feat(build): make the user-modpack build lane read its context (closes the last functional gap)

A submitted modpack was durable but unreadable: the uploads PVC cannot cross
namespaces (felis-api mounts it; Kaniko runs in felis-build) and the s3 lane
handed the sandboxed build Pod no credentials, so NO user build could ever
consume its context. The transport is now the API itself:

- submit: derived context refs become the internal-face URL
  /api/v1/internal/submissions/{id}/context (service-token gated), and Blobs
  gains Open (local + s3) with an ErrBlobNotFound sentinel for the route's 404.
- api: serves that route on the internal face only (openapi.yaml updated; the
  route-coverage test enforces it).
- build: an http(s) context renders a context-fetch initContainer (the felis
  image's new fetch-context entrypoint) that streams the blob with the
  namespace-local service-token Secret — never mounted into Kaniko — and
  extracts it under a zip-slip guard into a size-limited emptyDir that Kaniko
  reads read-only as --context=/context.
- platform/install: the api Deployment carries its own internal base URL; the
  build namespace gets the token Secret through the existing replica mechanism
  (bootstrap.sh + felis setup); the build egress lock opens exactly the control
  namespace on the internal port.
- cmd/felis: fetch-context entrypoint (registered, documented, unit-tested for
  escapes/symlinks/non-gzip).

Tests cover rendering, hardening, the s3/local Open paths, and the route's
404/503 mapping. Verified next on the real single-node cluster with Kaniko.

cmd/felis/api.go

@@ -18,6 +18,7 @@ import (
 	"felis.lolicon.best/internal/config"
 	"felis.lolicon.best/internal/fileedit"
 	"felis.lolicon.best/internal/mail"
+	"felis.lolicon.best/internal/naming"
 	"felis.lolicon.best/internal/panel"
 	"felis.lolicon.best/internal/passkey"
 	"felis.lolicon.best/internal/platform"
@@ -142,10 +143,14 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 	// build namespace and pushes to the internal registry. The build Pod never
 	// holds DB credentials — felis-api owns the PG store and admits scanned
 	// images, so the Builder is constructed here with both bindings.
+	buildCfg := buildConfig(cfg)
+	// The fetch initContainer runs THIS image's fetch-context entrypoint, so the
+	// build config carries the api's own image (the platform sets FELIS_IMAGE).
+	buildCfg.FelisImage = os.Getenv("FELIS_IMAGE")
 	builder := &build.Builder{
 		Store:  build.NewPGStore(drv.DB()),
-		Jobs:   build.NewK8sJobs(cl, buildConfig(cfg)),
+		Jobs:   build.NewK8sJobs(cl, buildCfg),
-		Config: buildConfig(cfg),
+		Config: buildCfg,
 	}
 
 	// User-modpack approval lane (user-directed extension over §16; see
@@ -159,14 +164,19 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 	// The blob upload transport is selected by the shape of user_uploads_context —
 	// the two backends the setup wizard chooses between. A local path wires
 	// LocalContextStore (the mounted uploads PVC); an s3:// base wires
-	// S3ContextStore when its credentials resolve. Either way the store's target is
+	// S3ContextStore when its credentials resolve. Anything else — or an s3:// base
-	// derived from the SAME config field the context ref uses, so the blob lands
+	// with no credentials configured — leaves Blobs nil so POST
-	// exactly where Kaniko's --context points. Anything else — or an s3:// base with
-	// no credentials configured — leaves Blobs nil so POST
 	// /me/submissions/{id}/context returns 503, honest like the restore executor
-	// when its PVC is not supplied. (Letting the sandboxed Kaniko build Pod READ the
+	// when its PVC is not supplied.
-	// context — PVC mount for local, creds+egress for S3 — is a separate deployment
+	//
-	// integration.)
+	// Reading the blob back is the API's job, not Kaniko's: the build Pod runs in
+	// another namespace and can neither mount the uploads PVC (a PVC does not cross
+	// namespaces) nor hold object-store credentials, so ContextBaseURL makes the
+	// derived context ref an internal-face URL that the build Job's fetch
+	// initContainer streams (cmd/felis fetch-context). The platform renders this
+	// address into the api Deployment (felis API base URL env); the fallback keeps
+	// a hand-rolled deployment working under the platform's default control
+	// namespace.
 	contextBase := cfg.Registry.UserUploadsContext
 	var blobs submit.Blobs
 	switch {
@@ -186,11 +196,12 @@ func cmdAPI(args []string, stdout, stderr io.Writer) int {
 		fmt.Fprintf(stderr, "felis api: user-uploads context %q is neither a local path nor an s3:// base — modpack upload transport disabled (POST /api/v1/me/submissions/{id}/context returns 503)\n", contextBase)
 	}
 	submissions := &submit.Manager{
-		Store:        submit.NewPGStore(drv.DB()),
+		Store:          submit.NewPGStore(drv.DB()),
-		Builds:       builder,
+		Builds:         builder,
-		Registry:     cfg.Registry.URL,
+		Registry:       cfg.Registry.URL,
-		ContextStore: contextBase,
+		ContextStore:   contextBase,
-		Blobs:        blobs,
+		ContextBaseURL: internalAPIBaseURL(),
+		Blobs:          blobs,
 	}
 
 	// Restore subsystem (spec §7): the weak-SA restore Job mounts the target
@@ -413,6 +424,17 @@ func buildConfig(cfg *config.Config) build.Config {
 	}
 }
 
+// internalAPIBaseURL resolves the platform's internal-face base URL: the address
+// the platform rendered into this pod (felis API base URL env), or — for a
+// hand-rolled deployment that set none — the platform default control namespace,
+// the same fallback setup.go uses to hand the login gate its address.
+func internalAPIBaseURL() string {
+	if base := os.Getenv(naming.EnvAPIBaseURL); base != "" {
+		return base
+	}
+	return platform.InternalAPIBaseURL(platform.DefaultControlNamespace)
+}
+
 // uploadsSchemeRE matches a leading URL scheme like "s3://" or "gs://".
 var uploadsSchemeRE = regexp.MustCompile(`^[a-zA-Z][a-zA-Z0-9+.-]*://`)
 

cmd/felis/fetchcontext.go

@@ -0,0 +1,150 @@
+package main
+
+import (
+	"archive/tar"
+	"compress/gzip"
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"os/signal"
+	"path/filepath"
+	"strings"
+	"syscall"
+	"time"
+)
+
+// cmdFetchContext is the in-Pod entrypoint the build Job's context-fetch
+// initContainer runs. It performs one read against the felis-api INTERNAL face —
+// the blob the platform stored for a submission — and extracts it into the shared
+// emptyDir the Kaniko container then builds from.
+//
+// Why this exists: the build Pod runs in the build namespace, where it can neither
+// mount the control-plane uploads PVC (a PVC does not cross namespaces) nor hold
+// object-store credentials, so the API that WROTE the blob is the transport. The
+// route is service-token-gated; the token arrives through a namespace-local Secret
+// mounted only into this initContainer, never into Kaniko's — so the untrusted
+// Dockerfile's build steps have no credential to read (their containers share no
+// environment, no PID namespace, and Kaniko itself mounts the context read-only).
+//
+// The extraction is deliberately paranoid: the tarball is attacker-controlled
+// input, so absolute paths, ".." escapes, links, and special files are refused
+// rather than sanitized. Kaniko treats the extracted tree as hostile regardless
+// (spec §16), but the pod's own filesystem still must not be written outside the
+// context directory it was given.
+func cmdFetchContext(args []string, _, stderr io.Writer) int {
+	fs := flag.NewFlagSet("fetch-context", flag.ContinueOnError)
+	fs.SetOutput(stderr)
+	url := fs.String("url", "", "internal-face URL of the submission's build-context tarball")
+	out := fs.String("out", "/context", "directory to extract the build context into")
+	if err := fs.Parse(args); err != nil {
+		return 2
+	}
+	if *url == "" {
+		fmt.Fprintln(stderr, "felis fetch-context: --url is required")
+		return 2
+	}
+	token := os.Getenv("FELIS_SERVICE_TOKEN")
+	if token == "" {
+		fmt.Fprintln(stderr, "felis fetch-context: FELIS_SERVICE_TOKEN is empty — the internal face rejects anonymous reads")
+		return 2
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
+	defer stop()
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, *url, nil)
+	if err != nil {
+		fmt.Fprintf(stderr, "felis fetch-context: bad --url: %v\n", err)
+		return 2
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	// No overall client timeout: a legitimate modpack context can be large and the
+	// Job's activeDeadlineSeconds is the real bound. The header timeout catches a
+	// wedged endpoint without capping a healthy download.
+	client := &http.Client{Transport: &http.Transport{ResponseHeaderTimeout: time.Minute}}
+	resp, err := client.Do(req)
+	if err != nil {
+		fmt.Fprintf(stderr, "felis fetch-context: GET failed: %v\n", err)
+		return 1
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		fmt.Fprintf(stderr, "felis fetch-context: %s\n", resp.Status)
+		return 1
+	}
+
+	if err := extractTarGz(resp.Body, *out); err != nil {
+		fmt.Fprintf(stderr, "felis fetch-context: %v\n", err)
+		return 1
+	}
+	return 0
+}
+
+// extractTarGz streams a gzip'd tarball into root, creating directories as
+// needed. Every entry is vetted BEFORE anything is written: a path that is
+// absolute or escapes root (via ".."), a link (symlink or hardlink), or any
+// special file kind aborts the whole extraction. Refusing rather than skipping is
+// deliberate — a context that needs one of those constructs is not a context this
+// transport carries, and silently dropping entries would build from a corpus the
+// submitter did not upload.
+func extractTarGz(r io.Reader, root string) error {
+	if err := os.MkdirAll(root, 0o755); err != nil {
+		return fmt.Errorf("create context dir: %w", err)
+	}
+	zr, err := gzip.NewReader(r)
+	if err != nil {
+		return fmt.Errorf("context is not a valid gzip tarball: %w", err)
+	}
+	defer zr.Close()
+	tr := tar.NewReader(zr)
+	for {
+		hdr, err := tr.Next()
+		if errors.Is(err, io.EOF) {
+			return nil
+		}
+		if err != nil {
+			return fmt.Errorf("read context tarball: %w", err)
+		}
+		name := filepath.Clean(hdr.Name)
+		if name == "." {
+			continue
+		}
+		// The zip-slip guard: reject, never rewrite. filepath.Clean collapses any
+		// "a/../../b", so these two checks are sufficient once Clean has run.
+		if filepath.IsAbs(name) || name == ".." || strings.HasPrefix(name, ".."+string(filepath.Separator)) {
+			return fmt.Errorf("context entry %q escapes the context directory", hdr.Name)
+		}
+		target := filepath.Join(root, name)
+		switch hdr.Typeflag {
+		case tar.TypeDir:
+			if err := os.MkdirAll(target, 0o755); err != nil {
+				return fmt.Errorf("create %q: %w", name, err)
+			}
+		case tar.TypeReg, tar.TypeRegA:
+			if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
+				return fmt.Errorf("create parent of %q: %w", name, err)
+			}
+			mode := os.FileMode(0o644)
+			if hdr.FileInfo().Mode()&0o111 != 0 {
+				mode = 0o755 // preserve executability (entrypoint scripts), nothing else
+			}
+			f, err := os.OpenFile(target, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, mode)
+			if err != nil {
+				return fmt.Errorf("create %q: %w", name, err)
+			}
+			if _, err := io.Copy(f, tr); err != nil {
+				_ = f.Close()
+				return fmt.Errorf("write %q: %w", name, err)
+			}
+			if err := f.Close(); err != nil {
+				return fmt.Errorf("close %q: %w", name, err)
+			}
+		default:
+			return fmt.Errorf("context entry %q has unsupported type %q (links and special files are refused)", hdr.Name, string(hdr.Typeflag))
+		}
+	}
+}

