Usage-limit recovery: resume with the original task, show Resume for limited subagents, fail over to another configured account

[BlackPhoenixSlo]

Problem

I run long Claude threads that delegate work to subagents, with Auto-resume limited threads turned on. When the session limit hits, three things go wrong.

1. The auto-resume prompt drops the task

When the limit resets, UsageLimitRecoveryWorker sends a fixed message, text: "Continue where you left off." (apps/server/src/orchestration-v2/UsageLimitRecoveryWorker.ts). In practice the agent picks the work back up for about two turns and then loses the goal, because the cache is cold and the original brief is far back in a long context. After that it drifts or stops, and I have to restart the task by hand.

Ask: make the resume message carry the thread's original task. For example, re-send the first user message of the thread, or of the run that was limited, with a short note:

You were interrupted by a usage limit. The original task is below. Check the current state of the workspace and continue from where you left off; don't start over.

A setting or template for the resume text would also work: one option for the current bare message, one for "original prompt + continue note", and one for custom text. If the user steered the thread after the first message, including the latest user message as well as the first one would cover that.

Related: #17854 (keep a small task brief across limit-triggered switches), #15128 (auto-fork with distilled context on quota cutoff), #17997 / #12825 (resume without injecting a bare "continue").

2. No Resume control when a subagent is the one that hit the limit

When a delegated subagent (child thread) hits the limit, the parent thread has no Resume / Resume-at-reset control once the limit resets. The work just sits there until I notice it and tell the parent by hand to restart the child.

Ask: when a subagent stops on a usage limit, the parent should show a resume control for it, or the subagent should be armed for auto-resume like a top-level thread, using the same resume text as in point 1.

Related open reports on the same path:

• [Bug]: Claude usage-limit reset shown in activity but missing auto-resume controls during subagent workflow #15665: reset time shown in activity, but no auto-resume controls during a subagent workflow
• [Bug]: Usage-limited Claude thread is never armed for auto-resume when background-task wake runs finish as completed right after the limit stop #16815: a limited Claude thread is never armed when background-task wake runs finish as completed (fix proposed in fix(server): background wakes no longer block usage-limit auto-resume #20804)
• [Bug]: Usage limit banner collapses behind the "Waiting on subagents" banner #18918: the usage-limit banner collapses behind "Waiting on subagents" (fix proposed in fix(web): urgent and warning notices no longer hide behind activity banners #18919)
• [Bug]: Nested Claude subagents stay Running in the UI after their parent subagent hits the usage limit, while the thread list reports them idle #19993: nested subagents stay "Running" after their parent subagent hits the limit
• [Bug]: A Claude subagent resumed after hitting the session limit keeps a Failed status and stops showing live updates #20920 (closed): a subagent resumed after the session limit keeps its Failed status
• Choose which usage-limited parent and delegated tasks resume at reset #20209: choosing which limited parent and delegated tasks resume at reset

3. Fail over to another configured account

If I have a second Claude provider instance (another subscription), a limit on one account should be able to move the thread to the other one automatically instead of waiting hours for the reset. This could be opt-in, with an ordered fallback list per provider and a visible handoff marker in the timeline. On failover, the incoming account should also get the original task (point 1), not only "continue".

Related: #18949 (continue a thread on another account when it hits a usage limit), #15846 (automatic hand-off to a fallback provider), #16512 (fallback provider for scheduled tasks), and PR #17753 (auto/specific fallback on usage limits; closed in the PR cleanup, not on its merits).

Summary

1. The auto-resume message should include the original task, not just "Continue where you left off."
2. Subagents stopped by a limit need a Resume control in the parent, or auto-resume.
3. Optional automatic failover to another configured account of the same provider when a limit hits.

Point 1 seems the smallest and has the biggest effect on unattended runs. I'm happy to test any of these.