78292e836f
FIXES: 1. Truncation was too aggressive (2KB threshold, truncate to 2000 chars) 2. Important data was being lost in medium-sized outputs (2-10KB) Changes: - Raise pass-through threshold from 2KB to 5KB - Medium outputs (5-20KB) now use hierarchical extraction instead of truncation - _extract_key_findings already handles chunking automatically - Better preservation of important data like service lists Benefits: - Full service lists will now be properly analyzed - No more missing services due to truncation - Macha can see the complete picture before responding
1012 lines
39 KiB
Python
1012 lines
39 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
AI Agent - Analyzes system state and proposes solutions using local LLMs
|
|
"""
|
|
|
|
import json
|
|
import requests
|
|
import subprocess
|
|
from typing import Dict, List, Any, Optional
|
|
from pathlib import Path
|
|
from datetime import datetime
|
|
|
|
from tools import SysadminTools
|
|
|
|
|
|
class MachaAgent:
|
|
"""AI agent that analyzes system issues and proposes fixes"""
|
|
|
|
# Load system prompt from file
|
|
@staticmethod
|
|
def _load_system_prompt() -> str:
|
|
"""Load the system prompt from file"""
|
|
prompt_file = Path(__file__).parent / "system_prompt.txt"
|
|
try:
|
|
return prompt_file.read_text()
|
|
except Exception as e:
|
|
print(f"Warning: Could not load system prompt from {prompt_file}: {e}")
|
|
return "You are Macha, an autonomous AI system maintenance agent."
|
|
|
|
SYSTEM_PROMPT = _load_system_prompt.__func__()
|
|
|
|
def __init__(
|
|
self,
|
|
ollama_host: str = "http://localhost:11434",
|
|
model: str = "gpt-oss:latest",
|
|
state_dir: Path = Path("/var/lib/macha"),
|
|
context_db = None,
|
|
config_repo: str = "git+https://git.coven.systems/lily/nixos-servers",
|
|
config_branch: str = "main",
|
|
enable_tools: bool = True,
|
|
use_queue: bool = True,
|
|
priority: str = "INTERACTIVE"
|
|
):
|
|
self.ollama_host = ollama_host
|
|
self.model = model
|
|
self.state_dir = state_dir
|
|
self.state_dir.mkdir(parents=True, exist_ok=True)
|
|
self.decision_log = self.state_dir / "decisions.jsonl"
|
|
self.context_db = context_db
|
|
self.config_repo = config_repo
|
|
self.config_branch = config_branch
|
|
self.enable_tools = enable_tools
|
|
|
|
# Queue settings
|
|
self.use_queue = use_queue
|
|
self.priority = priority
|
|
self.ollama_queue = None
|
|
|
|
if use_queue:
|
|
try:
|
|
from ollama_queue import OllamaQueue, Priority
|
|
self.ollama_queue = OllamaQueue()
|
|
self.priority_level = getattr(Priority, priority, Priority.INTERACTIVE)
|
|
except (PermissionError, OSError):
|
|
# Silently fall back to direct API calls when queue is not accessible
|
|
# (e.g., regular users don't have access to /var/lib/macha/queues)
|
|
self.use_queue = False
|
|
except Exception as e:
|
|
# Log unexpected errors but still fall back gracefully
|
|
import sys
|
|
print(f"Note: Ollama queue unavailable ({type(e).__name__}), using direct API", file=sys.stderr)
|
|
self.use_queue = False
|
|
|
|
# Initialize tools system
|
|
self.tools = SysadminTools(safe_mode=False) if enable_tools else None
|
|
|
|
# Tool output cache for hierarchical processing
|
|
self.tool_output_cache = {}
|
|
self.cache_dir = self.state_dir / "tool_cache"
|
|
|
|
# Only create cache dir if we have write access (running as macha user)
|
|
try:
|
|
self.cache_dir.mkdir(parents=True, exist_ok=True)
|
|
except (PermissionError, OSError):
|
|
# Running as unprivileged user (macha-chat), use temp dir instead
|
|
import tempfile
|
|
self.cache_dir = Path(tempfile.mkdtemp(prefix="macha_cache_"))
|
|
|
|
def _query_relevant_knowledge(self, query: str, limit: int = 3) -> str:
|
|
"""
|
|
Query knowledge base for relevant information
|
|
|
|
Returns formatted string of relevant knowledge to include in prompts
|
|
"""
|
|
if not self.context_db:
|
|
return ""
|
|
|
|
try:
|
|
knowledge_items = self.context_db.query_knowledge(query, limit=limit)
|
|
if not knowledge_items:
|
|
return ""
|
|
|
|
knowledge_text = "\n\nRELEVANT KNOWLEDGE FROM EXPERIENCE:\n"
|
|
for item in knowledge_items:
|
|
knowledge_text += f"\n• {item['topic']} ({item['category']}):\n"
|
|
knowledge_text += f" {item['knowledge']}\n"
|
|
knowledge_text += f" [Confidence: {item['confidence']}, Referenced: {item['times_referenced']} times]\n"
|
|
|
|
return knowledge_text
|
|
except Exception as e:
|
|
print(f"Error querying knowledge: {e}")
|
|
return ""
|
|
|
|
def store_learning(
|
|
self,
|
|
topic: str,
|
|
knowledge: str,
|
|
category: str = "experience",
|
|
confidence: str = "medium",
|
|
tags: list = None
|
|
) -> bool:
|
|
"""
|
|
Store a learned insight into the knowledge base
|
|
|
|
Args:
|
|
topic: What this is about
|
|
knowledge: The insight/pattern/learning
|
|
category: Type of knowledge
|
|
confidence: How confident we are
|
|
tags: Optional tags
|
|
|
|
Returns:
|
|
