LLM4HEP / supervisor_coder.py
ho22joshua's picture
initial commit
cfcbbc8
Raw
History Blame Contribute Delete
22.9 kB
import os
import re
import sys
import openai
import subprocess
import argparse
import numpy as np
import yaml
from datetime import datetime
base_dir = os.getcwd()
# Find the project root directory (where prompts/ directory is located)
# When running from results, we need to go up one level
project_root = os.path.dirname(base_dir) if 'results' in base_dir else base_dir
parser = argparse.ArgumentParser(description='supervisor/coder')
add_arg = parser.add_argument
add_arg('--prompt', help='prompt name')
add_arg('--out_dir', help='output directory')
add_arg('--input-files', nargs='*', help='List of input files from snakemake')
add_arg('--output-files', nargs='*', help='List of output files from snakemake')
add_arg('--config', help='Path to config file', default=os.path.join(project_root, 'config.yml'))
add_arg('--max_iterations', type=int, default=3, help='Maximum coder iterations for faster runs')
args = parser.parse_args()
# Read supervisor and coder from config.yml
config_path = args.config
if os.path.exists(config_path):
try:
with open(config_path, 'r') as f:
config = yaml.safe_load(f)
supervisor = config.get('supervisor')
if not supervisor:
print(f"ERROR: 'supervisor' not found in {config_path}")
sys.exit(1)
coder = config.get('coder')
if not coder:
print(f"ERROR: 'coder' not found in {config_path}")
sys.exit(1)
except Exception as e:
print(f"ERROR: Could not read {config_path}: {e}")
sys.exit(1)
else:
print(f"ERROR: config file not found at {config_path}")
sys.exit(1)
temperature = config.get('temperature', None)
def get_code(output): # convert Markdown to Python
match = re.search(r'```python\s*(.*?)\s*```', output, re.DOTALL)
if match:
return match.group(1).strip()
return output
name = args.prompt
out_dir = args.out_dir
os.makedirs(out_dir, exist_ok=True)
log_dir = os.path.join(out_dir, 'logs')
os.makedirs(log_dir, exist_ok=True)
code_dir = os.path.join(out_dir, 'generated_code')
os.makedirs(code_dir, exist_ok=True)
prompt_pair_dir = os.path.join(out_dir, 'prompt_pairs') # for saving supervisor/user prompt pairs
os.makedirs(prompt_pair_dir, exist_ok=True)
prompt_filepath = os.path.join(out_dir, f'prompts_temp/{name}.txt')
# If the prompt file doesn't exist in prompts_temp, try logs (for modified prompts)
if not os.path.exists(prompt_filepath):
prompt_filepath = os.path.join(out_dir, f'logs/{name}.txt')
with open(prompt_filepath, 'r') as file:
user_prompt = file.read()
first_supervisor_instructions_filepath = os.path.join(out_dir, 'prompts_temp/supervisor_first_call.txt')
with open(first_supervisor_instructions_filepath, 'r') as file:
first_supervisor_instructions = file.read()
supervisor_instructions_filepath = os.path.join(out_dir, 'prompts_temp/supervisor_call.txt')
with open(supervisor_instructions_filepath, 'r') as file:
supervisor_instructions = file.read()
# Build initial supervisor prompt with file context
file_context = ''
if args.input_files:
file_context += '\nInput files:\n' + '\n'.join(args.input_files)
if args.output_files:
file_context += '\nOutput files:\n' + '\n'.join(args.output_files)
supervisor_prompt = first_supervisor_instructions + file_context + '\n\n' + user_prompt
client = openai.OpenAI(
api_key = os.environ.get('CBORG_API_KEY'),
base_url = 'https://api.cborg.lbl.gov'
)
done_outer = False
count_outer = 0
total_calls = 0
input_tokens = 0
output_tokens = 0
tokens = {"User Prompt": 0, "Supervisor to Coder": 0, "Coder Outputs": 0, "Feedback to Supervisor": 0}
# Create comprehensive log (only log file we'll create)
comprehensive_log_path = os.path.join(log_dir, f'{name}_comprehensive_log.txt')
start_time = datetime.now()
# Initialize comprehensive log with header
with open(comprehensive_log_path, 'w') as comp_log:
comp_log.write("=" * 100 + "\n")
comp_log.write("🎯 COMPREHENSIVE SUPERVISOR-CODER LOG\n")
