Parallel subagents consume 642k tokens, then session ends mid-edit without handoff

Bug description

A single design-review request launched five parallel subagents. Their completion notifications reported 642,570 combined subagent_tokens within about ten minutes:

• 95,765
• 101,789
• 159,257
• 155,788
• 129,971

The parent later began an approved documentation edit. Immediately after a successful edit tool result, the next assistant message was only:

You've hit your session limit · resets 11:40pm (America/New_York)

There was no structured error, progress summary, saved handoff, or automatic continuation. Work resumed only after the reset, when the user manually sent Continue 3 hours 14 minutes later.

The subagents duplicated substantial work. All five independently read the same core documents, and five children plus the parent read the roadmap file before it changed.

Expected behavior

• Parallel subagent work should expose or enforce a shared usage budget before a large fan-out.
• The parent should show aggregate child usage while the agents are running.
• Approaching the session limit should stop new expensive work and preserve a usable handoff.
• If the limit interrupts a tool-driven edit, the UI should present structured continuation state rather than a raw assistant message.

Actual behavior

• Five children consumed 642,570 reported tokens without a visible shared budget or stop point.
• The parent continued into implementation.
• The session ended mid-task with only the raw limit text.
• The user had to wait for the reset and manually resume.

Reproduction outline

1. Start a design or repository review.
2. Dispatch five detailed parallel review agents over overlapping files.
3. Wait for all five results and continue the parent conversation.
4. Begin file edits near the resulting session limit.
5. Observe the raw session-limit message after a tool result and the absence of a handoff or automatic continuation.

Environment

• Platform: macOS
• Harness: Claude Code project session
• Exact historical Claude Code build: unavailable
• Model recorded in parent and child transcripts: claude-fable-5-1
• Session date: 2026-09-07 to 2026-09-08

I have a local evidence report with line-level transcript citations. I have not attached the transcript because it contains local paths and project content.