pi-max-context

A pi extension that limits how much context a conversation can use. It does three things:

1. Set a limit. /max-context 100K sets the limit and saves it in .pi.
2. Keep the limit. The limit is loaded again on every restart, new session, fork, and /reload.
3. Act on the limit. When context usage reaches the limit, pi compacts the conversation. Then the model is asked to continue where it left off.

Install

pi install /path/to/pi-max-context

The extension is now listed with pi list. Restart pi or run /reload.

Usage

/max-context            show the current limit and context usage
/max-context 100K       limit = 100,000 tokens        (K = 1,000)
/max-context 1.5M       limit = 1,500,000 tokens      (M = 1,000,000)
/max-context 250000     limit = 250,000 tokens (plain number)
/max-context off        disable the limit (also: 0, none)

Rules:

• Units are case-insensitive. 100k and 100K are the same.
• Decimals are allowed, for example 1.5M.
• Invalid values are rejected with a message.

While pi runs, the footer shows context usage: ctx <used>/<limit>.

Where the limit is saved

Project file (used when it exists):

<cwd>/.pi/max-context.json

{
  "maxContextTokens": 100000,
  "continuePrompt": "Continue the task. Use the compaction summary for earlier context."
}

• maxContextTokens — the limit in tokens. 0 means disabled.
• continuePrompt — the message sent to the model after compaction. Optional. The default is "continue where you left off".

If the project has no file, pi reads ~/.pi/agent/max-context.json instead. The project file always wins.

How it works

• The limit is checked before each model call and after each turn. Large tool results are caught too.
• When the limit is reached, pi compacts the conversation. After compaction, pi sends a follow-up message so the model continues the work.
• Safety rules stop loops:
  • 15 seconds must pass between two compactions.
  • After 3 compactions in a row that do not lower usage below the limit, auto-compaction pauses with a warning. It starts again when usage drops or you set a new limit.
• If the limit is higher than the model's real context window, pi warns you. The model's own window wins.

Files

pi-max-context/
├── index.ts          # the extension code
├── units.ts          # token parsing and formatting (K, M units)
├── package.json      # pi package manifest
└── README.md

License

MIT