flet-box 1.1.2 → 1.1.3

package/bin/commands/verify.js

@@ -0,0 +1,879 @@
+/**
+ * @file bin/commands/verify.js
+ * @description One command, one report, one exit code — every gate at once.
+ *
+ * An agent that writes code needs a single question answered: "may I say this
+ * is done?". Asking it once per gate (`npm run check`, then `npm run ui`, then
+ * `npm run api`) means the agent stops at the first green one, and the report
+ * it hands back names a gate instead of a line.
+ *
+ * So the gates run here, in parallel, and the runner only counts and prints:
+ *
+ *   syntax   — the parse gate: every module is handed to the Node parser, so a
+ *              file that cannot load is named by line and column before any
+ *              other gate reads it as text.
+ *   api      — the anti-hallucination gate: invented widgets and props,
+ *              positional arguments, foreign imports, broken imports.
+ *   types    — the editor's own question, asked outside the editor: the
+ *              project's `jsconfig.json`, run through the TypeScript checker,
+ *              so `estado.update()` on a widget and a prop the declaration
+ *              forgot are reported as `file:line:column`, not as a squiggle
+ *              somebody learns to ignore. Skipped when TypeScript is absent
+ *              (the editor brings its own, an installed app has neither).
+ *   ui       — the design gate: tokens, the 8-based scale, tap targets,
+ *              labels, contrast.
+ *   structure — how the app's own modules are wired: a cycle between two of
+ *              them, a module outside the folders the project is built from,
+ *              markup written by hand.
+ *   surface — the prop contract: a prop the widget reads but nothing declares
+ *              is an error (the surface lies); a prop declared but nobody
+ *              reads is a warning (the docs promise more than the code does).
+ *   tests    — the project's own suites (`test`, `test:browser`, `test:visual`),
+ *              run one after another; a failing suite's own last words are the
+ *              finding, because that is where its rule is named.
+ *
+ * A new gate is one entry in `GATES` and `GATE_ORDER` and nothing else.
+ *
+ * Node core only: this file ships inside every app `flet-box create` builds,
+ * so it may not reach for a devDependency.
+ */
+
+import fs from "node:fs";
+import path from "node:path";
+import { execFile } from "node:child_process";
+import { createRequire } from "node:module";
+import { checkApi, loadApi, SEVERITY } from "./checkApi.js";
+import { checkUi, ERROR_CODES, WARN_CODES, INFO_CODES } from "./checkUi.js";
+import { scanStructure } from "./checkStructure.js";
+import { scanSyntax } from "./checkSyntax.js";
+import { auditFix } from "./auditFix.js";
+import { metricsLog } from "./metrics.js";
+import { definitionIndex, propsSource, unpackWidget } from "./unpack.js";
+import { isFrameworkProject } from "../utils/migrationGuide.js";
+import { c, banner, section } from "../utils/colors.js";
+
+/**
+ * The severity a design-gate code carries. The gate knows which codes are
+ * errors and which are a reviewer's job; the report must not call a review
+ * an error, or an agent will "fix" a warning and break the screen.
+ *
+ * @param {string} code
+ * @returns {"error"|"warn"|"info"}
+ */
+const designSeverity = (code) => {
+  if (ERROR_CODES.has(code)) return "error";
+  if (WARN_CODES.has(code)) return "warn";
+  if (INFO_CODES.has(code)) return "info";
+  return "error";
+};
+
+/**
+ * Normalise a finding so the report does not care which gate produced it.
+ *
+ * @param {object} finding
+ * @param {"error"|"warn"|"info"} [severity]
+ * @returns {object}
+ */
+const shape = (finding, severity) => ({
+  code: finding.code || "UNKNOWN",
+  file: finding.file || null,
+  line: finding.line || 0,
+  column: finding.column || 0,
+  message: finding.message || "",
+  detail: finding.detail || "",
+  suggestion: finding.suggestion || null,
+  severity: severity || finding.severity || "error",
+});
+
+/**
+ * The path a finding names, relative to the project it belongs to.
