Ask for a token before spending the card (V-673)

mavgpud reverse-proxied every path to llama-server with no authentication on
a LAN port. Any client on the network could submit model work, hold the model
resident by touching the idle clock, and read /slots, which returns the
prompts of whoever else was using the card.

It now reads a bearer token from token_file and requires it on every request,
/health included: /health reports whether the card is loaded and free, which
is what someone deciding to take it would ask. A listen address reachable
from the network with no token is a startup failure rather than a downgrade
to loopback. homesrv is the client and it is on the LAN, so a loopback
default would look safe and take the model arm down.

Beyond the token: an allowlist of the five paths Maven calls, so a leaked
token buys the model API and not llama-server's admin surface; a body cap and
an in-flight cap on the proxy; and header and idle timeouts on the server.
No read or write timeout — a completion on this card legitimately takes
minutes.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ESv8hqNPseYt1CnotZpqDz

cmd/mavgpud/main.go

@@ -36,6 +36,21 @@ type config struct {
 	Listen    string `json:"listen"`     // what Maven talks to
 	LlamaAddr string `json:"llama_addr"` // where llama-server binds
 	LlamaBin  string `json:"llama_bin"`
+
+	// TokenFile holds the bearer token every request must carry. It is a path
+	// and never the token itself, the rule mavpoll, mavmaild and the CW2
+	// transcriber already follow: a secret in a committed config is a secret
+	// in the history. Empty is allowed only on a loopback Listen, and
+	// requireToken is where that is decided.
+	TokenFile string `json:"token_file,omitempty"`
+
+	// MaxBody bounds a proxied request body. A completion is a prompt, and a
+	// prompt that does not fit here would not fit the context window either.
+	MaxBody int64 `json:"max_body_bytes,omitempty"`
+	// MaxInflight bounds how many proxied requests reach llama-server at once.
+	// It runs with -np 1, so anything above a handful only queues inside the
+	// child while holding a connection and a body in memory here.
+	MaxInflight int `json:"max_inflight,omitempty"`
 	// LlamaArgs must include the flags that bind LlamaAddr. They are passed
 	// through untouched so the model, context size and layer count stay the
 	// owner's business and not this daemon's schema.
@@ -88,6 +103,8 @@ func defaults() config {
 		MinFreeVRAM: 15 << 30,
 		EvictAfter:  2,
 		StartAfter:  5,
+		MaxBody:     8 << 20,
+		MaxInflight: 4,
 	}
 }
 
@@ -123,6 +140,20 @@ func main() {
 		log.Fatal("mavgpud: llama_bin is required")
 	}
 
+	// A LAN listener with no token is refused rather than downgraded to
+	// loopback. Downgrading would look like a safe default and would take the
+	// model arm down instead: homesrv is the client and it is on the LAN.
+	var token string
+	if cfg.TokenFile != "" {
+		var err error
+		if token, err = readToken(cfg.TokenFile); err != nil {
+			log.Fatalf("mavgpud: %v", err)
+		}
+	} else if !loopbackListen(cfg.Listen) {
+		log.Fatalf("mavgpud: listen %s is reachable from the network and token_file is unset — "+
+			"set token_file, or listen on 127.0.0.1 and accept that Maven cannot reach it", cfg.Listen)
+	}
+
 	base := "http://" + cfg.LlamaAddr
 	run := newRunner("llama-server", cfg.LlamaBin, cfg.LlamaArgs, base+"/health")
 	sup := &supervisor{
@@ -145,7 +176,20 @@ func main() {
 	if err != nil {
 		log.Fatalf("mavgpud: llama_addr: %v", err)
 	}
-	srv := &http.Server{Addr: cfg.Listen, Handler: sup.handler(target)}
+	var h http.Handler = sup.handler(target)
+	if token != "" {
+		h = requireToken(token, cfg.MaxBody, h)
+	}
+	srv := &http.Server{
+		Addr:    cfg.Listen,
+		Handler: h,
+		// A slow-loris client holds a connection and a header buffer for free
+		// otherwise. No ReadTimeout or WriteTimeout: a completion legitimately
+		// takes minutes on this card, and either one would cut it off.
+		ReadHeaderTimeout: 10 * time.Second,
+		IdleTimeout:       60 * time.Second,
+		MaxHeaderBytes:    1 << 16,
+	}
 	go func() {
 		log.Printf("mavgpud: listening on %s, model %s", cfg.Listen, cfg.LlamaBin)
 		if err := srv.ListenAndServe(); err != nil && err != http.ErrServerClosed {
@@ -203,14 +247,14 @@ func (s *supervisor) handler(target *url.URL) http.Handler {
 		w.Header().Set("Content-Type", "application/json")
 		_, _ = w.Write([]byte(`{"status":"ok"}`))
 	})
-	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+	mux.Handle("/", allowlist(limitInflight(s.cfg.MaxInflight, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if !s.run.isReady() {
 			http.Error(w, "model not loaded", http.StatusServiceUnavailable)
 			return
 		}
 		s.touch()
 		proxy.ServeHTTP(w, r)
-	})
+	}))))
 	return mux
 }