/stack.json`
+// (`capture.loadingMarkers`). The capture script runs with cwd = project dir
+// (scenario_check.rs sets `.current_dir(project_dir)`), so this relative path
+// resolves to the project's own config. An app's loading copy ("Loading…",
+// "Please wait") is app-specific, so it lives in stack.json rather than being
+// hardcoded into the shared harness; the codeyam-harness defaults in
+// scenario-metrics.js always apply on top. Never throws — a missing or
+// malformed stack.json just yields no extra markers.
+function readStackLoadingMarkers() {
+  try {
+    const raw = fs.readFileSync(path.join(".codeyam", "stack.json"), "utf8");
+    const stack = JSON.parse(raw);
+    const markers = stack && stack.capture && stack.capture.loadingMarkers;
+    return Array.isArray(markers)
+      ? markers.filter((m) => typeof m === "string" && m.length > 0)
+      : [];
+  } catch (_) {
+    return [];
+  }
+}
+
+// Cold-start retry pause. waitForStablePage settles as soon as the page is
+// HTML-stable, which for a lazy/Suspense app is the empty `<div id="root">`
+// shell — stable for the ~3s the dynamic chunk takes to load (longer when the
+// scenario's mocks slow the boot). 500ms re-checked before the chunk resolved
+// and reported a false blank; this pause must comfortably exceed that window
+// while staying under the test runner's default per-case timeout.
+const BLANK_RETRY_DELAY_MS = 3000;
+
+// True when `url` targets a different origin than `appOrigin`. Used to decide
+// whether the codeyam capture markers must be stripped before the request
+// leaves (cross-origin) or may ride along (same-origin, the app's own dev
+// server). A malformed URL counts as same-origin (false) so we never strip
+// markers from a request we can't classify — the conservative default keeps
+// same-origin behavior unchanged. `url` may be a string or a URL-like object.
+function isCrossOriginRequest(url, appOrigin) {
+  try {
+    const href = typeof url === "string" ? url : url.href;
+    return new URL(href).origin !== appOrigin;
+  } catch (_) {
+    return false;
+  }
+}
+
+// Return a copy of `headers` with every name in `markerNames` removed.
+// Names are matched case-insensitively against the (lowercased) header keys
+// Playwright reports. Pure — never mutates its input — so unrelated headers
+// (Accept, User-Agent, a scenario's own requestHeaders) survive untouched.
+function stripMarkerHeaders(headers, markerNames) {
+  const out = { ...headers };
+  for (const name of markerNames) {
+    delete out[name.toLowerCase()];
+  }
+  return out;
+}
+
+// Apply the scenario's merged `browserState` to a Playwright context and
+// stamp the codeyam capture markers on every request the context makes.
+//
+// Cookies need a concrete URL to bind to (Playwright requires either
+// `url` or `domain`+`path`); we derive domain/path from the requested
+// capture URL when the scenario didn't pin them. Request headers go
+// through `setExtraHTTPHeaders` so every navigation and resource
+// request in the context carries them.
+//
+// Every capture-originated request carries `X-Codeyam-Capture: 1` (and
+// `X-Codeyam-Scenario: <slug>` when a scenario is active) so a dev-server
+// log can tell the headless capture apart from the operator's own browser
+// hitting the same route — they are otherwise indistinguishable. These are
+// defaults: a scenario's own `requestHeaders` are merged on top and win,
+// so a user can override or clear them.
+async function applyBrowserState(context, config) {
+  const state = (config && config.browserState) || {};
+  const cookies = state.cookies || {};
+  const cookieEntries = Object.entries(cookies);
+  if (cookieEntries.length > 0) {
+    let host = "127.0.0.1";
+    try {
+      host = new URL(config.url).hostname || host;
+    } catch (_) {
+      /* fall back to localhost if config.url is malformed */
+    }
+    const playwrightCookies = cookieEntries.map(([name, raw]) => {
+      const descriptor =
+        raw && typeof raw === "object" && !Array.isArray(raw) ? raw : {};
+      const value = typeof raw === "string" ? raw : descriptor.value || "";
+      return {
+        name,
+        value,
+        domain: descriptor.domain || host,
+        path: descriptor.path || "/",
+        sameSite: descriptor.sameSite || "Lax",
+        httpOnly: descriptor.httpOnly === true,
+        secure: descriptor.secure === true,
+      };
+    });
+    await context.addCookies(playwrightCookies);
+  }
+  const codeyamHeaders = {
+    "X-Codeyam-Capture": "1",
+    ...(config && config.scenarioId
+      ? { "X-Codeyam-Scenario": config.scenarioId }
+      : {}),
+  };
+  // Scenario `requestHeaders` are merged last so a user value overrides
+  // the codeyam default (e.g. setting `X-Codeyam-Capture: "0"` as an
+  // escape hatch). `codeyamHeaders` always has at least the capture marker,
+  // so this fires on every capture — including ones with no browserState.
+  const headers = { ...codeyamHeaders, ...(state.requestHeaders || {}) };
+  if (Object.keys(headers).length > 0) {
+    await context.setExtraHTTPHeaders(headers);
+  }
+
+  // Strip the codeyam capture markers (and any scenario request headers) from
+  // CROSS-ORIGIN requests. setExtraHTTPHeaders applies context-wide, so the
+  // custom `X-Codeyam-*` headers ride along on third-party subresource
+  // requests (Google Fonts, CDNs, external APIs) too — and a non-safelisted
+  // request header forces a CORS preflight those hosts reject
+  // (`Request header field x-codeyam-capture is not allowed`), which fails the
+  // whole capture. The markers are only meaningful to the app's OWN dev server
+  // (same-origin), so re-send cross-origin requests without them. Same-origin
+  // requests are never matched here, so dev-module/HMR loading is untouched.
+  let appOrigin = null;
+  try {
+    appOrigin = new URL(config.url).origin;
+  } catch (_) {
+    /* malformed capture URL — skip the cross-origin guard entirely */
+  }
+  if (appOrigin && typeof context.route === "function") {
+    // Only the codeyam markers are stripped — a scenario's own requestHeaders
+    // are the author's deliberate choice and left intact.
+    const markerNames = Object.keys(codeyamHeaders);
+    await context.route(
+      (url) => isCrossOriginRequest(url, appOrigin),
+      async (route) => {
+        const reqHeaders = stripMarkerHeaders(route.request().headers(), markerNames);
+        await route.continue({ headers: reqHeaders });
+      },
+    );
+  }
+}
 
