feat: add task weights to configuration and implement LLM-based inference agent

README.md

@@ -87,16 +87,18 @@ python inference.py --fast-mode --episodes 1
 
 You can run the same entrypoint directly with `python python/inference.py` (e.g. `python python/inference.py --fast-mode`); flags match the root `inference.py` wrapper.
 
-**LLM baseline** (requires Hugging Face or other OpenAI-compatible API credentials):
+**LLM baseline** (requires any OpenAI-compatible API credentials — HuggingFace, Groq, etc.):
 
 ```bash
 export ENV_URL=http://localhost:7860
 export API_BASE_URL=https://router.huggingface.co/v1
 export MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
-export HF_TOKEN=your_token_here
+export OPENAI_API_KEY=your_token_here   # or HF_TOKEN=your_token_here
 python inference.py --episodes 1 --llm-every 4
 ```
 
+> **Note:** The script accepts either `OPENAI_API_KEY` (hackathon standard) or `HF_TOKEN` (HuggingFace convention). You do **not** need a paid OpenAI key — any OpenAI-compatible provider works.
+
 Results are written to `baseline_scores.json` by default (`--output` to change).
 
 ---
@@ -160,7 +162,7 @@ There is **no** `--judge-mode` flag in this repository. Use the modes below.
 | Mode | Command pattern | Behavior |
 |------|-----------------|----------|
 | **Fast (heuristic)** | `python inference.py --fast-mode` | No LLM calls; deterministic given env seed; fastest for CI or smoke tests. |
-| **Default LLM** | `python inference.py` | Uses OpenAI-compatible API (`API_BASE_URL`, `MODEL_NAME`, `HF_TOKEN`); default `--llm-every 4` reuses each LLM action for 4 steps to limit API cost. |
+| **Default LLM** | `python inference.py` | Uses OpenAI Python client (`API_BASE_URL`, `MODEL_NAME`, `OPENAI_API_KEY` or `HF_TOKEN`); default `--llm-every 4` reuses each LLM action for 4 steps to limit API cost. |
 | **Recommended for automated evaluation / judging** | `python inference.py --fast-mode --episodes 1` | Recommended when automated pipelines need **reproducibility** and **no external API** dependency. |
 
 Other useful flags:

openenv.yaml

@@ -111,14 +111,25 @@ tasks:
     name: "Cost Minimization"
     description: "Minimize total energy cost over a 24-hour episode with no process constraints."
     difficulty: "easy"
+    weights:
+      cost: 1.0
   - id: 2
     name: "Constrained Temperature Management"
     description: "Minimize cost while keeping indoor temperature within ±2°C of setpoint at all times."
     difficulty: "medium"
+    weights:
+      cost: 0.6
+      temperature: 0.4
   - id: 3
     name: "Full Demand-Response with Batch Scheduling"
     description: "Minimize cost, maintain temperature, respond to grid stress events, schedule all batch jobs, and minimize carbon."
     difficulty: "hard"
+    weights:
+      cost: 0.28
+      temperature: 0.20
+      grid_response: 0.20
+      batch_deadline: 0.12
+      carbon: 0.20
 
 endpoints:
   health:
@@ -145,3 +156,7 @@ endpoints:
   tasks:
     path: /tasks
     method: GET
+  metrics:
+    path: /metrics
+    method: GET
+

python/inference.py

@@ -2,13 +2,19 @@
 GridMind-RL Baseline Inference Script
 --------------------------------------
 Runs an LLM agent against all 3 tasks for N episodes each.
-Uses OpenAI-compatible API via API_BASE_URL / MODEL_NAME / HF_TOKEN environment variables.
+Uses the OpenAI Python client pointed at any OpenAI-compatible endpoint.
+
+Required environment variables:
+    API_BASE_URL  — The API endpoint for the LLM (default: HuggingFace router)
+    MODEL_NAME    — The model identifier to use for inference
+    OPENAI_API_KEY or HF_TOKEN — API key for authentication (any provider)
 
 Usage:
+    export API_BASE_URL=https://router.huggingface.co/v1
     export MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
-    export HF_TOKEN=hf_xxxx
+    export OPENAI_API_KEY=hf_xxxx   # or HF_TOKEN=hf_xxxx
     python inference.py
-    # or: python python/inference.py [--episodes 1] [--llm-every 4] [--fast-mode]
+    # or: python inference.py --fast-mode --episodes 1
 """
 
 from __future__ import annotations
@@ -28,7 +34,8 @@ from openai import OpenAI
 ENV_URL = os.getenv("ENV_URL", "http://localhost:7860")
 MODEL_NAME = os.getenv("MODEL_NAME", "meta-llama/Llama-3.1-8B-Instruct")
 API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
-HF_TOKEN = os.getenv("HF_TOKEN", "")
+# Accept OPENAI_API_KEY (hackathon standard) or HF_TOKEN (HuggingFace convention)
+OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "") or os.getenv("HF_TOKEN", "")
 DEFAULT_EPISODES = 1
 DEFAULT_SEED_BASE = 1000
 MAX_RETRIES = 3
@@ -124,7 +131,7 @@ class LLMAgent:
     def __init__(self):
         self.client = OpenAI(
             base_url=API_BASE_URL,
-            api_key=HF_TOKEN if HF_TOKEN else "none",
+            api_key=OPENAI_API_KEY if OPENAI_API_KEY else "none",
         )
         self.model = MODEL_NAME
         self.fallback_mode = False