Keep long sessions affordable
Every reply re-reads the whole conversation. Yardmaster warns you when a session gets large, holds costly automatic replies, and can continue in a fresh session.
Every reply in a Claude session re-reads the whole conversation, so a very long session costs more with every turn. Yardmaster warns you when a session gets large and offers to continue in a small, fresh session that knows where things stand.
For a short while after each reply, the AI keeps the conversation handy, so the next reply is cheap. After that (usually about 5 minutes, sometimes an hour) the next reply re-sends everything at full price. On a large session, that one reply can cost dollars.
Continue in a fresh session#
Once a session passes 250,000 tokens (a measure of its length), a banner at the top says how large it is. It offers two buttons:
Fresh session with handoff writes a short note about what the session was doing, where it got to and what is left. It starts a new session in the same project with that note as its first message, archives the old one, and opens the new one.
Dismiss hides the banner. It comes back once when the session is twice as large.
The handoff waits until the session has finished its current work. If the new session cannot start, the old one stays exactly as it was.
Why is my session waiting with "Resume anyway"?#
Yardmaster sometimes sends a message for you: a scheduled message, a drip, a nudge to keep going, or a message from another agent. If that would wake a large session (over 300,000 tokens) that had finished and sat idle long enough to need the full re-read, Yardmaster holds it instead of paying quietly.
The session turns Needs You with "Waiting: large idle session", and one card appears in your Inbox. Choose:
Resume anyway sends the waiting message now and pays for the full re-read.
Fresh session with handoff continues cheaply in a new session, as above.
A message you type is never held. If you type into a held session, your message goes out, and the waiting automatic one is dropped with a note in the chat that quotes it.
Change or turn off these checks#
Both are on at first. Change them in Settings > Sessions:
To see what large and idle sessions have cost you, open Cost by cause in Statistics.
A spending stop for API-key sessions#
When a session bills a key for every message (an Anthropic API key, or a DeepSeek, Kimi, GLM, MiniMax or Meta key), a runaway session stops on its own. In Settings > Sessions > Spending limits, Budget Caps (per session) warns at $20 and stops the session at $100, and Daily limit (per account) warns at $10 and stops at $25 a day for each Anthropic API key. Sessions on your Claude plan are never stopped by these limits. More in Choose your AI engine.
Claude only. Sessions on other engines never get the size banner or a hold.
Restart counts as automatic. Restarting a large, finished, idle session is held the same way.
Archiving a held session ends the hold and removes its card.
Several automatic messages waiting on the same session join one hold, so you get one card.
Compacting (when a session summarizes itself and drops older messages) hides the size banner until the session grows again.
Related
- See what your agents cost and how you workStats shows what your sessions and AI helpers cost, how much time they took, and how you use Yardmaster, for any time range.
- Choose your AI engineClaude runs your sessions by default. Turn on other engines, such as Codex, Gemini or DeepSeek, and pick one per session or per project.
- Approve what your agents ask forAgents stop and ask before risky steps. Allow or deny them, approve requests in the Inbox, and choose how often they ask.