feat: Remove OpenRouter support, replace litellm with Chutes API

README.md

@@ -1,6 +1,6 @@
 # BaseAgent - SDK 3.0
 
-High-performance autonomous agent for [Term Challenge](https://term.challenge). **Does NOT use term_sdk** - fully autonomous with litellm.
+High-performance autonomous agent for [Term Challenge](https://term.challenge). **Does NOT use term_sdk** - fully autonomous with Chutes API.
 
 ## Installation
 
@@ -36,7 +36,7 @@ my-agent/
 │   │   ├── loop.py       # Main loop
 │   │   └── compaction.py # Context management (MANDATORY)
 │   ├── llm/
-│   │   └── client.py     # LLM client (litellm)
+│   │   └── client.py     # LLM client (Chutes API)
 │   └── tools/
 │       └── ...           # Available tools
 ├── requirements.txt      # Dependencies
@@ -77,13 +77,13 @@ AUTO_COMPACT_THRESHOLD = 0.85
 
 ## Features
 
-### LLM Client (litellm)
+### LLM Client (Chutes API)
 
 ```python
-from src.llm.client import LiteLLMClient
+from src.llm.client import LLMClient
 
-llm = LiteLLMClient(
-    model="openrouter/anthropic/claude-opus-4.5",
+llm = LLMClient(
+    model="deepseek/deepseek-chat",
     temperature=0.0,
     max_tokens=16384,
 )
@@ -129,7 +129,7 @@ See `src/config/defaults.py`:
 
 ```python
 CONFIG = {
-    "model": "openrouter/anthropic/claude-opus-4.5",
+    "model": "deepseek/deepseek-chat",
     "max_tokens": 16384,
     "max_iterations": 200,
     "auto_compact_threshold": 0.85,
@@ -142,7 +142,7 @@ CONFIG = {
 
 | Variable | Description |
 |----------|-------------|
-| `OPENROUTER_API_KEY` | OpenRouter API key |
+| `CHUTES_API_KEY` | Chutes API key |
 
 ## Documentation
 
@@ -151,7 +151,7 @@ CONFIG = {
 See [rules/](rules/) for comprehensive guides:
 
 - [Architecture Patterns](rules/02-architecture-patterns.md) - **Mandatory project structure**
-- [LLM Usage Guide](rules/06-llm-usage-guide.md) - **Using litellm**
+- [LLM Usage Guide](rules/06-llm-usage-guide.md) - **Using Chutes API**
 - [Best Practices](rules/05-best-practices.md)
 - [Error Handling](rules/08-error-handling.md)
 

agent.py

@@ -3,7 +3,7 @@
 SuperAgent for Term Challenge - Entry Point (SDK 3.0 Compatible).
 
 This agent accepts --instruction from the validator and runs autonomously.
-Uses litellm for LLM calls instead of term_sdk.
+Uses Chutes API for LLM calls instead of term_sdk.
 
 Installation:
     pip install .                    # via pyproject.toml
@@ -16,56 +16,61 @@
 from __future__ import annotations
 
 import argparse
-import sys
-import time
 import os
 import subprocess
+import sys
+import time
 from pathlib import Path
 
 # Add parent to path for imports
 sys.path.insert(0, str(Path(__file__).parent))
 
+
 # Auto-install dependencies if missing
 def ensure_dependencies():
     """Install dependencies if not present."""
     try:
-        import litellm
         import httpx
         import pydantic
     except ImportError:
         print("[setup] Installing dependencies...", file=sys.stderr)
         agent_dir = Path(__file__).parent
         req_file = agent_dir / "requirements.txt"
         if req_file.exists():
-            subprocess.run([sys.executable, "-m", "pip", "install", "-r", str(req_file), "-q"], check=True)
+            subprocess.run(
+                [sys.executable, "-m", "pip", "install", "-r", str(req_file), "-q"], check=True
+            )
         else:
-            subprocess.run([sys.executable, "-m", "pip", "install", str(agent_dir), "-q"], check=True)
+            subprocess.run(
+                [sys.executable, "-m", "pip", "install", str(agent_dir), "-q"], check=True
+            )
         print("[setup] Dependencies installed", file=sys.stderr)
 
