Port the kata daemon to Windows by njt · Pull Request #75 · kenn-io/kata

njt · 2026-05-31T10:03:11Z

Ports the kata daemon to Windows. go build, go vet, go test, and go mod tidy are clean on Windows and unchanged on Unix.

Supersedes #73 / #74 (the 0xc0000142 detach fix is folded in as its own commit).

What's in it

Daemon port — platform-specific pieces behind build tags:

Signaling — Windows has no cross-process SIGTERM/SIGHUP, so stop/reload use per-daemon named events (Local\kata-stop|reload-<dbhash>-<pid>) set by daemon stop/reload. Unix keeps SIGTERM/SIGHUP.
ProcessAlive — Windows uses OpenProcess + GetExitCodeProcess(STILL_ACTIVE); Unix keeps kill(0).
Namespace — no uid on Windows, so the per-user socket-dir tag derives from a sanitized USERNAME.

Console detach fix (the original bug) — the daemon runs git remote to resolve project identity. On Windows the auto-started daemon shared the caller's console; once that shell exited, git aborted at startup with STATUS_DLL_INIT_FAILED (exit status 0xc0000142). Now the daemon is created with DETACHED_PROCESS | CREATE_NEW_PROCESS_GROUP and routes stdout/stderr to daemon.log instead of the caller's stderr (which a surviving detached daemon would hold open, hanging output-capturing parents).

Test portability — stock main had ~25 Windows failures in a Unix-first suite:

token_identity e2e test tagged !windows like its siblings.
hookprobe helper built as hookprobe.exe; hook-config tests use bare-name / per-OS absolute commands.
Unix 0600 file-mode assertions skipped on Windows (ACL-based there).
shouldRefuseAutoStartDaemon normalized via filepath.ToSlash (a real bug: it looked for \go-build on Windows).
agent-output tests assert through agentValue; beads-from-live-bd tests skip on Windows; stop/reload tests stand up the named events a real daemon would.

CI — adds a windows job to .github/workflows/test.yml running go build, go vet, and go test on windows-latest. The existing ubuntu lanes never compile the *_windows.go files or !windows-tagged tests, so this is the regression gate for the port. No cgo, so no extra toolchain setup.

roborev-ci · 2026-05-31T10:06:58Z

roborev: Combined Review (`d4e05cb`)

High severity finding: Windows build still fails due to an unsupported signal constant.

High

cmd/kata/daemon_signaling_windows.go:97
The Windows-only file sends syscall.SIGHUP, but Go’s Windows syscall package does not define SIGHUP, so the Windows build still fails. Use a synthetic os.Signal value that exists on Windows, or define a small local signal type; runReloadLoop only needs a receive event and does not inspect the signal value.

Synthesized from 2 reviews (agents: codex | types: default, security)

wesm · 2026-05-31T13:25:40Z

Thanks. If you don't mind I can take it from here and I'll add CI so this doesn't regress?

Squashed Windows port branch. Additional commits folded in: - fix(daemon): detach auto-started daemon from caller console/stdio on Windows - fix(windows): tag token_identity e2e test !windows like its siblings - test(windows): stand up named signal events for faked daemons in stop/reload tests - fix(windows): make the test suite pass on Windows

Guard the Windows daemon port against regressions. The _windows.go files and !windows build tags are never compiled by the existing ubuntu-latest lanes, so breakage on Windows is invisible there. The new job runs go build, go vet, and go test -p 1 ./... on windows-latest. The project has no cgo, so unlike msgvault it needs no MSYS2 or toolchain setup. vet runs explicitly because the GOOS=linux golangci-lint lane never sees the Windows-tagged files.

The path is built from the daemon's own data dir and the fixed "daemon.log" constant, not user input. Annotate //nolint:gosec so the make-lint pre-commit hook (and CI lint lane) pass.

roborev-ci · 2026-05-31T23:39:07Z

roborev: Combined Review (`42a1fcd`)

Medium issue found; security review found no additional findings.

Medium

internal/client/ensure.go:152 - The automatic restart path for incompatible daemons still uses p.Signal(syscall.SIGTERM). Since the PR establishes that Windows requires the named stop event, EnsureRunning can detect an old live daemon on Windows, fail to stop it, wait 3 seconds, and return old daemon did not stop within 3s instead of replacing it.

Suggested fix: Move daemon stop signaling into shared platform-specific code, or add a Windows-specific stopRunningDaemons path that sets the same Local\kata-stop-<dbhash>-<pid> event, deriving the DB hash from the runtime record/config.

Synthesized from 2 reviews (agents: codex | types: default, security)

Git for Windows checks out text files as CRLF (core.autocrlf=true by default). The internal/tui snapshot tests compare golden files byte-for-byte against LF-rendered output, so on windows-latest all ~45 TestSnapshot_* cases failed with invisible want/got diffs. A repo-wide .gitattributes pins text files to LF on checkout for every platform.

TestClaimHubClientUsesUnixRuntimeForKataInvalid hardcoded os.MkdirTemp("/tmp", ...); on the GitHub windows-latest runner /tmp resolves to a nonexistent C:\tmp, failing with GetFileAttributesEx before the test ever reached its AF_UNIX listener. The test exercises the unix:// runtime-file fallback, which is Unix-only, so guard it with the same runtime.GOOS check its endpoint_test.go siblings already use. The file's other (platform-independent) test keeps running on Windows.

roborev-ci · 2026-06-01T00:45:52Z

roborev: Combined Review (`0f88398`)

Windows support is not ready to merge: one High compile break and one Medium restart-path bug remain.

