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

[flora131]

Problem

The Tests workflow failed on every run since #2064. The Windows job was killed at its 15-minute timeout-minutes budget, always inside the coding-agent vitest suite step. GitHub reports a timeout-minutes kill as cancelled, which is why this read as noise rather than a red build.

commit	Windows job	vitest step
#2057 fix(workflows)	success (7.5m)	125s
#2059 fix(auth)	success (6.7m)	117s
#2064 ModelRuntime	cancelled (15.1m)	never finished
#2065 fix(copilot)	cancelled (15.1m)	never finished

Root cause: a test launched a real browser on the CI runner

test/interactive-auth-login.test.ts drives a real LoginDialogComponent through the onAuth callback. showAuthUrl calls openBrowser(url), and on Windows that is:

rundll32 url.dll,FileProtocolHandler https://corp.invalid/login

spawned detached. Node's spawn on Windows creates the child with handle inheritance, so rundll32 inherits the test process's stdout/stderr write handles. The runner has no browser to hand the URL to, so rundll32 never exits.

A GitHub Actions run: step ends only when the child exits and every inherited stdout/stderr handle closes. So the whole visible chain finished normally — 2887 tests passed in ~135s, vitest exited, bun run exited, the wrapper's direct child exited — and the step still hung on pipes that could never reach EOF, until the 15-minute budget killed the job.

#2064 is the first bad commit because #2064 added that test case (honors transported callback metadata and resolves manual redirect input). At 86dd67d6c, the last green commit, this file never invoked onAuth.

Evidence

Run 30495631917, Windows job, with a temporary probe that dumped Win32_Process every 30s after the child exited:

22:25:38  [suite-probe child exited code=0; 2 inherited stream(s) still open]
          ProcessId ParentProcessId Name         Started  CommandLine
          2508      4772            rundll32.exe 22:24:23 rundll32 url.dll,FileProtocolHandler https://corp.invalid/login
...
22:33:xx  ##[error]The operation was canceled.

Every 30s sample through cancellation showed the same single non-infrastructure survivor, started mid-suite, with a dead parent. No bun.exe or node.exe descendant survived — the earlier "leaked engine/RPC child" theory is refuted; the two bun.exe and one node.exe seen at teardown in the first report were the live bun run → vitest chain itself.

Fix

Mock the launcher, exactly as pi's own login-dialog suites do (test/suite/regressions/5433-extension-oauth-prompt-input.test.ts):

vi.mock("../src/utils/open-browser.ts", () => ({ openBrowser: vi.fn() }));

and assert the dialog still hands the URL to it, so the launch path stays covered without spawning a process:

expect(vi.mocked(openBrowser)).toHaveBeenCalledWith("https://corp.invalid/login");

That assertion is the regression test: if a future change makes this suite reach the real launcher again, it fails rather than hanging Windows CI.

No shipped behavior changes, so there is no changelog entry. src/utils/open-browser.ts is untouched and still matches upstream pi.

Also in this PR

Three Windows-only test failures from #2064, which dropped platform-aware handling and made the bounded flake retry run the suite twice:

test	#2064 wrote	restored
footer-width — home abbreviation	"~/project"	`~${sep}project`
sdk-session-manager — session file prefix	`${expectedSessionDir}/`	`${expectedSessionDir}${sep}`
sdk-session-manager — prompt cwd	${sessionCwd}	sessionCwd.replaceAll("\\", "/")

Production code was correct throughout: formatCwdForFooter joins with sep, and system-prompt.ts renders cwd with POSIX separators.

scripts/run-flaky-test-suite.ts also buffered the child's entire output via new Response(child.stdout).text() and wrote it only after the child exited, so a child killed by the job timeout discarded every collected line — the cancelled step logged minutes of silence. It now tees both streams as they arrive. Retry, deterministic-file, and diagnostics-artifact behavior are unchanged (test/unit/flaky-test-suite-runner.test.ts passes). That change is what produced the evidence above and keeps the next timeout self-describing.

bash-session-metadata cleanup now retries around transient Windows directory locks.

