Tiers

A tier is how hard Tokand shrinks the request. You still call the same model. Set tier on the chat request. If you omit it, Tokand uses measured. On decide(), omitting tier leaves the state unchanged. Set it to shrink state.messages before that decision.

• safe

  Safe · noise only

  Drops repeated and leftover noise. The latest user message, system prompt, and tool results stay as you sent them.

• measured

  Measured · tool bodies

  The default. Shrinks bulky tool traces before the provider bills them, and leaves the request you care about intact.

• aggressive

  Aggressive · names only

  Keeps tool names and cuts the rest of the trace. Use it on hot agent loops where the body of old tool output is not needed.

{
  "model": "gpt-4.1",
  "tier": "measured",
  "messages": [...]
}