Make ClaudeTextGeneration tests hermetic on Windows

[mihneaptu]

What Changed

The ClaudeTextGeneration test fixture now works on Windows. The fake claude binary is a Node script with a platform-appropriate entry point — claude.cmd on Windows, a #!/bin/sh shim elsewhere — and PATH is built with the platform's separator instead of a hardcoded :.

One file, test-only. No production code changes.

Why

Fixes #4507.

On Windows all 5 tests in this file fail on unmodified main. The fixture never takes effect, so the real installed Claude CLI runs instead: assertions compare fixture strings against live model output, and the run makes real billed API calls.

Two POSIX-only assumptions caused it:

1. PATH was joined with a literal :. Windows uses ;, so the fixture's bin directory never became its own PATH entry.
2. The stub was written as extension-less bin/claude containing a #!/bin/sh script. Windows resolves executables through PATHEXT, so an extension-less file is skipped — and even with a correct extension a shebang script is not executable there.

Fixing only the separator is not enough, since the shebang stub still would not run. So the stub logic moved into Node (claude-stub.mjs) with a thin platform-specific launcher in front of it. The T3_FAKE_CLAUDE_* env contract and every exit code are unchanged.

No production change was needed: resolveSpawnCommand already resolves through PATHEXT and spawns .cmd with shell: true.

Validation

Windows 11, this file only:

	result
before	5 failed, 134.28s (live API calls)
after	5 passed, 4.09s (stub invoked)

The 134s to 4s drop is the signal that no network calls happen anymore.

vp fmt, vp lint --report-unused-disable-directives, git diff --check, and pnpm --filter t3 typecheck are all clean for this scope.

I do not have a POSIX machine, so the shell shim is verified behaviourally under sh rather than by running the suite: happy path, args-must-contain pass and fail, stdin-must-contain pass and fail, custom exit code, and config-dir mismatch all return the same exit codes as the old script (0, 0, 2, 4, 0, 7, 5). The full suite on Linux and macOS is unverified by me and worth confirming in CI.

One note so it is not a surprise: on Windows the run now emits DEP0190 (passing args with shell: true). That comes from the existing resolveSpawnCommand path and is not introduced here — it simply was not reachable from this test before, because the stub never resolved.

Not included

ClaudeTextGeneration.test.ts builds claudeConfigDir from process.cwd(), which leaves a .claude-work-test/ directory behind in the repo after the run. It is recorded in #4507 and left out here to keep this focused.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes — n/a, no UI changes
• I included a video for animation/interaction changes — n/a

---

Note

Low Risk
Changes are confined to one test file; no production code paths are modified.

Overview
Makes ClaudeTextGeneration tests use a portable fake Claude CLI so Windows runs hit the stub instead of the real CLI on PATH (which caused failures, live API calls, and slow runs).

The test fixture’s fake claude binary is no longer a single #!/bin/sh script. Stub logic lives in claude-stub.mjs (same T3_FAKE_CLAUDE_* env checks and exit codes). Windows gets claude.cmd that forwards to Node; Unix keeps a small executable shim that execs the same stub.

withFakeClaudeEnv prepends the fake bin to PATH using ; on Windows and : elsewhere, so the fixture directory is actually discoverable on Windows.

^{Reviewed by Cursor Bugbot for commit 68d1bb1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Make ClaudeTextGeneration tests hermetic on Windows

• Replaces the POSIX shell-based fake claude binary with a Node stub (claude-stub.mjs) that replicates the same validation logic using process APIs.
• On Windows, a claude.cmd batch file invokes the Node stub; on non-Windows, a shell launcher script is written and marked executable.
• Fixes PATH assembly in withFakeClaudeEnv to use ; as the delimiter on Windows instead of :.
• Platform detection uses isHostWindows from @t3tools/shared/hostProcess.

^{Macroscope summarized 68d1bb1.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 405d6947-5dd2-4b3a-9c63-88359a1cd2ff

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[macroscopeapp]

Approvability

Verdict: Approved 68d1bb1

Test-only changes that make the ClaudeTextGeneration tests work on Windows by replacing shell script stubs with cross-platform Node.js implementations. No production code is affected.

^{You can customize Macroscope's approvability policy. Learn more.}

[CDVolvik]

5/5 still green on Linux, so the POSIX path is unaffected by the rewrite. Moving the stub into Node and keeping a thin .cmd entry point is the right shape, and the PATHEXT note in the comment is the part most people miss.

One thing to change while it is being touched. The stub ends with:

process.stdout.write(process.env.T3_FAKE_CLAUDE_OUTPUT ?? "");
process.exit(Number(process.env.T3_FAKE_CLAUDE_EXIT_CODE ?? 0));

process.stdout.write to a pipe is asynchronous, and process.exit does not flush what is still queued. The shell version did not have this problem because printf blocks until the write completes, so this is a behaviour the port introduces rather than inherits.

Measured with the same write-then-exit shape, payload generated inside the child so nothing large goes through the environment:

payload=    1024   process.exit():     1024   |   process.exitCode:     1024
payload=   65536   process.exit():    65536   |   process.exitCode:    65536
payload=  262144   process.exit():   182720  TRUNCATED   |   process.exitCode:   262144
payload= 1048576   process.exit():   146176  TRUNCATED   |   process.exitCode:  1048576
payload= 4194304   process.exit():   146176  TRUNCATED   |   process.exitCode:  4194304

It is latent today rather than broken: T3_FAKE_CLAUDE_OUTPUT arrives through the environment, so current outputs sit under the threshold. But the cutoff lands between 64KB and 256KB, and it depends on pipe buffer state rather than on anything in the test, so the first person who needs a larger fixture gets a partial read with no error and a passing-then-failing test. That is an awkward thing to debug in a file whose stated purpose is hermeticity.

process.exitCode = Number(...) instead of process.exit(...) fixes it and still yields the right exit status, since the stub already drains stdin to EOF and has nothing else keeping the loop alive. The fail() helper has the same shape, though its messages are short enough that it will not bite in practice.

Minor: the new import sits above the @t3tools/contracts line rather than in sorted position, which may or may not matter depending on whether the import ordering rule is enforced here.

Dismissing prior approval to re-evaluate 68d1bb1