Verification

• Windows job green with no diagnostic scaffolding: run 30497271702 at 8918590d8 — test (blacksmith-4vcpu-windows-2025, windows-x64) succeeded in 8m30s (22:46:44 → 22:55:14) against the unchanged 15-minute budget. The coding-agent vitest suite step took 160s and ended on its own. Linux green in the same run.
• The temporary probes (packages/coding-agent/test-exit-diagnostic.ts, its globalSetup entry, and the wrapper's process-table dump) are removed; the green run above does not contain them.
• Local: bun run typecheck, bun run lint, bun run check:file-length, bun run test:unit (4406 pass / 0 fail) all green.
• No timeout-minutes was widened, no survivor is force-killed from CI YAML, and no suite is skipped or excluded.

Ruled out during triage

• Not a module-graph regression: local suite duration identical pre/post feat(coding-agent)!: adopt pi's ModelRuntime architecture for full model-provider parity #2064 (31.5s vs 32.0s).
• Not live-network contention: instrumenting globalThis.fetch across all 370 test files recorded exactly one outbound request.
• Not a leaked engine/RPC child and not a vitest event-loop leak: the vitest main process exits, and the post-exit process table names rundll32 as the only survivor.

Greptile Summary

Restores Windows CI reliability and improves diagnostics for timed-out test suites.

• Makes path-sensitive coding-agent assertions portable across Windows and POSIX.
• Prevents browser-launch and temporary-directory cleanup behavior from hanging Windows tests.
• Streams child stdout and stderr live while retaining output for retry diagnostics and artifacts.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the Windows timeout contract issue is fixed because the workflow retains the required 15-minute budget.

T-Rex Logs

What T-Rex did

• Observed the before-state where flaky-streaming-01-before.log recorded a timeout exit 124 with an empty live terminal-output section.
• Validated the after-state showing stdout and stderr markers at 22:49:11.495Z, followed by an external timeout completion at 22:49:14.477Z and exit 124.
• Verified the contract outcome in flaky-runner-contract-02-after.log reports 4 pass, 0 fail, exit 0.
• Noted that the harness source is retained to expose the exact recognizable markers and a persistent child process for review.
• Artifacts were uploaded to support verification of the above assertions.

_{Ran code and verified through T-Rex}

Important Files Changed

Filename	Overview
packages/coding-agent/test/bash-session-metadata.test.ts	Makes Windows temporary-session cleanup resilient to transient directory locks without affecting test assertions.
packages/coding-agent/test/footer-width.test.ts	Restores the platform-aware separator expectation for abbreviated home-directory paths.
packages/coding-agent/test/interactive-auth-login.test.ts	Mocks the platform browser launcher and verifies its URL, preventing detached Windows processes from retaining test output handles.
packages/coding-agent/test/sdk-session-manager.test.ts	Makes session paths and system-prompt expectations portable and verifies the shell working directory through a marker file.
scripts/run-flaky-test-suite.ts	Tees child output into live CI logs while preserving captured text for existing retry and diagnostic-artifact behavior.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  Start["Run test suite attempt"] --> Spawn["Spawn child with stdout/stderr pipes"]
  Spawn --> Tee["Stream both pipes to CI log and capture output"]
  Tee --> Exit{"Attempt succeeds?"}
  Exit -->|Yes| Done["Complete successfully"]
  Exit -->|No| Retry["Run bounded retry"]
  Retry --> TeeRetry["Stream retry output live"]
  TeeRetry --> Persist["Persist diagnostics artifact"]
  Persist --> Result{"Retry succeeds?"}
  Result -->|Yes| Done
  Result -->|No| Fail["Exit non-zero with visible diagnostics"]

Loading

_{Reviews (10): Last reviewed commit: "test: remove the Windows teardown diagno..." | Re-trigger Greptile}

Dismissed because a newer commit was pushed; Greptile will re-review the current head.

Dismissed because a newer commit was pushed; Greptile will re-review the current head.

Dismissed because a newer commit was pushed; Greptile will re-review the current head.