fix: random monitoring issues

...	...		@@ -73,6 +73,7 @@ type Cfg struct {
73	73		HumanOutput bool // human-readable output instead of JSONL / slog
74	74		Wait bool // block until job completes, print history and summary
75	75		EnvVars envList // custom environment variables passed via -e / -env
	76	+	SessionPrefix string // session prefix filter for monitor (default: "ci.")
76	77		}
77	78
78	79		type Event struct {

...	...		@@ -93,10 +94,12 @@ func NewCfg() (*Cfg, string, bool) {
93	94		var gcInterval time.Duration
94	95		var logLevel string
95	96		var envVars envList
	97	+	var sessionPrefix string
96	98		flag.StringVar(&keyLoc, "pk", "", "ssh private key used to authenticate with pico services")
97	99		flag.StringVar(&certLoc, "ck", "", "ssh certificate public key used to authenticate with pico services (only required if using ssh certificates)")
98	100		flag.StringVar(&artifactDir, "artifact-dir", "/tmp/pici-artifacts", "local directory to stage artifacts")
99	101		flag.StringVar(&event, "event", "", "event JSON to run (alternative to reading from stdin)")
	102	+	flag.StringVar(&sessionPrefix, "prefix", "ci.", "session prefix filter for monitor (default: ci.)")
100	103		flag.DurationVar(&monitorInterval, "monitor-interval", 5*time.Second, "interval for monitoring zmx sessions")
101	104		flag.DurationVar(&gcInterval, "gc-interval", 10*time.Minute, "interval for garbage collection in monitor (0 to disable)")
102	105		flag.StringVar(&logLevel, "log-level", "info", "log level: debug, info, warn, error")

...	...		@@ -134,6 +137,7 @@ func NewCfg() (*Cfg, string, bool) {
134	137		HumanOutput: human,
135	138		Wait: wait,
136	139		EnvVars: envVars,
	140	+	SessionPrefix: sessionPrefix,
137	141		}, cmd, wantHelp
138	142		}
139	143

...	...		@@ -614,8 +618,8 @@ func (eng *JobEngine) Setup() error {
614	618
615	619		func (eng *JobEngine) Run(manifest string) error {
616	620		domain := "ci"
617		-	if eng.Ev != nil && eng.Ev.Type == "local" {
618		-	domain = "local"
	621	+	if eng.Ev != nil && eng.Ev.Type != "" {
	622	+	domain = eng.Ev.Type
619	623		}
620	624		prefix := fmt.Sprintf("%s.%s.%s.", domain, eng.Ev.Name, eng.JobID)
621	625		// Child sessions use ".step." sub-prefix so zmx wait "*" inside pico.sh

...	...		@@ -693,8 +697,8 @@ func eventHandler(cfg Cfg, eventData Event) error {
693	697
694	698		eventBytes, _ := json.Marshal(eventData)
695	699		domain := "ci"
696		-	if eventData.Type == "local" {
697		-	domain = "local"
	700	+	if eventData.Type != "" {
	701	+	domain = eventData.Type
698	702		}
699	703		fmt.Fprintf(os.Stdout, "🚀 starting job %s.%s.%s\n", domain, eventData.Name, jobID) //nolint:errcheck
700	704		fmt.Fprintf(os.Stdout, " event: type=%s name=%s workspace=%s\n", eventData.Type, eventData.Name, eventData.Workspace) //nolint:errcheck

...	...		@@ -804,8 +808,8 @@ See: https://github.com/picosh/pici
804	808		}
805	809
806	810		domain = "ci"
807		-	if eventData.Type == "local" {
808		-	domain = "local"
	811	+	if eventData.Type != "" {
	812	+	domain = eventData.Type
809	813		}
810	814		session := fmt.Sprintf("%s.%s.%s.runner", domain, eventData.Name, jobID)
811	815		fmt.Fprintf(os.Stdout, " zmx tail %s\n", session) //nolint:errcheck

...	...		@@ -818,8 +822,8 @@ See: https://github.com/picosh/pici
818	822		// progress to stdout, then dumps session history and a final summary.
819	823		func waitAndReport(cfg Cfg, log slog.Logger, name, jobID, eventType string) error {
820	824		domain := "ci"
821		-	if eventType == "local" {
822		-	domain = "local"
	825	+	if eventType != "" {
	826	+	domain = eventType
823	827		}
824	828		prefix := domain + "." + name + "." + jobID + "."
825	829		interval := cfg.MonitorInterval

...	...		@@ -1180,7 +1184,7 @@ func monitorTick(cfg Cfg, log slog.Logger, output io.Writer, jobStates map[str
1180	1184		return fmt.Errorf("zmx list: %w", err)
1181	1185		}
1182	1186		sessions := parseZMXList(string(listOutput))
1183		-	ciSessions := filterCISessions(sessions)
	1187	+	ciSessions := filterCISessions(cfg, sessions)
1184	1188
1185	1189		if len(ciSessions) == 0 {
1186	1190		log.Debug("no ci.* sessions found")

