The release that changed what context recovery is allowed to do, and the change is a philosophy.
Before: to fit the window, Pengy deleted older turns from the provider's view and wrote a summary of them. Your chat history was untouched, but the model's view of your conversation was a summary of the first half. That costs a large request (a full prefill of the history to summarize), the model loses exact detail, and a summary is not the conversation.
Now: tool results are reduced, lossily, as they age, and history is kept. The newest result goes in full. A result from 4 turns ago keeps 40% of its budget, head and tail, with a truncation marker. A result from 8 turns ago keeps 10%. Older becomes a short stub. The model's own narration, its tool-call requests and IDs, user messages and assistant text are never reduced, because those are the turns, not attachments to them. The budget is estimated on the outgoing payload, and the tool schema count is included, because tool schemas are ~18.5k tokens and the old estimate ignored them.
The summary runs in the background. It runs asynchronously, off the critical path, gated on history size (12 turns and 60k tokens) and time (6 minutes), coalesced, with a 60-second budget, serialised through a queue so two summaries cannot interleave, and it falls back to the synchronous path when it cannot run.
The model's own working notes survive. render_plan is a tool the model uses to write out its
plan; reasoning_content is what a reasoning model says about a tool call. Some providers require
them back on the next request. They are in the provider's view now, gated by provider capability and
never in the saved transcript. Without them, a reasoning model over tool calls loses the chain of
thought and behaves as if it is starting again.