comp_log.write("=" * 100 + "\n\n")
comp_log.write(f"πŸ“‹ Task: {name}\n")
comp_log.write(f"πŸ€– Supervisor: {supervisor}\n")
comp_log.write(f"πŸ€– Coder: {coder}\n")
comp_log.write(f"πŸ• Start Time: {start_time}\n")
comp_log.write(f"πŸ“ Working Directory: {os.getcwd()}\n")
comp_log.write(f"πŸ“‚ Output Directory: {out_dir}\n\n")
# Log the original user prompt
comp_log.write("πŸ“ ORIGINAL USER PROMPT\n")
comp_log.write("-" * 50 + "\n")
comp_log.write(user_prompt)
comp_log.write("\n\n")
# Log supervisor instructions
# comp_log.write("πŸŽ“ SUPERVISOR INSTRUCTIONS\n")
# comp_log.write("-" * 50 + "\n")
# comp_log.write(first_supervisor_instructions)
# comp_log.write("\n\n")
# Log supervisor call template
# comp_log.write("πŸ“‹ SUPERVISOR CALL TEMPLATE\n")
# comp_log.write("-" * 50 + "\n")
# comp_log.write(supervisor_instructions)
# comp_log.write("\n\n")
comp_log.write("πŸš€ PROCESS START\n")
comp_log.write("-" * 50 + "\n\n")
# Function to log to comprehensive log
def log_comprehensive(message, section="", level="INFO", plain=False):
timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
with open(comprehensive_log_path, 'a') as comp_log:
if section:
# Use more prominent section headers
comp_log.write(f"\n\n{'='*25} {section.upper()} {'='*25}\n\n")
# Truncate very long messages to prevent memory issues
if len(message) > 50000: # 50KB limit
truncated_message = message[:50000] + f"\n\n[TRUNCATED: message was {len(message)} characters long]"
message = truncated_message
if plain:
# For logging raw content like prompts or code
comp_log.write(f"{message}\n")
else:
comp_log.write(f"[{timestamp}] {level}: {message}\n")
log_comprehensive("Starting supervisor-coder interaction", "πŸš€ INITIALIZATION")
print("=== SUPERVISOR-CODER MAIN LOOP ===")
log_comprehensive("=== MAIN LOOP STARTED ===", "πŸ”„ LOOP START")
print(f"πŸ”„ Calling supervisor {supervisor} (initial)")
log_comprehensive(f"Calling supervisor {supervisor} with initial prompt", "πŸ€– SUPERVISOR CALL")
try:
response = client.chat.completions.create(
model = supervisor,
messages = [
{
'role': 'user',
'content': supervisor_prompt
}
],
temperature = temperature
)
tokens['User Prompt'] += response.usage.prompt_tokens
input_tokens += response.usage.prompt_tokens
output_tokens += response.usage.completion_tokens
supervisor_response = response.choices[-1].message.content
# Log initial prompt tokens
print(f"Initial supervisor prompt tokens: {tokens['User Prompt']}")
log_comprehensive(f"Initial supervisor prompt tokens: {tokens['User Prompt']}", "TOKEN STATS")
# Log token stats and content under SUPERVISOR RESPONSE section
log_comprehensive(
f"Supervisor response received ({response.usage.prompt_tokens} input tokens, {response.usage.completion_tokens} output tokens)",
"SUPERVISOR RESPONSE"
)
log_comprehensive(supervisor_response, plain=True)
# Include input and output files under the same section without extra headers
log_comprehensive(file_context, section="", plain=True)
except Exception as e:
log_comprehensive(f"Supervisor API call failed: {e}", "❌ API ERROR", "ERROR")
print(f"OpenAI API error: {e}")
# Check if it's a rate limit or quota issue (temporary)
if "429" in str(e) or "rate" in str(e).lower() or "quota" in str(e).lower():
print("Rate limit or quota exceeded. This is a temporary issue.")
print("Stopping job to avoid wasting resources on temporary failures.")
log_comprehensive("Terminating due to rate limit/quota issue", "πŸ›‘ TERMINATION", "ERROR")
sys.exit(1) # Fail for rate limits
else:
print("Permanent API error (auth, network, etc.). Stopping job.")
log_comprehensive("Terminating due to permanent API error", "πŸ›‘ TERMINATION", "ERROR")
sys.exit(1) # Fail for permanent errors
if 'Call record:' in supervisor_response:
coder_prompt, record = supervisor_response.split('Call record:', 1)
else:
coder_prompt = supervisor_response
record = "Supervisor did not provide a record."