cmd/felis/fetchcontext_test.go

@@ -0,0 +1,171 @@
+package main
+
+import (
+	"archive/tar"
+	"bytes"
+	"compress/gzip"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+type tarEntry struct {
+	name     string
+	body     string
+	mode     int64
+	typ      byte
+	linkname string
+}
+
+// tgzBody builds an in-memory .tar.gz from entries, preserving each entry's type
+// and mode so the tests can exercise the guards with exactly the bytes an
+// attacker could upload.
+func tgzBody(t *testing.T, entries ...tarEntry) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	zw := gzip.NewWriter(&buf)
+	tw := tar.NewWriter(zw)
+	for _, e := range entries {
+		typ := e.typ
+		if typ == 0 {
+			typ = tar.TypeReg
+		}
+		mode := e.mode
+		if mode == 0 {
+			mode = 0o644
+		}
+		hdr := &tar.Header{Name: e.name, Typeflag: typ, Mode: mode, Size: int64(len(e.body))}
+		if typ == tar.TypeSymlink {
+			hdr.Linkname = e.linkname
+			hdr.Size = 0
+		}
+		if err := tw.WriteHeader(hdr); err != nil {
+			t.Fatalf("write header %q: %v", e.name, err)
+		}
+		if hdr.Size > 0 {
+			if _, err := tw.Write([]byte(e.body)); err != nil {
+				t.Fatalf("write body %q: %v", e.name, err)
+			}
+		}
+	}
+	if err := tw.Close(); err != nil {
+		t.Fatalf("close tar: %v", err)
+	}
+	if err := zw.Close(); err != nil {
+		t.Fatalf("close gzip: %v", err)
+	}
+	return buf.Bytes()
+}
+
+// A normal context extracts with its tree intact, and the executable bit that
+// modpack entrypoints rely on survives.
+func TestExtractTarGzRoundTrip(t *testing.T) {
+	dir := t.TempDir()
+	body := tgzBody(t,
+		tarEntry{name: "Dockerfile", body: "FROM scratch\n"},
+		tarEntry{name: "mods/example.jar", body: "jar-bytes"},
+		tarEntry{name: "start.sh", body: "#!/bin/sh\n", mode: 0o755},
+		tarEntry{name: "mods/", typ: tar.TypeDir, mode: 0o755},
+	)
+	if err := extractTarGz(bytes.NewReader(body), dir); err != nil {
+		t.Fatalf("extract: %v", err)
+	}
+	for name, want := range map[string]string{
+		"Dockerfile":       "FROM scratch\n",
+		"mods/example.jar": "jar-bytes",
+	} {
+		got, err := os.ReadFile(filepath.Join(dir, name))
+		if err != nil || string(got) != want {
+			t.Fatalf("%s = (%q, %v), want %q", name, got, err, want)
+		}
+	}
+	fi, err := os.Stat(filepath.Join(dir, "start.sh"))
+	if err != nil || fi.Mode()&0o111 == 0 {
+		t.Fatalf("entrypoint script lost its exec bit: %v (%v)", fi, err)
+	}
+}
+
+// The guards: "..", absolute paths, symlinks, and special files are refused whole
+// — nothing escapes, and nothing is silently skipped.
+func TestExtractTarGzRefusesEscapes(t *testing.T) {
+	cases := []struct {
+		name    string
+		entries []tarEntry
+	}{
+		{"dotdot", []tarEntry{{name: "../outside", body: "x"}}},
+		{"nested dotdot", []tarEntry{{name: "a/../../outside", body: "x"}}},
+		{"absolute", []tarEntry{{name: "/etc/outside", body: "x"}}},
+		{"symlink", []tarEntry{{name: "link", typ: tar.TypeSymlink, linkname: "/etc"}}},
+		{"hardlink", []tarEntry{{name: "hard", typ: tar.TypeLink, linkname: "somewhere"}}},
+		{"device", []tarEntry{{name: "dev", typ: tar.TypeChar}}},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			dir := t.TempDir()
+			if err := extractTarGz(bytes.NewReader(tgzBody(t, tc.entries...)), dir); err == nil {
+				t.Fatal("extract accepted a hostile entry, want an error")
+			}
+			// Nothing may have been written outside the target (or at all).
+			entries, _ := os.ReadDir(dir)
+			if len(entries) != 0 {
+				t.Fatalf("hostile archive left %d entries behind", len(entries))
+			}
+		})
+	}
+}
+
+// The command end to end: it dials the URL with the bearer token from the
+// environment, and refuses to run without it (the internal face would 401
+// anyway; failing at parse time is the honest earlier error).
+func TestCmdFetchContextFetchAndExtract(t *testing.T) {
+	body := tgzBody(t, tarEntry{name: "Dockerfile", body: "FROM scratch\n"})
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "Bearer test-token" {
+			w.WriteHeader(http.StatusUnauthorized)
+			return
+		}
+		w.Header().Set("Content-Type", "application/gzip")
+		_, _ = w.Write(body)
+	}))
+	defer srv.Close()
+
+	dir := t.TempDir()
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	if code := cmdFetchContext([]string{"--url=" + srv.URL + "/sub-1/context", "--out=" + dir}, io.Discard, io.Discard); code != 0 {
+		t.Fatalf("cmdFetchContext exit = %d, want 0", code)
+	}
+	if got, err := os.ReadFile(filepath.Join(dir, "Dockerfile")); err != nil || string(got) != "FROM scratch\n" {
+		t.Fatalf("extracted Dockerfile = (%q, %v)", got, err)
+	}
+
+	// No token: refuse before dialing.
+	t.Setenv("FELIS_SERVICE_TOKEN", "")
+	var stderr bytes.Buffer
+	if code := cmdFetchContext([]string{"--url=" + srv.URL + "/sub-1/context", "--out=" + t.TempDir()}, io.Discard, &stderr); code != 2 {
+		t.Fatalf("missing token exit = %d, want 2 (stderr %q)", code, stderr.String())
+	}
+
+	// A non-200 answer (e.g. the route's 404 for a never-uploaded context) fails.
+	srv404 := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer srv404.Close()
+	t.Setenv("FELIS_SERVICE_TOKEN", "test-token")
+	if code := cmdFetchContext([]string{"--url=" + srv404.URL + "/sub-1/context", "--out=" + t.TempDir()}, io.Discard, io.Discard); code != 1 {
+		t.Fatalf("404 exit = %d, want 1", code)
+	}
+}
+
+// A body that is not a gzip tarball must fail the extraction rather than produce
+// an empty (or partial) context Kaniko would then try to build.
+func TestExtractTarGzRejectsNonGzip(t *testing.T) {
+	dir := t.TempDir()
+	err := extractTarGz(strings.NewReader("not a tarball"), dir)
+	if err == nil || !strings.Contains(err.Error(), "gzip") {
+		t.Fatalf("err = %v, want a gzip complaint", err)
+	}
+}

cmd/felis/run.go

@@ -19,6 +19,7 @@ Commands:
   restore           Extract a world archive into a world volume (internal Job entrypoint)
   backup            Archive a world into the backup store and record it (internal Job entrypoint)
   files             List/read/write one file in a stopped server's world (internal Job entrypoint)
+  fetch-context     Fetch and extract a submission's build context (internal Job entrypoint)
   manifests         Render the control-plane RBAC + NetworkPolicy install bundle as YAML
   apply             Create a MinecraftServer CRD (direct K8s write; use -f server.json)
   setup             Run host bootstrap + first-run setup console (TUI; requires root/sudo)
@@ -47,6 +48,7 @@ var commands = map[string]func(args []string, stdout, stderr io.Writer) int{
 	"restore":          cmdRestore,
 	"backup":           cmdBackup,
 	"files":            cmdFiles,
+	"fetch-context":    cmdFetchContext,
 	"manifests":        cmdManifests,
 	"apply":            cmdApply,
 	"setup":            cmdSetup,

cmd/felis/setup.go

@@ -233,13 +233,25 @@ func provisionSystemServers(ctx context.Context, cfg *config.Config, out io.Writ
 	// (the on-demand BACKUP Job runs in the minecraft namespace and mounts it to
 	// self-record its world_backups row; without the replica the Job's volume
 	// mount fails and every backup request strands in the cluster).
+	// An empty build_namespace means the build system's compiled-in default; the
+	// replica must target the namespace the Jobs actually run in.
+	buildNS := cfg.Registry.BuildNamespace
+	if buildNS == "" {
+		buildNS = platform.DefaultBuildNamespace
+	}
 	secretOutcomes := []systemServerOutcome{
 		ensureSecretReplica(ctx, cl, controlNS, cfg.K8s.Namespace,
-			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token"),
+			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token", "minecraft ns"),
 		ensureSecretReplica(ctx, cl, controlNS, cfg.K8s.Namespace,
-			naming.ForwardingSecretName, naming.ForwardingSecretKey, "forwarding-secret"),
+			naming.ForwardingSecretName, naming.ForwardingSecretKey, "forwarding-secret", "minecraft ns"),
 		ensureSecretReplica(ctx, cl, controlNS, cfg.K8s.Namespace,
-			"felis-config", "felis.toml", "config"),
+			"felis-config", "felis.toml", "config", "minecraft ns"),
+		// The build namespace needs the same token: the build Job's fetch
+		// initContainer reads the submission context from the internal face. Best
+		// effort — a deployment that only installs the control plane simply never
+		// builds a user submission.
+		ensureSecretReplica(ctx, cl, controlNS, buildNS,
+			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token", "felis-build ns"),
 	}
 	outcomes := ensureSystemServers(ctx, cl, cfg.K8s.Namespace, cfg.Velocity.LoginImage, cfg.Velocity.LobbyImage, apiBaseURL, cfg.Server.RootDomain, defaultPanelHostname(cfg.Server.RootDomain, cfg.Auth.PanelHostname))
 	outcomes = append(secretOutcomes, outcomes...)

cmd/felis/systemservers.go

@@ -95,7 +95,7 @@ const felisLimboHealthPort int32 = 8080
 // fail-safes to readiness-only, so a login pod that has the URL/domain but not yet
 // the token is safe (it simply does not authenticate) rather than broken.
 const (
-	envAPIBaseURL    = "FELIS_API_BASE_URL"
+	envAPIBaseURL    = naming.EnvAPIBaseURL
 	envRootDomain    = "FELIS_ROOT_DOMAIN"
 	envPanelHostname = "FELIS_PANEL_HOSTNAME"
 	envLobbyServer   = "FELIS_LOBBY_SERVER"
@@ -471,26 +471,27 @@ func phaseOrPending(p v1alpha1.Phase) string {
 	return string(p)
 }
 
-// ensureSecretReplica copies one Secret from the control namespace into the minecraft
+// ensureSecretReplica copies one Secret from the control namespace into a workload
-// namespace so a backend pod can mount it via secretKeyRef. A secretKeyRef is
+// namespace (minecraft — or the build namespace, whose fetch initContainer reads the
-// namespace-local, but the backends run in the minecraft namespace while the sources
+// context from the felis-api internal face with the same token) so a pod can mount it
-// of truth live beside the control plane — so without this replica the operator's
+// via secretKeyRef. A secretKeyRef is namespace-local, but those workloads do not run
-// injected secretKeyRef would dangle and wedge the pod in CreateContainerConfigError.
+// beside the control plane — so without this replica the secretKeyRef would dangle and
+// wedge the pod in CreateContainerConfigError.
 //
-// Two Secrets need it, for different reasons: the service token (login only — it
+// Two Secrets need it, for different reasons: the service token (the login limbo and
-// authenticates the limbo plugin to the felis-api internal face) and the Velocity
+// the build Pod's context fetch — both authenticate to the felis-api internal face)
-// modern-forwarding secret (every backend — it is how a backend knows a login really
+// and the Velocity modern-forwarding secret (every backend — it is how a backend knows
-// came from the proxy, and so that the player's UUID is Mojang-verified rather than
+// a login really came from the proxy, and so that the player's UUID is Mojang-verified
-// offline-derived).
+// rather than offline-derived).
 //
 // It is create-if-absent: an existing replica is left untouched so a hand-rotated
-// value in the minecraft namespace is never clobbered (to rotate, delete the replica
+// value in the workload namespace is never clobbered (to rotate, delete the replica
 // and re-run setup). Best-effort like the rest of the provisioner: a missing source or
 // a create failure degrades to a reported outcome, never a hard setup failure. It
 // copies only Type and Data — never labels/annotations/ownerRefs — so the replica
 // carries no accidental GC owner or managed-by lineage.
-func ensureSecretReplica(ctx context.Context, cl client.Client, controlNamespace, minecraftNamespace, secretName, secretKey, label string) systemServerOutcome {
+func ensureSecretReplica(ctx context.Context, cl client.Client, controlNamespace, minecraftNamespace, secretName, secretKey, label, where string) systemServerOutcome {
-	name := label + " (minecraft ns)"
+	name := label + " (" + where + ")"
 	validate := func(secret *corev1.Secret, location, skipped string) systemServerOutcome {
 		if len(secret.Data[secretKey]) == 0 {
 			return systemServerOutcome{name: name, skipped: fmt.Sprintf(

cmd/felis/systemservers_test.go

@@ -156,7 +156,7 @@ func TestEnsureSecretReplica(t *testing.T) {
 	}
 	replicate := func(cl client.Client, controlNS, mcNS string) systemServerOutcome {
 		return ensureSecretReplica(ctx, cl, controlNS, mcNS,
-			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token")
+			naming.ServiceTokenSecretName, naming.ServiceTokenSecretKey, "service-token", "minecraft ns")
 	}
 
 	t.Run("replicates when absent", func(t *testing.T) {

deploy/bootstrap.sh

@@ -2148,6 +2148,10 @@ deploy_bundle() {
     --from-file=felis.toml="${STATE_DIR}/felis.pod.toml" \
     --dry-run=client -o yaml | kube apply -f -
   apply_literal_secret "$CONTROL_NS" felis-service-token token "$SERVICE_TOKEN"
+  # The build namespace needs the same token: the build Job's fetch initContainer
+  # streams a submission's build context from the felis-api internal face, and a
+  # secretKeyRef is namespace-local (a PVC cannot carry it across either).
+  apply_literal_secret "$BUILD_NS" felis-service-token token "$SERVICE_TOKEN"
   # The forwarding key every backend verifies the proxy's handshake with. `felis setup`
   # replicates it into the minecraft namespace (ensureSecretReplica) before it creates
   # the pods that mount it; the operator injects it into EVERY backend, because Velocity's

docs/deferred-seams.md

@@ -42,21 +42,22 @@ A grep across `*.md` and `*.go` returns both sets; only the Go ones are seams.
   does not exist. `felis update` runs with a zero window, under which every
   `Scheduled` component degrades to a notify, so no path can currently claim an
   apply is under way.
-- `internal/submit/blobstore.go:40` — the uploads PVC is mounted into felis-api but
+- `internal/submit/blobstore.go` — CLOSED 2026-09-22. The uploads PVC still cannot
-  not into the Kaniko build Pod, so a submitted context is durable at the derived
+  cross namespaces, so the transport went through the API instead of a mount: the
-  location without yet being readable by the build that consumes it. Audited
+  derived context ref is now the internal-face URL
-  2026-09-22: this is not a missing volume line — a PVC cannot cross namespaces
+  (`/api/v1/internal/submissions/{id}/context`, service-token gated), the build
-  (uploads live in the control namespace; build Pods run in `felis-build`), so the
+  Job's `context-fetch` initContainer streams it with `felis fetch-context` and
-  fix is a transport, not a mount. The `s3://` lane does not close it either: the
+  extracts under a zip-slip guard into a size-limited emptyDir, and Kaniko builds
-  build Job carries no AWS credentials (no env, and the weak SA's token is
+  `--context=/context`. The token reaches the build namespace through the same
-  deliberately unmounted, so no IAM either). Options on the table: (a) object
+  Secret-replica mechanism the login gate uses (bootstrap + `felis setup`), and the
-  storage with credentials plumbed into the build Pod as a per-build Secret plus an
+  build egress lock allows exactly the control namespace on the internal port.
-  egress allowance; (b) a context-handoff PVC/Job pair in `felis-build` fed from
+  Uniform for local and s3:// stores — neither hands the sandboxed build Pod a
-  the API side; (c) a node-local path both sides mount (single-node only, and it
+  filesystem view or object-store credentials. Kaniko/Trivy images are
-  hands an arbitrary Dockerfile a filesystem view — needs its own security review).
+  external-only by default; `[registry] kaniko_image / trivy_image /
-  Kaniko/Trivy images are external-only by default; `[registry] kaniko_image /
+  build_cpu_limit / build_mem_limit` override them for mirrored or air-gapped
-  trivy_image / build_cpu_limit / build_mem_limit` now override them for mirrored
+  installs, and Trivy's vulnerability DB download needs the same treatment (a
-  or air-gapped installs.
+  `package_source_cidrs` allowance or an internal `TRIVY_DB_REPOSITORY` mirror) or
+  the scan step fails closed on an egress-locked install.
 
