jkf87/jev-codex/jev-codex

Jev Codex

Use the bundled jev CLI to classify once with Jev and execute one selected model. The installed directory containing this file is the skill root. The user's project is the execution workspace, not the skill root.

Before running

• If JEV_CODEX_ACTIVE=1, do the assigned task directly. Never invoke this router from its own child execution.
• Check the user's intended request, relevant supplied context, target workspace, and whether file modification is already authorized. Do not ask again for permission already granted.
• Use the existing TYPESAFE_API_KEY environment variable or private router config. Never print, copy into the skill, or publish the key. Missing key means report the missing setup; do not invent a routing result.
• Run <skill-root>/jev --help if CLI details are needed. Setup and backend options are in references/setup.md.
• Explain that executing the skill inside an existing Codex task does not remove that parent task's usage. To route before starting a Codex task, start from the standalone CLI.

Execute

Preserve the actual request and necessary conversation context. Prefer request/context files outside the skill directory for multiline text; do not interpolate untrusted text into shell code.

<skill-root>/jev --request-file /path/to/request.txt --context-file /path/to/context.txt --workspace /path/to/project --json

Omit --context-file when no prior context is needed. The default sandbox is read-only. Add --sandbox workspace-write for already authorized edits. Use --plan when the user asks only which model should run. Use --max-tier fast or --max-tier standard for a requested capability ceiling; this is not a dollar limit.

The router uses fast, standard, and frontier. Model IDs and effort live in config. Jev's choice and confidence do not authorize external messages, publication, destructive commands, or broader permissions.

Report and stop

• Return the selected tier/model and the actual answer or produced artifact. Do not run the same task again in the parent after the selected executor completes it.
• Separate planned, completed, blocked, and execution_failed. A plan is not proof that the downstream model ran.
• Report measured usage/latency only. Do not claim fixed cost savings or that the CLI-confirmed requested model proves the server's exact resolved model.
• On routing failure, budget block, or execution failure, report the code and relevant recovery from the setup guide. Do not silently invoke a more expensive model.
• On execution timeout, inspect for partial changes before considering a retry. There is no automatic retry.

The CLI does not intercept all Codex app conversations. Skill discovery also is not a programmatic guarantee that every request will pass through Jev.