File size: 13,375 Bytes
ffd85e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
"""
Tool-using agent for WhipStudio.

This example demonstrates how to use WhipStudio's debugging tools
to iteratively analyze and fix bugs before submitting a final solution.

The agent uses:
1. execute_snippet - To test hypotheses about the code
2. inspect_tensor - To check tensor shapes, dtypes, and gradients
3. get_variable_state - To evaluate multiple expressions
4. run_training_probe - To test potential fixes
5. inspect_diff - To review changes before submission
6. submit_fix - Final submission

Usage:
    python examples/tool_agent.py --env-url http://localhost:7860 --task task1
    python examples/tool_agent.py --env-url https://your-space.hf.space --task task6
"""

import argparse
import json
import os
import re
import httpx
from openai import OpenAI


SYSTEM_PROMPT = """You are an expert PyTorch debugging agent that fixes buggy ML code.
You have debugging tools available, but your PRIMARY GOAL is to SUBMIT A FIX.

AVAILABLE TOOLS:
1. execute_snippet - Run Python code to test hypotheses
2. inspect_tensor - Check tensor shape, dtype, gradients, NaN/Inf
3. get_variable_state - Inspect multiple variable values  
4. run_training_probe - Run training steps to see loss curve
5. inspect_diff - Preview your changes before submitting
6. submit_fix - Submit your final fix (ALWAYS DO THIS BEFORE RUNNING OUT OF TURNS)

RESPONSE FORMAT - ALWAYS respond with valid JSON only:
{
  "reasoning": "Brief analysis",
  "action_type": "tool_name",
  "action_params": { ... }
}

For submit_fix, the fixed_code must be COMPLETE working Python code:
{
  "reasoning": "Fix explanation",
  "action_type": "submit_fix",
  "action_params": {
    "fixed_code": "import torch\\nimport torch.nn as nn\\n..."
  }
}

CRITICAL RULES:
1. You MUST call submit_fix before your turns run out
2. If you have 2 or fewer turns remaining, IMMEDIATELY submit your fix
3. Don't waste turns - analyze, test once if needed, then SUBMIT
4. Fixed code must print: LOSSES:[v1, v2, ...] or similar metrics
5. Keep torch.manual_seed() calls intact for reproducibility
6. Use \\n for newlines in code strings

EFFICIENT DEBUGGING:
- Turn 1-2: Analyze bug, maybe one quick test
- Turn 3+: SUBMIT YOUR FIX - don't keep testing!""".strip()


def get_client():
    """Initialize OpenAI-compatible client."""
    api_base = os.environ.get("API_BASE_URL", "https://router.huggingface.co/v1")
    api_key = os.environ.get("HF_TOKEN") or os.environ.get("OPENAI_API_KEY")
    
    if not api_key:
        raise ValueError("Set HF_TOKEN or OPENAI_API_KEY environment variable")
    
    return OpenAI(base_url=api_base, api_key=api_key)


def parse_agent_response(response: str) -> dict:
    """Parse JSON response from agent, handling common formatting issues."""
    response = response.strip()
    
    # Remove markdown code blocks
    if response.startswith("```json"):
        response = response[7:]
    elif response.startswith("```"):
        response = response[3:]
    if response.endswith("```"):
        response = response[:-3]
    response = response.strip()
    
    try:
        return json.loads(response)
    except json.JSONDecodeError:
        # Try to extract JSON from response
        match = re.search(r'\{[\s\S]*\}', response)
        if match:
            try:
                return json.loads(match.group())
            except json.JSONDecodeError:
                pass
    
    # Fallback
    return {
        "reasoning": "Fallback: could not parse response",
        "action_type": "submit_fix",
        "action_params": {"fixed_code": response}
    }


def format_tool_result(obs: dict, action_type: str) -> str:
    """Format tool result for display."""
    turn = obs.get("turn", 0)
    error = obs.get("error")
    
    if error:
        return f"[Turn {turn}] {action_type} ERROR: {error}"
    
    if action_type == "execute_snippet":
        stdout = obs.get("stdout", "")[:2000]  # Increased from 500
        stderr = obs.get("stderr", "")[:1000]  # Increased from 200
        exit_code = obs.get("exit_code", 0)
        result = f"[Turn {turn}] execute_snippet (exit={exit_code}):\n{stdout}"
        if stderr:
            result += f"\nSTDERR:\n{stderr}"
        return result
    