True if stored successfully
|
|
"""
|
|
if not self.context_db:
|
|
return False
|
|
|
|
try:
|
|
kid = self.context_db.store_knowledge(
|
|
topic=topic,
|
|
knowledge=knowledge,
|
|
category=category,
|
|
source="experience",
|
|
confidence=confidence,
|
|
tags=tags
|
|
)
|
|
if kid:
|
|
print(f"📚 Learned: {topic}")
|
|
return True
|
|
return False
|
|
except Exception as e:
|
|
print(f"Error storing learning: {e}")
|
|
return False
|
|
|
|
def reflect_and_learn(
|
|
self,
|
|
situation: str,
|
|
action_taken: str,
|
|
outcome: str,
|
|
success: bool
|
|
) -> None:
|
|
"""
|
|
Reflect on an operation and extract learnings to store
|
|
|
|
Args:
|
|
situation: What was the problem/situation
|
|
action_taken: What action was taken
|
|
outcome: What was the result
|
|
success: Whether it succeeded
|
|
"""
|
|
if not self.context_db:
|
|
return
|
|
|
|
# Only learn from successful operations for now
|
|
if not success:
|
|
return
|
|
|
|
# Build reflection prompt
|
|
prompt = f"""Based on this successful operation, extract key learnings to remember for the future.
|
|
|
|
SITUATION:
|
|
{situation}
|
|
|
|
ACTION TAKEN:
|
|
{action_taken}
|
|
|
|
OUTCOME:
|
|
{outcome}
|
|
|
|
Extract 1-2 specific, actionable learnings. For each learning provide:
|
|
1. topic: A concise topic name (e.g., "systemd service restart", "disk cleanup procedure")
|
|
2. knowledge: The specific insight or pattern (what worked, why, important details)
|
|
3. category: One of: command, pattern, troubleshooting, performance
|
|
|
|
Respond ONLY with valid JSON:
|
|
[
|
|
{{
|
|
"topic": "...",
|
|
"knowledge": "...",
|
|
"category": "...",
|
|
"confidence": "medium"
|
|
}}
|
|
]
|
|
"""
|
|
|
|
try:
|
|
response = self._query_ollama(prompt, temperature=0.3, timeout=30)
|
|
learnings = json.loads(response)
|
|
|
|
if isinstance(learnings, list):
|
|
for learning in learnings:
|
|
if all(k in learning for k in ['topic', 'knowledge', 'category']):
|
|
self.store_learning(
|
|
topic=learning['topic'],
|
|
knowledge=learning['knowledge'],
|
|
category=learning.get('category', 'experience'),
|
|
confidence=learning.get('confidence', 'medium')
|
|
)
|
|
except Exception as e:
|
|
# Reflection is optional - don't fail if it doesn't work
|
|
print(f"Note: Could not extract learnings: {e}")
|
|
|
|
def analyze_system_state(self, monitoring_data: Dict[str, Any], system_hostname: str = None, git_context=None) -> Dict[str, Any]:
|
|
"""Analyze system monitoring data and determine if action is needed"""
|
|
|
|
# Build context for the AI
|
|
context = self._build_analysis_context(monitoring_data)
|
|
|
|
# Get system infrastructure context if available
|
|
system_context = ""
|
|
if self.context_db and system_hostname:
|
|
system_context = self.context_db.get_system_context(system_hostname, git_context)
|
|
|
|
# Ask the AI to analyze
|
|
prompt = self._create_analysis_prompt(context, system_context)
|
|
|
|
response = self._query_ollama(prompt)
|
|
|
|
# Parse the AI's response
|
|
analysis = self._parse_analysis_response(response)
|
|
|
|
# Log the decision
|
|
self._log_decision(monitoring_data, analysis)
|
|
|
|
return analysis
|
|
|
|
def propose_fix(self, issue_description: str, context: Dict[str, Any]) -> Dict[str, Any]:
|
|
"""Propose a fix for a specific issue"""
|
|
|
|
# Query relevant config files if we have context_db
|
|
relevant_configs = []
|
|
if self.context_db:
|
|
try:
|
|
# Query for config files relevant to the issue
|
|
configs = self.context_db.query_config_files(
|
|
query=issue_description,
|
|
n_results=3
|
|
)
|
|
relevant_configs = configs
|
|
except Exception as e:
|
|
print(f"Warning: Could not query config files: {e}")
|
|
|
|
# Build config context section
|
|
config_context = ""
|
|
if relevant_configs:
|
|
config_context = "\n\nRELEVANT CONFIGURATION FILES:\n"
|
|
for cfg in relevant_configs:
|
|
config_context += f"\n--- {cfg['path']} (relevance: {cfg['relevance']:.2%}) ---\n"
|
|
config_context += cfg['content'][:1000] # First 1000 chars to avoid token limits
|
|
if len(cfg['content']) > 1000:
|
|
config_context += "\n... (truncated)"
|
|
config_context += "\n"
|
|
|
|
# Query relevant knowledge from experience
|
|
knowledge_context = self._query_relevant_knowledge(issue_description, limit=3)
|
|
|
|
# Build previous investigations context
|
|
previous_inv_context = ""
|
|
if context.get('previous_investigations'):
|
|
previous_inv_context = "\n\nPREVIOUS INVESTIGATIONS (DO NOT REPEAT THESE):\n"
|
|
for i, inv in enumerate(context['previous_investigations'][:3], 1): # Show up to 3
|
|
previous_inv_context += f"\nInvestigation #{i} ({inv['timestamp']}):\n"
|
|
previous_inv_context += f"Commands: {', '.join(inv['commands'])}\n"
|
|
previous_inv_context += f"Results:\n{inv['output'][:500]}...\n" # First 500 chars
|
|
previous_inv_context += "\n⚠️ You have already run these investigations. Do NOT propose them again."
