Docket worker cannot serialize litellm exceptions after task failure

[bsbodden]

When a background task fails with a litellm exception (rate limit, timeout, connection error, etc.), the docket worker cannot serialize the exception for the result queue.

What happens

docket/worker.py calls cloudpickle.dumps(e) on any exception from a failed task. The dumps succeeds, but cloudpickle.loads() later tries to reconstruct the exception by calling ExceptionClass.__init__() without arguments. litellm exception classes require message, model, and llm_provider as positional args, so this raises TypeError.

The worker can't store or report the error. The task silently disappears.

Affected classes

All litellm exception types: APIConnectionError, RateLimitError, Timeout, ServiceUnavailableError, BadRequestError, AuthenticationError, NotFoundError, ContentPolicyViolationError, InternalServerError, BadGatewayError, PermissionDeniedError, UnprocessableEntityError, APIError, APIResponseValidationError, ContextWindowExceededError.

Reproduction

import cloudpickle
import litellm

exc = litellm.exceptions.RateLimitError(
    message="rate limited", model="gpt-4", llm_provider="openai"
)
data = cloudpickle.dumps(exc)
cloudpickle.loads(data)  # TypeError: __init__() missing required positional arguments

Fix

Monkey-patch __reduce__ on litellm exception classes so cloudpickle reconstructs them via Exception.__new__() + __dict__ restoration, bypassing __init__. See agent_memory_server/litellm_pickle_compat.py.

The root cause is upstream in litellm — their exception classes don't implement pickle protocol methods. Reported separately there.