ci: run the Go agent E2E suite on Windows, and fail if it skips

.gitattributes

@@ -9,3 +9,8 @@
 # This is the mirror image of tailcat's own .gitattributes, which pins
 # build-tags.txt to eol=lf for the same class of reason.
 *.patch -text
+
+# Same class of bug for the shell scripts CI runs under Git Bash on Windows
+# (e2e/require-agent-e2e-ran.sh in windows-e2e): an autocrlf=true checkout
+# makes them CRLF, and bash then reads every line with a trailing \r.
+*.sh text eol=lf

.github/workflows/ci.yml

@@ -97,13 +97,14 @@ jobs:
 
       # Run the agent E2E suite only after the real binaries exist; the
       # earlier go test step intentionally runs before ./build.sh and those
-      # tests skip when dist/ is absent.
+      # tests skip when dist/ is absent. The wrapper fails the step if any
+      # of them still skipped for a missing binary, or none ran at all.
       - name: End-to-end test the agent against the binaries just built
         env:
           GOTOOLCHAIN: local
           MEOWSHELL: ${{ github.workspace }}/dist/meowshell_linux_amd64
           TAILCAT: ${{ github.workspace }}/dist/tailcat_linux_amd64
-        run: go test ./cmd/meowshell/ -count=1
+        run: ./e2e/require-agent-e2e-ran.sh go test -v ./cmd/meowshell/ -count=1
 
       - name: Scan the patched tailcat build input for known vulnerabilities
         env:
@@ -244,7 +245,9 @@ jobs:
   windows-e2e:
     needs: build
     runs-on: windows-latest
-    timeout-minutes: 15
+    # Was 15 before this job also ran the Go agent E2E suite, which recompiles
+    # and reruns cmd/meowshell's tests on top of the vet/test step.
+    timeout-minutes: 25
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
@@ -287,7 +290,8 @@ jobs:
       # validation, N9) only builds and runs under GOOS=windows, so it never
       # gets exercised by the "build" job's own go vet/go test, which run on
       # ubuntu-latest -- this is the only place in CI a real Windows host
-      # actually type-checks and runs that code.
+      # actually type-checks and runs that code. The agent E2E tests skip
+      # here, since the binaries are not fetched yet; they run further down.
       - name: Vet and test meowshell on Windows
         shell: bash
         env:
@@ -324,6 +328,24 @@ jobs:
           name: tailcat-binaries
           path: ${{ runner.temp }}/meowshell-dist
 
+      # The same agent E2E suite the "build" job runs after ./build.sh, here
+      # against the Windows binaries it built. Until this step existed the
+      # suite only ever ran in the vet/test step above, with no binaries and
+      # a fallback that only knew linux_amd64, so all of it skipped and the
+      # job went green. MEOWSHELL/TAILCAT stay explicit so a broken binary
+      # fails rather than skips; the wrapper fails the step if anything
+      # still skipped for a missing binary. Before the DERP relay step on
+      # purpose: that exports TAILCAT_DERPMAP_URL for later steps, and these
+      # tests run their servers in TS_DEBUG_TAILCAT_LOCAL_DERP mode instead,
+      # exactly as they do in the "build" job, which never sets it.
+      - name: End-to-end test the agent against the Windows binaries
+        shell: bash
+        env:
+          GOTOOLCHAIN: local
+          MEOWSHELL: ${{ runner.temp }}\meowshell-dist\meowshell_windows_amd64.exe
+          TAILCAT: ${{ runner.temp }}\meowshell-dist\tailcat_windows_amd64.exe
+        run: ./e2e/require-agent-e2e-ran.sh go test -v ./cmd/meowshell/ -count=1
+
       - name: Start a local DERP relay for the E2E tests
         shell: pwsh
         run: ./e2e/start-testderp.ps1

