Deploy Myco from CI

game/engine.py

@@ -44,7 +44,7 @@ def _get_model_and_tokenizer():
             _model = AutoModelForCausalLM.from_pretrained(
                 model_id,
                 token=token,
-                dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
+                torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
                 device_map=DEVICE,
                 trust_remote_code=True
             )
@@ -111,6 +111,13 @@ def _set_reply(text: str, tag: str = "chat"):
         _LAST_REPLY["text"] = "\n\n".join(_REPLY_LOG)
 
 
+def get_myco_log() -> str:
+    """Return the full running Myco log (narrative/action/chat) — non-blocking,
+    intended to be polled by a gr.Timer rather than streamed from an infinite
+    generator (which can starve the Gradio queue for other events)."""
+    return _LAST_REPLY["text"]
+
+
 # Shared store for the most recent raw LLM output, so the front end can
 # display it live for debugging (server-side prints aren't visible once deployed).
 _LAST_DEBUG = {"text": "🍄 Waiting for Myco's first thought..."}
@@ -911,39 +918,3 @@ def _llm_action(*args, **kwargs):
     """Action-mode LLM call. Delegates to the JSON-aware single-turn _llm()."""
     return _llm(*args, **kwargs)
     
-
-def narrate(event, score):
-    prompt = f"""
-You are Myco, a mystical forest spirit inside a game.
-
-RULES:
-- Respond with EXACTLY ONE sentence.
-- Maximum 20 words.
-- No line breaks or lists.
-
-Event: {event}
-Score: {score}
-"""
-
-    inputs = _tokenizer(prompt, return_tensors="pt").to(_model.device)
-
-    output = _model.generate(
-        **inputs,
-        max_new_tokens=60,
-        temperature=0.9,
-        do_sample=True,
-        eos_token_id=_tokenizer.eos_token_id,
-        pad_token_id=_tokenizer.eos_token_id,
-    )
-
-    text = _tokenizer.decode(
-        output[0][inputs["input_ids"].shape[-1]:],
-        skip_special_tokens=True
-    ).strip()
-
-    # hard enforce: single sentence + 20 words max
-    text = re.split(r'[.!?]', text)[0].strip()
-    text = " ".join(text.split()[:20])
-
-    return text
-    

ui/gradio_app.py

@@ -5,7 +5,6 @@ import json
 import copy
 import random
 import threading
-import time
 import logging
 from typing import Any, cast
 import gradio as gr
@@ -33,7 +32,7 @@ from ui.renderers import (
 )
 from ui.interface import render_shroom_tab
 
-from game.engine import _LAST_REPLY
+from game.engine import get_myco_log
 
 logger = logging.getLogger(__name__)
 EMPTY_COLLECTION = None
@@ -64,10 +63,7 @@ def _autonomous_tick(current, collection, active_mushrooms, position):
     with _ROAM_LOCK:
         x, y = _ROAM["position"]
         safe_mushrooms = list(active_mushrooms or [])
-        coll = list(collection or [])
-        
-        # 1. Define the variable here, before any logic happens
-        initial_coll_len = len(coll) 
+        coll           = list(collection or [])
 
         # Check pickup at current position first
         if safe_mushrooms:
@@ -88,10 +84,6 @@ def _autonomous_tick(current, collection, active_mushrooms, position):
                 nx, ny, current, coll, safe_mushrooms
             )
 
-        # 2. Now you can safely use it here because it was defined above
-        sound_to_play = "./assets/sounds/mushroom_pickup.wav" if len(coll) > initial_coll_len else None
-        print(f"DEBUG: Attempting to play sound: {sound_to_play}")
-
     return (
         list(pos),
         forest_scene(current, coll, pos, active_mushrooms=safe_mushrooms),
@@ -103,7 +95,6 @@ def _autonomous_tick(current, collection, active_mushrooms, position):
         gr.update(),
         dex_markdown(coll),
         progress_markdown(coll),
-        sound_to_play, # <-- Ensure this is in your return statement
     )
 
 # ---------------------------------------------------------------------------
@@ -304,19 +295,12 @@ def move_player(position, direction, current, collection, active_mushrooms):
 
     safe_mushrooms = list(active_mushrooms or [])
     coll           = list(collection or [])
-    
-    # --- FIX: Initialize this BEFORE calling check_auto_pickup ---
-    initial_coll_len = len(coll) 
-    # -----------------------------------------------------------
 
-    # Collision check
+    # Collision check — pure Python, no LLM
     current, coll, safe_mushrooms = check_auto_pickup(
         x, y, current, coll, safe_mushrooms
     )
 
-    # Now compare it to the length after the pickup
-    sound_to_play = "./assets/sounds/mushroom_pickup.wav" if len(coll) > initial_coll_len else "assets/sounds/bot_roam.wav"
-
     # Update roam tracker
     with _ROAM_LOCK:
         _ROAM["position"] = [x, y]
@@ -333,7 +317,6 @@ def move_player(position, direction, current, collection, active_mushrooms):
         gr.update(),              # leave chat untouched
         dex_markdown(coll),
         progress_markdown(coll),
-        sound_to_play,            # The sound output
     )
 
 # ---------------------------------------------------------------------------
@@ -420,16 +403,6 @@ def _safe_pos(position):
     x, y = position
     return (max(0,min(2,int(x))), max(0,min(2,int(y))))
 
-def _stream_myco():
-    """Streams only Myco's AI responses — no errors, no logs."""
-    seen = ""
-    while True:
-        current = _LAST_REPLY["text"]
-        if current != seen:
-            seen = current
-            yield seen
-        time.sleep(0.3)
-
 
 # ---------------------------------------------------------------------------
 # Guide text
@@ -478,8 +451,6 @@ def build_app():
         # Autonomous roam timer — every 5 seconds, pure Python movement
         roam_timer = gr.Timer(5)
 
-        sound_player = gr.Audio(visible=False, autoplay=True)
-        
         with gr.Tabs(selected="home") as tabs:
 
             with gr.Tab("🏠 Home", id="home"):
@@ -570,14 +541,21 @@ def build_app():
             elem_id="myco-live-log",
         )
 
-        demo.load(fn=_stream_myco, outputs=[myco_live])
+        # Poll the running Myco log (narrative/action/chat) once per second.
+        # NOTE: previously this used demo.load(fn=_stream_myco, ...) with an
+        # infinite `while True` generator. An infinite generator from
+        # demo.load permanently occupies a queue worker/stream for that
+        # session — other events (button clicks) can then execute their
+        # Python function fine (visible in server logs) but never get their
+        # result delivered back to the browser. A periodic Timer avoids this.
+        log_timer = gr.Timer(1)
+        log_timer.tick(fn=get_myco_log, outputs=myco_live)
                 
         # ── Shared output lists ──
         _move_outputs = [
             player_position, scene, status, discovery,
             current_mushroom, collection, active_mushrooms,
-            chat, dex, progress, 
-            sound_player
+            chat, dex, progress,
         ]
 
         # ── Event wiring ──