+ *
+ * @param {string|null} file
+ * @param {string} root
+ * @returns {string|null}
+ */
+const relative = (file, root) => {
+  if (!file) return null;
+  const absolute = path.isAbsolute(file) ? file : path.resolve(process.cwd(), file);
+  const rel = path.relative(root, absolute);
+  return rel && !rel.startsWith("..") ? rel : absolute;
+};
+
+/* ------------------------------------------------------------------ gates */
+
+const syntaxGate = async (root) => {
+  const findings = await scanSyntax(root);
+  return {
+    name: "syntax",
+    title: "Parse gate",
+    findings: findings.map((finding) => shape(finding)),
+    exitCode: findings.length ? 1 : 0,
+  };
+};
+
+const apiGate = async (root) => {
+  const { findings, exitCode } = await checkApi(["--root", root], { print: () => {} });
+  return {
+    name: "api",
+    title: "Anti-hallucination gate",
+    findings: findings.map((finding) =>
+      shape(finding, SEVERITY[finding.code] === "error" ? "error" : "warn"),
+    ),
+    exitCode,
+  };
+};
+
+/**
+ * The design findings this checkout has already decided about.
+ *
+ * The framework runs its own clean-UI gate over `src/`, and two of its findings
+ * are not defects: they are rules written for a screen, applied to a module that
+ * is not one. `scripts/check-framework-ui.mjs` already knows that, through
+ * `scripts/framework-ui-baseline.json`, where every entry carries its reason
+ * (and a new one cannot be recorded without one). `verify` did not, so every run
+ * printed those two under "what to review" — and a review list that never
+ * changes is a list an agent learns to skip, which costs it the finding that
+ * does matter.
+ *
+ * Only the framework checkout reads the file: a consumer's app has no such
+ * baseline, and its findings are its own to answer.
+ *
+ * @param {string} root
+ * @returns {Map<string, string>} `file:code` → the reason it was accepted
+ */
+const acceptedDesignFindings = (root) => {
+  if (!isFrameworkProject(root)) return new Map();
+  try {
+    const text = fs.readFileSync(path.join(root, "scripts", "framework-ui-baseline.json"), "utf8");
+    return new Map(Object.entries(JSON.parse(text)));
+  } catch {
+    return new Map();
+  }
+};
+
+const uiGate = async (root) => {
+  const { findings } = await checkUi(["--root", root], { print: () => {} });
+  const accepted = acceptedDesignFindings(root);
+  // The baseline key drops the line number on purpose: editing anything above a
+  // finding must not invalidate a decision about the file.
+  const keyOf = (finding) =>
+    finding.file ? `${relative(finding.file, root).split(path.sep).join("/")}:${finding.code}` : "";
+  const kept = [];
+  const decided = [];
+  for (const finding of findings) {
+    const why = accepted.get(keyOf(finding));
+    if (why === undefined) kept.push(finding);
+    else decided.push({ file: relative(finding.file, root), code: finding.code, why });
+  }
+  return {
+    name: "ui",
+    title: "Design gate",
+    findings: kept.map((finding) => shape(finding, designSeverity(finding.code))),
+    accepted: decided,
+    // An accepted finding is decided, so it cannot fail the run — including one
+    // whose code is an error elsewhere, which is what its written reason is for.
+    exitCode: kept.some((finding) => designSeverity(finding.code) === "error") ? 1 : 0,
+  };
+};
+
+const structureGate = async (root) => {
+  const findings = await scanStructure(root);
+  return {
+    name: "structure",
+    title: "Structural gate",
+    findings: findings.map((finding) => shape(finding)),
+    exitCode: findings.some((item) => item.severity === "error") ? 1 : 0,
+  };
+};
+
+const surfaceGate = async (root) => {
+  const api = await loadApi(root);
+  const definitions = definitionIndex(api.dir);
+  const widgets = api.exports
+    .filter((name) => /^[A-Z]/.test(name) && propsSource(api, name))
+    .sort();
+  const findings = [];
+  // The same pass that produces the findings also produces the per-widget
+  // ledger `flet-box report` prints. `unpackWidget` is the expensive half of
+  // this gate, and a report that unpacked the surface a second time would make
+  // the diagnosis cost more than the run it diagnoses.
