ADR 0037 — Chat context: explicit window, budgeted replay, compaction (v56-F1/F2/F3)
Date: 2026-07-18 · Status: accepted
Context
The Queen's turn loop resent the ENTIRE transcript every round — every
tool result as full verbatim JSON, forever — with no cap of any kind,
and never set num_ctx, so ollama truncated silently at its own small
default. skep's fixed floor (72 tool specs ≈ 38 KB + up to ~18 KB of
assembled system block) is ~14k tokens before one word of dialogue: on
a default window the model was blind-truncated from turn one, with no
error surfaced and a UI meter that measured none of it (it counted
thinking, which is never sent, against an arbitrary 24000-char divisor).
Field words: "fix the context on per chat."
Decision
-
The window is explicit.
llm_num_ctx(default 16384, floor 1024) rides every ollama chat call asoptions.num_ctx; openai-compat servers manage their own window and ignore it. -
The store is the audit trail; the replay is a view. Transcript rows are never truncated, edited, or deleted by context management. Only what is RESENT to the model each round is bounded: chars ≈ tokens × 4,
budget = max(8000, window·4 − system block − tool specs − 8000 response reserve), measured — not estimated — at assembly time. -
Prior-turn tool results replay capped (2000 chars + an honest marker naming the transcript); the current turn's tool results stay full — they are working data, not history.
-
Overflow compacts deterministically. Oldest replay overflow (never the newest 6 messages) folds into a per-chat digest (
chats.context_summary, cursorchats.compacted_through): one condensed line per user/assistant message, tool bursts counted, digest capped at 4000 chars dropping its oldest lines. No model calls — compaction works offline and is testable byte-for-byte. The digest rides the system prompt as "Earlier in this conversation (compacted)", followed by a line saying it is condensed, that its oldest lines drop once it is full, and thatsearch_chatswith this chat's id finds what was said (ES-F77: a 60-turn field test lost turn 2 and the model never knew). -
The meter is server truth. Chat detail carries
context(window, floor, history, budget, percent, compacted) computed from the same functions the replay uses; the composer renders it and stops guessing.
Consequences
- Long chats keep the thread: newest turns verbatim, older turns as a digest, the model never silently loses the start of the conversation.
- A replayed tool result can be shorter than what the tool returned —
the marker says so, and
get_chat_messages/the transcript always hold the full data. - The digest is lossy by design; if field tests show it loses the thread, an LLM-written summary can replace the line format behind the same columns (the trigger, not the default).