ManualMode: reserve a real repository task for manual practice alongside Codex

[itscloud0]

I’m building ManualMode for engineers who use coding agents but still want some hands-on implementation and debugging practice.

The Project Mode workflow is: let Codex propose a small task the repository actually needs, reserve it for yourself, and do that task manually in your IDE.

Under the hood, a static skill uses a remote MCP server:

1. ManualMode supplies a skill target and time budget.
2. The agent proposes 1–3 useful tasks from the current project.
3. A task is reserved, and the skill tells the agent to leave that scope to the developer.
4. Local verification produces redacted completion metadata for the next recommendation. If no useful task fits, there are browser drills instead.

The reservation is an instruction boundary, not a sandbox that physically prevents an agent from editing. Project evidence is metadata such as test counts and revision hashes; raw source, diffs and transcripts are not accepted as Project evidence. Your coding agent still has its usual model-provider data flow.

You can try the small five-minute patch review without an account, card or code upload. It supports JS/TS, Python, Go, Java and C++; this sample is a static review, not a benchmark of overall engineering ability. The signed-in browser Gym currently runs JavaScript.

How repository practice works

I’d like feedback from Codex users on one question: is deliberately reserving a useful task for yourself something you would keep doing, or does reviewing the agent’s work already give you enough practice?

Disclosure: I’m the creator. Three Gym reps and one Project rep are free; ongoing Pro is $20/month or $200/year. Independent product, not affiliated with OpenAI.