+  const ledger = [];
+  for (const name of widgets) {
+    const data = unpackWidget(api, name, definitions);
+    const declaration = data.interface || `${name}Props`;
+    ledger.push({
+      name,
+      interface: data.interface,
+      props: data.props.length,
+      invented: data.gaps.invented.map((prop) => prop.prop),
+      dead: data.gaps.dead.map((prop) => prop.prop),
+      notDeclared: data.gaps.hidden.map((prop) => prop.prop),
+      docOnly: data.gaps.docOnly,
+      codeOnly: data.gaps.codeOnly,
+    });
+    for (const prop of data.gaps.invented) {
+      findings.push(
+        shape({
+          code: "INVENTED_PROP",
+          file: data.sourceFile,
+          message: `${name}.${prop.prop} is read by the widget but declared nowhere — nobody can discover it`,
+          suggestion: `declare it in ${declaration} (src/index.d.ts)`,
+        }),
+      );
+    }
+    for (const prop of data.gaps.dead) {
+      findings.push(
+        shape({
+          code: "DEAD_PROP",
+          file: data.sourceFile,
+          message: `${name}.${prop.prop} is declared but nothing routes or reads it`,
+          suggestion: "wire it in the widget, or drop it from the declaration",
+          severity: "warn",
+        }),
+      );
+    }
+  }
+  const errors = findings.filter((item) => item.severity === "error");
+  return {
+    name: "surface",
+    title: "Prop contract",
+    findings,
+    widgets: ledger,
+    exitCode: errors.length ? 1 : 0,
+  };
+};
+
+/**
+ * The suites a project may declare, in the order they run.
+ *
+ * They are the project's own detection scripts — in this repository the static
+ * checkers, the browser harnesses and the visual baseline. The order is the
+ * cheap one first, so a broken rule is named before Chrome is launched.
+ */
+const SUITES = ["test", "test:browser", "test:visual"];
+
+/**
+ * The tail of a child's output, which is where a check says why it failed.
+ *
+ * @param {string} text
+ * @param {number} [count]
+ * @returns {string}
+ */
+const lastLines = (text, count = 15) =>
+  String(text)
+    .split(/\r?\n/)
+    .map((line) => line.trimEnd())
+    .filter(Boolean)
+    .slice(-count)
+    .join("\n");
+
+/**
+ * The project's own test scripts.
+ *
+ * A generated app declares none, so the gate says so instead of pretending to
+ * have run something. In this repository it is every suite `package.json`
+ * declares: the static checkers, the browser harnesses and the visual
+ * baseline — the command an agent runs after an edit has to be the same one
+ * that finds a parse error, a broken rule and a moved pixel, or the agent
+ * stops at the first green command it knows.
+ *
+ * The children run with `FLETBOX_VERIFY_NESTED` set, so a `test` script that
+ * itself calls `flet-box verify` cannot recurse: the nested run answers with
+ * the fast gates and returns.
