Token Tunnel keeps the prompt cache with stacked accounts

If you connected two or more accounts of the same provider to the Token Tunnel, for example two Claude subscriptions, the calls inside one conversation took turns between the accounts. A provider keeps its prompt cache per account, so each call landed on an account that had not seen the previous one and processed the whole conversation again from scratch. Long agent runs paid for that on every step with slower replies. Now a conversation stays on the account it started on, and every call reads the conversation so far from the cache.

How it works

• One conversation, one account. The Token Tunnel recognizes a conversation by its opening and sends each of its calls to the account its first call went to.
• Conversations still spread out. Each new conversation goes to the next account in turn, so stacked subscriptions still share the work.
• Limits still move you on. When the account a conversation is on reaches its usage limit or is signed out, the next call moves to another connected account, and the conversation stays there.
• Measured end to end. In a run with twelve parallel lookups on two Claude accounts, the second call used to read nothing from the cache. Now every call reads at least 90% of what the call before it sent.
• Desktop and hosted alike. Hosted subscriptions get it with this update, and the desktop app with its next release.