diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
new file mode 100644
index 0000000..ae6cd74
--- /dev/null
+++ b/.github/workflows/test.yml
@@ -0,0 +1,22 @@
+name: Tests
+on:
+  push:
+  pull_request:
+permissions:
+  contents: read
+jobs:
+  test:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+        node: [22, 24]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: ${{ matrix.node }}
+      - run: npm install --ignore-scripts --no-audit --no-fund
+      - run: npm run typecheck
+      - run: npm test
diff --git a/package.json b/package.json
index 781c55b..6bf9aaa 100644
--- a/package.json
+++ b/package.json
@@ -29,7 +29,7 @@
     "LICENSE"
   ],
   "scripts": {
-    "test": "tsx --test test/*.test.mjs",
+    "test": "node test-runner.mjs",
     "typecheck": "tsc --noEmit"
   },
   "peerDependencies": {
diff --git a/src/cli.ts b/src/cli.ts
index f8627ba..df448ff 100644
--- a/src/cli.ts
+++ b/src/cli.ts
@@ -1,4 +1,4 @@
-import { existsSync } from "node:fs";
+import { statSync } from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
 import { execFile, spawn } from "node:child_process";
@@ -6,21 +6,50 @@ import { promisify } from "node:util";
 
 const execFileAsync = promisify(execFile);
 
-const KNOWN_BINS = [
-  process.env.DEVIN_CLI,
-  join(homedir(), ".local/bin/devin"),
-  join(homedir(), ".devin/bin/devin"),
-  "/opt/homebrew/bin/devin",
-  "/usr/local/bin/devin",
-  "/Applications/Devin.app/Contents/Resources/app/extensions/windsurf/devin/bin/devin",
-].filter((p): p is string => Boolean(p));
+function candidateBins(): string[] {
+  const localAppData = process.env.LOCALAPPDATA ?? join(homedir(), "AppData/Local");
+  return [
+    // Explicit override always wins. Read lazily so a value set after module
+    // load (e.g. by the host process) is still honoured.
+    process.env.DEVIN_CLI,
+    ...(process.platform === "win32" ? [
+    // Official installer location (%LOCALAPPDATA%\devin\cli\bin\devin.exe)
+    join(localAppData, "devin/cli/bin/devin.exe"),
+    // Devin Desktop (Windsurf-based) bundled CLI, per-user install
+    join(localAppData, "Programs/Devin/resources/app/extensions/windsurf/devin/bin/devin.exe"),
+    // Devin Desktop bundled CLI, machine-wide install
+    process.env.ProgramFiles &&
+      join(process.env.ProgramFiles, "Devin/resources/app/extensions/windsurf/devin/bin/devin.exe"),
+    // Set by the official Windows installer when it completes normally.
+    process.env.LOCALAPPDATA && join(process.env.LOCALAPPDATA, "devin/bin/devin.exe"),
+    // Manual fallback for installs that could not finish while Devin was running.
+    join(homedir(), ".devin-cli/devin.exe"),
+    ] : [
+    join(homedir(), ".local/bin/devin"),
+    join(homedir(), ".devin/bin/devin"),
+    "/opt/homebrew/bin/devin",
+    "/usr/local/bin/devin",
+    "/Applications/Devin.app/Contents/Resources/app/extensions/windsurf/devin/bin/devin",
+    ]),
+  ].filter((p): p is string => Boolean(p));
+}
 
 let cachedBin: string | null | undefined;
 
+function isFile(path: string): boolean {
+  try {
+    return statSync(path).isFile();
+  } catch {
+    return false;
+  }
+}
+
 export function findDevinBin(): string | null {
+  // An explicit override must take effect even after a successful discovery.
+  if (process.env.DEVIN_CLI && isFile(process.env.DEVIN_CLI)) return process.env.DEVIN_CLI;
   if (cachedBin !== undefined) return cachedBin;
-  for (const bin of KNOWN_BINS) {
-    if (existsSync(bin)) {
+  for (const bin of candidateBins()) {
+    if (isFile(bin)) {
       cachedBin = bin;
       return bin;
     }
@@ -36,15 +65,24 @@ export function clearDevinBinCache(): void {
 export async function whichDevin(): Promise<string | null> {
   const known = findDevinBin();
   if (known) return known;
-  try {
-    const { stdout } = await execFileAsync("/usr/bin/which", ["devin"], { timeout: 5_000 });
-    const path = stdout.trim();
-    if (path && existsSync(path)) {
-      cachedBin = path;
-      return path;
+  // Windows has no `which`; use `where` and probe the returned paths directly.
+  const isWindows = process.platform === "win32";
+  const locator = isWindows ? "where.exe" : "/usr/bin/which";
+  const candidates = isWindows ? ["devin.exe", "devin.cmd", "devin"] : ["devin"];
+  for (const candidate of candidates) {
+    try {
+      const { stdout } = await execFileAsync(locator, [candidate], { timeout: 5_000 });
+      // `where` may return several lines; take the first existing one.
+      for (const line of stdout.split(/\r?\n/)) {
+        const path = line.trim();
+        if (path && isFile(path)) {
+          cachedBin = path;
+          return path;
+        }
+      }
+    } catch {
+      // not on PATH
     }
-  } catch {
-    // not on PATH
   }
   return null;
 }
@@ -60,9 +98,18 @@ export async function runDevin(
     );
   }
 
