mcp 0.5.0 - agent retry logic retries the wrong tool sometimes
tested mcp 0.5.0-patch2 yesterday and the retry logic is completely broken - when tool A fails the agent sometimes retries tool B instead saw this with a postgres tool (fails with timeout) and filesystem tool (succeeds). agent retries the filesystem tool 3 times even though it succeeded the first time. zero backoff, just spams the wrong tool
i hit this exact retry routing bug on mcp 0.5.0 patch 1 yesterday testing with postgres and stripe tools in our staging environment. the setup was: postgres tool queries user data, stripe tool charges the user based on query results. when the stripe tool fails (invalid card), the agent retry logic retries the POSTGRES tool instead of the stripe tool - so it just keeps re-querying the user table instead of fixing the payment issue. the retry target selection seems completely random or based on some internal state that doesn't match the actual failure point. honestly this breaks any multi-tool workflow where error recovery matters - you can't trust that the agent will retry the right operation
retry routing is broken across the board
lol of course it does. tested this exact thing tuesday - agent retries postgres tool but retries the stripe tool instead. zero dependency resolution, just vibes
retry logs show which tool failed. check those first.