dotsfeed
← Updates

Codex

Running a big refactor through Codex? Have the planner launch Codex CLI in separate worktrees, and add a second agent to push back

Observed· Oct 20, 2026Published Oct 20, 2026

A developer moved OpenClaw's web frontend from Lit to Solid 2.0 in about two weeks and ~600 PRs by having GPT-6 Astra plan, Codex CLI with GPT-6.1 Sol code in separate worktrees, and a second agent challenge the work. Field report, not independently verified.

What changed

On Oct 20, Peter Steinberger (@steipete) shared how he moved the web frontend of his open-source project OpenClaw from Lit to Solid 2.0 with a long-running agent loop. He started it about two weeks earlier, checked in roughly every other day, and says the process was mostly automatic. The project's public timeline puts the start at Oct 9, with Lit removed from the Control UI on Oct 16 (about 6 days and 18 hours later) and another stretch of cleanup to move the code to idiomatic Solid. It counts 255 PRs that explicitly name the migration plus 352 later UI refactors, which matches his "about 600 PRs." Asked how much human effort it took across prompting, review, and testing, he estimated about one workday in total.

The useful part is how he set it up:

  • GPT-6 Astra did the planning and GPT-6.1 Sol did the coding.
  • Instead of letting the planner spawn subagents, he prompted Astra to run the Codex CLI in separate git worktrees. He says this worked far better than relying on subagents.
  • A second agent watched the work and pushed back to keep the planner on track. He compares it to /goal, except a different agent challenges the work instead of the same agent checking its own progress against a text file. He built this as a plugin for OpenClaw.
  • Because Solid 2.0 is still a release candidate and too new for the models to know well, he gave the agent a cheat-sheet skill with the new APIs.

His result, in his words: a smaller JS and CSS footprint, finer-grained updates, and a UI that feels snappier. His team still found a few small regressions the agent missed, mostly button shapes and slight UX changes.

Who is affected

Developers using Codex (CLI or app) on a paid ChatGPT plan who want to hand a large, multi-day refactor or framework migration to agents. It's most relevant if you've tried subagents for big jobs and found them drifting or stepping on each other.

What to do now

  • Split the roles. Use a stronger planning model to break the migration into small PRs, and have it start a separate Codex CLI run in its own git worktree for each piece, so changes don't collide in one checkout.
  • Add a reviewer that isn't the planner. A second agent that reads the plan and the diffs and argues back catches drift that a single agent grading itself tends to miss.
  • If your target library or framework version is newer than the model's training data, write a short skill or cheat sheet with the current APIs and patterns and give it to the agent before it starts.
  • Check in on a schedule (he did every other day) instead of babysitting, and keep one human pass for visual and UX regressions, since that's what slipped through here.
  • Watch your usage. Hundreds of PRs over two weeks is a lot of tokens, so check your Codex limits before you start a loop this size.

What is not confirmed

This is one developer's field report and hasn't been independently verified. The PR counts and dates come from OpenClaw's own public timeline page, which says its numbers were checked but its interactive widget wasn't. The "smart goal" reviewer plugin wasn't linked publicly, and he didn't share the exact prompts. "Far better than subagents" is his own experience on this project, not a benchmark, and OpenAI hasn't published guidance comparing the two approaches.

Sources

Sources

Get updates like this every morning

  1. ① Email
  2. ② Card on Stripe
  3. ③ 7 days free

Then $2/month · cancel anytime in one click