Spend fewer tokens in long agent sessions without changing how you work.
Jev-powered optimization for Claude Code, plus model routing for OpenAI Codex.
Get started · Codex support · Documentation · Architecture · Contributing
Long agent sessions accumulate stale tool output, send oversized skill catalogs, and use the main model for work a smaller model could handle. Claude Code supports all three optimizers below. Codex supports the model router and an off-by-default advisory skill experiment.
Each tool works independently and fails open if unavailable. Measured cost reports remain Claude-only.
| Tool | What it does | Available in | Claude default | Learn more |
|---|---|---|---|---|
| jev-compacter | Removes stale tool calls and outputs instead of summarizing away the whole conversation | Claude Code | On | Guide |
| jev-router | Runs each subagent, Explore searches included, on the cheapest model that fits, within project rules that evolve with your workflow | Claude Code and Codex | Off | Guide |
| jev-skills | Sends a short, relevant skill list for each prompt instead of the entire catalog | Claude Code; experimental advisory mode in Codex | Off | Guide |
jev-compacter — preserve useful context
It works automatically with /compact and auto-compaction. Inspect the current session with:
/jev-compact stats
Want it to prune earlier or keep more? Pick a Compaction: style (balanced, save-tokens, or keep-context) in /config under optim-jev. See Pick a style.
jev-router — the right model for every subagent
Your main conversation keeps its model, so its prompt cache stays warm. The routing happens where it pays off: every subagent Claude spawns gets its own model and effort from Jev, Explore searches included. A quick lookup runs on haiku; a deep architecture review gets a stronger model.
Claude works as usual, editing directly and delegating when a task calls for it or when you ask. rules.json sets which models may edit each path; when a path needs a different model than your session's, Claude hands that change to a routed subagent.
/jev-route init "what matters in this project"
/jev-route on | off
/jev-route why | stats
Tell Claude to change a preference later—for example, “let haiku edit the tests”—and the project rules update with your workflow.
jev-skills — load only relevant skills
/jev-skills on | off
/jev-skills scope project | global
/jev-skills why | stats | reindex
- Claude Code with plugin support
- Python 3.10 or newer
- A TypeSafe API key for Jev
git clone https://github.com/Oqura-ai/optim-jev.git
claude plugin marketplace add ./optim-jev
claude plugin install optim-jev@optim-jevSet the key in Claude Code under /config → TypeSafe API key, or export it before starting Claude Code:
export TYPESAFE_API_KEY="your-key"Restart Claude Code. jev-compacter is ready immediately; enable the router and skill picker when you want them. See Getting started for every command, option, and troubleshooting step.
Note
The plugin uses only Python's standard library. Your API key is passed to child processes through the environment and is never written to project files or logs.
Start with the complete Codex guide for installation, usage, model-tier setup, verification, limitations, troubleshooting, and uninstall steps.
optim-jev can automatically route each Codex subagent to an account-specific fast, balanced, or
strong model. It also selects reasoning effort, applies project path rules, escalates after
failures, and can provide an optional advisory skill shortlist. Claude and Codex keep separate
model catalogues, policies, caches, and logs.
Important
Compaction is not supported in Codex. Codex hooks cannot replace live conversation history. The AbovePrompt dashboard, slash commands, and measured token and cost reports are also Claude-only.
Codex CLI 0.162.1 or newer can install optim-jev directly from GitHub. No PyPI release or manual clone is required.
py -3 -m pip install "https://github.com/Oqura-ai/optim-jev/archive/refs/heads/main.zip"
optim-jev codex install
$env:TYPESAFE_API_KEY = "your-key"
cd C:\path\to\your-project
optim-jev codex setup
optim-jev codex status
codexOn the first Codex launch after install or update, run /hooks, review and trust every optim-jev
hook, then restart Codex.
Inside Codex, subagent routing is automatic. Check the most recent decision or control project features with:
!optim-jev codex why
!optim-jev codex status
!optim-jev codex on
!optim-jev codex off
!optim-jev codex skills-on
!optim-jev codex skills-off
- Found a bug or have a feature idea? Open an issue.
- Want to help? Read the contributing guide and submit a focused pull request.
- For the wider Claude Code community, join the official Claude Developers Discord.
optim-jev is released under the MIT License.
Built by Oqura.ai for more efficient coding-agent workflows.