+ *
+ * @param {string} root
+ * @param {{print?: (line: string) => void}} [context]
+ * @returns {Promise<object>}
+ */
+const testsGate = async (root, context = {}) => {
+  const skipped = (why) => ({
+    name: "tests",
+    title: "Test suite",
+    findings: [],
+    exitCode: 0,
+    skipped: why,
+  });
+  if (process.env.FLETBOX_VERIFY_NESTED) return skipped("already inside a verify run");
+
+  let scripts = {};
+  try {
+    scripts = JSON.parse(fs.readFileSync(path.join(root, "package.json"), "utf8")).scripts ?? {};
+  } catch {
+    /* a project without a manifest declares no suite */
+  }
+  const declared = SUITES.filter((script) => scripts[script]);
+  if (!declared.length) return skipped("this project declares no test script");
+
+  const say = context.print ?? (() => {});
+  const findings = [];
+  let missingNpm = false;
+  for (const script of declared) {
+    say(c("gray", `   tests    npm run ${script} …`));
+    const started = Date.now();
+    const result = await runTestSuite(root, script);
+    const seconds = ((Date.now() - started) / 1000).toFixed(1);
+    if (result.missing) {
+      missingNpm = true;
+      say(c("gray", `   tests    npm run ${script} — skipped, npm is not on PATH`));
+      continue;
+    }
+    say(
+      result.code
+        ? c("red", `   tests    npm run ${script} — exit ${result.code} (${seconds}s)`)
+        : c("gray", `   tests    npm run ${script} — green (${seconds}s)`),
+    );
+    if (!result.code) continue;
+    findings.push(
+      shape({
+        code: "TEST_SUITE_FAILED",
+        message: `npm run ${script} exited ${result.code}`,
+        detail: lastLines(result.output),
+        suggestion: `run \`npm run ${script}\` alone — the check that failed names its own rule, and this is how its output ended`,
+      }),
+    );
+  }
+  if (missingNpm && !findings.length) return skipped("npm is not on PATH");
+  return {
+    name: "tests",
+    title: "Test suite",
+    findings,
+    exitCode: findings.length ? 1 : 0,
+  };
+};
+
+/**
+ * One diagnostic line from `tsc --pretty false`:
+ * `src/file.js(12,5): error TS2339: Property 'update' does not exist …`
+ *
+ * The message is kept with its `TS####` code: it is what the editor prints and
+ * what the reader can paste into a search, and it tells a missing property
+ * (2339) from a wrong union (2769) without parsing English.
+ */
+const TYPE_DIAGNOSTIC = /^(.+?)\((\d+),(\d+)\): error (TS\d+): (.*)$/;
+
+/**
+ * Resolve the compiler without depending on it.
+ *
+ * This file ships inside every app, so `typescript` is never a declared
+ * dependency of the project being checked: an editor brings its own, and an app
+ * installed from the registry has none. A missing compiler means the gate
+ * cannot answer and says so — it does not pass silently and it does not fail
+ * the build. `FLETBOX_TYPESCRIPT` points at a directory with `lib/tsc.js`.
+ *
+ * @returns {string|null} Absolute path to `tsc.js`.
+ */
+const resolveTsc = () => {
+  const explicit = process.env.FLETBOX_TYPESCRIPT;
+  if (explicit) {
+    const candidate = path.join(explicit, "lib", "tsc.js");
+    return fs.existsSync(candidate) ? candidate : null;
+  }
+  try {
+    const require = createRequire(import.meta.url);
+    const candidate = path.join(path.dirname(require.resolve("typescript")), "tsc.js");
+    return fs.existsSync(candidate) ? candidate : null;
+  } catch {
+    return null;
+  }
+};
+
+/**
+ * The type gate: the project's own `jsconfig.json` through the real compiler.
+ *
+ * The check the editor runs is the one that was missing from `verify`: a widget
+ * declared `HTMLElement` made the documented `widget.update({ … })` a red
+ * squiggle in eight files, and nothing in the pipeline noticed, because
+ * `check-app-types` only proves the five generated templates. This gate reads
+ * the same `jsconfig.json` the editor reads, so the app and the editor cannot
+ * disagree about whose fault an error is.
+ *
+ * @param {string} root Project root.
+ * @returns {Promise<object>}
+ */
+const typesGate = async (root) => {
+  const skipped = (why) => ({
+    name: "types",
+    title: "Type check",
+    findings: [],
+    exitCode: 0,
+    skipped: why,
+  });
+  const config = ["jsconfig.json", "tsconfig.json"].find((name) =>
+    fs.existsSync(path.join(root, name)),
+  );
+  if (!config) return skipped("no jsconfig.json — the editor has nothing to check");
+  const tsc = resolveTsc();
+  if (!tsc) return skipped("typescript is not installed — the editor brings its own");
+
+  const output = await new Promise((resolve) => {
+    execFile(
+      process.execPath,
+      [tsc, "-p", config, "--noEmit", "--pretty", "false"],
+      { cwd: root, maxBuffer: 32 * 1024 * 1024 },
+      (error, stdout, stderr) =>
+        resolve({
+          code: error ? (typeof error.code === "number" ? error.code : 1) : 0,
+          text: `${stdout ?? ""}${stderr ?? ""}`,
+        }),
+    );
+  });
+
+  const findings = [];
+  let last = null;
+  for (const raw of output.text.split("\n")) {
+    const line = raw.replace(/\r$/, "");
+    if (!line.trim()) continue;
+    const match = TYPE_DIAGNOSTIC.exec(line);
+    if (match) {
+      last = shape({
+        code: "TYPE_ERROR",
+        file: path.resolve(root, match[1]),
+        line: Number(match[2]),
+        column: Number(match[3]),
+        message: `${match[4]} ${match[5]}`,
+      });
+      findings.push(last);
+    } else if (last && /^\s/.test(line)) {
+      // tsc elaborates a type error over the next indented lines; they belong
+      // to the finding above, not to a finding of their own.