-async function runScenarioCheck(config) {
+// Emit a `redirect-mismatch` issue when the final iframe URL's path
+// differs from the requested path. This converts the silent
+// screenshot-of-/login failure (auth lost in the capture context) into
+// a typed, actionable diagnostic the agent can route to.
+function pushRedirectMismatchIssue(issues, requestedUrl, frame, response, config) {
+  let requestedPath;
+  try {
+    requestedPath = new URL(requestedUrl).pathname;
+  } catch (_) {
+    return;
+  }
+  const finalUrl = (frame && frame.url && frame.url()) || requestedUrl;
+  let finalPath;
+  try {
+    finalPath = new URL(finalUrl).pathname;
+  } catch (_) {
+    return;
+  }
+  if (requestedPath === finalPath) return;
+  const cookies =
+    (config && config.browserState && config.browserState.cookies) || {};
+  const hasCookies = Object.keys(cookies).length > 0;
+  const hint = hasCookies
+    ? " — scenario carries browserState.cookies; auth likely lost in the capture context (run `codeyam-editor editor scenario-explain <slug>` to verify)"
+    : "";
+  pushIssue(
+    issues,
+    createIssue(
+      "redirect-mismatch",
+      `Capture URL redirected from ${requestedPath} to ${finalPath}${hint}`,
+      {
+        url: finalUrl,
+        status: response && response.status ? response.status() : null,
+      },
+    ),
+  );
+}
+
+// `preflight` is injectable (defaulting to the real app-port reachability
+// check) so unit tests that mock the browser can stay network-free.
+async function runScenarioCheck(config, { preflight = assertAppPortReachable } = {}) {
   const { url, outputPath, width, height, httpMocks = {} } = config;
   const issues = [];
-  const browser = await chromium.launch();
+  const browser = await launchChromiumWithSelfHeal();
   const contextOptions = {
     viewport: { width: width || 1440, height: height || 900 },
   };
@@ -62,10 +294,23 @@ async function runScenarioCheck(config) {
   if (config.forcedColors) contextOptions.forcedColors = config.forcedColors;
   const context = await browser.newContext(contextOptions);
 
