llmgw/internal/web/web_test.go

package web

import (
	"context"
	"errors"
	"net/http"
	"net/http/httptest"
	"strings"
	"testing"
	"time"

	"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
	"gitea.mynodes.xyz/ilker/llmgw/internal/store"
)

func get(t *testing.T, h http.Handler, target string, headers map[string]string) (int, string) {
	t.Helper()
	r := httptest.NewRequest(http.MethodGet, target, nil)
	for k, v := range headers {
		r.Header.Set(k, v)
	}
	w := httptest.NewRecorder()
	h.ServeHTTP(w, r)
	return w.Code, w.Body.String()
}

func TestLandingPageTellsYouHowToConnect(t *testing.T) {
	code, body := get(t, Landing("v1.2.3", true, false, "", ""), "http://llmgw.example/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d", code)
	}
	for _, want := range []string{"http://llmgw.example/v1", "Authorization: Bearer", "v1.2.3"} {
		if !strings.Contains(body, want) {
			t.Errorf("the page does not mention %q", want)
		}
	}
}

func TestLandingPageSaysNothingAboutWhatIsHere(t *testing.T) {
	// It is served without a key, to anyone who can reach the address. Anything
	// on it is public, so nothing about the cluster belongs on it.
	_, body := get(t, Landing("v1", true, false, "", ""), "http://llmgw.example/", nil)
	for _, leak := range []string{"llama", "qwen", "10.0.0.", "replica", "node01"} {
		if strings.Contains(strings.ToLower(body), leak) {
			t.Errorf("the landing page leaks %q", leak)
		}
	}
}

func TestLandingPageWarnsWhenTheGatewayIsOpen(t *testing.T) {
	// Someone who reaches this page on an unauthenticated gateway should learn
	// that from the page rather than by discovering they can use it.
	_, open := get(t, Landing("v1", false, false, "", ""), "http://llmgw.example/", nil)
	if !strings.Contains(open, "no API keys configured") {
		t.Error("an open gateway does not say so")
	}
	_, closed := get(t, Landing("v1", true, false, "", ""), "http://llmgw.example/", nil)
	if strings.Contains(closed, "no API keys configured") {
		t.Error("a closed gateway claims to be open")
	}
}

func TestTheExampleURLSurvivesAReverseProxy(t *testing.T) {
	// The process sees http and an internal host; the client used https and a
	// public name. An example naming the internal address is one nobody can
	// paste anywhere.
	_, body := get(t, Landing("v1", true, false, "", ""), "http://10.0.0.140:8080/", map[string]string{
		"X-Forwarded-Proto": "https",
		"X-Forwarded-Host":  "llmgw.mynodes.xyz",
	})
	if !strings.Contains(body, "https://llmgw.mynodes.xyz/v1") {
		t.Error("the example does not use the address the client actually used")
	}
	if strings.Contains(body, "10.0.0.140") {
		t.Error("the internal address leaked into the page")
	}
}

func TestForwardedHeaderChainsTakeTheFirstEntry(t *testing.T) {
	// A chain of proxies appends rather than replaces.
	_, body := get(t, Landing("v1", true, false, "", ""), "http://internal/", map[string]string{
		"X-Forwarded-Proto": "https, http",
		"X-Forwarded-Host":  "llmgw.example, internal",
	})
	if !strings.Contains(body, "https://llmgw.example/v1") {
		t.Errorf("the base URL was built from the wrong entry")
	}
}

func TestOnlyTheRootIsTheLandingPage(t *testing.T) {
	// A catch-all would turn every mistyped path into this page, hiding real
	// 404s from anyone debugging a client.
	code, _ := get(t, Landing("v1", true, false, "", ""), "http://x/v1/chat/completionz", nil)
	if code != http.StatusNotFound {
		t.Errorf("a mistyped path returned %d", code)
	}
}