CLAUDE.md

@@ -51,13 +51,19 @@ fail once before believing it.
 Both E2E suites are wired into CI, but only just: the `dotnet` job sets
 `DOTNET_E2E_*`, and the `build` job runs the Go agent E2E tests in a step
 *after* `./build.sh`, because the `go test ./...` step before it has no `dist/`
-yet and skipped all 22 for as long as they existed. Those two steps are the
-only reason any of it runs — `findE2EBinary`'s `../../dist` fallback skips on a
-missing file, while an explicit `$MEOWSHELL`/`$TAILCAT` is returned unchecked,
-so keep setting them and a broken binary fails loudly instead of going quiet.
-Windows is still dark: `windows-e2e` also runs `go test` before fetching the
-artifact, and the dist name `findE2EBinary` falls back to is hardcoded
-`linux_amd64`.
+yet and skipped all 22 for as long as they existed. `windows-e2e` does the
+same against the Windows binaries, in a step after it downloads the build
+artifact (its own earlier `go test` step skips them too). Those steps are the
+only reason any of it runs — `findE2EBinary`'s `../../dist` fallback (named
+for the host's GOOS/GOARCH as `./build.sh` names it, `.exe` on Windows) skips
+on a missing file, while an explicit `$MEOWSHELL`/`$TAILCAT` is returned
+unchecked, so keep setting them and a broken binary fails loudly instead of
+going quiet. Both Go agent E2E steps run through `e2e/require-agent-e2e-ran.sh`,
+which fails the step if any test skipped for a missing binary or none resolved
+one; it needs `go test -v`, since it greps the markers `findE2EBinary` logs.
+The .NET E2E suite has no such guard and runs on Linux only: there is no .NET
+job on Windows, and `FindRealBinaries` also no-ops when a `DOTNET_E2E_*` path
+does not exist, so a wrong path there still goes quiet.
 
 ## E2E tests and the DERP relay
 

cmd/meowshell/agent_auth_e2e_test.go

@@ -79,7 +79,7 @@ func acceptHostKeyPrompt(t *testing.T, stdin *os.File, out *bufio.Reader) {
 }
 
 func TestAgentPasswordAuth(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	addr, _ := startAuthTestSSHServer(t, func(cfg *ssh.ServerConfig) {
 		cfg.PasswordCallback = func(conn ssh.ConnMetadata, password []byte) (*ssh.Permissions, error) {
@@ -127,7 +127,7 @@ func TestAgentPasswordAuth(t *testing.T) {
 }
 
 func TestAgentSuppliedPrivateKeyAuth(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 	privatePEM, publicKey := newTestKeyPair(t)
 
 	addr, _ := startAuthTestSSHServer(t, func(cfg *ssh.ServerConfig) {
@@ -189,7 +189,7 @@ func driveKeystoreAuth(t *testing.T, stdin *os.File, out *bufio.Reader, signer s
 }
 
 func TestAgentKeystoreKeyAuth(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	// An RSA key is the case the key-type-as-signature-algorithm shortcut got
 	// wrong: ssh-rsa means SHA-1, which every current server refuses. The server
@@ -312,7 +312,7 @@ func newEncryptedTestKeyPair(t *testing.T, passphrase string) (encryptedPEM []by
 // complete. mustReadFrame's 30s deadline (see readFrameWithDeadline) turns a
 // regression here into a clean failure instead of a hung test.
 func TestAgentEncryptedSuppliedPrivateKeyAuth(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 	const passphrase = "correct-passphrase"
 	encryptedPEM, publicKey := newEncryptedTestKeyPair(t, passphrase)
 

cmd/meowshell/agent_e2e_test.go

@@ -9,13 +9,14 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"testing"
 	"time"
 )
 
 func TestAgentEndToEnd(t *testing.T) {
-	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat_linux_amd64")
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	home := t.TempDir()
 	t.Setenv("XDG_CONFIG_HOME", filepath.Join(home, "config"))
@@ -61,7 +62,15 @@ func TestAgentEndToEnd(t *testing.T) {
 	})
 
