Skip to content

Codex: delegating website tickets from Linear

Website tickets go to Codex, OpenAI's Codex-for-Linear agent: it runs in OpenAI's cloud, bills against your ChatGPT plan, and drafts a PR. Everything else stays on the Factory, which needs ./bootstrap.py --branch and a local subscription identity.

Environments belong to a ChatGPT account and we use personal accounts, so each person who delegates sets up their own. Do it when you need it.

Short version

  1. Codex settings → Environments → Create environment on inboundx/inboundx.
  2. Env var CODEX_ENV_NODE_VERSION=24. Setup and maintenance script both bash website/bootstrap.sh --sparse. No secrets, internet at default.
  3. Install the Codex integration from Codex settings, then @Codex once on any Linear issue to link your account.
  4. Delegate: @Codex <task> in inboundx/inboundx.

Steps

1. Create the environment

At chatgpt.com/codex/settings/environments, click Create environment. One environment per repository.

Setting Value
Repository inboundx/inboundx
Environment variable CODEX_ENV_NODE_VERSION=24
Setup script bash website/bootstrap.sh --sparse
Maintenance script bash website/bootstrap.sh --sparse
Internet access (agent phase) default (disabled)
Secrets none

Register the script as both. Setup runs on container create, maintenance on cache resume — setup-only means the scoping lapses on every resumed container.

No secrets needed: the site builds with NEXT_PUBLIC_ROSE_API_KEY and PUBLIC_KEYSTATIC_GITHUB_APP_SLUG unset. Internet stays off because setup scripts always have network and astro build runs offline.

website/bootstrap.sh --sparse installs deps and sparse-checkouts the tree to website/, so backend/ and frontend/ are not on disk. It fails the build if that did not apply. It is a guardrail, not a sandbox — git sparse-checkout disable undoes it.

Install the Codex integration from chatgpt.com/codex/settings, then mention @Codex once in a comment on any issue.

3. Delegate

Assign the issue to Codex, or comment:

@Codex <task description> in inboundx/inboundx

Pin the repo — environment selection is ambiguous otherwise. Codex bases work on the default branch (develop), so drafted PRs target correctly.

Preview URLs

Every PR touching website/** gets one from the preview-website job in .github/workflows/deploy-website.yml: wrangler versions upload --preview-alias pr-<n> publishes a version without deploying it, so live staging is untouched and only the aliased URL serves the PR. Posted as a sticky comment, re-points on every push. Builds with CLOUDFLARE_ENV=staging, which is baked into dist/server/wrangler.json — a preview cannot reach production.

If the URL is missing: preview_urls is trigger-class config that versions upload cannot apply. It takes effect only after a full deploy:

gh workflow run deploy-website.yml --ref <branch> -f environment=test

Known gaps

  • Codex cloud screenshot support is unverified. Linear's own coding sessions do capture them; if Codex does not, visual evidence stays Factory-only.
  • Local and cloud Codex tasks share one 5-hour quota window.
  • The environment spans the whole monorepo; only the working tree is scoped. Watch the first few diffs.