Prompt cache TTL too short for neurodivergent users and complex workflows

Problem

The 5-minute prompt cache TTL expires while users are reading and processing Claude Code output. When the cache expires, the next interaction requires a full context re-read, which:

1. Adds 1-2+ minutes of "cogitation" time before the model responds
2. Consumes additional rate limit budget re-processing the full conversation
3. Punishes users who need more than 5 minutes to read, process, and think about complex output

Who this affects

• Neurodivergent users (ADHD, autism) who may need more processing time for complex technical output
• Hardware/firmware engineers reviewing detailed specs, wiring guides, or diagnostic data before responding
• Anyone working with long conversations where the re-read cost is significant

Real-world impact

In a 324-hour usage month across 240 sessions, a significant portion of rate limit consumption comes from cache misses during normal reading pauses. The user isn't idle — they're reading a spec, checking a datasheet, or processing diagnostic output. The 5-minute window doesn't match the actual workflow.

Request

Either:

• Make the cache TTL user-configurable (e.g., 5m / 15m / 30m / 1h options)
• Extend the default TTL to something more reasonable for complex workflows (15-30 minutes)
• Add a "keep-alive" mechanism that doesn't require sending a message (e.g., a keystroke or UI indicator that the user is still engaged)

Current workaround

Sending a throwaway message ("still reading", ".") before the 5-minute window expires. This works but it's disruptive and easy to forget.

---
Filed via Claude Code on behalf of a user.