 ## Built; only its I/O is unverifiable from this repo
 

docs/openapi.yaml

@@ -607,6 +607,34 @@ paths:
         '404':
           $ref: '#/components/responses/NotFound'
 
+  /api/v1/internal/submissions/{id}/context:
+    get:
+      tags: [submissions-internal]
+      operationId: internalSubmissionContext
+      summary: Stream a submission's stored build-context tarball to the build Pod.
+      description: >-
+        The build Job's fetch initContainer cannot mount the control-plane uploads
+        PVC (a PVC does not cross namespaces) and holds no object-store
+        credentials, so the API that stored the blob streams it here. Served on
+        the internal face (service token, no Zero Trust).
+      x-felis-face: [internal]
+      x-felis-tier: service
+      security: [{ serviceToken: [] }]
+      parameters:
+        - { name: id, in: path, required: true, schema: { type: string } }
+      responses:
+        '200':
+          description: The stored gzip tarball, verbatim.
+          content:
+            application/gzip:
+              schema: { type: string, format: binary }
+        '401':
+          $ref: '#/components/responses/Unauthorized'
+        '404':
+          $ref: '#/components/responses/NotFound'
+        '503':
+          $ref: '#/components/responses/ServiceUnavailable'
+
   /api/v1/internal/servers/{name}/join-event:
     post:
       tags: [servers-internal]

internal/api/api.go

@@ -271,6 +271,9 @@ func (a *API) internalAPIRoutes() []apiRoute {
 		{Method: "GET", Pattern: "/readyz", Public: true, h: a.handleReadyz},
 
 		{Method: "GET", Pattern: "/api/v1/servers", h: a.handleListServers},
+		// The build Pod's context-fetch initContainer streams a submission's stored
+		// modpack through this route (build namespace cannot mount the uploads PVC).
+		{Method: "GET", Pattern: "/api/v1/internal/submissions/{id}/context", h: a.handleInternalSubmissionContext},
 		{Method: "POST", Pattern: "/api/v1/internal/servers/{name}/ready", h: a.handleReady},
 		{Method: "POST", Pattern: "/api/v1/internal/servers/{name}/join-event", h: a.handleJoinEvent},
 		// Domain-autostart (spec §9.1, §14): velocity drives the wake lever and polls

internal/api/submissions.go

@@ -42,6 +42,11 @@ type SubmissionService interface {
 	// Reject is the admin's other verdict: pending_review -> rejected with a
 	// required reason; it starts no build.
 	Reject(ctx context.Context, id, reviewedBy, reason string) (*submit.Submission, error)
+	// OpenContext returns the stored build-context blob for the internal
+	// context-fetch route: the build Pod's initContainer cannot mount the uploads
+	// PVC across namespaces and holds no object-store credentials, so it streams
+	// the blob from the API over the service-token-gated internal face instead.
+	OpenContext(ctx context.Context, id string) (io.ReadCloser, error)
 }
 
 // createSubmissionRequest is the POST /me/submissions body. The user
@@ -211,6 +216,8 @@ func writeSubmitError(w http.ResponseWriter, r *http.Request, err error) {
 	case errors.Is(err, submit.ErrAlreadyReviewed):
 		writeError(w, r, newError(http.StatusConflict, "already_reviewed",
 			"submission has already been reviewed"))
+	case errors.Is(err, submit.ErrBlobNotFound):
+		writeError(w, r, newError(http.StatusNotFound, "not_found", "no context uploaded for this submission"))
 	case errors.Is(err, submit.ErrUploadsUnavailable):
 		writeError(w, r, newError(http.StatusServiceUnavailable, "uploads_unavailable",
 			"modpack upload transport is not configured"))
@@ -219,5 +226,32 @@ func writeSubmitError(w http.ResponseWriter, r *http.Request, err error) {
 	}
 }
 
+// handleInternalSubmissionContext streams a submission's stored build-context
+// tarball to the build Pod's `felis fetch-context` initContainer. It lives on the
+// internal face (service-token, no Zero Trust) because its only caller is
+// in-cluster infrastructure: the build Job runs in the build namespace, where it
+// can neither mount the uploads PVC nor hold object-store credentials, so the API
+// — which wrote the blob — is the transport. The blob is served verbatim; the
+// fetcher extracts it under a zip-slip guard, and Kaniko treats the result as
+// hostile regardless (spec §16).
+func (a *API) handleInternalSubmissionContext(w http.ResponseWriter, r *http.Request) {
+	if a.Submissions == nil {
+		writeError(w, r, errSubmissionsUnavailable)
+		return
+	}
+	rc, err := a.Submissions.OpenContext(r.Context(), r.PathValue("id"))
+	if err != nil {
+		writeSubmitError(w, r, err)
+		return
+	}
+	defer rc.Close()
+	w.Header().Set("Content-Type", "application/gzip")
+	if _, err := io.Copy(w, rc); err != nil {
+		// The status is already committed; the client sees a truncated stream and
+		// the fetch fails on size/extract, so there is nothing left to write here.
+		return
+	}
+}
+
 // Compile-time proof that the production Manager satisfies the API interface.
 var _ SubmissionService = (*submit.Manager)(nil)

internal/api/submissions_test.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"strings"
 	"testing"
 
 	"felis.lolicon.best/internal/submit"
@@ -35,6 +36,9 @@ type fakeSubmissions struct {
 	rejectedBy string
 	rejectReas string
 	rejectErr  error
+	openedID   string
+	openBody   string
+	openErr    error
 }
 
 func (f *fakeSubmissions) Create(_ context.Context, req submit.CreateRequest) (*submit.Submission, error) {
@@ -82,6 +86,16 @@ func (f *fakeSubmissions) Reject(_ context.Context, id, reviewedBy, reason strin
 	return &submit.Submission{ID: id, Status: submit.StatusRejected, ReviewedBy: reviewedBy, RejectReason: reason}, nil
 }
 
+// openErr injects the OpenContext outcome; the body recorder lets the internal
+// route test assert byte-exact streaming and the 404 mapping.
+func (f *fakeSubmissions) OpenContext(_ context.Context, id string) (io.ReadCloser, error) {
+	f.openedID = id
+	if f.openErr != nil {
+		return nil, f.openErr
+	}
+	return io.NopCloser(strings.NewReader(f.openBody)), nil
+}
+
 // appSubAPI wires a submissions service behind an ordinary user principal (the
 // app tier — /me/submissions). [email protected] / .test are deliberately not the
 // deployment domain.
@@ -396,3 +410,46 @@ func TestSubmissionRoutesWithoutServiceAre503(t *testing.T) {
 		t.Fatalf("admin route: code = %d, want 503", w.Code)
 	}
 }
+
+// The internal context route is the build Pod's only read path to a submission's
+// blob: it streams the bytes verbatim, and its error mapping distinguishes a
+// missing blob (404) from an unwired transport (503).
+func TestInternalSubmissionContextRoute(t *testing.T) {
+	newAPI := func(s SubmissionService) *API {
+		api := newTestAPI(newFakeRepo(), newFakeCluster())
+		api.Submissions = s
+		return api
+	}
+
+	t.Run("streams the blob", func(t *testing.T) {
+		fs := &fakeSubmissions{openBody: "\x1f\x8b\x08\x00blob"}
+		w := do(newAPI(fs).InternalHandler(), "GET", "/api/v1/internal/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusOK {
+			t.Fatalf("code = %d body %s", w.Code, w.Body.String())
+		}
+		if w.Body.String() != fs.openBody {
+			t.Fatalf("body = %q, want the stored blob %q", w.Body.String(), fs.openBody)
+		}
+		if fs.openedID != "sub-7" {
+			t.Fatalf("opened id = %q, want the path id", fs.openedID)
+		}
+		if ct := w.Header().Get("Content-Type"); ct != "application/gzip" {
+			t.Fatalf("content-type = %q, want application/gzip", ct)
+		}
+	})
+
+	t.Run("missing blob is 404", func(t *testing.T) {
+		fs := &fakeSubmissions{openErr: fmt.Errorf("%w: gone", submit.ErrBlobNotFound)}
+		w := do(newAPI(fs).InternalHandler(), "GET", "/api/v1/internal/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusNotFound {
+			t.Fatalf("code = %d, want 404", w.Code)
+		}
+	})
+
+	t.Run("unwired transport is 503", func(t *testing.T) {
+		w := do(newAPI(nil).InternalHandler(), "GET", "/api/v1/internal/submissions/sub-7/context", "", nil)
+		if w.Code != http.StatusServiceUnavailable {
+			t.Fatalf("code = %d, want 503", w.Code)
+		}
+	})
+}

internal/build/build.go

@@ -215,6 +215,11 @@ type Config struct {
 	// RegistryURL is the internal registry the build pushes to and Trivy scans
 	// (spec §17). Image refs are validated to be under it.
 	RegistryURL string
+	// FelisImage is the platform image whose `fetch-context` entrypoint streams a
+	// submission's context from the internal face into the build Pod. Required
+	// only when a build's ContextRef is an http(s) URL (the submit lane's derived
+	// shape); an install that never builds user submissions can leave it empty.
+	FelisImage string
 	// KanikoImage / TrivyImage are the executor images.
 	KanikoImage string
 	TrivyImage  string
@@ -355,6 +360,7 @@ func (b *Builder) jobParams(bld *Build, cfg Config) JobParams {
 		Namespace:      cfg.Namespace,
 		ServiceAccount: cfg.ServiceAccount,
 		RegistryURL:    cfg.RegistryURL,
+		FelisImage:     cfg.FelisImage,
 		KanikoImage:    cfg.KanikoImage,
 		TrivyImage:     cfg.TrivyImage,
 		Deadline:       cfg.Deadline,

internal/build/jobspec.go

@@ -2,8 +2,10 @@ package build
 
 import (
 	"fmt"
+	"strings"
 	"time"
 
+	"felis.lolicon.best/internal/naming"
 	batchv1 "k8s.io/api/batch/v1"
 	corev1 "k8s.io/api/core/v1"
 	networkingv1 "k8s.io/api/networking/v1"
@@ -32,8 +34,23 @@ const (
 const (
 	ContainerKaniko = "kaniko"
 	ContainerTrivy  = "trivy"
+	// ContainerFetch is the initContainer that pulls a submission's build context
+	// from the felis-api internal face and extracts it into the shared emptyDir.
+	// It exists only for an http(s) ContextRef (see BuildJob); a ref Kaniko can
+	// read natively (s3://) or a pre-mounted path renders no such container.
+	ContainerFetch = "context-fetch"
+
+	// contextVolume/contextMountPath carry a fetched build context: the fetch
+	// initContainer writes the extracted tree there, Kaniko reads it read-only.
+	contextVolume    = "context"
+	contextMountPath = "/context"
 )
 
