Normalize task IDs and artifact folder prefixes

Servers with NVIDIA compute GPUs (H100 etc.) have no display output,
so KMS blanks the console. nomodeset disables kernel modesetting and
lets the NVIDIA proprietary driver handle display via Xorg.

KMS variant moved to advanced submenu for cases where it is needed.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audit/internal/platform/benchmark.go

@@ -679,7 +679,10 @@ func runBenchmarkInterconnect(ctx context.Context, verboseLog, runDir string, gp
 		"-g", strconv.Itoa(len(gpuIndices)),
 		"--iters", strconv.Itoa(maxInt(20, spec.NCCLSec/10)),
 	}
-	env := []string{"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices)}
+	env := []string{
+		"CUDA_DEVICE_ORDER=PCI_BUS_ID",
+		"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices),
+	}
 	logFunc(fmt.Sprintf("NCCL interconnect: gpus=%s", joinIndexList(gpuIndices)))
 	out, err := runSATCommandCtx(ctx, verboseLog, "nccl-all-reduce.log", cmd, env, logFunc)
 	_ = os.WriteFile(filepath.Join(runDir, "nccl-all-reduce.log"), out, 0644)

audit/internal/platform/kill_workers.go

@@ -15,6 +15,10 @@ var workerPatterns = []string{
 	"stress-ng",
 	"stressapptest",
 	"memtester",
+	// DCGM diagnostic workers — nvvs is spawned by dcgmi diag and survives
+	// if dcgmi is killed mid-run, leaving the GPU occupied (DCGM_ST_IN_USE).
+	"nvvs",
+	"dcgmi",
 }
 
 // KilledProcess describes a process that was sent SIGKILL.

audit/internal/platform/nvidia_stress.go

@@ -16,12 +16,12 @@ func (s *System) RunNvidiaStressPack(ctx context.Context, baseDir string, opts N
 		return "", err
 	}
 
-	return runAcceptancePackCtx(ctx, baseDir, nvidiaStressArchivePrefix(opts.Loader), []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-nvidia-smi-list.log", cmd: []string{"nvidia-smi", "-L"}},
+	return runAcceptancePackCtx(ctx, baseDir, nvidiaStressArchivePrefix(opts.Loader), withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-nvidia-smi-list.log", cmd: []string{"nvidia-smi", "-L"}},
 		job,
-		{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func nvidiaStressArchivePrefix(loader string) string {

audit/internal/platform/platform_stress.go

@@ -110,7 +110,7 @@ func (s *System) RunPlatformStress(
 			wg.Add(1)
 			go func() {
 				defer wg.Done()
-				gpuCmd := buildGPUStressCmd(loadCtx, vendor)
+				gpuCmd := buildGPUStressCmd(loadCtx, vendor, cycle.LoadSec)
 				if gpuCmd == nil {
 					return
 				}
@@ -392,6 +392,13 @@ func buildCPUStressCmd(ctx context.Context) (*exec.Cmd, error) {
 		cmdArgs = append(cmdArgs, "-M", strconv.Itoa(mb))
 	}
 	cmd := exec.CommandContext(ctx, path, cmdArgs...)
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		}
+		return nil
+	}
 	cmd.Stdout = nil
 	cmd.Stderr = nil
 	if err := startLowPriorityCmd(cmd, 15); err != nil {
@@ -402,28 +409,28 @@ func buildCPUStressCmd(ctx context.Context) (*exec.Cmd, error) {
 
 // buildGPUStressCmd creates a GPU stress command appropriate for the detected vendor.
 // Returns nil if no GPU stress tool is available (CPU-only cycling still useful).
-func buildGPUStressCmd(ctx context.Context, vendor string) *exec.Cmd {
+func buildGPUStressCmd(ctx context.Context, vendor string, durSec int) *exec.Cmd {
 	switch strings.ToLower(vendor) {
 	case "amd":
-		return buildAMDGPUStressCmd(ctx)
+		return buildAMDGPUStressCmd(ctx, durSec)
 	case "nvidia":
-		return buildNvidiaGPUStressCmd(ctx)
+		return buildNvidiaGPUStressCmd(ctx, durSec)
 	}
 	return nil
 }
 
-func buildAMDGPUStressCmd(ctx context.Context) *exec.Cmd {
+func buildAMDGPUStressCmd(ctx context.Context, durSec int) *exec.Cmd {
 	rvsArgs, err := resolveRVSCommand()
 	if err != nil {
 		return nil
 	}
 	rvsPath := rvsArgs[0]
-	cfg := `actions:
+	cfg := fmt.Sprintf(`actions:
 - name: gst_platform
   device: all
   module: gst
   parallel: true
-  duration: 86400000
+  duration: %d`, durSec*1000) + `
   copy_matrix: false
   target_stress: 90
   matrix_size_a: 8640
@@ -433,13 +440,20 @@ func buildAMDGPUStressCmd(ctx context.Context) *exec.Cmd {
 	cfgFile := "/tmp/bee-platform-gst.conf"
 	_ = os.WriteFile(cfgFile, []byte(cfg), 0644)
 	cmd := exec.CommandContext(ctx, rvsPath, "-c", cfgFile)
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		}
+		return nil
+	}
 	cmd.Stdout = nil
 	cmd.Stderr = nil
 	_ = startLowPriorityCmd(cmd, 10)
 	return cmd
 }
 
-func buildNvidiaGPUStressCmd(ctx context.Context) *exec.Cmd {
+func buildNvidiaGPUStressCmd(ctx context.Context, durSec int) *exec.Cmd {
 	path, err := satLookPath("bee-gpu-burn")
 	if err != nil {
 		path, err = satLookPath("bee-gpu-stress")
@@ -447,7 +461,17 @@ func buildNvidiaGPUStressCmd(ctx context.Context) *exec.Cmd {
 	if err != nil {
 		return nil
 	}
-	cmd := exec.CommandContext(ctx, path, "--seconds", "86400")
+	// Pass exact duration so bee-gpu-burn exits on its own when the cycle ends.
+	// Process group kill via Setpgid+Cancel is kept as a safety net for cases
+	// where the context is cancelled early (user stop, parent timeout).
+	cmd := exec.CommandContext(ctx, path, "--seconds", strconv.Itoa(durSec))
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		}
+		return nil
+	}
 	cmd.Stdout = nil
 	cmd.Stderr = nil
 	_ = startLowPriorityCmd(cmd, 10)

audit/internal/platform/runtime.go

@@ -173,6 +173,22 @@ func (s *System) collectGPURuntimeHealth(vendor string, health *schema.RuntimeHe
 
 	switch vendor {
 	case "nvidia":
+		if raw, err := os.ReadFile("/run/bee-nvidia-mode"); err == nil {
+			health.NvidiaGSPMode = strings.TrimSpace(string(raw))
+			if health.NvidiaGSPMode == "gsp-stuck" {
+				health.Issues = append(health.Issues, schema.RuntimeIssue{
+					Code:        "nvidia_gsp_stuck",
+					Severity:    "critical",
+					Description: "NVIDIA GSP firmware init timed out and the kernel module is stuck. Reboot and select 'GSP=off' in the boot menu.",
+				})
+			} else if health.NvidiaGSPMode == "gsp-off" {
+				health.Issues = append(health.Issues, schema.RuntimeIssue{
+					Code:        "nvidia_gsp_disabled",
+					Severity:    "warning",
+					Description: "NVIDIA GSP firmware disabled (fallback). Power management runs via CPU path — power draw readings may differ from reference hardware.",
+				})
+			}
+		}
 		health.DriverReady = strings.Contains(lsmodText, "nvidia ")
 		if !health.DriverReady {
 			health.Issues = append(health.Issues, schema.RuntimeIssue{

audit/internal/platform/sat.go

@@ -21,10 +21,11 @@ import (
 )
 
 var (
-	satExecCommand = exec.Command
-	satLookPath    = exec.LookPath
-	satGlob        = filepath.Glob
-	satStat        = os.Stat
+	satExecCommand  = exec.Command
+	satLookPath     = exec.LookPath
+	satGlob         = filepath.Glob
+	satStat         = os.Stat
+	satFreeMemBytes = freeMemBytes
 
 	rocmSMIExecutableGlobs = []string{
 		"/opt/rocm/bin/rocm-smi",
@@ -262,6 +263,9 @@ func (s *System) ListNvidiaGPUs() ([]NvidiaGPU, error) {
 			MemoryMB: memMB,
 		})
 	}
+	sort.Slice(gpus, func(i, j int) bool {
+		return gpus[i].Index < gpus[j].Index
+	})
 	return gpus, nil
 }
 
@@ -274,13 +278,13 @@ func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(
 	if gpuCount < 1 {
 		gpuCount = 1
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "nccl-tests", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-all-reduce-perf.log", cmd: []string{
+	return runAcceptancePackCtx(ctx, baseDir, "nccl-tests", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-all-reduce-perf.log", cmd: []string{
 			"all_reduce_perf", "-b", "512M", "-e", "4G", "-f", "2",
 			"-g", strconv.Itoa(gpuCount), "--iters", "20",
 		}},
-	}, logFunc)
+	), logFunc)
 }
 
 func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -292,18 +296,18 @@ func (s *System) RunNvidiaOfficialComputePack(ctx context.Context, baseDir strin
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dcgmi-version.log", cmd: []string{"dcgmi", "-v"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-compute", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dcgmi-version.log", cmd: []string{"dcgmi", "-v"}},
+		satJob{
 			name:       "03-dcgmproftester.log",
 			cmd:        profCmd,
 			env:        nvidiaVisibleDevicesEnv(selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "04-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -311,16 +315,16 @@ func (s *System) RunNvidiaTargetedPowerPack(ctx context.Context, baseDir string,
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-power", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-power", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-targeted-power.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("targeted_power", normalizeNvidiaBurnDuration(durationSec), selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaPulseTestPack(ctx context.Context, baseDir string, durationSec int, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -328,16 +332,16 @@ func (s *System) RunNvidiaPulseTestPack(ctx context.Context, baseDir string, dur
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-pulse", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-pulse", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-pulse-test.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("pulse_test", normalizeNvidiaBurnDuration(durationSec), selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
@@ -345,16 +349,16 @@ func (s *System) RunNvidiaBandwidthPack(ctx context.Context, baseDir string, gpu
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-bandwidth", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-bandwidth", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-nvbandwidth.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("nvbandwidth", 0, selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func (s *System) RunNvidiaAcceptancePack(baseDir string, logFunc func(string)) (string, error) {
@@ -378,16 +382,23 @@ func (s *System) RunNvidiaTargetedStressValidatePack(ctx context.Context, baseDi
 	if err != nil {
 		return "", err
 	}
-	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-stress", []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{
+	// Kill any lingering nvvs/dcgmi processes from a previous interrupted run
+	// before starting — otherwise dcgmi diag fails with DCGM_ST_IN_USE (-34).
+	if killed := KillTestWorkers(); len(killed) > 0 && logFunc != nil {
+		for _, p := range killed {
+			logFunc(fmt.Sprintf("pre-flight: killed stale worker pid=%d name=%s", p.PID, p.Name))
+		}
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "gpu-nvidia-targeted-stress", withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{
 			name:       "02-dcgmi-targeted-stress.log",
 			cmd:        nvidiaDCGMNamedDiagCommand("targeted_stress", normalizeNvidiaBurnDuration(durationSec), selected),
 			collectGPU: true,
 			gpuIndices: selected,
 		},
-		{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
-	}, logFunc)
+		satJob{name: "03-nvidia-smi-after.log", cmd: []string{"nvidia-smi", "--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"}},
+	), logFunc)
 }
 
 func resolveDCGMGPUIndices(gpuIndices []int) ([]int, error) {
@@ -404,6 +415,25 @@ func resolveDCGMGPUIndices(gpuIndices []int) ([]int, error) {
 	return all, nil
 }
 
+func memoryStressSizeArg() string {
+	if mb := envInt("BEE_VM_STRESS_SIZE_MB", 0); mb > 0 {
+		return fmt.Sprintf("%dM", mb)
+	}
+	availBytes := satFreeMemBytes()
+	if availBytes <= 0 {
+		return "80%"
+	}
+	availMB := availBytes / (1024 * 1024)
+	targetMB := (availMB * 2) / 3
+	if targetMB >= 256 {
+		targetMB = (targetMB / 256) * 256
+	}
+	if targetMB <= 0 {
+		return "80%"
+	}
+	return fmt.Sprintf("%dM", targetMB)
+}
+
 func (s *System) RunMemoryAcceptancePack(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
 	sizeMB := envInt("BEE_MEMTESTER_SIZE_MB", 128)
 	passes := envInt("BEE_MEMTESTER_PASSES", 1)
@@ -419,11 +449,9 @@ func (s *System) RunMemoryStressPack(ctx context.Context, baseDir string, durati
 	if seconds <= 0 {
 		seconds = envInt("BEE_VM_STRESS_SECONDS", 300)
 	}
-	// Use 80% of RAM by default; override with BEE_VM_STRESS_SIZE_MB.
-	sizeArg := "80%"
-	if mb := envInt("BEE_VM_STRESS_SIZE_MB", 0); mb > 0 {
-		sizeArg = fmt.Sprintf("%dM", mb)
-	}
+	// Base the default on current MemAvailable and keep headroom for the OS and
+	// concurrent stressors so mixed burn runs do not trip the OOM killer.
+	sizeArg := memoryStressSizeArg()
 	return runAcceptancePackCtx(ctx, baseDir, "memory-stress", []satJob{
 		{name: "01-free-before.log", cmd: []string{"free", "-h"}},
 		{name: "02-stress-ng-vm.log", cmd: []string{
@@ -540,14 +568,24 @@ type satStats struct {
 	Unsupported int
 }
 
+func withNvidiaPersistenceMode(jobs ...satJob) []satJob {
+	out := make([]satJob, 0, len(jobs)+1)
+	out = append(out, satJob{
+		name: "00-nvidia-smi-persistence-mode.log",
+		cmd:  []string{"nvidia-smi", "-pm", "1"},
+	})
+	out = append(out, jobs...)
+	return out
+}
+
 func nvidiaSATJobs() []satJob {
-	return []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
-		{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
-		{name: "04-nvidia-bug-report.log", cmd: []string{"nvidia-bug-report.sh", "--output-file", "{{run_dir}}/nvidia-bug-report.log"}},
-		{name: "05-bee-gpu-burn.log", cmd: []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}},
-	}
+	return withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
+		satJob{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
+		satJob{name: "04-nvidia-bug-report.log", cmd: []string{"nvidia-bug-report.sh", "--output-file", "{{run_dir}}/nvidia-bug-report.log"}},
+		satJob{name: "05-bee-gpu-burn.log", cmd: []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}},
+	)
 }
 
 func nvidiaDCGMJobs(diagLevel int, gpuIndices []int) []satJob {
@@ -562,12 +600,12 @@ func nvidiaDCGMJobs(diagLevel int, gpuIndices []int) []satJob {
 		}
 		diagArgs = append(diagArgs, "-i", strings.Join(ids, ","))
 	}
-	return []satJob{
-		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
-		{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
-		{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
-		{name: "04-dcgmi-diag.log", cmd: diagArgs},
-	}
+	return withNvidiaPersistenceMode(
+		satJob{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		satJob{name: "02-dmidecode-baseboard.log", cmd: []string{"dmidecode", "-t", "baseboard"}},
+		satJob{name: "03-dmidecode-system.log", cmd: []string{"dmidecode", "-t", "system"}},
+		satJob{name: "04-dcgmi-diag.log", cmd: diagArgs},
+	)
 }
 
 func nvidiaDCGMNamedDiagCommand(name string, durationSec int, gpuIndices []int) []string {
@@ -592,7 +630,10 @@ func nvidiaVisibleDevicesEnv(gpuIndices []int) []string {
 	if len(gpuIndices) == 0 {
 		return nil
 	}
-	return []string{"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices)}
+	return []string{
+		"CUDA_DEVICE_ORDER=PCI_BUS_ID",
+		"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices),
+	}
 }
 
 func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []satJob, logFunc func(string)) (string, error) {
@@ -633,6 +674,9 @@ func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []sa
 		if writeErr := os.WriteFile(filepath.Join(runDir, job.name), out, 0644); writeErr != nil {
 			return "", writeErr
 		}
+		if ctx.Err() != nil {
+			return "", ctx.Err()
+		}
 		status, rc := classifySATResult(job.name, out, err)
 		stats.Add(status)
 		key := strings.TrimSuffix(strings.TrimPrefix(job.name, "0"), ".log")

audit/internal/platform/sat_test.go

@@ -1,12 +1,14 @@
 package platform
 
 import (
+	"context"
 	"errors"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"strings"
 	"testing"
+	"time"
 )
 
 func TestStorageSATCommands(t *testing.T) {
@@ -28,13 +30,19 @@ func TestRunNvidiaAcceptancePackIncludesGPUStress(t *testing.T) {
 
 	jobs := nvidiaSATJobs()
 
-	if len(jobs) != 5 {
-		t.Fatalf("jobs=%d want 5", len(jobs))
+	if len(jobs) != 6 {
+		t.Fatalf("jobs=%d want 6", len(jobs))
 	}
-	if got := jobs[4].cmd[0]; got != "bee-gpu-burn" {
+	if got := jobs[0].cmd[0]; got != "nvidia-smi" {
+		t.Fatalf("preflight command=%q want nvidia-smi", got)
+	}
+	if got := strings.Join(jobs[0].cmd, " "); got != "nvidia-smi -pm 1" {
+		t.Fatalf("preflight=%q want %q", got, "nvidia-smi -pm 1")
+	}
+	if got := jobs[5].cmd[0]; got != "bee-gpu-burn" {
 		t.Fatalf("gpu stress command=%q want bee-gpu-burn", got)
 	}
-	if got := jobs[3].cmd[1]; got != "--output-file" {
+	if got := jobs[4].cmd[1]; got != "--output-file" {
 		t.Fatalf("bug report flag=%q want --output-file", got)
 	}
 }
@@ -82,7 +90,7 @@ func TestAMDStressJobsIncludeBandwidthAndGST(t *testing.T) {
 
 func TestNvidiaSATJobsUseBuiltinBurnDefaults(t *testing.T) {
 	jobs := nvidiaSATJobs()
-	got := jobs[4].cmd
+	got := jobs[5].cmd
 	want := []string{"bee-gpu-burn", "--seconds", "5", "--size-mb", "64"}
 	if len(got) != len(want) {
 		t.Fatalf("cmd len=%d want %d", len(got), len(want))
@@ -94,6 +102,19 @@ func TestNvidiaSATJobsUseBuiltinBurnDefaults(t *testing.T) {
 	}
 }
 
