Merge origin/main into feature/update-vocal_system

Integrates action/target turn system, journal UI, painter and prompt
improvements from main while preserving Kokoro TTS voice system.
Regenerated uv.lock to include kokoro-onnx + soundfile TTS dependencies.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -179,8 +179,8 @@ def build_server():
 
     @app.api(name="turn")
     @gpu
-    def turn(player_input: str) -> dict:
-        return ENGINE.play_turn(player_input).model_dump()
+    def turn(player_input: str, action: str = "talk", target: str = "") -> dict:
+        return ENGINE.play_turn(player_input, action=action, target=target).model_dump()
 
     @app.api(name="transcribe")
     @gpu

frontend/index.html

@@ -58,8 +58,11 @@
   .char-card-name{font-family:var(--display); font-size:1.15rem; color:var(--gold); margin-bottom:4px}
   .char-card-bio{font-size:.82rem; color:var(--ink-dim); margin-bottom:10px; line-height:1.4}
   .rel-row{display:flex; align-items:center; gap:8px; margin-bottom:10px}
-  .rel-bar{flex:1; height:5px; background:rgba(255,255,255,.12); border-radius:3px; overflow:hidden}
-  .rel-fill{height:100%; border-radius:3px; transition:width .5s ease}
+  .rel-bar{flex:1; height:5px; background:rgba(255,255,255,.10); border-radius:3px; overflow:hidden; position:relative}
+  .rel-fill-pos{position:absolute; left:0; top:0; bottom:0; width:0;
+    background:var(--moss); border-radius:3px; transition:width .5s ease}
+  .rel-fill-neg{position:absolute; right:0; top:0; bottom:0; width:0;
+    background:var(--rose); border-radius:3px; transition:width .5s ease}
   .rel-val{font-size:.78rem; color:var(--ink-dim); min-width:28px; text-align:right}
   .traits-section{margin-bottom:8px}
   .traits-label{font-size:.72rem; text-transform:uppercase; letter-spacing:.8px;
@@ -98,13 +101,43 @@
   #place{position:absolute; top:18px; left:22px; z-index:4; font-family:var(--display);
     font-size:1.05rem; color:var(--ink-dim); letter-spacing:1px; text-transform:uppercase;
     text-shadow:0 1px 8px #000}
+  /* ---- target row (who to address) ---- */
+  #target-row{
+    display:flex; align-items:center; gap:6px; margin-bottom:8px; flex-wrap:wrap;
+  }
+  .target-label{font-size:.75rem; color:var(--ink-dim); letter-spacing:.4px; flex-shrink:0}
+  .target-chip{
+    font-family:var(--body); cursor:pointer;
+    border:1px solid rgba(255,255,255,.15); border-radius:20px;
+    padding:4px 12px; font-size:.83rem;
+    background:rgba(255,255,255,.04); color:var(--ink-dim);
+    transition:background .15s, border-color .15s, color .15s;
+  }
+  .target-chip:hover{background:rgba(233,166,166,.1); border-color:rgba(233,166,166,.4); color:var(--ink)}
+  .target-chip.active{background:rgba(233,166,166,.18); border-color:var(--rose); color:var(--rose); font-weight:600}
+  .target-chip:disabled{opacity:.4; cursor:wait}
+  /* ---- action chips ---- */
+  #actions{display:flex; gap:7px; margin-bottom:10px; flex-wrap:wrap}
+  .action-chip{
+    font-family:var(--body); cursor:pointer; border:1px solid rgba(255,255,255,.18);
+    border-radius:20px; padding:6px 13px; font-size:.88rem; font-weight:500;
+    background:rgba(255,255,255,.05); color:var(--ink-dim);
+    transition:background .15s, border-color .15s, color .15s; white-space:nowrap;
+  }
+  .action-chip:hover{background:rgba(232,193,114,.1); border-color:rgba(232,193,114,.4); color:var(--ink)}
+  .action-chip.active{
+    background:rgba(232,193,114,.18); border-color:var(--gold);
+    color:var(--gold); font-weight:600;
+  }
+  .action-chip:disabled{opacity:.4; cursor:wait}
   /* ---- input row ---- */
-  #inputrow{display:flex; gap:10px; margin-top:12px}
+  #inputrow{display:flex; gap:10px; margin-top:4px}
   #say{flex:1; background:rgba(255,255,255,.06); border:1px solid rgba(255,255,255,.15);
     color:var(--ink); border-radius:10px; padding:11px 14px; font-family:var(--body); font-size:1rem}
   #say:focus{outline:none; border-color:var(--gold)}
   button{font-family:var(--body); cursor:pointer; border:none; border-radius:10px;
     padding:11px 16px; color:#10131a; background:var(--gold); font-weight:600; font-size:.98rem}
+  #send{background:var(--moss); color:#0d1a0e; font-size:1.2rem; padding:11px 18px; min-width:46px}
   button.ghost{background:transparent; color:var(--ink); border:1px solid rgba(255,255,255,.22)}
   button:disabled{opacity:.45; cursor:wait}
   @keyframes pulse-rec{0%,100%{opacity:1}50%{opacity:.6}}
@@ -174,6 +207,90 @@
   #enter{width:100%; margin-top:16px; padding:14px; font-size:1.05rem}
   .shimmer{color:var(--ink-dim); font-style:italic}
   .hidden{display:none !important}
+  /* ---- journal button (top-right) ---- */
+  #journal-btn{
+    position:fixed; top:16px; right:16px; z-index:8;
+    display:flex; align-items:center; gap:7px;
+    background:rgba(10,13,24,.82); border:1px solid rgba(232,193,114,.30);
+    border-radius:10px; padding:9px 15px;
+    color:var(--gold); font-family:var(--display); font-size:1rem;
+    backdrop-filter:blur(6px); cursor:pointer;
+    transition:background .15s, border-color .15s, transform .15s;
+    position:fixed; /* re-state for specificity */
+  }
+  #journal-btn:hover{
+    background:rgba(232,193,114,.12); border-color:rgba(232,193,114,.55);
+    transform:translateY(-1px);
+  }
+  #journal-btn .jb-icon{font-size:1.15rem; line-height:1}
+  .je-badge{
+    display:none; position:absolute; top:-5px; right:-5px;
+    width:10px; height:10px; background:var(--rose); border-radius:50%;
+    border:2px solid var(--dusk-0);
+  }
+  #journal-btn.has-new .je-badge{display:block}
+  /* ---- journal side panel ---- */
+  #journal-overlay{
+    position:fixed; top:0; right:0; bottom:0;
+    width:min(400px,92vw); z-index:25;
+    display:flex; flex-direction:column;
+    background:rgba(8,10,18,.97);
+    border-left:1px solid rgba(232,193,114,.18);
+    backdrop-filter:blur(14px);
+    transform:translateX(105%);
+    transition:transform .35s cubic-bezier(.4,0,.2,1);
+    box-shadow:-14px 0 48px rgba(0,0,0,.55);
+  }
+  #journal-overlay.open{transform:translateX(0)}
+  #journal-header{
+    display:flex; align-items:center; justify-content:space-between;
+    padding:18px 20px 14px;
+    border-bottom:1px solid rgba(232,193,114,.13);
+    flex-shrink:0;
+  }
+  #journal-header h3{
+    font-family:var(--display); font-size:1.35rem;
+    color:var(--gold); margin:0; letter-spacing:.3px;
+  }
+  #journal-close{
+    background:transparent; border:none; color:var(--ink-dim);
+    font-size:1.2rem; cursor:pointer; padding:4px 9px;
+    border-radius:6px; transition:color .15s, background .15s;
+  }
+  #journal-close:hover{color:var(--ink); background:rgba(255,255,255,.07)}
+  #journal-list{
+    flex:1; overflow-y:auto; padding:16px 20px 28px;
+    scroll-behavior:smooth;
+  }
+  #journal-list::-webkit-scrollbar{width:4px}
+  #journal-list::-webkit-scrollbar-track{background:transparent}
+  #journal-list::-webkit-scrollbar-thumb{background:rgba(232,193,114,.22); border-radius:2px}
+  .journal-entry{margin-bottom:4px}
+  .je-meta{
+    font-size:.72rem; color:var(--ink-dim); letter-spacing:.6px;
+    text-transform:uppercase; margin-bottom:9px;
+  }
+  .je-meta em{font-style:normal; color:rgba(232,193,114,.55)}
+  .je-player{
+    font-size:.9rem; color:rgba(200,190,175,.6);
+    font-style:italic; margin-bottom:10px;
+    padding:6px 10px; border-left:2px solid rgba(255,255,255,.1);
+    border-radius:0 4px 4px 0;
+  }
+  .je-you{font-style:normal; color:rgba(210,200,185,.9); font-weight:500; margin-right:5px}
+  .je-speaker{
+    font-family:var(--display); font-size:1.1rem;
+    color:var(--gold); margin-bottom:3px;
+  }
+  .je-dialogue{font-size:.95rem; line-height:1.65; color:var(--ink)}
+  .je-sep{
+    text-align:center; color:rgba(232,193,114,.22);
+    font-size:.68rem; letter-spacing:7px; margin:18px 0;
+  }
+  .je-empty{
+    color:var(--ink-dim); font-style:italic; font-family:var(--display);
+    text-align:center; margin-top:50px; font-size:1.05rem; line-height:1.7;
+  }
 </style>
 </head>
 <body>
@@ -185,15 +302,42 @@
     <div id="dialogue">
       <h2 id="speaker">…</h2>
       <div id="line"><span class="shimmer">the story is awakening…</span></div>
+      <div id="target-row" class="hidden"></div>
+      <div id="actions">
+        <button class="action-chip active" data-action="talk">💬 Talk</button>
+        <button class="action-chip" data-action="touch">🤝 Touch</button>
+        <button class="action-chip" data-action="silent">😶 Silence</button>
+        <button class="action-chip" data-action="give">🎁 Give</button>
+        <button class="action-chip" data-action="move">🚪 Move</button>
+        <button class="action-chip" data-action="scout">👁️ Look around</button>
+      </div>
       <div id="inputrow">
         <input id="say" placeholder="say something…" autocomplete="off" />
-        <button id="send">Speak</button>
+        <button id="send">✓</button>
         <button id="mic" class="ghost" title="hold to speak">🎙️</button>
         <button id="tts-toggle" title="toggle voice output">🔇</button>
       </div>
     </div>
   </div>
 