// fakeSource is a cluster for the dashboard to render.
type fakeSource struct{}

func (fakeSource) Nodes() []adminapi.NodeView {
	return []adminapi.NodeView{{
		ID: "llmgw-node01", State: "ready", GPUTotal: 8, GPUFree: 4,
		Classes: []string{"h100-80g"}, Fabrics: []string{"ib-a"},
		Replicas: 1, LastSeenAgo: "2s", AgentVersion: "v0.5.0",
		Unhealthy: []string{"gpu 3: ECC errors"},
	}}
}

func (fakeUsage) UsageByUser(context.Context, time.Time) ([]store.UserUsage, error) {
	return []store.UserUsage{
		{UserID: "usr-1", Email: "ilker@example.com", Keys: 2,
			Requests: 40, Prompt: 8000, Completion: 4000, CostMicros: 4_500_000, LastSeen: time.Now()},
		{UserID: "", Keys: 1, Requests: 3, Prompt: 90, Completion: 40, Errors: 1, LastSeen: time.Now()},
	}, nil
}

func (fakeUsage) UsageByModel(context.Context, time.Time) ([]store.ModelUsage, error) {
	return []store.ModelUsage{
		{Model: "qwen-coder", Requests: 30, Prompt: 6000, Completion: 3000, AvgLatency: 240},
		{Model: "qwen", Requests: 13, Prompt: 2090, Completion: 1040, Errors: 1, AvgLatency: 90},
	}, nil
}

func (fakeSource) Summary() adminapi.Summary {
	return adminapi.Summary{
		Nodes: adminapi.NodeTotals{Total: 1, Ready: 1},
		GPUs: []adminapi.ClassUse{
			{Class: "h100-80g", Total: 8, Used: 3, Unhealthy: 1},
		},
		Fabrics: []adminapi.FabricUse{
			{Fabric: "ib-a", Nodes: 1, LargestFreeBlock: 0, NodesWithWeights: 1, Models: 2},
		},
		IdleReplicas: []adminapi.IdleReplica{
			{ID: "served-abc", Model: "served", IdleSec: 3600, GPUs: 4},
		},
	}
}

func (fakeSource) Models() []adminapi.ModelView {
	return []adminapi.ModelView{
		{Name: "served", Ready: 1, Min: 1, Max: 2, Profiles: []string{"tp4"}},
		{Name: "stuck", Ready: 0, Min: 1, Max: 1, Profiles: []string{"tp8"}},
	}
}

func (fakeSource) Replicas(string) []adminapi.ReplicaView {
	return []adminapi.ReplicaView{{
		ID: "served-1", Model: "served", State: "ready", Pool: "ib-a/h100-80g",
		Endpoint: "http://10.0.0.141:20000", AgeSec: 3600,
		Shards: []adminapi.ShardView{{ID: "served-1-r0", Node: "llmgw-node01"}},
	}}
}

func (fakeSource) Explain(model string) adminapi.ExplainView {
	if model != "stuck" {
		return adminapi.ExplainView{Model: model, Placed: true}
	}
	return adminapi.ExplainView{Model: "stuck", Rejections: []adminapi.RejectionView{
		{Profile: "tp8", Node: "llmgw-node01", Reason: "has 4 free h100-80g GPUs, needs 8"},
	}}
}

type fakeUsage struct {
	err     error
	buckets []store.Bucket
}

func (f fakeUsage) UsageBuckets(context.Context, time.Time, time.Duration) ([]store.Bucket, error) {
	if f.err != nil {
		return nil, f.err
	}
	return f.buckets, nil
}

func (f fakeUsage) UsageSince(context.Context, time.Time) ([]store.KeyUsage, error) {
	if f.err != nil {
		return nil, f.err
	}
	return []store.KeyUsage{{
		KeyID: "key-1", KeyName: "laptop", KeyPrefix: "sk-llmgw-ab12…",
		ProjectID: "chat", Requests: 42, Errors: 2,
		Prompt: 1000, Completion: 3000, CostMicros: 4_500_000,
		LastAt: time.Now().Add(-time.Minute),
	}}, nil
}