...	...		@@ -1420,11 +1424,18 @@ func fmtDurationTs(started, ended int64) string {
1420	1424		return fmt.Sprintf("%.1fs", secs)
1421	1425		}
1422	1426
1423		-	// filterCISessions returns only sessions with ci. prefix.
1424		-	func filterCISessions(sessions []SessionInfo) []SessionInfo {
	1427	+	// filterCISessions returns only sessions matching cfg.SessionPrefix (default "ci.").
	1428	+	func filterCISessions(cfg *Cfg, sessions []SessionInfo) []SessionInfo {
	1429	+	prefix := "ci."
	1430	+	if cfg != nil && cfg.SessionPrefix != "" {
	1431	+	prefix = cfg.SessionPrefix
	1432	+	}
	1433	+	if !strings.HasSuffix(prefix, ".") {
	1434	+	prefix = prefix + "."
	1435	+	}
1425	1436		var filtered []SessionInfo
1426	1437		for _, s := range sessions {
1427		-	if strings.HasPrefix(s.Name, "ci.") {
	1438	+	if strings.HasPrefix(s.Name, prefix) {
1428	1439		filtered = append(filtered, s)
1429	1440		}
1430	1441		}

...	...		@@ -48,10 +48,10 @@ zmx run step2 echo "hello from step2"
48	48		ctx, cancel := context.WithCancel(context.Background())
49	49		testJobID := fmt.Sprintf("j-%d", time.Now().UnixNano()%100000000)
50	50		t.Cleanup(func() {
51		-	_ = exec.Command("zmx", "kill", "-f", fmt.Sprintf("ci.test-repo.%s", testJobID)).Run()
	51	+	_ = exec.Command("zmx", "kill", "-f", fmt.Sprintf("test-ci.test-repo.%s", testJobID)).Run()
52	52		})
53	53		event := Event{
54		-	Type: "ci",
	54	+	Type: "test-ci",
55	55		Name: "test-repo",
56	56		JobID: testJobID,
57	57		Workspace: workspaceDir,

...	...		@@ -68,6 +68,7 @@ zmx run step2 echo "hello from step2"
68	68		NewWorkspace: defaultWorkspaceFactory,
69	69		StatusOutput: statusBuf,
70	70		IncludeRunning: true,
	71	+	SessionPrefix: "test-ci.",
71	72		}
72	73
73	74		// 3. Run the runner (fire-and-forget, exits quickly)

...	...		@@ -184,8 +185,8 @@ zmx run step2 echo "hello from step2"
184	185		if finalPayload == nil {
185	186		t.Fatal("no final payload")
186	187		}
187		-	// Filter to only sessions with our job prefix to avoid picking up unrelated ci.* sessions
188		-	expectedPrefix := fmt.Sprintf("ci.%s.%s.", finalPayload.Name, finalPayload.JobID)
	188	+	// Filter to only sessions with our job prefix to avoid picking up unrelated test-ci.* sessions
	189	+	expectedPrefix := fmt.Sprintf("test-ci.%s.%s.", finalPayload.Name, finalPayload.JobID)
189	190		sessions := make([]SessionInfo, 0, len(finalPayload.Sessions))
190	191		for _, s := range finalPayload.Sessions {
191	192		if strings.HasPrefix(s.Name, expectedPrefix) {

...	...		@@ -756,6 +757,13 @@ func TestDetectGitCommit(t *testing.T) {
756	757		}
757	758
758	759		func TestWaitAndReport_Cancellation(t *testing.T) {
	760	+	if testing.Short() {
	761	+	t.Skip("skip integration test")
	762	+	}
	763	+	if _, err := exec.LookPath("zmx"); err != nil {
	764	+	t.Skip("zmx not found, skipping integration test")
	765	+	}
	766	+
759	767		ctx, cancel := context.WithCancel(context.Background())
760	768		cfg := &Cfg{
761	769		Ctx: ctx,

...	...		@@ -764,9 +772,13 @@ func TestWaitAndReport_Cancellation(t *testing.T) {
764	772		}
765	773
766	774		jobID := fmt.Sprintf("canceltest-%d", time.Now().UnixNano())
767		-	prefix := "local.testrepo." + jobID + "."
	775	+	prefix := "test-local.testrepo." + jobID + "."
768	776		runnerSession := prefix + "runner"
769	777
	778	+	t.Cleanup(func() {
	779	+	_ = exec.Command("zmx", "kill", "-f", runnerSession).Run()
	780	+	})
	781	+
770	782		cmd := exec.Command("zmx", "run", runnerSession, "-d", "sleep", "30")
771	783		if err := cmd.Run(); err != nil {
772	784		t.Fatalf("failed to start zmx session: %v", err)

...	...		@@ -777,7 +789,7 @@ func TestWaitAndReport_Cancellation(t *testing.T) {
777	789		cancel()
778	790		}()
779	791
780		-	err := waitAndReport(cfg, nil, "testrepo", jobID, "local")
	792	+	err := waitAndReport(cfg, nil, "testrepo", jobID, "test-local")
781	793		if err == nil {
782	794		t.Error("expected error when waitAndReport is cancelled, got nil")
783	795		}

Commit 8759922