The overnight batch does not die from a bad model, it dies when an MCP server stops answering

The design detail worth stealing is where the decision gets made. The hook attaches to two Claude Code events, documented in the script's own header comment: PreToolUse, which probes the MCP server before the tool runs, and PostToolUseFailure, which marks the result ^[13]. The probe is a plain HTTP check, and its outcome routes to one of three actions: block immediately, retry after backoff, or run a reconnect command and probe again ^[7]. The verdict then goes to `~/.claude/mcp-health-cache.json`, and on the following turn the hook reads that file and refuses the call outright until `nextRetryAt` has passed, without spending a probe to confirm what it already knows ^[8][10].

That last part is the whole trick. A long session gets its conversation history compacted, and a note from earlier saying a server was down does not survive into the compacted context, so the agent goes back to calling something it has already established is unhealthy ^[9]. The author says he tried the obvious alternative first, a system prompt instructing the model to try another approach on error. It handles a single error acceptably. When failures arrive back to back, the model spends heavily on trying, and by the next turn it is aimed at the same dead server again ^[11].

The status-code half of the problem is separate and just as expensive. He describes a 429 that stopped the run, a re-call thirty seconds later, another 429, and a loop with no exit ^[5]. At that interval a stuck loop makes 120 attempts an hour against a server that has already said no ^[16]. Elsewhere a 503 was treated as if it were a 401, so the run went off and performed a re-authentication that nothing had asked for ^[6]. MCP's transport layer will retry, but it does not read status codes and switch strategy, and 401 or 403 wants credentials rather than patience ^[12]. Anyone assuming the protocol covers this is buying retries, not policy.

Read the sourcing honestly: this is one operator's account of his own tooling, and the supporting evidence is autobiographical. He puts the environment at 1.2M yen a month in revenue and says time spent on mechanisms that stop things from stopping returns more over the long run than code written to earn more ^[14]. At 600K a month, half of the current figure, he believed the opposite, that more tasks meant more income, and revised it after a layoff took the business to zero and he rebuilt ^[15][17]. There is no outage frequency here and no before-and-after completion rate, so the claim that no overnight batch has been killed by an MCP outage since is his to make and yours to test ^[2].

What generalises is not the script. It is the observation that an agent's memory of its own failures is stored in the least durable place available ^[9], and that the fix is a file ^[18].