# Commented out parsing warning to silence logs
# log_comprehensive("Supervisor response missing 'Call record:'. Treating entire response as coder prompt.", "PARSING WARNING", "WARNING")
supervisor_record = os.path.join(log_dir, 'supervisor_record.txt')
with open(supervisor_record, 'w') as f:
f.write('Call 1 record: ' + record)
f.write(f'WARNING: running supervisor/coder with supervisor {supervisor} and coder {coder}')
# log_comprehensive(f"Initial supervisor record: {record}", "πŸ“ SUPERVISOR RECORD")
done_inner = False
count_inner = 0
old_coder_prompt = "Default coder prompt" # Initialize to avoid unbound variable
while not done_inner and count_inner < args.max_iterations:
count_inner += 1
print(f"πŸ€– Calling coder {coder} (attempt {count_inner}/{args.max_iterations})")
log_comprehensive(f"Calling coder {coder} (attempt {count_inner}/{args.max_iterations})", f"πŸ€– CODER CALL #{count_inner}")
# log_comprehensive(f"Coder Prompt:\n{coder_prompt}")
try:
print("DEBUG: Making coder API call...")
response = client.chat.completions.create(
model = coder,
messages = [
{
'role': 'user',
'content': coder_prompt
}
],
temperature = temperature
)
print("DEBUG: Coder API call completed successfully")
input_tokens += response.usage.prompt_tokens
output_tokens += response.usage.completion_tokens
tokens['Supervisor to Coder'] += response.usage.prompt_tokens
tokens['Coder Outputs'] += response.usage.completion_tokens
supervisor_response = response.choices[-1].message.content
output = response.choices[-1].message.content
print(f"DEBUG: Extracted output, type: {type(output)}, length: {len(output) if output else 0}")
log_comprehensive(f"Coder response received ({response.usage.prompt_tokens} input tokens, {response.usage.completion_tokens} output tokens)", f"CODER RESPONSE #{count_inner}")
log_comprehensive("Generated Code:", plain=True)
log_comprehensive(output, plain=True)
print("DEBUG: Logged coder response successfully")
except Exception as e:
log_comprehensive(f"Coder API call failed: {e}", f"❌ CODER ERROR #{count_inner}", "ERROR")
print(f"OpenAI API error: {e}")
if "429" in str(e) or "rate" in str(e).lower() or "quota" in str(e).lower():
print("Rate limit or quota exceeded. This is a temporary issue.")
print("Stopping job to avoid wasting resources on temporary failures.")
log_comprehensive("Terminating due to rate limit/quota issue", "πŸ›‘ TERMINATION", "ERROR")
sys.exit(1) # Fail for rate limits
else:
print("Permanent API error (auth, network, etc.). Stopping job.")
log_comprehensive("Terminating due to permanent API error", "πŸ›‘ TERMINATION", "ERROR")
sys.exit(1) # Fail for permanent errors
code = get_code(output)
# log_comprehensive(f"Extracted code from response:\n{code}", f"πŸ”§ CODE EXTRACTION #{count_inner}")
code_filepath = os.path.join(code_dir, f'{name}.py')
with open(code_filepath, 'w') as f:
f.write(code)
print("⚑ Executing code from coder")
# log_comprehensive("Executing generated code", f"⚑ CODE EXECUTION #{count_inner}")
# Execute code and capture output directly to comprehensive log
cmd = ['python', code_filepath]
