Before Agencies Use Codex Desktop: Diff Review, PRs, and Release Checks

In a web agency, a diff view, pull request, staging URL, and client approval note often sit in different tabs. Asking Codex to review everything does not prevent a bad release unless the team separates diff review, PR work, and production approval.

This guide uses “Codex Desktop” as an agency shorthand, not as a claim about one official product name. The actual working surfaces are the ChatGPT desktop app, its review pane, Codex CLI, the IDE extension, and /review. The goal is to decide which evidence Codex prepares and which release decisions remain with the account owner, director, or engineer on duty.

Where agency Codex use goes wrong

• Treat Codex Desktop as a practical label for Codex work in the ChatGPT desktop app, review pane, CLI, and IDE extension.
• Do not mix diff review, PR feedback, staging URL checks, and production approval in one chat.
• Let Codex summarize diffs, run /review, list risks, and draft fixes; keep client approval, pricing, legal copy, publish timing, and production approval human.
• Put practical release rules in AGENTS.md: mobile width, form submit, CTA, OGP, tracking tag, rollback note.
• Measure rework, release incidents, form failures, and client review time, not lines changed by AI.

The source base is OpenAI documentation for ChatGPT desktop app, Code review, AGENTS.md, Subagents, Best practices, and Agent approvals and security. For a related release workflow, see the agency Azure DevOps release article.

What Codex can do and what humans decide

Start with an input contract, not another “review everything” prompt. Codex should receive a defined diff scope, base branch, staging URL, PR URL, changed files, human checks, and rollback note. It should not receive authority to approve pricing, legal copy, client acceptance, publish timing, or production release.

Humans keep the business decision. Codex prepares review evidence. This split is the practical safety line for agency work because a technically clean change can still be wrong for the client.

Keep one release record with three explicit gates. A finding may move from the first gate to the second, but it cannot skip the human approval in the third.

Gate	Evidence Codex may prepare	Human decision
Diff and PR	changed files, risk-ranked findings, test output, unresolved comments	whether the change still matches the agreed scope
Staging	viewport checks, form result, links, metadata, screenshots	whether copy, design, tracking, and client feedback are acceptable
Production	blockers, rollback steps, owner, proposed publish time	client acceptance and final go/no-go

Three use cases

Use case 1

Input: PR URL, base branch, changed files, staging URL, change request, client note. Output: a three-column table for code review, staging review, and client approval. Human review: pricing, ad copy, publish timing, client OK, production release.

Use this when a campaign page has passed implementation review but contains business-sensitive copy. Ask Codex to identify the exact files and lines behind each finding, then put client wording and release timing in a separate “human decision” column. The useful output is not “approved”; it is a short list of evidence that a director can accept or reject.

Use case 2

Input: PR comments, target files, touched lines, files not to touch, validation command. Output: one fix plan per comment, changed files, verification result, and remaining decision note. Human review: client scope, design shift, estimate impact, launch date.

Use this when review comments arrive from several people. Process one comment at a time and require Codex to name the allowed files before editing. A request that changes the agreed layout, estimate, or client-approved message becomes a blocker instead of being silently folded into a technical fix.

Use case 3

Input: staging URL, viewport widths, form example, main CTA, tracking tag, rollback note, reviewer. Output: desktop/mobile checklist, form result, link list, screenshots, and launch blockers. Human review: client approval, ad timing, contact address, production go/no-go.

Use this after the PR is technically ready. The staging pass should record the tested route and viewport, whether submission reached the expected success state, and which links were opened. Do not paste real customer data into the prompt; use synthetic names, addresses, and inquiry text, then let the human reviewer confirm the destination mailbox and production analytics property.

Live page interaction requires the Browser capability with access to the staging site, including any approval required before a sensitive action such as form submission. Without that capability, Codex should prepare the checklist and mark browser evidence as missing; a human performs the checks and supplies the screenshots and form result.

Copy-paste prompt

You are reviewing a web agency workflow before using Codex Desktop.
Inputs: PR URL, base branch, staging URL, changed files, client approval note, release checklist, and rollback note.
Output:
1. Work Codex can do: diff summary, /review, risk list, fix draft, verification command.
2. Work humans decide: pricing, ad copy, legal wording, client OK, publish timing, production approval.
3. Three-column table: PR fixes, staging checks, production blockers.
4. Short AGENTS.md rules for this agency.
5. One first action that can be done in 30 minutes.
Rules: do not treat review findings as production approval. Do not rewrite client-approved copy without explicit review.