+  <!-- Journal button — hidden until game starts -->
+  <button id="journal-btn" class="hidden" title="Story journal">
+    <span class="jb-icon">📖</span>
+    <span>Journal</span>
+    <span class="je-badge"></span>
+  </button>
+
+  <!-- Journal side panel -->
+  <div id="journal-overlay">
+    <div id="journal-header">
+      <h3>📖 Story Journal</h3>
+      <button id="journal-close" title="Close">✕</button>
+    </div>
+    <div id="journal-list">
+      <div class="je-empty">The story has not yet begun…</div>
+    </div>
+  </div>
+
   <!-- Book intro overlay — shown after /start if intro_text is non-empty -->
   <div id="intro-book" class="hidden">
     <div class="book-wrap">
@@ -235,6 +379,55 @@ let client;
   } catch (e) { console.warn("themes endpoint not ready", e); }
 })();
 
+// --- journal ---
+let history = [];
+let journalTurn = 0;
+
+const ACTION_ICONS = { talk:"💬", flirt:"🌸", touch:"🤝", ask:"🤔", silent:"😶", give:"🎁", move:"🚪" };
+
+function addToHistory(playerText, v, action = "talk") {
+  if (!v?.dialogue || !v?.speaker) return;
+  history.push({
+    turn: ++journalTurn,
+    place: v.place || "",
+    player: playerText || "",
+    action: action,
+    speaker: v.speaker,
+    dialogue: v.dialogue,
+  });
+  if ($("journal-overlay").classList.contains("open")) {
+    renderJournal();
+  } else {
+    $("journal-btn").classList.add("has-new");
+  }
+}
+
+function renderJournal() {
+  const list = $("journal-list");
+  if (history.length === 0) {
+    list.innerHTML = '<div class="je-empty">The story has not yet begun…</div>';
+    return;
+  }
+  list.innerHTML = history.map((e, i) => `
+    ${i > 0 ? '<div class="je-sep">· · ·</div>' : ""}
+    <div class="journal-entry">
+      <div class="je-meta">Turn ${e.turn}${e.place ? ` &nbsp;·&nbsp; <em>${e.place}</em>` : ""}</div>
+      ${(e.player || e.action !== "talk") ? `<div class="je-player"><span class="je-you">${ACTION_ICONS[e.action] || "💬"} You</span>${e.player ? ` "${e.player}"` : ""}</div>` : ""}
+      <div class="je-speaker">${e.speaker}</div>
+      <div class="je-dialogue">${e.dialogue}</div>
+    </div>
+  `).join("");
+  setTimeout(() => { list.scrollTop = list.scrollHeight; }, 30);
+}
+
+$("journal-btn").addEventListener("click", () => {
+  const open = $("journal-overlay").classList.toggle("open");
+  if (open) { renderJournal(); $("journal-btn").classList.remove("has-new"); }
+});
+$("journal-close").addEventListener("click", () => {
+  $("journal-overlay").classList.remove("open");
+});
+
 // --- typewriter ---
 function typeOut(el, text){
   el.textContent = ""; let i = 0;
@@ -267,10 +460,10 @@ function render(v){
     const img = document.createElement("img");
     img.className = "sprite"; img.src = p.sprite_url; img.alt = p.name;
 
