utim-cli 1.47.4__tar.gz → 1.47.6__tar.gz

{utim_cli-1.47.4 → utim_cli-1.47.6}/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## [1.47.6] - 2026-07-19
+
+### Added
+- **Asynchronous Local LLM Setup**: Moved the local Qwen model downloading and setup pipeline into a background daemon thread, allowing UTIM to run completions instantly via Cloud API during initial setup/installation without blocking user queries.
+
+## [1.47.5] - 2026-07-19
+
+### Optimized
+- **Instant startup / rules summarization**: Prioritized Cloud API completion calls (via UTIM completions server or direct OpenRouter) over local Qwen LLM execution in `summarize_experiences_with_model`. This completely eliminates the 1m 49s latency caused by importing PyTorch/Transformers and loading the 0.5B model from scratch, falling back to local execution only in offline mode.
+
 ## [1.47.4] - 2026-07-19
 
 ### Fixed

{utim_cli-1.47.4/utim_cli.egg-info → utim_cli-1.47.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: utim-cli
-Version: 1.47.4
+Version: 1.47.6
 Summary: UTIM – Universal Terminal Intelligence Manager. A powerful agentic AI coding assistant for your terminal.
 License: MIT
 Project-URL: Homepage, https://utim.dev

{utim_cli-1.47.4 → utim_cli-1.47.6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "utim-cli"
-version = "1.47.4"
+version = "1.47.6"
 description = "UTIM – Universal Terminal Intelligence Manager. A powerful agentic AI coding assistant for your terminal."
 readme = "README.md"
 requires-python = ">=3.9"

{utim_cli-1.47.4 → utim_cli-1.47.6}/utim_cli/__init__.py

@@ -5,7 +5,7 @@ try:
     from utim_cli._version import VERSION as __version__
 except ImportError:
     # Fallback during first-install / editable-install edge cases
-    __version__ = "1.47.4"
+    __version__ = "1.47.6"
 
 import builtins, os, pathlib, urllib.parse, re
 

{utim_cli-1.47.4 → utim_cli-1.47.6}/utim_cli/_version.py

@@ -11,4 +11,4 @@
 # from here so every surface stays in sync automatically.
 # ─────────────────────────────────────────────────────────────────────────────
 
-VERSION = "1.47.4"
+VERSION = "1.47.6"

{utim_cli-1.47.4 → utim_cli-1.47.6}/utim_cli/orchestrator.py

@@ -350,6 +350,42 @@ def compress_experience(text: str) -> str:
 
 _qwen_model = None
 _qwen_tokenizer = None
+_qwen_downloading = False
+
+def _trigger_qwen_download_bg():
+    """Trigger background download and caching of Qwen local model if not already loaded or downloading."""
+    global _qwen_model, _qwen_tokenizer, _qwen_downloading
+    if _qwen_model is None and not _qwen_downloading:
+        import sys
+        if "pytest" in sys.modules:
+            return
+        
+        _qwen_downloading = True
+        import threading
+        
+        def download_model_bg():
+            global _qwen_model, _qwen_tokenizer, _qwen_downloading
+            try:
+                import torch
+                from transformers import AutoModelForCausalLM, AutoTokenizer
+                model_id = "Qwen/Qwen2.5-0.5B-Instruct"
+                tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True, local_files_only=False)
+                mod = AutoModelForCausalLM.from_pretrained(
+                    model_id,
+                    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+                    device_map="auto",
+                    trust_remote_code=True,
+                    local_files_only=False
+                )
+                _qwen_tokenizer = tok
+                _qwen_model = mod
+            except Exception:
+                pass
+            finally:
+                _qwen_downloading = False
+                
+        t = threading.Thread(target=download_model_bg, daemon=True)
+        t.start()
 
 def summarize_experiences_with_model(experiences: list[dict], existing_cache: str = "") -> str:
     """Summarize list of experiences into a clean list of rules under 2000 tokens using local Qwen model first, falling back to LLM API."""
@@ -409,25 +445,40 @@ def summarize_experiences_with_model(experiences: list[dict], existing_cache: st
             "Output ONLY the clean markdown list of rules (do not include introductory or concluding text):"
         )
 
