Overview
Astra-Ares (github.com/miuuyy/Astra-Ares) is miuuyy's experimental bridge that installs a separate pinned Codex build, then asks Jev before each eligible generation how hard the next step looks and how many generations that effort should cover. Jev reads bounded task context (recent tool results, public progress, retained user goals) and returns choices mapped to native GPT-6 reasoning effort plus a lease of 1, 2, 5, or 10 generations. Codex applies settings through OpenAI's configuration_update path so prompt prefixes stay cache-friendly. OpenRouter is the default Jev provider on fresh installs; transcript lines show APPLIED when native effort changes stick. README labels the project a reference implementation, not a polished daily driver.
Problem: Codex sessions on GPT-6 Astra, Sol, or Luna often burn tokens on high reasoning effort for trivial next steps.
Built for: Power users running a patched Codex CLI who want mid-task reasoning effort changes without swapping models or invalidating prompt cache prefixes.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- miuuyy
- Handle
- @miuuyy
How Jev is used
- Role in the product flow
- Mid-run reasoning effort and lease duration routing inside patched Codex generations
- Primitives
- ChoiceScore
- State in
- Evaluator packet with original task, public plans, last six tool call pairs, and truncated tool results under documented token guards (not the full encrypted reasoning stream).
- Decision out
- Selected reasoning effort tier and generation lease count applied natively before the next model generation.
- Detect checkpoint when a new generation is due and lease expired or context changed
- Send bounded state to configured Jev provider (OpenRouter default)
- Map Jev choices to Codex native effort settings
- Hold effort across leased generations without extra Jev calls until lease ends or user interrupts
Ares is model routing by reasoning depth, not by vendor swap. Hermes skill packs might pick which model answers; Ares keeps Astra, Sol, or Luna fixed and only moves effort up or down based on what Jev thinks the next step needs. That matches TypeSafe's pitch that structured decisions belong on the hot path while the big model writes code. Failures are explicit: README promises no silent provider fallback and logs decisions under ~/.local/share/astra-ares/runs. Treat leases as a cost knob: ten-generation leases amortize Jev latency, but tool failures force a fresh assessment.
Sourced performance claims
- Example transcript in README shows a Jev decision in about 321 ms with a two-generation lease.Source: github.com/miuuyy/Astra-Ares README
- Evaluator context caps recent tool results at 1,000 local tokens each and 28,000 tokens overall per configuration docs.Source: github.com/miuuyy/Astra-Ares docs/configuration.md
- Public GitHub repo miuuyy/Astra-Ares had about two hundred thirty two stars when this listing was drafted.Source: GitHub star count September 2026
Features and stack
Features
- Pinned Codex patch with npm run setup build pipeline
- Native GPT-6 effort changes with prefix-preserving cache story
- ares doctor and --probe for local config plus billable Jev ping
- Separate Ares Codex profiles without touching stock codex binary
Stack
- TypeScript
- Node.js 22+
- Patched Codex CLI
- OpenRouter or other Jev providers
- Rust toolchain for Codex build
Pricing: MIT bridge plus Apache-2.0 patched Codex sources; Jev calls bill through your configured provider, Codex usage bills through OpenAI as usual.
Links
FAQ
- Does this replace my normal codex command?
- No. Setup installs a separate binary and profiles. Your existing Codex install stays untouched per README installation notes.
- Which Jev provider works out of the box?
- OpenRouter is default after ares configure. Other providers are documented under docs/configuration.md with explicit env vars.
- Is it production ready?
- README marks Astra-Ares as an experimental reference for adaptive reasoning effort, primarily for integrators experimenting with GPT-6 native effort APIs.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.