+func TestNvidiaDCGMJobsEnablePersistenceModeBeforeDiag(t *testing.T) {
+	jobs := nvidiaDCGMJobs(3, []int{2, 0})
+	if len(jobs) != 5 {
+		t.Fatalf("jobs=%d want 5", len(jobs))
+	}
+	if got := strings.Join(jobs[0].cmd, " "); got != "nvidia-smi -pm 1" {
+		t.Fatalf("preflight=%q want %q", got, "nvidia-smi -pm 1")
+	}
+	if got := strings.Join(jobs[4].cmd, " "); got != "dcgmi diag -r 3 -i 2,0" {
+		t.Fatalf("diag=%q want %q", got, "dcgmi diag -r 3 -i 2,0")
+	}
+}
+
 func TestBuildNvidiaStressJobUsesSelectedLoaderAndDevices(t *testing.T) {
 	t.Parallel()
 
@@ -234,11 +255,14 @@ func TestNvidiaDCGMNamedDiagCommandUsesDurationAndSelection(t *testing.T) {
 
 func TestNvidiaVisibleDevicesEnvUsesSelectedGPUs(t *testing.T) {
 	env := nvidiaVisibleDevicesEnv([]int{0, 2, 4})
-	if len(env) != 1 {
-		t.Fatalf("env len=%d want 1 (%v)", len(env), env)
+	if len(env) != 2 {
+		t.Fatalf("env len=%d want 2 (%v)", len(env), env)
 	}
-	if env[0] != "CUDA_VISIBLE_DEVICES=0,2,4" {
-		t.Fatalf("env[0]=%q want CUDA_VISIBLE_DEVICES=0,2,4", env[0])
+	if env[0] != "CUDA_DEVICE_ORDER=PCI_BUS_ID" {
+		t.Fatalf("env[0]=%q want CUDA_DEVICE_ORDER=PCI_BUS_ID", env[0])
+	}
+	if env[1] != "CUDA_VISIBLE_DEVICES=0,2,4" {
+		t.Fatalf("env[1]=%q want CUDA_VISIBLE_DEVICES=0,2,4", env[1])
 	}
 }
 
@@ -276,6 +300,37 @@ func TestEnvIntFallback(t *testing.T) {
 	}
 }
 
+func TestMemoryStressSizeArgUsesAvailableMemory(t *testing.T) {
+	oldFreeMemBytes := satFreeMemBytes
+	satFreeMemBytes = func() int64 { return 96 * 1024 * 1024 * 1024 }
+	t.Cleanup(func() { satFreeMemBytes = oldFreeMemBytes })
+
+	if got := memoryStressSizeArg(); got != "65536M" {
+		t.Fatalf("sizeArg=%q want 65536M", got)
+	}
+}
+
+func TestMemoryStressSizeArgRespectsOverride(t *testing.T) {
+	oldFreeMemBytes := satFreeMemBytes
+	satFreeMemBytes = func() int64 { return 96 * 1024 * 1024 * 1024 }
+	t.Cleanup(func() { satFreeMemBytes = oldFreeMemBytes })
+	t.Setenv("BEE_VM_STRESS_SIZE_MB", "4096")
+
+	if got := memoryStressSizeArg(); got != "4096M" {
+		t.Fatalf("sizeArg=%q want 4096M", got)
+	}
+}
+
+func TestMemoryStressSizeArgFallsBackWhenFreeMemoryUnknown(t *testing.T) {
+	oldFreeMemBytes := satFreeMemBytes
+	satFreeMemBytes = func() int64 { return 0 }
+	t.Cleanup(func() { satFreeMemBytes = oldFreeMemBytes })
+
+	if got := memoryStressSizeArg(); got != "80%" {
+		t.Fatalf("sizeArg=%q want 80%%", got)
+	}
+}
+
 func TestClassifySATResult(t *testing.T) {
 	tests := []struct {
 		name   string
@@ -300,6 +355,38 @@ func TestClassifySATResult(t *testing.T) {
 	}
 }
 
+func TestRunAcceptancePackCtxReturnsContextErrorWithoutArchive(t *testing.T) {
+	dir := t.TempDir()
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+
+	done := make(chan struct{})
+	go func() {
+		time.Sleep(100 * time.Millisecond)
+		cancel()
+		close(done)
+	}()
+
+	archive, err := runAcceptancePackCtx(ctx, dir, "cancelled-pack", []satJob{
+		{name: "01-sleep.log", cmd: []string{"sh", "-c", "sleep 5"}},
+	}, nil)
+	<-done
+
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("err=%v want context.Canceled", err)
+	}
+	if archive != "" {
+		t.Fatalf("archive=%q want empty", archive)
+	}
+	matches, globErr := filepath.Glob(filepath.Join(dir, "cancelled-pack-*.tar.gz"))
+	if globErr != nil {
+		t.Fatalf("Glob error: %v", globErr)
+	}
+	if len(matches) != 0 {
+		t.Fatalf("archives=%v want none", matches)
+	}
+}
+
 func TestParseStorageDevicesSkipsUSBDisks(t *testing.T) {
 	t.Parallel()
 

audit/internal/platform/services.go

@@ -61,7 +61,9 @@ func (s *System) ServiceState(name string) string {
 }
 
 func (s *System) ServiceDo(name string, action ServiceAction) (string, error) {
-	raw, err := exec.Command("systemctl", string(action), name).CombinedOutput()
+	// bee-web runs as the bee user; sudo is required to control system services.
+	// /etc/sudoers.d/bee grants bee NOPASSWD:ALL.
+	raw, err := exec.Command("sudo", "systemctl", string(action), name).CombinedOutput()
 	return string(raw), err
 }
 

audit/internal/schema/hardware.go

@@ -20,6 +20,7 @@ type RuntimeHealth struct {
 	ExportDir     string                 `json:"export_dir,omitempty"`
 	DriverReady   bool                   `json:"driver_ready,omitempty"`
 	CUDAReady     bool                   `json:"cuda_ready,omitempty"`
+	NvidiaGSPMode string                 `json:"nvidia_gsp_mode,omitempty"` // "gsp-on", "gsp-off", "gsp-stuck"
 	NetworkStatus string                 `json:"network_status,omitempty"`
 	Issues        []RuntimeIssue         `json:"issues,omitempty"`
 	Tools         []RuntimeToolStatus    `json:"tools,omitempty"`

audit/internal/webui/api.go

@@ -11,6 +11,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"regexp"
+	"sort"
 	"strings"
 	"sync/atomic"
 	"syscall"
@@ -21,13 +22,238 @@ import (
 )
 
 var ansiEscapeRE = regexp.MustCompile(`\x1b\[[0-9;]*[a-zA-Z]|\x1b[()][A-Z0-9]|\x1b[DABC]`)
+var apiListNvidiaGPUs = func(a *app.App) ([]platform.NvidiaGPU, error) {
+	if a == nil {
+		return nil, fmt.Errorf("app not configured")
+	}
+	return a.ListNvidiaGPUs()
+}
 
 // ── Job ID counter ────────────────────────────────────────────────────────────
 
 var jobCounter atomic.Uint64
 
-func newJobID(prefix string) string {
-	return fmt.Sprintf("%s-%d", prefix, jobCounter.Add(1))
+func newJobID(_ string) string {
+	start := int((jobCounter.Add(1) - 1) % 1000)
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	for offset := 0; offset < 1000; offset++ {
+		n := (start + offset) % 1000
+		id := fmt.Sprintf("TASK-%03d", n)
+		if !taskIDInUseLocked(id) {
+			return id
+		}
+	}
+	return fmt.Sprintf("TASK-%03d", start)
+}
+
+func taskIDInUseLocked(id string) bool {
+	for _, t := range globalQueue.tasks {
+		if t != nil && t.ID == id {
+			return true
+		}
+	}
+	return false
+}
+
+type taskRunResponse struct {
+	TaskID    string   `json:"task_id,omitempty"`
+	JobID     string   `json:"job_id,omitempty"`
+	TaskIDs   []string `json:"task_ids,omitempty"`
+	JobIDs    []string `json:"job_ids,omitempty"`
+	TaskCount int      `json:"task_count,omitempty"`
+}
+
+type nvidiaTaskSelection struct {
+	GPUIndices []int
+	Label      string
+}
+
+func writeTaskRunResponse(w http.ResponseWriter, tasks []*Task) {
+	if len(tasks) == 0 {
+		writeJSON(w, taskRunResponse{})
+		return
+	}
+	ids := make([]string, 0, len(tasks))
+	for _, t := range tasks {
+		if t == nil || strings.TrimSpace(t.ID) == "" {
+			continue
+		}
+		ids = append(ids, t.ID)
+	}
+	resp := taskRunResponse{TaskCount: len(ids)}
+	if len(ids) > 0 {
+		resp.TaskID = ids[0]
+		resp.JobID = ids[0]
+		resp.TaskIDs = ids
+		resp.JobIDs = ids
+	}
+	writeJSON(w, resp)
+}
+
+func shouldSplitHomogeneousNvidiaTarget(target string) bool {
+	switch strings.TrimSpace(target) {
+	case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute",
+		"nvidia-targeted-power", "nvidia-pulse", "nvidia-interconnect",
+		"nvidia-bandwidth", "nvidia-stress":
+		return true
+	default:
+		return false
+	}
+}
+
+func expandHomogeneousNvidiaSelections(gpus []platform.NvidiaGPU, include, exclude []int) ([]nvidiaTaskSelection, error) {
+	if len(gpus) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs detected")
+	}
+	indexed := make(map[int]platform.NvidiaGPU, len(gpus))
+	allIndices := make([]int, 0, len(gpus))
+	for _, gpu := range gpus {
+		indexed[gpu.Index] = gpu
+		allIndices = append(allIndices, gpu.Index)
+	}
+	sort.Ints(allIndices)
+
+	selected := allIndices
+	if len(include) > 0 {
+		selected = make([]int, 0, len(include))
+		seen := make(map[int]struct{}, len(include))
+		for _, idx := range include {
+			if _, ok := indexed[idx]; !ok {
+				continue
+			}
+			if _, dup := seen[idx]; dup {
+				continue
+			}
+			seen[idx] = struct{}{}
+			selected = append(selected, idx)
+		}
+		sort.Ints(selected)
+	}
+	if len(exclude) > 0 {
+		skip := make(map[int]struct{}, len(exclude))
+		for _, idx := range exclude {
+			skip[idx] = struct{}{}
+		}
+		filtered := selected[:0]
+		for _, idx := range selected {
+			if _, ok := skip[idx]; ok {
+				continue
+			}
+			filtered = append(filtered, idx)
+		}
+		selected = filtered
+	}
+	if len(selected) == 0 {
+		return nil, fmt.Errorf("no NVIDIA GPUs selected")
+	}
+
+	modelGroups := make(map[string][]platform.NvidiaGPU)
+	modelOrder := make([]string, 0)
+	for _, idx := range selected {
+		gpu := indexed[idx]
+		model := strings.TrimSpace(gpu.Name)
+		if model == "" {
+			model = fmt.Sprintf("GPU %d", gpu.Index)
+		}
+		if _, ok := modelGroups[model]; !ok {
+			modelOrder = append(modelOrder, model)
+		}
+		modelGroups[model] = append(modelGroups[model], gpu)
+	}
+	sort.Slice(modelOrder, func(i, j int) bool {
+		left := modelGroups[modelOrder[i]]
+		right := modelGroups[modelOrder[j]]
+		if len(left) == 0 || len(right) == 0 {
+			return modelOrder[i] < modelOrder[j]
+		}
+		return left[0].Index < right[0].Index
+	})
+
+	var groups []nvidiaTaskSelection
+	var singles []nvidiaTaskSelection
+	for _, model := range modelOrder {
+		group := modelGroups[model]
+		sort.Slice(group, func(i, j int) bool { return group[i].Index < group[j].Index })
+		indices := make([]int, 0, len(group))
+		for _, gpu := range group {
+			indices = append(indices, gpu.Index)
+		}
+		if len(indices) >= 2 {
+			groups = append(groups, nvidiaTaskSelection{
+				GPUIndices: indices,
+				Label:      fmt.Sprintf("%s; GPUs %s", model, joinTaskIndices(indices)),
+			})
+			continue
+		}
+		gpu := group[0]
+		singles = append(singles, nvidiaTaskSelection{
+			GPUIndices: []int{gpu.Index},
+			Label:      fmt.Sprintf("GPU %d — %s", gpu.Index, model),
+		})
+	}
+	return append(groups, singles...), nil
+}
+
+func joinTaskIndices(indices []int) string {
+	parts := make([]string, 0, len(indices))
+	for _, idx := range indices {
+		parts = append(parts, fmt.Sprintf("%d", idx))
+	}
+	return strings.Join(parts, ",")
+}
+
+func formatSplitTaskName(baseName, selectionLabel string) string {
+	baseName = strings.TrimSpace(baseName)
+	selectionLabel = strings.TrimSpace(selectionLabel)
+	if baseName == "" {
+		return selectionLabel
+	}
+	if selectionLabel == "" {
+		return baseName
+	}
+	return baseName + " (" + selectionLabel + ")"
+}
+
+func buildNvidiaTaskSet(target string, priority int, createdAt time.Time, params taskParams, baseName string, appRef *app.App, idPrefix string) ([]*Task, error) {
+	if !shouldSplitHomogeneousNvidiaTarget(target) {
+		t := &Task{
+			ID:        newJobID(idPrefix),
+			Name:      baseName,
+			Target:    target,
+			Priority:  priority,
+			Status:    TaskPending,
+			CreatedAt: createdAt,
+			params:    params,
+		}
+		return []*Task{t}, nil
+	}
+	gpus, err := apiListNvidiaGPUs(appRef)
+	if err != nil {
+		return nil, err
+	}
+	selections, err := expandHomogeneousNvidiaSelections(gpus, params.GPUIndices, params.ExcludeGPUIndices)
+	if err != nil {
+		return nil, err
+	}
+	tasks := make([]*Task, 0, len(selections))
+	for _, selection := range selections {
+		taskParamsCopy := params
+		taskParamsCopy.GPUIndices = append([]int(nil), selection.GPUIndices...)
+		taskParamsCopy.ExcludeGPUIndices = nil
+		displayName := formatSplitTaskName(baseName, selection.Label)
+		taskParamsCopy.DisplayName = displayName
+		tasks = append(tasks, &Task{
+			ID:        newJobID(idPrefix),
+			Name:      displayName,
+			Target:    target,
+			Priority:  priority,
+			Status:    TaskPending,
+			CreatedAt: createdAt,
+			params:    taskParamsCopy,
+		})
+	}
+	return tasks, nil
 }
 
 // ── SSE helpers ───────────────────────────────────────────────────────────────
@@ -207,28 +433,28 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 		}
 
 		name := taskDisplayName(target, body.Profile, body.Loader)
-		t := &Task{
-			ID:        newJobID("sat-" + target),
-			Name:      name,
-			Target:    target,
-			Status:    TaskPending,
-			CreatedAt: time.Now(),
-			params: taskParams{
-				Duration:           body.Duration,
-				DiagLevel:          body.DiagLevel,
-				GPUIndices:         body.GPUIndices,
-				ExcludeGPUIndices:  body.ExcludeGPUIndices,
-				Loader:             body.Loader,
-				BurnProfile:        body.Profile,
-				DisplayName:        body.DisplayName,
-				PlatformComponents: body.PlatformComponents,
-			},
-		}
 		if strings.TrimSpace(body.DisplayName) != "" {
-			t.Name = body.DisplayName
+			name = body.DisplayName
 		}
-		globalQueue.enqueue(t)
-		writeJSON(w, map[string]string{"task_id": t.ID, "job_id": t.ID})
+		params := taskParams{
+			Duration:           body.Duration,
+			DiagLevel:          body.DiagLevel,
+			GPUIndices:         body.GPUIndices,
+			ExcludeGPUIndices:  body.ExcludeGPUIndices,
+			Loader:             body.Loader,
+			BurnProfile:        body.Profile,
+			DisplayName:        body.DisplayName,
+			PlatformComponents: body.PlatformComponents,
+		}
+		tasks, err := buildNvidiaTaskSet(target, 0, time.Now(), params, name, h.opts.App, "sat-"+target)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, err.Error())
+			return
+		}
+		for _, t := range tasks {
+			globalQueue.enqueue(t)
+		}
+		writeTaskRunResponse(w, tasks)
 	}
 }
 
@@ -257,27 +483,26 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 	if body.RunNCCL != nil {
 		runNCCL = *body.RunNCCL
 	}
-	t := &Task{
-		ID:        newJobID("benchmark-nvidia"),
-		Name:      taskDisplayName("nvidia-benchmark", "", ""),
-		Target:    "nvidia-benchmark",
-		Priority:  15,
-		Status:    TaskPending,
-		CreatedAt: time.Now(),
-		params: taskParams{
-			GPUIndices:        body.GPUIndices,
-			ExcludeGPUIndices: body.ExcludeGPUIndices,
-			SizeMB:            body.SizeMB,
-			BenchmarkProfile:  body.Profile,
-			RunNCCL:           runNCCL,
-			DisplayName:       body.DisplayName,
-		},
-	}
+	name := taskDisplayName("nvidia-benchmark", "", "")
 	if strings.TrimSpace(body.DisplayName) != "" {
-		t.Name = body.DisplayName
+		name = body.DisplayName
 	}
-	globalQueue.enqueue(t)
-	writeJSON(w, map[string]string{"task_id": t.ID, "job_id": t.ID})
+	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", 15, time.Now(), taskParams{
+		GPUIndices:        body.GPUIndices,
+		ExcludeGPUIndices: body.ExcludeGPUIndices,
+		SizeMB:            body.SizeMB,
+		BenchmarkProfile:  body.Profile,
+		RunNCCL:           runNCCL,
+		DisplayName:       body.DisplayName,
+	}, name, h.opts.App, "benchmark-nvidia")
+	if err != nil {
+		writeError(w, http.StatusBadRequest, err.Error())
+		return
+	}
+	for _, t := range tasks {
+		globalQueue.enqueue(t)
+	}
+	writeTaskRunResponse(w, tasks)
 }
 
 func (h *handler) handleAPISATStream(w http.ResponseWriter, r *http.Request) {
@@ -383,11 +608,13 @@ func (h *handler) handleAPIServicesAction(w http.ResponseWriter, r *http.Request
 		return
 	}
 	result, err := h.opts.App.ServiceActionResult(req.Name, action)
+	status := "ok"
 	if err != nil {
-		writeError(w, http.StatusInternalServerError, err.Error())
-		return
+		status = "error"
 	}
-	writeJSON(w, map[string]string{"status": "ok", "output": result.Body})
+	// Always return 200 with output so the frontend can display the actual
+	// systemctl error message instead of a generic "exit status 1".
+	writeJSON(w, map[string]string{"status": status, "output": result.Body})
 }
 
 // ── Network ───────────────────────────────────────────────────────────────────

audit/internal/webui/api_test.go

@@ -1,6 +1,7 @@
 package webui
 
 import (
+	"encoding/json"
 	"net/http/httptest"
 	"strings"
 	"testing"
@@ -74,6 +75,14 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 		globalQueue.tasks = originalTasks
 		globalQueue.mu.Unlock()
 	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 3, Name: "NVIDIA H100 PCIe"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
 
 	h := &handler{opts: HandlerOptions{App: &app.App{}}}
 	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[1,3],"run_nccl":false}`))
@@ -101,6 +110,97 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 	}
 }
 
+func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
+	globalQueue.mu.Lock()
+	originalTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 0, Name: "NVIDIA H100 PCIe"},
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 2, Name: "NVIDIA H200 NVL"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/benchmark/nvidia/run", strings.NewReader(`{"profile":"standard","gpu_indices":[0,1,2],"run_nccl":false}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPIBenchmarkNvidiaRun(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	var resp taskRunResponse
+	if err := json.Unmarshal(rec.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if len(resp.TaskIDs) != 2 {
+		t.Fatalf("task_ids=%v want 2 items", resp.TaskIDs)
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 2 {
+		t.Fatalf("tasks=%d want 2", len(globalQueue.tasks))
+	}
+	if got := globalQueue.tasks[0].params.GPUIndices; len(got) != 2 || got[0] != 0 || got[1] != 1 {
+		t.Fatalf("task[0] gpu indices=%v want [0 1]", got)
+	}
+	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
+		t.Fatalf("task[1] gpu indices=%v want [2]", got)
+	}
+}
+
+func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(t *testing.T) {
+	globalQueue.mu.Lock()
+	originalTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = originalTasks
+		globalQueue.mu.Unlock()
+	})
+	prevList := apiListNvidiaGPUs
+	apiListNvidiaGPUs = func(_ *app.App) ([]platform.NvidiaGPU, error) {
+		return []platform.NvidiaGPU{
+			{Index: 0, Name: "NVIDIA H100 PCIe"},
+			{Index: 1, Name: "NVIDIA H100 PCIe"},
+			{Index: 2, Name: "NVIDIA H200 NVL"},
+		}, nil
+	}
+	t.Cleanup(func() { apiListNvidiaGPUs = prevList })
+
+	h := &handler{opts: HandlerOptions{App: &app.App{}}}
+	req := httptest.NewRequest("POST", "/api/sat/nvidia-targeted-power/run", strings.NewReader(`{"profile":"acceptance","gpu_indices":[0,1,2]}`))
+	rec := httptest.NewRecorder()
+
+	h.handleAPISATRun("nvidia-targeted-power").ServeHTTP(rec, req)
+
+	if rec.Code != 200 {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	globalQueue.mu.Lock()
+	defer globalQueue.mu.Unlock()
+	if len(globalQueue.tasks) != 2 {
+		t.Fatalf("tasks=%d want 2", len(globalQueue.tasks))
+	}
+	if got := globalQueue.tasks[0].params.GPUIndices; len(got) != 2 || got[0] != 0 || got[1] != 1 {
+		t.Fatalf("task[0] gpu indices=%v want [0 1]", got)
+	}
+	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
+		t.Fatalf("task[1] gpu indices=%v want [2]", got)
+	}
+}
+
 func TestPushFanRingsTracksByNameAndCarriesForwardMissingSamples(t *testing.T) {
 	h := &handler{}
 	h.pushFanRings([]platform.FanReading{

audit/internal/webui/charts_svg.go

@@ -6,6 +6,7 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"bee/audit/internal/platform"
@@ -52,6 +53,12 @@ var metricChartPalette = []string{
 	"#ffbe5c",
 }
 
+var gpuLabelCache struct {
+	mu       sync.Mutex
+	loadedAt time.Time
+	byIndex  map[int]string
+}
+
 func renderMetricChartSVG(title string, labels []string, times []time.Time, datasets [][]float64, names []string, yMin, yMax *float64, canvasHeight int, timeline []chartTimelineSegment) ([]byte, error) {
 	pointCount := len(labels)
 	if len(times) > pointCount {
@@ -76,15 +83,7 @@ func renderMetricChartSVG(title string, labels []string, times []time.Time, data
 		}
 	}
 
-	mn, avg, mx := globalStats(datasets)
-	if mx > 0 {
-		title = fmt.Sprintf("%s    ↓%s  ~%s  ↑%s",
-			title,
-			chartLegendNumber(mn),
-			chartLegendNumber(avg),
-			chartLegendNumber(mx),
-		)
-	}
+	statsLabel := chartStatsLabel(datasets)
 
 	legendItems := []metricChartSeries{}
 	for i, name := range names {
@@ -106,7 +105,7 @@ func renderMetricChartSVG(title string, labels []string, times []time.Time, data
 
 	var b strings.Builder
 	writeSVGOpen(&b, layout.Width, layout.Height)
-	writeChartFrame(&b, title, layout.Width, layout.Height)
+	writeChartFrame(&b, title, statsLabel, layout.Width, layout.Height)
 	writeTimelineIdleSpans(&b, layout, start, end, timeline)
 	writeVerticalGrid(&b, layout, times, pointCount, 8)
 	writeHorizontalGrid(&b, layout, scale)
@@ -126,21 +125,19 @@ func renderGPUOverviewChartSVG(idx int, samples []platform.LiveMetricSample, tim
 	temp := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.TempC })
 	power := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.PowerW })
 	coreClock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.ClockMHz })
-	memClock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.MemClockMHz })
-	if temp == nil && power == nil && coreClock == nil && memClock == nil {
+	if temp == nil && power == nil && coreClock == nil {
 		return nil, false, nil
 	}
 	labels := sampleTimeLabels(samples)
 	times := sampleTimes(samples)
 	svg, err := drawGPUOverviewChartSVG(
-		fmt.Sprintf("GPU %d Overview", idx),
+		gpuDisplayLabel(idx)+" Overview",
 		labels,
 		times,
 		[]metricChartSeries{
 			{Name: "Temp C", Values: coalesceDataset(temp, len(labels)), Color: "#f05a5a", AxisTitle: "Temp C"},
 			{Name: "Power W", Values: coalesceDataset(power, len(labels)), Color: "#ffb357", AxisTitle: "Power W"},
 			{Name: "Core Clock MHz", Values: coalesceDataset(coreClock, len(labels)), Color: "#73bf69", AxisTitle: "Core MHz"},
-			{Name: "Memory Clock MHz", Values: coalesceDataset(memClock, len(labels)), Color: "#5794f2", AxisTitle: "Memory MHz"},
 		},
 		timeline,
 	)
@@ -151,8 +148,8 @@ func renderGPUOverviewChartSVG(idx int, samples []platform.LiveMetricSample, tim
 }
 
 func drawGPUOverviewChartSVG(title string, labels []string, times []time.Time, series []metricChartSeries, timeline []chartTimelineSegment) ([]byte, error) {
-	if len(series) != 4 {
-		return nil, fmt.Errorf("gpu overview requires 4 series, got %d", len(series))
+	if len(series) != 3 {
+		return nil, fmt.Errorf("gpu overview requires 3 series, got %d", len(series))
 	}
 	const (
 		width      = 1400
@@ -166,7 +163,6 @@ func drawGPUOverviewChartSVG(title string, labels []string, times []time.Time, s
 		leftOuterAxis  = 72
 		leftInnerAxis  = 132
 		rightInnerAxis = 1268
-		rightOuterAxis = 1328
 	)
 	layout := chartLayout{
 		Width:      width,
@@ -176,7 +172,7 @@ func drawGPUOverviewChartSVG(title string, labels []string, times []time.Time, s
 		PlotTop:    plotTop,
 		PlotBottom: plotBottom,
 	}
-	axisX := []int{leftOuterAxis, leftInnerAxis, rightInnerAxis, rightOuterAxis}
+	axisX := []int{leftOuterAxis, leftInnerAxis, rightInnerAxis}
 	pointCount := len(labels)
 	if len(times) > pointCount {
 		pointCount = len(times)
@@ -214,7 +210,7 @@ func drawGPUOverviewChartSVG(title string, labels []string, times []time.Time, s
 
 	var b strings.Builder
 	writeSVGOpen(&b, width, height)
-	writeChartFrame(&b, title, width, height)
+	writeChartFrame(&b, title, "", width, height)
 	writeTimelineIdleSpans(&b, layout, start, end, timeline)
 	writeVerticalGrid(&b, layout, times, pointCount, 8)
 	writeHorizontalGrid(&b, layout, scales[0])
@@ -457,10 +453,14 @@ func writeSVGClose(b *strings.Builder) {
 	b.WriteString("</svg>\n")
 }
 
-func writeChartFrame(b *strings.Builder, title string, width, height int) {
+func writeChartFrame(b *strings.Builder, title, subtitle string, width, height int) {
 	fmt.Fprintf(b, `<rect width="%d" height="%d" rx="10" ry="10" fill="#ffffff" stroke="#d7e0ea"/>`+"\n", width, height)
 	fmt.Fprintf(b, `<text x="%d" y="30" text-anchor="middle" font-family="sans-serif" font-size="16" font-weight="700" fill="#1f2937">%s</text>`+"\n",
 		width/2, sanitizeChartText(title))
+	if strings.TrimSpace(subtitle) != "" {
+		fmt.Fprintf(b, `<text x="%d" y="50" text-anchor="middle" font-family="sans-serif" font-size="12" font-weight="600" fill="#64748b">%s</text>`+"\n",
+			width/2, sanitizeChartText(subtitle))
+	}
 }
 
 func writePlotBorder(b *strings.Builder, layout chartLayout) {
@@ -545,7 +545,21 @@ func writeSeriesPolyline(b *strings.Builder, layout chartLayout, times []time.Ti
 		x := chartXForTime(chartPointTime(times, 0), start, end, layout.PlotLeft, layout.PlotRight)
 		y := chartYForValue(values[0], scale, layout.PlotTop, layout.PlotBottom)
 		fmt.Fprintf(b, `<circle cx="%.1f" cy="%.1f" r="3.5" fill="%s"/>`+"\n", x, y, color)
+		return
 	}
+	peakIdx := 0
+	peakValue := values[0]
+	for idx, value := range values[1:] {
+		if value >= peakValue {
+			peakIdx = idx + 1
+			peakValue = value
+		}
+	}
+	x := chartXForTime(chartPointTime(times, peakIdx), start, end, layout.PlotLeft, layout.PlotRight)
+	y := chartYForValue(peakValue, scale, layout.PlotTop, layout.PlotBottom)
+	fmt.Fprintf(b, `<circle cx="%.1f" cy="%.1f" r="4.2" fill="%s" stroke="#ffffff" stroke-width="1.6"/>`+"\n", x, y, color)
+	fmt.Fprintf(b, `<path d="M %.1f %.1f L %.1f %.1f L %.1f %.1f Z" fill="%s" opacity="0.9"/>`+"\n",
+		x, y-10, x-5, y-18, x+5, y-18, color)
 }
 
 func writeLegend(b *strings.Builder, layout chartLayout, series []metricChartSeries) {
@@ -711,3 +725,49 @@ func valueClamp(value float64, scale chartScale) float64 {
 	}
 	return value
 }
+
+func chartStatsLabel(datasets [][]float64) string {
+	mn, avg, mx := globalStats(datasets)
+	if mx <= 0 && avg <= 0 && mn <= 0 {
+		return ""
+	}
+	return fmt.Sprintf("min %s   avg %s   max %s",
+		chartLegendNumber(mn),
+		chartLegendNumber(avg),
+		chartLegendNumber(mx),
+	)
+}
+
+func gpuDisplayLabel(idx int) string {
+	if name := gpuModelNameByIndex(idx); name != "" {
+		return fmt.Sprintf("GPU %d — %s", idx, name)
+	}
+	return fmt.Sprintf("GPU %d", idx)
+}
+
+func gpuModelNameByIndex(idx int) string {
+	now := time.Now()
+	gpuLabelCache.mu.Lock()
+	if now.Sub(gpuLabelCache.loadedAt) > 30*time.Second || gpuLabelCache.byIndex == nil {
+		gpuLabelCache.loadedAt = now
+		gpuLabelCache.byIndex = loadGPUModelNames()
+	}
+	name := strings.TrimSpace(gpuLabelCache.byIndex[idx])
+	gpuLabelCache.mu.Unlock()
+	return name
+}
+
+func loadGPUModelNames() map[int]string {
+	out := map[int]string{}
+	gpus, err := platform.New().ListNvidiaGPUs()
+	if err != nil {
+		return out
+	}
+	for _, gpu := range gpus {
+		name := strings.TrimSpace(gpu.Name)
+		if name != "" {
+			out[gpu.Index] = name
+		}
+	}
+	return out
+}

audit/internal/webui/jobs.go

@@ -9,13 +9,14 @@ import (
 
 // jobState holds the output lines and completion status of an async job.
 type jobState struct {
-	lines   []string
-	done    bool
-	err     string
-	mu      sync.Mutex
-	subs    []chan string
-	cancel  func() // optional cancel function; nil if job is not cancellable
-	logPath string
+	lines        []string
+	done         bool
+	err          string
+	mu           sync.Mutex
+	subs         []chan string
+	cancel       func() // optional cancel function; nil if job is not cancellable
+	logPath      string
+	serialPrefix string
 }
 
 // abort cancels the job if it has a cancel function and is not yet done.
@@ -36,6 +37,9 @@ func (j *jobState) append(line string) {
 	if j.logPath != "" {
 		appendJobLog(j.logPath, line)
 	}
+	if j.serialPrefix != "" {
+		taskSerialWriteLine(j.serialPrefix + line)
+	}
 	for _, ch := range j.subs {
 		select {
 		case ch <- line:
@@ -107,8 +111,11 @@ func (m *jobManager) get(id string) (*jobState, bool) {
 	return j, ok
 }
 
-func newTaskJobState(logPath string) *jobState {
+func newTaskJobState(logPath string, serialPrefix ...string) *jobState {
 	j := &jobState{logPath: logPath}
+	if len(serialPrefix) > 0 {
+		j.serialPrefix = serialPrefix[0]
+	}
 	if logPath == "" {
 		return j
 	}

audit/internal/webui/metricsdb.go

@@ -22,6 +22,13 @@ type MetricsDB struct {
 	db *sql.DB
 }
 
+func (m *MetricsDB) Close() error {
+	if m == nil || m.db == nil {
+		return nil
+	}
+	return m.db.Close()
+}
+
 // openMetricsDB opens (or creates) the metrics database at the given path.
 func openMetricsDB(path string) (*MetricsDB, error) {
 	if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
@@ -164,6 +171,23 @@ func (m *MetricsDB) LoadAll() ([]platform.LiveMetricSample, error) {
 	return m.loadSamples(`SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM sys_metrics ORDER BY ts`, nil)
 }
 
+// LoadBetween returns samples in chronological order within the given time window.
+func (m *MetricsDB) LoadBetween(start, end time.Time) ([]platform.LiveMetricSample, error) {
+	if m == nil {
+		return nil, nil
+	}
+	if start.IsZero() || end.IsZero() {
+		return nil, nil
+	}
+	if end.Before(start) {
+		start, end = end, start
+	}
+	return m.loadSamples(
+		`SELECT ts,cpu_load_pct,mem_load_pct,power_w FROM sys_metrics WHERE ts>=? AND ts<=? ORDER BY ts`,
+		start.Unix(), end.Unix(),
+	)
+}
+
 // loadSamples reconstructs LiveMetricSample rows from the normalized tables.
 func (m *MetricsDB) loadSamples(query string, args ...any) ([]platform.LiveMetricSample, error) {
 	rows, err := m.db.Query(query, args...)
@@ -364,9 +388,6 @@ func (m *MetricsDB) ExportCSV(w io.Writer) error {
 	return cw.Error()
 }
 
-// Close closes the database.
-func (m *MetricsDB) Close() { _ = m.db.Close() }
-
 func nullFloat(v float64) sql.NullFloat64 {
 	return sql.NullFloat64{Float64: v, Valid: true}
 }

audit/internal/webui/metricsdb_test.go

@@ -143,3 +143,32 @@ CREATE TABLE temp_metrics (
 		t.Fatalf("MemClockMHz=%v want 2600", got)
 	}
 }
+
+func TestMetricsDBLoadBetweenFiltersWindow(t *testing.T) {
+	db, err := openMetricsDB(filepath.Join(t.TempDir(), "metrics.db"))
+	if err != nil {
+		t.Fatalf("openMetricsDB: %v", err)
+	}
+	defer db.Close()
+
+	base := time.Unix(1_700_000_000, 0).UTC()
+	for i := 0; i < 5; i++ {
+		if err := db.Write(platform.LiveMetricSample{
+			Timestamp:  base.Add(time.Duration(i) * time.Minute),
+			CPULoadPct: float64(i),
+		}); err != nil {
+			t.Fatalf("Write(%d): %v", i, err)
+		}
+	}
+
+	got, err := db.LoadBetween(base.Add(1*time.Minute), base.Add(3*time.Minute))
+	if err != nil {
+		t.Fatalf("LoadBetween: %v", err)
+	}
+	if len(got) != 3 {
+		t.Fatalf("LoadBetween len=%d want 3", len(got))
+	}
+	if !got[0].Timestamp.Equal(base.Add(1*time.Minute)) || !got[2].Timestamp.Equal(base.Add(3*time.Minute)) {
+		t.Fatalf("window=%v..%v", got[0].Timestamp, got[2].Timestamp)
+	}
+}

audit/internal/webui/serial_console.go

@@ -0,0 +1,41 @@
+package webui
+
+import (
+	"fmt"
+	"os"
+	"strings"
+	"time"
+)
+
+var taskSerialWriteLine = writeTaskSerialLine
+
+func writeTaskSerialLine(line string) {
+	line = strings.TrimSpace(line)
+	if line == "" {
+		return
+	}
+	payload := fmt.Sprintf("%s %s\n", time.Now().UTC().Format("2006-01-02 15:04:05Z"), line)
+	for _, path := range []string{"/dev/ttyS0", "/dev/ttyS1", "/dev/console"} {
+		f, err := os.OpenFile(path, os.O_WRONLY|os.O_APPEND, 0)
+		if err != nil {
+			continue
+		}
+		_, _ = f.WriteString(payload)
+		_ = f.Close()
+		return
+	}
+}
+
+func taskSerialPrefix(t *Task) string {
+	if t == nil {
+		return "[task] "
+	}
+	return fmt.Sprintf("[task %s %s] ", t.ID, t.Name)
+}
+
+func taskSerialEvent(t *Task, event string) {
+	if t == nil {
+		return
+	}
+	taskSerialWriteLine(fmt.Sprintf("%s%s", taskSerialPrefix(t), strings.TrimSpace(event)))
+}

audit/internal/webui/server.go

@@ -221,6 +221,11 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	// ── Infrastructure ──────────────────────────────────────────────────────
 	mux.HandleFunc("GET /healthz", h.handleHealthz)
 	mux.HandleFunc("GET /api/ready", h.handleReady)
+	mux.HandleFunc("GET /loading", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Cache-Control", "no-store")
+		w.Header().Set("Content-Type", "text/html; charset=utf-8")
+		_, _ = w.Write([]byte(loadingPageHTML))
+	})
 
