Recuperación de errores para agentes

Los errores ocurren. Las redes fallan, las APIs devuelven 500s, los Mind Keys expiran. Esta guía muestra cómo los agentes LLM deben manejar errores de forma elegante y aprender de ellos.

Principios de manejo de errores

1. Reintente con backoff — los errores transitorios suelen resolverse
2. Almacene el error — aprenda de los patrones
3. Degrade de forma elegante — no bloquee toda la sesión
4. Notifique al humano — para errores que no pueda resolver

Manejo de errores HTTP

Reintento con backoff exponencial

import time
import requests

def call_with_retry(url, max_retries=3, backoff_base=2):
    """Call URL with exponential backoff."""
    for attempt in range(max_retries):
        try:
            r = requests.get(url, headers={"Authorization": f"Bearer {KEY}"})
            
            # Success
            if r.status_code < 400:
                return r
            
            # Rate limited — wait and retry
            if r.status_code == 429:
                wait = int(r.headers.get("Retry-After", 60))
                print(f"Rate limited. Waiting {wait}s...")
                time.sleep(wait)
                continue
            
            # Server error — retry with backoff
            if r.status_code >= 500:
                wait = backoff_base ** attempt
                print(f"Server error {r.status_code}. Retrying in {wait}s...")
                time.sleep(wait)
                continue
            
            # Client error — don't retry
            if 400 <= r.status_code < 500:
                raise ClientError(f"{r.status_code}: {r.text}")
        
        except requests.RequestException as e:
            # Network error — retry
            wait = backoff_base ** attempt
            print(f"Network error: {e}. Retrying in {wait}s...")
            time.sleep(wait)
    
    raise MaxRetriesError(f"Failed after {max_retries} retries")

Manejo de errores de auth

def safe_call(url):
    """Call with auth error handling."""
    try:
        return call_with_retry(url)
    except ClientError as e:
        if "401" in str(e):
            # Mind Key invalid — critical, can't recover
            store_error("auth_invalid", str(e), "Check Mind Key")
            notify_human("My Mind Key is invalid. Please update.")
            raise AuthError("Cannot continue without valid Mind Key")
        elif "403" in str(e):
            store_error("forbidden", str(e), "Wrong token type?")
            raise ForbiddenError(str(e))
        elif "404" in str(e):
            # Path doesn't exist — don't retry
            store_error("not_found", str(e), "Check endpoint path")
            raise NotFoundError(str(e))
        else:
            raise

Almacenar errores como memorias

Cuando ocurren errores, almacénelos para que las sesiones futuras puedan aprender:

def store_error(error_type, error_message, recovery_hint=""):
    """Store an error as a memory for future reference."""
    requests.post(f"{URL}/memory",
        headers={"Authorization": f"Bearer {KEY}",
                 "Content-Type": "application/json"},
        json={
            "category": "mistake",
            "key": f"error_{error_type}_{int(time.time())}",
            "content": f"Error: {error_type}\nMessage: {error_message}\nRecovery: {recovery_hint}",
            "tags": ["error", error_type],
            "priority": "high"
        })

# Example
try:
    deploy()
except DeployError as e:
    store_error("deploy_failed", str(e), 
                "Check CI logs, verify Docker image exists")
    raise

Escenarios de error comunes

Escenario 1: Mind Key inválido

# Detection: 401 on every call
# Recovery: Cannot recover — need human intervention

def handle_invalid_mind_key():
    store_error("mind_key_invalid", 
                "All API calls returning 401",
                "Mind Key may be revoked. Need new key.")
    
    # Notify human via chat (if possible)
    try:
        reply("⚠️ My Mind Key is invalid. I cannot access memories. "
              "Please check and update SYNAPSE_MIND_KEY.")
    except:
        pass  # Chat may also fail with bad key
    
    # Exit gracefully
    raise CriticalError("Cannot continue without valid Mind Key")

Escenario 2: Error de red

# Detection: ConnectionError, Timeout
# Recovery: Retry with backoff, then degrade

def handle_network_error(url, retry=3):
    for attempt in range(retry):
        try:
            return requests.get(url, timeout=10)
        except (requests.ConnectionError, requests.Timeout) as e:
            wait = 2 ** attempt
            print(f"Network error, retrying in {wait}s: {e}")
            time.sleep(wait)
    
    # All retries failed — degrade
    store_error("network_failure", 
                f"Cannot reach {url}",
                "Check internet connection, Synapse may be down")
    
    # Work offline if possible
    return work_offline()

Escenario 3: Limitado por frecuencia

# Detection: 429 with Retry-After header
# Recovery: Wait and retry, or switch to header auth

def handle_rate_limit(response):
    retry_after = int(response.headers.get("Retry-After", 60))
    print(f"Rate limited. Waiting {retry_after}s...")
    time.sleep(retry_after)
    
    # If this keeps happening, suggest switching to header auth
    if has_query_param_auth(url):
        store_error("rate_limited", 
                    "Frequent 429s with ?key= auth",
                    "Switch to Authorization: Bearer header (no rate limit)")

Escenario 4: Error de servidor (5xx)

# Detection: 500, 502, 503
# Recovery: Retry with backoff, check /health

def handle_server_error(url):
    # Check if server is up
    health = requests.get(f"{URL}/health")
    if health.status_code != 200:
        store_error("server_down", 
                    "Synapse health check failing",
                    "Wait for server recovery")
        raise ServerDownError()
    
    # Retry with backoff
    return call_with_retry(url, max_retries=5)

Escenario 5: La llamada a herramienta falla

# Detection: Tool returns error content
# Recovery: Try alternative approach, store failure

def call_tool_safely(tool_name, args, alternatives=None):
    try:
        result = call_tool(tool_name, args)
        if result.get("isError"):
            raise ToolError(result["content"])
        return result
    except ToolError as e:
        store_error(f"tool_{tool_name}_failed",
                    f"Args: {args}\nError: {e}",
                    f"Try: {alternatives or 'no alternatives'}")
        
        # Try alternatives
        if alternatives:
            for alt in alternatives:
                try:
                    return call_tool(alt, args)
                except:
                    continue
        
        raise

Patrón: Circuit Breaker

Para fallos repetidos, deje de intentarlo temporalmente:

class CircuitBreaker:
    def __init__(self, threshold=5, reset_time=300):
        self.failures = 0
        self.threshold = threshold
        self.reset_time = reset_time
        self.last_failure = 0
    
    def call(self, fn, *args, **kwargs):
        if self.failures >= self.threshold:
            if time.time() - self.last_failure < self.reset_time:
                raise CircuitOpenError("Circuit breaker open")
            else:
                # Reset
                self.failures = 0
        
        try:
            result = fn(*args, **kwargs)
            self.failures = 0  # Reset on success
            return result
        except:
            self.failures += 1
            self.last_failure = time.time()
            raise

# Usage
breaker = CircuitBreaker(threshold=5)
try:
    result = breaker.call(api_call, url)
except CircuitOpenError:
    print("Too many failures, waiting before retry")

Mejores prácticas

Próximos pasos

• Referencia de la API de errores
• Patrón de inicio de sesión
• Tests autoreparables