How Compaction Works in Pi

If you've ever had a long coding session in a coding agent like Pi, Claude Code, or Codex, you will have triggered a compaction. That is because large language models (LLMs) have limited context windows.

The context window is what the model can "see" while producing a response. Transformer architecture limits how much input an LLM can process. The input for a coding agent session includes all the previous messages and tool calls, and this keeps growing as you work. Once it exceeds the context window, the LLM rejects the request.

In this post, we will discuss when compaction is needed, and how it works in Pi. Compaction is also a useful tool for managing the size of the context window, which helps reduce both the cost of LLM requests and context rot.

An LLM conversation

When working interactively with a coding agent like Pi, the agent sends requests to an LLM and receives responses back. Each request to an LLM contains initial context including a system prompt, as well as some additional input. This is typically files loaded into the context such as AGENTS.md, and tool definitions.

A coding agent's first LLM request contains this initial context, along with a first user message.

request 1:
[system][tools][user]

This starts a turn. The LLM may first return an assistant message containing tool calls. The agent program executes them and sends their results back to the LLM, which can then return another assistant message. The turn is finished when the assistant has completed generating output.

after request 1:
[system][tools][user][assistant: tool call][tool result][assistant]
                     <------------------->     ^        <--------->
                     returned by LLM           |        returned by LLM
                                               |
                                     produced by the agent

We continue working, and send another message.

request 2:
[system][tools][user][assistant: tool call][tool result][assistant][user]
                                                                     ^
                                                               new user message

Each turn expands the conversation. Eventually, the history exceeds the context limit. The next request then returns an error such as Request exceeds the maximum size.

[system][tools][user][assistant][....][tool result][user]
                                                      ^
                                             exceeds context window

Handling context overflow

When we cannot continue with the existing conversation as-is, we have two choices.

1. We can start a new, empty conversation without the accumulated context. If we know what we were going to do next, this may often work, but will inevitably lose the conversation history. It might still be a good idea to do, because it is observable that the performance of LLM outputs decrease as the context size grows.
2. We can create a smaller representation of the conversation context, since we want to keep this conversation going. That is what compaction does.

Compaction

In theory, there are many ways to implement compaction. For example, we can write a deterministic function which keeps some of what is in the conversation and discards the rest. In practice, though, implementations of compaction use an LLM request to summarize the conversation history.

Regardless of the method, after compaction, the context should have been compressed such that we have room for many new messages and tool calls.

[system][tools][compaction result][user]
                                    ^
                               new message

Pi's implementation

Let's look more closely at how Pi specifically implements compaction.

When conversations grow too long, Pi uses compaction to summarize older content while preserving recent work. Compaction is triggered when the context limit is nearing the total size of the context window. It can also be manually triggered using the /compact command.

Pi checks for auto-compaction after a turn ends. Until then, each request extends the existing prompt and can reuse its cached prefix. Pi may also compact mid-turn, if it encounters a context overflow error.

When compacting, Pi retains some number of recent messages unchanged.

before compaction:
[system + tools][older turns][recent retained messages]

How many messages are retained vary by session, but it's determined by a configurable number of tokens. Pi's current default of 20 thousand tokens comes out to roughly 5-20 turns.

All the messages before this cut point are extracted and serialized, and will be summarixed. To keep the compaction request within the context limit, Pi truncates tool call results in the history to 2,000 characters. Without reducing the history somehow, the request would already be above the context limit. Tool outputs are a reasonable place to cut because they tend to be intermediate results.

The compaction request that Pi sends differs from regular conversational requests.

1. The system prompt used in the standalone compaction request is different. Instead of telling the LLM "you are an expert coding assistant", we tell the LLM "you are a context summarization assistant".
2. The user message in the compaction request is also different. It requests "a structured summary of this conversation branch for context when returning later". The prompt specifies sections for goal, progress and key decisions.
3. It is a standalone request that doesn't use any of the existing conversation history, which means it can use a different LLM model without incurring any unnecessary cost.

The result of the compaction is appended to the Pi session as a compaction entry, and the session can now continue. After the compaction request, the context has been compressed.

after compaction:
[system][tools][summary][recent turns][new user message]

There is now room in the conversation context for many more messages.

Compaction and prompt caching

Prompt caching is used by LLM providers to make repeated requests in the same conversation more cost efficient. In an active coding session, we pay less for the context that has already been generated by the model. This caching requires an exact prefix match, so compacting a session will break the prompt cache.

cached before compaction:
[system][tools][older history][recent retained turns]
<-------------------- cached prefix -------------------->

first request after compaction:
[system][tools][summary][recent retained turns][new user message]
<-- reusable -->^
                |
        first changed token
                |
                +-- everything after this point must be recomputed

The retained turns contain the same tokens, but they now follow a different prefix. Their previous cached state therefore cannot be reused.

New requests after compaction will benefit from prompt caching again.

Experiment

Since Pi is extensible and malleable, you can replace its compaction with your own. To test a different compaction mechanism, ask Pi to create an extension with your own custom compaction prompt.