CompactKeeper — a black box for agent compaction

Flag critical tool output before an agent compacts context, so stack traces and test results don't get flattened

Problem

When an agent's context fills up and compaction triggers, built-in summarization is indiscriminate: test results and stack traces get compressed into a vague 'all tests passed', the model then works from that paraphrase or reruns tools, wasting time and tokens, and the user has no idea what was dropped.

Solution

A local pre-compaction guard that scores tool output in the message stream by type and signal strength, marks messages that must not be lost (errors, test output, diffs, command results), produces a structured summary that preserves key fragments, and injects it back into the agent context via a hook or wrapper.

Users

Full-time and indie developers who live inside Codex, Claude Code, or GLM Coding Plan, especially those juggling multiple repos in a single session.

MVP

Read session logs from Codex and Claude Code; score tool output and extract key fragments with rules; generate an injectable 'compact retention pack'; show a diff view so users confirm what is kept or dropped; run entirely locally with no code upload.

Revenue

One-time purchase ($19–$39) with an optional annual plan for multi-agent support; seat-based team licenses.

Why now

Context management and compaction are becoming acute pain points for coding agents, yet official hook protocols are incomplete, leaving room for an indie wrapper layer.