// the take
pi-context-prune is an extension for the Pi coding agent that summarizes and strips verbose tool-call outputs out of future LLM context while keeping the originals retrievable on demand. It's for people running long agentic coding sessions in Pi who are burning context/tokens on tool output they don't need verbatim after the first read.
It doesn't actually delete anything — originals stay in a session-backed index and are recoverable via context_tree_query, so pruning is reversible by design, not destructive. The docs show real understanding of provider-side prompt caching: multiple trigger modes exist specifically to control how often the prompt prefix gets rewritten, since rewriting it busts the cache and can cost more than it saves. It also checks whether a generated summary is actually smaller than the raw output before pruning, so a verbose summarizer can't make things worse. The summarizer model/thinking level is configurable independently of your main model, so you're not wasting a frontier model's tokens writing bullet-point summaries.
It only works inside Pi coding agent (badlogic/pi-mono) — it's not a general context-management tool, so it's useless outside that specific harness. Summarization is lossy by construction and the README admits the default mode (agentic-auto) depends on the model's own judgment about when to prune, which it calls 'less predictable' than the manual alternatives — that's a correctness guarantee resting on model discipline, not code. If a cheap summarizer model drops something load-bearing from a big diff or log, you only find out if you remember to query the original. It's a single-maintainer project (253 stars) that's one of seven near-identical extensions the same author ships for this one agent, so there's no real evidence yet of multi-contributor maintenance or longevity beyond one person's attention.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →