352 lines
11 KiB
Python
352 lines
11 KiB
Python
|
"""
|
||
|
|
Fallback and Graceful Degradation Module
|
||
|
|
|
||
|
|
Provides fallback strategies when components fail.
|
||
|
|
"""
|
||
|
|
|
||
|
|
import asyncio
|
||
|
|
from dataclasses import dataclass, field
|
||
|
|
from enum import Enum
|
||
|
|
from typing import Any, Awaitable, Callable, Dict, List, Optional, TypeVar
|
||
|
|
|
||
|
|
from .errors import (
|
||
|
|
AgentError,
|
||
|
|
LLMError,
|
||
|
|
LLMRateLimitError,
|
||
|
|
LLMTimeoutError,
|
||
|
|
LLMContextLengthError,
|
||
|
|
ToolError,
|
||
|
|
ExternalToolError,
|
||
|
|
RecoveryStrategy,
|
||
|
|
)
|
||
|
|
from .logging import get_logger
|
||
|
|
|
||
|
|
logger = get_logger("fallback")
|
||
|
|
|
||
|
|
T = TypeVar("T")
|
||
|
|
|
||
|
|
|
||
|
|
class FallbackAction(str, Enum):
|
||
|
|
"""Actions to take when fallback is triggered"""
|
||
|
|
RETRY = "retry"
|
||
|
|
RETRY_WITH_REDUCED_CONTEXT = "retry_reduced"
|
||
|
|
USE_FALLBACK_TOOL = "use_fallback"
|
||
|
|
SKIP = "skip"
|
||
|
|
CONTINUE_PARTIAL = "continue_partial"
|
||
|
|
ABORT = "abort"
|
||
|
|
|
||
|
|
|
||
|
|
@dataclass
|
||
|
|
class FallbackResult:
|
||
|
|
"""Result of a fallback operation"""
|
||
|
|
action: FallbackAction
|
||
|
|
success: bool
|
||
|
|
result: Optional[Any] = None
|
||
|
|
error: Optional[Exception] = None
|
||
|
|
fallback_used: Optional[str] = None
|
||
|
|
message: str = ""
|
||
|
|
|
||
|
|
|
||
|
|
@dataclass
|
||
|
|
class FallbackConfig:
|
||
|
|
"""Configuration for fallback behavior"""
|
||
|
|
enabled: bool = True
|
||
|
|
max_context_reduction_ratio: float = 0.5 # Reduce context to 50%
|
||
|
|
continue_on_partial: bool = True
|
||
|
|
tool_fallbacks: Dict[str, str] = field(default_factory=lambda: {
|
||
|
|
"semgrep_scan": "pattern_match",
|
||
|
|
"bandit_scan": "pattern_match",
|
||
|
|
"gitleaks_scan": "search_code",
|
||
|
|
"npm_audit": "search_code",
|
||
|
|
})
|
||
|
|
|
||
|
|
|
||
|
|
class FallbackHandler:
|
||
|
|
"""
|
||
|
|
Handles fallback logic for various failure scenarios.
|
||
|
|
|
||
|
|
Usage:
|
||
|
|
handler = FallbackHandler(config)
|
||
|
|
result = await handler.handle_llm_failure(error, context)
|
||
|
|
"""
|
||
|
|
|
||
|
|
def __init__(self, config: Optional[FallbackConfig] = None):
|
||
|
|
self.config = config or FallbackConfig()
|
||
|
|
|
||
|
|
async def handle_llm_failure(
|
||
|
|
self,
|
||
|
|
error: Exception,
|
||
|
|
context: Dict[str, Any],
|
||
|
|
retry_func: Optional[Callable[[], Awaitable[Any]]] = None,
|
||
|
|
) -> FallbackResult:
|
||
|
|
"""
|
||
|
|
Handle LLM call failure.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
error: The exception that occurred
|
||
|
|
context: Context including messages, iteration, etc.
