Steinberger moved OpenClaw to Solid 2 in about 600 PRs
A framework migration run mostly on autopilot, with check-ins every other day and a handful of button-shaped regressions left behind.
Peter Steinberger kicked off an agent loop two weeks ago to move OpenClaw's Lit-based web frontend to Solid 2.0, and reports it finished in about 600 PRs. Lit was gone after roughly six days. The second week went to moving patterns to idiomatic Solid.
I found that the workflow of prompting Astra to run codex cli in worktrees works FAR better than relying on subagents.
He says he checked in with the session every other day and the process was mostly automatic. The result, by his account, is a smaller JS and CSS footprint and much more fine-grained update control, with the app feeling snappier. The misses were small. He found a few tiny regressions the agent skipped, mostly different shapes on buttons and slight UX regressions.
The setup
Two models split the work. Astra did the planning, Sol 6.1 did the coding. The part Steinberger singles out is how the planner dispatched that coding work.
He says prompting Astra to run codex cli in worktrees (separate checkouts of the same repo, so parallel branches do not fight over one working directory) worked far better than relying on subagents. On top of that he ran a second agent to oversee the first and keep it on track, which he describes as being like a /goal command except that instead of the same agent reviewing a text file, a separate agent challenges the work. He wrote an OpenClaw plugin for the feature and calls it smart goal.
This echoes a pattern other builders have been converging on, including dax running one pinned session that farms work out to subagents. Steinberger's version swaps the subagents for full worktrees driven by a CLI.
Why it lands differently
Most long-horizon agent receipts are greenfield. This one is a framework migration on a codebase that already existed and already worked, which is the harder case, because the bar is not working code, it is working code that behaves the same as before.
Ryan Carniato, who posts about Solid, replied that he is never really sure how to weigh different opinions online about how good agents are at writing Solid, let alone Solid 2.0, but that success cases make him happy. He also joked that these guys might be responsible for a two week spike in his chart.
Treat it as one engineer's report on one codebase. The regression list is self-reported, and nobody outside the project has audited the 600 PRs.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.