When a usage limit cuts your subagents off, they pick up where they stopped. Nothing gets redone.

npx agent-limit-retry   # one question, then the plugin and status line tap are installed
alr selftest                     # simulated limit hit and resume against a local fake API, no quota spent
alr status                       # your Claude Code and Codex quota windows

alr init backs up ~/.claude/settings.json before editing it and keeps an existing status line. alr uninstall undoes everything.

If the limit hits while the main agent has parallel subagents running, every subagent fails at once. Each one still has its full context on disk and can be woken by id. After the reset the main agent almost never does that: it sends the same task again from scratch, paying quota for work that was already half done, or forgets the subagent entirely.

Claude Code's own auto-continue covers less than it seems. It continues the main conversation, in an interactive session, when the reset is under 24 hours away. Not the subagents, not weekly limits, not headless claude -p runs. Codex has no auto-continue at all, and none of its hooks fire on a limit error.

476

subagents cut off by a usage limit across 5 of the author's own sessions

449

never touched again after the cut: re-dispatched from scratch or forgotten

27

got a message afterwards, 19 of them while testing this tool

The data to fix this is already there. Claude Code writes a quotaLimits.resetsAt field with the exact reset time; Codex writes a token_count.rate_limits snapshot right before the failure. Each subagent's transcript says what it was doing, which files it changed and what its last steps were.

1. Watch the windowsClaude Code passes your 5-hour and 7-day percentages to the status line; alr saves them. For Codex a per-thread watcher reads the rollout file.

2. Hand off at 90%The agent is told to finish its current step and write HANDOFF.md: goal, done, in flight (including subagents), next steps, gotchas. Once per window.

3. Record the limitThe exact reset time is taken from the transcript and you get a desktop notification. Pro also maps the killed subagents from their transcripts.

4. After the resetFree: alr run resumes headless runs. Pro: the main agent is told to wake each subagent by id, and Codex gets a queued "continue".

In Codex the same message asks for followup_task on each child thread's path. In ZCode it asks for SendMessage, or ReadSessionContext plus a fresh agent if ZCode was restarted since. When the vendor already handled a case, alr says nothing.

• It cannot raise your limits or change any vendor's policy. It only helps you use the plan you already have.
• It never rotates accounts, shares sessions, or touches your credentials.
• Claude Code handles the quit-or-crash case itself since 2.1.270. alr stays quiet there and only speaks up for subagents Claude Code didn't report.
• Verified end to end on macOS. Windows support for Codex is new and comes from early testers. The ZCode adapter was written from ZCode's source and fixtures; live reports are still coming in.