-    // relation bar: 0..100 maps to 0..100%; colour shifts gold→green as affection rises
+    // relation bar: positive → green left→right | negative → red right→left
     const rel = p.relationship ?? 0;
-    const relPct = Math.min(100, Math.max(0, rel)).toFixed(1);
-    const relColor = `hsl(${40 + rel * 0.8}, 65%, 52%)`; // gold at 0, warm-green at 100
+    const relPosW = rel > 0 ? rel.toFixed(1) : "0";
+    const relNegW = rel < 0 ? Math.abs(rel).toFixed(1) : "0";
     const relLabel = rel > 0 ? `+${rel}` : `${rel}`;
 
     // traits rows
@@ -293,7 +486,10 @@ function render(v){
     card.innerHTML = `
       <div class="char-card-name">${p.name}</div>
       <div class="rel-row">
-        <div class="rel-bar"><div class="rel-fill" style="width:${relPct}%;background:${relColor}"></div></div>
+        <div class="rel-bar">
+          <div class="rel-fill-neg" style="width:${relNegW}%"></div>
+          <div class="rel-fill-pos" style="width:${relPosW}%"></div>
+        </div>
         <span class="rel-val">${relLabel}</span>
       </div>
       <div class="traits-section">
@@ -311,6 +507,9 @@ function render(v){
 
   // notifications
   for (const msg of (v.notifications || [])) showNotif(msg);
+
+  // target chips — show only when multiple characters are on stage
+  updateTargetRow(v.present);
 }
 
 function showNotif(msg){
@@ -324,7 +523,84 @@ function showNotif(msg){
   }, 5000);
 }
 
-function busy(state){ $("send").disabled = state; $("say").disabled = state; $("mic").disabled = state; }
+// --- target selector ---
+let currentTarget = ""; // "" = anyone
+
+function updateTargetRow(present) {
+  const row = $("target-row");
+  if (!present || present.length <= 1) {
+    row.classList.add("hidden");
+    currentTarget = "";
+    return;
+  }
+  row.classList.remove("hidden");
+  row.innerHTML = `
+    <span class="target-label">To:</span>
+    <button class="target-chip active" data-name="">Everyone</button>
+    ${present.map(p => `<button class="target-chip" data-name="${p.name}">${p.name}</button>`).join("")}
+  `;
+  row.querySelectorAll(".target-chip").forEach(btn => {
+    btn.addEventListener("click", () => {
+      if (btn.disabled) return;
+      row.querySelectorAll(".target-chip").forEach(b => b.classList.remove("active"));
+      btn.classList.add("active");
+      currentTarget = btn.dataset.name;
+    });
+  });
+}
+
+function resetTarget() {
+  currentTarget = "";
+  const row = $("target-row");
+  row.querySelectorAll(".target-chip").forEach(b => b.classList.remove("active"));
+  const everyone = row.querySelector('.target-chip[data-name=""]');
+  if (everyone) everyone.classList.add("active");
+}
+
+// --- action chips ---
+
+const ACTION_PLACEHOLDERS = {
+  talk:   "say something…",
+  flirt:  "add a flirty line… (or leave blank)",
+  touch:  "add words… (optional)",
+  ask:    "what do you want to know about them?",
+  silent: "",
+  give:   "what do you offer?",
+  move:   "where? (leave blank and the story decides)",
+  scout:  "name someone specific… (or leave blank)",
+};
+let currentAction = "talk";
+
+document.querySelectorAll(".action-chip").forEach(btn => {
+  btn.addEventListener("click", () => {
+    if (btn.disabled) return;
+    document.querySelectorAll(".action-chip").forEach(b => b.classList.remove("active"));
+    btn.classList.add("active");
+    currentAction = btn.dataset.action;
+    if (currentAction === "silent") {
+      $("say").value = "";
+      $("say").disabled = true;
+    } else {
+      $("say").disabled = false;
+      $("say").placeholder = ACTION_PLACEHOLDERS[currentAction] || "say something…";
+      $("say").focus();
+    }
+  });
+});
+
+function resetAction() {
+  currentAction = "talk";
+  document.querySelectorAll(".action-chip").forEach(b => b.classList.remove("active"));
+  document.querySelector('.action-chip[data-action="talk"]').classList.add("active");
+  $("say").disabled = false;
+  $("say").placeholder = ACTION_PLACEHOLDERS.talk;
+}
+
+function busy(state){
+  $("send").disabled = state; $("say").disabled = state; $("mic").disabled = state;
+  document.querySelectorAll(".action-chip").forEach(b => b.disabled = state);
+  document.querySelectorAll(".target-chip").forEach(b => b.disabled = state);
+}
 