 	// ── Existing read-only endpoints (preserved for compatibility) ──────────
 	mux.HandleFunc("GET /audit.json", h.handleAuditJSON)
@@ -265,6 +270,9 @@ func NewHandler(opts HandlerOptions) http.Handler {
 	mux.HandleFunc("POST /api/tasks/{id}/cancel", h.handleAPITasksCancel)
 	mux.HandleFunc("POST /api/tasks/{id}/priority", h.handleAPITasksPriority)
 	mux.HandleFunc("GET /api/tasks/{id}/stream", h.handleAPITasksStream)
+	mux.HandleFunc("GET /api/tasks/{id}/charts", h.handleAPITaskChartsIndex)
+	mux.HandleFunc("GET /api/tasks/{id}/chart/", h.handleAPITaskChartSVG)
+	mux.HandleFunc("GET /tasks/{id}", h.handleTaskPage)
 
 	// Services
 	mux.HandleFunc("GET /api/services", h.handleAPIServicesList)
@@ -703,7 +711,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 		}
 		switch sub {
 		case "load":
-			title = fmt.Sprintf("GPU %d Load", idx)
+			title = gpuDisplayLabel(idx) + " Load"
 			util := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.UsagePct })
 			mem := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.MemUsagePct })
 			if util == nil && mem == nil {
@@ -714,7 +722,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			yMin = floatPtr(0)
 			yMax = floatPtr(100)
 		case "temp":
-			title = fmt.Sprintf("GPU %d Temperature", idx)
+			title = gpuDisplayLabel(idx) + " Temperature"
 			temp := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.TempC })
 			if temp == nil {
 				return nil, nil, nil, "", nil, nil, false
@@ -724,7 +732,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			yMin = floatPtr(0)
 			yMax = autoMax120(temp)
 		case "clock":
-			title = fmt.Sprintf("GPU %d Core Clock", idx)
+			title = gpuDisplayLabel(idx) + " Core Clock"
 			clock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.ClockMHz })
 			if clock == nil {
 				return nil, nil, nil, "", nil, nil, false
@@ -733,7 +741,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			names = []string{"Core Clock MHz"}
 			yMin, yMax = autoBounds120(clock)
 		case "memclock":
-			title = fmt.Sprintf("GPU %d Memory Clock", idx)
+			title = gpuDisplayLabel(idx) + " Memory Clock"
 			clock := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.MemClockMHz })
 			if clock == nil {
 				return nil, nil, nil, "", nil, nil, false
@@ -742,7 +750,7 @@ func chartDataFromSamples(path string, samples []platform.LiveMetricSample) ([][
 			names = []string{"Memory Clock MHz"}
 			yMin, yMax = autoBounds120(clock)
 		default:
-			title = fmt.Sprintf("GPU %d Power", idx)
+			title = gpuDisplayLabel(idx) + " Power"
 			power := gpuDatasetByIndex(samples, idx, func(g platform.GPUMetricRow) float64 { return g.PowerW })
 			if power == nil {
 				return nil, nil, nil, "", nil, nil, false
@@ -871,7 +879,7 @@ func gpuDatasets(samples []platform.LiveMetricSample, pick func(platform.GPUMetr
 			continue
 		}
 		datasets = append(datasets, ds)
-		names = append(names, fmt.Sprintf("GPU %d", idx))
+		names = append(names, gpuDisplayLabel(idx))
 	}
 	return datasets, names
 }
@@ -1206,37 +1214,106 @@ const loadingPageHTML = `<!DOCTYPE html>
 <html lang="en">
 <head>
 <meta charset="UTF-8">
-<title>EASY-BEE</title>
+<title>EASY-BEE — Starting</title>
 <style>
 *{margin:0;padding:0;box-sizing:border-box}
 html,body{height:100%;background:#0f1117;display:flex;align-items:center;justify-content:center;font-family:'Courier New',monospace;color:#e2e8f0}
-.logo{font-size:13px;line-height:1.4;color:#f6c90e;margin-bottom:48px;white-space:pre}
-.spinner{width:48px;height:48px;border:4px solid #2d3748;border-top-color:#f6c90e;border-radius:50%;animation:spin .8s linear infinite;margin:0 auto 24px}
+.wrap{text-align:center;width:420px}
+.logo{font-size:11px;line-height:1.4;color:#f6c90e;margin-bottom:6px;white-space:pre;text-align:left}
+.subtitle{font-size:12px;color:#a0aec0;text-align:left;margin-bottom:24px;padding-left:2px}
+.spinner{width:36px;height:36px;border:3px solid #2d3748;border-top-color:#f6c90e;border-radius:50%;animation:spin .8s linear infinite;margin:0 auto 14px}
+.spinner.hidden{display:none}
 @keyframes spin{to{transform:rotate(360deg)}}
-.status{font-size:14px;color:#a0aec0;letter-spacing:.05em}
+.status{font-size:13px;color:#a0aec0;margin-bottom:20px;min-height:18px}
+table{width:100%;border-collapse:collapse;font-size:12px;margin-bottom:20px;display:none}
+td{padding:3px 6px;text-align:left}
+td:first-child{color:#718096;width:55%}
+.ok{color:#68d391}
+.run{color:#f6c90e}
+.fail{color:#fc8181}
+.dim{color:#4a5568}
+.btn{background:#1a202c;color:#a0aec0;border:1px solid #2d3748;padding:7px 18px;font-size:12px;cursor:pointer;font-family:inherit;display:none}
+.btn:hover{border-color:#718096;color:#e2e8f0}
 </style>
 </head>
 <body>
-<div style="text-align:center">
+<div class="wrap">
   <div class="logo">  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗
   ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝
   █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗
   ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝
   ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗
   ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝</div>
-  <div class="spinner"></div>
-  <div class="status" id="s">Starting up...</div>
+  <div class="subtitle">Hardware Audit LiveCD</div>
+  <div class="spinner" id="spin"></div>
+  <div class="status" id="st">Connecting to bee-web...</div>
+  <table id="tbl"></table>
+  <button class="btn" id="btn" onclick="go()">Open app now</button>
 </div>
 <script>
-function probe(){
-  fetch('/api/ready',{cache:'no-store'})
-    .then(function(r){
-      if(r.ok){window.location.replace('/');}
-      else{setTimeout(probe,1000);}
+(function(){
+var gone = false;
+function go(){ if(!gone){gone=true;window.location.replace('/');} }
+
+function icon(s){
+  if(s==='active')   return '<span class="ok">&#9679; active</span>';
+  if(s==='failed')   return '<span class="fail">&#10005; failed</span>';
+  if(s==='activating'||s==='reloading') return '<span class="run">&#9675; starting</span>';
+  if(s==='inactive') return '<span class="dim">&#9675; inactive</span>';
+  return '<span class="dim">'+s+'</span>';
+}
+
+function allSettled(svcs){
+  for(var i=0;i<svcs.length;i++){
+    var s=svcs[i].state;
+    if(s!=='active'&&s!=='failed'&&s!=='inactive') return false;
+  }
+  return true;
+}
+
+var pollTimer=null;
+
+function pollServices(){
+  fetch('/api/services',{cache:'no-store'})
+    .then(function(r){return r.json();})
+    .then(function(svcs){
+      if(!svcs||!svcs.length) return;
+      var tbl=document.getElementById('tbl');
+      tbl.style.display='';
+      var html='';
+      for(var i=0;i<svcs.length;i++)
+        html+='<tr><td>'+svcs[i].name+'</td><td>'+icon(svcs[i].state)+'</td></tr>';
+      tbl.innerHTML=html;
+      if(allSettled(svcs)){
+        clearInterval(pollTimer);
+        document.getElementById('spin').className='spinner hidden';
+        document.getElementById('st').textContent='Ready \u2014 opening...';
+        setTimeout(go,800);
+      }
     })
-    .catch(function(){setTimeout(probe,1000);});
+    .catch(function(){});
+}
+
+function probe(){
+  fetch('/healthz',{cache:'no-store'})
+    .then(function(r){
+      if(r.ok){
+        document.getElementById('st').textContent='bee-web running \u2014 checking services...';
+        document.getElementById('btn').style.display='';
+        pollServices();
+        pollTimer=setInterval(pollServices,1500);
+      } else {
+        document.getElementById('st').textContent='bee-web starting (status '+r.status+')...';
+        setTimeout(probe,500);
+      }
+    })
+    .catch(function(){
+      document.getElementById('st').textContent='Waiting for bee-web to start...';
+      setTimeout(probe,500);
+    });
 }
 probe();
+})();
 </script>
 </body>
 </html>`

audit/internal/webui/server_test.go

@@ -1,6 +1,7 @@
 package webui
 
 import (
+	"encoding/json"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -184,15 +185,15 @@ func TestChartDataFromSamplesIncludesGPUClockCharts(t *testing.T) {
 		{
 			Timestamp: time.Now().Add(-2 * time.Minute),
 			GPUs: []platform.GPUMetricRow{
-				{GPUIndex: 0, ClockMHz: 1400, MemClockMHz: 2600},
-				{GPUIndex: 3, ClockMHz: 1500, MemClockMHz: 2800},
+				{GPUIndex: 0, ClockMHz: 1400},
+				{GPUIndex: 3, ClockMHz: 1500},
 			},
 		},
 		{
 			Timestamp: time.Now().Add(-1 * time.Minute),
 			GPUs: []platform.GPUMetricRow{
-				{GPUIndex: 0, ClockMHz: 1410, MemClockMHz: 2610},
-				{GPUIndex: 3, ClockMHz: 1510, MemClockMHz: 2810},
+				{GPUIndex: 0, ClockMHz: 1410},
+				{GPUIndex: 3, ClockMHz: 1510},
 			},
 		},
 	}
@@ -210,20 +211,6 @@ func TestChartDataFromSamplesIncludesGPUClockCharts(t *testing.T) {
 	if got := datasets[1][1]; got != 1510 {
 		t.Fatalf("GPU 3 core clock=%v want 1510", got)
 	}
-
-	datasets, names, _, title, _, _, ok = chartDataFromSamples("gpu-all-memclock", samples)
-	if !ok {
-		t.Fatal("gpu-all-memclock returned ok=false")
-	}
-	if title != "GPU Memory Clock" {
-		t.Fatalf("title=%q", title)
-	}
-	if len(names) != 2 || names[0] != "GPU 0" || names[1] != "GPU 3" {
-		t.Fatalf("names=%v", names)
-	}
-	if got := datasets[0][0]; got != 2600 {
-		t.Fatalf("GPU 0 memory clock=%v want 2600", got)
-	}
 }
 
 func TestNormalizePowerSeriesHoldsLastPositive(t *testing.T) {
@@ -256,10 +243,10 @@ func TestRenderMetricsUsesBufferedChartRefresh(t *testing.T) {
 	if !strings.Contains(body, `/api/metrics/chart/gpu-all-clock.svg`) {
 		t.Fatalf("metrics page should include GPU core clock chart: %s", body)
 	}
-	if !strings.Contains(body, `/api/metrics/chart/gpu-all-memclock.svg`) {
-		t.Fatalf("metrics page should include GPU memory clock chart: %s", body)
+	if strings.Contains(body, `/api/metrics/chart/gpu-all-memclock.svg`) {
+		t.Fatalf("metrics page should not include GPU memory clock chart: %s", body)
 	}
-	if !strings.Contains(body, `renderGPUOverviewCards(indices)`) {
+	if !strings.Contains(body, `renderGPUOverviewCards(indices, names)`) {
 		t.Fatalf("metrics page should build per-GPU chart cards dynamically: %s", body)
 	}
 }
@@ -585,7 +572,7 @@ func TestAuditPageRendersViewerFrameAndActions(t *testing.T) {
 	}
 }
 
-func TestTasksPageRendersLogModalAndPaginationControls(t *testing.T) {
+func TestTasksPageRendersOpenLinksAndPaginationControls(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()
 	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks", nil))
@@ -593,8 +580,8 @@ func TestTasksPageRendersLogModalAndPaginationControls(t *testing.T) {
 		t.Fatalf("status=%d", rec.Code)
 	}
 	body := rec.Body.String()
-	if !strings.Contains(body, `id="task-log-overlay"`) {
-		t.Fatalf("tasks page missing log modal overlay: %s", body)
+	if !strings.Contains(body, `Open a task to view its saved logs and charts.`) {
+		t.Fatalf("tasks page missing task report hint: %s", body)
 	}
 	if !strings.Contains(body, `_taskPageSize = 50`) {
 		t.Fatalf("tasks page missing pagination size config: %s", body)
@@ -615,8 +602,8 @@ func TestToolsPageRendersRestartGPUDriversButton(t *testing.T) {
 	if !strings.Contains(body, `Restart GPU Drivers`) {
 		t.Fatalf("tools page missing restart gpu drivers button: %s", body)
 	}
-	if !strings.Contains(body, `svcAction('bee-nvidia', 'restart')`) {
-		t.Fatalf("tools page missing bee-nvidia restart action: %s", body)
+	if !strings.Contains(body, `restartGPUDrivers()`) {
+		t.Fatalf("tools page missing restartGPUDrivers action: %s", body)
 	}
 	if !strings.Contains(body, `id="boot-source-text"`) {
 		t.Fatalf("tools page missing boot source field: %s", body)
@@ -650,6 +637,66 @@ func TestBenchmarkPageRendersGPUSelectionControls(t *testing.T) {
 	}
 }
 
+func TestBenchmarkPageRendersSavedResultsTable(t *testing.T) {
+	dir := t.TempDir()
+	exportDir := filepath.Join(dir, "export")
+	runDir := filepath.Join(exportDir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	if err := os.MkdirAll(runDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	result := platform.NvidiaBenchmarkResult{
+		GeneratedAt:      time.Date(2026, time.April, 6, 12, 0, 0, 0, time.UTC),
+		BenchmarkProfile: "standard",
+		OverallStatus:    "OK",
+		GPUs: []platform.BenchmarkGPUResult{
+			{
+				Index: 0,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1176.25,
+				},
+			},
+			{
+				Index: 1,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1168.50,
+				},
+			},
+		},
+	}
+	raw, err := json.Marshal(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runDir, "result.json"), raw, 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	handler := NewHandler(HandlerOptions{ExportDir: exportDir})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/benchmark", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	wantTime := result.GeneratedAt.Local().Format("2006-01-02 15:04:05")
+	for _, needle := range []string{
+		`Benchmark Results`,
+		`Composite score by saved benchmark run and GPU.`,
+		`NVIDIA H100 PCIe / GPU 0`,
+		`NVIDIA H100 PCIe / GPU 1`,
+		`#1`,
+		wantTime,
+		`1176.25`,
+		`1168.50`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("benchmark page missing %q: %s", needle, body)
+		}
+	}
+}
+
 func TestValidatePageRendersNvidiaTargetedStressCard(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()
@@ -663,6 +710,8 @@ func TestValidatePageRendersNvidiaTargetedStressCard(t *testing.T) {
 		`nvidia-targeted-stress`,
 		`controlled NVIDIA DCGM load`,
 		`<code>dcgmi diag targeted_stress</code>`,
+		`NVIDIA GPU Selection`,
+		`id="sat-gpu-list"`,
 	} {
 		if !strings.Contains(body, needle) {
 			t.Fatalf("validate page missing %q: %s", needle, body)
@@ -691,37 +740,154 @@ func TestBurnPageRendersGoalBasedNVIDIACards(t *testing.T) {
 	}
 }
 
-func TestTasksPageRendersScrollableLogModal(t *testing.T) {
+func TestTaskDetailPageRendersSavedReport(t *testing.T) {
 	dir := t.TempDir()
-	path := filepath.Join(dir, "audit.json")
 	exportDir := filepath.Join(dir, "export")
-	if err := os.MkdirAll(exportDir, 0755); err != nil {
+	reportDir := filepath.Join(exportDir, "tasks", "task-1_cpu_sat_done")
+	if err := os.MkdirAll(reportDir, 0755); err != nil {
 		t.Fatal(err)
 	}
-	if err := os.WriteFile(path, []byte(`{"collected_at":"2026-03-15T00:00:00Z"}`), 0644); err != nil {
+	reportPath := filepath.Join(reportDir, "report.html")
+	if err := os.WriteFile(reportPath, []byte(`<div class="card"><div class="card-head">Task Report</div><div class="card-body">saved report</div></div>`), 0644); err != nil {
 		t.Fatal(err)
 	}
 
-	handler := NewHandler(HandlerOptions{
-		Title:     "Bee Hardware Audit",
-		AuditPath: path,
-		ExportDir: exportDir,
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:             "task-1",
+		Name:           "CPU SAT",
+		Target:         "cpu",
+		Status:         TaskDone,
+		CreatedAt:      time.Now(),
+		ArtifactsDir:   reportDir,
+		ReportHTMLPath: reportPath,
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
 	})
 
