Imidef Blog

Hands-on with Codex in VS Code: A Practical Look at GPT‑5 Thinking Modes

Codex Deep Dive Image

In the previous article, I walked through using GPT‑5 with Codex—OpenAI’s coding agent—in VS Code and how to switch between its thinking modes.
In this follow‑up, we’ll look closer at how each thinking mode behaves, how approval levels keep things safe, and how to push work to the cloud when it makes sense.

How the Codex Extension Works

  • Install from the VS Code Marketplace.
  • Sign in with your OpenAI account (no manual API key).
  • Work locally in the editor, or delegate bigger jobs to a managed cloud runtime.

This two‑layer model makes it easy to move from quick edits to large refactors without breaking flow.

GPT‑5 Thinking Modes, in Practice

Here’s how Instant, Balanced, and Thinking feel in day‑to‑day use:

  • Instant

    • Snappy responses with minimal exploration
    • Ideal for code completion and short explanations
    • Feels like “asking a colleague”

  • Balanced

    • Middle ground between speed and depth
    • Great for code reviews and drafting content
    • Feels like a reliable pair‑programmer

  • Thinking

    • Slower replies, but much deeper reasoning
    • Strong at design exploration and debugging
    • Feels like consulting a senior engineer

The key is choosing your trade‑off between speed and depth.
Switching modes to match the task has a real productivity payoff.

Approval Levels and Safety

Codex uses clear privilege levels so you stay in control:

  • Chat: Conversation only — no code or environment changes.
  • Agent: Can edit files in the workspace; external access needs approval.
  • Agent (Full Access): Allows network and broader system access; best for sandboxed or throwaway environments.

This reduces the “I don’t want AI touching my project” anxiety.

Cloud Offloading: What It Enables

Codex can offload heavy work to a cloud runtime as well as run locally.

  • Generate large test suites and run big refactors remotely
  • Review outputs first, then apply locally with confidence
  • Keep thinking and editing while background jobs run

It supports a healthy split: you set direction, AI turns the crank.

Practical Workflows

  1. Clarify specs → Thinking
    State assumptions and constraints to reduce gaps.
  2. First implementation → Balanced
    Get a working baseline with minimal tests.
  3. Micro‑fix pass → Instant
    Tweak naming and comments in seconds.
  4. Bug hunt → Thinking
    Trace side effects and identify test points.
  5. Offload heavy work → Balanced/Thinking
    Push big jobs to the cloud, pull back only the results.

Takeaway: Shift Gears with AI

Using the Codex extension, GPT‑5 becomes something you can “gear shift” between speed and depth:

  • Instant for lightweight tasks
  • Balanced for the day‑to‑day
  • Thinking for design and research

Layer in approval levels and cloud offloading, and you get a safe, efficient coding companion.

👉 Related Articles: