pi-muse-compact

Standalone spark-compact for pi — fixes OpenAI compaction (tool_choice auto only) on FreeBSD/Linux

Packages

Package details

extension

Install pi-muse-compact from npm and Pi will load the resources declared by the package manifest.

npm report

$ pi install npm:pi-muse-compact

Package: pi-muse-compact
Version: 1.0.0
Published: Aug 26, 2026
Downloads: 142/mo · 142/wk
Author: muswawir
License: unknown
Types: extension
Size: 12 KB
Dependencies: 0 dependencies · 0 peers

Pi manifest JSON

{
  "extensions": [
    "./extensions"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-muse-compact — WORKAROUND for failing pi compaction

This is a workaround, not a fix upstream. Pi's built-in compaction is broken on your setup (OpenAI-compatible proxy that only supports tool_choice:"auto"). This package works around it so /compact and auto-compaction don't crash.

Root cause: pi hardcodes toolChoice:"none" in completeSummarization (chunk-E5KXRMZK.js: requestOptions={...,toolChoice:"none"}) → your proxy returns 400 only "auto" is supported for tool_choice. The workaround intercepts session_before_compact and re-does summarization with no tools so the request stays "auto".

Upstream fix would be pi/pi-ai defaulting to "auto" when no tools are needed, or making compaction configurable. Until then, use this.

Install

pi install ~/pi-muse-compact
# or test without installing:
pi -e ~/pi-muse-compact/extensions

Use

Auto: just use pi normally — /compact and auto-compaction (overflow) are intercepted and re-run with auto.

Manual standalone:

/spark-compact
/spark-compact focus on auth changes
# tool for Spark:
# spark_compact(instructions="focus on ...")

How the workaround works

session_before_compact → serializes messagesToSummarize via serializeConversation(convertToLlm(...)), calls ctx.modelRegistry.complete(model, {messages}, {cacheRetention:"none", sessionId:uuidv7()}) with no tools → stays tool_choice:"auto".
Prefers a non-OpenAI model for summarization (google/gemini-2.5-flash) when available; else uses current model.
Handles text + thinking content; if LLM still fails/empty (e.g. Insufficient Balance), falls back to a deterministic summary from preparation.fileOps + tail messages so compaction never falls back to pi's broken tool_choice:"none" path.
Returns {compaction:{summary, firstKeptEntryId, tokensBefore, usage}} so pi skips its default compaction.
Also registers tool spark_compact + command /spark-compact as standalone entry points that call ctx.compact().

Limitations

Workaround only — if the summary is LLM-generated you get a real summary; if the provider is down/billing fails you get a minimal deterministic fallback (still compacts, but lossier). Remove this package once upstream pi fixes the toolChoice:"none" default.

Layout

~/pi-muse-compact/
  package.json          # pi package manifest (extensions: ["./extensions"])
  extensions/index.ts   # the extension
  README.md

Uninstall: pi remove ~/pi-muse-compact or pi remove pi-muse-compact