Automate Codex Workflows with Scripts and 59API
Why automate Codex workflows?
Codex is useful for code generation, refactoring, test creation, and repository analysis, but repeating the same instructions manually wastes time. A small script can turn a one-off prompt into a reliable workflow: collect a task, provide the right project context, call the model, save the result, and return a useful exit status.
This approach is especially valuable for maintenance tasks such as generating tests for changed files, reviewing a pull request, updating documentation, or checking whether a migration is complete. The goal is not to let an agent make uncontrolled changes. The goal is to create a predictable automation layer around Codex.
1. Configure a compatible API endpoint
Start by creating an API key with 59API and exporting it locally. 59API provides pay-as-you-go access to native Claude and GPT models through an OpenAI-compatible interface, so existing Codex and OpenAI SDK workflows can use the same general configuration.
export OPENAI_API_KEY="your-59api-key"
export OPENAI_BASE_URL="https://api.59api.com"
Keep the key in your shell environment, a local secret manager, or your CI platform's encrypted variables. Do not commit it to a repository. Before automating a large job, make one small request through Codex or your SDK to confirm that authentication, the endpoint, and the selected model work correctly.
2. Create a focused task script
Use a script that accepts a clear input instead of embedding a vague, permanent prompt. The following Python pattern uses the OpenAI SDK and can be adapted to your installed client version:
from openai import OpenAI
import os, sys
client = OpenAI(api_key=os.environ["OPENAI_API_KEY"], base_url=os.getenv("OPENAI_BASE_URL", "https://api.59api.com"))
task = " ".join(sys.argv[1:])
response = client.responses.create(model="gpt-4o-mini", input=task)
print(response.output_text)
Save it as codex_task.py, then run it with a specific instruction such as python codex_task.py "Write unit tests for the parser and explain any assumptions". Check the model names available in your 59API account and replace the example model when necessary.
3. Add repository context safely
Models produce better results when they receive relevant files, but sending an entire repository increases cost and noise. In a shell wrapper, collect only the files needed for the task:
git diff -- '*.py' > /tmp/change.diff
find tests -maxdepth 2 -type f -name '*.py' -print
You can then include the diff and selected file contents in the prompt. Exclude environment files, credentials, build directories, and large generated assets. A useful instruction is: “Review this diff, identify risks, and suggest tests. Do not modify files or invent APIs.” Explicit constraints make automated output easier to review.
4. Turn the script into a repeatable workflow
Wrap the Python command in a shell script that validates inputs, records output, and fails clearly:
- set -euo pipefail to stop on missing variables or failed commands.
- Write results to a timestamped directory such as artifacts/codex/.
- Store the prompt, model name, and git commit alongside the response.
- Use a nonzero exit code when the API request fails, so CI can detect the problem.
For long-running jobs, add limited retries with exponential backoff. Retry temporary network errors, but do not repeatedly retry authentication failures or invalid requests. Also set a maximum prompt size and a timeout so a broken workflow cannot consume your entire budget.
5. Connect Codex automation to CI
A practical first CI task is an advisory pull-request review. The job can generate a diff, send it to Codex, and publish the response as an artifact for a human to inspect. Keep the workflow read-only at first. Require approval before applying patches, merging code, or running commands suggested by the model.
Use separate API keys for local development and CI, restrict permissions where possible, and monitor request volume. 59API's low-cost, pay-as-you-go pricing makes frequent small automation jobs easier to budget than maintaining a large always-on setup. Its support for official-quality Claude and GPT models also lets you choose a model for each task without rewriting your integration. A referral rebate may provide an additional saving when available.
6. Improve prompts from actual results
Review failed or unhelpful runs and refine the workflow rather than simply increasing model size. Include the expected output format, relevant coding standards, test commands, and a definition of “done.” Log enough metadata to reproduce a result, but remove secrets and personal data from stored prompts.
When the script is stable, you can sign up for 59API and connect it to your existing Codex or OpenAI SDK setup. Start with a small, read-only workflow, measure cost and accuracy, then gradually automate higher-value steps.