This is a throwaway local prototype for one question: can Jev reliably send complex or uncertain coding tasks to Codex Astra while routing contained mechanical work to Grok Build?
The router uses Jev only for the decision. It asks three typed questions in one request:
- Choice: Grok Build or Codex Astra
- Score: complexity from mechanical to high consequence
- Boolean: whether planning or investigation should happen before code changes
The final policy is intentionally conservative. If Jev is uncertain, the complexity score is high, or planning is likely, the task goes to Astra.
- Grok Build: installed locally, default model
grok-4.6 - Codex Astra: installed locally as model
gpt-6-astra
cd /Users/simsies/github/jev-model-router-demo
npm install
Add a Vercel AI Gateway key to the already gitignored .env.local:
AI_GATEWAY_API_KEY=your_key_here
npm run demo
This classifies six tasks in parallel and prints the chosen harness, route probability, complexity score, planning probability, and latency. It does not launch either coding agent.
npm run route -- "Add an aria-label to the close button and run its existing test."
npm run route -- \
"Trace the intermittent authentication failure without weakening security." \
--cwd /Users/simsies/github/example-project \
--launch
Launches are deliberately safe for the demo:
- Codex Astra runs ephemerally with a read-only sandbox.
- The Codex launch accepts an explicitly selected non-Git demo directory.
- Grok Build runs in plan mode with web search and subagents disabled.
Neither route edits code in this prototype. After the routing policy is calibrated on real tasks, execution permissions can be designed separately.
Jev requests enable Zero Data Retention and disallow prompt training. Do not paste secrets, credentials, private customer data, or unredacted logs into a filmed task.