func TestDashboardShowsTheCluster(t *testing.T) {
	h := Dashboard(fakeSource{}, fakeUsage{}, "v1", time.Second, true, nil)
	code, body := get(t, h, "http://127.0.0.1:9090/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d", code)
	}
	for _, want := range []string{
		"llmgw-node01", "h100-80g", "served-1", "http://10.0.0.141:20000",
		"ECC errors", "laptop", "4.50", // spend shown in units, not micro-units
	} {
		if !strings.Contains(body, want) {
			t.Errorf("the dashboard does not show %q", want)
		}
	}
}

func TestDashboardExplainsOnlyWhatIsUnplaced(t *testing.T) {
	// Asking why a satisfied model was not placed gives an empty answer, and a
	// page full of empty answers hides the one that matters.
	h := Dashboard(fakeSource{}, nil, "v1", time.Second, true, nil)
	_, body := get(t, h, "http://127.0.0.1:9090/", nil)

	if !strings.Contains(body, "needs 8") {
		t.Error("the unplaced model has no explanation")
	}
	if strings.Count(body, "Why nothing was placed") != 1 {
		t.Error("the explanation section is missing or duplicated")
	}
}

func TestDashboardSurvivesAFailingUsageQuery(t *testing.T) {
	// A reporting query failing must not take the page with it: the cluster
	// state is what someone is most likely here for.
	h := Dashboard(fakeSource{}, fakeUsage{err: context.DeadlineExceeded}, "v1", time.Second, true, nil)
	code, body := get(t, h, "http://127.0.0.1:9090/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d", code)
	}
	if !strings.Contains(body, "llmgw-node01") {
		t.Error("the cluster state was lost with the usage query")
	}
}

func TestDashboardSaysWhenThereIsNothingToAccount(t *testing.T) {
	h := Dashboard(fakeSource{}, nil, "v1", time.Second, true, nil)
	_, body := get(t, h, "http://127.0.0.1:9090/", nil)
	if !strings.Contains(body, "Authentication is off") {
		t.Error("the page does not explain the empty usage section")
	}
}

func TestDashboardRendersAnEmptyCluster(t *testing.T) {
	// The state a fresh install is in, and the one most likely to be looked at
	// with a template that has never seen it.
	h := Dashboard(emptySource{}, nil, "v1", time.Second, true, nil)
	code, body := get(t, h, "http://127.0.0.1:9090/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d", code)
	}
	if !strings.Contains(body, "No agents have registered") {
		t.Error("an empty cluster renders nothing useful")
	}
}

type emptySource struct{}

func (emptySource) Nodes() []adminapi.NodeView             { return nil }
func (emptySource) Models() []adminapi.ModelView           { return nil }
func (emptySource) Replicas(string) []adminapi.ReplicaView { return nil }
func (emptySource) Explain(string) adminapi.ExplainView    { return adminapi.ExplainView{} }
func (emptySource) Summary() adminapi.Summary              { return adminapi.Summary{} }

func TestStaticServesAssetsAndNothingElse(t *testing.T) {
	// The assets live in their own embed. A file server over the tree that also
	// carries the templates would serve those as well, and a directory listing
	// would publish the shape of the tree to nobody's benefit.
	h := Static()
	for _, path := range []string{
		"/_static/base.css",
		"/_static/katex/katex.min.css",
		"/_static/katex/katex.min.js",
		"/_static/katex/auto-render.min.js",
		"/_static/katex/fonts/KaTeX_Main-Regular.woff2",
	} {
		if code, _ := get(t, h, "http://x"+path, nil); code != http.StatusOK {
			t.Errorf("%s is not served: %d", path, code)
		}
	}
	for _, path := range []string{
		"/_static/dashboard.html", "/_static/chat.html",
		"/_static/", "/_static/katex/", "/_static/katex/fonts/",
		"/_static/../web.go",
	} {
		if code, _ := get(t, h, "http://x"+path, nil); code == http.StatusOK {
			t.Errorf("%s was served", path)
		}
	}
}

