$ cat ~/map/inception.md

inception

A curator LLM replaces pi's compaction, rewriting old history every few turns to keep the context dense.

extensions✓ Oct 23

pi --extension /path/to/inception/.pi/extensions/inception/index.ts

## TL;DR

An extension that replaces pi's compaction with a Context Gardener: every N turns a curator model rewrites the history older than a protected head into a compact factual replacement, keeping the newest turns verbatim.

## Key points

• $Scheduled curation runs on turn_end: when the frequency is reached and history exists outside the protected head, the curator model is called and a compaction entry is committed from the turn boundary, so in-flight runs stay alive.
• $Flags: --inception-frequency (default 3; each model call counts as a turn), --inception-protected-head (default 5 newest turns left untouched), and --inception-model to pick a cheaper curator as provider/modelId.
• $pi's own threshold and overflow compactions and /compact are intercepted via session_before_compact and curated the same way; if curation fails or returns nothing, pi's default routine runs instead.
• $The curator sees the newest turns read-only, never performs the objective, and never writes code; its prompt prioritizes the objective, implementation state, discoveries, decisions, and open questions.
• $Each pass is iterative — the previous curated state is the input, so discoveries survive repeated compactions.
• $Curator calls disable prompt-cache writes and use a fresh session id (one-off prompts).
• $The protected head is a soft guarantee: under hard context pressure it may yield, because letting pi's default cut run beats overflowing.
• $Install by copying both index.ts and boundary.ts into .pi/extensions/inception/ — one location per machine, or pi rejects the duplicate --inception-* flags; tests run under plain node with node --test.

## FAQ

## References

• README (curator prompt, install, flags, how it works) ↗
• Initial test results (linked from the README) ↗

Last updated: Oct 23, 2026