    elif action_type == "inspect_tensor":
        return f"""[Turn {turn}] inspect_tensor:
  shape: {obs.get('shape')}
  dtype: {obs.get('dtype')}
  requires_grad: {obs.get('requires_grad')}
  grad_is_none: {obs.get('grad_is_none')}
  min/max/mean: {obs.get('min_val')}/{obs.get('max_val')}/{obs.get('mean_val')}
  is_nan: {obs.get('is_nan')}, is_inf: {obs.get('is_inf')}"""
    
    elif action_type == "run_training_probe":
        losses = obs.get("losses", [])[:10]
        final_loss = obs.get("final_loss")
        grad_norms = obs.get("grad_norms", {})
        return f"[Turn {turn}] run_training_probe:\n  losses: {losses}\n  final_loss: {final_loss}\n  grad_norms: {grad_norms}"
    
    elif action_type == "get_variable_state":
        results = obs.get("results", {})
        lines = [f"[Turn {turn}] get_variable_state:"]
        for expr, res in results.items():
            if res.get("error"):
                lines.append(f"  {expr}: ERROR - {res['error']}")
            else:
                val = res.get("repr", str(res.get("value", "?")))[:200]  # Increased from 80
                lines.append(f"  {expr}: {val}")
        return "\n".join(lines)
    
    elif action_type == "inspect_diff":
        lines_changed = obs.get("lines_changed", 0)
        additions = obs.get("additions", 0)
        deletions = obs.get("deletions", 0)
        diff = obs.get("diff", "")[:500]
        return f"[Turn {turn}] inspect_diff: {lines_changed} lines changed (+{additions}/-{deletions})\n{diff}"
    
    elif action_type == "submit_fix":
        reward = obs.get("reward", 0.0)
        return f"[Turn {turn}] submit_fix: reward={reward}"
    
    return f"[Turn {turn}] {action_type}: {json.dumps(obs, default=str)[:500]}"


def run_tool_agent(env_url: str, task_id: str, client, max_turns: int = 10) -> float:
    """Run a tool-using agent on a single task."""
    print(f"\n{'='*60}")
    print(f"Tool Agent: {task_id}")
    print(f"{'='*60}")
    
    model = os.environ.get("MODEL_NAME", "Qwen/Qwen2.5-Coder-3B-Instruct")
    
    # Reset environment
    with httpx.Client(timeout=60.0) as http_client:
        resp = http_client.post(f"{env_url}/reset", json={"task_id": task_id})
        resp.raise_for_status()
        obs = resp.json().get("observation", resp.json())
    
    buggy_code = obs.get("buggy_code", "")
    task_description = obs.get("task_description", "")
    episode_id = obs.get("episode_id", "")  # Track episode_id for session persistence
    
    print(f"Task: {task_description[:100]}...")
    print(f"Episode ID: {episode_id[:16]}..." if episode_id else "No episode_id")
    
    tool_history = []
    best_reward = 0.0
    
    for turn in range(1, max_turns + 1):
        print(f"\n--- Turn {turn}/{max_turns} ---")
        
        turns_remaining = max_turns - turn
        
        # Build context
        history_text = "\n".join(tool_history[-5:]) if tool_history else "No previous tool calls."
        
        # Force submission on last turn
        if turns_remaining == 0:
            urgency = "\n⚠️ THIS IS YOUR LAST TURN! You MUST call submit_fix NOW with your best fix."
        elif turns_remaining <= 2:
            urgency = f"\n⚠️ ONLY {turns_remaining} TURN(S) LEFT! Submit your fix soon!"
        else:
            urgency = ""
        
        user_prompt = f"""Task: {task_description}

Buggy Code:
```python
{buggy_code}
```

Turn {turn}/{max_turns}
Best reward so far: {best_reward}
Turns remaining: {turns_remaining}{urgency}

Tool History:
{history_text}

Analyze the code and submit your fix. Don't waste turns on unnecessary testing.""".strip()
        
        # Get LLM response
        try:
            response = client.chat.completions.create(
                model=model,
                messages=[
                    {"role": "system", "content": SYSTEM_PROMPT},
                    {"role": "user", "content": user_prompt},
                ],
                max_tokens=4096,
                temperature=0.2,
            )
            response_text = response.choices[0].message.content.strip()
            parsed = parse_agent_response(response_text)
        except Exception as e:
            print(f"LLM Error: {e}")
            tool_history.append(f"[Turn {turn}] LLM ERROR: {e}")
            continue
        
        action_type = parsed.get("action_type", "submit_fix")
        action_params = parsed.get("action_params", {})
        reasoning = parsed.get("reasoning", "")[:100]
        
        # Force submit_fix on last turn if agent didn't choose it
        if turns_remaining == 0 and action_type != "submit_fix":
            print(f"[OVERRIDE] Last turn - forcing submit_fix instead of {action_type}")
            action_type = "submit_fix"
            # Use fixed_code from action_params if available, otherwise use any code param
            fixed_code = action_params.get("fixed_code") or action_params.get("code") or action_params.get("proposed_code") or buggy_code
            action_params = {"fixed_code": fixed_code}
        
        print(f"Action: {action_type}")
        print(f"Reasoning: {reasoning}...")
        
        # Build action payload - ALWAYS include episode_id for session tracking
        action = {
            "action_type": action_type,
            "episode_id": episode_id,  # Critical for session persistence in HTTP mode
        }
        
        if action_type == "execute_snippet":
            action["code"] = action_params.get("code", "print('test')")
        elif action_type == "inspect_tensor":
            action["setup_code"] = action_params.get("setup_code", "")[:8000]
            action["target_expression"] = action_params.get("target_expression", "")
        elif action_type == "run_training_probe":
            action["code"] = action_params.get("code", buggy_code)[:8000]
            action["steps"] = min(action_params.get("steps", 5), 10)
        elif action_type == "get_variable_state":
            action["setup_code"] = action_params.get("setup_code", "")[:8000]
            action["expressions"] = action_params.get("expressions", [])[:10]
        elif action_type == "inspect_diff":
            action["proposed_code"] = action_params.get("proposed_code", "")
        elif action_type == "submit_fix":
            fixed_code = action_params.get("fixed_code", "")
            # Clean markdown
            if "```python" in fixed_code:
                fixed_code = fixed_code.split("```python", 1)[1].split("```", 1)[0].strip()
            elif "```" in fixed_code:
                fixed_code = fixed_code.split("```", 1)[1].split("```", 1)[0].strip()
            action["fixed_code"] = fixed_code
        
        # Execute action
        try:
            with httpx.Client(timeout=120.0) as http_client:
                resp = http_client.post(f"{env_url}/step", json={"action": action})
                resp.raise_for_status()
                result = resp.json()
        except Exception as e:
            print(f"API Error: {e}")
            tool_history.append(f"[Turn {turn}] API ERROR: {e}")
            continue
        
        obs = result.get("observation", {})
        reward = float(result.get("reward", 0.0) or 0.0)
        done = result.get("done", False) or obs.get("episode_done", False)
        
        # Format and store result
        tool_result = format_tool_result(obs, action_type)
        tool_history.append(tool_result)
        print(tool_result)
        
        if reward > best_reward:
            best_reward = reward
        
        if action_type == "submit_fix":
            print(f"Reward: {reward:.4f}")
            if reward >= 0.95 or done:
                break
        
        if done:
            break
    
    print(f"\nFinal reward for {task_id}: {best_reward:.4f}")
    return best_reward


def main():
    parser = argparse.ArgumentParser(description="Tool-using WhipStudio Agent")
    parser.add_argument("--env-url", default="http://localhost:7860", help="Environment URL")
    parser.add_argument("--task", default="task1", help="Task ID to run")
    parser.add_argument("--all-tasks", action="store_true", help="Run all tasks")
    parser.add_argument("--max-turns", type=int, default=10, help="Max turns per task")
    args = parser.parse_args()
    
    client = get_client()
    
    tasks = ["task1", "task2", "task3", "task4", "task5", "task6"] if args.all_tasks else [args.task]
    
    results = {}
    for task_id in tasks:
        try:
            score = run_tool_agent(args.env_url, task_id, client, args.max_turns)
            results[task_id] = score
        except Exception as e:
            print(f"Error on {task_id}: {e}")
            results[task_id] = 0.0
    
    if len(results) > 1:
        print("\n" + "="*60)
        print("FINAL RESULTS")
        print("="*60)
        
        total = 0.0
        for task_id, score in results.items():
            emoji = "✅" if score >= 0.7 else ("📈" if score >= 0.3 else "❌")
            print(f"{emoji} {task_id}: {score:.4f}")
            total += score
        
        avg = total / len(results) if results else 0.0
        print(f"\nAverage: {avg:.4f}")


if __name__ == "__main__":
    main()