From 04616dcf07f25627c592f68060508f9ba3a269cc Mon Sep 17 00:00:00 2001
From: Max Xing <mxing@nvidia.com>
Date: Fri, 24 Jul 2026 11:27:46 -0700
Subject: [PATCH] fix(worker-llm-credentials): write worker-token 0644 so pylon
 can read it

Pylon (llm-router-client) reads /var/run/llm/worker-token to authenticate
its router registration, but it runs as a different non-root UID than the
credential-manager (root) on the shared emptyDir, so it could not read the
0600 token. Registration failed and gateway invokes returned
no_eligible_candidates.

Write the token 0644 and chmod the tmp file to force the mode: os.WriteFile
honors umask and skips the mode when tmp already exists, so a restrictive
container umask or a stale tmp could otherwise leave it 0600. The llm
emptyDir is mounted only by the two infra sidecars, not the customer
workload container.

NO-REF
---
 .../internal/worker/worker.go                    | 12 +++++++++++-
 .../internal/worker/worker_coverage_test.go      | 16 ++++++++++++++++
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/src/compute-plane-services/worker-llm-credentials/internal/worker/worker.go b/src/compute-plane-services/worker-llm-credentials/internal/worker/worker.go
index 7b287a285..86b6c7b80 100644
--- a/src/compute-plane-services/worker-llm-credentials/internal/worker/worker.go
+++ b/src/compute-plane-services/worker-llm-credentials/internal/worker/worker.go
@@ -98,7 +98,17 @@ func (w *Worker) Run(ctx context.Context) error {
 				}
 			}
 			tmp := w.config.WorkerTokenPath + ".tmp"
-			if err := os.WriteFile(tmp, []byte(t.Token), 0600); err != nil {
+			// 0644 so the pylon sidecar (a different non-root UID) can read the
+			// token from the shared emptyDir; only these two infra sidecars
+			// mount it, not the customer workload container.
+			//nolint:gosec // G306: intentionally readable by the pylon sidecar
+			if err := os.WriteFile(tmp, []byte(t.Token), 0644); err != nil {
+				return err
+			}
+			// WriteFile honors umask and skips the mode when tmp already exists;
+			// chmod guarantees 0644 so the sidecar can read it.
+			//nolint:gosec // G302: intentionally readable by the pylon sidecar
+			if err := os.Chmod(tmp, 0644); err != nil {
 				return err
 			}
 			return os.Rename(tmp, w.config.WorkerTokenPath)
diff --git a/src/compute-plane-services/worker-llm-credentials/internal/worker/worker_coverage_test.go b/src/compute-plane-services/worker-llm-credentials/internal/worker/worker_coverage_test.go
index 2829bf54a..f1dc64519 100644
--- a/src/compute-plane-services/worker-llm-credentials/internal/worker/worker_coverage_test.go
+++ b/src/compute-plane-services/worker-llm-credentials/internal/worker/worker_coverage_test.go
@@ -21,6 +21,7 @@ import (
 	"context"
 	"os"
 	"path/filepath"
+	"syscall"
 	"testing"
 	"time"
 
@@ -134,6 +135,11 @@ func TestNew_DefaultsSharedConfigDir(t *testing.T) {
 // subdirectory that does not yet exist, so os.Stat returns ErrNotExist and the
 // callback must create the directory tree before writing the token.
 func TestRun_CreatesMissingTokenDir(t *testing.T) {
+	// A restrictive umask must not strip the token's read bits: Run chmods the
+	// file after writing so the pylon sidecar can always read it.
+	oldUmask := syscall.Umask(0o077)
+	defer syscall.Umask(oldUmask)
+
 	addr := startMockNVCFServer(t)
 	tmpDir := t.TempDir()
 	// Nested, non-existent subdirectory of the temp dir.
@@ -198,6 +204,16 @@ func TestRun_CreatesMissingTokenDir(t *testing.T) {
 	if string(content) != testWorkerToken {
 		t.Fatalf("expected token %q, got %q", testWorkerToken, string(content))
 	}
+
+	// The token must be readable by the pylon sidecar, which runs as a
+	// different UID on the shared emptyDir.
+	tokenInfo, err := os.Stat(workerTokenPath)
+	if err != nil {
+		t.Fatalf("token file not found: %v", err)
+	}
+	if got := tokenInfo.Mode().Perm(); got != 0644 {
+		t.Fatalf("expected token file mode 0644, got %o", got)
+	}
 }
 
 // TestRun_TokenDirCreateFails exercises the error-return of the directory