|
|
previous_inv_context += "\n⚠️ Based on the investigation results above, propose an ACTUAL FIX, not more investigation.\n"
|
|
|
|
prompt = f"""{self.SYSTEM_PROMPT}
|
|
|
|
TASK: PROPOSE FIX
|
|
================================================================================
|
|
|
|
ISSUE TO ADDRESS:
|
|
{issue_description}
|
|
|
|
SYSTEM CONTEXT:
|
|
{json.dumps(context, indent=2)}{config_context}{knowledge_context}{previous_inv_context}
|
|
|
|
REPOSITORY INFO:
|
|
- Git Repository: {self.config_repo}
|
|
- Branch: {self.config_branch}
|
|
|
|
YOUR RESPONSE MUST BE VALID JSON:
|
|
{{
|
|
"diagnosis": "brief description of what you think is wrong",
|
|
"proposed_action": "specific action to take",
|
|
"action_type": "one of: systemd_restart, nix_rebuild, config_change, cleanup, investigation",
|
|
"risk_level": "one of: low, medium, high",
|
|
"commands": ["list", "of", "shell", "commands"],
|
|
"config_changes": {{
|
|
"file": "path/to/config.nix in the repository",
|
|
"change": "description of change needed"
|
|
}},
|
|
"reasoning": "why this fix should work",
|
|
"rollback_plan": "how to undo if it doesn't work"
|
|
}}
|
|
|
|
RESPOND WITH ONLY THE JSON, NO OTHER TEXT.
|
|
"""
|
|
|
|
response = self._query_ollama(prompt)
|
|
|
|
try:
|
|
# Try to extract JSON from response
|
|
# LLMs sometimes add extra text, so we need to find the JSON part
|
|
import re
|
|
json_match = re.search(r'\{.*\}', response, re.DOTALL)
|
|
if json_match:
|
|
return json.loads(json_match.group())
|
|
else:
|
|
return {
|
|
"diagnosis": "Failed to parse AI response",
|
|
"proposed_action": "manual investigation required",
|
|
"action_type": "investigation",
|
|
"risk_level": "high",
|
|
"reasoning": "AI response was not in expected format"
|
|
}
|
|
except json.JSONDecodeError:
|
|
return {
|
|
"diagnosis": "Failed to parse AI response",
|
|
"proposed_action": "manual investigation required",
|
|
"action_type": "investigation",
|
|
"risk_level": "high",
|
|
"reasoning": f"Raw response: {response[:500]}"
|
|
}
|
|
|
|
def _build_analysis_context(self, data: Dict[str, Any]) -> str:
|
|
"""Build a concise context string for the AI"""
|
|
lines = []
|
|
|
|
# System resources
|
|
res = data.get("resources", {})
|
|
lines.append(f"CPU: {res.get('cpu_percent', 0):.1f}%, Memory: {res.get('memory_percent', 0):.1f}%, Load: {res.get('load_average', {}).get('1min', 0):.2f}")
|
|
|
|
# Disk usage
|
|
disk = data.get("disk", {})
|
|
for part in disk.get("partitions", []):
|
|
if part.get("percent_used", 0) > 80: # Only mention if >80% full
|
|
lines.append(f"⚠️ Disk {part['mountpoint']}: {part['percent_used']:.1f}% full")
|
|
|
|
# Failed services
|
|
systemd = data.get("systemd", {})
|
|
if systemd.get("failed_count", 0) > 0:
|
|
lines.append(f"\n⚠️ {systemd['failed_count']} failed systemd services:")
|
|
for svc in systemd.get("failed_services", [])[:10]:
|
|
lines.append(f" - {svc.get('unit', 'unknown')}: {svc.get('sub', 'unknown')}")
|
|
|
|
# Recent errors
|
|
logs = data.get("logs", {})
|
|
error_count = logs.get("error_count_1h", 0)
|
|
if error_count > 0:
|
|
lines.append(f"\n{error_count} errors in last hour")
|
|
# Group errors by service
|
|
errors_by_service = {}
|
|
for err in logs.get("recent_errors", [])[:20]:
|
|
svc = err.get("SYSLOG_IDENTIFIER", "unknown")
|
|
errors_by_service[svc] = errors_by_service.get(svc, 0) + 1
|
|
for svc, count in sorted(errors_by_service.items(), key=lambda x: x[1], reverse=True)[:5]:
|
|
lines.append(f" - {svc}: {count} errors")
|
|
|
|
# Network
|
|
net = data.get("network", {})
|
|
if not net.get("internet_reachable", True):
|
|
lines.append("\n⚠️ No internet connectivity")
|
|
|
|
return "\n".join(lines)
|
|
|
|
def _create_analysis_prompt(self, context: str, system_context: str = "") -> str:
|
|
"""Create the analysis prompt for the AI"""
|
|
|
|
prompt = f"""{self.SYSTEM_PROMPT}
|
|
|
|
TASK: ANALYZE SYSTEM HEALTH
|
|
================================================================================
|
|
|
|
OBJECTIVE:
|
|
Analyze the current system state and determine if any action is needed.
|
|
Be thorough but not alarmist. Only recommend action if truly necessary.
|
|
"""
|
|
|
|
if system_context:
|
|
prompt += f"\n\nSYSTEM INFRASTRUCTURE:\n{system_context}"
|
|
|
|
prompt += f"""
|
|
|
|
CURRENT SYSTEM STATE:
|
|
{context}
|
|
|
|
YOUR RESPONSE MUST BE VALID JSON:
|
|
{{
|
|
"status": "one of: healthy, attention_needed, intervention_required",
|
|
"issues": [
|
|
{{
|
|
"severity": "one of: info, warning, critical",
|
|
"category": "one of: resources, services, disk, network, logs",
|
|
"description": "brief description of the issue",
|
|
"requires_action": true/false
|
|
}}
|
|
],
|
|
"overall_assessment": "brief summary of system health",
|
|
"recommended_actions": ["list of recommended actions, if any"]
|
|
}}
|
|
|
|
RESPOND WITH ONLY THE JSON, NO OTHER TEXT.
|
|
"""
|
|
|
|
return prompt
|
|
|
|
def _auto_diagnose_ollama(self) -> str:
|
|
"""Automatically diagnose Ollama issues"""
|
|
diagnostics = []
|
|
|
|
diagnostics.append("=== OLLAMA SELF-DIAGNOSTIC ===")
|
|
|
|
# Check if Ollama service is running
|
|
try:
|
|
result = subprocess.run(
|
|
['systemctl', 'is-active', 'ollama.service'],
|
|
capture_output=True,
|
|
text=True,
|
|
timeout=5
|
|
)
|
|
if result.returncode == 0:
|
|
diagnostics.append("✅ Ollama service is active")
|
|
else:
|
|
diagnostics.append(f"❌ Ollama service is NOT active: {result.stdout.strip()}")
|
|
except Exception as e:
|
|
diagnostics.append(f"⚠️ Could not check service status: {e}")
|
|
|
|
# Check memory usage
|
|
try:
|
|
result = subprocess.run(['free', '-h'], capture_output=True, text=True, timeout=5)
|
|
diagnostics.append(f"\nMemory:\n{result.stdout}")
|
|
except Exception as e:
|
|
diagnostics.append(f"⚠️ Could not check memory: {e}")
|
|
|
|
# Check which models are loaded
|
|
try:
|
|
response = requests.get(f"{self.ollama_host}/api/tags", timeout=5)
|
|
if response.status_code == 200:
|
|
models = response.json().get('models', [])
|
|
diagnostics.append(f"\nLoaded models: {len(models)}")
|
|
for model in models:
|
|
name = model.get('name', 'unknown')
|
|
size = model.get('size', 0) / (1024**3)
|
|
is_target = "← TARGET" if name == self.model else ""
|
|
diagnostics.append(f" • {name} ({size:.1f} GB) {is_target}")
|
|
|
|
# Check if target model is loaded
|
|
model_names = [m.get('name') for m in models]
|
|
if self.model not in model_names:
|
|
diagnostics.append(f"\n❌ TARGET MODEL NOT LOADED: {self.model}")
|
|
diagnostics.append(f" Available: {', '.join(model_names)}")
|
|
else:
|
|
diagnostics.append(f"❌ Ollama API returned {response.status_code}")
|
|
except Exception as e:
|
|
diagnostics.append(f"⚠️ Could not query Ollama API: {e}")
|
|
|
|
# Check recent Ollama logs
|
|
try:
|
|
result = subprocess.run(
|
|
['journalctl', '-u', 'ollama.service', '-n', '20', '--no-pager'],
|
|
capture_output=True,
|
|
text=True,
|
|
timeout=5
|
|
)
|
|
if result.stdout:
|
|
diagnostics.append(f"\nRecent logs:\n{result.stdout}")
|
|
except Exception as e:
|
|
diagnostics.append(f"⚠️ Could not check logs: {e}")
|
|
|
|
return "\n".join(diagnostics)
|
|
|
|
def _query_ollama(self, prompt: str, temperature: float = 0.3) -> str:
|
|
"""Query Ollama API (with optional queue)"""
|
|
# If queue is enabled, submit to queue and wait
|
|
if self.use_queue and self.ollama_queue:
|
|
try:
|
|
payload = {
|
|
"model": self.model,
|
|
"prompt": prompt,
|
|
"stream": False,
|
|
"temperature": temperature,
|
|
"timeout": 120
|
|
}
|
|
|
|
request_id = self.ollama_queue.submit(
|
|
request_type="generate",
|
|
payload=payload,
|
|
priority=self.priority_level
|
|
)
|
|
|
|
result = self.ollama_queue.wait_for_result(request_id, timeout=300)
|
|
return result.get("response", "")
|
|
|
|
except Exception as e:
|
|
print(f"Warning: Queue request failed, falling back to direct: {e}")
|
|
# Fall through to direct query
|
|
|
|
# Direct query (no queue or queue failed)
|
|
try:
|
|
response = requests.post(
|
|
f"{self.ollama_host}/api/generate",
|
|
json={
|
|
"model": self.model,
|
|
"prompt": prompt,
|
|
"stream": False,
|
|
"temperature": temperature,
|
|
},
|
|
timeout=120 # 2 minute timeout for large models
|
|
)
|
|
response.raise_for_status()
|
|
return response.json().get("response", "")
|
|
except requests.exceptions.HTTPError as e:
|
|
error_detail = ""
|
|
try:
|
|
error_detail = f" - {response.text}"
|
|
except:
|
|
pass
|
|
print(f"ERROR: Ollama HTTP error {response.status_code}{error_detail}")
|
|
print(f"Model requested: {self.model}")
|
|
print(f"Ollama host: {self.ollama_host}")
|
|
# Run diagnostics
|
|
diagnostics = self._auto_diagnose_ollama()
|
|
print(diagnostics)
|
|
return json.dumps({
|
|
"error": f"Ollama HTTP {response.status_code}: {str(e)}{error_detail}",
|
|
"diagnosis": f"Ollama API error - check if model '{self.model}' is available",
|
|
"action_type": "investigation",
|
|
"risk_level": "high"
|
|
})
|
|
except Exception as e:
|
|
print(f"ERROR: Failed to query Ollama: {str(e)}")
|
|
print(f"Model requested: {self.model}")
|
|
print(f"Ollama host: {self.ollama_host}")
|
|
# Run diagnostics
|
|
diagnostics = self._auto_diagnose_ollama()
|
|
print(diagnostics)
|
|
return json.dumps({
|
|
"error": f"Failed to query Ollama: {str(e)}",
|
|
"diagnosis": "Ollama API unavailable",
|
|
"action_type": "investigation",
|
|
"risk_level": "high"
|
|
})
|
|
|
|
def _estimate_tokens(self, text: str) -> int:
|
|
"""Rough token estimation: ~4 chars per token"""
|
|
return len(text) // 4
|
|
|
|
def _extract_key_findings(self, tool_name: str, raw_output: str, progress_callback=None) -> str:
|
|
"""
|
|
Extract key findings from large tool output using chunked map-reduce.
|
|
Processes large outputs in smaller chunks to prevent Ollama overload.
|
|
"""
|
|
output_size = len(raw_output)
|
|
chunk_size = 8000 # ~2000 tokens per chunk, safe size
|
|
|
|
# Store full output in cache for potential deep dive
|
|
cache_id = f"{tool_name}_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
|
|
try:
|
|
cache_file = self.cache_dir / f"{cache_id}.txt"
|
|
cache_file.write_text(raw_output)
|
|
except (PermissionError, OSError):
|
|
# Fallback to temp directory if cache dir not writable
|
|
import tempfile
|
|
cache_file = Path(tempfile.gettempdir()) / f"macha_{cache_id}.txt"
|
|
cache_file.write_text(raw_output)
|
|
|
|
# If output is small enough, process in one go
|
|
if output_size <= chunk_size:
|
|
try:
|
|
extraction_prompt = f"""Analyze this output from '{tool_name}'.
|
|
|
|
Extract: key findings, errors/warnings, metrics, actionable insights.
|
|
|
|
Output:
|
|
{raw_output}
|
|
|
|
Provide concise summary (max 600 chars)."""
|
|
|
|
summary = self._query_ollama(extraction_prompt, temperature=0.1)
|
|
return f"[Summary of {tool_name}]:\n{summary}\n\n[Full output: {output_size:,} chars cached as {cache_id}]"
|
|
except Exception as e:
|
|
print(f"Warning: Failed to extract findings: {e}")
|
|
return self._simple_truncate(raw_output, 2000)
|
|
|
|
# Large output: chunk and process with map-reduce
|
|
try:
|
|
chunks = []
|
|
num_chunks = (output_size + chunk_size - 1) // chunk_size
|
|
|
|
for i in range(0, output_size, chunk_size):
|
|
chunk = raw_output[i:i+chunk_size]
|
|
chunks.append(chunk)
|
|
|
|
# Phase 1: Map - Summarize each chunk
|
|
chunk_summaries = []
|
|
for idx, chunk in enumerate(chunks):
|
|
chunk_num = idx + 1
|
|
|
|
# Progress feedback
|
|
if progress_callback:
|
|
progress_callback(f" Processing chunk {chunk_num}/{num_chunks}...")
|
|
else:
|
|
print(f" → Processing chunk {chunk_num}/{num_chunks}...", flush=True)
|
|
|
|
extraction_prompt = f"""Analyze chunk {chunk_num}/{num_chunks} from '{tool_name}'.
|
|
|
|
Extract: key findings, errors/warnings, metrics, insights.
|
|
|
|
Chunk:
|
|
{chunk}
|
|
|
|
Concise summary (max 400 chars)."""
|
|
|
|
chunk_summary = self._query_ollama(extraction_prompt, temperature=0.1)
|
|
chunk_summaries.append(f"[Chunk {chunk_num}]: {chunk_summary}")
|
|
|
|
# Phase 2: Reduce - Combine chunk summaries if many chunks
|
|
if len(chunk_summaries) > 5:
|
|
if progress_callback:
|
|
progress_callback(f" Synthesizing {len(chunk_summaries)} chunk summaries...")
|
|
else:
|
|
print(f" → Synthesizing {len(chunk_summaries)} chunk summaries...", flush=True)
|
|
|
|
combined = "\n".join(chunk_summaries)
|
|
reduce_prompt = f"""Synthesize these chunk summaries from '{tool_name}':
|
|
|
|
{combined}
|
|
|
|
Provide unified summary (max 800 chars) covering all key points."""
|
|
|
|
final_summary = self._query_ollama(reduce_prompt, temperature=0.1)
|
|
return f"""[Chunked analysis of {tool_name}]:
|
|
{final_summary}
|
|
|
|
[Processed {num_chunks} chunks, {output_size:,} chars total, cached as {cache_id}]"""
|
|
|
|
else:
|
|
# Few chunks: just concatenate summaries
|
|
combined_summary = "\n".join(chunk_summaries)
|
|
return f"""[Chunked analysis of {tool_name}]:
|
|
{combined_summary}
|
|
|
|
[Processed {num_chunks} chunks, {output_size:,} chars total, cached as {cache_id}]"""
|
|
|
|
except Exception as e:
|
|
print(f"Warning: Chunked extraction failed for {tool_name}: {e}")
|
|
return self._simple_truncate(raw_output, 2000)
|
|
|
|
def _simple_truncate(self, text: str, max_chars: int) -> str:
|
|
"""Simple head+tail truncation"""
|
|
if len(text) <= max_chars:
|
|
return text
|
|
|
|
half = max_chars // 2
|
|
return (
|
|
text[:half] +
|
|
f"\n... [TRUNCATED: {len(text) - max_chars} chars omitted] ...\n" +
|
|
text[-half:]
|
|
)
|
|
|
|
def _process_tool_result_hierarchical(self, tool_name: str, result: Any) -> str:
|
|
"""
|
|
Intelligently process tool results based on size:
|
|
- Small (< 5KB): Pass through directly
|
|
- Medium (5-20KB): Hierarchical extraction with single-pass summarization
|
|
- Large (> 20KB): Hierarchical extraction with chunked processing
|
|
"""
|
|
result_str = json.dumps(result) if not isinstance(result, str) else result
|
|
size = len(result_str)
|
|
|
|
# Small outputs: pass through directly
|
|
if size < 5000:
|
|
print(f" [Tool result: {size} chars, passing through]")
|
|
return result_str
|
|
|
|
# Medium and large outputs: hierarchical extraction with chunking
|
|
else:
|
|
print(f" [Tool result: {size} chars, extracting key findings...]")
|
|
# _extract_key_findings automatically chunks large outputs
|
|
return self._extract_key_findings(tool_name, result_str)
|
|
|
|
def _prune_messages(self, messages: List[Dict], max_context_tokens: int = 80000) -> List[Dict]:
|
|
"""
|
|
Prune message history to stay within context limits.
|
|
Keeps: system prompt + recent conversation window
|
|
"""
|
|
if not messages:
|
|
return messages
|
|
|
|
# Separate system message from conversation
|
|
system_msg = None
|
|
conversation = []
|
|
|
|
for msg in messages:
|
|
if msg["role"] == "system":
|
|
system_msg = msg
|
|
else:
|
|
conversation.append(msg)
|
|
|
|
# Calculate current token count
|
|
total_tokens = 0
|
|
if system_msg:
|
|
total_tokens += self._estimate_tokens(system_msg["content"])
|
|
|
|
for msg in conversation:
|
|
content = msg.get("content", "")
|
|
total_tokens += self._estimate_tokens(str(content))
|
|
|
|
# If under limit, return as-is
|
|
if total_tokens <= max_context_tokens:
|
|
result = []
|
|
if system_msg:
|
|
result.append(system_msg)
|
|
result.extend(conversation)
|
|
print(f"[Context: {total_tokens:,} tokens, {len(conversation)} messages]")
|
|
return result
|
|
|
|
# Need to prune - keep sliding window of recent messages
|
|
# Strategy: Keep last 20 messages (10 exchanges) which should be ~40K tokens max
|
|
print(f"[Context pruning: {total_tokens:,} tokens → keeping last 20 messages]")
|
|
|
|
pruned_conversation = conversation[-20:]
|
|
|
|
result = []
|
|
if system_msg:
|
|
result.append(system_msg)
|
|
result.extend(pruned_conversation)
|
|
|
|
# Calculate new token count
|
|
new_tokens = self._estimate_tokens(system_msg["content"]) if system_msg else 0
|
|
for msg in pruned_conversation:
|
|
new_tokens += self._estimate_tokens(str(msg.get("content", "")))
|
|
|
|
print(f"[Context after pruning: {new_tokens:,} tokens, {len(pruned_conversation)} messages]")
|
|
|
|
return result
|
|
|
|
def _query_ollama_with_tools(
|
|
self,
|
|
messages: List[Dict[str, str]],
|
|
temperature: float = 0.3,
|
|
max_iterations: int = 30
|
|
) -> str:
|
|
"""
|
|
Query Ollama using chat API with tool support.
|
|
Handles tool calls and returns final response.
|
|
|
|
Args:
|
|
messages: List of chat messages [{"role": "user", "content": "..."}]
|
|
temperature: Generation temperature
|
|
max_iterations: Maximum number of tool-calling iterations (default 30 for complex system operations)
|
|
|
|
Returns:
|
|
Final text response from the model
|
|
"""
|
|
if not self.enable_tools or not self.tools:
|
|
# Fallback to regular query
|
|
user_content = " ".join([m["content"] for m in messages if m["role"] == "user"])
|
|
return self._query_ollama(user_content, temperature)
|
|
|
|
# Add system message if not present
|
|
if not any(m["role"] == "system" for m in messages):
|
|
messages = [{"role": "system", "content": self.SYSTEM_PROMPT}] + messages
|
|
|
|
tool_definitions = self.tools.get_tool_definitions()
|
|
|
|
for iteration in range(max_iterations):
|
|
try:
|
|
# Prune messages before sending to avoid context overflow
|
|
pruned_messages = self._prune_messages(messages, max_context_tokens=80000)
|
|
|
|
# Use queue if enabled
|
|
if self.use_queue and self.ollama_queue:
|
|
try:
|
|
payload = {
|
|
"model": self.model,
|
|
"messages": pruned_messages,
|
|
"stream": False,
|
|
"temperature": temperature,
|
|
"tools": tool_definitions,
|
|
"timeout": 120
|
|
}
|
|
|
|
request_id = self.ollama_queue.submit(
|
|
request_type="chat_with_tools",
|
|
payload=payload,
|
|
priority=self.priority_level
|
|
)
|
|
|
|
resp_data = self.ollama_queue.wait_for_result(request_id, timeout=300)
|
|
|
|
except Exception as e:
|
|
print(f"Warning: Queue request failed, falling back to direct: {e}")
|
|
# Fall through to direct query
|
|
response = requests.post(
|
|
f"{self.ollama_host}/api/chat",
|
|
json={
|
|
"model": self.model,
|
|
"messages": pruned_messages,
|
|
"stream": False,
|
|
"temperature": temperature,
|
|
"tools": tool_definitions
|
|
},
|
|
timeout=120
|
|
)
|
|
response.raise_for_status()
|
|
resp_data = response.json()
|
|
else:
|
|
# Direct query (no queue)
|
|
response = requests.post(
|
|
f"{self.ollama_host}/api/chat",
|
|
json={
|
|
"model": self.model,
|
|
"messages": pruned_messages,
|
|
"stream": False,
|
|
"temperature": temperature,
|
|
"tools": tool_definitions
|
|
},
|
|
timeout=120
|
|
)
|
|
response.raise_for_status()
|
|
resp_data = response.json()
|
|
|
|
message = resp_data.get("message", {})
|
|
|
|
# Check if model wants to call tools
|
|
tool_calls = message.get("tool_calls", [])
|
|
|
|
if not tool_calls:
|
|
# No tools to call, return the text response
|
|
return message.get("content", "")
|
|
|
|
# Add assistant's message to history
|
|
messages.append(message)
|
|
|
|
# Execute each tool call
|
|
for tool_call in tool_calls:
|
|
function_name = tool_call["function"]["name"]
|
|
arguments = tool_call["function"]["arguments"]
|
|
|
|
print(f" → Tool call: {function_name}({arguments})")
|
|
|
|
# Execute the tool
|
|
tool_result = self.tools.execute_tool(function_name, arguments)
|
|
|
|
# Process result hierarchically based on size
|
|
processed_result = self._process_tool_result_hierarchical(function_name, tool_result)
|
|
|
|
# Add processed result to messages
|
|
messages.append({
|
|
"role": "tool",
|
|
"content": processed_result
|
|
})
|
|
|
|
# Continue loop to let model process tool results
|
|
|
|
except requests.exceptions.HTTPError as e:
|
|
error_body = ""
|
|
try:
|
|
error_body = response.text
|
|
except:
|
|
pass
|
|
|
|
# Check if this is a context length error
|
|
if "context length" in error_body.lower() or "too long" in error_body.lower():
|
|
print(f"ERROR: Context length exceeded. Attempting recovery...")
|
|
# Emergency pruning - keep only system + last user message
|
|
system_msg = next((m for m in messages if m["role"] == "system"), None)
|
|
last_user_msg = next((m for m in reversed(messages) if m["role"] == "user"), None)
|
|
|
|
if system_msg and last_user_msg:
|
|
messages = [system_msg, last_user_msg]
|
|
print(f"[Emergency context reset: keeping only system + last user message]")
|
|
continue # Retry with minimal context
|
|
|
|
print(f"ERROR: Ollama chat API error: {e}")
|
|
diagnostics = self._auto_diagnose_ollama()
|
|
print(diagnostics)
|
|
return json.dumps({
|
|
"error": f"Ollama chat API error: {str(e)}",
|
|
"diagnosis": "Failed to communicate with Ollama",
|
|
"action_type": "investigation",
|
|
"risk_level": "high"
|
|
})
|
|
except Exception as e:
|
|
print(f"ERROR: Tool calling failed: {e}")
|
|
return json.dumps({
|
|
"error": f"Tool calling error: {str(e)}",
|
|
"diagnosis": "Failed during tool execution",
|
|
"action_type": "investigation",
|
|
"risk_level": "high"
|
|
})
|
|
|
|
# If we hit max iterations, return what we have
|
|
return "Maximum tool calling iterations reached. Unable to complete request."
|
|
|
|
def _parse_analysis_response(self, response: str) -> Dict[str, Any]:
|
|
"""Parse the AI's analysis response"""
|
|
import re
|
|
|
|
# Log the raw response for debugging
|
|
self._log(f"AI raw response (first 1000 chars): {response[:1000]}")
|
|
|
|
try:
|
|
json_match = re.search(r'\{.*\}', response, re.DOTALL)
|
|
if json_match:
|
|
parsed = json.loads(json_match.group())
|
|
self._log(f"Successfully parsed AI response: {parsed.get('status', 'unknown')}")
|
|
return parsed
|
|
else:
|
|
self._log("ERROR: No JSON found in AI response")
|
|
except Exception as e:
|
|
self._log(f"ERROR parsing AI response: {e}")
|
|
|
|
# Fallback
|
|
self._log("Falling back to default response")
|
|
return {
|
|
"status": "healthy",
|
|
"issues": [],
|
|
"overall_assessment": "Unable to parse AI response",
|
|
"recommended_actions": []
|
|
}
|
|
|
|
def _log(self, message: str):
|
|
"""Log a message to the orchestrator log"""
|
|
# This will go to the orchestrator log via print
|
|
print(f"[AGENT] {message}")
|
|
|
|
def _log_decision(self, monitoring_data: Dict[str, Any], analysis: Dict[str, Any]):
|
|
"""Log AI decisions for auditing"""
|
|
log_entry = {
|
|
"timestamp": datetime.now().isoformat(),
|
|
"monitoring_summary": {
|
|
"cpu": monitoring_data.get("resources", {}).get("cpu_percent"),
|
|
"memory": monitoring_data.get("resources", {}).get("memory_percent"),
|
|
"failed_services": monitoring_data.get("systemd", {}).get("failed_count"),
|
|
"error_count": monitoring_data.get("logs", {}).get("error_count_1h"),
|
|
},
|
|
"analysis": analysis,
|
|
}
|
|
|
|
with open(self.decision_log, 'a') as f:
|
|
f.write(json.dumps(log_entry) + '\n')
|
|
|
|
def get_recent_decisions(self, count: int = 10) -> List[Dict[str, Any]]:
|
|
"""Get recent decision history"""
|
|
if not self.decision_log.exists():
|
|
return []
|
|
|
|
decisions = []
|
|
with open(self.decision_log, 'r') as f:
|
|
for line in f:
|
|
if line.strip():
|
|
try:
|
|
decisions.append(json.loads(line))
|
|
except:
|
|
pass
|
|
|
|
return decisions[-count:]
|
|
|
|
|
|
if __name__ == "__main__":
|
|
import sys
|
|
|
|
# Test the agent
|
|
agent = MachaAgent()
|
|
|
|
if len(sys.argv) > 1 and sys.argv[1] == "test":
|
|
# Test with sample data
|
|
test_data = {
|
|
"systemd": {"failed_count": 2, "failed_services": [
|
|
{"unit": "test-service.service", "sub": "failed"}
|
|
]},
|
|
"resources": {"cpu_percent": 25.0, "memory_percent": 45.0, "load_average": {"1min": 1.5}},
|
|
"logs": {"error_count_1h": 10},
|
|
"network": {"internet_reachable": True}
|
|
}
|
|
|
|
print("Testing agent analysis...")
|
|
analysis = agent.analyze_system_state(test_data)
|
|
print(json.dumps(analysis, indent=2))
|
|
|
|
if analysis.get("issues"):
|
|
print("\nTesting fix proposal...")
|
|
fix = agent.propose_fix(
|
|
analysis["issues"][0]["description"],
|
|
test_data
|
|
)
|
|
print(json.dumps(fix, indent=2))
|