+	handler := NewHandler(HandlerOptions{Title: "Bee Hardware Audit", ExportDir: exportDir})
+
 	rec := httptest.NewRecorder()
-	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks", nil))
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks/task-1", nil))
 	if rec.Code != http.StatusOK {
 		t.Fatalf("status=%d", rec.Code)
 	}
 	body := rec.Body.String()
-	if !strings.Contains(body, `height:calc(100vh - 32px)`) {
-		t.Fatalf("tasks page missing bounded log modal height: %s", body)
+	if !strings.Contains(body, `saved report`) {
+		t.Fatalf("task detail page missing saved report: %s", body)
 	}
-	if !strings.Contains(body, `flex:1;min-height:0;overflow:hidden`) {
-		t.Fatalf("tasks page missing log modal overflow guard: %s", body)
+	if !strings.Contains(body, `Back to Tasks`) {
+		t.Fatalf("task detail page missing back link: %s", body)
 	}
-	if !strings.Contains(body, `height:100%;min-height:0;overflow:auto`) {
-		t.Fatalf("tasks page missing scrollable log wrapper: %s", body)
+}
+
+func TestTaskDetailPageRendersCancelForRunningTask(t *testing.T) {
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:        "task-live-1",
+		Name:      "CPU SAT",
+		Target:    "cpu",
+		Status:    TaskRunning,
+		CreatedAt: time.Now(),
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+	})
+
+	handler := NewHandler(HandlerOptions{Title: "Bee Hardware Audit"})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/tasks/task-live-1", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, `Cancel</button>`) {
+		t.Fatalf("task detail page missing cancel button: %s", body)
+	}
+	if !strings.Contains(body, `function cancelTaskDetail(id)`) {
+		t.Fatalf("task detail page missing cancel handler: %s", body)
+	}
+	if !strings.Contains(body, `/api/tasks/' + id + '/cancel`) {
+		t.Fatalf("task detail page missing cancel endpoint: %s", body)
+	}
+	if !strings.Contains(body, `id="task-live-charts"`) {
+		t.Fatalf("task detail page missing live charts container: %s", body)
+	}
+	if !strings.Contains(body, `/api/tasks/' + taskId + '/charts`) {
+		t.Fatalf("task detail page missing live charts index endpoint: %s", body)
+	}
+}
+
+func TestTaskChartSVGUsesTaskTimeWindow(t *testing.T) {
+	dir := t.TempDir()
+	metricsPath := filepath.Join(dir, "metrics.db")
+	prevMetricsPath := taskReportMetricsDBPath
+	taskReportMetricsDBPath = metricsPath
+	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
+
+	db, err := openMetricsDB(metricsPath)
+	if err != nil {
+		t.Fatalf("openMetricsDB: %v", err)
+	}
+	base := time.Now().UTC()
+	samples := []platform.LiveMetricSample{
+		{Timestamp: base.Add(-3 * time.Minute), PowerW: 100},
+		{Timestamp: base.Add(-2 * time.Minute), PowerW: 200},
+		{Timestamp: base.Add(-1 * time.Minute), PowerW: 300},
+	}
+	for _, sample := range samples {
+		if err := db.Write(sample); err != nil {
+			t.Fatalf("Write: %v", err)
+		}
+	}
+	_ = db.Close()
+
+	started := base.Add(-2*time.Minute - 5*time.Second)
+	done := base.Add(-1*time.Minute + 5*time.Second)
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = []*Task{{
+		ID:        "task-chart-1",
+		Name:      "Power Window",
+		Target:    "cpu",
+		Status:    TaskDone,
+		CreatedAt: started.Add(-10 * time.Second),
+		StartedAt: &started,
+		DoneAt:    &done,
+	}}
+	globalQueue.mu.Unlock()
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+	})
+
+	handler := NewHandler(HandlerOptions{Title: "Bee Hardware Audit"})
+	req := httptest.NewRequest(http.MethodGet, "/api/tasks/task-chart-1/chart/server-power.svg", nil)
+	req.SetPathValue("id", "task-chart-1")
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	body := rec.Body.String()
+	if !strings.Contains(body, "System Power") {
+		t.Fatalf("task chart missing expected title: %s", body)
+	}
+	if !strings.Contains(body, "min 200") {
+		t.Fatalf("task chart stats should start from in-window sample: %s", body)
+	}
+	if strings.Contains(body, "min 100") {
+		t.Fatalf("task chart should not include pre-task sample in stats: %s", body)
 	}
 }
 

audit/internal/webui/task_page.go

@@ -0,0 +1,267 @@
+package webui
+
+import (
+	"encoding/json"
+	"fmt"
+	"html"
+	"net/http"
+	"os"
+	"strings"
+	"time"
+
+	"bee/audit/internal/platform"
+)
+
+func (h *handler) handleTaskPage(w http.ResponseWriter, r *http.Request) {
+	id := r.PathValue("id")
+	task, ok := globalQueue.findByID(id)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	snapshot := *task
+	body := renderTaskDetailPage(h.opts, snapshot)
+	w.Header().Set("Cache-Control", "no-store")
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	_, _ = w.Write([]byte(body))
+}
+
+func (h *handler) handleAPITaskChartsIndex(w http.ResponseWriter, r *http.Request) {
+	task, samples, _, _, ok := h.taskSamplesForRequest(r)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	type taskChartIndexEntry struct {
+		Title string `json:"title"`
+		File  string `json:"file"`
+	}
+	entries := make([]taskChartIndexEntry, 0)
+	for _, spec := range taskChartSpecsForSamples(samples) {
+		title, _, ok := renderTaskChartSVG(spec.Path, samples, taskTimelineForTask(task))
+		if !ok {
+			continue
+		}
+		entries = append(entries, taskChartIndexEntry{Title: title, File: spec.File})
+	}
+	w.Header().Set("Cache-Control", "no-store")
+	w.Header().Set("Content-Type", "application/json; charset=utf-8")
+	_ = json.NewEncoder(w).Encode(entries)
+}
+
+func (h *handler) handleAPITaskChartSVG(w http.ResponseWriter, r *http.Request) {
+	task, samples, _, _, ok := h.taskSamplesForRequest(r)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	file := strings.TrimPrefix(r.URL.Path, "/api/tasks/"+task.ID+"/chart/")
+	path, ok := taskChartPathFromFile(file)
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	title, buf, hasData := renderTaskChartSVG(path, samples, taskTimelineForTask(task))
+	if !hasData || len(buf) == 0 || strings.TrimSpace(title) == "" {
+		http.Error(w, "metrics history unavailable", http.StatusServiceUnavailable)
+		return
+	}
+	w.Header().Set("Content-Type", "image/svg+xml")
+	w.Header().Set("Cache-Control", "no-store")
+	_, _ = w.Write(buf)
+}
+
+func renderTaskDetailPage(opts HandlerOptions, task Task) string {
+	title := task.Name
+	if strings.TrimSpace(title) == "" {
+		title = task.ID
+	}
+	var body strings.Builder
+	body.WriteString(`<div style="display:flex;align-items:center;gap:12px;margin-bottom:16px;flex-wrap:wrap">`)
+	body.WriteString(`<a class="btn btn-secondary btn-sm" href="/tasks">Back to Tasks</a>`)
+	if task.Status == TaskRunning || task.Status == TaskPending {
+		body.WriteString(`<button class="btn btn-danger btn-sm" onclick="cancelTaskDetail('` + html.EscapeString(task.ID) + `')">Cancel</button>`)
+	}
+	body.WriteString(`<span style="font-size:12px;color:var(--muted)">Artifacts are saved in the task folder under <code>./tasks</code>.</span>`)
+	body.WriteString(`</div>`)
+
+	if report := loadTaskReportFragment(task); report != "" {
+		body.WriteString(report)
+	} else {
+		body.WriteString(`<div class="card"><div class="card-head">Task Summary</div><div class="card-body">`)
+		body.WriteString(`<div style="font-size:18px;font-weight:700">` + html.EscapeString(title) + `</div>`)
+		body.WriteString(`<div style="margin-top:8px">` + renderTaskStatusBadge(task.Status) + `</div>`)
+		if strings.TrimSpace(task.ErrMsg) != "" {
+			body.WriteString(`<div style="margin-top:8px;color:var(--crit-fg)">` + html.EscapeString(task.ErrMsg) + `</div>`)
+		}
+		body.WriteString(`</div></div>`)
+	}
+
+	if task.Status == TaskRunning {
+		body.WriteString(`<div class="card"><div class="card-head">Live Charts</div><div class="card-body">`)
+		body.WriteString(`<div id="task-live-charts" style="display:flex;flex-direction:column;gap:16px;color:var(--muted);font-size:13px">Loading charts...</div>`)
+		body.WriteString(`</div></div>`)
+	}
+
+	if task.Status == TaskRunning || task.Status == TaskPending {
+		body.WriteString(`<div class="card"><div class="card-head">Live Logs</div><div class="card-body">`)
+		body.WriteString(`<div id="task-live-log" class="terminal" style="max-height:none;white-space:pre-wrap">Connecting...</div>`)
+		body.WriteString(`</div></div>`)
+		body.WriteString(`<script>
+function cancelTaskDetail(id) {
+  fetch('/api/tasks/' + id + '/cancel', {method:'POST'}).then(function(){
+    var term = document.getElementById('task-live-log');
+    if (term) {
+      term.textContent += '\nCancel requested.\n';
+      term.scrollTop = term.scrollHeight;
+    }
+  });
+}
+function renderTaskLiveCharts(taskId, charts) {
+  const host = document.getElementById('task-live-charts');
+  if (!host) return;
+  if (!Array.isArray(charts) || charts.length === 0) {
+    host.innerHTML = 'Waiting for metric samples...';
+    return;
+  }
+  const seen = {};
+  charts.forEach(function(chart) {
+    seen[chart.file] = true;
+    let img = host.querySelector('img[data-chart-file="' + chart.file + '"]');
+    if (img) {
+      const card = img.closest('.card');
+      if (card) {
+        const title = card.querySelector('.card-head');
+        if (title) title.textContent = chart.title;
+      }
+      return;
+    }
+    const card = document.createElement('div');
+    card.className = 'card';
+    card.style.margin = '0';
+    card.innerHTML = '<div class="card-head"></div><div class="card-body" style="padding:12px"></div>';
+    card.querySelector('.card-head').textContent = chart.title;
+    const body = card.querySelector('.card-body');
+    img = document.createElement('img');
+    img.setAttribute('data-task-chart', '1');
+    img.setAttribute('data-chart-file', chart.file);
+    img.setAttribute('data-base-src', '/api/tasks/' + taskId + '/chart/' + chart.file);
+    img.src = '/api/tasks/' + taskId + '/chart/' + chart.file + '?t=' + Date.now();
+    img.style.width = '100%';
+    img.style.display = 'block';
+    img.style.borderRadius = '6px';
+    img.alt = chart.title;
+    body.appendChild(img);
+    host.appendChild(card);
+  });
+  Array.from(host.querySelectorAll('img[data-task-chart="1"]')).forEach(function(img) {
+    const file = img.getAttribute('data-chart-file') || '';
+    if (seen[file]) return;
+    const card = img.closest('.card');
+    if (card) card.remove();
+  });
+}
+function loadTaskLiveCharts(taskId) {
+  fetch('/api/tasks/' + taskId + '/charts').then(function(r){ return r.json(); }).then(function(charts){
+    renderTaskLiveCharts(taskId, charts);
+  }).catch(function(){
+    const host = document.getElementById('task-live-charts');
+    if (host) host.innerHTML = 'Task charts are unavailable.';
+  });
+}
+function refreshTaskLiveCharts() {
+  document.querySelectorAll('img[data-task-chart="1"]').forEach(function(img){
+    const base = img.dataset.baseSrc;
+    if (!base) return;
+    img.src = base + '?t=' + Date.now();
+  });
+}
+var _taskDetailES = new EventSource('/api/tasks/` + html.EscapeString(task.ID) + `/stream');
+var _taskDetailTerm = document.getElementById('task-live-log');
+var _taskChartTimer = null;
+var _taskChartsFrozen = false;
+_taskDetailES.onopen = function(){ _taskDetailTerm.textContent = ''; };
+_taskDetailES.onmessage = function(e){ _taskDetailTerm.textContent += e.data + "\n"; _taskDetailTerm.scrollTop = _taskDetailTerm.scrollHeight; };
+_taskDetailES.addEventListener('done', function(e){
+  if (_taskChartTimer) clearInterval(_taskChartTimer);
+  _taskDetailES.close();
+  _taskDetailES = null;
+  _taskChartsFrozen = true;
+  _taskDetailTerm.textContent += (e.data ? '\nTask finished with error.\n' : '\nTask finished.\n');
+  _taskDetailTerm.scrollTop = _taskDetailTerm.scrollHeight;
+  refreshTaskLiveCharts();
+});
+_taskDetailES.onerror = function(){
+  if (_taskChartTimer) clearInterval(_taskChartTimer);
+  if (_taskDetailES) {
+    _taskDetailES.close();
+    _taskDetailES = null;
+  }
+};
+loadTaskLiveCharts('` + html.EscapeString(task.ID) + `');
+_taskChartTimer = setInterval(function(){
+  if (_taskChartsFrozen) return;
+  loadTaskLiveCharts('` + html.EscapeString(task.ID) + `');
+  refreshTaskLiveCharts();
+}, 2000);
+</script>`)
+	}
+
+	return layoutHead(opts.Title+" — "+title) +
+		layoutNav("tasks", opts.BuildLabel) +
+		`<div class="main"><div class="topbar"><h1>` + html.EscapeString(title) + `</h1></div><div class="content">` +
+		body.String() +
+		`</div></div></body></html>`
+}
+
+func loadTaskReportFragment(task Task) string {
+	if strings.TrimSpace(task.ReportHTMLPath) == "" {
+		return ""
+	}
+	data, err := os.ReadFile(task.ReportHTMLPath)
+	if err != nil || len(data) == 0 {
+		return ""
+	}
+	return string(data)
+}
+
+func taskArtifactDownloadLink(task Task, absPath string) string {
+	if strings.TrimSpace(absPath) == "" {
+		return ""
+	}
+	return fmt.Sprintf(`/export/file?path=%s`, absPath)
+}
+
+func (h *handler) taskSamplesForRequest(r *http.Request) (Task, []platform.LiveMetricSample, time.Time, time.Time, bool) {
+	id := r.PathValue("id")
+	taskPtr, ok := globalQueue.findByID(id)
+	if !ok {
+		return Task{}, nil, time.Time{}, time.Time{}, false
+	}
+	task := *taskPtr
+	start, end := taskTimeWindow(&task)
+	samples, err := loadTaskMetricSamples(start, end)
+	if err != nil {
+		return task, nil, start, end, true
+	}
+	return task, samples, start, end, true
+}
+
+func taskTimelineForTask(task Task) []chartTimelineSegment {
+	start, end := taskTimeWindow(&task)
+	return []chartTimelineSegment{{Start: start, End: end, Active: true}}
+}
+
+func taskChartPathFromFile(file string) (string, bool) {
+	file = strings.TrimSpace(file)
+	for _, spec := range taskDashboardChartSpecs {
+		if spec.File == file {
+			return spec.Path, true
+		}
+	}
+	if strings.HasPrefix(file, "gpu-") && strings.HasSuffix(file, "-overview.svg") {
+		id := strings.TrimSuffix(strings.TrimPrefix(file, "gpu-"), "-overview.svg")
+		return "gpu/" + id + "-overview", true
+	}
+	return "", false
+}

