nekowasabi/jev-routingArchived
User repository on GitHub
Go Jev harness for Claude Code, Codex, and Grok Build. No npx. Not an MCP server.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Agents & MCP
Jev Routing
Archived Go proxy that sat between a coding-agent CLI and the upstream API to rewrite tool schemas and history with Jev, kept for the negative result: tool steering cut tokens on one model and added 37% and 51% on two others.
- Stars
- 1
- Language
- Go
- License
- MIT
- Published by
- Community
- Author
- nekowasabi
- Last verified
- Oct 2026
About this project
Goal: reduce upstream tokens (parent + child input and output) used by AI agents such as Claude Code and Codex, while preserving quality. Jev (TypeSafe) usage is cheap and is not included in this metric. Jev-based tool selection and steering did not become a generic reduction method. The effect depends heavily on the model, and the same approach can swing in the opposite direction. On the models that increased, forcing the tool roughly doubled the number of upstream requests at most.
Quoted from the project's README. Read the full README.
- Languages
- Go
Highlights
- Pre-registered 6-pair runs: forcing tool_choice cut median tokens 19.30% on one Codex model and increased them on two others.
- On Claude Code the same advice approach moved tokens by 1.67% at most, with one model showing no upstream difference.
- What did work was shrinking context: Anthropic-native context editing plus a clearing gate gave a median 11.1% net reduction.
- The author says most early effects were measurement bugs, and only pre-registered, paired comparisons survived scrutiny.
Watch out
The repository is archived, so this is a record of a negative result rather than maintained tooling. The Claude Code context-editing path is off by default and is judged on net reduction along that path rather than on totals across runs, because the path differs sharply depending on whether clearing occurred, and the Codex truncation result's interval crosses zero under load testing. Steering also proved model-specific: the same approach saved 19% on one Codex model while raising usage on two others, roughly doubling upstream requests at worst. The lessons section notes Anthropic rejects forced tool_choice on Opus 5.5 and under manual extended thinking, and that changing tools or tool_choice drops prompt-cache entries. The bench also lacked the roughly 80k-token fixed prefix that real sessions carry. 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.
Because the measured conclusion was that Jev-based tool selection and steering is not a generic token-reduction method. The effect depends heavily on the model: forcing tool_choice cut median tokens 19.30% on one Codex model, but the same approach increased usage 37.08% and 50.78% on two others, and on Claude Code the advice variant moved 1.67% at most. The author kept the repository to share the trial-and-error record and the full data.
Yes, but it was context shrinking rather than decision replacement. Anthropic-native context editing with a Jev clearing gate gave a median 11.1% net reduction on the same path, off by default behind JEV_CLAUDE_CLEAR_TOOL_USES=on. Deterministic truncation of large tool results, omitting the middle of anything over 20,000 bytes, is on by default and can be disabled, with a median 8.19% under load testing where the interval crosses zero. Jev-based Read replacement, Jev-free tool-call synthesis and compaction tuning all showed no effect.
Very possibly because the author hit the same measurement traps. The lessons section says most early apparent effects were bugs — zero-filled missing usage, miscounted requests and miscounted subagents — and that identical runs varied from -271% to +90%. Only pre-registered, paired comparisons held up, which is why the published table gives both a per-pair median and an improved/worse count.
Yes, the host and API contract notes, which were learned after implementing rather than before. Forced tool_choice is rejected on Opus 5.5 and under manual extended thinking, changing tools or tool_choice drops prompt-cache entries, and some hosts report no usage data at all. The advice is to estimate the ceiling on real sessions before building: two of these approaches were finished before an offline replay showed ceilings of 0.1-0.3% and below zero.
It is a Go proxy that sat between a coding-agent CLI and the upstream API, rewriting tool schemas and history with Jev. The repository is archived.
The bench lacked the roughly 80k-token fixed prefix that real sessions carry, so the measured token effects may not transfer to production sessions.
Roughly doubling upstream requests at worst, alongside token increases of 37.08% and 50.78% on two models. That overhead is the main reason the approach was not iterated on.
More Jev resources like this
3GitHub stars
Go MCP server exposing Jev through OpenRouter or TypeSafe as five tools: decide, classify, verify, rerank and list_models. Each tool call is one paid inference request, answers are validated against the question asked, and the unchanged provider response is returned with its usage cost.
Jev Agents & MCP#community#jevbest#catalog-import
Community
freepik-company/jev-mcpGo
freepik-company/jev-mcpGo51GitHub stars
Local LinkedIn recruiting workspace on Browser Use's Jev Ultrafast: Jev screens titles, drives navigation, and assesses profile excerpts against your brief while a live browser stays visible; matches and evidence export as JSON.
Jev Agents & MCP#community#jevbest#catalog-import
Community
skeptrunedev/jev-recruiterPython
skeptrunedev/jev-recruiterPython33GitHub stars
Codex Skill plus stdlib-only Python CLI that recommends which installed skill to read next for a task, allowing no-skill or uncertain answers and never executing candidates.
Jev Agents & MCP#community#jevbest#catalog-import
Community
win4r/jev-skill-suggesterPython
win4r/jev-skill-suggesterPython72GitHub stars
MCP server over stdio that gives coding agents fourteen Jev judgment tools, turning evidence plus an enumerable question into auto, review, or escalate actions from calibrated probabilities.
Jev Agents & MCP#community#jevbest#catalog-import
Community
PyModel/jev-judge-mcpPython
PyModel/jev-judge-mcpPython21GitHub stars
Research runtime that pages a large logical tool and context space through a small bounded Jev decision surface with explicit approvals.
Jev Agents & MCP#community#jevbest#catalog-import
35GitHub stars
Rust terminal tool that asks Jev noul, choice and score questions about piped text or files and returns calibrated probabilities with stable exit codes for shell scripts, CI jobs and agents.
Jev Agents & MCP#community#jevbest#catalog-import
Community
shaharia-lab/jev-cliRust
shaharia-lab/jev-cliRustFrom the Jev community
Posts from builders shipping with Jev right now.