+
 ensure_dependencies()
 
 from src.config.defaults import CONFIG
 from src.core.loop import run_agent_loop
+from src.llm.client import CostLimitExceeded, LLMClient
+from src.output.jsonl import ErrorEvent, emit
 from src.tools.registry import ToolRegistry
-from src.output.jsonl import emit, ErrorEvent
-from src.llm.client import LiteLLMClient, CostLimitExceeded
 
 
 class AgentContext:
     """Minimal context for agent execution (replaces term_sdk.AgentContext)."""
-    
+
     def __init__(self, instruction: str, cwd: str = None):
         self.instruction = instruction
         self.cwd = cwd or os.getcwd()
         self.step = 0
         self.is_done = False
         self.history = []
         self._start_time = time.time()
-    
+
     @property
     def elapsed_secs(self) -> float:
         return time.time() - self._start_time
-    
+
     def shell(self, cmd: str, timeout: int = 120) -> "ShellResult":
         """Execute a shell command."""
         self.step += 1
@@ -86,20 +91,22 @@ def shell(self, cmd: str, timeout: int = 120) -> "ShellResult":
         except Exception as e:
             output = f"[ERROR] {e}"
             exit_code = -1
-        
+
         shell_result = ShellResult(output=output, exit_code=exit_code)
-        self.history.append({
-            "step": self.step,
-            "command": cmd,
-            "output": output[:1000],
-            "exit_code": exit_code,
-        })
+        self.history.append(
+            {
+                "step": self.step,
+                "command": cmd,
+                "output": output[:1000],
+                "exit_code": exit_code,
+            }
+        )
         return shell_result
-    
+
     def done(self):
         """Mark task as complete."""
         self.is_done = True
-    
+
     def log(self, msg: str):
         """Log a message."""
         timestamp = time.strftime("%H:%M:%S")
@@ -108,13 +115,13 @@ def log(self, msg: str):
 
 class ShellResult:
     """Result from shell command."""
-    
+
     def __init__(self, output: str, exit_code: int):
         self.output = output
         self.stdout = output
         self.stderr = ""
         self.exit_code = exit_code
-    
+
     def has(self, text: str) -> bool:
         return text in self.output
 
@@ -129,29 +136,29 @@ def main():
     parser = argparse.ArgumentParser(description="SuperAgent for Term Challenge SDK 3.0")
     parser.add_argument("--instruction", required=True, help="Task instruction from validator")
     args = parser.parse_args()
-    
+
     _log("=" * 60)
-    _log("SuperAgent Starting (SDK 3.0 - litellm)")
+    _log("SuperAgent Starting (SDK 3.0 - Chutes API)")
     _log("=" * 60)
     _log(f"Model: {CONFIG['model']}")
     _log(f"Reasoning effort: {CONFIG.get('reasoning_effort', 'default')}")
     _log(f"Instruction: {args.instruction[:200]}...")
     _log("-" * 60)
-    
+
     # Initialize components
     start_time = time.time()
-    
-    llm = LiteLLMClient(
+
+    llm = LLMClient(
         model=CONFIG["model"],
         temperature=CONFIG.get("temperature"),
         max_tokens=CONFIG.get("max_tokens", 16384),
     )
-    
+
     tools = ToolRegistry()
     ctx = AgentContext(instruction=args.instruction)
-    
+
     _log("Components initialized")
-    
+
     try:
         run_agent_loop(
             llm=llm,

astuces/08-cost-optimization.md

@@ -2,21 +2,20 @@
 
 ## Cost Breakdown
 
-For Claude Sonnet via OpenRouter:
+Typical LLM pricing (varies by model):
 
-| Token Type | Cost per 1M |
-|------------|-------------|
-| Input tokens | $3.00 |
-| Cached input | $0.30 (90% off) |
-| Output tokens | $15.00 |
+| Token Type | Typical Cost per 1M |
+|------------|---------------------|
+| Input tokens | $1.00 - $15.00 |
+| Cached input | 10-50% of input |
+| Output tokens | $2.00 - $60.00 |
 
 For a typical task:
 - 50 turns
 - 100k context average
 - 500 output tokens per turn
 
-**Without optimization**: 50 × 100k × $3/1M = **$15 per task**
-**With 90% caching**: 50 × 100k × $0.30/1M = **$1.50 per task**
+Costs vary significantly by model choice. DeepSeek models are typically more cost-effective than Claude or GPT-4.
 
 ## Optimization Strategies
 

pyproject.toml

@@ -26,7 +26,6 @@ dependencies = [
     "tomli-w>=1.0",
     "rich>=13.0",
     "typer>=0.12.0",
-    "litellm>=1.50.0",
 ]
 
 [project.optional-dependencies]

requirements.txt

@@ -4,4 +4,3 @@ tomli>=2.0;python_version<'3.11'
 tomli-w>=1.0
 rich>=13.0
 typer>=0.12.0
-litellm>=1.50.0

rules/02-architecture-patterns.md

@@ -20,7 +20,7 @@ my-agent/
 │   │   ├── loop.py       # Main loop
 │   │   └── compaction.py # Context management (MANDATORY)
 │   ├── llm/
-│   │   └── client.py     # LLM client (litellm)
+│   │   └── client.py     # LLM client (Chutes API)
 │   └── tools/
 │       └── ...           # Tools
 ├── requirements.txt      # Dependencies
@@ -275,7 +275,7 @@ flowchart TB
 ### Implementation
 
 ```python
-# Définition des outils (format OpenAI/litellm)
+# Tool definition (OpenAI-compatible format)
 TOOLS = [
     {
         "name": "run_command",