fix(doctor): per-OS compute remedies + resources-set-max drift nudge (#400)

internal/cli/doctor.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"net/http"
 	"os"
+	"runtime"
 	"strings"
 	"time"
 
@@ -297,6 +298,24 @@ func tokenLabel(tok tokenState) string {
 // stays in --verbose. When the owner isn't connected — for ANY reason — a
 // healthy local cluster still can't run training, so readiness degrades honestly
 // to "can't check" rather than showing a reassuring ✔ next to a Connected ✖.
+// computeRemedy is the "not enough compute" remedy for the host OS (#400). On
+// Windows the default Docker backend is WSL2, where Docker Desktop has NO
+// Resources→Memory slider — the VM's memory comes from %UserProfile%\.wslconfig.
+// macOS keeps the slider; bare Linux has no Docker Desktop at all. Every
+// variant ends with the drift fix: size runs to the machine (backend#1236's
+// install-time auto-size can go stale when a machine shrinks).
+func computeRemedy(goos string) string {
+	resize := fmt.Sprintf("size runs to this machine with `%s resources set max`", launcher())
+	switch goos {
+	case "windows":
+ return fmt.Sprintf("Free some up, give Docker more memory (WSL2 backend: `[wsl2] memory=…` in %%UserProfile%%\\.wslconfig then `wsl --shutdown`; Hyper-V backend: Docker Desktop → Settings → Resources → Advanced), or %s.", resize)
+	case "darwin":
+ return fmt.Sprintf("Free some up, raise the machine's allocation in Docker Desktop → Settings → Resources → Advanced, or %s.", resize)
+	default:
+ return fmt.Sprintf("Free some up on this machine, or %s.", resize)
+	}
+}
+
 func summarizeDoctor(results []doctor.Result, tok tokenState) (connected, ready healthLine) {
 	by := make(map[string]doctor.Result, len(results))
 	for _, r := range results {
@@ -370,7 +389,7 @@ func summarizeDoctor(results []doctor.Result, tok tokenState) (connected, ready
  // the "Everything looks good" verdict (Bugbot).
  ready = healthLine{doctor.StatusFail,
  "Not ready — part of your secure environment can't start yet.",
- fmt.Sprintf("Some pods are stuck starting — usually not enough free compute, or a training image that can't be pulled. Free some up in Docker Desktop → Resources, then re-run `%s doctor`; if it persists, email support@tracebloc.io with `%s doctor --diagnose`.", launcher(), launcher())}
+ fmt.Sprintf("Some pods are stuck starting — usually not enough free compute, or a training image that can't be pulled. %s Then re-run `%s doctor`; if it persists, email support@tracebloc.io with `%s doctor --diagnose`.", computeRemedy(runtime.GOOS), launcher(), launcher())}
 	case by["Image pull secret"].Status == doctor.StatusFail:
  ready = healthLine{doctor.StatusFail,
  "Not ready — the training images can't be pulled.",
@@ -394,7 +413,7 @@ func summarizeDoctor(results []doctor.Result, tok tokenState) (connected, ready
 	case by["Node capacity"].Status == doctor.StatusFail:
  ready = healthLine{doctor.StatusFail,
  "Not ready — not enough free compute to start a training.",
-"Free some up, or raise the machine's allocation in Docker Desktop → Resources."}
+computeRemedy(runtime.GOOS)}
 	default:
  ready = healthLine{doctor.StatusOK, "Ready to run training", ""}
 	}

internal/cli/doctor_test.go

@@ -311,8 +311,33 @@ func TestSummarizeDoctor(t *testing.T) {
 
 	t.Run("no compute → ready Fail", func(t *testing.T) {
  _, r := summarizeDoctor(with(allOK, "Node capacity", doctor.StatusFail), tokenOK)
- if r.status != doctor.StatusFail || !strings.Contains(r.remedy, "Docker Desktop") {
- t.Errorf("want ready Fail with a raise-allocation remedy, got %v remedy=%q", r.status, r.remedy)
+ // GOOS-independent invariant: every variant carries the resize fix
+ // (#400 — the old string pinned "Docker Desktop", wrong on WSL2/linux).
+ if r.status != doctor.StatusFail || !strings.Contains(r.remedy, "resources set max") {
+ t.Errorf("want ready Fail with a resize remedy, got %v remedy=%q", r.status, r.remedy)
+ }
+	})
+
+	// computeRemedy (#400): the compute remedy must match the host's actual
+	// memory lever — Docker Desktop's Resources slider does not exist on the
+	// WSL2 backend, and bare Linux has no Docker Desktop at all.
+	t.Run("computeRemedy per GOOS", func(t *testing.T) {
+ win := computeRemedy("windows")
+ if !strings.Contains(win, ".wslconfig") || !strings.Contains(win, "Hyper-V") {
+ t.Fatalf("windows remedy must name both levers: %q", win)
+ }
+ mac := computeRemedy("darwin")
+ if !strings.Contains(mac, "Docker Desktop → Settings → Resources → Advanced") {
+ t.Fatalf("darwin remedy keeps the slider: %q", mac)
+ }
+ lin := computeRemedy("linux")
+ if strings.Contains(lin, "Docker Desktop") || strings.Contains(lin, "wslconfig") {
+ t.Fatalf("linux remedy must not name Docker Desktop/WSL: %q", lin)
+ }
+ for _, r := range []string{win, mac, lin} {
+ if !strings.Contains(r, "resources set max") {
+ t.Fatalf("every remedy carries the resize fix: %q", r)
+ }
  }
 	})
 

internal/cli/testdata/golden/zz-all-strings.golden

@@ -165,7 +165,9 @@ screen. %s/%d are runtime placeholders.
 "Follow it later with: kubectl logs -f -n %s job/%s"
 "For help: https://docs.tracebloc.io/create-use-case/prepare-dataset"
 "Found labels.csv and a %s folder — this looks like text data."
-"Free some up, or raise the machine's allocation in Docker Desktop → Resources."
+"Free some up on this machine, or %s."
+"Free some up, give Docker more memory (WSL2 backend: `[wsl2] memory=…` in %%UserProfile%%\\.wslconfig then `wsl --shutdown`; Hyper-V backend: Docker Desktop → Settings → Resources → Advanced), or %s."
+"Free some up, raise the machine's allocation in Docker Desktop → Settings → Resources → Advanced, or %s."
 "Full log: %s"
 "GPU access removed — training runs will use CPU only."
 "Give the path to a file or a folder — whichever holds your data:"
@@ -275,7 +277,7 @@ screen. %s/%d are runtime placeholders.
 "Signed out locally, but couldn't revoke the token server-side (%v). Revoke from the dashboard if this was a shared machine."
 "Signed out."
 "Signed-in token was rejected by the backend — run `tracebloc login`."
-"Some pods are stuck starting — usually not enough free compute, or a training image that can't be pulled. Free some up in Docker Desktop → Resources, then re-run `%s doctor`; if it persists, email support@tracebloc.io with `%s doctor --diagnose`."
+"Some pods are stuck starting — usually not enough free compute, or a training image that can't be pulled. %s Then re-run `%s doctor`; if it persists, email support@tracebloc.io with `%s doctor --diagnose`."
 "Some tracebloc images couldn't be reclaimed (harmless) — remove them later with `docker rmi $(docker images --filter=reference='ghcr.io/tracebloc/*' --format '{{.Repository}}:{{.Tag}}')`."
 "Still stuck? Email support@tracebloc.io with the output of `%s doctor --diagnose`."
 "Stopped following after 1 hour — the ingestion is still running and will finish on its own."
@@ -565,6 +567,7 @@ screen. %s/%d are runtime placeholders.
 "shared PVC: %s (%s)"
 "sign-in was denied in the browser"
 "signed in"
+"size runs to this machine with `%s resources set max`"
 "skipped"
 "source"
 "stage Pod %s/%s did not become Ready within %s%s"
@@ -627,3 +630,4 @@ screen. %s/%d are runtime placeholders.
 "~%s (requested; server may cap shorter)"
 "· %d GPU"
 "· %d classes"
+"— this machine could give a run up to cpu=%d,memory=%dGi ('tracebloc resources set max')"

internal/doctor/doctor.go

@@ -1,3 +1,3 @@
 // Package doctor implements the checks behind `tracebloc cluster doctor`:
 // a read-only, best-effort health sweep of a running tracebloc client
 // cluster. Each check reports ✔/⚠/✖ plus a one-line remedy, so a customer
@@ -30,6 +30,7 @@
 	"k8s.io/client-go/kubernetes"
 
 	"github.com/tracebloc/cli/internal/cluster"
+	"github.com/tracebloc/cli/internal/resources"
 )
 
 // Status is a single check's severity. Ordered so the numerically-greatest
@@ -596,12 +597,21 @@
 	// node as a whole — never OR cpu/mem and GPU across different nodes, which
 	// would pass even when no single node can run the job (Bugbot on PR #91).
 	var cpuMemFits, fullFits bool
+	// Largest Ready node for the drift nudge — CPU-major with memory as the
+	// tie-break, EXACTLY like resources.nodeLarger, so the advertised ceiling
+	// always matches what `resources set max` will actually apply (Bugbot).
+	var bestCPU, bestMem resource.Quantity
 	for i := range nodes.Items {
  n := nodes.Items[i]
  if !nodeReady(n) {
  continue
  }
  alloc := n.Status.Allocatable
+ if alloc.Cpu().Cmp(bestCPU) > 0 ||
+ (alloc.Cpu().Cmp(bestCPU) == 0 && alloc.Memory().Cmp(bestMem) > 0) {
+ bestCPU = *alloc.Cpu()
+ bestMem = *alloc.Memory()
+ }
  nodeCPUMem := alloc.Cpu().Cmp(cpuReq) >= 0 && alloc.Memory().Cmp(memReq) >= 0
  nodeGPU := !gpuRequested
  if gpuRequested {
@@ -633,7 +643,19 @@
  Remedy: "If GPU training is expected, ensure one node has both the compute and the GPU capacity, with its device plugin.",
  }
 	default:
- return Result{Name: name, Status: StatusOK, Detail: fmt.Sprintf("a Ready node can schedule a training job (%s)", req)}
+ detail := fmt.Sprintf("a Ready node can schedule a training job (%s)", req)
+ // Drift nudge (#400 / backend#1236): the install-time auto-size goes
+ // stale when a machine GROWS. When the configured budget uses no more
+ // than half of what this machine could give one run (largest node −
+ // platform overhead), say so.
+ m := resources.Machine{CPU: bestCPU, Mem: bestMem}
+ maxCores, maxGiB := resources.MaxRunCores(m), resources.MaxRunGiB(m)
+ if maxCores >= 1 && maxGiB >= 2 &&
+ cpuReq.MilliValue()*2 <= int64(maxCores)*1000 &&
+ memReq.Value()*2 <= int64(maxGiB)<<30 {
+ detail += fmt.Sprintf(" — this machine could give a run up to cpu=%d,memory=%dGi ('tracebloc resources set max')", maxCores, maxGiB)
+ }
+ return Result{Name: name, Status: StatusOK, Detail: detail}
 	}
 }
 

internal/doctor/doctor_test.go

@@ -657,6 +657,36 @@ func TestCheckNodeFit(t *testing.T) {
  t.Fatalf("=> %v (%q), want warn", r.Status, r.Detail)
  }
 	})
+	// Drift nudge (#400 / backend#1236): a budget using ≤ half of what the
+	// largest node could give one run gets a resources-set-max pointer; a snug
+	// fit stays quiet.
+	t.Run("big machine + small budget -> ok with the resize nudge", func(t *testing.T) {
+ cs := fake.NewClientset(node("n1", "32", "64Gi"))
+ r := checkNodeFit(bg(), cs, cpuOnly) // 2/8Gi vs max 31/61
+ if r.Status != StatusOK || !strings.Contains(r.Detail, "resources set max") {
+ t.Fatalf("=> %v (%q), want ok with nudge", r.Status, r.Detail)
+ }
+	})
+	t.Run("snug fit -> ok without the nudge", func(t *testing.T) {
+ cs := fake.NewClientset(node("n1", "4", "12Gi"))
+ r := checkNodeFit(bg(), cs, cpuOnly) // 2/8Gi vs max 3/9: memory over half
+ if r.Status != StatusOK || strings.Contains(r.Detail, "resources set max") {
+ t.Fatalf("=> %v (%q), want ok without nudge", r.Status, r.Detail)
+ }
+	})
+	t.Run("heterogeneous nodes: nudge quotes the CPU-major node, matching set max (Bugbot)", func(t *testing.T) {
+ // resources.LargestReadyNode (what `set max` applies) is CPU-major:
+ // it picks cpuBig (32/64Gi -> max 31/61Gi), not memBig (8/128Gi ->
+ // 7/125Gi). The advertised ceiling must be the SAME node's.
+ cs := fake.NewClientset(
+ node("cpuBig", "32", "64Gi"),
+ node("memBig", "8", "128Gi"),
+ )
+ r := checkNodeFit(bg(), cs, cpuOnly)
+ if r.Status != StatusOK || !strings.Contains(r.Detail, "cpu=31,memory=61Gi") {
+ t.Fatalf("=> %v (%q), want the CPU-major node's ceiling cpu=31,memory=61Gi", r.Status, r.Detail)
+ }
+	})
 }
 
 func dockerSecret(name string, data []byte) *corev1.Secret {