What does astra-jev-harness do?

It reduces Astra token consumption on coding tasks: a local harness snapshots eligible files and narrows oversized repos, Jev judges which candidates the coding model needs, and the model implements and tests within recorded scope.

Which clients does it support?

Codex Desktop via the astra-jev-coding Skill, Claude Code via the claude-jev-coding Skill, and Codex CLI through a separate harness that runs Codex itself.

How does Jev get used?

Jev assesses relevance over the bounded candidate files, retains uncertain context and dependencies, and the run records what was kept and why; the README's concept illustration is not a cost or speed benchmark.

What did the benchmark measure?

A historical v0.3.0 comparison: one synthetic CLI task at Astra Extra High used 27.0% fewer Astra input tokens and 27.2% less at equivalent Standard API rates including Jev, with both modes passing six checks.

What changed in v0.7.2?

When Jev planning or authentication fails, Desktop and Claude Code now state that selection is unverified and continue authorized investigation, implementation and tests instead of looping back to planning.

Is the token saving guaranteed?

No. The README labels the numbers as one trial per mode from v0.3.0 and says elapsed time was 7.3% shorter in that pair but 34.6% longer in a separate medium pair; Claude Code Skill savings are not yet measured.

Do I need Jev credentials to use it?

Jev judgments need access to the service, and v0.7.2 documents behavior when credentials are unavailable: the affected operation stops, receipts are kept, and authorized work continues without automatic retries.

Jev directory

CommunityJev Agents & MCP

Astra + Jev Harness

Agent Skills for Codex Desktop and Claude Code plus a Codex CLI harness: the harness narrows large repositories locally, Jev judges the bounded candidate files, and the coding model implements and tests the change. A Go architecture arrived in v0.7.0, current at v0.7.2.

See highlights Open github.com

Stars: 3
Language: Python
License: MIT
Reading time: 2 min
Published by: Community
Author: oranquelui
Last verified: Oct 2026

oranquelui/astra-jev-harness

User repository on GitHub

View on GitHub

Agent Skills for Codex Desktop and Claude Code, plus a Codex CLI harness. Jev-assisted context selection. / Codex Desktop・Claude Code向けAgent SkillsとCodex CLI用ハーネス。Jevでコンテキストを選別。

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity catalog-import logicrw codex context-selection

Works forJev for Code Jev for Ranking

Used inJev for Software & DevTools

Languages: Python Go
Platforms: CLI
Capabilities: Open sourceRequires API key

Highlights

Narrow large repositories locally, let Jev judge the bounded candidates, then let the coding model write the code, keeping dependencies and recorded reasons.
Ships the astra-jev-coding Skill for Codex Desktop, a separate Codex CLI harness and a claude-jev-coding Skill for Claude Code; current version 0.7.2.
In a historical v0.3.0 pair on one synthetic CLI task at Astra Extra High, input tokens fell 27.0% and cost 27.2% at Standard rates including Jev.
v0.7.2 lets authorized work continue when Jev is unavailable, marking selection unverified; a user-chosen no-Jev workflow persists across in-scope follow-ups.
The README calls the project experimental: benchmark pairs are single trials, and Claude Code Skill savings are not yet measured.

Watch out

Grounded on the truncated README and GitHub metadata reviewed 2026-10-15; the README's last section was cut mid-sentence and it is bilingual English/Japanese. The token-saving numbers are historical v0.3.0 single-trial pairs per mode, not a general speedup or a Codex Desktop result.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

MCP tool routing: Jev in front of the tool call

Oct 16, 2026

Tool menus can outgrow the context that reads them. The catalog's Jev-based routers and gates narrow the list: benchmarked tool search, per-tool scoring, fail-closed approvals, and WebMCP in the browser.

Jev Agents & MCP#mcp#tool-routing#agents

Read article

Approval policies for coding agents: hooks, guards, and gateways