An efficiency audit for the Pi coding agent, bottled as an extension

What it does

SoL-Pi is a standalone TypeScript extension for the Pi coding agent that applies four opt-in efficiency mechanisms to long-running sessions. It reduces repeated model turns, context replay, oversized observations, and unnecessary log reading while preserving the evidence an agent needs to finish a task. Every mechanism is disabled by default and hooks into Pi through public extension APIs without modifying the core.

The interesting bit

The four mechanisms were originally discovered by scaling auto-research loops—essentially letting agents optimize their own harness—and then hardened into a standalone package. The Evidence-Preserving Reducer is especially particular: it will only swap a long diagnostic log for a compact receipt if every retained quotation exactly matches an archived local copy.

Key highlights

• Action Fusion lets a file edit or write trigger its follow-up validation in the same tool call.
• ObservationPack replaces repeated large text results with stable handles that support exact paged recall.
• Evidence-Preserving Reducer compresses logs only when every quote is verifiable against local session archives.
• Online Context Compact marks completed plan steps for Pi’s native compaction and resumes the task in a new turn.
• Uses only public Pi APIs; auth, model choice, and shell behavior stay under Pi’s control.

Caveats

• evidencePreservingReducer and onlineContextCompact make extra model calls or can stop and restart an active run, so the README advises reviewing their configuration and security implications before enabling them.
• ObservationPack and the reducer leave archived copies in the session directory; these are not automatically deleted when Pi exits.
• The reducer may send diagnostic-log content to a remote model via Pi-managed authentication, which is worth checking if your logs must stay local.

Verdict

Grab it if you already run Pi on lengthy coding tasks and want to trim token burn without gutting verification. If you do not use Pi or your sessions rarely hit context limits, this is not for you.