+// contextSizeLimit bounds the extracted (attacker-controlled) context tree so a
+// tarball bomb wedges the build pod instead of the node's disk. The compressed
+// upload is capped at 1 GiB by the submit lane; 4 GiB leaves expansion room.
+var contextSizeLimit = resource.MustParse("4Gi")
+
 // JobParams are the rendered inputs to a build Job. They are derived from a
 // Build + Config by the Builder; jobspec is a pure function of them so the
 // security-critical Job shape is unit-tested without a cluster.
@@ -44,11 +61,14 @@ type JobParams struct {
 	Namespace      string
 	ServiceAccount string
 	RegistryURL    string
-	KanikoImage    string
+	// FelisImage runs the context-fetch initContainer (the felis binary's
-	TrivyImage     string
+	// fetch-context entrypoint). Required when ContextRef is an http(s) URL.
-	Deadline       time.Duration
+	FelisImage  string
-	CPULimit       string
+	KanikoImage string
-	MemLimit       string
+	TrivyImage  string
+	Deadline    time.Duration
+	CPULimit    string
+	MemLimit    string
 }
 
 // BuildJobName is the deterministic Job name for a build id.
@@ -100,21 +120,75 @@ func BuildJob(p JobParams) (*batchv1.Job, error) {
 		Capabilities:             &corev1.Capabilities{Drop: []corev1.Capability{"ALL"}},
 	}
 
+	// The context Kaniko reads. An http(s) ref (the submit lane's derived ref: the
+	// API streams the blob on its internal face, because the build Pod can neither
+	// mount the control-plane uploads PVC across namespaces nor hold object-store
+	// credentials) is first fetched into a shared emptyDir; a ref Kaniko can read
+	// in place (s3://, or a path an installer pre-mounted) passes through untouched.
+	contextPath := p.ContextRef
+	initContainers := []corev1.Container{}
+	var kanikoMounts []corev1.VolumeMount
+	var podVolumes []corev1.Volume
+	if isHTTPContextRef(p.ContextRef) {
+		if p.FelisImage == "" {
+			return nil, fmt.Errorf("build: context ref %q needs FelisImage for the fetch initContainer", p.ContextRef)
+		}
+		contextPath = contextMountPath
+		fetch := corev1.Container{
+			Name:  ContainerFetch,
+			Image: p.FelisImage,
+			Args: []string{
+				"fetch-context",
+				"--url=" + p.ContextRef,
+				"--out=" + contextMountPath,
+			},
+			// The internal face is service-token gated, and the token is read from a
+			// Secret the installer materializes in THIS namespace (secretKeyRef is
+			// namespace-local). It is mounted into this initContainer only: the Kaniko
+			// container executes the untrusted Dockerfile and must never hold it, and
+			// pod containers share neither environment nor PID namespace.
+			Env: []corev1.EnvVar{{
+				Name: "FELIS_SERVICE_TOKEN",
+				ValueFrom: &corev1.EnvVarSource{SecretKeyRef: &corev1.SecretKeySelector{
+					LocalObjectReference: corev1.LocalObjectReference{Name: naming.ServiceTokenSecretName},
+					Key:                  naming.ServiceTokenSecretKey,
+				}},
+			}},
+			VolumeMounts:    []corev1.VolumeMount{{Name: contextVolume, MountPath: contextMountPath}},
+			Resources:       corev1.ResourceRequirements{Limits: limits, Requests: limits},
+			SecurityContext: sec,
+		}
+		initContainers = append(initContainers, fetch)
+		kanikoMounts = []corev1.VolumeMount{{Name: contextVolume, MountPath: contextMountPath, ReadOnly: true}}
+		podVolumes = []corev1.Volume{{
+			Name: contextVolume,
+			VolumeSource: corev1.VolumeSource{EmptyDir: &corev1.EmptyDirVolumeSource{
+				// The extracted tree is attacker-controlled; bound it so a tarball
+				// bomb wedges THIS pod (admitted failure) instead of filling the
+				// node's disk. The compressed upload is capped at 1 GiB by the
+				// submit lane, and 4 GiB leaves room for a typical expansion.
+				SizeLimit: sizeLimitPtr(),
+			}},
+		}}
+	}
+
 	kaniko := corev1.Container{
 		Name:  ContainerKaniko,
 		Image: p.KanikoImage,
 		Args: []string{
 			"--dockerfile=Dockerfile",
-			"--context=" + p.ContextRef,
+			"--context=" + contextPath,
 			"--destination=" + p.ImageRef,
 			// The internal registry is in-cluster only and may serve plain HTTP;
 			// it is never a public ingress (spec §17).
 			"--insecure",
 			"--skip-tls-verify",
 		},
+		VolumeMounts:    kanikoMounts,
 		Resources:       corev1.ResourceRequirements{Limits: limits, Requests: limits},
 		SecurityContext: sec,
 	}
+	initContainers = append(initContainers, kaniko)
 
 	trivy := corev1.Container{
 		Name:  ContainerTrivy,
@@ -147,8 +221,9 @@ func BuildJob(p JobParams) (*batchv1.Job, error) {
 					RestartPolicy:                corev1.RestartPolicyNever,
 					ServiceAccountName:           p.ServiceAccount,
 					AutomountServiceAccountToken: boolPtr(false),
-					InitContainers:               []corev1.Container{kaniko},
+					InitContainers:               initContainers,
 					Containers:                   []corev1.Container{trivy},
+					Volumes:                      podVolumes,
 				},
 			},
 		},
@@ -156,11 +231,24 @@ func BuildJob(p JobParams) (*batchv1.Job, error) {
 	return job, nil
 }
 
+// isHTTPContextRef reports whether ref is an http(s) URL — the shape the submit
+// lane derives when the API is the blob transport — i.e. a context only the
+// fetch initContainer can turn into a local path for Kaniko.
+func isHTTPContextRef(ref string) bool {
+	return strings.HasPrefix(ref, "http://") || strings.HasPrefix(ref, "https://")
+}
+
 // NetPolParams parameterises the build-namespace egress lock.
 type NetPolParams struct {
 	Namespace         string
 	RegistryNamespace string
 	RegistryPort      int32
+	// ControlNamespace and APIPort are where the felis-api internal face lives:
+	// the fetch initContainer's only egress besides DNS and the registry. Both
+	// defaults (felis, 8081) match platform.DefaultControlNamespace and the
+	// internal listener, so an unset Params is still the safe shape.
+	ControlNamespace string
+	APIPort          int32
 	// PackageSourceCIDRs is an optional, explicit allowlist of external package
 	// mirrors (spec §16: egress 仅 registry + 包源). Empty means the most
 	// locked-down default — no internet egress at all (默认拒外网).
@@ -177,10 +265,19 @@ func BuildNetworkPolicy(p NetPolParams) *networkingv1.NetworkPolicy {
 	if port == 0 {
 		port = 5000
 	}
+	controlNS := p.ControlNamespace
+	if controlNS == "" {
+		controlNS = "felis"
+	}
+	apiPort := p.APIPort
+	if apiPort == 0 {
+		apiPort = 8081
+	}
 	dnsUDP := corev1.ProtocolUDP
 	dnsTCP := corev1.ProtocolTCP
 	dns53 := intstr.FromInt32(53)
 	regPort := intstr.FromInt32(port)
+	ctxPort := intstr.FromInt32(apiPort)
 
 	egress := []networkingv1.NetworkPolicyEgressRule{
 		// DNS resolution: port-restricted to 53, so this is not an open-internet
@@ -203,6 +300,21 @@ func BuildNetworkPolicy(p NetPolParams) *networkingv1.NetworkPolicy {
 				{Protocol: &dnsTCP, Port: &regPort},
 			},
 		},
+		// felis-api's internal face, where the fetch initContainer streams the
+		// submission's build context from. Without this rule the build Pod could
+		// not read the context and every user build would fail in its first init
+		// step — the default-deny here is exactly why the transport had to be
+		// planned, not assumed.
+		{
+			To: []networkingv1.NetworkPolicyPeer{{
+				NamespaceSelector: &metav1.LabelSelector{
+					MatchLabels: map[string]string{"kubernetes.io/metadata.name": controlNS},
+				},
+			}},
+			Ports: []networkingv1.NetworkPolicyPort{
+				{Protocol: &dnsTCP, Port: &ctxPort},
+			},
+		},
 	}
 	// Explicit package-mirror CIDRs, when configured. No CIDR ⇒ no internet.
 	for _, cidr := range p.PackageSourceCIDRs {
@@ -281,4 +393,12 @@ func resourceLimits(cpu, mem string) (corev1.ResourceList, error) {
 
 func boolPtr(b bool) *bool    { return &b }
 func int32Ptr(i int32) *int32 { return &i }
+
+// sizeLimitPtr returns a copy of contextSizeLimit for a VolumeSource (the API
+// object only ever gets serialized, but a shared pointer across rendered Jobs
+// invites accidental aliasing).
+func sizeLimitPtr() *resource.Quantity {
+	q := contextSizeLimit
+	return &q
+}
 func int64Ptr(i int64) *int64 { return &i }

