pi-agentic-compaction

A pi extension that provides conversation compaction using a virtual filesystem approach.

Installation

``bash
pi install npm:pi-agentic-compaction
`

Or add to your ~/.pi/agent/settings.json:

`json
{
"packages": ["npm:pi-agentic-compaction"]
}
`

How it works

When pi triggers compaction (either manually via /compact or automatically when approaching context limits), this extension:

1. Converts the conversation to JSON and mounts it at /conversation.json in a virtual filesystem
2. Spawns a summarizer agent with bash/jq tools to explore the conversation
3. The summarizer queries specific parts of the conversation (beginning, end, file modifications, user feedback)
4. Returns a structured summary to pi

$3

pi's default compaction sends the entire conversation to an LLM in one pass. This works well for shorter sessions, but for long conversations (50k+ tokens), you pay for all those input tokens and the model may miss details in the middle.

This extension's approach lets a small, fast model explore the conversation by running queries. Only the queried portions enter the summarizer's context, making it cheaper and more focused for long sessions.

Trade-offs:
- Cheaper for very long conversations
- May miss context that a full-pass approach would catch
- Requires multiple LLM calls (but with a fast model like Cerebras, this is still quick)

Usage

Compaction happens automatically when needed, or manually with /compact.

You can pass a note to guide the summarizer:

`
/compact focus on the authentication changes
`

Configuration

Edit the constants at the top of index.ts:

`typescript
// Models to try for compaction, in order of preference
const COMPACTION_MODELS = [
{ provider: "cerebras", id: "zai-glm-4.7" },
{ provider: "anthropic", id: "claude-haiku-4-5" },
];

// Debug mode - saves compaction data to ~/.pi/agent/compactions/
const DEBUG_COMPACTIONS = false;
``

License

MIT