#35303 [Bug]: A single malformed tool_calls[].function.arguments DoSes a proxy replica — pre-network transformation error is mapped to retryable APIConnectionError, and the cyclic fallback graph is walked without cycle detection (79,858 attempts against num_retries=2, ~2.3 GB of logs in ~7 min, event loop starved, pod killed with exit 137)
### Check for existing issues
- [x] I have searched the existing issues and checked that my issue is not a duplicate.
### What happened?
> **Title (paste into the GitHub issue title field):** > > `[Bug]: A single malformed tool_calls[].function.arguments DoSes a proxy replica — pre-network transformation error is mapped to retryable APIConnectionError, and the cyclic fallback graph is walked without cycle detection (79,858 attempts against num_retries=2, ~2.3 GB of logs in ~7 min, event loop starved, pod killed with exit 137)` > > Suggested labels: `bug`, `bedrock`, `router`, `availability`
---
## What happened?
A client replayed a conversation history in which one assistant message contained a **truncated** `tool_calls[].function.arguments` JSON string. LiteLLM's Bedrock Converse transformation tried to `json.loads()` that string, failed, and raised a **bare `Exception`**. `exception_mapping_utils.exception_type()` classified it as **`litellm.APIConnectionError`** — a *retryable* class.
Because the failure occurs inside `_async_transform_request()`, i.e. **before any network I/O**, each retry costs ~0.56 ms of pure CPU. The router's retry + fallback machinery therefore dege…