Working check code

const releaseReview = {
  project: "campaign-lp",
  headBranch: "feature/lp-copy",
  baseBranch: "main",
  diffScope: "feature/lp-copy vs main",
  reviewCommand: "codex review --base main",
  stagingUrl: "https://staging.example.com/campaign",
  pullRequestUrl: "https://github.com/example/site/pull/42",
  filesChanged: ["src/pages/campaign.astro", "src/components/LeadForm.tsx"],
  humanChecks: ["mobile layout", "form submit", "tracking tag", "client approval", "rollback note"],
  productionRequirements: ["client approval", "rollback note", "analytics check"],
  clientApprovalReference: "approval-example-001",
  rollbackPlan: "Redeploy the previous production commit, then rerun form and analytics checks.",
  releaseOwner: "human release owner",
  codexCanDo: ["summarize diff", "run review", "list risks", "draft fix prompts"],
  humanMustDecide: ["publish timing", "legal wording", "pricing claim", "production approval"]
};

const hasText = (value) => typeof value === "string" && value.trim().length > 0;
const hasChecklist = (value) =>
  Array.isArray(value) && value.length > 0 && value.every(hasText);
const expectedDiffScope = `${releaseReview.headBranch} vs ${releaseReview.baseBranch}`;
const expectedReviewCommand = `codex review --base ${releaseReview.baseBranch}`;

const required = [
  ["head branch", hasText(releaseReview.headBranch)],
  ["base branch", hasText(releaseReview.baseBranch)],
  ["diff scope matches branches", releaseReview.diffScope === expectedDiffScope],
  ["review command matches base branch", releaseReview.reviewCommand === expectedReviewCommand],
  ["staging URL", hasText(releaseReview.stagingUrl)],
  ["pull request URL", hasText(releaseReview.pullRequestUrl)],
  ["changed files", hasChecklist(releaseReview.filesChanged)],
  ["human checks", hasChecklist(releaseReview.humanChecks)],
  ["production requirements", hasChecklist(releaseReview.productionRequirements)],
  ["client approval reference", hasText(releaseReview.clientApprovalReference)],
  ["rollback plan", hasText(releaseReview.rollbackPlan)],
  ["release owner", hasText(releaseReview.releaseOwner)],
  ["production approval stays human", hasChecklist(releaseReview.humanMustDecide) && releaseReview.humanMustDecide.includes("production approval")]
];

const findings = required
  .filter(([, ok]) => !ok)
  .map(([item]) => ({ item, fix: "add this before using Codex for release review" }));

console.table(findings);
if (findings.length > 0) process.exitCode = 1;

Save the snippet as release-review-check.mjs and run node release-review-check.mjs. An empty table and exit code 0 mean that the minimum review inputs are present; they do not mean the site is approved for production. Set rollbackPlan to an empty string to exercise the failure path: the script prints the missing item and returns a non-zero exit code.

Before attaching the output to a PR, add the evidence that code cannot infer: the reviewer’s name, the client approval reference, the planned release window, and the person who can execute the rollback. That turns a generic AI summary into a release record the next person can audit.

Pitfall: common mistakes

The first cause is reading /review as production approval. The fix is to name the output as diff findings and keep go/no-go with a human.

The second cause is vague AGENTS.md guidance. The fix is to write practical agency gates: mobile width, form submit, CTA, OGP, tracking tag, and rollback note.

The third cause is parallel editing of the same page. The fix is to delegate read-heavy checks, then return all edits to one main thread for review.

The fourth cause is skipping the staging URL. The fix is to keep PR diff, staging URL, screenshots, form result, and rollback note in one release table.

FAQ

Q. Is Codex Desktop the official product name? A. Official docs mainly say ChatGPT desktop app, Codex CLI, IDE extension, and review pane. This article uses Codex Desktop as a practical agency label for that working surface.

Q. Should an agency start with desktop app or CLI? A. Directors often start with the desktop app review pane. Implementers often use CLI or IDE extension for local verification.

Q. Is /review enough before release? A. No. It helps with diff review. Client approval, ad timing, pricing, and production approval stay human.

Q. Should agencies use subagents? A. Yes for read-heavy checks such as links, logs, and summaries. Avoid concurrent edits to the same LP.

Training path

For agencies, the win is not just faster edits. It is fewer releases that need apology emails. Bring a recent PR, staging URL, AGENTS.md, and release checklist to the consultation page. Use Claude Code training and consultation when the team needs help turning PRs, staging URLs, AGENTS.md, and release checklists into a repeatable review workflow.

What I verified

I checked the OpenAI Codex manual, ChatGPT desktop app docs, Code review, AGENTS.md, Subagents, Agent approvals and security, the JavaScript checklist’s success and missing-rollback paths, internal link, CTA, frontmatter, and slug. The first action is to split diff review, PR follow-up, and release approval into three columns.