 	t.Run("exec channel with a nonzero exit reports it as a structured value", func(t *testing.T) {
-		send(t, stdin, 0, controlMessage{Msg: "open_channel", Kind: "exec", Command: []string{"sh", "-c", "'exit 42'"}})
+		// The server runs an exec command through its user's shell: sh on
+		// Unix, but `pwsh -Command` on Windows, which turns a failing native
+		// command's exit code into 1 -- so `sh -c 'exit 42'` read back 1
+		// there. `exit 42` is what each shell itself exits 42 on.
+		command := []string{"sh", "-c", "'exit 42'"}
+		if runtime.GOOS == "windows" {
+			command = []string{"exit", "42"}
+		}
+		send(t, stdin, 0, controlMessage{Msg: "open_channel", Kind: "exec", Command: command})
 		id := expectChannelOpened(t, out)
 
 		exitCode := readExitOnly(t, out, id)
@@ -77,34 +86,92 @@ func TestAgentEndToEnd(t *testing.T) {
 
 		send(t, stdin, id, controlMessage{Msg: "resize", Cols: 120, Rows: 40})
 
-		mustWriteFrame(t, stdin, frame{Type: frameTypeData, ChannelID: id, Payload: []byte("echo shell-marker-e2e\n")})
+		// Enter is \r, as a real terminal (and the app's xterm.js) sends it.
+		// A Unix pty turns it into \n, which is why \n used to work here; the
+		// Windows server's ConPTY does not, and PowerShell took \n as
+		// Ctrl+Enter -- a new line, never a submitted command -- so "exit"
+		// sat unrun until the frame read timed out.
+		mustWriteFrame(t, stdin, frame{Type: frameTypeData, ChannelID: id, Payload: []byte("echo shell-marker-e2e\r")})
 
 		got := readUntil(t, out, id, "shell-marker-e2e", 20*time.Second)
 		if !bytes.Contains(got, []byte("shell-marker-e2e")) {
 			t.Errorf("shell output = %q, want it to contain the echoed marker", got)
 		}
 
-		mustWriteFrame(t, stdin, frame{Type: frameTypeData, ChannelID: id, Payload: []byte("exit\n")})
+		mustWriteFrame(t, stdin, frame{Type: frameTypeData, ChannelID: id, Payload: []byte("exit\r")})
 		readUntilExit(t, out, id)
 	})
 }
 
-func findE2EBinary(t *testing.T, envVar, distName string) string {
+// CI greps the `go test -v` log of every step meant to run these tests for
+// both markers (e2e/require-agent-e2e-ran.sh): any "missing" line, or no
+// "resolved" line at all, fails the step. Without that, a job whose binaries
+// were never wired up passes green with every E2E test skipped -- which is
+// exactly how windows-e2e ran none of them for as long as they existed.
+const (
+	e2eBinaryMissingMarker  = "E2E binary missing:"
+	e2eBinaryResolvedMarker = "E2E binary resolved:"
+)
+
+// findE2EBinary returns the real <name> binary for this test: $envVar if set,
+// else the one ./build.sh wrote for this host under ../../dist.
+//
+// An explicit $envVar is returned unchecked on purpose: CI always sets it, so
+// a missing or broken binary fails the test loudly instead of skipping.
+func findE2EBinary(t *testing.T, envVar, name string) string {
 	t.Helper()
 	if p := os.Getenv(envVar); p != "" {
+		t.Logf("%s $%s=%s", e2eBinaryResolvedMarker, envVar, p)
 		return p
 	}
+	distName := e2eDistName(name, runtime.GOOS, runtime.GOARCH)
 	p := filepath.Join("..", "..", "dist", distName)
 	if _, err := os.Stat(p); err != nil {
-		t.Skipf("no %s (looked for $%s and %s); build.sh must run first", distName, envVar, p)
+		t.Skipf("%s no %s (looked for $%s and %s); build.sh must run first", e2eBinaryMissingMarker, distName, envVar, p)
 	}
 	abs, err := filepath.Abs(p)
 	if err != nil {
 		t.Fatal(err)
 	}
+	t.Logf("%s %s", e2eBinaryResolvedMarker, abs)
 	return abs
 }
 