# General-purpose argument passing from Snakemake to the generated script.
# Pass all input and output files to the generated script.
if args.input_files:
cmd.extend(['--input-files', *args.input_files])
if args.output_files:
cmd.extend(['--output-files', *args.output_files])
# Log the execution command
log_comprehensive("Executing generated code.", f"CODE EXECUTION #{count_inner}")
log_comprehensive(f"Command: {' '.join(cmd)}", level="DEBUG")
# Ensure generated code writes job-scoped artifacts
env = os.environ.copy()
if out_dir:
env['OUTPUT_DIR'] = out_dir
process = subprocess.run(
cmd,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
env=env,
)
# Log execution results to comprehensive log
log_comprehensive(f"Execution completed with exit code {process.returncode}", f"EXECUTION RESULT #{count_inner}")
if process.returncode != 0:
log_comprehensive("Execution failed. Output:", level="ERROR")
log_comprehensive(process.stdout, plain=True)
else:
log_comprehensive("Execution successful. Output:", level="INFO")
log_comprehensive(process.stdout, plain=True)
# Use the execution output directly for supervisor feedback
command_line_output = process.stdout or ""
# Verify that the output files were actually created
if args.output_files:
for output_file in args.output_files:
if not os.path.exists(output_file):
error_message = f"\nERROR: Output file was not created: {output_file}"
print(error_message)
command_line_output += error_message
log_comprehensive(f"Output file not found after execution: {output_file}", f"EXECUTION RESULT #{count_inner}", "ERROR")
elif os.path.getsize(output_file) == 0:
error_message = f"\nERROR: Output file exists but is empty: {output_file}"
print(error_message)
command_line_output += error_message
log_comprehensive(f"Output file exists but is empty: {output_file}", f"EXECUTION RESULT #{count_inner}", "ERROR")
else:
# File exists and has content - verify it was recently created/modified
import time
mtime = os.path.getmtime(output_file)
current_time = time.time()
if current_time - mtime > 60: # Modified more than 1 minute ago
error_message = f"\nWARNING: Output file exists but was not recently modified: {output_file} (modified {current_time - mtime:.1f} seconds ago)"
print(error_message)
command_line_output += error_message
log_comprehensive(f"Output file exists but not recently modified: {output_file}", f"EXECUTION RESULT #{count_inner}", "WARNING")
# construct new supervisor prompt
with open(supervisor_record, 'r') as f:
record = f.read()
before_user, after_user = supervisor_instructions.split('User prompt:')
# Rebuild file context for feedback prompts
file_context = ''
if args.input_files:
file_context += '\nInput files:\n' + '\n'.join(args.input_files)
if args.output_files:
file_context += '\nOutput files:\n' + '\n'.join(args.output_files)
new_supervisor_prompt = (
before_user + file_context + '\nUser prompt:\n' + user_prompt +
'\nGenerated code:\n' + code +
'\nCommand line output:\n' + command_line_output +
'\nRecord:\n' + record
)
# SUPERVISOR_PROMPT_FEEDBACK logging removed to avoid duplicating the built prompt
print(f"Calling supervisor {supervisor} (iteration {count_inner})")
log_comprehensive(f"Calling supervisor {supervisor} for feedback (iteration {count_inner})", f"πŸ€– SUPERVISOR FEEDBACK CALL #{count_inner}")
supervisor_response_inner = "" # Initialize to avoid unbound variable
try:
response = client.chat.completions.create(
model = supervisor,
messages = [
{
'role': 'user',
'content': new_supervisor_prompt
}
],
temperature = temperature
)
input_tokens += response.usage.prompt_tokens
output_tokens += response.usage.completion_tokens
tokens['Feedback to Supervisor'] += response.usage.prompt_tokens
supervisor_response_inner = response.choices[-1].message.content
log_comprehensive(f"Supervisor feedback received ({response.usage.prompt_tokens} input tokens, {response.usage.completion_tokens} output tokens)", f"SUPERVISOR FEEDBACK RESPONSE #{count_inner}")
log_comprehensive("Supervisor Feedback:", plain=True)
log_comprehensive(supervisor_response_inner, plain=True)
except Exception as e:
log_comprehensive(f"Supervisor feedback API call failed: {e}", f"❌ SUPERVISOR ERROR #{count_inner}", "ERROR")
print(f"OpenAI API error: {e}")
if "429" in str(e) or "rate" in str(e).lower() or "quota" in str(e).lower():
print("Rate limit or quota exceeded. This is a temporary issue.")
print("Stopping job to avoid wasting resources on temporary failures.")
log_comprehensive("Terminating due to rate limit/quota issue", "πŸ›‘ TERMINATION", "ERROR")
sys.exit(1) # Fail for rate limits
else:
print("Permanent API error (auth, network, etc.). Stopping job.")
log_comprehensive("Terminating due to permanent API error", "πŸ›‘ TERMINATION", "ERROR")
sys.exit(1) # Fail for permanent errors
# Set the response for processing
supervisor_response = supervisor_response_inner
if 'Call record:' in supervisor_response:
try:
parts = supervisor_response.split('Call record:', 1)
coder_prompt = parts[0]
record = parts[-1]
with open(supervisor_record, 'a') as f:
f.write(f'\nCall {count_inner} record: ' + record)
except Exception as e:
log_comprehensive(f"Error parsing supervisor response: {e}", f"⚠️ PARSING WARNING #{count_inner}", "WARNING")
print(f"Error parsing supervisor response: {e}")
coder_prompt = supervisor_response
else:
coder_prompt = supervisor_response
# Suppress parsing warning
# log_comprehensive("Supervisor response missing 'Call record:'. Treating entire response as coder prompt.", f"⚠️ PARSING WARNING #{count_inner}", "WARNING")
old_coder_prompt = coder_prompt
if 'Supervisor is satisfied with current results' in coder_prompt:
done_inner = True
print(f"βœ… Supervisor satisfied after {count_inner} iterations!")
