File size: 6,481 Bytes
9b30c36 b9a8029 9b30c36 b9a8029 9b30c36 071b72b 9b30c36 03d4a62 9b30c36 071b72b 9b30c36 071b72b 9b30c36 abcbc9f 9b30c36 071b72b 9b30c36 45e918d 9b30c36 071b72b 45e918d 9b30c36 b9a8029 9b30c36 b9a8029 6704b62 9b30c36 6704b62 9b30c36 11c4bc3 9b30c36 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 | """
CodeSentinel Baseline Inference Script
Uses the OpenAI client to run gpt-4o against all 3 tasks.
Reads API key from OPENAI_API_KEY env var.
Usage: python baseline/run_baseline.py [--base-url http://localhost:7860] [--seed 42]
"""
import os
import json
import argparse
import sys
try:
import httpx
except ImportError:
import urllib.request as _urllib
httpx = None
from openai import OpenAI
BASE_URL = "http://localhost:7860"
SYSTEM_PROMPT = """\
You are an expert software engineer performing code review.
You will be shown code and must identify issues by calling the correct action.
Available actions (respond ONLY with valid JSON, no prose, no markdown fences):
flag_bug: {"action_type":"flag_bug","line":<int>,"severity":<low|medium|high|critical>,"message":"<description>"}
flag_vulnerability: {"action_type":"flag_vulnerability","line":<int>,"cve_category":"<OWASP cat>","severity":<level>,"message":"<desc>"}
add_comment: {"action_type":"add_comment","line":<int>,"message":"<comment>"}
request_changes: {"action_type":"request_changes","reason":"<summary of all findings>"}
approve: {"action_type":"approve","reason":"<reason — only if truly no issues>"}
submit_report: {"action_type":"submit_report","report":{"security":[...],"reliability":[...],"api_design":[...],"observability":[...],"scalability":[...]}}
Respond with exactly ONE JSON action per turn. No prose, no markdown fences.
"""
def _post(url: str, data: dict) -> dict:
try:
body = json.dumps(data).encode()
if httpx:
return httpx.post(url, content=body, headers={"Content-Type": "application/json"}).json()
req = __import__("urllib.request").request.Request(
url, data=body, headers={"Content-Type": "application/json"}, method="POST"
)
with __import__("urllib.request").request.urlopen(req) as r:
return json.loads(r.read())
except Exception as e:
raise RuntimeError(f"Network/parsing error in _post to {url}: {e}")
def _get(url: str) -> dict:
try:
if httpx:
return httpx.get(url).json()
with __import__("urllib.request").request.urlopen(url) as r:
return json.loads(r.read())
except Exception as e:
raise RuntimeError(f"Network/parsing error in _get from {url}: {e}")
def obs_to_prompt(obs: dict) -> str:
lines = [
f"Task: {obs['task_id']}",
f"Step: {obs['step']}/{obs['max_steps']}",
]
if obs.get("hint"):
lines.append(f"Hint: {obs['hint']}")
lines.append(f"\nObjective:\n{obs['task_description']}\n")
for f in obs["files"]:
lines.append(f"=== {f['filename']} ===\n{f['content']}")
if obs["history"]:
lines.append("\nYour actions so far:")
for h in obs["history"][-6:]:
lines.append(f" Step {h['step']}: {h['action']} line={h.get('line')} — {h['message']}")
return "\n".join(lines)
def run_task(client: OpenAI, task_id: str, seed: int = 42) -> float:
print(f"\n{'='*60}")
print(f" Task: {task_id.upper()}")
print(f"{'='*60}")
print(f"[START] task={task_id}", flush=True)
reset_resp = _post(f"{BASE_URL}/reset", {"task_id": task_id, "seed": seed})
obs = reset_resp
messages = [{"role": "system", "content": SYSTEM_PROMPT}]
final_score = 0.0
step_index = 0
for step_index in range(obs["max_steps"]):
messages.append({"role": "user", "content": obs_to_prompt(obs)})
try:
completion = client.chat.completions.create(
model = "gpt-4o",
messages = messages,
temperature = 0.0,
max_tokens = 512,
)
raw_action = completion.choices[0].message.content.strip()
except Exception as e:
print(f" [ERROR] LLM generation failed: {e}")
break
print(f" Step {obs['step']+1}: {raw_action[:100]}")
messages.append({"role": "assistant", "content": raw_action})
try:
action = json.loads(raw_action)
except json.JSONDecodeError:
print(" [WARN] Invalid JSON — skipping step")
continue
step_resp = _post(f"{BASE_URL}/step", action)
obs = step_resp["observation"]
reward = step_resp["reward"]
done = step_resp["done"]
final_score = step_resp["info"].get("final_score") or reward["value"]
print(f"[STEP] step={step_index + 1} reward={reward['value']:.3f}", flush=True)
print(f" reward={reward['value']:.3f} delta={reward['delta']:+.3f} done={done}")
if done:
break
final_score = final_score or 0.001
print(f"\n FINAL SCORE: {final_score:.4f}")
print(f"[END] task={task_id} score={final_score:.4f} steps={step_index + 1}", flush=True)
return final_score
def ascii_bar(score: float, width: int = 20) -> str:
filled = round(score * width)
return "█" * filled + "░" * (width - filled)
def main():
global BASE_URL
parser = argparse.ArgumentParser(description="CodeSentinel GPT-4o baseline")
parser.add_argument("--base-url", default=BASE_URL)
parser.add_argument("--seed", type=int, default=42)
args = parser.parse_args()
BASE_URL = args.base_url
try:
_get(f"{BASE_URL}/validate")
except Exception as e:
print(f"ERROR: Server at {BASE_URL} is unreachable: {e}", file=sys.stderr)
return
api_key = os.environ.get("API_KEY") or os.environ.get("OPENAI_API_KEY", "dummy_key_for_validation")
api_base_url = os.environ.get("API_BASE_URL")
if api_base_url:
client = OpenAI(api_key=api_key, base_url=api_base_url)
else:
client = OpenAI(api_key=api_key)
tasks = ["bug_detection", "security_audit", "architecture_review"]
scores = {}
for task_id in tasks:
try:
scores[task_id] = run_task(client, task_id, seed=args.seed)
except Exception as exc:
print(f" [ERROR] {exc}")
scores[task_id] = 0.001
print(f"\n{'='*60}")
print(" BASELINE RESULTS")
print(f"{'='*60}")
for task_id, score in scores.items():
print(f" {task_id:<25} [{ascii_bar(score)}] {score:.4f}")
avg = sum(scores.values()) / len(scores)
print(f"\n Average score: {avg:.4f}")
print(f"{'='*60}\n")
if __name__ == "__main__":
main()
|