 // --- TTS toggle ---
 let ttsEnabled = true;
@@ -362,6 +638,8 @@ $("enter").addEventListener("click", async () => {
   const { data } = await client.predict("/start", { theme: $("theme").value, tone: $("tone").value });
   $("setup").classList.add("hidden");
   render(data[0]);
+  addToHistory("", data[0]);
+  $("journal-btn").classList.remove("hidden");
   // Show the book intro overlay if the engine sent an intro text
   if (data[0]?.intro_text) {
     $("book-text").textContent = data[0].intro_text;
@@ -372,12 +650,20 @@ $("enter").addEventListener("click", async () => {
 // --- a turn ---
 async function sendTurn(){
   const text = $("say").value.trim();
-  if (!text) return;
+  // allow empty text for non-talk actions (the action hint is enough)
+  if (!text && currentAction === "talk") return;
+  const action = currentAction;
   $("say").value = ""; busy(true);
   try {
-    const { data } = await client.predict("/turn", { player_input: text });
+    const { data } = await client.predict("/turn", { player_input: text, action, target: currentTarget });
     render(data[0]);
-  } finally { busy(false); $("say").focus(); }
+    addToHistory(text, data[0], action);
+  } finally {
+    busy(false);
+    resetAction();
+    resetTarget();
+    $("say").focus();
+  }
 }
 $("send").addEventListener("click", sendTurn);
 $("say").addEventListener("keydown", (e) => { if (e.key === "Enter") sendTurn(); });

modal_app.py

@@ -41,9 +41,12 @@ painter_image = (
         "diffusers==0.35.1",
         "transformers==4.47.1",
         "accelerate==1.8.1",
+        "peft",        # required by diffusers load_lora_weights / fuse_lora
         "safetensors",
         "Pillow",
         "huggingface_hub",
+        "rembg",       # background removal for sprites
+        "onnxruntime", # rembg runtime dep
     )
 )
 
@@ -65,7 +68,11 @@ app = modal.App("vn-app")
 
 GGUF_REPO = "Qwen/Qwen3-14B-GGUF"
 GGUF_FILE = "Qwen3-14B-Q8_0.gguf"
-SDXL_MODEL = "stabilityai/sdxl-turbo"
+# Painter: SDXL-base-1.0 + ByteDance Lightning LoRA (matches local SdxlLightningPainter)
+SDXL_BASE_REPO  = "stabilityai/stable-diffusion-xl-base-1.0"
+SDXL_BASE_LOCAL = "sdxl-base"
+LIGHTNING_LORA_REPO = "ByteDance/SDXL-Lightning"
+LIGHTNING_LORA_FILE = "sdxl_lightning_4step_lora.safetensors"
 
 
 # --------------------------------------------------------------------------- #
@@ -77,26 +84,31 @@ SDXL_MODEL = "stabilityai/sdxl-turbo"
 def download_model() -> None:
     from huggingface_hub import hf_hub_download
 
-    print(f"Downloading {GGUF_REPO}/{GGUF_FILE} …")
+    print(f"Downloading {GGUF_REPO}/{GGUF_FILE} ...")
     path = hf_hub_download(repo_id=GGUF_REPO, filename=GGUF_FILE, local_dir=LLM_DIR)
     llm_volume.commit()
-    print(f"Saved → {path}")
+    print(f"Saved -> {path}")
 
 
 @app.function(
     image=painter_image, volumes={IMAGE_DIR: image_volume}, timeout=900, secrets=[hf_secret]
 )
 def download_image_model() -> None:
+    """Download SDXL-base-1.0 weights (~6.5 GB) to the volume.
+
+    The Lightning LoRA (ByteDance/SDXL-Lightning) is small (~400 MB) and loaded
+    at inference time directly from HF Hub — no need to pre-download it.
+    """
     from huggingface_hub import snapshot_download
 
-    print(f"Downloading {SDXL_MODEL} …")
-    snapshot_download(SDXL_MODEL, local_dir=f"{IMAGE_DIR}/sdxl-turbo")
+    print(f"Downloading {SDXL_BASE_REPO} -> {IMAGE_DIR}/{SDXL_BASE_LOCAL} ...")
+    snapshot_download(SDXL_BASE_REPO, local_dir=f"{IMAGE_DIR}/{SDXL_BASE_LOCAL}")
     image_volume.commit()
-    print("SDXL-Turbo saved to volume.")
+    print("SDXL-base saved to volume.")
 
 
 # --------------------------------------------------------------------------- #
-#  LLM backend  (llama.cpp + Qwen3-4B on A10G)
+#  LLM backend  (llama.cpp + Qwen3-14B on A10G)
 # --------------------------------------------------------------------------- #
 
 
@@ -143,7 +155,7 @@ class ModalLLMBackend:
 
 
 # --------------------------------------------------------------------------- #
-#  Painter backend  (SDXL-Turbo on A10G)
+#  Painter backend  (SDXL-base-1.0 + Lightning LoRA on A10G)
 # --------------------------------------------------------------------------- #
 
 
@@ -152,20 +164,46 @@ class ModalLLMBackend:
     gpu="A10G",
     volumes={IMAGE_DIR: image_volume},
     timeout=120,
-    scaledown_window=60,
+    scaledown_window=600,  # keep warm 10 min — avoids cold-start during a play session
 )
 class ModalPainterBackend:
+    """SDXL-base-1.0 + ByteDance SDXL-Lightning 4-step LoRA.
+
+    Mirrors the local SdxlLightningPainter exactly:
+      - EulerDiscreteScheduler with trailing timestep spacing
+      - LoRA fused into weights (fuse_lora) for faster inference
+      - guidance_scale=0.0 (required by Lightning distillation)
+      - 4 inference steps
+      - Optional rembg background removal for sprites
+    """
+
     @modal.enter()
     def load(self) -> None:
         import torch
-        from diffusers import StableDiffusionXLPipeline
+        from diffusers import EulerDiscreteScheduler, StableDiffusionXLPipeline
 
+        print(f"[modal] Loading SDXL-base from {IMAGE_DIR}/{SDXL_BASE_LOCAL} ...")
         self.pipe = StableDiffusionXLPipeline.from_pretrained(
-            f"{IMAGE_DIR}/sdxl-turbo",
+            f"{IMAGE_DIR}/{SDXL_BASE_LOCAL}",
             torch_dtype=torch.float16,
             variant="fp16",
         ).to("cuda")
-        print("[modal] SDXL-Turbo loaded on GPU")
+
+        # Lightning requires EulerDiscrete with trailing timestep spacing
+        self.pipe.scheduler = EulerDiscreteScheduler.from_config(
+            self.pipe.scheduler.config, timestep_spacing="trailing"
+        )
+
+        # Load Lightning LoRA from HF Hub (~400 MB, fast)
+        print(f"[modal] Loading Lightning LoRA from {LIGHTNING_LORA_REPO} ...")
+        self.pipe.load_lora_weights(LIGHTNING_LORA_REPO, weight_name=LIGHTNING_LORA_FILE)
+        self.pipe.fuse_lora()  # bake into weights for faster inference
+
+        # Fix fp16 VAE instability ("weird pixels" artifact)
+        self.pipe.upcast_vae()
+
+        self.torch = torch
+        print("[modal] SDXL-Lightning ready on GPU")
 
