Your agents
worked.
Get caught up.

Automatic preparation and compaction.

New installations automatically index enabled history, summarize new work and prepare a briefing while the server runs. Turn this off with Automatically prepare conversations in Settings if you prefer manual preparation. Compaction creates cached summaries; it does not edit or delete original agent logs. Matching summaries and ready briefings are reused, rather than starting over every time.

Track preparation spending estimates.

Settings’ Preparation usage dashboard shows estimated USD cost, completed batches and model requests. Choose a 6-hour, 1-day, 7-day or 14-day window and hourly, 6-hour or daily buckets. Labelled cost/batch axes and bucket details make values readable. Manual preparation, automatic work, retries and failures stay counted.

Missing prices stay unavailable or partial; older timing is approximate. These are not invoices, total spending or account quota. Chat, voice, standalone post-conversation recaps and local machine/network costs are excluded.

You choose when models can work.

Automatic mode is on by default for new installations and uses your configured summary and briefing models, including quota or API billing, within the saved invocation limit. Existing installations keep their previous behavior, including off or no saved preference. Provider failures pause for an explicit retry. Older local-index-only permission does not silently authorize model usage. No automatic microphone, call, saved decision or review. Opening the explanation or cost dashboard only reads metadata.

It won’t change your projects.

It doesn’t run project commands, edit your code, start coding agents or make Git changes.

Keep your original logs.

Links in the summary lead back to your agent logs. If a file is missing or changed, Draken tells you. It doesn’t keep a hidden copy of the original text.

Some content leaves your computer.

Reading and searching logs happen locally. For summaries and text replies, pi sends content to the model provider you use. If you use voice, your audio and some history go to OpenAI.

Reading isn’t reviewing.

Draken won’t treat an AI suggestion as your decision, or mark work reviewed for you.

Wrap a long line, not the whole batch.

Recap prose can now wrap at existing spaces without dropping words or qualifications. A line just over the width limit no longer needs a model correction if the result still fits five lines and 600 characters. Unbreakable tokens and other invalid output remain checked; retries are still bounded.

Know why a recap stopped.

Preparation metrics retain content-free validation reasons and line lengths, not rejected model text. If both responses fail validation, the error explains the failing rule. Older rejected responses cannot be reconstructed after the fact.

Unassigned is not the same as prepared.

Home distinguishes entries not yet in a sync from entries already held in an unfinished batch. A stopped batch shows its entry count and stage. Zero unassigned entries does not mean preparation succeeded. The same automatic settings, source permissions, model limits and explicit failure retries remain.

A quick question gets a short answer.

Chat defaults to one paragraph or up to three bullets, usually 60–100 words plus citations. Ask for a detailed breakdown when you want more. The same evidence and later-report checks remain; the length target is guidance to the model, not a hard cutoff or an extra model request. Opening briefings keep their own format.

Keep your existing history.

No new tables, columns or dependencies from 0.5.0. The default is stored only when initializing an empty database. Conversations, notes, review state, summaries and compatible search vectors are retained. Back up logs and Draken data, finish active work, stop the server, upgrade and restart.