Runaway duplicate turns reread ~500k-token context and exhausted 5-hour usage limit in minutes

Summary

A single short user prompt in an existing long Claude Code conversation triggered a rapid series of duplicate/repeated Opus 4.8 turns. Each turn reread roughly 500k cached input tokens. The account's five-hour usage allowance was exhausted in under ten minutes.

Environment

• Claude Code / Claude desktop integration on macOS
• Model recorded in the session log: claude-opus-4-8
• Incident time: July 12, 2026, approximately 5:21–5:25 AM CDT
• Paid Max subscription
• Conversation had accumulated a very large context before the incident

Observed behavior

Local Claude session telemetry shows:

• User prompt length: 212 characters
• Approximately 9.3 million processed tokens in the 5:20–5:29 AM CDT window
• Most usage was cache_read_input_tokens, not new prompt input
• Individual assistant turns repeatedly read approximately 506k–522k cached tokens
• Several assistant responses/tool turns were emitted within seconds of one another
• The repeated sequence included Bash, Write, Skill, and Workflow turns
• The five-hour usage limit was consumed within minutes

Examples from local usage records:

• cache_read_input_tokens: 521,829 on repeated assistant turns
• cache_read_input_tokens: 518,749 on repeated assistant turns
• cache_read_input_tokens: 517,884 on repeated assistant turns
• Ten-minute aggregate: approximately 9.3 million tokens

The prompt itself was ordinary UI feedback and did not request a loop, parallel agents, or repeated execution.

Expected behavior

Claude Code should:

1. Deduplicate or serialize turns originating from one user prompt.
2. Stop repeated tool/workflow execution when no new user input has arrived.
3. Warn or compact before repeatedly submitting a context around 500k tokens.
4. Apply a per-prompt/token circuit breaker so one prompt cannot consume an entire five-hour subscription allowance.
5. Expose which workflow/subagent caused the usage spike.

Impact

• Entire five-hour Claude usage allowance exhausted in minutes.
• Local resource pressure increased because repeated agent/tool processes were launched.
• User had to terminate processes and abandon the affected conversation.

Requested investigation

Please investigate whether Workflow execution, queued turns, or duplicate event handling can submit the same large cached context multiple times from one prompt. Please also consider restoring the usage allowance consumed by this incident.

I can provide sanitized session timestamps and usage records privately if needed, but I am not posting the full conversation log publicly because it contains business context.