The LMD GHOST consensus protocol is a critical component of proof-of-stake Ethereum. In its current form, this protocol is brittle, as evidenced by recent attacks and patching attempts. We propose Goldfish, a new protocol that satisfies key properties required of a drop-in replacement for LMD GHOST: Goldfish is secure in the sleepy model, assuming a majority of the validators follows the protocol. Goldfish is reorg resilient so that honestly produced blocks are guaranteed inclusion in the ledger, and it supports fast confirmation with expected confirmation latency independent of the desired security level. Subsampling validators can improve the communication efficiency of Goldfish, and Goldfish is composable with finality/accountability gadgets. Crucially, Goldfish is structurally similar to LMD GHOST, providing a credible path to adoption in Ethereum. Attacks on LMD GHOST exploit lack of coordination among honest validators, typically provided by a locking mechanism in classical BFT protocols. However, locking requires votes from a quorum of all participants and is not compatible with fluctuating participation. Goldfish is powered by a novel coordination mechanism to synchronize the honest validators' actions. Experiments with our prototype implementation of Goldfish suggest practicality.