     @modal.method()
     def render(
@@ -175,23 +213,40 @@ class ModalPainterBackend:
         seed: int,
         size: int,
         steps: int,
+        guidance_scale: float = 0.0,
+        remove_bg: bool = False,
     ) -> bytes:
+        """Generate an image and return PNG bytes.
+
+        Args:
+            prompt: Positive prompt.
+            negative_prompt: Negative prompt.
+            seed: RNG seed for reproducibility.
+            size: Square image side in pixels.
+            steps: Inference steps (4 for Lightning).
+            guidance_scale: CFG scale — 0.0 for Lightning sprites, >1.0 for backdrops.
+            remove_bg: Run rembg on the output (sprites only).
+        """
         import io
 
-        import torch
-
-        gen = torch.Generator(device="cuda").manual_seed(seed)
+        gen = self.torch.Generator(device="cuda").manual_seed(seed)
         result = self.pipe(
             prompt=prompt,
-            negative_prompt=negative_prompt,
+            negative_prompt=negative_prompt or None,
             num_inference_steps=steps,
-            guidance_scale=0.0,
+            guidance_scale=guidance_scale,
             height=size,
             width=size,
             generator=gen,
         )
+        img = result.images[0]
+
+        if remove_bg:
+            from rembg import remove  # noqa: PLC0415
+            img = remove(img)  # returns RGBA PIL image
+
         buf = io.BytesIO()
-        result.images[0].save(buf, format="PNG")
+        img.save(buf, format="PNG")  # PNG supports RGBA transparency
         return buf.getvalue()
 
 
@@ -216,13 +271,31 @@ def smoke_painter() -> None:
     from PIL import Image
 
     backend = ModalPainterBackend()
+
+    # Backdrop test
     png = backend.render.remote(
-        prompt="Japanese anime forest, glowing mushrooms",
-        negative_prompt="text, watermark",
+        prompt="Japanese anime forest, glowing mushrooms, painterly background",
+        negative_prompt="text, watermark, characters, person",
         seed=42,
         size=512,
-        steps=2,
+        steps=4,
+        guidance_scale=0.0,
+        remove_bg=False,
     )
     img = Image.open(io.BytesIO(png))
-    img.save("smoke_painter.png")
-    print(f"Painter smoke OK → smoke_painter.png  {img.size}")
+    img.save("smoke_backdrop.png")
+    print(f"Backdrop smoke OK -> smoke_backdrop.png  {img.size}")
+
+    # Sprite test (with rembg background removal)
+    png_sprite = backend.render.remote(
+        prompt="anime girl, school uniform, happy expression, white background",
+        negative_prompt="text, watermark, scenery, complex background",
+        seed=7,
+        size=512,
+        steps=4,
+        guidance_scale=0.0,
+        remove_bg=True,
+    )
+    img_sprite = Image.open(io.BytesIO(png_sprite))
+    img_sprite.save("smoke_sprite.png")
+    print(f"Sprite smoke OK  -> smoke_sprite.png   {img_sprite.size}  mode={img_sprite.mode}")

