mcp 0.5.0 - tool error recovery just retries infinitely with same inputs
tested error recovery on mcp 0.5.0 yesterday with a stripe tool that fails on invalid card number format. agent sends malformed card number {cardNumber: "1234-5678-9012-345"}, tool returns validation error, agent retries with EXACT same malformed input 14 times before we killed it. zero backoff, zero input variation, just infinite retry loop with the same broken data. does error recovery even work or is this a known issue? 🤔
we hit this exact failure mode with a postgres tool that does bulk inserts. tool fails because of connection timeout, agent retries with identical params 14 times before we added manual circuit breaker logic in our mcp wrapper. the retry logic doesn't inspect the error type at all - treats transient rate limits the same as permanent auth failures
lol this is exactly what we saw in staging last week with mcp 0.5.0 and a github api tool. tool fails because of rate limit, agent just retries with the exact same params 18 times in a row until we killed it. no exponential backoff, no param variation, just brute force retries. feels like the retry logic doesn't even look at the error response from the tool
does the retry logic at least check if the error is retryable (rate limit vs auth failure) or does it just blindly retry everything....
it doesn't check. blindly retries everything including auth failures.
tested with github api tool that sometimes fails (rate limits, network timeouts). agent just retries with the exact same inputs over and over - saw it retry 23 times with identical params before i killed it. does the retry logic have any backoff or input variation at all?
23 retries.... does it at least exponential backoff or just spam instantly
lol 23 retries with identical params is insane.... does it at least log the retry count somewhere or do you just have to watch your tool execution logs to notice this is happening
lol of course it does. does the retry count get exposed anywhere or is it just invisible until you check logs? need to understand if this is even observable for monitoring