func TestTheLandingPageSaysWhereToGetAKey(t *testing.T) {
	// A page that explains the API and leaves the reader with no route to using
	// it is where onboarding stops.
	_, body := get(t, Landing("v1", true, false, "platform@example.com", ""), "http://x/", nil)
	if !strings.Contains(body, `href="mailto:platform@example.com"`) {
		t.Error("an address was not turned into a mailto link")
	}
	// With signup closed, the route is a person; the page still has to say what
	// arriving looks like.
	if !strings.Contains(body, "shown once") {
		t.Error("the page does not say a key cannot be recovered")
	}
}

func TestContactRenderingByKind(t *testing.T) {
	cases := []struct{ in, text, href string }{
		{"platform@example.com", "platform@example.com", "mailto:platform@example.com"},
		{"mailto:ops@example.com", "mailto:ops@example.com", "mailto:ops@example.com"},
		{"https://wiki.example/llmgw", "https://wiki.example/llmgw", "https://wiki.example/llmgw"},
		{"http://wiki.example", "http://wiki.example", "http://wiki.example"},
		// Not addresses and not URLs: shown, not linked.
		{"#platform on Slack", "#platform on Slack", ""},
		{"Ayse", "Ayse", ""},
		{"", "", ""},
		{"   ", "", ""},
	}
	for _, c := range cases {
		text, href := contactLink(c.in)
		if text != c.text || href != c.href {
			t.Errorf("contactLink(%q) = (%q, %q), want (%q, %q)", c.in, text, href, c.text, c.href)
		}
	}
}

func TestADangerousContactSchemeIsNeverLinked(t *testing.T) {
	// This string comes from configuration, and configuration is not always
	// written by whoever runs the gateway. A javascript: href on a page served
	// to every visitor is a stored XSS with a very short path to it; escaping
	// the text does not make the scheme safe.
	for _, bad := range []string{
		"javascript:alert(1)",
		"JavaScript:alert(1)",
		"data:text/html,<script>alert(1)</script>",
		"vbscript:msgbox(1)",
		"file:///etc/passwd",
	} {
		text, href := contactLink(bad)
		if href != "" {
			t.Errorf("contactLink(%q) produced href %q", bad, href)
		}
		if text == "" {
			t.Errorf("contactLink(%q) dropped the text entirely", bad)
		}

		_, body := get(t, Landing("v1", true, false, bad, ""), "http://x/", nil)
		if strings.Contains(body, "href=\"javascript") || strings.Contains(body, "href=\"data:") {
			t.Errorf("%q reached the page as a link", bad)
		}
	}
}

func TestTheNoticeIsEscaped(t *testing.T) {
	_, body := get(t, Landing("v1", true, false, "", "<script>alert(1)</script>"), "http://x/", nil)
	if strings.Contains(body, "<script>alert(1)</script>") {
		t.Error("the notice was rendered as markup")
	}
	if !strings.Contains(body, "&lt;script&gt;") {
		t.Error("the notice was dropped rather than escaped")
	}
}

func TestWithoutAContactThePageStillSaysSomethingUseful(t *testing.T) {
	_, body := get(t, Landing("v1", true, false, "", ""), "http://x/", nil)
	if !strings.Contains(body, "whoever operates this gateway") {
		t.Error("an unconfigured gateway leaves the reader with nothing")
	}
}

func TestWithSignupOpenTheLandingPagePointsAtIt(t *testing.T) {
	// The route from "I found this address" to "I have a key" has to be on the
	// page, or the reader is stuck whatever else it says.
	_, body := get(t, Landing("v1", true, true, "", ""), "http://x/", nil)
	for _, want := range []string{`href="/signup"`, `href="/login"`, `href="/profile"`} {
		if !strings.Contains(body, want) {
			t.Errorf("the page does not link to %s", want)
		}
	}
}

func TestTheDashboardDrawsTrafficWhenThereIsSome(t *testing.T) {
	now := time.Now().Truncate(time.Hour)
	usage := fakeUsage{buckets: []store.Bucket{
		{Start: now.Add(-2 * time.Hour), Requests: 40, Prompt: 4000, Completion: 900, LatencyMs: 120},
		{Start: now.Add(-time.Hour)},
		{Start: now, Requests: 95, Prompt: 9000, Completion: 2100, LatencyMs: 310},
	}}
	h := Dashboard(fakeSource{}, usage, "v1", time.Second, true, nil)
	_, body := get(t, h, "http://127.0.0.1:9090/", nil)

	if !strings.Contains(body, "<svg") {
		t.Fatal("no chart was drawn")
	}
	// The empty hour in the middle must still be a slot, or the shape lies
	// about when traffic happened.
	if n := strings.Count(body, `<rect`); n < 4 {
		t.Errorf("%d bars across three charts and three hours", n)
	}
	if !strings.Contains(body, "peak 95 requests/h") {
		t.Error("the request peak is not labelled")
	}
}

func TestTheDashboardSaysWhenThereIsNoTrafficYet(t *testing.T) {
	h := Dashboard(fakeSource{}, fakeUsage{}, "v1", time.Second, true, nil)
	_, body := get(t, h, "http://127.0.0.1:9090/", nil)
	if !strings.Contains(body, "No traffic recorded yet") {
		t.Error("an empty history renders nothing at all")
	}
}

func TestTheDashboardOffersOnlyLinksThatExistWhereItIsServed(t *testing.T) {
	// The same page is served twice with different neighbours. Metrics, pprof
	// and the management API exist only on the admin listener; account
	// administration and the profile only on the client-facing one.
	//
	// Getting this wrong was not a 404: on the public port those paths fall
	// through to the API-key middleware, so a visitor clicking "metrics" was
	// told "no api key was provided", which is a far more confusing way to say
	// the link does not belong there.
	_, admin := get(t, Dashboard(fakeSource{}, nil, "v1", time.Second, true, nil),
		"http://127.0.0.1:9090/", nil)
	for _, want := range []string{`href="/metrics"`, `href="/debug/pprof/"`, `href="/api/v1/nodes"`} {
		if !strings.Contains(admin, want) {
			t.Errorf("the admin listener's page is missing %s", want)
		}
	}
	if strings.Contains(admin, `href="/admin/users"`) {
		t.Error("the admin listener's page links to a page only the public port serves")
	}

	_, public := get(t, Dashboard(fakeSource{}, nil, "v1", time.Second, false, nil),
		"https://llmgw.example/admin", nil)
	for _, unwanted := range []string{`href="/metrics"`, `href="/debug/pprof/"`, `href="/api/v1/nodes"`} {
		if strings.Contains(public, unwanted) {
			t.Errorf("the public page links to %s, which is not served there", unwanted)
		}
	}
	if !strings.Contains(public, `href="/admin/users"`) {
		t.Error("the public page does not link to account administration")
	}
	// And it says where the missing ones are, rather than pretending they do
	// not exist.
	if !strings.Contains(public, "admin listener") {
		t.Error("the public page does not say where metrics and pprof live")
	}
}

func TestTheDashboardShowsCapacityWithoutPrometheus(t *testing.T) {
	// The whole point of putting this on the page: knowing how full a cluster
	// is should not require standing up a monitoring stack beside it. The
	// tables say what exists; this says how much of it is in use, which is the
	// question asked first.
	h := Dashboard(fakeSource{}, fakeUsage{}, "v1", time.Second, true, nil)
	code, page := get(t, h, "http://127.0.0.1:9090/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d", code)
	}
	for _, want := range []string{
		"Capacity", "h100-80g", "3/8", // used of total
		"Fabrics", "ib-a",
		"Holding hardware nothing is asking for", "served-abc",
	} {
		if !strings.Contains(page, want) {
			t.Errorf("the dashboard does not show %q", want)
		}
	}
}