High

cmd/kata/daemon_signaling_windows.go:88: The Windows-only reload event path sends syscall.SIGHUP, but SIGHUP is not available in the Windows syscall API, so the new windows-latest build will fail to compile. Use a cross-platform sentinel signal value such as os.Interrupt, since runReloadLoop ignores the value, or define a small private os.Signal implementation for synthetic reload events.

Medium

internal/client/ensure.go:137: The auto-start compatibility path still stops incompatible daemons with p.Signal(syscall.SIGTERM). The new Windows stop mechanism uses named events because SIGTERM cannot be delivered cross-process there, so a Windows client that discovers an incompatible daemon will fail to stop it and restart cleanly. Move platform-specific stop signaling into a shared internal helper and use it from both kata daemon stop and stopRunningDaemons.

Synthesized from 2 reviews (agents: codex | types: default, security)

vercel · 2026-06-01T01:23:48Z

@wesm is attempting to deploy a commit to the Kenn Software Team on Vercel.

A member of the Team first needs to authorize it.

roborev-ci · 2026-06-01T01:45:09Z

roborev: Combined Review (`405f282`)

Medium risk: one lifecycle race can block daemon upgrades on Windows or during normal shutdown.

Medium

internal/client/ensure.go:149
stopRunningDaemons treats any SignalDaemonStop failure as fatal after a prior liveness/probe check. The daemon can exit between the probe and the signal/open-event call, leaving a stale runtime record; on Windows the named event can also disappear during shutdown. In that case, an upgrade path that should treat the old daemon as already stopped instead fails and refuses to start the replacement.

Suggested fix: After a signal error, re-check the runtime record/process or probe result and ignore/clean up the record when the daemon is no longer live. Only return an error when the daemon is still confirmed running but cannot be signaled.

Synthesized from 2 reviews (agents: codex | types: default, security)

This reverts commit ae35f50.

roborev-ci · 2026-06-01T02:46:54Z

roborev: Combined Review (`aa985bc`)

High severity issue: Windows CI will fail to compile due to a Unix-only signal in a shared test.

High

cmd/kata/daemon_reload_loop_test.go:63
The new Windows CI job will compile this test package on Windows, but the test still sends syscall.SIGHUP, which is Unix-only. As a result, go test ./... on windows-latest will fail at compile time.
Fix: Use a platform-independent os.Signal test value, or send any cross-platform signal-like value, since runReloadLoop only reacts to channel delivery and does not inspect the signal.

Synthesized from 2 reviews (agents: codex | types: default, security)

njt mentioned this pull request May 31, 2026

fix(daemon): detach auto-started daemon from caller console/stdio on Windows #74

Closed

Nat Torkington and others added 3 commits May 31, 2026 18:15

fix(lint): silence gosec G304 on the daemon.log writer

42a1fcd

The path is built from the daemon's own data dir and the fixed "daemon.log" constant, not user input. Annotate //nolint:gosec so the make-lint pre-commit hook (and CI lint lane) pass.

wesm force-pushed the windows-port branch from d4e05cb to 42a1fcd Compare May 31, 2026 23:35

wesm added 2 commits May 31, 2026 19:23

wesm added 2 commits May 31, 2026 20:05

fix(windows): share daemon stop signaling

13e7f75

test(windows): skip DB close checkpoints in test harness

8ff5550

test(windows): use fast SQLite sync in test harness

405f282

wesm added 3 commits May 31, 2026 20:46

ci(windows): skip duplicate go test vet

3314b81

test(windows): keep daemon harness DBs in memory

ae35f50

Revert "test(windows): keep daemon harness DBs in memory"

aa985bc

This reverts commit ae35f50.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Port the kata daemon to Windows#75

Port the kata daemon to Windows#75
njt wants to merge 11 commits into
kenn-io:mainfrom
njt:windows-port

njt commented May 31, 2026 •

edited by wesm

Loading

Uh oh!

roborev-ci Bot commented May 31, 2026

Uh oh!

wesm commented May 31, 2026

Uh oh!

roborev-ci Bot commented May 31, 2026

Uh oh!

roborev-ci Bot commented Jun 1, 2026

Uh oh!

vercel Bot commented Jun 1, 2026

Uh oh!

roborev-ci Bot commented Jun 1, 2026

Uh oh!

roborev-ci Bot commented Jun 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

njt commented May 31, 2026 • edited by wesm Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What's in it

Uh oh!

roborev-ci Bot commented May 31, 2026

roborev: Combined Review (d4e05cb)

High

Uh oh!

wesm commented May 31, 2026

Uh oh!

roborev-ci Bot commented May 31, 2026

roborev: Combined Review (42a1fcd)

Medium

Uh oh!

roborev-ci Bot commented Jun 1, 2026

roborev: Combined Review (0f88398)

High

Medium

Uh oh!

vercel Bot commented Jun 1, 2026

Uh oh!

roborev-ci Bot commented Jun 1, 2026

roborev: Combined Review (405f282)

Medium

Uh oh!

roborev-ci Bot commented Jun 1, 2026

roborev: Combined Review (aa985bc)

High

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

njt commented May 31, 2026 •

edited by wesm

Loading

roborev: Combined Review (`d4e05cb`)

roborev: Combined Review (`42a1fcd`)

roborev: Combined Review (`0f88398`)

roborev: Combined Review (`405f282`)

roborev: Combined Review (`aa985bc`)