User repository on GitHub
Adaptive reasoning effort for GPT-6 during Codex tasks, powered by Jev to reduce token usage.
- Forks
- 24
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Astra-Ares
A source-only preview that builds a separate patched Codex CLI so Jev can raise or lower GPT-6 reasoning effort mid-task and decide how many generations the choice holds. Needs Rust, Node 22 and a native toolchain.
- Stars
- 304
- Language
- JavaScript
- License
- MIT
- Published by
- Community
- Author
- miuuyy
- Last verified
- Oct 2026
About this project
Adaptive Reasoning Effort Selection for GPT-6 Astra, Sol and Luna while your Codex task runs. Get started · How it works · Configuration · Troubleshooting.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- Setup downloads a pinned Codex build, applies the included patch and compiles it, leaving your existing codex command and the desktop app untouched.
- Jev picks the effort plus a lease of one, two, five or ten generations, and the steps inside a lease cost no further Jev requests.
- Bounded evaluator context: last six tool calls, 1,000 local tokens of result text per call with head and tail truncation, and a 28,000-token request guard.
- Only Apple Silicon macOS was built and tested locally; Intel and Linux have build paths but no acceptance testing, and the socket integration rules out Windows.
Watch out
The token-saving premise is unmeasured: the README says fixture tests confirm that settings are applied and the prompt prefix is preserved, but cache hit rates and savings against fixed effort have not been measured on real workloads, and every decision adds a Jev round trip. The build wants roughly 10 GB of free disk, several minutes of compile time, and Rust, so this is not a quick install, and a new user needs funded OpenRouter credits because that is the default Jev provider. The captured README does not cover uninstall steps, which live in a separate installation doc. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Neither. It installs a separate patched Codex build, and the bridge stays outside the selected model's network path, so Codex still talks directly to OpenAI with the same model, conversation and prompt-cache prefix intact. Your normal codex command and desktop app stay as they are, and resume applies to the separate Ares profile's sessions rather than your everyday ones.
Bounded task context only: the original and current task plus retained previous user requests, public progress, plans and published reasoning summaries, the last six tool calls with paired results clipped to 1,000 local tokens each, any existing summary and saved-artifact names. System prompts, private or encrypted reasoning, images and complete transcripts are excluded, oversized requests stop explicitly at the 28,000-token guard, and the whole payload goes to whichever Jev provider you configured.
Not on a real workload, and the README admits it. Native fixture tests verify that the settings are applied and the prefix survives, but cache hit rates and savings against fixed effort have not been measured yet. It also costs an extra Jev round trip and local checkpoint processing per decision, so treat the saving as a design goal rather than a result.
A 429 is read as a rate or capacity rejection rather than proof that the context is too large, and transient HTTP errors get at most three attempts against the same provider inside a 30-second deadline. Once that is exhausted the turn stops visibly, and the tool never switches providers, substitutes another model or invents an effort value. ares doctor checks the install locally and ares doctor --probe makes one small billable request, with decision logs written per run.
Node 22+, npm, Git, curl, tar, a native C/C++ toolchain, and Rust via rustup. Allow about 10 GB free and several minutes to compile, since setup downloads a pinned Codex version, applies the included patch, and builds it. macOS may need the Xcode Command Line Tools; Windows needs rustup's MSVC toolchain with the Visual Studio C++ workload and Windows SDK.
Jev chooses an effort and a lease of one, two, five, or ten generations. A step is one model generation that may produce several tool calls, and the decision happens before the next generation once tool results are in. With a ten-generation lease, Ares asks at step 1 and step 11; steps 2 through 10 reuse the accepted choice with no extra Jev requests.
Each model entry keeps its underlying model fixed while Jev chooses the effort: Astra Ares, Sol Ares, and Luna Ares, appearing when that model is in your Codex catalog, with new profiles defaulting to Astra Ares. Jev is the default evaluator, but you can select another OpenRouter decision model or connect a separately operated local decision service.
More Jev resources like this
118GitHub stars
A Claude Code skill that moves bulk judgment off Claude and onto Jev. You filter files, classify items, rank candidates or locate lines, and it returns typed verdicts, a shortlist with borderline markers, and a receipt of the tokens Claude did not read.
Jev Agents & MCP#community#jevbest
Community
UditAkhourii/quicksilverJavaScript
UditAkhourii/quicksilverJavaScript66GitHub stars
Claude Code plugin that asks Jev one yes/no question per rule and passes Claude only the matches, so large rule folders stay out of context until relevant.
Jev Agents & MCP#community#jevbest#catalog-import
Community
EliaAlberti/jev-rulesJavaScript
EliaAlberti/jev-rulesJavaScript47GitHub stars
MCP server that screens files, public webpages, or inline text against a query and returns compact relevance probabilities, so the main agent opens only promising items.
Jev Agents & MCP#community#jevbest#catalog-import
Community
kbhuw/jev-siftJavaScript
kbhuw/jev-siftJavaScript16GitHub stars
Standalone terminal agent that puts Jev inside its planning loop, proposing text-model steps and scoring candidates against independent rubrics before a deterministic policy executes one.
Jev Agents & MCP#community#jevbest#catalog-import
19GitHub stars
Capability resolver that uses Jev to search a nested command catalog and return selected instruction context to an agent through one MCP tool, keeping the search outside agent context.
Jev Agents & MCP#community#jevbest#catalog-import
Community
NiazMorshed2007/jcrJavaScript
NiazMorshed2007/jcrJavaScript7GitHub stars
MCP server exposing one evaluate tool for Jev Choice, Noul, and Score questions, plus a separate eval harness that compares Jev against LLMs on shared labeled sets.
Jev Agents & MCP#community#jevbest#catalog-import
Community
arunav25/jev-mcpJavaScript
arunav25/jev-mcpJavaScriptFrom the Jev community
Posts from builders shipping with Jev right now.