-    # 1. Attempt local execution using Qwen2.5-0.5B-Instruct
-    try:
-        import sys
-        if "pytest" not in sys.modules:
-            global _qwen_model, _qwen_tokenizer
+    # Proactively trigger Qwen local model download in a background thread if needed
+    _trigger_qwen_download_bg()
+
+    # 1. Attempt Cloud API completions (via UTIM completions server or direct OpenRouter)
+    import os
+    api_key = config.get("api_key")
+    llm_key = os.environ.get("OPENROUTER_API_KEY") or os.environ.get("UTIM_API_KEY")
+    
+    if api_key or llm_key:
+        from utim_cli.reflection import REFLECTION_MODELS
+        from utim_cli.client_utils import proxy_openrouter_request
+        
+        for model in REFLECTION_MODELS:
+            try:
+                resp = proxy_openrouter_request({
+                    "model": model,
+                    "messages": [
+                        {"role": "system", "content": "You are a context compressor. Output rules list directly."},
+                        {"role": "user", "content": prompt}
+                    ],
+                    "max_tokens": 1000
+                })
+                if resp.status_code == 200:
+                    content = resp.json()["choices"][0]["message"]["content"].strip()
+                    if content:
+                        return content
+            except Exception:
+                continue
+
+    # 2. Fallback to local execution using Qwen2.5-0.5B-Instruct if background loader has completed
+    global _qwen_model, _qwen_tokenizer
+    if _qwen_model is not None:
+        try:
             import torch
-            from transformers import AutoModelForCausalLM, AutoTokenizer
-            
-            model_id = "Qwen/Qwen2.5-0.5B-Instruct"
-            
-            if _qwen_model is None:
-                _qwen_tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-                _qwen_model = AutoModelForCausalLM.from_pretrained(
-                    model_id,
-                    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-                    device_map="auto",
-                    trust_remote_code=True
-                )
-                
             messages = [
                 {"role": "system", "content": "You are a context compressor. Output rules list directly."},
                 {"role": "user", "content": prompt}
@@ -454,36 +505,9 @@ def summarize_experiences_with_model(experiences: list[dict], existing_cache: st
             response = _qwen_tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
             if response:
                 return response
-                
-    except Exception:
-        pass
-
-    # 2. Fallback to OpenRouter Cloud API
-    import os
-    llm_key = os.environ.get("OPENROUTER_API_KEY") or os.environ.get("UTIM_API_KEY")
-    if not llm_key:
-        return existing_cache or "\n".join(raw_texts[:15])
-        
-    from utim_cli.reflection import REFLECTION_MODELS
-    from utim_cli.client_utils import proxy_openrouter_request
-    
-    for model in REFLECTION_MODELS:
-        try:
-            resp = proxy_openrouter_request({
-                "model": model,
-                "messages": [
-                    {"role": "system", "content": "You are a context compressor. Output rules list directly."},
-                    {"role": "user", "content": prompt}
-                ],
-                "max_tokens": 1000
-            })
-            if resp.status_code == 200:
-                content = resp.json()["choices"][0]["message"]["content"].strip()
-                if content:
-                    return content
         except Exception:
-            continue
-            
+            pass
+
     return existing_cache or "\n".join(raw_texts[:15])
 
 def update_experience_summary_cache(user_prompt: str) -> str:

{utim_cli-1.47.4 → utim_cli-1.47.6}/utim_cli/utim.py

@@ -952,7 +952,7 @@ def _print_animated_banner(animated: bool = True):
     else:
         console.print(f"[bold {color}]{BANNER_BIG}[/bold {color}]")
             
-    console.print(f" [bold white]U Think I Make[/bold white] [dim]v1.47.4[/dim]")
+    console.print(f" [bold white]U Think I Make[/bold white] [dim]v1.47.6[/dim]")
     user_email = config.email
     user_type = "UTIM Community"
     api_key = config.get("api_key")
@@ -3586,7 +3586,7 @@ def check_and_update_background():
                 return
 
             latest_ver = resp.json().get("version", "")
-            current_ver = "1.47.4"
+            current_ver = "1.47.6"
 
             def parse_ver(v):
                 return [int(x) for x in v.split(".") if x.isdigit()]

{utim_cli-1.47.4 → utim_cli-1.47.6/utim_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: utim-cli
-Version: 1.47.4
+Version: 1.47.6
 Summary: UTIM – Universal Terminal Intelligence Manager. A powerful agentic AI coding assistant for your terminal.
 License: MIT
 Project-URL: Homepage, https://utim.dev