+  // Batch wrappers need cmd.exe. Quote paths with spaces, and reject shell
+  // expansion characters rather than interpreting them as part of a command.
+  const wrapper = process.platform === "win32" && /\.(cmd|bat)$/i.test(bin);
+  if (wrapper && [bin, ...args].some((part) => /["%!\r\n]/.test(part))) {
+    throw new Error("Unsupported shell characters in Devin wrapper path or arguments. Set DEVIN_CLI to devin.exe.");
+  }
+  const executable = wrapper ? process.env.ComSpec || "cmd.exe" : bin;
+  const commandArgs = wrapper ? ["/d", "/s", "/c", `"${[bin, ...args].map((part) => `"${part}"`).join(" ")}"`] : args;
+
   if (opts.inheritStdio) {
     return await new Promise((resolve, reject) => {
-      const child = spawn(bin, args, { stdio: "inherit" });
+      const child = spawn(executable, commandArgs, { stdio: "inherit", windowsVerbatimArguments: wrapper });
       child.on("error", reject);
       child.on("close", (code) => {
         resolve({ stdout: "", stderr: "", code: code ?? 1 });
@@ -71,9 +118,10 @@ export async function runDevin(
   }
 
   try {
-    const { stdout, stderr } = await execFileAsync(bin, args, {
+    const { stdout, stderr } = await execFileAsync(executable, commandArgs, {
       timeout: opts.timeoutMs ?? 30_000,
       maxBuffer: 8 * 1024 * 1024,
+      windowsVerbatimArguments: wrapper,
     });
     return { stdout, stderr, code: 0 };
   } catch (error) {
diff --git a/test-runner.mjs b/test-runner.mjs
new file mode 100644
index 0000000..6b25dfe
--- /dev/null
+++ b/test-runner.mjs
@@ -0,0 +1,8 @@
+import { readdirSync } from "node:fs";
+import { spawnSync } from "node:child_process";
+import { join } from "node:path";
+
+const files = readdirSync("test").filter((file) => file.endsWith(".test.mjs")).sort().map((file) => join("test", file));
+const result = spawnSync(process.execPath, ["--import", "tsx", "--test", ...files], { stdio: "inherit" });
+if (result.error) throw result.error;
+process.exit(result.status ?? 1);
diff --git a/test/catalog-cache.test.mjs b/test/catalog-cache.test.mjs
index 4669244..f2db4ed 100644
--- a/test/catalog-cache.test.mjs
+++ b/test/catalog-cache.test.mjs
@@ -52,8 +52,8 @@ test("checks cache age and future timestamps", () => {
 });
 
 test("uses XDG cache directory when configured", () => {
-  assert.equal(catalogCachePath({ XDG_CACHE_HOME: "/cache" }, "/home/test"), "/cache/pi-devin/models.json");
-  assert.equal(catalogCachePath({}, "/home/test"), "/home/test/.cache/pi-devin/models.json");
+  assert.equal(catalogCachePath({ XDG_CACHE_HOME: "/cache" }, "/home/test"), join("/cache", "pi-devin", "models.json"));
+  assert.equal(catalogCachePath({}, "/home/test"), join("/home/test", ".cache", "pi-devin", "models.json"));
 });
 
 for (const [field, value] of Object.entries({ cost_summary: {}, max_context_tokens: "1000", max_output_tokens: -1, model_uid: "", is_beta: "yes" })) {
diff --git a/test/catalog-startup.test.mjs b/test/catalog-startup.test.mjs
index 2a4f8f4..6243f24 100644
--- a/test/catalog-startup.test.mjs
+++ b/test/catalog-startup.test.mjs
@@ -8,10 +8,12 @@ import { catalogCachePath, writeCatalogCache } from "../src/catalog-cache.ts";
 
