fast-jev-codex

Overview & result

fast-jev-codex uses Jev to score tool history and lets the plugin restore verbatim context after Codex compacts a long session.

From the original post

Context compaction shouldn't eat the error you're trying to fix

I built fast-jev-codex: Jev scores tool history, then the plugin restores verbatim context after Codex compacts.

Open source. Bring your own Jev key.

https://github.com/chrismathew3/fast-jev-codex

How Jev fits in the loop

1. The plugin observes prior conversation and tool-call history.
2. Jev judges which historical items remain useful for the current task.
3. The plugin restores selected original text so Codex can continue with the preserved evidence.

How to reproduce

1. Read the repository README and review how it hooks into Codex context handling.
2. Run a deliberately long coding session containing tool calls, errors, and an important earlier constraint.
3. Check whether the important context remains verbatim after compaction and record any omissions.

Why this build matters

It isolates a concrete failure mode of long-running coding agents: context compaction can be lossy even when the exact error or path matters.

Limitations

• The short post does not specify the scoring rubric or a retention benchmark.
• Context pruning can still remove information if the state representation or threshold is wrong.
• Treat the repository behavior as community observed until JevForAgents reproduces it.

Patterns