astra-jev-harness
Codex and Claude Code skills that narrow a repository locally and let Jev judge which candidate files the coding model needs.
Overview
astra-jev-harness provides the astra-jev-coding skill for Codex Desktop, a separate harness for Codex CLI, and a claude-jev-coding skill for Claude Code. Given a task, it snapshots eligible files, narrows oversized repositories locally, asks Jev which candidates the coding model actually needs, preserves dependencies and records what was kept and why. The coding model then writes the code.
The goal is lower token use and inference cost without losing correctness or adding turnaround. In a historical v0.3.0 comparison, one synthetic task at the highest effort used 27.0% fewer input tokens; Claude Code savings are not yet measured. The default batch policy keeps a whole batch if any judgment is uncertain; an experimental per-file policy keeps uncertain files and drops confidently irrelevant siblings. Since v0.7.2, if Jev is unavailable the skills record that selection is unverified and continue authorized work rather than retrying.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does the coding model need this candidate file for the task?(inferred) | yes · no |
How it works
- Snapshot eligible files for the task
- Narrow large repos locally
- Jev judges bounded candidates
- Keep needed files plus dependencies; record reasons
- Coding model implements with the selected context
Strengths
- Reported 27% fewer input tokens in one historical test
- Conservative policy keeps batches when uncertain
- Continues safely when Jev is unavailable
Caveats
- Experimental; savings measured on one synthetic task
- Claude Code savings unmeasured
Getting started
Install the astra-jev-coding skill for Codex Desktop or claude-jev-coding for Claude Code from the repository.At a glance
- Type
- Project
- Status
- prototype
- Language
- Go
- Runtime
- agent skills + CLI harness
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
The default \`batch\` policy retains an entire batch when any judgment is uncertain. The experimental \`select --policy per-file\` retains uncertain/unjudged files while omitting confidently irrelevant siblings. Dependencies and the global no-match fallback still apply. Compare before changing policy; fewer bytes alone do not establish correctness.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev decide?
Which candidate files the coding model needs for a task.
What if Jev is uncertain?
The default policy keeps the whole batch.
How much does it save?
One historical synthetic test showed 27% fewer input tokens; broader savings are unmeasured.
Related projects
- winnow★ 110
Calibrated context sieve for Claude Code: Jev judges each block of large tool output and hides confident-no blocks behind recallable stubs.
Projectbeta Local compaction proxy for Claude Code and Codex that compacts more often with local rules and keeps full history searchable.
ProxybetaGoClaude Code and Codex plugin where Jev judges each turn for model routing, delegation, evidence search and compaction.
ProjectbetaTypeScriptClaude Code (and Codex) plugin that trims long Bash output chunk by chunk with Jev before it reaches the model, without summarising.
PluginbetaTypeScriptGo CLI that gives coding agents Jev-powered lookups and calibrated yes/no or choice verdicts instead of reading whole files.
CLIbetaGoAgent skill that has Jev pick the few relevant lines from big logs before your coding agent reads them, with a reversible reduce and ledger.
ProjectbetaPython