Codex leaving lint errors behind? Ultracite v8's Codex hook sends what autofix can't fix straight back to the agent

#What changed

Hayden Bleasel (@haydenbleasel) released Ultracite v8, which he describes as the linter and formatter for agentic development. According to Ultracite's changelog and docs:

• v8 writes Ultracite's rules into AGENTS.md once, whichever agents you pick, and only adds separate rule files for the few agents that need their own.
• The Codex hook (written to .codex/hooks.json) matches on apply_patch, the tool Codex edits files with. After each patch it runs ultracite fix --hook on just the files that patch added, changed, or moved, not the whole project.
• Anything autofix can't settle is printed with exit code 2. Codex shows that output to the agent in place of the patch result, so Codex fixes it on its next step.
• ultracite fix --codex runs the normal autofix pass, then hands the remaining diagnostics to the Codex CLI.
• The announcement also lists config upgrades with Codex, leaner ESLint presets, and git hooks that fix staged files.

#Who is affected

Codex CLI and Codex app users working in JavaScript or TypeScript repos that use Oxlint and Oxfmt, Biome, or ESLint with Prettier. It is not tied to any ChatGPT plan.

#What to do now

1. Run npx ultracite@latest init and choose Codex when it asks which agent hooks to enable. In a project that already uses Ultracite, npx ultracite init --hooks codex adds just the hook.
2. Trust the project in Codex and review the hook in /hooks. Codex won't run a project's hooks until you do, and it asks again whenever re-running init changes the hook command.
3. If you keep a long AGENTS.md of your own, check it still fits under Codex's 32 KiB loading limit after Ultracite adds its rules (see https://dotsfeed.com/updates/codex-agents-md-32-kib-limit-project-doc-max-bytes).
4. For a one-off cleanup of an existing repo, run npx ultracite fix --codex.

#What is not confirmed

We haven't run v8 ourselves; the behavior above comes from Ultracite's docs and changelog plus the announcement post. The hook runs after Codex's patch is already applied, so it can't stop a bad edit, only flag it for the next step. The full v8 release notes shown in the announcement image weren't checked line by line.

Sources: Hayden Bleasel on X (https://x.com/haydenbleasel/status/2109691326229053710) and Ultracite docs on agent hooks (https://www.ultracite.ai/hooks).