func TestAFaultyCardIsShownAsNeitherUsedNorFree(t *testing.T) {
	// Counting a faulty card as free is how a cluster looks like it has room it
	// does not, and the page must not disagree with the scheduler about that.
	h := Dashboard(fakeSource{}, fakeUsage{}, "v1", time.Second, true, nil)
	code, page := get(t, h, "http://127.0.0.1:9090/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d", code)
	}
	if !strings.Contains(page, "Faulty") {
		t.Error("faulty cards are not shown")
	}
	// 8 total, 3 used, 1 faulty, so 4 free.
	if !strings.Contains(page, ">4<") {
		t.Errorf("free count is wrong or missing:\n%s", page)
	}
}

func TestAUsageMeterCannotOverrunItsColumn(t *testing.T) {
	// The two counts it divides can disagree for an instant — a GPU freed
	// between one read and the next — and a bar wider than its cell wrecks the
	// table.
	pct, ok := funcs["pct"].(func(int, int) int)
	if !ok {
		t.Fatal("the pct helper is missing")
	}
	for _, tc := range [][2]int{{0, 0}, {9, 8}, {-1, 8}, {100, 8}} {
		if got := pct(tc[0], tc[1]); got < 0 || got > 100 {
			t.Errorf("pct(%d,%d) = %d", tc[0], tc[1], got)
		}
	}
}

func TestTheDashboardShowsWhoIsUsingTheCluster(t *testing.T) {
	// Per key is the right unit for revoking something and the wrong one for
	// answering who is using this cluster: one person holds a laptop key, a CI
	// key and a notebook key, and adding three rows up by eye gets skipped.
	h := Dashboard(fakeSource{}, fakeUsage{}, "v1", time.Second, true, nil)
	code, page := get(t, h, "http://127.0.0.1:9090/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d", code)
	}
	for _, want := range []string{"People", "ilker@example.com", "12.0k"} {
		if !strings.Contains(page, want) {
			t.Errorf("the dashboard does not show %q", want)
		}
	}
	// A key with no owner predates accounts. Calling it "unknown" would imply
	// something is wrong; it is an operator-issued key doing its job.
	if !strings.Contains(page, "operator-issued keys") {
		t.Error("an unowned key is not explained")
	}
}

func TestTheDashboardShowsWhatIsActuallyRunning(t *testing.T) {
	h := Dashboard(fakeSource{}, fakeUsage{}, "v1", time.Second, true, nil)
	_, page := get(t, h, "http://127.0.0.1:9090/", nil)
	for _, want := range []string{"What is being run", "qwen-coder", "240 ms"} {
		if !strings.Contains(page, want) {
			t.Errorf("the dashboard does not show %q", want)
		}
	}
}

func TestAReportingFailureDoesNotTakeThePageWithIt(t *testing.T) {
	// The cluster state is what someone is most likely here for, and a usage
	// query is the part most likely to be slow or broken.
	h := Dashboard(fakeSource{}, brokenUsage{}, "v1", time.Second, true, nil)
	code, page := get(t, h, "http://127.0.0.1:9090/", nil)
	if code != http.StatusOK {
		t.Fatalf("status %d; a reporting failure took the page down", code)
	}
	if !strings.Contains(page, "llmgw-node01") {
		t.Error("the cluster state was lost along with the usage query")
	}
}

// brokenUsage fails every query, as a database under pressure would.
type brokenUsage struct{}

func (brokenUsage) UsageSince(context.Context, time.Time) ([]store.KeyUsage, error) {
	return nil, errUsage
}

func (brokenUsage) UsageBuckets(context.Context, time.Time, time.Duration) ([]store.Bucket, error) {
	return nil, errUsage
}

func (brokenUsage) UsageByUser(context.Context, time.Time) ([]store.UserUsage, error) {
	return nil, errUsage
}

func (brokenUsage) UsageByModel(context.Context, time.Time) ([]store.ModelUsage, error) {
	return nil, errUsage
}

var errUsage = errors.New("the database is unhappy")