fix(ci): restore green Windows CI and make timed-out suites diagnosable

packages/coding-agent/test/bash-session-metadata.test.ts

@@ -45,7 +45,17 @@ async function createSession(options: { persisted?: boolean; stage?: string } =
 }
 
 afterEach(() => {
-	for (const root of roots.splice(0)) if (existsSync(root)) rmSync(root, { recursive: true, force: true });
+	// Windows locks a directory while a spawned shell still has it as its cwd, so a
+	// detached async job can outlive its test and make cleanup fail with EBUSY.
+	// Retry, then give up: reclaiming an OS temp dir is never the assertion.
+	for (const root of roots.splice(0)) {
+		if (!existsSync(root)) continue;
+		try {
+			rmSync(root, { recursive: true, force: true, maxRetries: 20, retryDelay: 50 });
+		} catch {
+			// Left for the OS to reclaim.
+		}
+	}
 });
 
 describe("session-aware bash environment", () => {

packages/coding-agent/test/footer-width.test.ts

@@ -1,3 +1,4 @@
+import { sep } from "node:path";
 import { visibleWidth } from "@earendil-works/pi-tui";
 import { beforeAll, describe, expect, it } from "vitest";
 import type { AgentSession } from "../src/core/agent-session.ts";
@@ -115,7 +116,7 @@ describe("formatCwdForFooter", () => {
 
 	it("abbreviates the home directory and descendants", () => {
 		expect(formatCwdForFooter("/home/user", "/home/user")).toBe("~");
-		expect(formatCwdForFooter("/home/user/project", "/home/user")).toBe("~/project");
+		expect(formatCwdForFooter("/home/user/project", "/home/user")).toBe(`~${sep}project`);
 	});
 });
 

packages/coding-agent/test/interactive-auth-login.test.ts

@@ -4,6 +4,14 @@ import { LoginDialogComponent } from "../src/modes/interactive/components/login-
 import { InteractiveModeBase } from "../src/modes/interactive/interactive-mode-base.ts";
 import { initTheme } from "../src/modes/interactive/theme/theme.ts";
 import "../src/modes/interactive/interactive-auth-login.ts";
+import { openBrowser } from "../src/utils/open-browser.ts";
+
+// Driving the real LoginDialogComponent through onAuth opens the platform browser.
+// On Windows that spawns a detached `rundll32 url.dll,FileProtocolHandler`, which
+// inherits this process's stdout/stderr write handles; on a runner with no browser
+// it never exits, the suite's pipes never reach EOF, and the CI step hangs until the
+// job budget expires. Mock the launcher, as pi's login-dialog suites do.
+vi.mock("../src/utils/open-browser.ts", () => ({ openBrowser: vi.fn() }));
 
 beforeAll(() => {
 	initTheme("dark");
@@ -153,6 +161,7 @@ describe("interactive OAuth cancellation", () => {
 		const dialog = addedChildren[0] as LoginDialogComponent;
 		expect(dialog.render(100).join("\n")).toContain("Sign in to Corp");
 		expect(completeProviderAuthentication).toHaveBeenCalledOnce();
+		expect(vi.mocked(openBrowser)).toHaveBeenCalledWith("https://corp.invalid/login");
 	}, 1_000);
 
 	it("keeps a post-login refresh AbortError visible", async () => {

packages/coding-agent/test/sdk-session-manager.test.ts

@@ -1,6 +1,6 @@
-import { existsSync, mkdirSync, realpathSync, rmSync } from "node:fs";
+import { existsSync, mkdirSync, rmSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
-import { join } from "node:path";
+import { join, sep } from "node:path";
 import { getModel } from "@earendil-works/pi-ai/compat";
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 import { AuthStorage } from "../src/core/auth-storage.ts";
@@ -92,7 +92,7 @@ describe("createAgentSession session manager defaults", () => {
 		const sessionFile = session.sessionManager.getSessionFile();
 
 		expect(sessionDir).toBe(expectedSessionDir);
-		expect(sessionFile?.startsWith(`${expectedSessionDir}/`)).toBe(true);
+		expect(sessionFile?.startsWith(`${expectedSessionDir}${sep}`)).toBe(true);
 
 		session.dispose();
 	});
@@ -129,17 +129,22 @@ describe("createAgentSession session manager defaults", () => {
 		});
 
 		expect(session.sessionManager).toBe(sessionManager);
-		expect(session.systemPrompt).toContain(`Current working directory: ${sessionCwd}`);
+		// The system prompt always renders the cwd with POSIX separators (system-prompt.ts).
+		expect(session.systemPrompt).toContain(`Current working directory: ${sessionCwd.replaceAll("\\", "/")}`);
 
 		const bashTool = session.agent.state.tools.find((tool) => tool.name === "bash");
 		expect(bashTool).toBeTruthy();
-		const result = await bashTool!.execute("test", { command: "pwd" });
+		// Prove the tool's working directory by reading a marker only resolvable from
+		// sessionCwd. Comparing `pwd` text is not portable: on Windows the bash tool runs
+		// under Git Bash, which reports MSYS paths (/tmp/...) that Node cannot resolve.
+		writeFileSync(join(sessionCwd, "cwd-marker.txt"), "session-cwd-marker");
+		const result = await bashTool!.execute("test", { command: "cat cwd-marker.txt" });
 		const output = result.content
 			.filter((item): item is { type: "text"; text: string } => item.type === "text")
 			.map((item) => item.text)
 			.join("");
 
-		expect(realpathSync(output.trim())).toBe(realpathSync(sessionCwd));
+		expect(output.trim()).toBe("session-cwd-marker");
 
 		session.dispose();
 	});

scripts/run-flaky-test-suite.ts

@@ -33,16 +33,39 @@ function safeName(value: string): string {
   return value.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-|-$/g, "") || "tests";
 }
 
+/** Tee one child stream to the live step log while capturing it for diagnostics. */
+async function pump(stream: ReadableStream<Uint8Array> | undefined, sink: NodeJS.WriteStream): Promise<string> {
+  if (!stream) return "";
+  const decoder = new TextDecoder();
+  let text = "";
+  for await (const chunk of stream) {
+    const piece = decoder.decode(chunk, { stream: true });
+    text += piece;
+    sink.write(piece);
+  }
+  const tail = decoder.decode();
+  if (tail) {
+    text += tail;
+    sink.write(tail);
+  }
+  return text;
+}
+
+/**
+ * Stream child output as it arrives instead of buffering it until the child exits.
+ * A suite that overruns the job's `timeout-minutes` budget is killed mid-run, and a
+ * buffered attempt discards every line it had already collected, leaving the cancelled
+ * step with no record of which test stalled. Teeing preserves the retry and diagnostics
+ * contract while making a timed-out attempt self-describing in the step log.
+ */
 async function runAttempt(command: string[], logPath: string, persist: boolean): Promise<{ code: number; output: string }> {
   const child = Bun.spawn(command, { stdout: "pipe", stderr: "pipe", env: process.env });
   const [stdout, stderr, code] = await Promise.all([
-    new Response(child.stdout).text(),
-    new Response(child.stderr).text(),
+    pump(child.stdout, process.stdout),
+    pump(child.stderr, process.stderr),
     child.exited,
   ]);
   const output = `${stdout}${stderr}`;
-  process.stdout.write(stdout);
-  process.stderr.write(stderr);
   if (persist) writeFileSync(logPath, output);
   return { code, output };
 }