Context compaction for the OpenAI Responses API (OpenAI and Azure OpenAI).

Compaction asks the service to fold a long conversation into one or more opaque compaction items that carry the essential prior state in far fewer tokens. Those items come back as :provider_block content parts on the assistant message of the returned response and are replayed verbatim by the Responses encoder on the next request.

Usage

{:ok, first} = ReqLLM.generate_text(model, "Draft a landing page.")

{:ok, compacted} = ReqLLM.compact_context(model, first.context)

next = ReqLLM.Context.append(compacted.context, ReqLLM.Context.user("Add a booking form."))
{:ok, follow_up} = ReqLLM.generate_text(model, next)

compacted.context holds the complete returned window in one message. Its metadata.responses_replay preserves all API items, including retained user messages, assistant messages, and tool items, in their original order. Do not prune this window: the next request must replay it as returned by the service. Append the next user message to continue. The service can also compact a stored response instead of a replayed context:

{:ok, compacted} = ReqLLM.compact_context(model, nil, previous_response_id: first.id)

Server-side compaction on ordinary requests is enabled with the context_management provider option; the resulting compaction items land on the response message the same way and replay automatically.

The compacted message keeps the compaction response id under metadata.compaction_response_id rather than metadata.response_id, so the next turn replays the compaction items instead of chaining through previous_response_id.