Context
The ring in the header shows the estimated prompt size against the model's window — counting messages, attachments, tool definitions and call arguments, so it is the honest number, not just the visible text.
What happens as it fills
- Stale tool results and older attachments are trimmed (stubbed) automatically after a few turns, so the window does not clog up.
- When the limit is near, older turns fold into a summary. The summary can be expanded or undone.
- Compaction runs in chunks, with timeouts, and is limited by an automatic cap you can change — so a big history never stalls the server.
Thinking effort
Thinking effort (none, low, medium, high) is per chat — and per agent, set in Settings. Raise it for the hard question, lower it for the quick one; the cost is time and tokens, and the ring shows the result.
The stats panel
Your average speed, and the server's cache, memory and thermal state — so you can tell what a slow reply is doing before you do.
Reading ahead
A model server skips text it has already read, as far as a request matches the last one from its first word. Trimming and compaction change earlier text on purpose, so after either one the server would read everything from that point again at the start of your next message, while you wait.
TrailSail does that reading in the pause. When a reply ends and something earlier is due to change, it sends the next request's opening to the server to be read, asking for nothing back. Your next message then starts almost at once: on a 32 GB Mac mini with a 27B model, 12,000 tokens took 36 seconds to read cold and under half a second once read.
- It is sent once per reply. If you go and make a coffee, nothing more is sent until the next reply.
- It is not sent while a reply is being written, in this chat or another: reading slows writing to about a third.
- It is skipped when there is too little to read to be worth it, which is most turns.
- It works with Splash. Turn it off in Settings → Context → Read ahead after a reply.
Found a problem, or is something here unclear? Report it.