Skip to content

SUPPORTING KNOWLEDGE · 0010

Context, compaction and cost

A model can only hold so much of a conversation at once: the context window, measured in tokens. Berd shows how full it is in the chat toolbar as a ring and a count, {tokens} / {limit} tokens used, with how much is left. When a session runs long, the window fills with earlier messages, tool output and file contents, and something has to give.

Compaction is what gives. Older context is summarised so the window has room again; the transcript keeps your messages and a notice says Conversation compacted. Older context was summarized. You can compact on demand from the toolbar, and for Goose sessions Berd can compact automatically at a threshold you set in Settings (100% turns auto-compaction off). Auto-compaction is labelled Goose-only; other harnesses manage their own windows.

Cost is the third number. When the setting Show session cost is on, the toolbar can show an estimated running cost for the session, described as displayed "when available". That phrase carries a lot, and the sources do not unpack it.

Why it matters to you

Compaction is lossy by design. The summary keeps the shape of what happened, not every detail, so an agent after compaction may no longer have the exact wording of a source you pasted an hour ago. The moment to check is when the notice appears.

There is also a failure notice worth recognising: Couldn't verify the compacted conversation because refreshed history wasn't received. Your previous messages are still shown. Try reloading the session. Berd is telling you it summarised but could not confirm the result matches what the harness now holds. Your transcript is intact; the agent's view of it may not be.

Cost is the number a consultant needs and the one this page can say least about. What the sources establish: it is an estimate, per session, shown when available, switched on in the behaviour settings. What they do not say: how it is calculated, which providers report it, whether it includes tool calls and compaction, and how closely it tracks the invoice. Until verified in the app, the toolbar figure is a trend, not a bill.

Switching model mid-chat is allowed and is one lever on cost. One documented limit: a chat whose earlier reasoning history is no longer in a form a Claude model accepts cannot continue with Claude; the notice says to start a new chat or switch to a non-Claude model.

How to apply it

Watch the ring, not the clock. When it is high, compact deliberately at a natural break rather than letting the threshold pick the moment for you.

After any compaction, re-anchor. Re-attach or re-mention the file that matters most; @mentioning a path costs little and removes the doubt.

Fork rather than push on. A long session that has compacted twice is a candidate for starting a fresh chat with the conclusions, especially before delegating anything.

Turn cost display on and treat it as a comparison between sessions and models, not a ledger. Compare it with the provider's own usage report once before you quote a number to anyone.

Compaction is a harness behaviour Berd surfaces, which is why the auto-compact setting names Goose. See App vs harness.

If you ignore this

The agent confidently cites a detail that was in the summarised part of the conversation, slightly wrong, and nothing flags it. Or you promise a client a per-task cost from the toolbar and discover the estimate did not cover what you thought.

Examples

A meeting that has run three hours. Someone reads back a summary of the first two so the room can go on. Useful, and also the moment to say "the figure was 14, not 12".

A review session at 85% context. You compact at the end of section two, re-mention sources/interview-3.md, and continue. The notice appears once, the ring drops, and the next answer still cites the right interview.

Reading session cost $0.42 and telling the client that a review costs 42 cents. The number is an estimate shown when available, with no documented model behind it.

Do it: Run a working session

Last checked: Berd v0.6.4, 2026-09-11.

Reference

Verified
Berd v0.6.4 · 2026-09-11
NEVER BESTUCK AGAINCLICK ME

Guide built with Berd, Claude and Codex · verified against Berd v0.6.4, 2026-09-11 · read from source, not yet tested in the app · Buzz Guide · About · not affiliated with Block