nautahakk/jev-codex-router

Jev Codex Router is a plugin purpose-built for DeepSeek Harness that consults TypeSafe's Jev service on every substantive prompt to pick a Codex model and reasoning effort among Luna, Terra, Sol, and Astra, then rewrites the request so the chosen route runs while the user stays inside Harness's normal web UI. Installation demands Node.js 22 or newer, Git, pnpm, a ChatGPT account with Codex access, and a TypeSafe API key, and proceeds by cloning the repository, running npm install, registering the third-party @syncended/dsh-codex provider at version 0.0.10 with npx dsh plugin --profile web add, then executing npm run router:install followed by npm run harness. A new Jev Router settings pane stores the TypeSafe key through Harness's credential service, and the README insists this is an unofficial community project with no endorsement from DeepSeek, TypeSafe, or OpenAI.

The intended audience is Harness users who want automatic, per-turn model selection without manually editing configuration between requests, and the workflow centers on a single settings page where operators cap the maximum eligible model (for example setting Sol to exclude Astra from Jev's choices, backups, and emergency fallbacks) and tune the Astra automatic threshold up to which it can run unattended, with any higher Astra effort asking for that turn's approval and a declined approval falling back to the best permitted model. Each completed Codex turn displays the actual model and reasoning effort used beneath the response, short acknowledgements such as "yes" or "continue" keep the previous route unless the maximum was lowered, and turning routing off restores Harness's manual selector, which makes the plugin attractive for users benchmarking routing policies, managing subscription costs, or sharing a Harness instance across teammates with different model budgets.

Hard dependencies include Node.js 22+, Git, and pnpm, while the repo pins @deepseek-ai/dsh to 0.1.6-alpha.2, the TypeSafe SDK to 0.6.0, and the Codex provider to 0.0.10; Jev receives at most the first 8,000 characters and last 4,000 characters of the latest user-authored prompt plus attachment type and display-name metadata, never file contents, tool output, OAuth tokens, or the full conversation. Licensing is MIT under the standard LICENSE file, but operators must budget TypeSafe usage separately from the ChatGPT subscription, and a documented upstream advisory warns that the pinned Harness stack pulls fflate@0.8.2 through its office-to-PDF pipeline, which can stall on malformed ZIP64 files, so the README recommends avoiding untrusted document uploads until upstream ships a fix. The first-run caveat is to restart Harness after installing or updating any plugin and to keep the checkout in place so that the Windows npm run launcher:install desktop shortcut and the locally linked plugin continue to resolve.