Evidence-first tooling for coding agents: context compaction that keeps the facts, measured on blind held-out data.
| Project | What it does | Latest |
|---|---|---|
| jev-factkeep-compaction | Claude Code plugin, plus a Codex hook. Jev-guided context compaction that never erases a tool call: reproducible reads shrink to a note, observations keep their errors, ids, codes and counts; after a Codex compaction, a fact sheet brings them back. A fork of tamaratran/fast-jev-compaction. | v0.3.0-astra.29 |
| hermes-jev-compaction | The factkeep engine as a context engine for Hermes Agent: reproducible reads shrink to a note, observations keep errors, ids, codes and counts; deciders resolve through a provider registry (TypeSafe Jev, OpenAI gpt-6-luna). |
v0.90.0 |
| jev-watch | Claude Code plugin that logs what the compaction does (outcomes, TypeSafe requests, compactions, hook failures) to JSONL files. | 0.1.2 |
On blind held-out rounds the compaction rules kept 304 of 336 preregistered facts, against 36 for the original engine. The analysis, with per-call data and a script that recomputes every number: why facts are lost.
- Claims come with data. Measurements are preregistered before any run, repeated, and published with the tables and scripts needed to recompute them.
- Upstream when it is alive. Fixes go back to a maintained original as pull requests; the original of these forks has had no commits since 2026-09-18, so they are developed here.
- Forks stay compatible. A fork keeps the upstream license, history and plugin names, so it can replace the original without reconfiguration.
- Known limits are stated. Every release lists what it still loses.