+  // Apply the scenario's merged `browserState` (cookies + request
+  // headers) to the capture context BEFORE the first navigation.
+  // Belt-and-suspenders with the proxy's `Set-Cookie` injection: the
+  // proxy handles upstream-bound forwards, this branch handles the
+  // capture context's own request headers so an auth-gated route does
+  // not redirect to `/login` when the proxy is bypassed.
+  await applyBrowserState(context, config);
+
   // Context-level init script runs in ALL frames (including cross-origin iframes)
   await context.addInitScript(getInitScript());
 
   const page = await context.newPage();
+  // Attach the network tracker BEFORE navigation so every request (the document
+  // and every client-side data fetch) is counted. Used after DOM stability to
+  // wait out an in-flight fetch that would otherwise be screenshotted as a
+  // loading skeleton.
+  const networkTracker = createNetworkTracker(page);
   await attachHttpMocks(page, httpMocks);
 
   page.on("pageerror", (error) => {
@@ -80,15 +325,28 @@ async function runScenarioCheck(config) {
   });
 
   page.on("requestfailed", (request) => {
-    pushIssue(issues, handleRequestFailed(request));
+    const issue = handleRequestFailed(request);
+    if (issue) {
+      pushIssue(issues, issue);
+    }
   });
 
   let loaded = false;
 
   try {
-    const { frame, response } = await loadScenarioInIframe(page, url, {
-      background: config.iframeBackground,
-    });
+    // Application/route captures navigate at the top level so the
+    // first-party session cookie is sent (auth-gated routes render the
+    // authenticated page instead of /login); component captures keep the
+    // iframe harness for its background/sizing control. The backend signals
+    // the choice via `config.navigation` ("topLevel"); absent (the default)
+    // means the iframe harness, so existing callers are unchanged.
+    const { frame, response } =
+      config.navigation === "topLevel"
+        ? await loadScenarioTopLevel(page, url, { preflight })
+        : await loadScenarioInIframe(page, url, {
+            background: config.iframeBackground,
+            preflight,
+          });
     loaded = true;
 
     if (response && response.status() >= 400) {
@@ -101,7 +359,22 @@ async function runScenarioCheck(config) {
       );
     }
 
-    await waitForStablePage(page, frame);
+    pushRedirectMismatchIssue(issues, url, frame, response, config);
+
+    // Project loading markers come from config when a caller injects them
+    // (unit tests), otherwise from stack.json — so a stable-but-loading app
+    // screen is not mistaken for settled content and captured mid-hydration.
+    const loadingMarkers = Array.isArray(config.loadingMarkers)
+      ? config.loadingMarkers
+      : readStackLoadingMarkers();
+    await waitForStablePage(page, frame, 10000, loadingMarkers);
+
+    // DOM-stable does not mean done: a client-side data fetch can still be in
+    // flight (the loading skeleton cleared but its replacement content hasn't
+    // landed). Wait for the network to go quiet — bounded, so a streaming /
+    // long-poll endpoint that never idles caps out and captures anyway rather
+    // than hanging.
+    await waitForNetworkQuiet(networkTracker);
 
     const rejectionMessages = await frame.evaluate(
       () => window.__codeyamUnhandledRejections || [],
@@ -115,11 +388,12 @@ async function runScenarioCheck(config) {
       );
     }
 
-    // Cold-start retry: a single re-collect after a short pause covers
-    // the React.lazy / Suspense-fallback race that flagged 4-of-59
-    // scenarios as blank against a cold Vite dev server on 2026-04-30.
-    // One retry is enough — the Suspense window is sub-second in
-    // practice, longer waits just slow down genuine blank-render bugs.
+    // Cold-start retry: a single re-collect after BLANK_RETRY_DELAY_MS covers
+    // the React.lazy / Suspense-fallback race where waitForStablePage settles
+    // on the still-empty `<div id="root">` shell before the dynamic chunk
+    // resolves. One retry is enough — the pause is sized to outlast the
+    // chunk-load window; only a genuinely blank page falls through to the
+    // blank issue below.
     let contentState = await collectContentState(frame);
     let hasContent = hasRenderableContent(contentState);
     if (!hasContent) {
@@ -140,7 +414,7 @@ async function runScenarioCheck(config) {
     }
 