log_comprehensive(f"SUCCESS: Supervisor satisfied after {count_inner} iterations!", "πŸŽ‰ SUCCESS")
else:
log_comprehensive(f"Supervisor not satisfied, continuing to iteration {count_inner + 1}", f"πŸ”„ CONTINUING TO ITERATION #{count_inner + 1}")
# Ensure coder_prompt is set for the next iteration
if 'Call record:' not in supervisor_response:
coder_prompt = supervisor_response
total_calls = 1 + 2*count_inner # Initial supervisor + coder calls + feedback supervisor calls
print(f"πŸ“Š Total API calls made: {total_calls}")
log_comprehensive(f"Total API calls made: {total_calls}", "πŸ“Š FINAL STATISTICS")
log_comprehensive(f"Final token counts: {input_tokens} input, {output_tokens} output")
# Final summary
end_time = datetime.now()
duration = end_time - start_time
log_comprehensive(f"Interaction completed in {duration}", "🏁 PROCESS COMPLETED")
log_comprehensive("=== END OF COMPREHENSIVE LOG ===", "🏁 END OF LOG")
# Add final summary to comprehensive log
with open(comprehensive_log_path, 'a') as comp_log:
comp_log.write(f"\n{'='*100}\n")
comp_log.write("πŸ“Š FINAL SUMMARY\n")
comp_log.write(f"{'='*100}\n")
comp_log.write(f"Task: {name}\n")
comp_log.write(f"Supervisor: {supervisor}\n")
comp_log.write(f"Coder: {coder}\n")
comp_log.write(f"Total Iterations: {count_inner if 'count_inner' in locals() else 0}\n")
comp_log.write(f"Start Time: {start_time}\n")
comp_log.write(f"End Time: {end_time}\n")
comp_log.write(f"Duration: {duration}\n")
comp_log.write(f"Total API Calls: {total_calls}\n")
comp_log.write(f"Total Input Tokens: {input_tokens}\n")
comp_log.write(f"Total Output Tokens: {output_tokens}\n")
comp_log.write(f"User Prompt Tokens: {tokens['User Prompt']}\n")
comp_log.write(f"Supervisor to Coder Tokens: {tokens['Supervisor to Coder']}\n")
comp_log.write(f"Coder Output Tokens: {tokens['Coder Outputs']}\n")
comp_log.write(f"Feedback to Supervisor Tokens: {tokens['Feedback to Supervisor']}\n")
final_status = 'SUCCESS' if ('done_inner' in locals() and done_inner) else 'INCOMPLETE'
comp_log.write(f"Final Status: {final_status}\n")
comp_log.write(f"{'='*100}\n")
prompt_pair = f'User Prompt: \n {user_prompt} \n Supervisor Prompt: \n {old_coder_prompt}'
prompt_pair_path = os.path.join(prompt_pair_dir, f'{name}.txt')
with open(prompt_pair_path, 'w') as f:
f.write(prompt_pair)
# Save metrics arrays under the run's logs directory
calls_filepath = os.path.join(out_dir, 'logs', 'calls.npy')
if os.path.exists(calls_filepath):
old_calls = np.load(calls_filepath)
new_calls = np.append(old_calls, total_calls)
else:
new_calls = np.array([total_calls])
np.save(calls_filepath, new_calls)
input_tokens_filepath = os.path.join(out_dir, 'logs', 'input_tokens.npy')
if os.path.exists(input_tokens_filepath):
old_tokens = np.load(input_tokens_filepath)
new_tokens = np.append(old_tokens, input_tokens)
else:
new_tokens = np.array([input_tokens])
np.save(input_tokens_filepath, new_tokens)
output_tokens_filepath = os.path.join(out_dir, 'logs', 'output_tokens.npy')
if os.path.exists(output_tokens_filepath):
old_tokens = np.load(output_tokens_filepath)
new_tokens = np.append(old_tokens, output_tokens)
else:
new_tokens = np.array([output_tokens])
np.save(output_tokens_filepath, new_tokens)
# Before final success messages, verify output files exist and are non-empty
if args.output_files:
missing = [f for f in args.output_files if not os.path.exists(f) or os.path.getsize(f) == 0]
if missing:
print(f"βœ— Missing or empty output files: {missing}")
sys.exit(1)
else:
print("βœ… All expected output files created.")
print("πŸŽ‰ Supervisor-coder interaction completed!")
print(f"πŸ“ˆ Final stats: {total_calls} total API calls, {input_tokens} input tokens, {output_tokens} output tokens")
print(f"πŸ“‹ Comprehensive log saved to: {comprehensive_log_path}")