internal/build/jobspec_test.go

@@ -177,6 +177,132 @@ func TestBuildNetworkPolicyIsDefaultDeny(t *testing.T) {
 	if !egressAllowsPort(np, 53) {
 		t.Error("egress must allow DNS (port 53)")
 	}
+	// The context fetch: build Pods stream submissions from the control
+	// namespace's internal face (defaults: felis + 8081).
+	if !egressAllowsNamespace(np, "felis") {
+		t.Error("egress must allow the control namespace (context fetch)")
+	}
+	if !egressAllowsPort(np, 8081) {
+		t.Error("egress must allow the internal face's port (8081)")
+	}
+}
+
+// An http(s) context ref — the submit lane's derived shape — must render the
+// fetch initContainer ahead of Kaniko, with the service token mounted ONLY into
+// that container, and hand Kaniko the extracted local directory.
+func TestBuildJobFetchesHTTPContext(t *testing.T) {
+	p := sampleJobParams()
+	p.ContextRef = "http://felis-api-internal.felis.svc.cluster.local:8081/api/v1/internal/submissions/sub-abc/context"
+	p.FelisImage = "felis:test"
+	job, err := BuildJob(p)
+	if err != nil {
+		t.Fatalf("BuildJob: %v", err)
+	}
+	inits := job.Spec.Template.Spec.InitContainers
+	if len(inits) != 2 || inits[0].Name != ContainerFetch || inits[1].Name != ContainerKaniko {
+		t.Fatalf("initContainers = %v, want [%s %s]", initNames(inits), ContainerFetch, ContainerKaniko)
+	}
+	fetch, kaniko := inits[0], inits[1]
+	if fetch.Image != p.FelisImage {
+		t.Errorf("fetch image = %q, want the platform image %q", fetch.Image, p.FelisImage)
+	}
+	if !hasArg(fetch.Args, "fetch-context") || !hasArg(fetch.Args, "--url="+p.ContextRef) ||
+		!hasArg(fetch.Args, "--out="+contextMountPath) {
+		t.Errorf("fetch args = %v, want fetch-context --url=%s --out=%s", fetch.Args, p.ContextRef, contextMountPath)
+	}
+	// The token comes from the namespace-local Secret and is mounted into the
+	// fetch container only — never into Kaniko, which executes the untrusted
+	// Dockerfile.
+	var fetchToken *corev1.EnvVar
+	for i := range fetch.Env {
+		if fetch.Env[i].Name == "FELIS_SERVICE_TOKEN" {
+			fetchToken = &fetch.Env[i]
+		}
+	}
+	if fetchToken == nil || fetchToken.ValueFrom == nil || fetchToken.ValueFrom.SecretKeyRef == nil {
+		t.Fatalf("fetch container must read FELIS_SERVICE_TOKEN from a secretKeyRef, got %#v", fetchToken)
+	}
+	if fetchToken.Value != "" {
+		t.Error("fetch container must not carry a literal token")
+	}
+	if len(kaniko.Env) != 0 {
+		t.Errorf("kaniko must carry no env (especially no token), got %v", kaniko.Env)
+	}
+	if !hasArg(kaniko.Args, "--context="+contextMountPath) {
+		t.Errorf("kaniko context = %v, want the fetched local dir %s", kaniko.Args, contextMountPath)
+	}
+	// The shared emptyDir must exist, be bounded, and be read-only to Kaniko.
+	var ctxVol *corev1.Volume
+	for i := range job.Spec.Template.Spec.Volumes {
+		if job.Spec.Template.Spec.Volumes[i].Name == contextVolume {
+			ctxVol = &job.Spec.Template.Spec.Volumes[i]
+		}
+	}
+	if ctxVol == nil || ctxVol.EmptyDir == nil || ctxVol.EmptyDir.SizeLimit == nil {
+		t.Fatalf("context volume must be a size-limited emptyDir, got %#v", ctxVol)
+	}
+	mountedRO := false
+	for _, m := range kaniko.VolumeMounts {
+		if m.Name == contextVolume && m.MountPath == contextMountPath && m.ReadOnly {
+			mountedRO = true
+		}
+	}
+	if !mountedRO {
+		t.Errorf("kaniko must mount the context read-only at %s, got %v", contextMountPath, kaniko.VolumeMounts)
+	}
+}
+
+// Without the platform image the fetch initContainer cannot run, so rendering an
+// http(s) context must fail loudly at Job-creation time, not with an ImagePull
+// error at 3am.
+func TestBuildJobHTTPContextNeedsFelisImage(t *testing.T) {
+	p := sampleJobParams()
+	p.ContextRef = "https://example.invalid/sub-abc/context"
+	if _, err := BuildJob(p); err == nil {
+		t.Fatal("http(s) context without FelisImage must fail to render")
+	}
+}
+
+// A ref Kaniko reads natively (or an installer pre-mounted) must NOT grow the
+// fetch initContainer: the transport is for http(s) only.
+func TestBuildJobNativeContextNeedsNoFetch(t *testing.T) {
+	p := sampleJobParams()
+	p.ContextRef = "s3://bucket/prefix/context.tar.gz"
+	job, err := BuildJob(p)
+	if err != nil {
+		t.Fatalf("BuildJob: %v", err)
+	}
+	if len(job.Spec.Template.Spec.InitContainers) != 1 || job.Spec.Template.Spec.InitContainers[0].Name != ContainerKaniko {
+		t.Errorf("a native ref must render just kaniko, got %v", initNames(job.Spec.Template.Spec.InitContainers))
+	}
+	if len(job.Spec.Template.Spec.Volumes) != 0 {
+		t.Errorf("a native ref must render no context volume, got %v", job.Spec.Template.Spec.Volumes)
+	}
+}
+
+func initNames(cs []corev1.Container) []string {
+	names := make([]string, 0, len(cs))
+	for _, c := range cs {
+		names = append(names, c.Name)
+	}
+	return names
+}
+
+// An explicit control namespace/port override must reach the egress rule (a
+// renamed control namespace otherwise silently blocks every context fetch).
+func TestBuildNetworkPolicyHonoursControlNamespaceOverride(t *testing.T) {
+	np := BuildNetworkPolicy(NetPolParams{
+		Namespace:         "felis-build",
+		RegistryNamespace: "felis-system",
+		ControlNamespace:  "control-plane",
+		APIPort:           9081,
+	})
+	if !egressAllowsNamespace(np, "control-plane") {
+		t.Error("egress must allow the overridden control namespace")
+	}
+	if !egressAllowsPort(np, 9081) {
+		t.Error("egress must allow the overridden api port")
+	}
 }
 
 // With no package-source CIDRs configured, there must be zero IPBlock egress —

internal/naming/naming.go

@@ -56,6 +56,10 @@ const (
 const (
 	ServiceTokenSecretName = "felis-service-token"
 	ServiceTokenSecretKey  = "token"
+	// EnvAPIBaseURL carries the internal-face base URL (platform.InternalAPIBaseURL)
+	// into a pod: the login gate dials it, and the api reads it to derive the build
+	// contexts' fetch URLs, so both sides name the same address for the same face.
+	EnvAPIBaseURL = "FELIS_API_BASE_URL"
 )
 
 // ForwardingSecretName / ForwardingSecretKey name the Velocity modern player-info

internal/platform/bundle.go

@@ -78,9 +78,12 @@ func Objects(p Params) []Object {
 
 	// Build-namespace egress lock (reused from internal/build; TypeMeta stamped).
 	buildNP := build.BuildNetworkPolicy(build.NetPolParams{
-		Namespace:          p.BuildNamespace,
+		Namespace:         p.BuildNamespace,
-		RegistryNamespace:  p.RegistryNamespace,
+		RegistryNamespace: p.RegistryNamespace,
-		RegistryPort:       p.RegistryPort,
+		RegistryPort:      p.RegistryPort,
+		ControlNamespace:  p.ControlNamespace,
+		// The internal face's port, single-sourced with the api Deployment below.
+		APIPort:            apiInternalPort,
 		PackageSourceCIDRs: p.PackageSourceCIDRs,
 	})
 	buildNP.TypeMeta = metav1.TypeMeta{APIVersion: "networking.k8s.io/v1", Kind: "NetworkPolicy"}

internal/platform/workloads.go

@@ -274,6 +274,10 @@ func APIDeployment(p Params) *appsv1.Deployment {
 			},
 		},
 		{Name: "FELIS_IMAGE", Value: p.FelisImage},
+		// The api's own internal-face base URL, so it derives the submission
+		// context URLs that build Pods fetch through it. Same value the login gate
+		// is handed; one address for one face.
+		{Name: naming.EnvAPIBaseURL, Value: InternalAPIBaseURL(p.ControlNamespace)},
 	}
 	if p.BackupPVC != "" {
 		env = append(env, corev1.EnvVar{Name: "FELIS_BACKUP_PVC", Value: p.BackupPVC})

internal/platform/workloads_test.go

@@ -179,6 +179,11 @@ func TestAPIDeployment_Wiring(t *testing.T) {
 	if v := envValue(c.Env, "FELIS_IMAGE"); v != p.FelisImage {
 		t.Errorf("FELIS_IMAGE = %q, want %q", v, p.FelisImage)
 	}
+	// The internal-face base URL the api derives submission context-fetch URLs
+	// from — the same address the login gate is handed.
+	if v := envValue(c.Env, "FELIS_API_BASE_URL"); v != InternalAPIBaseURL(p.ControlNamespace) {
+		t.Errorf("FELIS_API_BASE_URL = %q, want %q", v, InternalAPIBaseURL(p.ControlNamespace))
+	}
 	// FELIS_SERVICE_TOKEN must come from a Secret, never a literal value.
 	tok := envVar(c.Env, "FELIS_SERVICE_TOKEN")
 	if tok == nil || tok.ValueFrom == nil || tok.ValueFrom.SecretKeyRef == nil {

internal/submit/blobstore.go

@@ -111,5 +111,23 @@ func (s *LocalContextStore) Exists(_ context.Context, id string) (bool, error) {
 	}
 }
 
