Expand description
The session wake-up briefing and one-shot efficiency hints. The read-path drivers (session-start preload, related hints, silent import preload, prefetch, auto-dedup) were retired on 2026-10-17: an A/B on the retrieval corpus returned byte-identical responses and equal R@5/MRR with them off, while the session start cost 2-14 s more.
Structs§
- Autonomy
State - Tracks autonomous action state independently of the MCP tool layer.
- Search
History - Per-key stats for progressive search hints (
ctx_search/ctx_semantic_search).
Constants§
- AUTO_
CONTEXT_ MAX_ TOKENS - The largest
--- AUTO CONTEXT ---block a reply carries; larger blocks are dropped, so the briefing is fitted to it.
Functions§
- apply_
auto_ load_ packages_ once - Applies the auto-load context packages silently on the first call of a session that does not deliver the briefing; the result text is unchanged and the briefing names them later.
- delivers_
session_ briefing - Whether this call is the one that may carry the session briefing and the
shared-context hint:
ctx_sessionwithload,statusorshow, also throughctx_call. Every other tool result stays a deterministic function of its inputs (#498); the briefing carries per-run peer ids and memory. - session_
wakeup_ briefing - The wake-up briefing (last task, recent decisions, project facts, active
peers), as
--- AUTO CONTEXT ---, on the first briefing-delivering call of a session (seedelivers_session_briefing). - shell_
efficiency_ hint - Suggests a more token-efficient lean-ctx tool when shell compression is low.
- should_
auto_ consolidate - Returns true if enough tool calls have elapsed to trigger auto-consolidation.