+      last.detail = last.detail ? `${last.detail}\n${line.trim()}` : line.trim();
+    } else if (!findings.length) {
+      // A failure that is not about one line (a bad `include`, a missing file)
+      // still has to be named, or a red gate would print no reason.
+      last = shape({ code: "TYPE_ERROR", message: line.trim() });
+      findings.push(last);
+    }
+  }
+  if (output.code !== 0 && !findings.length) {
+    findings.push(shape({ code: "TYPE_ERROR", message: `tsc exited ${output.code}` }));
+  }
+  return {
+    name: "types",
+    title: "Type check",
+    findings,
+    exitCode: output.code === 0 ? 0 : 1,
+  };
+};
+
+/**
+ * Every gate, by name.
+ *
+ * Exported with `GATE_ORDER` because `scripts/selftest-gates.mjs` holds the
+ * inventory: it proves the two lists agree (a gate in only one of them is a
+ * gate that never runs, or a name that runs nothing) and that every gate has a
+ * fixture. Without that, the coverage could shrink one entry at a time.
+ */
+export const GATES = {
+  syntax: syntaxGate,
+  api: apiGate,
+  types: typesGate,
+  ui: uiGate,
+  structure: structureGate,
+  surface: surfaceGate,
+  tests: testsGate,
+};
+
+/** Every gate, in the order the report prints them. */
+export const GATE_ORDER = ["syntax", "api", "types", "ui", "structure", "surface", "tests"];
+
+/* ----------------------------------------------------------------- report */
+
+/**
+ * How many review findings the report prints before it points at `--json`.
+ *
+ * A warning is a reading, not a failure, and a report nobody can scan is one
+ * nobody reads: past this many, the count still prints.
+ */
+const REVIEW_LIMIT = 12;
+
+/**
+ * The report: one line per gate, then the findings of the gates that failed,
+ * then the ones that are a reading rather than a failure.
+ *
+ * A warning is printed with its file, its line and its suggestion too. It used
+ * to be counted and never named — `ui 2 warning(s)` told an agent that
+ * something was wrong somewhere, which is the one thing it cannot act on.
+ *
+ * @param {object} report
+ * @returns {string[]}
+ */
+export const formatReport = (report) => {
+  const lines = [];
+  const count = (severity) =>
+    report.steps.reduce(
+      (n, step) => n + step.findings.filter((item) => item.severity === severity).length,
+      0,
+    );
+  const errors = count("error");
+  const warnings = count("warn");
+  const infos = count("info");
+  const failed = report.steps.filter((step) => step.exitCode);
+  const ok = failed.length === 0;
+
+  for (const step of report.steps) {
+    const stepErrors = step.findings.filter((item) => item.severity === "error").length;
+    const stepWarns = step.findings.filter((item) => item.severity === "warn").length;
+    const stepInfos = step.findings.filter((item) => item.severity === "info").length;
+    const mark = step.exitCode ? c("red", "✗") : c("green", "✓");
+    const counts = [
+      stepErrors ? c("red", `${stepErrors} error(s)`) : "",
+      stepWarns ? c("yellow", `${stepWarns} warning(s)`) : "",
+      stepInfos ? c("gray", `${stepInfos} to review`) : "",
+    ]
+      .filter(Boolean)
+      .join(", ");
+    // An accepted finding is a decision with a reason in the baseline file, not
+    // work: it is counted once, by name, so the number stays visible without
+    // asking the same question on every run.