|
||
|
|
retry_func: Optional function to retry with modified params
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
FallbackResult with action and result
|
||
|
|
"""
|
||
|
|
if not self.config.enabled:
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.ABORT,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message="Fallback disabled",
|
||
|
|
)
|
||
|
|
|
||
|
|
# Handle specific error types
|
||
|
|
if isinstance(error, LLMRateLimitError):
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.RETRY,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message=f"Rate limited, retry after {error.retry_after}s",
|
||
|
|
)
|
||
|
|
|
||
|
|
if isinstance(error, LLMTimeoutError):
|
||
|
|
# Try with reduced context
|
||
|
|
if retry_func and context.get("can_reduce_context", True):
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.RETRY_WITH_REDUCED_CONTEXT,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message="Timeout, will retry with reduced context",
|
||
|
|
)
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.CONTINUE_PARTIAL,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message="Timeout, continuing with partial results",
|
||
|
|
)
|
||
|
|
|
||
|
|
if isinstance(error, LLMContextLengthError):
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.RETRY_WITH_REDUCED_CONTEXT,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message="Context too long, reducing and retrying",
|
||
|
|
)
|
||
|
|
|
||
|
|
# Generic LLM error
|
||
|
|
if isinstance(error, LLMError):
|
||
|
|
if error.recoverable:
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.RETRY,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message=f"LLM error (recoverable): {error.message}",
|
||
|
|
)
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.ABORT,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message=f"LLM error (non-recoverable): {error.message}",
|
||
|
|
)
|
||
|
|
|
||
|
|
# Unknown error
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.ABORT,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message=f"Unknown LLM error: {error}",
|
||
|
|
)
|
||
|
|
|
||
|
|
async def handle_tool_failure(
|
||
|
|
self,
|
||
|
|
tool_name: str,
|
||
|
|
error: Exception,
|
||
|
|
tool_input: Dict[str, Any],
|
||
|
|
fallback_executor: Optional[Callable[[str, Dict], Awaitable[Any]]] = None,
|
||
|
|
) -> FallbackResult:
|
||
|
|
"""
|
||
|
|
Handle tool execution failure.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
tool_name: Name of the failed tool
|
||
|
|
error: The exception that occurred
|
||
|
|
tool_input: Original tool input
|
||
|
|
fallback_executor: Function to execute fallback tool
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
FallbackResult with action and result
|
||
|
|
"""
|
||
|
|
if not self.config.enabled:
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.ABORT,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message="Fallback disabled",
|
||
|
|
)
|
||
|
|
|
||
|
|
# Check if there's a fallback tool
|
||
|
|
fallback_tool = self.config.tool_fallbacks.get(tool_name)
|
||
|
|
|
||
|
|
if fallback_tool and fallback_executor:
|
||
|
|
try:
|
||
|
|
logger.info(
|
||
|
|
f"Using fallback tool {fallback_tool} for failed {tool_name}",
|
||
|
|
tool_name=tool_name,
|
||
|
|
fallback_tool=fallback_tool,
|
||
|
|
)
|
||
|
|
|
||
|
|
result = await fallback_executor(fallback_tool, tool_input)
|
||
|
|
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.USE_FALLBACK_TOOL,
|
||
|
|
success=True,
|
||
|
|
result=result,
|
||
|
|
fallback_used=fallback_tool,
|
||
|
|
message=f"Used fallback tool: {fallback_tool}",
|
||
|
|
)
|
||
|
|
except Exception as fallback_error:
|
||
|
|
logger.warning(
|
||
|
|
f"Fallback tool {fallback_tool} also failed: {fallback_error}",
|
||
|
|
tool_name=tool_name,
|
||
|
|
fallback_tool=fallback_tool,
|
||
|
|
)
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.SKIP,
|
||
|
|
success=False,
|
||
|
|
error=fallback_error,
|
||
|
|
fallback_used=fallback_tool,
|
||
|
|
message=f"Fallback tool also failed: {fallback_error}",
|
||
|
|
)
|
||
|
|
|
||
|
|
# No fallback available
|
||
|
|
if isinstance(error, (ToolError, ExternalToolError)):
|
||
|
|
if error.recoverable:
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.RETRY,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message=f"Tool error (recoverable): {error}",
|
||
|
|
)
|
||
|
|
|
||
|
|
return FallbackResult(
|
||
|
|
action=FallbackAction.SKIP,
|
||
|
|
success=False,
|
||
|
|
error=error,
|
||
|
|
message=f"Tool failed, skipping: {error}",
|
||
|
|
)
|
||
|
|
|
||
|
|
def reduce_context(
|
||
|
|
self,
|
||
|
|
messages: List[Dict[str, Any]],
|
||
|
|
reduction_ratio: Optional[float] = None,
|
||
|
|
) -> List[Dict[str, Any]]:
|
||
|
|
"""
|
||
|
|
Reduce context size by removing older messages.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
messages: List of conversation messages
|
||
|
|
reduction_ratio: How much to keep (0.5 = keep 50%)
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
Reduced list of messages
|
||
|
|
"""
|
||
|
|
ratio = reduction_ratio or self.config.max_context_reduction_ratio
|
||
|
|
|
||
|
|
if len(messages) <= 2:
|
||
|
|
return messages
|
||
|
|
|
||
|
|
# Always keep system message and last user message
|
||
|
|
system_messages = [m for m in messages if m.get("role") == "system"]
|
||
|
|
other_messages = [m for m in messages if m.get("role") != "system"]
|
||
|
|
|
||
|
|
# Keep ratio of non-system messages (at least the last one)
|
||
|
|
keep_count = max(1, int(len(other_messages) * ratio))
|
||
|
|
kept_messages = other_messages[-keep_count:]
|
||
|
|
|
||
|
|
return system_messages + kept_messages
|
||
|
|
|
||
|
|
def truncate_content(
|
||
|
|
self,
|
||
|
|
content: str,
|
||
|
|
max_length: int = 50000,
|
||
|
|
keep_start: int = 20000,
|
||
|
|
keep_end: int = 20000,
|
||
|
|
) -> str:
|
||
|
|
"""
|
||
|
|
Truncate content while keeping start and end.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
content: Content to truncate
|
||
|
|
max_length: Maximum length
|
||
|
|
keep_start: Characters to keep from start
|
||
|
|
keep_end: Characters to keep from end
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
Truncated content
|
||
|
|
"""
|
||
|
|
if len(content) <= max_length:
|
||
|
|
return content
|
||
|
|
|
||
|
|
truncation_notice = "\n\n... [CONTENT TRUNCATED] ...\n\n"
|
||
|
|
available = max_length - len(truncation_notice)
|
||
|
|
|
||
|
|
start_len = min(keep_start, available // 2)
|
||
|
|
end_len = min(keep_end, available - start_len)
|
||
|
|
|
||
|
|
return content[:start_len] + truncation_notice + content[-end_len:]
|
||
|
|
|
||
|
|
|
||
|
|
# ============ Fallback Decorator ============
|
||
|
|
|
||
|
|
def with_fallback(
|
||
|
|
fallback_func: Optional[Callable[..., Awaitable[T]]] = None,
|
||
|
|
on_error: Optional[Callable[[Exception], Awaitable[None]]] = None,
|
||
|
|
default_value: Optional[T] = None,
|
||
|
|
):
|
||
|
|
"""
|
||
|
|
Decorator to add fallback behavior to async functions.
|
||
|
|
|
||
|
|
Usage:
|
||
|
|
@with_fallback(fallback_func=backup_llm_call, default_value="")
|
||
|
|
async def call_llm(prompt: str):
|
||
|
|
...
|
||
|
|
"""
|
||
|
|
def decorator(func: Callable[..., Awaitable[T]]) -> Callable[..., Awaitable[T]]:
|
||
|
|
async def wrapper(*args, **kwargs) -> T:
|
||
|
|
try:
|
||
|
|
return await func(*args, **kwargs)
|
||
|
|
except Exception as e:
|
||
|
|
if on_error:
|
||
|
|
await on_error(e)
|
||
|
|
|
||
|
|
if fallback_func:
|
||
|
|
try:
|
||
|
|
return await fallback_func(*args, **kwargs)
|
||
|
|
except Exception:
|
||
|
|
pass
|
||
|
|
|
||
|
|
if default_value is not None:
|
||
|
|
return default_value
|
||
|
|
|
||
|
|
raise
|
||
|
|
|
||
|
|
return wrapper
|
||
|
|
return decorator
|
||
|
|
|
||
|
|
|
||
|
|
# ============ Global Handler ============
|
||
|
|
|
||
|
|
_global_handler: Optional[FallbackHandler] = None
|
||
|
|
|
||
|
|
|
||
|
|
def get_fallback_handler() -> FallbackHandler:
|
||
|
|
"""Get global fallback handler"""
|
||
|
|
global _global_handler
|
||
|
|
if _global_handler is None:
|
||
|
|
_global_handler = FallbackHandler()
|
||
|
|
return _global_handler
|
||
|
|
|
||
|
|
|
||
|
|
def configure_fallback(config: FallbackConfig) -> None:
|
||
|
|
"""Configure global fallback handler"""
|
||
|
|
global _global_handler
|
||
|
|
_global_handler = FallbackHandler(config)
|