 const directory = fs.mkdtempSync(join(tmpdir(), "pi-devin-catalog-test-"));
 after(() => fs.rmSync(directory, { recursive: true, force: true }));
-const cli = join(directory, "devin");
+const cli = join(directory, process.platform === "win32" ? "devin.cmd" : "devin");
 const counter = join(directory, "calls");
 const catalog = { families: [{ family_label: "Test", family_uid: "test", slug: "test", variants: [{ model_uid: "test-high", label: "Test high", cost_summary: "$1/MTok In, $2/MTok Out" }] }] };
-fs.writeFileSync(cli, `#!${process.execPath}\nconst fs = require('node:fs');\nfs.appendFileSync(${JSON.stringify(counter)}, 'call\\n');\nsetTimeout(() => { console.log(process.env.TEST_CATALOG); process.exit(Number(process.env.TEST_CLI_EXIT || 0)); }, 30);\n`, { mode: 0o700 });
+const script = join(directory, "catalog.cjs");
+fs.writeFileSync(script, `const fs = require('node:fs');\nfs.appendFileSync(${JSON.stringify(counter)}, 'call\\n');\nsetTimeout(() => { console.log(process.env.TEST_CATALOG); process.exit(Number(process.env.TEST_CLI_EXIT || 0)); }, 30);\n`);
+fs.writeFileSync(cli, process.platform === "win32" ? `@echo off\r\n"${process.execPath}" "${script}"\r\n` : `#!/bin/sh\nexec "${process.execPath}" "${script}"\n`, { mode: 0o700 });
 const originalEnv = { ...process.env };
 process.env.DEVIN_CLI = cli;
 const { default: extension } = await import("../extensions/index.ts");
@@ -28,7 +30,7 @@ function setup(t) {
   delete process.env.TEST_CLI_EXIT;
   fs.writeFileSync(counter, "");
   const read = fs.readFileSync, exists = fs.existsSync;
-  const isCredentials = (path) => String(path).endsWith("devin/credentials.toml");
+  const isCredentials = (path) => String(path).replaceAll("\\", "/").endsWith("devin/credentials.toml");
   t.mock.method(fs, "readFileSync", (path, ...args) => isCredentials(path) ? 'api_key = "synthetic-test-key"\n' : read(path, ...args));
   t.mock.method(fs, "existsSync", (path) => isCredentials(path) || exists(path));
   t.mock.method(console, "warn", () => {});
diff --git a/test/cli.test.mjs b/test/cli.test.mjs
new file mode 100644
index 0000000..3d8c633
--- /dev/null
+++ b/test/cli.test.mjs
@@ -0,0 +1,63 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { mkdtempSync, mkdirSync, writeFileSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { dirname, join } from "node:path";
+import { clearDevinBinCache, findDevinBin, whichDevin, runDevin } from "../src/cli.ts";
+
+function fixture(t) {
+  const directory = mkdtempSync(join(tmpdir(), "pi devin cli "));
+  const original = { ...process.env };
+  clearDevinBinCache();
+  t.after(() => {
+    for (const key of ["DEVIN_CLI", "LOCALAPPDATA", "ProgramFiles", "PATH"]) {
+      if (original[key] === undefined) delete process.env[key]; else process.env[key] = original[key];
+    }
+    clearDevinBinCache(); rmSync(directory, { recursive: true, force: true });
+  });
+  return directory;
+}
+
+test("honors an override set after the discovery cache has already been populated", async (t) => {
+  const directory = fixture(t);
+  const first = join(directory, "first"), second = join(directory, "second");
+  writeFileSync(first, ""); writeFileSync(second, "");
+  process.env.DEVIN_CLI = first; assert.equal(findDevinBin(), first);
+  process.env.DEVIN_CLI = second; assert.equal(await whichDevin(), second);
+});
+
+test("runs a native executable directly with argument boundaries intact", async (t) => {
+  fixture(t); process.env.DEVIN_CLI = process.execPath;
+  const value = 'spaces & punctuation "stay literal"';
+  const result = await runDevin(["-e", "process.stdout.write(process.argv[1])", value]);
+  assert.equal(result.code, 0); assert.equal(result.stdout, value);
+});
+
+test("finds the Windows installer binary and ignores a directory override", { skip: process.platform !== "win32" }, (t) => {
+  const directory = fixture(t); process.env.LOCALAPPDATA = directory;
+  process.env.DEVIN_CLI = directory;
+  const binary = join(directory, "devin", "cli", "bin", "devin.exe");
+  mkdirSync(dirname(binary), { recursive: true }); writeFileSync(binary, "");
+  assert.equal(findDevinBin(), binary);
+});
+
+test("Windows PATH lookup uses where.exe and reads multiple results", { skip: process.platform !== "win32" }, async (t) => {
+  const directory = fixture(t); delete process.env.DEVIN_CLI;
+  process.env.LOCALAPPDATA = directory; process.env.ProgramFiles = directory;
+  const bin = join(directory, "path bin"); mkdirSync(bin);
+  const binary = join(bin, "devin.exe"); writeFileSync(binary, "");
+  process.env.PATH = `${bin};${process.env.PATH}`;
+  assert.equal(await whichDevin(), binary);
+});
+
+test("Windows batch wrappers work in paths with spaces, including inherited stdio", { skip: process.platform !== "win32" }, async (t) => {
+  const directory = fixture(t), script = join(directory, "echo.cjs"), wrapper = join(directory, "devin.cmd");
+  writeFileSync(script, 'process.stdout.write(JSON.stringify(process.argv.slice(2)));');
+  writeFileSync(wrapper, `@echo off\r\n"${process.execPath}" "${script}" %*\r\n`);
+  process.env.DEVIN_CLI = wrapper;
+  const result = await runDevin(["models", "list", "--format", "json", "a b"]);
+  assert.equal(result.code, 0, result.stderr);
+  assert.deepEqual(JSON.parse(result.stdout), ["models", "list", "--format", "json", "a b"]);
+  assert.equal((await runDevin(["version"], { inheritStdio: true })).code, 0);
+  await assert.rejects(runDevin(["%PATH%"]), /Unsupported shell characters/);
+});