# Error-Recovery für Agenten Fehler passieren. Netzwerke fallen aus, APIs liefern 500er, Mind Keys laufen ab. Dieser Guide zeigt, wie LLM-Agenten Fehler graceful behandeln und daraus lernen. ## Fehlerbehandlungs-Prinzipien 1. **Retry mit Backoff** — transiente Fehler lösen sich oft von selbst 2. **Den Fehler speichern** — aus Mustern lernen 3. **Graceful degradieren** — nicht die ganze Session crashen 4. **Den Menschen benachrichtigen** — bei Fehlern, die du nicht beheben kannst ## HTTP-Fehlerbehandlung ### Retry mit exponentiellem Backoff ```python 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") ``` ### Auth-Fehlerbehandlung ```python 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 ``` ## Fehler als Memories speichern Wenn Fehler auftreten, speichere sie, damit zukünftige Sessions lernen können: ```python 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 ``` ## Häufige Fehler-Szenarien ### Szenario 1: Mind Key ungültig ```python # 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") ``` ### Szenario 2: Netzwerkfehler ```python # 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() ``` ### Szenario 3: Rate Limit ```python # 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)") ``` ### Szenario 4: Serverfehler (5xx) ```python # 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) ``` ### Szenario 5: Tool-Aufruf schlägt fehl ```python # 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 ``` ## Pattern: Circuit Breaker Bei wiederholten Fehlern temporär aufhören zu versuchen: ```python 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") ``` ## Best Practices > [!TIP] > - **Transiente Fehler immer retryen** — Netzwerke fallen aus, Server hicksen > - **Client-Fehler (4xx) nicht retryen** — sie beheben sich nicht von selbst > - **Fehler als Memories speichern** — aus Mustern lernen > - **Menschen bei kritischen Fehlern benachrichtigen** — sie müssen es wissen > - **Graceful degradieren** — partielle Arbeit ist besser als Crash > - **Circuit Breaker verwenden** — keinen fehlerhaften Dienst hämmern ## Nächste Schritte - [Errors-API-Referenz](/docs/api/errors) - [Session-Start-Pattern](/docs/llm-cookbook/session-start-pattern) - [Self-Healing-Tests](/docs/guides/self-healing-tests)