Context management

Context management is off by default. Turn on summarization when an agent needs to keep working through long conversations.

const agent = new AgentDock({
  model,
  contextManagement: {
    summarization: {
      trigger: "auto",
    },
  },
});

With trigger: "auto", Agentdock uses the primary model's context size. It starts summarizing at 75% and keeps the newest 25% of the conversation. Supported AgentDockModel models provide the context profile.

Choose a different summary model

contextManagement: {
  summarization: {
    trigger: "auto",
    summaryModel: AgentDockModel.openAI({ model: "gpt-5.4-nano" }),
  },
}

The summary model is optional. If you omit it, Agentdock uses the primary model.

Set your own limits

Use a token limit or message limit instead of "auto":

contextManagement: {
  summarization: {
    trigger: { tokens: 96_000 },
    keep: { tokens: 24_000 },
  },
}

maxSteps limits model calls in a run. It does not control conversation size. Agentdock keeps system instructions and recent messages when it summarizes. Your tool ctx is not added to the prompt.