+    const decided = step.accepted?.length
+      ? c("gray", `${step.accepted.length} accepted (framework-ui-baseline.json)`)
+      : "";
+    const status = step.skipped
+      ? c("gray", `skipped — ${step.skipped}`)
+      : [counts, decided].filter(Boolean).join(", ") || c("gray", "clean");
+    lines.push(`  ${mark} ${c("brightCyan", step.name.padEnd(8))} ${status}`);
+  }
+
+  const failing = failed.flatMap((step) =>
+    step.findings
+      .filter((item) => item.severity === "error")
+      .map((item) => ({ ...item, gate: step.name })),
+  );
+  const printFinding = (item) => {
+    const at = item.column ? `${item.line}:${item.column}` : `${item.line}`;
+    const where = item.file
+      ? c("brightCyan", `${relative(item.file, report.root)}:${at}`)
+      : c("gray", item.gate);
+    lines.push(`  ${where}  ${c("yellow", item.code)} — ${item.message}`);
+    if (item.detail) {
+      for (const text of String(item.detail).split("\n")) lines.push(`     ${c("gray", text)}`);
+    }
+    if (item.suggestion) lines.push(`     ${c("gray", "→ " + item.suggestion)}`);
+  };
+
+  if (failing.length) {
+    lines.push("");
+    lines.push(section("•", "what to fix", "#f87171"));
+    for (const item of failing) printFinding(item);
+  }
+
+  const review = report.steps.flatMap((step) =>
+    step.findings
+      .filter((item) => item.severity !== "error")
+      .map((item) => ({ ...item, gate: step.name })),
+  );
+  if (review.length) {
+    lines.push("");
+    lines.push(section("•", "what to review", "#fbbf24"));
+    for (const item of review.slice(0, REVIEW_LIMIT)) printFinding(item);
+    if (review.length > REVIEW_LIMIT) {
+      lines.push(
+        `  ${c("gray", `… ${review.length - REVIEW_LIMIT} more — \`flet-box verify --json\` carries them all`)}`,
+      );
+    }
+  }
+
+  const suitesRan = report.steps.some((step) => step.name === "tests" && !step.skipped);
+  const counted = [`${errors} error(s)`];
+  if (warnings) counted.push(`${warnings} warning(s)`);
+  if (infos) counted.push(`${infos} to review`);
+  lines.push("");
+  lines.push(
+    ok
+      ? c(
+          "green",
+          `✅ ${counted.join(", ")} — the code parses, is legal, on the scale and on the tokens${suitesRan ? ", and every suite is green" : ""}.`,
+        )
+      : c(
+          "red",
+          `❌ ${errors} error(s), ${warnings} warning(s) — fix the line named above, then run it again.`,
+        ),
+  );
+  return lines;
+};
+
+/**
+ * The machine-readable report.
+ *
+ * @param {object} report
+ * @returns {object}
+ */
+const asJson = (report) => ({
+  root: report.root,
+  package: report.package,
+  ok: report.steps.every((step) => !step.exitCode),
+  steps: report.steps.map((step) => ({
+    name: step.name,
+    title: step.title,
+    exitCode: step.exitCode,
+    skipped: step.skipped || null,
+    findings: step.findings,
+    // Present only for a gate that can accept a finding: the entries this
+    // checkout has decided about, with the reason it wrote down.
+    ...(step.accepted?.length ? { accepted: step.accepted } : {}),
+  })),
+});
+
+/* --------------------------------------------------------------- the log */
+
+/**
+ * Append one row to the metrics log.
+ *
+ * The row is the run, not the session: when it happened, which project, whether
+ * the gates were green, which gates failed and which codes they reported. The
+ * metric script groups the rows into generations — a run of failing rows closed
+ * by the first green one — and reads the north metric off that.