     // Check for known error states in the rendered content
-    const bodyText = await frame.evaluate(() => document.body.innerText || "");
+    const bodyText = await frame.evaluate(() => document.body?.innerText || "");
     const matchedPattern = findErrorPattern(bodyText);
     if (matchedPattern) {
       const contextSnippet = buildErrorContextSnippet(bodyText, matchedPattern);
@@ -158,6 +432,29 @@ async function runScenarioCheck(config) {
       );
     }
 
+    // Hydration / interactivity gate: a page can render content and log no
+    // errors yet never have hydrated, leaving every control dead. Read-only
+    // (it inspects framework-attachment markers, never clicks), so it is safe
+    // to run before the screenshot. Stack-gated and fail-safe — see
+    // scenario-interactivity.js — so backend / static / unknown-framework
+    // captures are an automatic pass.
+    const hydrationIssue = await probeInteractivity(frame, {
+      url: page.url() || url,
+    });
+    if (hydrationIssue) {
+      pushIssue(issues, hydrationIssue);
+    }
+
+    // Drive the requested interaction (if any) against the settled frame,
+    // then re-settle, so `preview-interact` captures the RESULT of a click /
+    // fill / press (expanded accordion, open modal) without editing app
+    // source. A no-match target throws here and is caught below as a failed
+    // capture with the candidate-labels hint — never a silent blank shot.
+    if (config.interaction) {
+      await performInteraction(frame, config.interaction);
+      await waitForStablePage(page, frame, 5000, loadingMarkers);
+    }
+
     if (outputPath && loaded) {
       fs.mkdirSync(path.dirname(outputPath), { recursive: true });
       await page.screenshot({ path: outputPath, fullPage: false });
@@ -188,6 +485,11 @@ async function runScenarioCheck(config) {
   }
 }
 
+/**
+ * npm wrapper entry point for the scenario-check binary: parses the
+ * JSON config from argv, drives Playwright to capture the configured
+ * URL, and writes the resulting screenshot.
+ */
 async function main() {
   const config = JSON.parse(process.argv[2] || "{}");
 
@@ -204,7 +506,14 @@ async function main() {
 
 module.exports = {
   runScenarioCheck,
+  readStackLoadingMarkers,
+  applyBrowserState,
+  isCrossOriginRequest,
+  stripMarkerHeaders,
   main,
+  launchChromiumWithSelfHeal,
+  PLAYWRIGHT_INSTALL_COMMAND,
+  PLAYWRIGHT_MISSING_BROWSER_PATTERN,
 };
 
 if (require.main === module) {

package/npm/scenario-handlers.js

@@ -56,11 +56,22 @@ function handlePageError(error) {
 }
 
 function handleRequestFailed(request) {
-  return createIssue(
-    "requestfailed",
-    request.failure()?.errorText || "Request failed",
-    { url: request.url() }
-  );
+  const errorText = request.failure()?.errorText || "Request failed";
+
+  // Filter benign request cancellations. `net::ERR_ABORTED` is what Playwright
+  // emits when a request is in-flight at the moment the page is closed (or the
+  // iframe is destroyed). For scenarios whose pages fetch large payloads (the
+  // editor's own EditorShell mounts a 2.2MB `/api/tests` fetch), the
+  // browser.close() at the end of capture races the fetch and produces this
+  // event AFTER the screenshot has already been taken — there is no real
+  // failure to surface. Genuine network failures arrive under different
+  // codes (net::ERR_CONNECTION_REFUSED, net::ERR_NAME_NOT_RESOLVED, etc.)
+  // and continue to be reported.
+  if (errorText.includes("net::ERR_ABORTED")) {
+    return null;
+  }
+
+  return createIssue("requestfailed", errorText, { url: request.url() });
 }
 
 module.exports = {