Compaction

Compactor defines the async compaction seam. PruneThenSummarize is the built-in implementation; configure it with CompactionConfig and pass it as Runtime(compactor=...).

Compactor.compact(ctx) -> list[SessionEvent] remains the custom-compactor contract. TurnContext.sample_request contains the complete prepared request. Returned events are appended before the request is rebuilt.

CompactionConfig fields and defaults:

Field	Default	Meaning
buffer	4_096	Safety reserve below the hard context ceiling
prune_pool_min	40_000	Minimum candidate tool-output pool before pruning
prune_gain_min	20_000	Minimum estimated gain before pruning
tail_turns	2	Preferred whole-turn minimum, subject to the token cap
summarize_at	0.95	Post-compaction target relative to the trigger
trigger_at	0.80	Context-window ratio that triggers compaction
recent_tokens	20_000	Maximum estimated recent-history window
summary_max_output	4_096	Maximum summary-generation output

The effective trigger is min(floor(context_window * trigger_at), context_window - max_output - buffer). Estimation covers the full SampleRequest; prior provider usage is an additional lower bound.

PruneThenSummarize preserves valid tool pairs, never prunes skill output, folds the prior summary into the next summary, and writes CompactionApplied to the current actor journal. The event selects a summary and retained floor during future assembly without rewriting history. Prompt blocks, execution-environment blocks, and tool schemas are not journal-derived and are not changed.

The current input must fit with the fixed request payload. Post-compaction requests at or above the trigger fail. An explicit pre-output context overflow can force one additional compaction and retry; no other provider error or partial stream is retried through this path.