+ *
+ * @param {string} root
+ * @param {object} report
+ * @returns {string|null} the row that was written
+ */
+export const recordGeneration = (root, report) => {
+  const file = metricsLog(root);
+  if (!file) return null;
+  const failed = report.steps.filter((step) => step.exitCode);
+  const codes = [
+    ...new Set(
+      report.steps.flatMap((step) =>
+        step.findings.filter((item) => item.severity === "error").map((item) => item.code),
+      ),
+    ),
+  ];
+  const row = [
+    new Date().toISOString(),
+    path.basename(root),
+    failed.length ? "✗" : "✓",
+    failed.length ? failed.map((step) => step.name).join(", ") : "—",
+    codes.length ? codes.join(", ") : "—",
+  ].join(" | ");
+  try {
+    const text = fs.readFileSync(file, "utf8");
+    fs.appendFileSync(file, `${text.endsWith("\n") ? "" : "\n"}| ${row} |\n`, "utf8");
+  } catch {
+    return null;
+  }
+  return row;
+};
+
+/* -------------------------------------------------------------------- run */
+
+/**
+ * Run every gate, in parallel, and collect what they say.
+ *
+ * @param {string} root Project root.
+ * @param {string[]} names Gate names to run.
+ * @param {{print?: (line: string) => void}} [context] Passed to a gate that reports progress.
+ * @returns {Promise<object>}
+ */
+export const runGates = async (root, names, context = {}) => {
+  const steps = await Promise.all(
+    names.map(async (name) => {
+      const gate = GATES[name];
+      try {
+        const result = await gate(root, context);
+        return { ...result, name };
+      } catch (error) {
+        return {
+          name,
+          title: name,
+          findings: [
+            shape({
+              code: "GATE_FAILED",
+              message: error.message,
+              suggestion: "the gate could not run — check that flet-box is installed",
+            }),
+          ],
+          exitCode: 1,
+        };
+      }
+    }),
+  );
+  let version = "?";
+  try {
+    version = (await loadApi(root)).pkg.version;
+  } catch {
+    /* the gates report the missing package themselves */
+  }
+  return { root, package: version, steps };
+};
+
+/**
+ * The `flet-box verify` command.
+ *
+ * @param {string[]} argv Flags after the command name.
+ * @param {{print?: (line: string) => void, root?: string}} [io]
+ * @returns {Promise<{exitCode: number, report: object}>}
+ */
+export const verify = async (argv = [], io = {}) => {
+  const print = io.print ?? ((line) => console.log(line));
+  const rootIndex = argv.indexOf("--root");
+  const root = path.resolve(
+    rootIndex >= 0 && argv[rootIndex + 1] ? argv[rootIndex + 1] : io.root || process.cwd(),
+  );
+  const only = argv.includes("--only") ? argv[argv.indexOf("--only") + 1] : null;
+  const watching = argv.includes("--watch");
+  // `--watch` starts the loop without the suites: the gate an editor reruns on
+  // every keystroke cannot spend a minute in Chrome. `--no-tests` is the same
+  // choice stated by the caller — a CI job that runs the suites in their own
+  // jobs uses it, so the report says what happened instead of borrowing the
+  // "already inside a verify run" note that only tells the truth about nesting.
+  // A plain `verify` still runs them, and `--only tests` runs them alone.
+  const skipTests = watching || argv.includes("--no-tests");
+  const names = only
+    ? GATE_ORDER.filter((name) => name === only)
+    : GATE_ORDER.filter((name) => !skipTests || name !== "tests");
+
+  if (!names.length) {
+    print(c("red", `Unknown gate "${only}" — the gates are: ${GATE_ORDER.join(", ")}`));
+    return { exitCode: 1, report: null };
+  }
+
+  // `--fix` runs the mechanical repair first, so a mistyped prop or a wrong
+  // relative path is corrected by the tool that can prove the correction and
+  // never reaches the agent as work to do.
+  if (argv.includes("--fix")) {
+    const fix = await auditFix(["--root", root], { print });
+    if (fix.changed.length) {
+      print(
+        c("gray", `   ${fix.changed.length} file(s) rewritten — the gates below read the result.`),
+      );
+    }
+    print("");
+  }
+
+  const report = await runGates(root, names, argv.includes("--json") ? {} : { print });
+  const failed = report.steps.some((step) => step.exitCode);
+  if (argv.includes("--json")) {
+    print(JSON.stringify(asJson(report), null, 2));
+  } else {
+    print(
+      banner(
+        failed ? "❌ FletBox verify" : "✅ FletBox verify",
+        `${path.basename(root)} · flet-box@${report.package}`,
+      ),
+    );
+    print("");
+    for (const line of formatReport(report)) print(line);
+    print("");
+  }
+
+  // The log is written for every run, not only for the green ones: a metric
+  // that counts only the successes cannot tell a correction from a first try.
