Making the Most of Codex Agents - Part 1
I have been wanting to document a practical Codex workflow for a while, so this is the beginning of a new series.
In this Part 1, I will cover:
- Why the $20 plan is a sweet spot for hobby projects
- How to get comfortable with the Codex interface
- A real first task: music transcription + pull requests
- Integrating Codex with OpenClaw via OpenAI web URL login
- What I am planning next in this journey
Why I chose the $20 plan
If you’re an independent builder, the $20 plan is a great cost saver.
For me, it balances:
- Enough usage for evenings/weekend hacking
- Room for multiple iterations on one task
- A predictable monthly budget
Unless you are running heavy daily automations or large team workflows, this tier is usually enough to get consistent progress without overcommitting.
Getting into the interface quickly
My onboarding was simple:
- Start with one repository that has clear structure.
- Give Codex one scoped task at a time.
- Ask it to explain changes before committing.
- Review diffs like you would review a teammate’s PR.
The key is to treat Codex as an engineering partner, not a magic black box.
First task: building transcribemusic with iterative Codex development
By “transcribe music,” I mean the project we built iteratively with Codex:
- Repo: richhiey/transcribemusic
The workflow was intentionally iterative:
- Start from a very small goal (single-file transcription path).
- Run the code and inspect outputs.
- Ask Codex for focused improvements (error handling, CLI options, output formatting, docs).
- Re-run and verify.
- Repeat until the feature felt stable.
That iterative loop was the real unlock. Instead of trying to get everything perfect in one prompt, we used short feedback cycles and kept momentum.
How the pull requests were created
For each iteration, we followed a repeatable PR workflow with Codex:
- Ask Codex to summarize exactly what changed.
- Stage only relevant files.
- Commit with a scoped message.
- Generate a PR title + body that includes:
- problem/context
- implementation details
- validation steps
- known limitations and next tasks
This process kept each change reviewable and made the project history easier to understand.
Integrating Codex with OpenClaw using OpenAI web URL login
The correct setup I use here is based on OpenClaw docs -> Providers -> OpenAI -> Option B: OpenAI Code (Codex Subscription):
- https://docs.openclaw.ai/providers/openai#option-b-openai-code-codex-subscription
What Option B means
Option B is the path for people using a Codex subscription login flow (web login), instead of directly wiring a standard API-key-only provider flow.
Exact OpenClaw flow for Option B (UI + CLI combined)
- Start with the OpenClaw onboarding wizard and pick Codex OAuth:
# Run Codex OAuth in the wizard
openclaw onboard --auth-choice openai-codex
- If you skip the wizard or need to re-authenticate later, run OAuth directly:
# Or run OAuth directly
openclaw models auth login --provider openai-codex
- Open the OpenClaw TUI and confirm the provider/model selection is available:
# Open the OpenClaw TUI
openclaw tui
- In the UI path (Providers -> OpenAI), select Option B: OpenAI Code (Codex Subscription) if prompted and confirm web-login status.
- Choose the Codex-capable model exposed by that login path.
- Restart the OpenClaw gateway to reload provider/auth state:
# Restart OpenClaw gateway after provider/login updates
openclaw gateway restart
- Run a tiny task (for example: “edit one markdown line”) to verify end-to-end read/write workflow.
Practical checks for Option B
- If login succeeds but models are empty, re-open Providers and refresh/reselect Option B.
- If agent replies but cannot commit/edit, check workspace mount + repo permissions in your OpenClaw runtime.
- If session seems stale, restart OpenClaw and repeat Option B web login once.
This keeps the integration aligned with OpenClaw’s Codex-subscription flow instead of generic OpenAI API-key wiring.
Jekyll + GitHub Actions publishing loop
For this blog itself, my loop is simple:
- Draft/update post locally.
- Commit and push to the website repo.
- Let GitHub Pages deploy through Actions.
- Verify the live page once the workflow completes.
As this series continues, I plan to automate more of this loop with Codex while keeping final human review before merge.
Next steps in this series
In Part 2, I plan to cover:
- Prompt patterns that reduce rework
- Guardrails for safer autonomous edits
- CI-aware task planning
- Better review loops for agent-authored pull requests
If you’re also experimenting with Codex agents, this is a good moment to start small, build confidence, and iterate fast.