[opticon454]

What

A bare npx vitest run writes test fixtures into the developer's real ~/.codeman.

Why

There is no root vitest config: the configs live in config/ and npm test passes --config. A bare invocation therefore has no setupFiles, test/setup.ts never runs, HOME is not redirected and the suite works against the real data dir. On a machine that runs Codeman this left a clis.json containing a test entry, a fixture users.json, webhook.json, intents.json, .env, settings.json values and a .secret-fixture file; the running instance loaded them on its next restart, which looks exactly like "my settings were reset".

How

• A root vitest.config.ts re-exports config/vitest.ci.config.ts, so the default invocation is the isolated one. npm test and the browser, mobile and perf scripts keep passing their own --config.
• The shared temp-HOME teardown keeps removing the folder (up to ~5 s) until a straggling writer goes quiet, and leaves it to the exit backstop instead of failing a file that passed. On a machine with a real agent CLI installed, a test's code path probes or spawns claude, which flushes its own ~/.claude.json and ~/.claude/backups/ into the temp HOME after the test ends; rmSync then raced it and failed the whole file with ENOTEMPTY after every test had passed (about one run in three for session-routes-claude-model on unmodified master; CI has no such CLI). Twelve consecutive runs of that file now pass.
• Docs: CLAUDE.md and Contributing state that a bare npx vitest run is isolated now, and what it used to do.

Tests

test/test-env-isolation.test.ts pins that the root config exists and loads test/setup.ts. Checked by running a test that used to write into the data dir with a bare npx vitest run and confirming nothing appeared there. Changeset included.

Fixes #640

[Ark0N]

Thanks for this, @opticon454. It closes a nasty trap: a bare npx vitest run used to skip test/setup.ts and write fixtures into the real ~/.codeman. The root vitest.config.ts re-export fixes that, and the stricter deletion guards close the hole that let a run without setup delete real state. I checked vitest's config lookup (it walks up from cwd), ran a bare npx vitest run from the repo root and from test/ (both land in the sandboxed HOME), and confirmed packages/xterm-zerolag-input still picks up its own config. CI is green.

A few small things I will apply at merge time, so no action is needed unless you want to do them yourself:

1. test/setup.ts:130: the retry loop has no wall-clock bound, and afterAll runs under vitest's default 10 s hookTimeout. Each failed attempt is the 250 ms wait plus rmSync's own retry sleeps (50 + 100 + 150 ms), so 20 attempts can run past 10 s. A temporary test that left a writer filling $HOME/.claude/backups for 20 s failed with Hook timed out in 10000ms after its test passed, so the "leave it to the exit backstop" branch is never reached. I will bound the loop with a deadline (around 5 s) and keep the comment's claim in line with it.
2. CLAUDE.md:338 and docs/architecture-invariants.md:1175 still say a raw npx vitest skips setup.ts, which now contradicts your updated note at CLAUDE.md:334. I will reword both.
3. docs/architecture-invariants.md:1173 still describes safeRmHomeTree as refusing only paths outside HOME. I will add the sandbox-HOME requirement, the route-harness assert and the 2026-10-10 incident there.
4. CLAUDE.md:9: the preamble lists every file allowed at the repo root and why. I will add vitest.config.ts (vitest resolves its config by walking up from cwd) so a future tidy-up does not move it into config/ and reopen Bare npx vitest run writes test fixtures into the real ~/.codeman #640.

Smaller wording items: the assertSandboxedHome JSDoc (test/mocks/test-helpers.ts:64) says every route test refuses to start, but only the 39 route files on createRouteTestHarness do (the 30 that build their own app from createMockRouteContext are not covered), and the changeset does not mention the stricter guards. I will adjust both.

Thanks again for tracking down both the root cause and the ENOTEMPTY flake.

[Ark0N]

Merged and shipped in 1.44.0. Thank you @opticon454! This closes a trap that had already cost real state once. The release gate even proved the point: the one file that failed on a branch without your change failed on exactly the ENOTEMPTY teardown flake this PR fixes.

At merge I bounded the teardown retry by a deadline of about 5 s instead of an attempt count, so a stuck writer can no longer push afterAll past vitest's 10 s hook timeout (a temporary test with a writer running for 20 s failed with a hook timeout before and passes in under 6 s now). I also brought CLAUDE.md and the invariants in line with your note: a bare npx vitest is isolated now, vitest.config.ts is listed among the files that must stay at the repo root, and the deletion guards and the 2026-10-10 incident are described. The assertSandboxedHome JSDoc now says which route tests it covers. Closes #640.