+// e2eDistName is the file name ./build.sh's build() gives <name> for
+// goos/goarch. This used to be the literal "..._linux_amd64" at every call
+// site, so on any other host -- Windows above all, where the file also ends
+// in .exe -- the fallback looked for a binary that was never there and every
+// E2E test skipped. build.sh only builds arm as GOARM=7, hence "armv7".
+func e2eDistName(name, goos, goarch string) string {
+	s := name + "_" + goos + "_" + goarch
+	if goarch == "arm" {
+		s += "v7"
+	}
+	if goos == "windows" {
+		s += ".exe"
+	}
+	return s
+}
+
+// The wanted names are copied from what ./build.sh writes (its build() and
+// targets_for), not derived from e2eDistName's own logic: the point is that
+// the fallback finds build.sh's output on the host running the tests.
+func TestE2EDistNameMatchesBuildSh(t *testing.T) {
+	for _, tc := range []struct{ name, goos, goarch, want string }{
+		{"meowshell", "linux", "amd64", "meowshell_linux_amd64"},
+		{"tailcat", "linux", "arm64", "tailcat_linux_arm64"},
+		{"meowshell", "linux", "arm", "meowshell_linux_armv7"},
+		{"tailcat", "linux", "386", "tailcat_linux_386"},
+		{"meowshell", "windows", "amd64", "meowshell_windows_amd64.exe"},
+		{"tailcat", "windows", "arm64", "tailcat_windows_arm64.exe"},
+		{"meowshell", "android", "arm64", "meowshell_android_arm64"},
+	} {
+		if got := e2eDistName(tc.name, tc.goos, tc.goarch); got != tc.want {
+			t.Errorf("e2eDistName(%q, %q, %q) = %q, want %q", tc.name, tc.goos, tc.goarch, got, tc.want)
+		}
+	}
+}
+
 func startE2EServer(t *testing.T, tailcatBin, meowshellBin, home string) string {
 	t.Helper()
 	addrFile := filepath.Join(home, "addr")

cmd/meowshell/agent_forward_e2e_test.go

@@ -10,7 +10,7 @@ import (
 )
 
 func TestAgentLocalForwardEndToEnd(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	backendLn, err := net.Listen("tcp", "127.0.0.1:0")
 	if err != nil {
@@ -65,7 +65,7 @@ func TestAgentLocalForwardEndToEnd(t *testing.T) {
 }
 
 func TestAgentSOCKSForwardEndToEnd(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	backendLn, err := net.Listen("tcp", "127.0.0.1:0")
 	if err != nil {

cmd/meowshell/agent_health_e2e_test.go

@@ -44,8 +44,8 @@ import (
 //     legitimately report while it's still settling in a sandboxed test
 //     environment.
 func TestAgentRelayHealthReporterAttachesToARealTailcatConnection(t *testing.T) {
-	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat_linux_amd64")
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	home := t.TempDir()
 	t.Setenv("XDG_CONFIG_HOME", filepath.Join(home, "config"))

cmd/meowshell/agent_security_e2e_test.go

@@ -11,7 +11,7 @@ import (
 )
 
 func TestAgentRejectsNonLoopbackBindByDefault(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	addr, _, _ := startTestSSHServer(t, echoCommandHandler)
 	knownHosts := filepath.Join(t.TempDir(), "known_hosts")
@@ -63,7 +63,7 @@ func TestAgentUnixSocketForward(t *testing.T) {
 	if os.PathSeparator == '\\' {
 		t.Skip("unix domain sockets aren't this test's concern on Windows")
 	}
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	backendLn, err := net.Listen("tcp", "127.0.0.1:0")
 	if err != nil {
@@ -122,7 +122,7 @@ func TestAgentUnixSocketForward(t *testing.T) {
 }
 
