fix(zerogpu): persist phase state, because zerogpu run each function in

dedicated worker, so we need persistant data so workers can grab data
from worker to worker.

visualnovel/engine.py

@@ -19,6 +19,35 @@ import json
 import time
 from pathlib import Path
 
+# ZeroGPU runs each @spaces.GPU call in a separate worker subprocess; in-memory
+# state set by start_text() is invisible to start_images().  We bridge the gap
+# by writing/reading a temp file between the two phases.
+_PHASE_FILE = Path("/tmp/vn_phase_state.json")
+
+
+def _persist_phase(state: GameState, out: DirectorOutput, intro: str) -> None:
+    _PHASE_FILE.write_text(
+        json.dumps({"state": state.model_dump(), "out": out.model_dump(), "intro": intro}),
+        encoding="utf-8",
+    )
+
+
+def _restore_phase() -> tuple[GameState, DirectorOutput, str] | None:
+    if not _PHASE_FILE.exists():
+        return None
+    try:
+        d = json.loads(_PHASE_FILE.read_text(encoding="utf-8"))
+        from .schemas import DirectorOutput, GameState  # noqa: PLC0415
+
+        return (
+            GameState.model_validate(d["state"]),
+            DirectorOutput.model_validate(d["out"]),
+            d.get("intro", ""),
+        )
+    except Exception:
+        return None
+
+
 from . import config, memory, orchestrator, state
 from .llm import LLMBackend, get_llm
 from .metrics import collector
@@ -97,9 +126,10 @@ class Engine:
         state.save_memory(self.state)
         self.tracer.log(event="start", setup=setup.model_dump(), opening=opening.model_dump())
         intro = self.state.flags.get("situation_intro", "")
-        # Stash for start_images()
+        # Stash for start_images() — also written to disk so ZeroGPU workers can pick it up
         self._pending_out: DirectorOutput | None = opening
         self._pending_intro: str = intro
+        _persist_phase(self.state, opening, intro)
         # Build a text-only ViewState so the frontend can show dialogue immediately
         s = self.state
         speaker_ch = s.characters.get(opening.speaker)
@@ -138,6 +168,11 @@ class Engine:
     def start_images(self) -> ViewState:
         """Phase 2 — paint backdrop + sprite, TTS. Call after start_text().
         Returns the full ViewState."""
+        if self.state is None or self._pending_out is None:
+            # Restore state persisted by start_text() (handles ZeroGPU worker isolation)
+            restored = _restore_phase()
+            if restored is not None:
+                self.state, self._pending_out, self._pending_intro = restored
         assert self.state is not None and self._pending_out is not None, "call start_text() first"
         return self._view(self._pending_out, intro_text=self._pending_intro)
 

visualnovel/llm.py

@@ -158,8 +158,7 @@ class TransformersLLM:
             + "\nNo markdown fences. No explanation. Only the JSON object."
         )
         augmented = [
-            {**m, "content": m["content"] + hint} if m["role"] == "system" else m
-            for m in messages
+            {**m, "content": m["content"] + hint} if m["role"] == "system" else m for m in messages
         ]
         text = self.complete(augmented, **kw)
         return json.loads(text[text.find("{") : text.rfind("}") + 1])