Context management

ContextPolicy controls how much work stays inline in a deep-agent conversation. Its defaults limit iterations, inline tool-result characters and estimated tokens, and retained history characters and estimated tokens.

When a tool result exceeds either inline limit, the harness writes the complete result under context/tool-results/ in the current workspace, returns a short preview to the model, and emits ContextOffloaded. The model can then use read_file to inspect the complete result only when it needs it.

var policy = new ContextPolicy(
        30, 16_000, 100_000,
        4_000, 25_000,
        128_000, 1_000);

var agent = DeepAgent.builder(model).contextPolicy(policy).build();

Set contextWindowTokens when the deployment’s actual context window is known. The effective history budget reserves room for a subsequent completion. Keep the configured workspace writable when offloading may occur.

Before every model turn, the harness compacts history. If a provider nevertheless reports a context-window overflow, it asks the model for a conversation summary and retries the original turn once with that compacted context. Other provider errors are propagated normally.

PromptCachePolicy is enabled by default and adds Cortavyn’s cache marker to static system, memory, and skill context. Provider adapters decide whether and how to use that marker. Call promptCachePolicy(PromptCachePolicy.disabled()) for a deployment where static-prompt caching is undesirable.