audit/internal/webui/task_report.go

@@ -0,0 +1,343 @@
+package webui
+
+import (
+	"encoding/json"
+	"fmt"
+	"html"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"time"
+
+	"bee/audit/internal/platform"
+)
+
+var taskReportMetricsDBPath = metricsDBPath
+
+type taskReport struct {
+	ID          string            `json:"id"`
+	Name        string            `json:"name"`
+	Target      string            `json:"target"`
+	Status      string            `json:"status"`
+	CreatedAt   time.Time         `json:"created_at"`
+	StartedAt   *time.Time        `json:"started_at,omitempty"`
+	DoneAt      *time.Time        `json:"done_at,omitempty"`
+	DurationSec int               `json:"duration_sec,omitempty"`
+	Error       string            `json:"error,omitempty"`
+	LogFile     string            `json:"log_file,omitempty"`
+	Charts      []taskReportChart `json:"charts,omitempty"`
+	GeneratedAt time.Time         `json:"generated_at"`
+}
+
+type taskReportChart struct {
+	Title string `json:"title"`
+	File  string `json:"file"`
+}
+
+type taskChartSpec struct {
+	Path string
+	File string
+}
+
+var taskDashboardChartSpecs = []taskChartSpec{
+	{Path: "server-load", File: "server-load.svg"},
+	{Path: "server-temp-cpu", File: "server-temp-cpu.svg"},
+	{Path: "server-temp-ambient", File: "server-temp-ambient.svg"},
+	{Path: "server-power", File: "server-power.svg"},
+	{Path: "server-fans", File: "server-fans.svg"},
+	{Path: "gpu-all-load", File: "gpu-all-load.svg"},
+	{Path: "gpu-all-memload", File: "gpu-all-memload.svg"},
+	{Path: "gpu-all-clock", File: "gpu-all-clock.svg"},
+	{Path: "gpu-all-power", File: "gpu-all-power.svg"},
+	{Path: "gpu-all-temp", File: "gpu-all-temp.svg"},
+}
+
+func taskChartSpecsForSamples(samples []platform.LiveMetricSample) []taskChartSpec {
+	specs := make([]taskChartSpec, 0, len(taskDashboardChartSpecs)+len(taskGPUIndices(samples)))
+	specs = append(specs, taskDashboardChartSpecs...)
+	for _, idx := range taskGPUIndices(samples) {
+		specs = append(specs, taskChartSpec{
+			Path: fmt.Sprintf("gpu/%d-overview", idx),
+			File: fmt.Sprintf("gpu-%d-overview.svg", idx),
+		})
+	}
+	return specs
+}
+
+func writeTaskReportArtifacts(t *Task) error {
+	if t == nil {
+		return nil
+	}
+	ensureTaskReportPaths(t)
+	if strings.TrimSpace(t.ArtifactsDir) == "" {
+		return nil
+	}
+	if err := os.MkdirAll(t.ArtifactsDir, 0755); err != nil {
+		return err
+	}
+
+	start, end := taskTimeWindow(t)
+	samples, _ := loadTaskMetricSamples(start, end)
+	charts, inlineCharts := writeTaskCharts(t.ArtifactsDir, start, end, samples)
+
+	logText := ""
+	if data, err := os.ReadFile(t.LogPath); err == nil {
+		logText = string(data)
+	}
+
+	report := taskReport{
+		ID:          t.ID,
+		Name:        t.Name,
+		Target:      t.Target,
+		Status:      t.Status,
+		CreatedAt:   t.CreatedAt,
+		StartedAt:   t.StartedAt,
+		DoneAt:      t.DoneAt,
+		DurationSec: taskElapsedSec(t, reportDoneTime(t)),
+		Error:       t.ErrMsg,
+		LogFile:     filepath.Base(t.LogPath),
+		Charts:      charts,
+		GeneratedAt: time.Now().UTC(),
+	}
+	if err := writeJSONFile(t.ReportJSONPath, report); err != nil {
+		return err
+	}
+	return os.WriteFile(t.ReportHTMLPath, []byte(renderTaskReportFragment(report, inlineCharts, logText)), 0644)
+}
+
+func reportDoneTime(t *Task) time.Time {
+	if t != nil && t.DoneAt != nil && !t.DoneAt.IsZero() {
+		return *t.DoneAt
+	}
+	return time.Now()
+}
+
+func taskTimeWindow(t *Task) (time.Time, time.Time) {
+	if t == nil {
+		now := time.Now().UTC()
+		return now, now
+	}
+	start := t.CreatedAt.UTC()
+	if t.StartedAt != nil && !t.StartedAt.IsZero() {
+		start = t.StartedAt.UTC()
+	}
+	end := time.Now().UTC()
+	if t.DoneAt != nil && !t.DoneAt.IsZero() {
+		end = t.DoneAt.UTC()
+	}
+	if end.Before(start) {
+		end = start
+	}
+	return start, end
+}
+
+func loadTaskMetricSamples(start, end time.Time) ([]platform.LiveMetricSample, error) {
+	db, err := openMetricsDB(taskReportMetricsDBPath)
+	if err != nil {
+		return nil, err
+	}
+	defer db.Close()
+	return db.LoadBetween(start, end)
+}
+
+func writeTaskCharts(dir string, start, end time.Time, samples []platform.LiveMetricSample) ([]taskReportChart, map[string]string) {
+	if len(samples) == 0 {
+		return nil, nil
+	}
+	timeline := []chartTimelineSegment{{Start: start, End: end, Active: true}}
+	var charts []taskReportChart
+	inline := make(map[string]string)
+	for _, spec := range taskChartSpecsForSamples(samples) {
+		title, svg, ok := renderTaskChartSVG(spec.Path, samples, timeline)
+		if !ok || len(svg) == 0 {
+			continue
+		}
+		path := filepath.Join(dir, spec.File)
+		if err := os.WriteFile(path, svg, 0644); err != nil {
+			continue
+		}
+		charts = append(charts, taskReportChart{Title: title, File: spec.File})
+		inline[spec.File] = string(svg)
+	}
+	return charts, inline
+}
+
+func renderTaskChartSVG(path string, samples []platform.LiveMetricSample, timeline []chartTimelineSegment) (string, []byte, bool) {
+	if idx, sub, ok := parseGPUChartPath(path); ok && sub == "overview" {
+		buf, hasData, err := renderGPUOverviewChartSVG(idx, samples, timeline)
+		if err != nil || !hasData {
+			return "", nil, false
+		}
+		return gpuDisplayLabel(idx) + " Overview", buf, true
+	}
+	datasets, names, labels, title, yMin, yMax, ok := chartDataFromSamples(path, samples)
+	if !ok {
+		return "", nil, false
+	}
+	buf, err := renderMetricChartSVG(
+		title,
+		labels,
+		sampleTimes(samples),
+		datasets,
+		names,
+		yMin,
+		yMax,
+		chartCanvasHeightForPath(path, len(names)),
+		timeline,
+	)
+	if err != nil {
+		return "", nil, false
+	}
+	return title, buf, true
+}
+
+func taskGPUIndices(samples []platform.LiveMetricSample) []int {
+	seen := map[int]bool{}
+	var out []int
+	for _, s := range samples {
+		for _, g := range s.GPUs {
+			if seen[g.GPUIndex] {
+				continue
+			}
+			seen[g.GPUIndex] = true
+			out = append(out, g.GPUIndex)
+		}
+	}
+	sort.Ints(out)
+	return out
+}
+
+func writeJSONFile(path string, v any) error {
+	data, err := json.MarshalIndent(v, "", "  ")
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(path, data, 0644)
+}
+
+func renderTaskReportFragment(report taskReport, charts map[string]string, logText string) string {
+	var b strings.Builder
+	b.WriteString(`<div class="card"><div class="card-head">Task Report</div><div class="card-body">`)
+	b.WriteString(`<div class="grid2">`)
+	b.WriteString(`<div><div style="font-size:12px;color:var(--muted);margin-bottom:6px">Task</div><div style="font-size:16px;font-weight:700">` + html.EscapeString(report.Name) + `</div>`)
+	b.WriteString(`<div style="font-size:13px;color:var(--muted)">` + html.EscapeString(report.Target) + `</div></div>`)
+	b.WriteString(`<div><div style="font-size:12px;color:var(--muted);margin-bottom:6px">Status</div><div>` + renderTaskStatusBadge(report.Status) + `</div>`)
+	if strings.TrimSpace(report.Error) != "" {
+		b.WriteString(`<div style="margin-top:8px;font-size:13px;color:var(--crit-fg)">` + html.EscapeString(report.Error) + `</div>`)
+	}
+	b.WriteString(`</div></div>`)
+	b.WriteString(`<div style="margin-top:14px;font-size:13px;color:var(--muted)">`)
+	b.WriteString(`Started: ` + formatTaskTime(report.StartedAt, report.CreatedAt) + ` | Finished: ` + formatTaskTime(report.DoneAt, time.Time{}) + ` | Duration: ` + formatTaskDuration(report.DurationSec))
+	b.WriteString(`</div></div></div>`)
+	if benchmarkCard := renderTaskBenchmarkResultsCard(report.Target, logText); benchmarkCard != "" {
+		b.WriteString(benchmarkCard)
+	}
+
+	if len(report.Charts) > 0 {
+		for _, chart := range report.Charts {
+			b.WriteString(`<div class="card"><div class="card-head">` + html.EscapeString(chart.Title) + `</div><div class="card-body" style="padding:12px">`)
+			b.WriteString(charts[chart.File])
+			b.WriteString(`</div></div>`)
+		}
+	} else {
+		b.WriteString(`<div class="alert alert-info">No metric samples were captured during this task window.</div>`)
+	}
+
+	b.WriteString(`<div class="card"><div class="card-head">Logs</div><div class="card-body">`)
+	b.WriteString(`<div class="terminal" style="max-height:none;white-space:pre-wrap">` + html.EscapeString(strings.TrimSpace(logText)) + `</div>`)
+	b.WriteString(`</div></div>`)
+	return b.String()
+}
+
+func renderTaskBenchmarkResultsCard(target, logText string) string {
+	if strings.TrimSpace(target) != "nvidia-benchmark" {
+		return ""
+	}
+	resultPath := taskBenchmarkResultPath(logText)
+	if strings.TrimSpace(resultPath) == "" {
+		return ""
+	}
+	columns, runs := loadBenchmarkHistoryFromPaths([]string{resultPath})
+	if len(runs) == 0 {
+		return ""
+	}
+	return renderBenchmarkResultsCardFromRuns(
+		"Benchmark Results",
+		"Composite score for this benchmark task.",
+		"No benchmark results were saved for this task.",
+		columns,
+		runs,
+	)
+}
+
+func taskBenchmarkResultPath(logText string) string {
+	archivePath := taskArchivePathFromLog(logText)
+	if archivePath == "" {
+		return ""
+	}
+	runDir := strings.TrimSuffix(archivePath, ".tar.gz")
+	if runDir == archivePath {
+		return ""
+	}
+	return filepath.Join(runDir, "result.json")
+}
+
+func taskArchivePathFromLog(logText string) string {
+	lines := strings.Split(logText, "\n")
+	for i := len(lines) - 1; i >= 0; i-- {
+		line := strings.TrimSpace(lines[i])
+		if line == "" || !strings.HasPrefix(line, "Archive:") {
+			continue
+		}
+		path := strings.TrimSpace(strings.TrimPrefix(line, "Archive:"))
+		if strings.HasPrefix(path, "Archive written to ") {
+			path = strings.TrimSpace(strings.TrimPrefix(path, "Archive written to "))
+		}
+		if strings.HasSuffix(path, ".tar.gz") {
+			return path
+		}
+	}
+	return ""
+}
+
+func renderTaskStatusBadge(status string) string {
+	className := map[string]string{
+		TaskRunning:   "badge-ok",
+		TaskPending:   "badge-unknown",
+		TaskDone:      "badge-ok",
+		TaskFailed:    "badge-err",
+		TaskCancelled: "badge-unknown",
+	}[status]
+	if className == "" {
+		className = "badge-unknown"
+	}
+	label := strings.TrimSpace(status)
+	if label == "" {
+		label = "unknown"
+	}
+	return `<span class="badge ` + className + `">` + html.EscapeString(label) + `</span>`
+}
+
+func formatTaskTime(ts *time.Time, fallback time.Time) string {
+	if ts != nil && !ts.IsZero() {
+		return ts.Local().Format("2006-01-02 15:04:05")
+	}
+	if !fallback.IsZero() {
+		return fallback.Local().Format("2006-01-02 15:04:05")
+	}
+	return "n/a"
+}
+
+func formatTaskDuration(sec int) string {
+	if sec <= 0 {
+		return "n/a"
+	}
+	if sec < 60 {
+		return fmt.Sprintf("%ds", sec)
+	}
+	if sec < 3600 {
+		return fmt.Sprintf("%dm %02ds", sec/60, sec%60)
+	}
+	return fmt.Sprintf("%dh %02dm %02ds", sec/3600, (sec%3600)/60, sec%60)
+}

audit/internal/webui/tasks.go

@@ -92,17 +92,20 @@ func taskDisplayName(target, profile, loader string) string {
 
 // Task represents one unit of work in the queue.
 type Task struct {
-	ID         string     `json:"id"`
-	Name       string     `json:"name"`
-	Target     string     `json:"target"`
-	Priority   int        `json:"priority"`
-	Status     string     `json:"status"`
-	CreatedAt  time.Time  `json:"created_at"`
-	StartedAt  *time.Time `json:"started_at,omitempty"`
-	DoneAt     *time.Time `json:"done_at,omitempty"`
-	ElapsedSec int        `json:"elapsed_sec,omitempty"`
-	ErrMsg     string     `json:"error,omitempty"`
-	LogPath    string     `json:"log_path,omitempty"`
+	ID             string     `json:"id"`
+	Name           string     `json:"name"`
+	Target         string     `json:"target"`
+	Priority       int        `json:"priority"`
+	Status         string     `json:"status"`
+	CreatedAt      time.Time  `json:"created_at"`
+	StartedAt      *time.Time `json:"started_at,omitempty"`
+	DoneAt         *time.Time `json:"done_at,omitempty"`
+	ElapsedSec     int        `json:"elapsed_sec,omitempty"`
+	ErrMsg         string     `json:"error,omitempty"`
+	LogPath        string     `json:"log_path,omitempty"`
+	ArtifactsDir   string     `json:"artifacts_dir,omitempty"`
+	ReportJSONPath string     `json:"report_json_path,omitempty"`
+	ReportHTMLPath string     `json:"report_html_path,omitempty"`
 
 	// runtime fields (not serialised)
 	job    *jobState
@@ -126,17 +129,20 @@ type taskParams struct {
 }
 
 type persistedTask struct {
-	ID        string     `json:"id"`
-	Name      string     `json:"name"`
-	Target    string     `json:"target"`
-	Priority  int        `json:"priority"`
-	Status    string     `json:"status"`
-	CreatedAt time.Time  `json:"created_at"`
-	StartedAt *time.Time `json:"started_at,omitempty"`
-	DoneAt    *time.Time `json:"done_at,omitempty"`
-	ErrMsg    string     `json:"error,omitempty"`
-	LogPath   string     `json:"log_path,omitempty"`
-	Params    taskParams `json:"params,omitempty"`
+	ID             string     `json:"id"`
+	Name           string     `json:"name"`
+	Target         string     `json:"target"`
+	Priority       int        `json:"priority"`
+	Status         string     `json:"status"`
+	CreatedAt      time.Time  `json:"created_at"`
+	StartedAt      *time.Time `json:"started_at,omitempty"`
+	DoneAt         *time.Time `json:"done_at,omitempty"`
+	ErrMsg         string     `json:"error,omitempty"`
+	LogPath        string     `json:"log_path,omitempty"`
+	ArtifactsDir   string     `json:"artifacts_dir,omitempty"`
+	ReportJSONPath string     `json:"report_json_path,omitempty"`
+	ReportHTMLPath string     `json:"report_html_path,omitempty"`
+	Params         taskParams `json:"params,omitempty"`
 }
 
 type burnPreset struct {
@@ -252,6 +258,7 @@ func (q *taskQueue) enqueue(t *Task) {
 	q.prune()
 	q.persistLocked()
 	q.mu.Unlock()
+	taskSerialEvent(t, "queued")
 	select {
 	case q.trigger <- struct{}{}:
 	default:
@@ -416,44 +423,30 @@ func (q *taskQueue) worker() {
 			setCPUGovernor("performance")
 			defer setCPUGovernor("powersave")
 
-			// Drain all pending tasks and start them in parallel.
-			q.mu.Lock()
-			var batch []*Task
 			for {
+				q.mu.Lock()
 				t := q.nextPending()
 				if t == nil {
-					break
+					q.prune()
+					q.persistLocked()
+					q.mu.Unlock()
+					return
 				}
 				now := time.Now()
 				t.Status = TaskRunning
 				t.StartedAt = &now
 				t.DoneAt = nil
 				t.ErrMsg = ""
-				j := newTaskJobState(t.LogPath)
+				j := newTaskJobState(t.LogPath, taskSerialPrefix(t))
 				t.job = j
-				batch = append(batch, t)
-			}
-			if len(batch) > 0 {
 				q.persistLocked()
-			}
-			q.mu.Unlock()
+				q.mu.Unlock()
 
-			var wg sync.WaitGroup
-			for _, t := range batch {
-				t := t
-				j := t.job
 				taskCtx, taskCancel := context.WithCancel(context.Background())
 				j.cancel = taskCancel
-				wg.Add(1)
-				goRecoverOnce("task "+t.Target, func() {
-					defer wg.Done()
-					defer taskCancel()
-					q.executeTask(t, j, taskCtx)
-				})
-			}
-			wg.Wait()
+				q.executeTask(t, j, taskCtx)
+				taskCancel()
 
-			if len(batch) > 0 {
 				q.mu.Lock()
 				q.prune()
 				q.persistLocked()
@@ -496,8 +489,6 @@ func (q *taskQueue) executeTask(t *Task, j *jobState, ctx context.Context) {
 
 func (q *taskQueue) finalizeTaskRun(t *Task, j *jobState) {
 	q.mu.Lock()
-	defer q.mu.Unlock()
-
 	now := time.Now()
 	t.DoneAt = &now
 	if t.Status == TaskRunning {
@@ -509,7 +500,18 @@ func (q *taskQueue) finalizeTaskRun(t *Task, j *jobState) {
 			t.ErrMsg = ""
 		}
 	}
+	q.finalizeTaskArtifactPathsLocked(t)
 	q.persistLocked()
+	q.mu.Unlock()
+
+	if err := writeTaskReportArtifacts(t); err != nil {
+		appendJobLog(t.LogPath, "WARN: task report generation failed: "+err.Error())
+	}
+	if t.ErrMsg != "" {
+		taskSerialEvent(t, "finished with status="+t.Status+" error="+t.ErrMsg)
+		return
+	}
+	taskSerialEvent(t, "finished with status="+t.Status)
 }
 
 // setCPUGovernor writes the given governor to all CPU scaling_governor sysfs files.
@@ -848,6 +850,7 @@ func (h *handler) handleAPITasksCancel(w http.ResponseWriter, r *http.Request) {
 		now := time.Now()
 		t.DoneAt = &now
 		globalQueue.persistLocked()
+		taskSerialEvent(t, "finished with status="+t.Status)
 		writeJSON(w, map[string]string{"status": "cancelled"})
 	case TaskRunning:
 		if t.job != nil {
@@ -857,6 +860,7 @@ func (h *handler) handleAPITasksCancel(w http.ResponseWriter, r *http.Request) {
 		now := time.Now()
 		t.DoneAt = &now
 		globalQueue.persistLocked()
+		taskSerialEvent(t, "finished with status="+t.Status)
 		writeJSON(w, map[string]string{"status": "cancelled"})
 	default:
 		writeError(w, http.StatusConflict, "task is not running or pending")
@@ -897,6 +901,7 @@ func (h *handler) handleAPITasksCancelAll(w http.ResponseWriter, _ *http.Request
 		case TaskPending:
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			n++
 		case TaskRunning:
 			if t.job != nil {
@@ -904,6 +909,7 @@ func (h *handler) handleAPITasksCancelAll(w http.ResponseWriter, _ *http.Request
 			}
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			n++
 		}
 	}
@@ -922,6 +928,7 @@ func (h *handler) handleAPITasksKillWorkers(w http.ResponseWriter, _ *http.Reque
 		case TaskPending:
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			cancelled++
 		case TaskRunning:
 			if t.job != nil {
@@ -929,6 +936,7 @@ func (h *handler) handleAPITasksKillWorkers(w http.ResponseWriter, _ *http.Reque
 			}
 			t.Status = TaskCancelled
 			t.DoneAt = &now
+			taskSerialEvent(t, "finished with status="+t.Status)
 			cancelled++
 		}
 	}
@@ -992,10 +1000,10 @@ func (h *handler) handleAPITasksStream(w http.ResponseWriter, r *http.Request) {
 }
 
 func (q *taskQueue) assignTaskLogPathLocked(t *Task) {
-	if t.LogPath != "" || q.logsDir == "" || t.ID == "" {
+	if q.logsDir == "" || t.ID == "" {
 		return
 	}
-	t.LogPath = filepath.Join(q.logsDir, t.ID+".log")
+	q.ensureTaskArtifactPathsLocked(t)
 }
 
 func (q *taskQueue) loadLocked() {
@@ -1012,17 +1020,20 @@ func (q *taskQueue) loadLocked() {
 	}
 	for _, pt := range persisted {
 		t := &Task{
-			ID:        pt.ID,
-			Name:      pt.Name,
-			Target:    pt.Target,
-			Priority:  pt.Priority,
-			Status:    pt.Status,
-			CreatedAt: pt.CreatedAt,
-			StartedAt: pt.StartedAt,
-			DoneAt:    pt.DoneAt,
-			ErrMsg:    pt.ErrMsg,
-			LogPath:   pt.LogPath,
-			params:    pt.Params,
+			ID:             pt.ID,
+			Name:           pt.Name,
+			Target:         pt.Target,
+			Priority:       pt.Priority,
+			Status:         pt.Status,
+			CreatedAt:      pt.CreatedAt,
+			StartedAt:      pt.StartedAt,
+			DoneAt:         pt.DoneAt,
+			ErrMsg:         pt.ErrMsg,
+			LogPath:        pt.LogPath,
+			ArtifactsDir:   pt.ArtifactsDir,
+			ReportJSONPath: pt.ReportJSONPath,
+			ReportHTMLPath: pt.ReportHTMLPath,
+			params:         pt.Params,
 		}
 		q.assignTaskLogPathLocked(t)
 		if t.Status == TaskRunning {
@@ -1053,17 +1064,20 @@ func (q *taskQueue) persistLocked() {
 	state := make([]persistedTask, 0, len(q.tasks))
 	for _, t := range q.tasks {
 		state = append(state, persistedTask{
-			ID:        t.ID,
-			Name:      t.Name,
-			Target:    t.Target,
-			Priority:  t.Priority,
-			Status:    t.Status,
-			CreatedAt: t.CreatedAt,
-			StartedAt: t.StartedAt,
-			DoneAt:    t.DoneAt,
-			ErrMsg:    t.ErrMsg,
-			LogPath:   t.LogPath,
-			Params:    t.params,
+			ID:             t.ID,
+			Name:           t.Name,
+			Target:         t.Target,
+			Priority:       t.Priority,
+			Status:         t.Status,
+			CreatedAt:      t.CreatedAt,
+			StartedAt:      t.StartedAt,
+			DoneAt:         t.DoneAt,
+			ErrMsg:         t.ErrMsg,
+			LogPath:        t.LogPath,
+			ArtifactsDir:   t.ArtifactsDir,
+			ReportJSONPath: t.ReportJSONPath,
+			ReportHTMLPath: t.ReportHTMLPath,
+			Params:         t.params,
 		})
 	}
 	data, err := json.MarshalIndent(state, "", "  ")
@@ -1094,3 +1108,113 @@ func taskElapsedSec(t *Task, now time.Time) int {
 	}
 	return int(end.Sub(start).Round(time.Second) / time.Second)
 }
+
+func taskFolderStatus(status string) string {
+	status = strings.TrimSpace(strings.ToLower(status))
+	switch status {
+	case TaskRunning, TaskDone, TaskFailed, TaskCancelled:
+		return status
+	default:
+		return TaskPending
+	}
+}
+
+func sanitizeTaskFolderPart(s string) string {
+	s = strings.TrimSpace(strings.ToLower(s))
+	if s == "" {
+		return "task"
+	}
+	var b strings.Builder
+	lastDash := false
+	for _, r := range s {
+		isAlnum := (r >= 'a' && r <= 'z') || (r >= '0' && r <= '9')
+		if isAlnum {
+			b.WriteRune(r)
+			lastDash = false
+			continue
+		}
+		if !lastDash {
+			b.WriteByte('-')
+			lastDash = true
+		}
+	}
+	out := strings.Trim(b.String(), "-")
+	if out == "" {
+		return "task"
+	}
+	return out
+}
+
+func taskArtifactsDir(root string, t *Task, status string) string {
+	if strings.TrimSpace(root) == "" || t == nil {
+		return ""
+	}
+	prefix := taskFolderNumberPrefix(t.ID)
+	return filepath.Join(root, fmt.Sprintf("%s_%s_%s", prefix, sanitizeTaskFolderPart(t.Name), taskFolderStatus(status)))
+}
+
+func taskFolderNumberPrefix(taskID string) string {
+	taskID = strings.TrimSpace(taskID)
+	if strings.HasPrefix(taskID, "TASK-") && len(taskID) >= len("TASK-000") {
+		num := strings.TrimSpace(strings.TrimPrefix(taskID, "TASK-"))
+		if len(num) == 3 {
+			allDigits := true
+			for _, r := range num {
+				if r < '0' || r > '9' {
+					allDigits = false
+					break
+				}
+			}
+			if allDigits {
+				return num
+			}
+		}
+	}
+	fallback := sanitizeTaskFolderPart(taskID)
+	if fallback == "" {
+		return "000"
+	}
+	return fallback
+}
+
+func ensureTaskReportPaths(t *Task) {
+	if t == nil || strings.TrimSpace(t.ArtifactsDir) == "" {
+		return
+	}
+	if t.LogPath == "" || filepath.Base(t.LogPath) == "task.log" {
+		t.LogPath = filepath.Join(t.ArtifactsDir, "task.log")
+	}
+	t.ReportJSONPath = filepath.Join(t.ArtifactsDir, "report.json")
+	t.ReportHTMLPath = filepath.Join(t.ArtifactsDir, "report.html")
+}
+
+func (q *taskQueue) ensureTaskArtifactPathsLocked(t *Task) {
+	if t == nil || strings.TrimSpace(q.logsDir) == "" || strings.TrimSpace(t.ID) == "" {
+		return
+	}
+	if strings.TrimSpace(t.ArtifactsDir) == "" {
+		t.ArtifactsDir = taskArtifactsDir(q.logsDir, t, t.Status)
+	}
+	if t.ArtifactsDir != "" {
+		_ = os.MkdirAll(t.ArtifactsDir, 0755)
+	}
+	ensureTaskReportPaths(t)
+}
+
+func (q *taskQueue) finalizeTaskArtifactPathsLocked(t *Task) {
+	if t == nil || strings.TrimSpace(q.logsDir) == "" || strings.TrimSpace(t.ID) == "" {
+		return
+	}
+	q.ensureTaskArtifactPathsLocked(t)
+	dstDir := taskArtifactsDir(q.logsDir, t, t.Status)
+	if dstDir == "" {
+		return
+	}
+	if t.ArtifactsDir != "" && t.ArtifactsDir != dstDir {
+		if _, err := os.Stat(dstDir); err != nil {
+			_ = os.Rename(t.ArtifactsDir, dstDir)
+		}
+		t.ArtifactsDir = dstDir
+	}
+	ensureTaskReportPaths(t)
+}

audit/internal/webui/tasks_test.go

@@ -2,6 +2,7 @@ package webui
 
 import (
 	"context"
+	"encoding/json"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -12,6 +13,7 @@ import (
 	"time"
 
 	"bee/audit/internal/app"
+	"bee/audit/internal/platform"
 )
 
 func TestTaskQueuePersistsAndRecoversPendingTasks(t *testing.T) {
@@ -161,6 +163,40 @@ func TestTaskQueueSnapshotSortsNewestFirst(t *testing.T) {
 	}
 }
 
+func TestNewJobIDUsesTASKPrefixAndZeroPadding(t *testing.T) {
+	globalQueue.mu.Lock()
+	origTasks := globalQueue.tasks
+	globalQueue.tasks = nil
+	globalQueue.mu.Unlock()
+	origCounter := jobCounter.Load()
+	jobCounter.Store(0)
+	t.Cleanup(func() {
+		globalQueue.mu.Lock()
+		globalQueue.tasks = origTasks
+		globalQueue.mu.Unlock()
+		jobCounter.Store(origCounter)
+	})
+
+	if got := newJobID("ignored"); got != "TASK-000" {
+		t.Fatalf("id=%q want TASK-000", got)
+	}
+	if got := newJobID("ignored"); got != "TASK-001" {
+		t.Fatalf("id=%q want TASK-001", got)
+	}
+}
+
+func TestTaskArtifactsDirStartsWithTaskNumber(t *testing.T) {
+	root := t.TempDir()
+	task := &Task{
+		ID:   "TASK-007",
+		Name: "NVIDIA Benchmark",
+	}
+	got := filepath.Base(taskArtifactsDir(root, task, TaskDone))
+	if !strings.HasPrefix(got, "007_") {
+		t.Fatalf("artifacts dir=%q want prefix 007_", got)
+	}
+}
+
 func TestHandleAPITasksStreamReplaysPersistedLogWithoutLiveJob(t *testing.T) {
 	dir := t.TempDir()
 	logPath := filepath.Join(dir, "task.log")
@@ -248,6 +284,196 @@ func TestHandleAPITasksStreamPendingTaskStartsSSEImmediately(t *testing.T) {
 	t.Fatalf("stream did not emit queued status promptly, body=%q", rec.Body.String())
 }
 
+func TestFinalizeTaskRunCreatesReportFolderAndArtifacts(t *testing.T) {
+	dir := t.TempDir()
+	metricsPath := filepath.Join(dir, "metrics.db")
+	prevMetricsPath := taskReportMetricsDBPath
+	taskReportMetricsDBPath = metricsPath
+	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
+
+	db, err := openMetricsDB(metricsPath)
+	if err != nil {
+		t.Fatalf("openMetricsDB: %v", err)
+	}
+	base := time.Now().UTC().Add(-45 * time.Second)
+	if err := db.Write(platform.LiveMetricSample{
+		Timestamp:  base,
+		CPULoadPct: 42,
+		MemLoadPct: 35,
+		PowerW:     510,
+	}); err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+	_ = db.Close()
+
+	q := &taskQueue{
+		statePath: filepath.Join(dir, "tasks-state.json"),
+		logsDir:   filepath.Join(dir, "tasks"),
+		trigger:   make(chan struct{}, 1),
+	}
+	if err := os.MkdirAll(q.logsDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+
+	started := time.Now().UTC().Add(-90 * time.Second)
+	task := &Task{
+		ID:        "task-1",
+		Name:      "CPU SAT",
+		Target:    "cpu",
+		Status:    TaskRunning,
+		CreatedAt: started.Add(-10 * time.Second),
+		StartedAt: &started,
+	}
+	q.assignTaskLogPathLocked(task)
+	appendJobLog(task.LogPath, "line-1")
+
+	job := newTaskJobState(task.LogPath)
+	job.finish("")
+	q.finalizeTaskRun(task, job)
+
+	if task.Status != TaskDone {
+		t.Fatalf("status=%q want %q", task.Status, TaskDone)
+	}
+	if !strings.Contains(filepath.Base(task.ArtifactsDir), "_done") {
+		t.Fatalf("artifacts dir=%q", task.ArtifactsDir)
+	}
+	if _, err := os.Stat(task.ReportJSONPath); err != nil {
+		t.Fatalf("report json: %v", err)
+	}
+	if _, err := os.Stat(task.ReportHTMLPath); err != nil {
+		t.Fatalf("report html: %v", err)
+	}
+	var report taskReport
+	data, err := os.ReadFile(task.ReportJSONPath)
+	if err != nil {
+		t.Fatalf("ReadFile(report.json): %v", err)
+	}
+	if err := json.Unmarshal(data, &report); err != nil {
+		t.Fatalf("Unmarshal(report.json): %v", err)
+	}
+	if report.ID != task.ID || report.Status != TaskDone {
+		t.Fatalf("report=%+v", report)
+	}
+	if len(report.Charts) == 0 {
+		t.Fatalf("expected charts in report, got none")
+	}
+}
+
+func TestWriteTaskReportArtifactsIncludesBenchmarkResultsForTask(t *testing.T) {
+	dir := t.TempDir()
+	metricsPath := filepath.Join(dir, "metrics.db")
+	prevMetricsPath := taskReportMetricsDBPath
+	taskReportMetricsDBPath = metricsPath
+	t.Cleanup(func() { taskReportMetricsDBPath = prevMetricsPath })
+
+	benchmarkDir := filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000")
+	if err := os.MkdirAll(benchmarkDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	result := platform.NvidiaBenchmarkResult{
+		GeneratedAt:      time.Date(2026, time.April, 6, 12, 0, 0, 0, time.UTC),
+		BenchmarkProfile: "standard",
+		OverallStatus:    "OK",
+		GPUs: []platform.BenchmarkGPUResult{
+			{
+				Index: 0,
+				Name:  "NVIDIA H100 PCIe",
+				Scores: platform.BenchmarkScorecard{
+					CompositeScore: 1176.25,
+				},
+			},
+		},
+	}
+	raw, err := json.Marshal(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(benchmarkDir, "result.json"), raw, 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	artifactsDir := filepath.Join(dir, "tasks", "task-bench_done")
+	if err := os.MkdirAll(artifactsDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	task := &Task{
+		ID:           "task-bench",
+		Name:         "NVIDIA Benchmark",
+		Target:       "nvidia-benchmark",
+		Status:       TaskDone,
+		CreatedAt:    time.Now().UTC().Add(-time.Minute),
+		ArtifactsDir: artifactsDir,
+	}
+	ensureTaskReportPaths(task)
+	logText := "line-1\nArchive: " + filepath.Join(dir, "bee-benchmark", "gpu-benchmark-20260406-120000.tar.gz") + "\n"
+	if err := os.WriteFile(task.LogPath, []byte(logText), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := writeTaskReportArtifacts(task); err != nil {
+		t.Fatalf("writeTaskReportArtifacts: %v", err)
+	}
+
+	body, err := os.ReadFile(task.ReportHTMLPath)
+	if err != nil {
+		t.Fatalf("ReadFile(report.html): %v", err)
+	}
+	html := string(body)
+	for _, needle := range []string{
+		`Benchmark Results`,
+		`Composite score for this benchmark task.`,
+		`NVIDIA H100 PCIe / GPU 0`,
+		`1176.25`,
+	} {
+		if !strings.Contains(html, needle) {
+			t.Fatalf("report missing %q: %s", needle, html)
+		}
+	}
+}
+
+func TestTaskLifecycleMirrorsToSerialConsole(t *testing.T) {
+	var lines []string
+	prev := taskSerialWriteLine
+	taskSerialWriteLine = func(line string) { lines = append(lines, line) }
+	t.Cleanup(func() { taskSerialWriteLine = prev })
+
+	dir := t.TempDir()
+	q := &taskQueue{
+		statePath: filepath.Join(dir, "tasks-state.json"),
+		logsDir:   filepath.Join(dir, "tasks"),
+		trigger:   make(chan struct{}, 1),
+	}
+	task := &Task{
+		ID:        "task-serial-1",
+		Name:      "CPU SAT",
+		Target:    "cpu",
+		Status:    TaskPending,
+		CreatedAt: time.Now().UTC(),
+	}
+
+	q.enqueue(task)
+	started := time.Now().UTC()
+	task.Status = TaskRunning
+	task.StartedAt = &started
+	job := newTaskJobState(task.LogPath, taskSerialPrefix(task))
+	job.append("Starting CPU SAT...")
+	job.append("CPU stress duration: 60s")
+	job.finish("")
+	q.finalizeTaskRun(task, job)
+
+	joined := strings.Join(lines, "\n")
+	for _, needle := range []string{
+		"queued",
+		"Starting CPU SAT...",
+		"CPU stress duration: 60s",
+		"finished with status=done",
+	} {
+		if !strings.Contains(joined, needle) {
+			t.Fatalf("serial mirror missing %q in %q", needle, joined)
+		}
+	}
+}
+
 func TestResolveBurnPreset(t *testing.T) {
 	tests := []struct {
 		profile string

iso/builder/auto/config

@@ -32,7 +32,7 @@ lb config noauto \
     --memtest memtest86+ \
     --iso-volume "EASY_BEE_${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
     --iso-application "EASY-BEE-${BEE_GPU_VENDOR_UPPER:-NVIDIA}" \
-    --bootappend-live "boot=live components video=1920x1080 console=tty0 console=ttyS0,115200n8 loglevel=6 systemd.show_status=1 username=bee user-fullname=Bee modprobe.blacklist=nouveau,snd_hda_intel,snd_hda_codec_realtek,snd_hda_codec_generic,soundcore" \
+    --bootappend-live "boot=live components video=1920x1080 console=tty0 console=ttyS0,115200n8 loglevel=3 systemd.show_status=1 username=bee user-fullname=Bee modprobe.blacklist=nouveau,snd_hda_intel,snd_hda_codec_realtek,snd_hda_codec_generic,soundcore" \
     --apt-recommends false \
     --chroot-squashfs-compression-type zstd \
     "${@}"

iso/builder/config/bootloaders/grub-pc/grub.cfg

@@ -7,6 +7,7 @@ echo "  █████╗  ███████║███████╗ ╚
 echo "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝"
 echo "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗"
 echo "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝"
+echo "  Hardware Audit LiveCD"
 echo ""
 
 menuentry "EASY-BEE" {
@@ -14,29 +15,21 @@ menuentry "EASY-BEE" {
     initrd  @INITRD_LIVE@
 }
 
-menuentry "EASY-BEE (graphics/KMS)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
+submenu "EASY-BEE (advanced options) -->" {
+    menuentry "EASY-BEE — GSP=off" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        initrd  @INITRD_LIVE@
+    }
 
-menuentry "EASY-BEE (load to RAM)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ toram nomodeset bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
+    menuentry "EASY-BEE — KMS (no nomodeset)" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=normal net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
+        initrd  @INITRD_LIVE@
+    }
 
-menuentry "EASY-BEE (NVIDIA GSP=off)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
-
-menuentry "EASY-BEE (graphics/KMS, GSP=off)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.display=kms bee.nvidia.mode=gsp-off net.ifnames=0 biosdevname=0 mitigations=off transparent_hugepage=always numa_balancing=disable nowatchdog nosoftlockup
-    initrd  @INITRD_LIVE@
-}
-
-menuentry "EASY-BEE (fail-safe)" {
-    linux   @KERNEL_LIVE@ @APPEND_LIVE@ bee.nvidia.mode=gsp-off memtest noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
-    initrd  @INITRD_LIVE@
+    menuentry "EASY-BEE — fail-safe" {
+        linux   @KERNEL_LIVE@ @APPEND_LIVE@ nomodeset bee.nvidia.mode=gsp-off noapic noapm nodma nomce nolapic nosmp vga=normal net.ifnames=0 biosdevname=0
+        initrd  @INITRD_LIVE@
+    }
 }
 
 if [ "${grub_platform}" = "efi" ]; then

iso/builder/config/hooks/normal/9000-bee-setup.hook.chroot

@@ -31,6 +31,7 @@ systemctl enable bee-audit.service
 systemctl enable bee-web.service
 systemctl enable bee-sshsetup.service
 systemctl enable bee-selfheal.timer
+systemctl enable bee-boot-status.service
 systemctl enable ssh.service
 systemctl enable lightdm.service 2>/dev/null || true
 systemctl enable qemu-guest-agent.service 2>/dev/null || true
@@ -59,7 +60,8 @@ chmod +x /usr/local/bin/bee-sshsetup   2>/dev/null || true
 chmod +x /usr/local/bin/bee-smoketest  2>/dev/null || true
 chmod +x /usr/local/bin/bee            2>/dev/null || true
 chmod +x /usr/local/bin/bee-log-run    2>/dev/null || true
-chmod +x /usr/local/bin/bee-selfheal   2>/dev/null || true
+chmod +x /usr/local/bin/bee-selfheal      2>/dev/null || true
+chmod +x /usr/local/bin/bee-boot-status  2>/dev/null || true
 if [ "$GPU_VENDOR" = "nvidia" ]; then
     chmod +x /usr/local/bin/bee-nvidia-load 2>/dev/null || true
     chmod +x /usr/local/bin/bee-gpu-burn 2>/dev/null || true

iso/builder/config/hooks/normal/9001-wallpaper.hook.chroot

@@ -0,0 +1,76 @@
+#!/bin/sh
+# 9001-wallpaper.hook.chroot — generate /usr/share/bee/wallpaper.png inside chroot
+set -e
+echo "=== generating bee wallpaper ==="
+mkdir -p /usr/share/bee
+
+python3 - <<'PYEOF'
+from PIL import Image, ImageDraw, ImageFont
+import os
+
+W, H = 1920, 1080
+
+LOGO = """\
+  \u2588\u2588\u2588\u2588\u2588\u2588\u2557 \u2588\u2588\u2588\u2588\u2588\u2557 \u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2557   \u2588\u2588\u2557      \u2588\u2588\u2588\u2588\u2588\u2588\u2557 \u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557
+  \u2588\u2588\u2554\u2550\u2550\u2550\u2550\u255d\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557\u2588\u2588\u2554\u2550\u2550\u2550\u2550\u255d\u255a\u2588\u2588\u2557 \u2588\u2588\u2554\u255d      \u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557\u2588\u2588\u2554\u2550\u2550\u2550\u2550\u255d\u2588\u2588\u2554\u2550\u2550\u2550\u2550\u255d
+  \u2588\u2588\u2588\u2588\u2588\u2557  \u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2551\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557 \u255a\u2588\u2588\u2588\u2588\u2554\u255d \u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2588\u2588\u2554\u255d\u2588\u2588\u2588\u2588\u2588\u2557  \u2588\u2588\u2588\u2588\u2588\u2557
+  \u2588\u2588\u2554\u2550\u2550\u255d  \u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2551\u255a\u2550\u2550\u2550\u2550\u2588\u2588\u2551  \u255a\u2588\u2588\u2554\u255d  \u255a\u2550\u2550\u2550\u2550\u255d\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557\u2588\u2588\u2554\u2550\u2550\u255d  \u2588\u2588\u2554\u2550\u2550\u255d
+  \u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2551  \u2588\u2588\u2551\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2551   \u2588\u2588\u2551         \u2588\u2588\u2588\u2588\u2588\u2588\u2554\u255d\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557
+  \u255a\u2550\u2550\u2550\u2550\u2550\u2550\u255d\u255a\u2550\u255d  \u255a\u2550\u255d\u255a\u2550\u2550\u2550\u2550\u2550\u2550\u255d   \u255a\u2550\u255d         \u255a\u2550\u2550\u2550\u2550\u2550\u255d \u255a\u2550\u2550\u2550\u2550\u2550\u2550\u255d\u255a\u2550\u2550\u2550\u2550\u2550\u2550\u255d
+  Hardware Audit LiveCD"""
+
+# Find a monospace font that supports box-drawing characters
+FONT_CANDIDATES = [
+    '/usr/share/fonts/truetype/dejavu/DejaVuSansMono.ttf',
+    '/usr/share/fonts/truetype/liberation/LiberationMono-Regular.ttf',
+    '/usr/share/fonts/truetype/freefont/FreeMono.ttf',
+    '/usr/share/fonts/truetype/noto/NotoMono-Regular.ttf',
+]
+
+font_path = None
+for p in FONT_CANDIDATES:
+    if os.path.exists(p):
+        font_path = p
+        break
+
+SIZE = 22
+if font_path:
+    font_logo = ImageFont.truetype(font_path, SIZE)
+    font_sub  = ImageFont.truetype(font_path, SIZE)
+else:
+    font_logo = ImageFont.load_default()
+    font_sub  = font_logo
+
+img  = Image.new('RGB', (W, H), (0, 0, 0))
+draw = ImageDraw.Draw(img)
+
+# Measure logo block line by line to avoid font ascender offset
+lines = LOGO.split('\n')
+logo_lines = lines[:6]
+sub_line   = lines[6] if len(lines) > 6 else ''
+
+line_h = SIZE + 2
+block_h = len(logo_lines) * line_h + 8 + (SIZE if sub_line else 0)
+
+# Width: measure the widest logo line
+max_w = 0
+for line in logo_lines:
+    bb = draw.textbbox((0, 0), line, font=font_logo)
+    max_w = max(max_w, bb[2] - bb[0])
+
+x = (W - max_w) // 2
+y = (H - block_h) // 2
+
+cy = y
+for line in logo_lines:
+    draw.text((x, cy), line, font=font_logo, fill=(0xf6, 0xc9, 0x0e))
+    cy += line_h
+cy += 8
+if sub_line:
+    draw.text((x, cy), sub_line, font=font_sub, fill=(0x80, 0x68, 0x18))
+
+img.save('/usr/share/bee/wallpaper.png', optimize=True)
+print('wallpaper written: /usr/share/bee/wallpaper.png')
+PYEOF
+
+echo "=== wallpaper done ==="

iso/builder/config/package-lists/bee.list.chroot

@@ -60,9 +60,15 @@ qrencode
 # Local desktop (openbox + chromium kiosk)
 openbox
 tint2
+feh
+python3-pil
 xorg
 xterm
 chromium
+mousepad
+pcmanfm
+ristretto
+mupdf
 xserver-xorg-video-fbdev
 xserver-xorg-video-vesa
 lightdm

iso/overlay/etc/systemd/system/bee-boot-status.service

@@ -0,0 +1,18 @@
+[Unit]
+Description=Bee: boot status display
+After=systemd-user-sessions.service
+Before=getty@tty1.service
+
+[Service]
+Type=oneshot
+RemainAfterExit=no
+ExecStart=/usr/local/bin/bee-boot-status
+TTYPath=/dev/tty1
+StandardInput=tty
+StandardOutput=tty
+StandardError=tty
+TTYReset=yes
+TTYVHangup=yes
+
+[Install]
+WantedBy=multi-user.target

iso/overlay/etc/systemd/system/getty@tty1.service.d/wait-bee.conf

@@ -0,0 +1,2 @@
+[Unit]
+After=bee-boot-status.service

iso/overlay/etc/systemd/system/lightdm.service.d/bee-display-mode.conf

@@ -1,6 +1,4 @@
 [Unit]
-Wants=bee-preflight.service
-After=bee-preflight.service
 
 [Service]
 ExecStartPre=/usr/local/bin/bee-display-mode

iso/overlay/usr/local/bin/bee-boot-status

@@ -0,0 +1,89 @@
+#!/bin/sh
+# bee-boot-status — boot progress display on tty1.
+# Shows live service status until all bee services are done or failed,
+# then exits so getty can show the login prompt.
+
+CRITICAL="bee-preflight bee-nvidia bee-audit"
+ALL="bee-sshsetup ssh bee-network bee-nvidia bee-preflight bee-audit bee-web"
+
+svc_state() { systemctl is-active "$1.service" 2>/dev/null || echo "inactive"; }
+
+svc_icon() {
+    case "$(svc_state "$1")" in
+        active)       printf '\033[32m[  OK  ]\033[0m' ;;
+        failed)       printf '\033[31m[ FAIL ]\033[0m' ;;
+        activating)   printf '\033[33m[  ..  ]\033[0m' ;;
+        deactivating) printf '\033[33m[ stop ]\033[0m' ;;
+        inactive)     printf '\033[90m[      ]\033[0m' ;;
+        *)            printf '\033[90m[  ?   ]\033[0m' ;;
+    esac
+}
+
+svc_detail() {
+    local svc="$1" state
+    state="$(svc_state "$svc")"
+    case "$state" in
+        failed)
+            local res
+            res="$(systemctl show -p Result "$svc.service" 2>/dev/null | cut -d= -f2)"
+            [ -n "$res" ] && [ "$res" != "success" ] && printf '  \033[31m(%s)\033[0m' "$res"
+            ;;
+        activating)
+            local line
+            line="$(journalctl -u "$svc.service" -n 1 --no-pager --output=cat 2>/dev/null | cut -c1-55)"
+            [ -n "$line" ] && printf '  \033[90m%s\033[0m' "$line"
+            ;;
+    esac
+}
+
+all_critical_done() {
+    for svc in $CRITICAL; do
+        case "$(svc_state "$svc")" in
+            active|failed|inactive) ;;
+            *) return 1 ;;
+        esac
+    done
+    return 0
+}
+
+while true; do
+    # move to top-left and clear screen
+    printf '\033[H\033[2J'
+
+    printf '\n'
+    printf '  \033[33m███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗\033[0m\n'
+    printf '  \033[33m██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝\033[0m\n'
+    printf '  \033[33m█████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗\033[0m\n'
+    printf '  \033[33m██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝\033[0m\n'
+    printf '  \033[33m███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗\033[0m\n'
+    printf '  \033[33m╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝\033[0m\n'
+    printf '  Hardware Audit LiveCD\n'
+    printf '\n'
+
+    for svc in $ALL; do
+        printf '  %s  %-20s%s\n' "$(svc_icon "$svc")" "$svc" "$(svc_detail "$svc")"
+    done
+    printf '\n'
+
+    # Network
+    ips="$(ip -4 addr show scope global 2>/dev/null | awk '/inet /{printf "  %-16s %s\n", $NF, $2}')"
+    if [ -n "$ips" ]; then
+        printf '  \033[1mNetwork:\033[0m\n'
+        printf '%s\n' "$ips"
+        printf '\n'
+    fi
+
+    if all_critical_done; then
+        printf '  \033[1;32mSystem ready.\033[0m  Audit is running in the background.\n'
+        first_ip="$(ip -4 addr show scope global 2>/dev/null | awk '/inet /{print $2}' | cut -d/ -f1 | head -1)"
+        if [ -n "$first_ip" ]; then
+            printf '  Web UI: \033[1mhttp://%s/\033[0m\n' "$first_ip"
+        fi
+        printf '\n'
+        sleep 3
+        break
+    fi
+
+    printf '  \033[90mStarting up...\033[0m\n'
+    sleep 3
+done

iso/overlay/usr/local/bin/bee-gpu-burn

@@ -62,6 +62,8 @@ done
 echo "loader=bee-gpu-burn"
 echo "selected_gpus=${FINAL}"
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+
 TMP_DIR=$(mktemp -d)
 trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
 
@@ -78,7 +80,8 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
         fi
     fi
     echo "starting gpu ${id} size=${gpu_size_mb}MB"
-    "${WORKER}" --device "${id}" --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
+    CUDA_VISIBLE_DEVICES="${id}" \
+        "${WORKER}" --device 0 --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
     pid=$!
     WORKERS="${WORKERS} ${pid}:${id}:${log}"
 done

iso/overlay/usr/local/bin/bee-john-gpu-stress

@@ -152,14 +152,19 @@ done
 
 [ -n "${FINAL}" ] || { echo "no NVIDIA GPUs selected after filters" >&2; exit 1; }
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+export CUDA_VISIBLE_DEVICES="${FINAL}"
+
 JOHN_DEVICES=""
+local_id=1
 for id in $(echo "${FINAL}" | tr ',' ' '); do
-    opencl_id=$((id + 1))
+    opencl_id="${local_id}"
     if [ -z "${JOHN_DEVICES}" ]; then
         JOHN_DEVICES="${opencl_id}"
     else
         JOHN_DEVICES="${JOHN_DEVICES},${opencl_id}"
     fi
+    local_id=$((local_id + 1))
 done
 
 echo "loader=john"

iso/overlay/usr/local/bin/bee-nccl-gpu-stress

@@ -70,6 +70,8 @@ echo "gpu_count=${GPU_COUNT}"
 echo "range=${MIN_BYTES}..${MAX_BYTES}"
 echo "iters=${ITERS}"
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+
 deadline=$(( $(date +%s) + SECONDS ))
 round=0
 

iso/overlay/usr/local/bin/bee-nvidia-load

@@ -50,11 +50,93 @@ load_module() {
         log "WARN: not found: $ko"
         return 1
     fi
-    if insmod "$ko" "$@"; then
+    if timeout 90 insmod "$ko" "$@"; then
         log "loaded: $mod $*"
         return 0
     fi
-    log "WARN: failed to load: $mod"
+    log "WARN: failed to load: $mod (exit $?)"
+    dmesg | tail -n 10 | sed 's/^/  dmesg: /' || true
+    return 1
+}
+
+nvidia_is_functional() {
+    grep -q ' nvidiactl$' /proc/devices 2>/dev/null
+}
+
+load_module_with_gsp_fallback() {
+    ko="$NVIDIA_KO_DIR/nvidia.ko"
+    if [ ! -f "$ko" ]; then
+        log "ERROR: not found: $ko"
+        return 1
+    fi
+
+    # Run insmod in background — on some converted SXM→PCIe cards GSP enters an
+    # infinite crash/reload loop and insmod never returns. We check for successful
+    # initialization by polling /proc/devices for nvidiactl instead of waiting for
+    # insmod to exit.
+    log "loading nvidia (GSP enabled, timeout 90s)"
+    insmod "$ko" &
+    _insmod_pid=$!
+
+    _waited=0
+    while [ $_waited -lt 90 ]; do
+        if nvidia_is_functional; then
+            log "loaded: nvidia (GSP enabled, ${_waited}s)"
+            echo "gsp-on" > /run/bee-nvidia-mode
+            return 0
+        fi
+        # Check if insmod exited with an error before timeout
+        if ! kill -0 "$_insmod_pid" 2>/dev/null; then
+            wait "$_insmod_pid"
+            _rc=$?
+            if [ $_rc -ne 0 ]; then
+                log "nvidia load failed (exit $_rc)"
+                dmesg | tail -n 10 | sed 's/^/  dmesg: /' || true
+                return 1
+            fi
+            # insmod exited 0 but nvidiactl not yet in /proc/devices — give it a moment
+            sleep 2
+            if nvidia_is_functional; then
+                log "loaded: nvidia (GSP enabled, ${_waited}s)"
+                return 0
+            fi
+            log "insmod exited 0 but nvidiactl missing — treating as failure"
+            return 1
+        fi
+        sleep 1
+        _waited=$((_waited + 1))
+    done
+
+    # GSP init timed out — kill the hanging insmod and attempt gsp-off fallback
+    log "nvidia GSP init timed out after 90s"
+    kill "$_insmod_pid" 2>/dev/null || true
+    wait "$_insmod_pid" 2>/dev/null || true
+
+    # Attempt to unload the partially-initialized module
+    if ! rmmod nvidia 2>/dev/null; then
+        # Module is stuck in the kernel — cannot reload with different params.
+        # User must reboot and select bee.nvidia.mode=gsp-off at boot menu.
+        log "ERROR: rmmod nvidia failed (EBUSY) — module stuck in kernel"
+        log "ERROR: reboot and select 'EASY-BEE (advanced) -> GSP=off' in boot menu"
+        echo "gsp-stuck" > /run/bee-nvidia-mode
+        return 1
+    fi
+
+    sleep 2
+    log "retrying with NVreg_EnableGpuFirmware=0"
+    log "WARNING: GSP disabled — power management will run via CPU path, not GPU firmware"
+
+    if insmod "$ko" NVreg_EnableGpuFirmware=0; then
+        if nvidia_is_functional; then
+            log "loaded: nvidia (GSP disabled)"
+            echo "gsp-off" > /run/bee-nvidia-mode
+            return 0
+        fi
+        log "insmod gsp-off exited 0 but nvidiactl missing"
+        return 1
+    fi
+
+    log "nvidia load failed (GSP=off)"
     dmesg | tail -n 10 | sed 's/^/  dmesg: /' || true
     return 1
 }
@@ -70,7 +152,7 @@ load_host_module() {
 
 case "$nvidia_mode" in
     normal|full)
-        if ! load_module nvidia; then
+        if ! load_module_with_gsp_fallback; then
             exit 1
         fi
         # nvidia-modeset on some server kernels needs ACPI video helper symbols
@@ -127,6 +209,18 @@ fi
 ldconfig 2>/dev/null || true
 log "ldconfig refreshed"
 
+# Keep persistence mode enabled across the session so dcgmi / stress tools do
+# not fail with deployment warnings on otherwise healthy GPUs.
+if command -v nvidia-smi >/dev/null 2>&1; then
+    if nvidia-smi -pm 1 >/dev/null 2>&1; then
+        log "enabled NVIDIA persistence mode"
+    else
+        log "WARN: failed to enable NVIDIA persistence mode"
+    fi
+else
+    log "WARN: nvidia-smi not found — cannot enable persistence mode"
+fi
+
 # Start DCGM host engine so dcgmi can discover GPUs.
 # nv-hostengine must run after the NVIDIA modules and device nodes are ready.
 # If it started too early (for example via systemd before bee-nvidia-load), it can

iso/overlay/usr/local/bin/bee-openbox-session

@@ -7,16 +7,24 @@ xset s off
 xset -dpms
 xset s noblank
 
+# Set desktop background.
+if [ -f /usr/share/bee/wallpaper.png ]; then
+    feh --bg-fill /usr/share/bee/wallpaper.png
+else
+    xsetroot -solid '#f6c90e'
+fi
+
 tint2 &
 
-# Wait up to 120s for bee-web to bind. The web server starts immediately now
-# (audit is deferred), so this should succeed in a few seconds on most hardware.
-i=0
-while [ $i -lt 120 ]; do
-    if curl -sf http://localhost/healthz >/dev/null 2>&1; then break; fi
+# Wait up to 60s for bee-web before opening Chromium.
+# Without this Chromium gets connection-refused and shows a blank page.
+_i=0
+while [ $_i -lt 60 ]; do
+    curl -sf http://localhost/healthz >/dev/null 2>&1 && break
     sleep 1
-    i=$((i+1))
+    _i=$((_i+1))
 done
+unset _i
 
 chromium \
     --disable-infobars \
@@ -24,7 +32,8 @@ chromium \
     --no-first-run \
     --disable-session-crashed-bubble \
     --disable-features=TranslateUI \
+    --user-data-dir=/tmp/bee-chrome \
     --start-maximized \
-    http://localhost/ &
+    http://localhost/loading &
 
 exec openbox