Remote Control blocks Claude Code auto-continue after limit reset

The author's setup is aimed at a specific workflow: an Opus main session dispatching Sonnet subagents in parallel, running a hackathon project of 86 tickets. That run hit the session limit and did not auto-continue, so the author had to type "continue" manually at the 23:40 reset. The reported evidence comes from transcripts, where limit hits within 10 minutes were counted: 29 in total, with 10 of 10 failing to auto-continue while Remote Control was connected and 8 of 8 starting when it was not. One session that switched state shows the pattern cleanly — three hits before /remote-control started, three after did not. The author also says the bundled CLI code at version 2.1.289 returns immediately from the function that starts the wait when Remote Control is connected, while noting this is internal implementation that can change.

The workaround keeps the built-in auto-continue enabled and layers four hooks on top: Stop and StopFailure(rate_limit) run a resume step that waits for the reset and wakes the session with exit 2 under asyncRewake; UserPromptSubmit adds usage context and blocks the built-in continuation prompt at or above the weekly line; and a PreToolUse gate asks for confirmation before dispatching or continuing agents. Validation was two probes with real interactive sessions and 35 unit tests, though the author states the real 5-hour reset, behavior above 85% weekly, and a live rate_limit StopFailure were only exercised in unit tests with fake data, on Claude Code 2.1.289, macOS, and a Max plan.

The stakes here hinge on whether the built-in exclusion persists in future Claude Code versions and whether the hook path behaves the same outside the tested environment. For anyone relying on unattended Claude Code runs, the practical question is whether their sessions are silently attached to Remote Control — the author warns this can happen without the user ever typing the command — and whether their own guard rails are tested against a real limit reset rather than simulated data.