+// Open returns the stored context blob for id — the read side of the transport the
+// build Pod's fetch initContainer uses. A missing blob is ErrBlobNotFound (404 on
+// the route), never a bare os error, so the API keeps its status mapping.
+func (s *LocalContextStore) Open(_ context.Context, id string) (io.ReadCloser, error) {
+	dir, err := s.dir(id)
+	if err != nil {
+		return nil, err
+	}
+	f, err := os.Open(filepath.Join(dir, contextBlobName))
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, fmt.Errorf("%w: %v", ErrBlobNotFound, err)
+		}
+		return nil, fmt.Errorf("submit: open context blob: %w", err)
+	}
+	return f, nil
+}
+
 // Compile-time proof that the filesystem store satisfies the Blobs transport.
 var _ Blobs = (*LocalContextStore)(nil)

internal/submit/blobstore_test.go

@@ -2,6 +2,8 @@ package submit
 
 import (
 	"context"
+	"errors"
+	"io"
 	"os"
 	"path/filepath"
 	"strings"
@@ -54,6 +56,39 @@ func TestLocalContextStorePutAndExists(t *testing.T) {
 	}
 }
 
+// Open is the internal context-fetch route's read path: it serves exactly the
+// stored bytes, and a missing blob is ErrBlobNotFound (404), never a bare os error.
+func TestLocalContextStoreOpen(t *testing.T) {
+	base := t.TempDir()
+	s := &LocalContextStore{Base: base}
+	ctx := context.Background()
+
+	if _, err := s.Open(ctx, "sub-gone"); !errors.Is(err, ErrBlobNotFound) {
+		t.Fatalf("Open of a missing blob = %v, want ErrBlobNotFound", err)
+	}
+
+	payload := "\x1f\x8b\x08\x00the modpack context"
+	if _, err := s.Put(ctx, "sub-abc", strings.NewReader(payload)); err != nil {
+		t.Fatalf("Put: %v", err)
+	}
+	rc, err := s.Open(ctx, "sub-abc")
+	if err != nil {
+		t.Fatalf("Open: %v", err)
+	}
+	defer rc.Close()
+	got, err := io.ReadAll(rc)
+	if err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	if string(got) != payload {
+		t.Fatalf("Open served %q, want %q", got, payload)
+	}
+	// The same path guard as Put: an id that could escape Base is refused.
+	if _, err := s.Open(ctx, "../etc/passwd"); err == nil {
+		t.Fatal("Open must reject an unsafe id")
+	}
+}
+
 func TestLocalContextStorePutOverwrites(t *testing.T) {
 	base := t.TempDir()
 	s := &LocalContextStore{Base: base}

internal/submit/s3store.go

@@ -14,11 +14,30 @@ import (
 )
 
 // s3Client is the minimal object-store surface S3ContextStore needs. *minio.Client
-// satisfies it, and a fake satisfies it in tests — so the store's key derivation
+// satisfies it through minioStoreClient, and a fake satisfies it in tests — so the
-// and not-found handling are unit-verifiable without a live bucket.
+// store's key derivation and not-found handling are unit-verifiable without a live
+// bucket.
 type s3Client interface {
 	PutObject(ctx context.Context, bucket, object string, reader io.Reader, size int64, opts minio.PutObjectOptions) (minio.UploadInfo, error)
 	StatObject(ctx context.Context, bucket, object string, opts minio.StatObjectOptions) (minio.ObjectInfo, error)
+	GetObject(ctx context.Context, bucket, object string, opts minio.GetObjectOptions) (s3Object, error)
+}
+
+// s3Object is the handle GetObject yields: a stream whose Stat performs the HEAD
+// eagerly, so a missing object surfaces before the first byte is read.
+type s3Object interface {
+	io.ReadCloser
+	Stat() (minio.ObjectInfo, error)
+}
+
+// minioStoreClient adapts *minio.Client to s3Client. The adapter exists because a
+// method's return type cannot be narrowed by an interface: GetObject on the real
+// client returns a concrete *minio.Object, which does not satisfy a method declared
+// to return s3Object.
+type minioStoreClient struct{ *minio.Client }
+
+func (m minioStoreClient) GetObject(ctx context.Context, bucket, object string, opts minio.GetObjectOptions) (s3Object, error) {
+	return m.Client.GetObject(ctx, bucket, object, opts)
 }
 
 // S3ContextStore is the object-store-backed build-context blob store: it writes
@@ -27,16 +46,17 @@ type s3Client interface {
 // selected by cmd/felis when user_uploads_context is an s3:// base.
 //
 // The bucket + key prefix are parsed from that same base (parseS3Base), so an
-// object written here lands at exactly s3://{bucket}/{prefix}/{id}/context.tar.gz —
+// object written here lands at exactly s3://{bucket}/{prefix}/{id}/context.tar.gz.
-// the ref deriveContextRef records and Kaniko's native s3:// --context reads.
 // Credentials are static V4 keys resolved by cmd/felis from the environment (the
 // setup wizard injects them into felis-api from the felis-uploads-s3 Secret); they
 // never touch felis.toml.
 //
-// Kaniko reading the S3 context at build time needs its own credentials + egress
+// The sandboxed build Job never needs S3 credentials of its own: the api reads the
-// on the sandboxed build Job — a separate deployment integration, exactly like the
+// object back here (Open) and streams it over the internal face, which is the
-// LocalContextStore PVC mount. This transport only makes the upload durable at the
+// transport every in-cluster build uses (Manager.ContextBaseURL). Only a
-// derived location.
+// deployment that leaves ContextBaseURL empty would fall back to Kaniko reading
+// s3:// natively — and such a deployment would still need to hand the build Pod
+// credentials + egress itself.
 type S3ContextStore struct {
 	client s3Client
 	bucket string
@@ -79,7 +99,7 @@ func NewS3ContextStore(cfg S3StoreConfig) (*S3ContextStore, error) {
 	if err != nil {
 		return nil, fmt.Errorf("submit: s3 client: %w", err)
 	}
-	return &S3ContextStore{client: client, bucket: bucket, prefix: prefix}, nil
+	return &S3ContextStore{client: minioStoreClient{client}, bucket: bucket, prefix: prefix}, nil
 }
 
 // CheckS3Access verifies the S3 coordinates before they are committed to config:
@@ -167,6 +187,35 @@ func (s *S3ContextStore) Exists(ctx context.Context, id string) (bool, error) {
 	return true, nil
 }
 
+// Open returns the stored context blob for id — the read side of the transport the
+// build Pod's fetch initContainer uses. minio's GetObject returns only once the
+// server answered with an object (it surfaces NoSuchKey up front), so a missing
+// object maps to ErrBlobNotFound right here and the route answers 404.
+func (s *S3ContextStore) Open(ctx context.Context, id string) (io.ReadCloser, error) {
+	key, err := s.keyFor(id)
+	if err != nil {
+		return nil, err
+	}
+	obj, err := s.client.GetObject(ctx, s.bucket, key, minio.GetObjectOptions{})
+	if err != nil {
+		if isS3NotFound(err) {
+			return nil, fmt.Errorf("%w: %v", ErrBlobNotFound, err)
+		}
+		return nil, fmt.Errorf("submit: open context blob: %w", err)
+	}
+	// minio.Object is lazy: the first Read triggers the GET and is where a missing
+	// key actually surfaces, so stat it once here to translate that case eagerly
+	// (the caller can then trust the io.ReadCloser belongs to a real object).
+	if _, err := obj.Stat(); err != nil {
+		_ = obj.Close()
+		if isS3NotFound(err) {
+			return nil, fmt.Errorf("%w: %v", ErrBlobNotFound, err)
+		}
+		return nil, fmt.Errorf("submit: open context blob: %w", err)
+	}
+	return obj, nil
+}
+
 // isS3NotFound recognizes the "object is absent" outcome across S3
 // implementations: a GET-shaped NoSuchKey code or a bare 404 from the HEAD that
 // StatObject issues.

internal/submit/s3store_test.go

@@ -45,6 +45,41 @@ func (f *fakeS3) StatObject(_ context.Context, bucket, object string, _ minio.St
 	return minio.ObjectInfo{}, minio.ErrorResponse{Code: "NoSuchKey", StatusCode: http.StatusNotFound}
 }
 
+// fakeS3Object is the object handle fakeS3.GetObject yields: Stat mirrors
+// StatObject's not-found behaviour, Read serves the stored bytes.
+type fakeS3Object struct {
+	data []byte
+	err  error
+}
+
+func (o *fakeS3Object) Read(p []byte) (int, error) {
+	if o.err != nil {
+		return 0, o.err
+	}
+	if len(o.data) == 0 {
+		return 0, io.EOF
+	}
+	n := copy(p, o.data)
+	o.data = o.data[n:]
+	return n, nil
+}
+
+func (o *fakeS3Object) Close() error { return nil }
+
+func (o *fakeS3Object) Stat() (minio.ObjectInfo, error) {
+	if o.err != nil {
+		return minio.ObjectInfo{}, o.err
+	}
+	return minio.ObjectInfo{Size: int64(len(o.data))}, nil
+}
+
+func (f *fakeS3) GetObject(_ context.Context, bucket, object string, _ minio.GetObjectOptions) (s3Object, error) {
+	if data, ok := f.objects[bucket+"/"+object]; ok {
+		return &fakeS3Object{data: append([]byte(nil), data...)}, nil
+	}
+	return &fakeS3Object{err: minio.ErrorResponse{Code: "NoSuchKey", StatusCode: http.StatusNotFound}}, nil
+}
+
 func TestCheckBucketAccess(t *testing.T) {
 	ctx := context.Background()
 
