axnwork-cli 0.2.0__py3-none-any.whl

axon/llm.py

@@ -0,0 +1,214 @@
+"""LLM integration — call Anthropic/OpenAI/Ollama via litellm."""
+import re
+import os
+
+from axon.config import load_config
+
+
+def build_prompt(task, my_best_answer, my_best_score, platform_best_score, last_feedback=None, community_subs=None, my_past_subs=None):
+    direction_text = "higher is better" if task.get("direction") == "maximize" else "lower is better"
+    is_code = task.get("eval_type") == "code_output"
+
+    prompt = f"# Task: {task['title']}\n\n## Description\n{task['description']}\n\n"
+    prompt += f"## Evaluation\nScored using: {task['eval_type']}\nDirection: {task['direction']} ({direction_text})\n"
+    prompt += f"Completion threshold: {task['completion_threshold']}\n"
+    prompt += f"Platform best: {platform_best_score if platform_best_score is not None else 'None (no submissions yet)'}\n"
+
+    if community_subs:
+        prompt += "\n## Community Submissions (top answers from other miners)\n"
+        for i, sub in enumerate(community_subs[:3], 1):
+            score = sub.get("score")
+            score_str = f"{score:.4f}" if score is not None else "N/A"
+            prompt += f"\n### Submission #{i} (score: {score_str})\n"
+            if sub.get("answer"):
+                answer_preview = sub["answer"][:2000]
+                prompt += f"```\n{answer_preview}\n```\n"
+
+    if my_past_subs:
+        recent_subs = my_past_subs[-10:]
+        prompt += "\n## Your Past Submissions (DO NOT repeat these answers)\n"
+        for i, sub in enumerate(recent_subs, 1):
+            score = sub.get("score")
+            eval_status = sub.get("eval_status", "unknown")
+            error = sub.get("eval_error")
+            if error:
+                prompt += f"  #{i}  score=error   status={eval_status}  error={error[:120]}\n"
+            elif score is not None:
+                prompt += f"  #{i}  score={score:.4f}  status={eval_status}\n"
+            else:
+                prompt += f"  #{i}  score=N/A     status={eval_status}\n"
+            answer_text = sub.get("answer")
+            if answer_text:
+                preview = answer_text[:500]
+                if len(answer_text) > 500:
+                    preview += "... (truncated)"
+                prompt += f"    Answer: {preview}\n"
+        prompt += "\nAvoid submitting answers similar to the ones above. Try a different approach.\n"
+
+    if my_best_answer is not None and my_best_score is not None:
+        prompt += f"\n## Your Current Best\nScore: {my_best_score}\nAnswer:\n{my_best_answer}\n"
+
+    if last_feedback:
+        prompt += "\n## Last Round Feedback\n"
+        if last_feedback.get("error"):
+            prompt += f"Status: ERROR\nError: {last_feedback['error']}\n"
+            if last_feedback.get("details", {}).get("stderr"):
+                prompt += f"Stderr:\n{str(last_feedback['details']['stderr'])[:500]}\n"
+            prompt += f"\nYour submission that caused this error:\n{last_feedback.get('answer', '')}\n"
+            prompt += "\nFix the error and try again.\n"
+        else:
+            status = "improved (but not yet completed)" if last_feedback.get("improved") else "no improvement"
+            prompt += f"Score: {last_feedback.get('score')}\nStatus: {status}\n"
+            if last_feedback.get("details", {}).get("stdout"):
+                prompt += f"Eval output:\n{str(last_feedback['details']['stdout'])[:300]}\n"
+            prompt += "\nAnalyze the eval output and your score. Find ways to improve.\n"
+    elif not my_best_answer:
+        prompt += "\nThis is your first attempt. Think carefully.\n"
+
+    if is_code:
+        prompt += (
+            "\n## CRITICAL RULES\n"
+            "1. Do NOT hardcode test data. Input comes via function parameters.\n"
+            "2. Functions must work with ANY input.\n"
+            "3. The evaluator writes your submission directly to solution.py and executes it.\n"
+            "4. Inside <answer>, include ONLY raw executable Python code.\n"
+            "5. Do NOT include XML tags, prose, or markdown fences inside the submitted code.\n\n"
+        )
+
+    prompt += f"## OUTPUT FORMAT\n<thinking>your reasoning</thinking>\n<answer>{'RAW EXECUTABLE PYTHON CODE ONLY' if is_code else 'your answer'}</answer>\n"
+    return prompt
+
+
+def build_agent_prompt(task, my_best_answer, my_best_score, platform_best_score, last_feedback=None, community_subs=None, my_past_subs=None):
+    """Build prompt for CLI agent backends (claude-cli, codex-cli).
+
+    Unlike build_prompt(), this omits XML output format instructions (CLI backends
+    use --json-schema or embedded JSON format) and adds tool usage guidance.
+    """
+    direction_text = "higher is better" if task.get("direction") == "maximize" else "lower is better"
+    is_code = task.get("eval_type") == "code_output"
+    eval_type = task.get("eval_type", "")
+
+    prompt = f"# Task: {task['title']}\n\n## Description\n{task['description']}\n\n"
+    prompt += f"## Evaluation\nScored using: {eval_type}\nDirection: {task['direction']} ({direction_text})\n"
+    prompt += f"Completion threshold: {task['completion_threshold']}\n"
+    prompt += f"Platform best: {platform_best_score if platform_best_score is not None else 'None (no submissions yet)'}\n"
+
+    if community_subs:
+        prompt += "\n## Community Submissions (top answers from other miners)\n"
+        for i, sub in enumerate(community_subs[:3], 1):
+            score = sub.get("score")
+            score_str = f"{score:.4f}" if score is not None else "N/A"
+            prompt += f"\n### Submission #{i} (score: {score_str})\n"
+            if sub.get("answer"):
+                answer_preview = sub["answer"][:2000]
+                prompt += f"```\n{answer_preview}\n```\n"
+
+    if my_past_subs:
+        recent_subs = my_past_subs[-10:]
+        prompt += "\n## Your Past Submissions (DO NOT repeat these answers)\n"
+        for i, sub in enumerate(recent_subs, 1):
+            score = sub.get("score")
+            eval_status = sub.get("eval_status", "unknown")
+            error = sub.get("eval_error")
+            if error:
+                prompt += f"  #{i}  score=error   status={eval_status}  error={error[:120]}\n"
+            elif score is not None:
+                prompt += f"  #{i}  score={score:.4f}  status={eval_status}\n"
+            else:
+                prompt += f"  #{i}  score=N/A     status={eval_status}\n"
+            answer_text = sub.get("answer")
+            if answer_text:
+                preview = answer_text[:500]
+                if len(answer_text) > 500:
+                    preview += "... (truncated)"
+                prompt += f"    Answer: {preview}\n"
+        prompt += "\nAvoid submitting answers similar to the ones above. Try a different approach.\n"
+
+    if my_best_answer is not None and my_best_score is not None:
+        prompt += f"\n## Your Current Best\nScore: {my_best_score}\nAnswer:\n{my_best_answer}\n"
+
+    if last_feedback:
+        prompt += "\n## Last Round Feedback\n"
+        if last_feedback.get("error"):
+            prompt += f"Status: ERROR\nError: {last_feedback['error']}\n"
+            if last_feedback.get("details", {}).get("stderr"):
+                prompt += f"Stderr:\n{str(last_feedback['details']['stderr'])[:500]}\n"
+            prompt += f"\nYour submission that caused this error:\n{last_feedback.get('answer', '')}\n"
+            prompt += "\nFix the error and try again.\n"
+        else:
+            status = "improved (but not yet completed)" if last_feedback.get("improved") else "no improvement"
+            prompt += f"Score: {last_feedback.get('score')}\nStatus: {status}\n"
+            if last_feedback.get("details", {}).get("stdout"):
+                prompt += f"Eval output:\n{str(last_feedback['details']['stdout'])[:300]}\n"
+            prompt += "\nAnalyze the eval output and your score. Find ways to improve.\n"
+    elif not my_best_answer:
+        prompt += "\nThis is your first attempt. Think carefully.\n"
+
+    # Tool usage guidance and evaluator-aligned submission rules.
