What to Do When Your Codex Limit Is Reached: Use ChatGPT and GitHub to Keep Coding

The three roles in one view

Start here: confirm what actually ran out

A limit message is not always the same thing as a Codex limit. Open the Codex usage dashboard or Settings → Usage and record which allowance is exhausted, the reset time, and whether your account shows credits, an available reset, an upgrade, or a wait option. Do not assume that switching to ChatGPT is free: on supported plans, Codex and ChatGPT Work can use a shared allowance or credit pool, while regular ChatGPT has separate limits.

1. Save the exact limit message and the displayed reset time in the task notes.
2. Check whether the task stopped before implementation, during an active turn, or after a completed change.
3. If Codex stopped after planning, preserve the plan and continue the low-risk analysis work; do not restart discovery later.

1. Connect ChatGPT to only the GitHub repository you need

In ChatGPT, open Settings and find GitHub under Apps or Plugins—the label depends on the current product surface. Start the connection, install or authorize the ChatGPT GitHub app in GitHub, and choose the organization/account and repositories it may access.

1. Open Settings → Apps/Plugins → GitHub in ChatGPT.
2. Continue to GitHub, authorize the ChatGPT app, and choose the correct account or organization.
3. In GitHub app repository settings, select only the target repository, for example owner/project; request organization-admin approval if required.
4. Return to ChatGPT and verify access with a read-only request naming the repository and a known file path.
5. If the repository does not appear, check the GitHub App installation and selected repositories before disconnecting and reconnecting.

Repository: acme/storefront
Read-only verification:
“Find the checkout entry point in acme/storefront and return the relevant file paths and each file’s role. Do not modify anything.”

GitHub access is on-demand; connecting it does not create a synced administrator-managed repository index. Organization policy, plan, workspace, and product surface can change what is available.

2. Ask ChatGPT for an implementation package, not a vague answer

Tell ChatGPT the exact repository, desired user outcome, constraints, and acceptance checks. Ask it to inspect the authorized repository first, then return a plan that another coding agent can execute. If you also want a code draft, limit it to named files and ask for a patch or complete replacement—not an invented repository-wide rewrite.

You are only responsible for analysis and planning. Do not modify the GitHub repository.

Repository: acme/storefront
Goal: Show a sign-in entry point on checkout for logged-out users; keep the logged-in flow unchanged.
Constraints: Only modify src/checkout/ and its tests; follow the existing React/TypeScript style.

Return:
1. The files and call chain involved in the current implementation;
2. The smallest implementation plan;
3. Specific changes for each file;
4. An exact patch if needed;
5. Regression risks;
6. Test, build, and Preview acceptance commands;
7. Questions Codex must confirm.

3. Save ChatGPT’s result in GitHub before handing off

The read-only GitHub connection cannot be your durable write step. Copy the accepted plan into a GitHub Issue, or add it to the existing Issue/PR as a comment. Keep the acceptance criteria, file scope, proposed code, open questions, and exact verification commands together. This is the handoff object Codex will read.

1. Create one Issue for one user-visible outcome; use a short title that can be referenced from a Codex prompt.
2. Paste the accepted plan and code patch into the Issue body or a comment. Remove secrets and irrelevant code.
3. Record what must not change, the base branch, and the command that proves success.
4. If ChatGPT produced code, label it as a draft until Codex has checked it against the actual repository and tests.

## Codex handoff
Goal: [one outcome]
Source: ChatGPT analysis on [date]
In scope: [files]
Out of scope: [files/behavior]
Plan: [numbered steps]
Draft patch: [only if needed]
Acceptance:
- [check 1]
- [check 2]
Verify: pnpm test --filter checkout
Deploy: Preview only; Production requires approval

4. Open Codex on the same repository and implement the handoff

In Codex, connect GitHub and select the repository environment, or open the local checkout if you are using Codex locally. Ask Codex to read the Issue first, inspect the real files, and challenge the plan when the repository disagrees with the draft. The plan is input—not permission to blindly apply every line of generated code.

Read GitHub Issue #123 as the task brief.

First:
1. Check the current branch, working tree, and project instructions;
2. Audit the Issue plan and patch against the real files;
3. If the plan disagrees with the repository, report the differences before guessing.

After confirmation:
4. Create an isolated branch;
5. Modify only the agreed scope;
6. Add or update regression tests;
7. Run tests, typecheck, and build;
8. Report the diff, verification results, and unresolved issues;
9. Do not merge or deploy to Production without explicit approval.

5. Let Codex produce a reviewable PR, then pass deployment gates

Codex should finish with a small branch commit and a PR that links the Issue. The PR must show the actual diff and checks, not just repeat ChatGPT’s plan. Preview deployment can be triggered by the repository’s existing CI/CD workflow. Production deployment is a separate gate: it needs the project’s deployment credentials, environment configuration, and an explicit approval or merge rule.

1. Codex runs the smallest relevant tests first, then the project’s typecheck/build commands.
2. Codex pushes a branch and opens or updates a PR; the PR records changed files, checks, and remaining risk.
3. CI/CD creates Preview if the repository is configured for it; verify the real Preview URL and behavior.
4. Only after human review and the project’s release rule should the change reach Production.
5. After deployment, verify the public route and write the result back to the PR or Issue.

Do not tell Codex “deploy” without naming the environment. Use “prepare Preview and stop” by default; Production is never implied by a successful local build or a merged PR.

The shortest reliable version

1. ChatGPT + GitHub: read the repository and create the plan.
2. Human: save the plan as a GitHub Issue or PR comment.
3. Codex: audit the plan, implement the real change, run checks, and push a PR.
4. CI/CD or approved Codex action: deploy Preview, verify it, then separately release Production.

Frequently Asked Questions