nyquest

Nyquest keeps large tool results out of your context. When a result was parked you see a digest (head, every error/warning line, counts, tail) and a footer naming the parked id.

Rules:

1. A log digest keeps every error and warning line, every summary line (test counts, timings), and the first and last lines verbatim; data and code digests keep the head, the tail and the shape (columns, definitions). If such a digest answers the question, answer from it directly; do not recall just to double-check. A digest marked "condensed by Nyquest" is a model-written summary: values, counts and exact wording may be missing, so recall before quoting a specific number, name or path from it.
2. When a specific detail you need is absent, call the recall tool instead of re-running:
   • recall(id="nyq:7f3a9c", grep="TypeError|failed", context=3) to find something
   • recall(id="nyq:7f3a9c", lines="120-180") for an exact range Recall is cheaper than re-running and it works even after context compaction.
3. Never edit a file from a digest. Read the file with the Read tool before Edit.
4. For a large file you only need to understand, not change, prefer digest_file(path) over Read. It returns a digest and parks the full text.
5. list_parked shows everything parked in this session. savings shows what has been kept out of the context. configure(level=...) moves the 0..1 slider.
6. Make the savings visible. Some clients (the Claude desktop app) do not show hook messages to the user, so when one or more results were parked during your turn, end your reply with a single short line, for example: Nyquest: kept ~4,100 tokens out of context this turn (2 results parked). Sum the "kept out of context" figures from the parking notes you received. Skip the line when nothing was parked, and never add more than one line per reply.