uv.lock

@@ -2684,16 +2684,17 @@ wheels = [
 
 [[package]]
 name = "protobuf"
-version = "4.25.9"
+version = "6.33.6"
 source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/d9/8e/d08c41a8c004e1d437ef467e7c4f9c3295cd784eba48ed5d1d01f94b1dad/protobuf-4.25.9.tar.gz", hash = "sha256:b0dc7e7c68de8b1ce831dacb12fb407e838edbb8b6cc0dc3a2a6b4cbf6de9cff", size = 381040, upload-time = "2026-03-25T23:09:36.423Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/66/70/e908e9c5e52ef7c3a6c7902c9dfbb34c7e29c25d2f81ade3856445fd5c94/protobuf-6.33.6.tar.gz", hash = "sha256:a6768d25248312c297558af96a9f9c929e8c4cee0659cb07e780731095f38135", size = 444531, upload-time = "2026-03-18T19:05:00.988Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/a8/e9/59435bd04bdd46cb38c42a336b22f9843e8e586ff83c35a5423f8b14704e/protobuf-4.25.9-cp310-abi3-win32.whl", hash = "sha256:bde396f568b0b46fc8fbfe9f02facf25b6755b2578a3b8ac61e74b9d69499e03", size = 392879, upload-time = "2026-03-25T23:09:21.32Z" },
-    { url = "https://files.pythonhosted.org/packages/f3/16/42a5c7f1001783d2b5bfcecde10127f09010f78982c86ae409122ce3ece6/protobuf-4.25.9-cp310-abi3-win_amd64.whl", hash = "sha256:3683c05154252206f7cb2d371626514b3708199d9bcf683b503dabf3a2e38e06", size = 413900, upload-time = "2026-03-25T23:09:23.589Z" },
-    { url = "https://files.pythonhosted.org/packages/56/5b/0074a0a9eb01f3d1c4648ca5e81b22090c811b210b61df9018ac6d6c5cda/protobuf-4.25.9-cp37-abi3-macosx_10_9_universal2.whl", hash = "sha256:9560813560e6ee72c11ca8873878bdb7ee003c96a57ebb013245fe84e2540904", size = 394826, upload-time = "2026-03-25T23:09:25.194Z" },
-    { url = "https://files.pythonhosted.org/packages/54/aa/b2dba856f64c36b2a06c67be1472de98cca07a2322d0f0cbf03279a40e5b/protobuf-4.25.9-cp37-abi3-manylinux2014_aarch64.whl", hash = "sha256:999146ef02e7fa6a692477badd1528bcd7268df211852a3df2d834ba2b480791", size = 294191, upload-time = "2026-03-25T23:09:26.613Z" },
-    { url = "https://files.pythonhosted.org/packages/a8/5c/53f18822017b8bda6bd8bb4e02048e911fdc79a3dafdc83ab994fe922a84/protobuf-4.25.9-cp37-abi3-manylinux2014_x86_64.whl", hash = "sha256:438c636de8fb706a0de94a12a268ef1ae8f5ba5ae655a7671fcda5968ba3c9be", size = 295178, upload-time = "2026-03-25T23:09:27.839Z" },
-    { url = "https://files.pythonhosted.org/packages/16/28/d5065b212685875d3924bcdb3201cbf467cb4d58a18aa19a8dfd99ea80a9/protobuf-4.25.9-py3-none-any.whl", hash = "sha256:d49b615e7c935194ac161f0965699ac84df6112c378e05ec53da65d2e4cbb6d4", size = 156822, upload-time = "2026-03-25T23:09:34.957Z" },
+    { url = "https://files.pythonhosted.org/packages/fc/9f/2f509339e89cfa6f6a4c4ff50438db9ca488dec341f7e454adad60150b00/protobuf-6.33.6-cp310-abi3-win32.whl", hash = "sha256:7d29d9b65f8afef196f8334e80d6bc1d5d4adedb449971fefd3723824e6e77d3", size = 425739, upload-time = "2026-03-18T19:04:48.373Z" },
+    { url = "https://files.pythonhosted.org/packages/76/5d/683efcd4798e0030c1bab27374fd13a89f7c2515fb1f3123efdfaa5eab57/protobuf-6.33.6-cp310-abi3-win_amd64.whl", hash = "sha256:0cd27b587afca21b7cfa59a74dcbd48a50f0a6400cfb59391340ad729d91d326", size = 437089, upload-time = "2026-03-18T19:04:50.381Z" },
+    { url = "https://files.pythonhosted.org/packages/5c/01/a3c3ed5cd186f39e7880f8303cc51385a198a81469d53d0fdecf1f64d929/protobuf-6.33.6-cp39-abi3-macosx_10_9_universal2.whl", hash = "sha256:9720e6961b251bde64edfdab7d500725a2af5280f3f4c87e57c0208376aa8c3a", size = 427737, upload-time = "2026-03-18T19:04:51.866Z" },
+    { url = "https://files.pythonhosted.org/packages/ee/90/b3c01fdec7d2f627b3a6884243ba328c1217ed2d978def5c12dc50d328a3/protobuf-6.33.6-cp39-abi3-manylinux2014_aarch64.whl", hash = "sha256:e2afbae9b8e1825e3529f88d514754e094278bb95eadc0e199751cdd9a2e82a2", size = 324610, upload-time = "2026-03-18T19:04:53.096Z" },
+    { url = "https://files.pythonhosted.org/packages/9b/ca/25afc144934014700c52e05103c2421997482d561f3101ff352e1292fb81/protobuf-6.33.6-cp39-abi3-manylinux2014_s390x.whl", hash = "sha256:c96c37eec15086b79762ed265d59ab204dabc53056e3443e702d2681f4b39ce3", size = 339381, upload-time = "2026-03-18T19:04:54.616Z" },
+    { url = "https://files.pythonhosted.org/packages/16/92/d1e32e3e0d894fe00b15ce28ad4944ab692713f2e7f0a99787405e43533a/protobuf-6.33.6-cp39-abi3-manylinux2014_x86_64.whl", hash = "sha256:e9db7e292e0ab79dd108d7f1a94fe31601ce1ee3f7b79e0692043423020b0593", size = 323436, upload-time = "2026-03-18T19:04:55.768Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/72/02445137af02769918a93807b2b7890047c32bfb9f90371cbc12688819eb/protobuf-6.33.6-py3-none-any.whl", hash = "sha256:77179e006c476e69bf8e8ce866640091ec42e1beb80b213c3900006ecfba6901", size = 170656, upload-time = "2026-03-18T19:04:59.826Z" },
 ]
 
 [[package]]
@@ -3819,7 +3820,7 @@ wheels = [
 
 [[package]]
 name = "transformers"
-version = "5.10.2"
+version = "5.9.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "huggingface-hub" },
@@ -3832,9 +3833,9 @@ dependencies = [
     { name = "tqdm" },
     { name = "typer" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/8d/38/d5f978bd5091019e89aef29b9a831f5cd70f2598963a3ead8b9570cab592/transformers-5.10.2.tar.gz", hash = "sha256:f9a44b9c8ca9ab1156b467f574d832ea066284299c2fd0ed84641ccb592751fc", size = 8799687, upload-time = "2026-06-04T18:43:49.119Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/51/58/7f843608f2e8421f86bb97060b54649be6239ec612b82bf9d41e65c26c00/transformers-5.9.0.tar.gz", hash = "sha256:25997cb8fa6053533171634b6162d7df54346530ec2aa9b42bb834e63668c842", size = 8642240, upload-time = "2026-05-20T14:50:49.278Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/73/6f/e1564b0cc182afa05e219a8e09a8e770ffaab879b6b824b56c819bd221da/transformers-5.10.2-py3-none-any.whl", hash = "sha256:8a669db546f82c7c3618cb46ceb0f0afd89292bc70f319c058f8332ec63e268d", size = 11003830, upload-time = "2026-06-04T18:43:45.303Z" },
+    { url = "https://files.pythonhosted.org/packages/02/ca/2eaa5359f2ccb8c2e1656bc26305ad0cf438aa392ce4b29ae67a315c186e/transformers-5.9.0-py3-none-any.whl", hash = "sha256:1d19509bcff7028ebc6b277d71caa712e8353778463d38764237d14b42b52788", size = 10787648, upload-time = "2026-05-20T14:50:45.337Z" },
 ]
 
 [[package]]

visualnovel/config.py

@@ -71,8 +71,6 @@ LLM_GGUF_FILE = os.getenv("VN_LLM_GGUF_FILE", "Qwen3-14B-Q4_K_M.gguf")
 IMAGE_MODEL   = os.getenv("VN_IMAGE_MODEL", "stabilityai/sdxl-turbo")
 # Path or HF repo of YOUR fine-tuned anime-style LoRA (the Well-Tuned bonus). Empty = none.
 IMAGE_LORA    = os.getenv("VN_IMAGE_LORA", "")
-# sdxl → SDXL-Turbo (1–2 steps)  |  lightning → SDXL-Lightning (4 steps, better quality)
-IMAGE_BACKEND = os.getenv("VN_IMAGE_BACKEND", "sdxl")
 
