Account Pooler: 30-minute session pin is shorter than the 1-hour prompt cache #5262

Summary

The Account Pooler forgets which account a session used after 30 idle minutes (AFFINITY_IDLE_TTL_MS = 30 * 60 * 1_000, plugins/account-pool/src/hub.ts:57). Since #5130, pooled Claude Code threads use the 1-hour prompt cache. A prompt cache belongs to one account, so a thread that pauses for 30 to 60 minutes can be routed to another account and must write its whole context again, although its cache on the first account is still alive.

Why it matters

Orchestrating threads often wait 30 to 60 minutes on a child thread, a CI run or a person, and they carry large contexts. On one host with the pooler on (bb 0.45.0, so still before #5130), we measured 619 pauses of 5 to 60 minutes across Claude threads in two days; the median request was about 225,000 prompt tokens. With the 1-hour cache in place, every pause in the 30 to 60 minute range that lands on a different account still costs one full context write.

Expected

A session stays on its account for at least as long as its prompt cache can live (60 minutes for the 1-hour cache, plus a margin), unless that account is exhausted or removed.

Suggested change

Raise AFFINITY_IDLE_TTL_MS above the cache lifetime (for example 65 or 70 minutes), or derive it from the cache lifetime the pooled thread uses.

Notes

• We have not measured this after Restore the 1-hour prompt cache for pooled Claude Code threads #5130, because the installed stable release does not contain that fix yet and we keep the pooler off until it does. The reasoning follows from Anthropic's caching rules (caches are isolated per organization; a read refreshes the lifetime).
• A public report describes the same behaviour from a user's side: "it forgets which account a session was using after 30 mins" (https://x.com/kevinkern/status/2108559704263721121, in the replies).