 func TestAgentSocksAuthToken(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	backendLn, err := net.Listen("tcp", "127.0.0.1:0")
 	if err != nil {
@@ -207,7 +207,7 @@ func TestAgentSocksAuthToken(t *testing.T) {
 }
 
 func TestAgentConfigureCarriesProxyURL(t *testing.T) {
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 	addr, _, _ := startTestSSHServer(t, echoCommandHandler)
 
 	proxyLn, err := net.Listen("tcp", "127.0.0.1:0")

cmd/meowshell/agent_sftp_e2e_test.go

@@ -7,13 +7,30 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"testing"
 	"time"
 )
 
+// servedFileMode is the permission bits a server on this platform reports
+// for a file chmod-ed to mode. Windows has no Unix permissions: Go's
+// os.Chmod there sets or clears only the read-only attribute, from the owner
+// write bit, and os.Stat reports 0666 or 0444. Expecting 0640 back failed
+// the first time these tests ran on Windows; the exact bits are the Linux
+// job's to check.
+func servedFileMode(mode uint32) uint32 {
+	if runtime.GOOS != "windows" {
+		return mode
+	}
+	if mode&0o200 != 0 {
+		return 0o666
+	}
+	return 0o444
+}
+
 func TestAgentSFTPEndToEnd(t *testing.T) {
-	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat_linux_amd64")
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	home := t.TempDir()
 	t.Setenv("XDG_CONFIG_HOME", filepath.Join(home, "config"))
@@ -46,8 +63,8 @@ func TestAgentSFTPEndToEnd(t *testing.T) {
 
 		sftpOp(t, stdin, out, controlMessage{Op: "chmod", Path: "adir/file.txt", Mode: 0o640})
 		stat = sftpOp(t, stdin, out, controlMessage{Op: "stat", Path: "adir/file.txt"})
-		if got := stat.Entries[0].Mode & 0o777; got != 0o640 {
-			t.Errorf("mode after chmod = %o, want 0640", got)
+		if got, want := stat.Entries[0].Mode&0o777, servedFileMode(0o640); got != want {
+			t.Errorf("mode after chmod = %o, want %o", got, want)
 		}
 
 		sftpOp(t, stdin, out, controlMessage{Op: "rename", Path: "adir/file.txt", NewPath: "adir/renamed.txt"})
@@ -71,8 +88,8 @@ func TestAgentSFTPEndToEnd(t *testing.T) {
 		if len(stat.Entries) != 1 {
 			t.Fatalf("stat = %+v", stat.Entries)
 		}
-		if got := stat.Entries[0].Mode & 0o777; got != 0o600 {
-			t.Errorf("preserved mode = %o, want 0600", got)
+		if got, want := stat.Entries[0].Mode&0o777, servedFileMode(0o600); got != want {
+			t.Errorf("preserved mode = %o, want %o", got, want)
 		}
 		if got := stat.Entries[0].ModTime; got != mtime.Unix() {
 			t.Errorf("preserved mtime = %d, want %d", got, mtime.Unix())

cmd/meowshell/agent_tailcat_forward_e2e_test.go

@@ -13,8 +13,8 @@ import (
 )
 
 func TestAgentForwardsThroughTailcatDestination(t *testing.T) {
-	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat_linux_amd64")
-	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell_linux_amd64")
+	tailcatBin := findE2EBinary(t, "TAILCAT", "tailcat")
+	meowshellBin := findE2EBinary(t, "MEOWSHELL", "meowshell")
 
 	home := t.TempDir()
 	t.Setenv("XDG_CONFIG_HOME", filepath.Join(home, "config"))