+    if is_code:
+        prompt += (
+            "\n## CRITICAL RULES\n"
+            "1. Do NOT hardcode test data. Input comes via function parameters.\n"
+            "2. Functions must work with ANY input.\n"
+            "3. Use Bash to test your code before submitting your final answer.\n"
+            "4. Your final answer must be raw executable Python code only.\n"
+            "5. Do NOT wrap it in XML tags, markdown fences, or prose.\n"
+            "6. The evaluator writes your submission directly to solution.py and executes it.\n\n"
+        )
+    elif eval_type == "llm_judge":
+        prompt += "\n## APPROACH\nUse WebSearch and WebFetch to research the topic. Verify facts before answering.\n\n"
+
+    return prompt
+
+
+def _parse_response(text):
+    think_match = re.search(r"<thinking>(.*?)</thinking>", text, re.DOTALL)
+    answer_match = re.search(r"<answer>(.*?)</answer>", text, re.DOTALL)
+    thinking = think_match.group(1).strip() if think_match else text
+    answer = answer_match.group(1).strip() if answer_match else text.strip()
+    # Strip code fences
+    answer = re.sub(r"```[\w]*\s*\n?", "", answer)
+    answer = re.sub(r"\n?\s*```", "", answer)
+    return thinking, answer.strip()
+
+
+def call_llm(prompt, model, api_base=""):
+    """Call LLM via litellm. Returns (thinking, answer, usage)."""
+    # Set API keys from config
+    config = load_config()
+    keys = config.get("api_keys", {})
+    for provider, env_var in [("anthropic", "ANTHROPIC_API_KEY"), ("openai", "OPENAI_API_KEY"), ("deepseek", "DEEPSEEK_API_KEY")]:
+        if keys.get(provider) and not os.environ.get(env_var):
+            os.environ[env_var] = keys[provider]
+
+    # Guard: litellm's proxy_cli.py calls os.getcwd() at import time,
+    # which crashes if the CWD no longer exists (e.g. deleted temp dir).
+    try:
+        os.getcwd()
+    except (FileNotFoundError, OSError):
+        os.chdir(os.path.expanduser("~"))
+
+    import litellm
+    litellm.suppress_debug_info = True
+    from litellm import completion, completion_cost
+
+    kwargs = {"model": model, "messages": [{"role": "user", "content": prompt}], "max_tokens": 4096}
+    if api_base:
+        kwargs["api_base"] = api_base
+
+    response = completion(**kwargs)
+    text = response.choices[0].message.content
+
+    usage = {}
+    if hasattr(response, "usage") and response.usage:
+        usage["prompt_tokens"] = getattr(response.usage, "prompt_tokens", 0) or 0
+        usage["completion_tokens"] = getattr(response.usage, "completion_tokens", 0) or 0
+        usage["total_tokens"] = usage["prompt_tokens"] + usage["completion_tokens"]
+    try:
+        usage["cost"] = completion_cost(completion_response=response) or 0.0
+    except Exception:
+        usage["cost"] = 0.0
+
+    thinking, answer = _parse_response(text)
+    return thinking, answer, usage

axon/log.py

@@ -0,0 +1,44 @@
+"""Centralized logging — all errors auto-saved to ~/.axon/logs/."""
+import logging
+import sys
+
+from axon.config import AXON_HOME
+
+LOG_DIR = AXON_HOME / "logs"
+LOG_FILE = LOG_DIR / "axon.log"
+
+
+def setup_logging():
+    """Configure logging: errors go to file, always. Call once at startup."""
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+
+    file_handler = logging.FileHandler(LOG_FILE, encoding="utf-8")
+    file_handler.setLevel(logging.DEBUG)
+    file_handler.setFormatter(logging.Formatter(
+        "%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+    ))
+
+    root = logging.getLogger()
+    root.setLevel(logging.DEBUG)
+    root.addHandler(file_handler)
+
+    # Suppress noisy third-party loggers from leaking to stdout
+    for name in ("LiteLLM", "litellm", "httpx", "httpcore"):
+        logging.getLogger(name).setLevel(logging.WARNING)
+
+    # Also catch unhandled exceptions
+    _original_excepthook = sys.excepthook
+
+    def _excepthook(exc_type, exc_value, exc_tb):
+        if not issubclass(exc_type, (KeyboardInterrupt, SystemExit)):
+            logging.getLogger("axon").critical(
+                "Unhandled exception", exc_info=(exc_type, exc_value, exc_tb),
+            )
+        _original_excepthook(exc_type, exc_value, exc_tb)
+
+    sys.excepthook = _excepthook
+
+
+def get_logger(name: str = "axon") -> logging.Logger:
+    return logging.getLogger(name)