@@ -107,6 +142,34 @@ func TestS3ContextStorePutAndExists(t *testing.T) {
 	}
 }
 
+// Open serves the stored object's bytes and maps a missing key to ErrBlobNotFound
+// (the internal fetch route's 404), eagerly — before the caller reads a byte.
+func TestS3ContextStoreOpen(t *testing.T) {
+	fake := &fakeS3{}
+	s := &S3ContextStore{client: fake, bucket: "felis-uploads", prefix: "builds"}
+	ctx := context.Background()
+
+	if _, err := s.Open(ctx, "sub-gone"); !errors.Is(err, ErrBlobNotFound) {
+		t.Fatalf("Open of a missing object = %v, want ErrBlobNotFound", err)
+	}
+	payload := "\x1f\x8b\x08\x00the modpack context"
+	if _, err := s.Put(ctx, "sub-abc", strings.NewReader(payload)); err != nil {
+		t.Fatalf("Put: %v", err)
+	}
+	rc, err := s.Open(ctx, "sub-abc")
+	if err != nil {
+		t.Fatalf("Open: %v", err)
+	}
+	defer rc.Close()
+	got, err := io.ReadAll(rc)
+	if err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	if string(got) != payload {
+		t.Fatalf("Open served %q, want %q", got, payload)
+	}
+}
+
 func TestS3ContextStoreEmptyPrefix(t *testing.T) {
 	fake := &fakeS3{}
 	s := &S3ContextStore{client: fake, bucket: "b", prefix: ""}

internal/submit/submit.go

@@ -37,7 +37,8 @@
 //     chosen ref would let an untrusted origin point the build at an arbitrary
 //     source. By deriving it from the submission id the user selects nothing
 //     that reaches the executor — only the modpack blob behind the pinned,
-//     id-namespaced location (uploaded by a separate, deferred transport).
+//     id-namespaced location (uploaded through the Blobs transport, and served
+//     back to the build Pod over the service-token-gated internal face).
 //
 // Source of truth. submissions is a NEW Postgres business-truth domain, added to
 // §1 invariant 2's enumeration (owner/claim/accounts/audit/quota/images/builds/
@@ -91,6 +92,10 @@ var (
 	// an honest 503, never a 500, exactly as the restore executor does when its
 	// integration is not wired.
 	ErrUploadsUnavailable = errors.New("submit: context upload transport not configured")
+	// ErrBlobNotFound reports that a submission has no stored context blob (or it
+	// was never uploaded). The internal context-fetch route maps it to 404, the
+	// same distinction Exists draws for Approve.
+	ErrBlobNotFound = errors.New("submit: context blob not found")
 )
 
 // invalidf wraps ErrInvalid so every malformed-request case maps to one 400.
@@ -176,11 +181,12 @@ type Builds interface {
 
 // Blobs is the build-context blob transport the lane depends on to place a
 // submitter's uploaded modpack at the platform-derived, id-namespaced location
-// deriveContextRef points Kaniko at. It is the piece the package doc calls a
+// deriveContextRef points Kaniko at. Creation only derives and records the ref;
-// "separate, deferred transport": creation only derives and records the ref, and
+// the bytes behind it arrive through Put here, and the build Pod reads them back
-// the bytes behind it arrive through Put here. It is an interface so the Manager
+// through Open (the manager exposes it as OpenContext, which the API's internal
-// is unit-tested against an in-memory fake; the production implementation is the
+// context route serves). It is an interface so the Manager is unit-tested against
-// filesystem-backed LocalContextStore.
+// an in-memory fake; the production implementations are LocalContextStore
+// (filesystem) and S3ContextStore (object store).
 //
 // Both methods key off the submission id, never a caller-supplied path, so the
 // write target is as platform-pinned as the derived ref itself. Put stores (and
@@ -190,6 +196,11 @@ type Builds interface {
 type Blobs interface {
 	Put(ctx context.Context, id string, r io.Reader) (int64, error)
 	Exists(ctx context.Context, id string) (bool, error)
+	// Open returns the stored blob's bytes for the internal context-fetch route
+	// the build Pod's initContainer dials (cmd/felis fetch-context). It returns an
+	// error wrapping ErrBlobNotFound when no blob exists, so the route can answer
+	// 404 without leaking which ids do exist.
+	Open(ctx context.Context, id string) (io.ReadCloser, error)
 }
 
 // Manager orchestrates the approval lane. It holds no mutable state; the clock
@@ -210,6 +221,15 @@ type Manager struct {
 	// "s3://felis-user-uploads" (an object store) or a local uploads PVC path. The
 	// derived context ref is {ContextStore}/{id}/context.tar.gz.
 	ContextStore string
+	// ContextBaseURL, when set, is the platform's internal-face base URL
+	// (platform.InternalAPIBaseURL). It makes the derived context ref an HTTP URL
+	// on that face — {ContextBaseURL}/api/v1/internal/submissions/{id}/context —
+	// instead of a filesystem/object-store location: the build Pod cannot mount
+	// the uploads PVC (builds run in another namespace) and carries no object-store
+	// credentials, so the API streams the blob it stored at ContextStore over the
+	// service-token-gated internal face. Empty keeps the legacy ref shape for a
+	// deployment that predates the transport.
+	ContextBaseURL string
 	// Blobs is the upload transport that persists the modpack behind the derived
 	// context ref. When nil (a store with no implemented transport, e.g. an
 	// object-store base with no client), UploadContext returns ErrUploadsUnavailable
@@ -269,9 +289,23 @@ func (m *Manager) deriveImageRef(id string) string {
 // selects nothing that reaches Kaniko's --context argument; only the blob behind
 // this pinned, id-namespaced location (placed by the upload transport) varies.
 func (m *Manager) deriveContextRef(id string) string {
+	if m.ContextBaseURL != "" {
+		return fmt.Sprintf("%s/api/v1/internal/submissions/%s/context", strings.TrimRight(m.ContextBaseURL, "/"), id)
+	}
 	return fmt.Sprintf("%s/%s/%s", strings.TrimRight(m.ContextStore, "/"), id, contextBlobName)
 }
 
+// OpenContext returns the stored build context for id — the read path behind the
+// internal context-fetch route. It requires the upload transport (Blobs): with no
+// transport there is no blob to read, so it reports ErrUploadsUnavailable, the
+// same honest 503 the upload endpoint gives.
+func (m *Manager) OpenContext(ctx context.Context, id string) (io.ReadCloser, error) {
+	if m.Blobs == nil {
+		return nil, ErrUploadsUnavailable
+	}
+	return m.Blobs.Open(ctx, id)
+}
+
 // auditDockerfile is the audit-archive Dockerfile recorded on the build row. It
 // is NOT what Kaniko executes — Kaniko reads the real Dockerfile from inside the
 // uploaded context (build/jobspec.go) — so this honestly documents the

internal/submit/submit_test.go

@@ -1,8 +1,10 @@
 package submit
 
 import (
+	"bytes"
 	"context"
 	"errors"
+	"fmt"
 	"io"
 	"strings"
 	"testing"
@@ -50,6 +52,14 @@ func (f *fakeBlobs) Exists(_ context.Context, id string) (bool, error) {
 	return ok, nil
 }
 
+func (f *fakeBlobs) Open(_ context.Context, id string) (io.ReadCloser, error) {
+	b, ok := f.stored[id]
+	if !ok {
+		return nil, fmt.Errorf("%w: no blob for %s", ErrBlobNotFound, id)
+	}
+	return io.NopCloser(bytes.NewReader(b)), nil
+}
+
 // testNow is the frozen clock for hermetic assertions.
 var testNow = time.Date(2026, 1, 1, 12, 0, 0, 0, time.UTC)
 
@@ -217,6 +227,52 @@ func TestCreatePendingDoesNotBuild(t *testing.T) {
 	}
 }
 
+// With a ContextBaseURL the derived ref is the internal-face URL the build Pod's
+// fetch initContainer dials — not a filesystem path it could never read across
+// namespaces. OpenContext then serves whatever the Blobs transport stored.
+func TestContextRefIsFetchURLAndOpenContextServesIt(t *testing.T) {
+	m, _, _ := newManager()
+	m.ContextBaseURL = "http://felis-api-internal.felis.svc.cluster.local:8081/"
+	m.Blobs = newFakeBlobs()
+	ctx := context.Background()
+
+	sub, err := m.Create(ctx, CreateRequest{DisplayName: "Pack", SubmittedBy: "user-1"})
+	if err != nil {
+		t.Fatalf("Create: %v", err)
+	}
+	want := "http://felis-api-internal.felis.svc.cluster.local:8081/api/v1/internal/submissions/sub-1/context"
+	if sub.ContextRef != want {
+		t.Fatalf("context_ref = %q, want the internal fetch URL %q", sub.ContextRef, want)
+	}
+
+	// Before any upload the read path reports not-found (the route's 404).
+	if _, err := m.OpenContext(ctx, sub.ID); !errors.Is(err, ErrBlobNotFound) {
+		t.Fatalf("OpenContext before upload = %v, want ErrBlobNotFound", err)
+	}
+	payload := "\x1f\x8b\x08\x00payload"
+	if _, err := m.UploadContext(ctx, sub.ID, "user-1", strings.NewReader(payload)); err != nil {
+		t.Fatalf("UploadContext: %v", err)
+	}
+	rc, err := m.OpenContext(ctx, sub.ID)
+	if err != nil {
+		t.Fatalf("OpenContext: %v", err)
+	}
+	defer rc.Close()
+	got, _ := io.ReadAll(rc)
+	if string(got) != payload {
+		t.Fatalf("OpenContext served %q, want %q", got, payload)
+	}
+}
+
+// No upload transport ⇒ no readable blob: the route reports the same 503 the
+// upload endpoint does, rather than a misleading 404.
+func TestOpenContextWithoutTransportIsUnavailable(t *testing.T) {
+	m, _, _ := newManager()
+	if _, err := m.OpenContext(context.Background(), "sub-1"); !errors.Is(err, ErrUploadsUnavailable) {
+		t.Fatalf("OpenContext with nil Blobs = %v, want ErrUploadsUnavailable", err)
+	}
+}
+
 func TestCreateValidation(t *testing.T) {
 	cases := []struct {
 		name string