 WHISPER_SIZE = os.getenv(
     "VN_WHISPER_SIZE", "large-v3-turbo"

visualnovel/engine.py

@@ -56,7 +56,7 @@ class Engine:
     def transcribe(self, audio_path: str) -> str:
         return self.stt.transcribe(audio_path)
 
-    def play_turn(self, player_input: str, audio_path: str | None = None) -> ViewState:
+    def play_turn(self, player_input: str, action: str = "talk", target: str = "", audio_path: str | None = None) -> ViewState:
         assert self.state is not None, "call start() first"
         # turn_num matches what the tracer logs (post-increment value)
         turn_num = self.state.turn_index + 1
@@ -67,7 +67,7 @@ class Engine:
                     player_input = self.stt.transcribe(audio_path) or player_input
 
             with collector.measure("llm_direct", turn=turn_num):
-                out = orchestrator.direct_turn(self.llm, self.state, player_input)
+                out = orchestrator.direct_turn(self.llm, self.state, player_input, action=action, target=target)
 
             with collector.measure("apply_directives", turn=turn_num):
                 effects = state.apply_directives(self.state, out)

visualnovel/memory.py

@@ -23,6 +23,17 @@ def assemble_context(state: GameState, player_input: str) -> str:
         f"- description: {state.scene.description}\n- mood: {state.scene.mood}"
     )
     parts.append(f"SPIRITS ON STAGE:\n{characters.present_sheets_block(state)}")
+
+    # Characters the player has already met but who are currently off-stage.
+    # The LLM can bring them back via new_character (same id) — data is preserved.
+    off_stage = [
+        f"  - {ch.name} (id: {ch.id}, rel: {ch.relationship:+d})"
+        for cid, ch in state.characters.items()
+        if cid not in state.scene.present
+    ]
+    if off_stage:
+        parts.append("KNOWN BUT OFF STAGE (can return — use their existing id):\n" + "\n".join(off_stage))
+
     if state.flags:
         parts.append("FACTS: " + ", ".join(f"{k}={v}" for k, v in state.flags.items()))
 

visualnovel/orchestrator.py

@@ -65,6 +65,7 @@ def init_world(llm: LLMBackend, setup: SetupForm) -> tuple[GameState, DirectorOu
         init = InitOutput.model_validate(raw)
 
     first = init.first_character
+    first.id = _slug(first.id) or _slug(first.name)  # sanitize LLM-supplied id (Windows-safe)
     scene = Scene(
         id=_slug(init.scene.place),
         place=init.scene.place,
@@ -104,11 +105,15 @@ def init_world(llm: LLMBackend, setup: SetupForm) -> tuple[GameState, DirectorOu
 # --------------------------------------------------------------------------- #
 #  direct_turn
 # --------------------------------------------------------------------------- #
-def direct_turn(llm: LLMBackend, state: GameState, player_input: str) -> DirectorOutput:
+def direct_turn(
+    llm: LLMBackend, state: GameState, player_input: str,
+    action: str = "talk", target: str = "",
+) -> DirectorOutput:
     if config.USE_MOCK:
         return _mock_turn(state, player_input)
 
-    context = memory.assemble_context(state, player_input)
+    formatted = prompts.format_player_turn(player_input, action, target)
+    context = memory.assemble_context(state, formatted)
     raw = llm.complete_json(
         messages=[
             {"role": "system", "content": prompts.DIRECTOR_SYSTEM},

visualnovel/painter.py

@@ -56,26 +56,45 @@ class Painter(Protocol):
 # --------------------------------------------------------------------------- #
 #  Prompt composition (shared)
 # --------------------------------------------------------------------------- #
+
+# CLIP hard limit is 77 tokens (~4 chars/token).  Keep variable parts short so the
+# fixed prefix + suffix stay within budget.  Rough budgets per prompt:
+#   backdrop : 15 (prefix) + 18 (desc) + 5 (mood) + 12 (style) + 2 (suffix) ≈ 52
+#   sprite   : 12 (prefix) + 10 (style) + 20 (appearance) + 5 (mood+suffix)  ≈ 47
+_DESC_WORDS  = 18   # scene description budget
+_STYLE_WORDS = 12   # style_guide budget in prompts
+_APP_WORDS   = 20   # character appearance budget
+
+
+def _w(text: str, n: int) -> str:
+    """Return at most *n* words from *text* (space-split, no tokenizer needed)."""
+    words = text.split()
+    return " ".join(words[:n]) if len(words) > n else text
+
+
 def _backdrop_style(style_guide: str) -> str:
-    """Strip character-triggering tokens from the style guide for backdrop renders."""
+    """Keep only the most impactful style tokens and strip character-specific ones."""
     skip = {"expressive eyes", "clean linework"}
     tokens = [t.strip() for t in style_guide.split(",") if t.strip() not in skip]
-    return ", ".join(tokens)
+    return _w(", ".join(tokens), _STYLE_WORDS)
 
 
 def backdrop_prompt(state: GameState) -> str:
     style = _backdrop_style(state.style_guide)
+    desc  = _w(state.scene.description, _DESC_WORDS)
     return (
         f"background art, empty scenic environment, wide establishing shot, no humans, "
-        f"uninhabited location, {state.scene.description}, {state.scene.mood} atmosphere, "
+        f"uninhabited location, {desc}, {state.scene.mood} atmosphere, "
         f"{style}, no text"
     )
 
 
 def sprite_prompt(state: GameState, ch: Character) -> str:
+    style = _w(state.style_guide, _STYLE_WORDS)
+    app   = _w(ch.appearance, _APP_WORDS)
     return (
         f"vtuber character design, solo, single character, full-body, centered, "
-        f"{state.style_guide}, {ch.appearance}, {ch.mood} expression, "
+        f"{style}, {app}, {ch.mood} expression, "
         f"pure white background, simple background, no scenery, no text"
     )
 
@@ -299,6 +318,38 @@ class ModalPainter(PainterBase):
 
         self._backend = modal.Cls.from_name("vn-app", "ModalPainterBackend")()
 
