Context, compaction and the task record
See what fills a long conversation's context and how the app compacts it at once, with nothing lost, using the record it keeps of each task.
A model only has so much conversation in view: its context. When it fills up, NoBreiniers compacts it: the old part leaves the model's view and the record the app keeps of each task stands in its place. You lose nothing: the chat keeps every message.
See what fills the context
- Under the message box, at the right of the token bar, press the context gauge with its percentage. You can also type
/contextand choose «Conversation space». - «Context» opens. At the top you read one sentence («Conversation space: … used. … summarizes by itself when it fills.») and a bar with how much room is used.
- Press «Technical details» (closed at first): «What fills this conversation's context, by part, and what it costs.», and how much is used and the point where it compacts by itself.
- There, each part with its tokens: «Instructions (system prompt)», «Owner and project rules», «Memory», «Skills index», «Tools», «Conversation», «Files read», «Other tool results» and «Pictures».
- Then what a request costs at that size, «Largest» and «Suggestions» for trimming.
The cost line also says how much of this chat's input came from the provider's cache and what it saved, like this: «… of this chat's input came from the provider's cache (… tokens), … saved». Each turn says its share «from the cache». NoBreiniers knows how each AI provider's cache works and uses it on every request, without touching the quality of the answers.
Compact
By itself. When the context reaches its point, the app compacts it with nothing for you to do. In the chat you see «Conversation shortened · the model continues from this summary». Open it to read what the model carries.
Whenever you want. There are three ways:
- «Compact now», in the «Context» window.
/compactin the composer.- The composer's «Add files, or type / for commands and @ for files» button › «Shorten the chat».
The task record
While the agent works, the app notes each task down from what really happened: what you asked, the files read and edited, the commands and tests with their result, the errors and what fixed them, the pages consulted and what the agent said along the way. No model writes it.
At a compaction, that record takes the old conversation's place:
- It is instant. You do not wait for a model to write a summary, and that call is not paid for.
- It says what happened. It comes from the chat's facts. Before the model reads it, the app checks it against the disk: if a file changed afterwards, or a test is from before the last edit, it says so.
- Nothing is lost. The conversation stays word for word in the chat's history. In the chat, the row «See … earlier messages» opens it; in «Steps», «See earlier history» brings the earlier steps.
- It works with any model. The record is yours and lives only in your profile.
When a summary is written, as before
The record covers work on files and commands. In these cases the app asks the model for a summary of the old part:
- The conversation is mostly talk, or pages and documents read (research).
- You type
/compactfollowed by a few words, to focus the compaction on that. - You have a
PreCompacthook. See Hooks.
For you it looks the same: «Conversation shortened · …», and the whole chat is still there.
Open the record
You do not need to open it to work. It is there for when you want to know what the agent remembers.
- Above the conversation, press «Steps».
- Press the «Task record» button, beside the number of events.
- Each task shows its status («In progress», «Done», «Stopped», «Failed» or «Cut off») and these parts:
| Part | What it keeps |
|---|---|
| «What you asked» | Your words, as you wrote them. |
| «How it ended» | The task's final answer. |
| «Files» | Each file read, edited, created or deleted, and whether it is still «as it was left», «changed since» or «no longer there». |
| «Commands and tests» | Each command once, with its latest result, and whether the test «ran after the last edit». |
| «Errors and what fixed them» | Each failure, «open» or «fixed», and what fixed it. |
| «Pages and sources» | Where what was consulted came from. |
| «Notes» and «What it said along the way» | What the agent left said on the way. |
| «Subagents» | The work it delegated. |
| «What the model carries» | The same record as the model reads it when the old conversation leaves its view. |
What else you gain
- The next step comes from the record. Beside «Next» you see buttons such as «Go on with: …», «Fix what fails in: …» or «Run again, the files changed since: …». See One step ahead.
- It picks up where it was. If the model's provider fails or the app closes in the middle of a task, the agent goes on knowing which errors are still open, which checks to run again and which files changed.
- Another chat finds it. A finished task is kept in memory. In a new chat you can ask in plain words, for example «what did we do about the sign-in?».
- Subagents hand theirs over. A subagent returns its record beside its answer.
What the agent has in view
- A new task starts with its own work. When your message does not continue the last task, the earlier tasks' work leaves the model's view and their record takes its place; their pages and searches stay as sources. Working on one social network no longer carries what was done on another.
- Long work on a website keeps what it read. The conversation is no longer compacted every few minutes while an agent studies a site, so it does not have to look again for what it had already seen. It is compacted near a price step or when the window fills. See Price steps.
- A command's or a tool's result comes in whole up to about 8,000 tokens (about 32 KB) when the model has a window of 128,000 tokens or more; on a local 32,000 model, about 2,000. It was 3,000 before, and the agent read the rest in pieces, each piece one more step that sent the whole context again.
- The app's skills go by name. All of them are named on a single line, so the «Skills index» (in «Technical details» of «Context») weighs much less; before, each had its own sentence and together they were more than a third of the first request of a new chat. When one is needed, Breini reads its description and loads it whole, as always. The agent's, the business's and the project's skills keep their description.
- After a checkpoint, what the app reports is read again; what you asked for stays. The date and time, the working folder, the branch, the model in use and what is connected are what the app says now, not the copy in the summary. A folder to work in, a branch, a model or a deadline that you asked for stays in force until you change it, and so does a branch the work made or an operation it left open.
If something fails
| You see | What to do |
|---|---|
| «Nothing new since the conversation was last shortened.», in «Steps» | The chat was just compacted and there are no new steps yet. «See earlier history» shows the earlier ones. |
| «One of your hooks did not let the context be compacted: …» | A PreCompact hook blocked it. See Hooks. |
| «This chat has no work on record yet.» | The record appears when the agent takes its first step. |
