Use Jev in Codex with the Official TypeSafe Skill

This tutorial follows the skill-based workflow in the referenced Chinese article and checks implementation details against TypeSafe documentation. Install typesafe-ai from TypeSafe’s own typesafe-ai/skills repository. The skill guides question design and API integration; a live judgment still needs credentials and an API request. Codex remains the coding agent, while Jev supplies a bounded decision.

With Node.js and npm available, open a terminal in your project and run the command below. Select Codex when the installer asks for an agent. Installation is project-local by default; add -g for a personal installation across projects. Use one installation method. You can also ask Codex to install the official typesafe-ai skill using this command.

npx skills add typesafe-ai/skills --skill typesafe-ai

Start a new Codex session in that project and explicitly request the skill by name. Ask it to read the installed SKILL.md and summarize the relevant setup. This checks discovery without making a paid API request. Installing the skill does not force Jev to run on every turn. If discovery fails, verify that installation targeted Codex and the intended project, then restart.

Use the typesafe-ai skill. Read its installed SKILL.md and explain how to make one Jev Choice request. Do not call the API yet.

Create a key in the TypeSafe Console (linked below). Set TYPESAFE_API_KEY locally, then launch Codex from that same terminal. These examples use placeholders. The Bash assignment has no spaces around the equals sign; Windows PowerShell uses $env:. These settings last for this terminal session. Do not paste a real key into chat or commit it. A JevModel website key is issued for our separate API and cannot replace the TypeSafe key.

# macOS / Linux (Bash or Zsh)
export TYPESAFE_API_KEY="YOUR_TYPESAFE_API_KEY"
codex

# Windows PowerShell
$env:TYPESAFE_API_KEY = "YOUR_TYPESAFE_API_KEY"
codex

Try the prompt below with a non-sensitive example. It asks for one real Choice request and a dry run of the next action. It defines candidate actions, a demo confidence threshold, and failure handling. Ask for the actual response and request evidence so that a written analysis is not mistaken for an API call. The 0.8 threshold is a starting example to evaluate, not a universal guarantee.

Use the typesafe-ai skill and read the current official API docs.
Make exactly one real Jev Choice request using model jev-latest.
Read TYPESAFE_API_KEY from the environment; never print it.
If the key is missing or the request fails, stop and report the error.

state:
{"userRequest":"Find the latest release notes for this project’s framework.","knownFacts":[],"availableActions":["answer_directly","web_search","inspect_code","ask_user","human_review"]}

Question: Which next action is needed to reliably complete this request?
Candidates:
- answer_directly: the known facts are sufficient.
- web_search: current external information is needed.
- inspect_code: local project evidence is needed first.
- ask_user: essential information is missing.
- human_review: no suitable action or review is required.

Use a demo confidence threshold of 0.8.
If confidence is below 0.8, set finalAction to human_review.
Report the actual returned model, usage, selected choice,
confidence, full probabilities, and finalAction.
Show redacted request evidence. Do not execute finalAction yet.

The official HTTP endpoint is POST https://api.typesafe.ai/v1/systemone. A request contains model, state, and a questions map. For a question named next_action, inspect answers.next_action plus the returned model and usage. Keep the original probabilities visible. finalAction is a value computed by your workflow, not a required Jev API response field. Only execute it when the tool exists and the action is authorized.

Choice selects from defined options; its confidence summarizes probability concentration, not end-to-end correctness. Noul is a yes/no probability from 0 to 1 and has no separate confidence field. Score uses ordered rubric levels, so it is not inherently a 0–1 score. Keep thresholds and permissions in code, and test them on representative cases.

If Codex sees the skill but authentication fails, check whether its process inherited the key. A desktop app launched from a shortcut may not inherit terminal variables. If a request is rejected, compare the payload with the live API reference. Update a skills.sh installation with npx skills update. Do not retry blindly or invent a successful response.

npx skills update

The installation and API guidance was checked against official sources on September 30, 2026. The prompt is an original dry-run example, not a recorded live result. This article does not claim measured speedups, lower Codex quota usage, or a successful paid API run. Inspect your own request, response, latency, and costs before adopting the workflow.