Attention slider: micro management to yolo

Not every layer costs the same. Ingestion, redaction, the structural tree and the search index are cheap and incremental, so they always run near realtime and the data is always complete. Semantic analysis through claude -p is the expensive layer: seconds per call, thousands of tokens, and it competes with your own session for the same subscription rate limit. Re-summarizing a part of the session that is still growing, after every turn, costs tokens again and again.

The slider therefore decides only which summaries the engine writes ahead of time. It never affects ingestion, and it never limits what you can read: any summary the slider skips is written when you open that node in the reader.

The four modes

A part of the session is closed once newer work follows it, or once the session has been idle for 30 minutes. The newest turn and the newest phase are open while the session runs.

Mode	Written ahead of time	Lag of the summaries	Rough cost
Micro management	Title and abstract of every turn and every phase, and of the session, including the open ones, refreshed as they grow	seconds	≈ 3 calls per turn
Attentive	A title for every closed turn; title and abstract for every closed phase	about one turn	≈ 1 call per turn
Balanced (default)	Title and abstract for every closed phase	about one phase	≈ 1 call per 4 turns
Yolo	Nothing	on demand	only what you open

Phases are split structurally, without a model: at compactions, after long idle gaps, and at prompts that change the topic.

Design rules

• The reader is the demand signal. Whatever the mode, opening a node that has no summary yet asks the engine for one. The slider only decides how much is ready before you look.
• Short text is never summarized. A one-line prompt or a single tool call is already as short as a title, so it is shown as it is, without a model call.
• Same input, same summary. Every summary is keyed by the hash of its input text, the granularity and the analyzer. A closed phase has the same input in every mode, so its summary is identical whether it was written eagerly or on demand, and it is only written once.
• Open parts stay local. Summaries of turns and phases that are still growing go to the gitignored .yolo-twice/cache/, never into the committed sessions/ folder.
• The cost is shown next to the slider, as an estimate of summarization calls per hour for the current session, together with the expected lag.
• The setting is personal. The team default is attention in .yolo-twice/config.json (balanced). Your own choice lives in the gitignored .yolo-twice/local.json and can be changed live from the UI or with /yolo:twice mode <micro|attentive|balanced|yolo>.
• The model is configurable. All summaries use the model set in analyzers.model in config.json (default: haiku), whatever the mode.

Catch me up

The "Catch me up" button summarizes everything that happened in the session since this browser last showed it. It is most useful in yolo mode: let the agent work, come back, and read one paragraph instead of the whole transcript. The position is remembered per session in the browser, so each viewer gets their own "since you last looked".