+    def sprite(self, state: GameState, ch: Character) -> Path:
+        """Override to request server-side background removal via rembg."""
+        import io  # noqa: PLC0415
+
+        from PIL import Image  # noqa: PLC0415
+
+        prompt = sprite_prompt(state, ch)
+        neg = _SPRITE_NEGATIVE
+        seed = ch.sprite_seed
+        kind, key = "sprite", f"{ch.id}.{ch.mood}"
+
+        # Include remove_bg in the cache key so backdrop/sprite hashes never collide.
+        h = hashlib.sha1(f"{kind}|{prompt}|{neg}|{seed}|rembg".encode()).hexdigest()[:12]
+        path = config.CACHE_DIR / f"{kind}_{key}_{h}.png"
+        if path.exists():
+            collector.record_cache(hit=True)
+            return path
+
+        collector.record_cache(hit=False)
+        png_bytes = self._backend.render.remote(
+            prompt=prompt,
+            negative_prompt=neg,
+            seed=seed,
+            size=config.IMAGE_SIZE,
+            steps=config.IMAGE_STEPS,
+            guidance_scale=0.0,
+            remove_bg=True,
+        )
+        img = Image.open(io.BytesIO(png_bytes))
+        img.save(path)
+        return path
+
     def _render(self, prompt: str, seed: int, size: int,
                 negative_prompt: str = "", guidance_scale: float = 0.0):
         import io  # noqa: PLC0415
@@ -312,6 +363,7 @@ class ModalPainter(PainterBase):
             size=size,
             steps=config.IMAGE_STEPS,
             guidance_scale=guidance_scale,
+            remove_bg=False,
         )
         return Image.open(io.BytesIO(png_bytes))
 

visualnovel/prompts.py

@@ -35,14 +35,23 @@ This turn:
 Rules:
   - Stay true to each character's established personality, speech quirks, and appearance.
     A character only reacts to what they could realistically know.
-  - Player moves to a new place -> set scene_change.
+  - Player moves to a new place -> ALWAYS emit scene_change. The current character stays
+    behind. Their goodbye is their last line — they never follow unless the player invites them.
   - Player meets someone new -> set new_character with a vivid one_line, a STABLE appearance
     (used for every illustration), a distinct voice, 2-3 personality traits, and a secret/goal.
     Maximum ~3 characters on stage at once.
-  - Romantic tension and friction build via relationship_delta for the speaking character.
-  - When affection is high enough, allow confessions, hand-holding moments, blushing reactions.
+  - Relationship_delta ranges from -100 to +100. Romantic warmth builds it up; coldness,
+    insults, or rejection bring it down. React authentically to negative moments — hurt, cold,
+    quietly withdrawing. Do NOT loop the same response — escalate or shift the dynamic.
+  - If a character explicitly says goodbye, storms off, or leaves for good -> emit
+    exit_character with their exact id to remove them from the stage.
+  - With multiple characters on stage, they may react to each other (banter, jealousy, a
+    knowing glance) — but never more than 2 NPC-to-NPC exchanges in a row. After that the
+    next line MUST address the player directly.
   - If a detail is missing, invent something charming and consistent — never contradict an
     established fact.
+  - NEVER repeat a line of dialogue verbatim from the recent exchanges. Every response must
+    move the scene forward — emotionally, narratively, or tonally.
   - Output MUST be a single JSON object matching the schema. Nothing else.
 """
 
@@ -76,6 +85,64 @@ exact wording. Neutral narrator tone, plain prose, no lists. Output ONLY the sum
 """
 
 
+# --------------------------------------------------------------------------- #
+#  Player action hints
+#  These are injected into the context so the Weaver knows what kind of action
+#  the player is taking, independently of the text they typed.
+# --------------------------------------------------------------------------- #
+_ACTION_HINTS: dict[str, str] = {
+    # move absorbs "leave" — destination optional, the LLM picks freely if blank
+    "move": (
+        "*The wanderer moves to a different location. "
+        "If a destination was named, go there; otherwise pick a fitting place freely. "
+        "EMIT scene_change. "
+        "The current character stays behind — they do NOT follow. "
+        "Their farewell is their last line; the wanderer is already walking away.*"
+    ),
+    "touch": (
+        "*The wanderer gently touches the character — "
+        "react emotionally (surprise, warmth, or shy flustering). "
+        "adjust relationship_delta (positive if welcomed, negative if unwanted).*"
+    ),
+    "give": (
+        "*The wanderer offers something to the character — "
+        "react warmly or curiously, adjust relationship_delta.*"
+    ),
+    "flirt": (
+        "*The wanderer flirts directly — bold, romantic, slightly daring. "
+        "The character reacts with flustering, excitement, or playful deflection. "
+        "Apply a significant positive relationship_delta.*"
+    ),
+    "ask": (
+        "*The wanderer asks a personal question about the character — "
+        "they reveal something about their personality, past, or hidden feelings. "
+        "This may unlock a trait; apply a small positive relationship_delta.*"
+    ),
+    "silent": (
+        "*The wanderer says nothing — just watches in silence. "
+        "The character fills the silence: nervous rambling, a shy glance, "
+        "or a quiet moment of shared understanding. "
+        "No relationship_delta unless the silence clearly resonates (then small positive).*"
+    ),
+    "scout": (
+        "*The wanderer looks around with curiosity, taking in the surroundings. "
+        "This is a natural moment for someone new to appear — a passerby, a student, "
+        "a stranger who catches the eye. EMIT new_character to introduce them. "
+        "If the player named a specific person (e.g. 'Emily'), use that name and invent "
+        "a fitting personality for them. The existing character may notice too and react.*"
+    ),
+}
+
+
+def format_player_turn(player_input: str, action: str = "talk", target: str = "") -> str:
+    """Combine target + action hint + player text into a single context line."""
+    hint = _ACTION_HINTS.get(action, "")
+    text = player_input.strip()
+    target_hint = f"*The wanderer addresses {target} specifically — only {target} should reply.*\n" if target else ""
+    parts = [p for p in [target_hint, hint, text] if p]
+    return "\n".join(parts) or "…"
+
+
 # --------------------------------------------------------------------------- #
 #  Helpers
 # --------------------------------------------------------------------------- #

visualnovel/state.py

@@ -52,7 +52,7 @@ def apply_directives(state: GameState, out: DirectorOutput) -> list[str]:
     # new character arrives
     if d.new_character:
         nc = d.new_character
-        cid = nc.id or _slug(nc.name)
+        cid = _slug(nc.id) or _slug(nc.name)  # always sanitize — LLM ids can contain invalid chars
         if cid not in state.characters:
             state.characters[cid] = Character(
                 id=cid,