+  const row = recordGeneration(root, report);
+  if (row && !argv.includes("--json")) {
+    print(c("gray", `   logged → ${metricsLog(root)}`));
+    print("");
+  }
+
+  if (argv.includes("--watch")) {
+    print(
+      c(
+        "dim",
+        "👀 watching src/ — every edit is checked again (500 ms debounce); Ctrl+C to stop",
+      ),
+    );
+    watchVerify({
+      root,
+      onReport: (live) => {
+        console.log("");
+        for (const line of formatReport(live)) console.log(line);
+        console.log("");
+      },
+    });
+    return { exitCode: 0, report };
+  }
+
+  return { exitCode: failed ? 1 : 0, report };
+};
+
+/* ------------------------------------------------------------------ watch */
+
+/**
+ * Watch the project and re-run the gates on every change.
+ *
+ * The debounce is 500 ms: an agent writes a file and the editor writes a
+ * second one, and two runs would report the same finding twice.
+ *
+ * @param {{root: string, onReport: (report: object) => void, debounce?: number}} options
+ * @returns {fs.FSWatcher}
+ */
+export const watchVerify = ({ root, onReport, debounce = 500 }) => {
+  const watchDir = fs.existsSync(path.join(root, "src")) ? path.join(root, "src") : root;
+  let timer = null;
+  let running = false;
+  let queued = false;
+  let previous = "";
+
+  const tick = async () => {
+    if (running) {
+      queued = true;
+      return;
+    }
+    running = true;
+    const report = await runGates(root, GATE_ORDER.filter((name) => name !== "tests"));
+    running = false;
+    const signature = report.steps
+      .map((step) =>
+        `${step.name}:${step.exitCode}:${step.findings
+          .map((item) => `${item.code}@${item.file}:${item.line}`)
+          .join(",")}`,
+      )
+      .join("|");
+    if (signature !== previous) {
+      previous = signature;
+      onReport(report);
+    }
+    if (queued) {
+      queued = false;
+      await tick();
+    }
+  };
+
+  const watcher = fs.watch(watchDir, { recursive: true }, (_event, filename) => {
+    if (!filename) return;
+    const name = String(filename);
+    if (name.includes("node_modules") || name.startsWith(".")) return;
+    if (!/\.(m?js|css|html|json)$/.test(name)) return;
+    if (timer) clearTimeout(timer);
+    timer = setTimeout(() => {
+      timer = null;
+      tick();
+    }, debounce);
+  });
+  return watcher;
+};
+
+/**
+ * Run one of the project's own npm scripts and keep what it said.
+ *
+ * `--silent` matters: without it npm prints its own banner and, on failure, a
+ * block of `npm error` lines that bury the one line the check wrote. With it
+ * the output is the check's own words, which is the whole point of running the
+ * suite from here.
+ *
+ * @param {string} root
+ * @param {string} script The npm script name.
+ * @returns {Promise<{code: number|null, output: string, missing?: string}>}
+ */
+export const runTestSuite = (root, script = "test") =>
+  new Promise((resolve) => {
+    execFile(
+      process.platform === "win32" ? "npm.cmd" : "npm",
+      ["run", script, "--silent"],
+      {
+        cwd: root,
+        env: { ...process.env, FLETBOX_VERIFY_NESTED: "1" },
+        timeout: 15 * 60 * 1000,
+        maxBuffer: 32 * 1024 * 1024,
+        shell: process.platform === "win32",
+      },
+      (error, stdout, stderr) => {
+        const output = `${stdout}${stderr}`;
+        if (!error) return resolve({ code: 0, output });
+        if (error.code === "ENOENT") return resolve({ code: null, output, missing: "npm" });
+        resolve({ code: typeof error.code === "number" ? error.code : 1, output });
+      },
+    );
+  });
+
+export default verify;