DEMO / RECORDING
Overview & Result
vechen 在 Codex 任务执行中途用 Jev 动态切换 GPT-6 的推理强度，遇到瓶颈时增强、常规步骤时降低，作者报告节省 50% Astra 成本。
How Jev fits in the loop
- Jev evaluates the current task step to classify it as routine or complex/stuck.
- For routine steps, Jev routes to low reasoning effort to minimize token cost.
- When the agent appears stuck, Jev switches GPT-6 to high reasoning effort for deeper problem-solving.
- The effort-level transition is designed to preserve prompt caching, avoiding cache invalidation costs.
How to reproduce
- Read the author's original post on X at https://x.com/miu21590/status/2101857866378362926.
- Integrate Jev into your Codex agent loop to evaluate task complexity at each step.
- Use Jev decisions to switch GPT-6 reasoning effort: high when stuck, low for routine steps.
- Verify that prompt caching is preserved across effort-level transitions.
Why this build matters
Shows Jev enabling in-flight model configuration changes within a running agent session, a pattern that could make long coding agent runs significantly cheaper without sacrificing quality.
Limitations
- Implementation code is not publicly released; this is a community demonstration post.
- 50% cost reduction is author-reported from personal tests and may vary by task type.
- GPT-6 reasoning effort API availability may depend on Codex access tier.