Aleksei Krasnoperov · October 2026
Everything I build ships through one factory: seven live products and the factory itself. Each agent works its own line: patchrelay from Linear, Claude Code and Codex from sessions I start. Every pull request now goes through review-quill, and merge-steward merges all of them into one queue. The figure below is a simulation, not a live feed.
Issues live in Linear. Delegating one to the patchrelay app lets it write the feature; undelegating pauses it. Once a PR is approved and green, patchrelay may still repair its merge candidate, never the branch.
I triage and plan in an ordinary Claude Code or Codex session: split the work into issues and set blockedBy between the ones that touch the same ground. When patchrelay picks an issue up, a short Codex pass only decides whether it is implementation or an umbrella that coordinates child issues.
One durable worktree per issue. patchrelay drives Codex straight from Linear; I pick up the rest in Claude Code or Codex. Every path ends in a pull request.
review-quill checks out the exact head, reads it against the repo’s REVIEW_WORKFLOW.md, and posts a plain GitHub approve or request changes. Requested changes come back as a review_fix run when patchrelay wrote the code, or land in the waiting ship-pr session when I did.
merge-steward admits a PR approved on its exact head with green checks, then builds future-main candidates, each on top of the one ahead, and runs CI on those SHAs in parallel. A candidate that conflicts with the ones below, or goes red, is repaired in place on its candidate ref, rechecked by a narrow integration review, and tested again; the approved PR head stays frozen, and nothing is stacked on a conflict until it has a candidate.
Landing is a fast-forward of main to the SHA that passed. If main moved first, the candidate is rebuilt on the new tip and tested again, not reviewed again.
A push to main runs the repo’s own Actions: stage and production on Cloudflare Workers, from the same SHA at once. On UserTold each runs a smoke test, and the issue is Done only when both pass.
UserTold interviews the people using what shipped. Interviews become evidence, evidence becomes findings, and the findings I send land in Linear’s Triage as the next work on the line.
Every repair loop (review_fix, ci_repair, queue and integration repair) has a budget, ten tries each on my repos. When one runs out, the issue moves to Human Needed. The agents are told to stop and ask before touching auth, migrations, bindings or dependencies.
The services never call each other. A webhook only wakes them; every decision is re-read from GitHub: head SHA, reviews, checks, refs.
Eleven frozen cases from seven PRs in UserTold, MakeFX, Inventory and Subtitles, replayed through the real prompt in a throwaway worktree. Each is graded on the verdict, how many concerns it raised, no nits where none belong, findings anchored to changed lines, and required or forbidden markers.
patchrelay has no model eval yet. What it has around it is 34 merge-steward scenario suites (speculative cascades, a CI failure mid-chain, crash recovery, a broken main) and 134 test files in its own repo.
Since Opus 5.5, a Claude Code session carries an issue to a pull request on its own, and those sessions have taken over most of what patchrelay’s Codex runs used to do: in October every agent PR came from one. The delivery pipeline stayed. Review, the merge queue and the deploy are still what land the work, whoever wrote it. The spring and summer peak includes building two products twice; both first versions are still counted.
Voice interviews with your real users, inside your product.
Images, video and audio made with your agent on one canvas.
A Spanish notebook for speaking, not just understanding.
A cozy farm game for Spanish vocabulary.
A landscape of founders, drawn only from what they wrote themselves.
A game-master app for Mafia and Werewolf nights.
This site, figure included.
The factory itself: patchrelay, review-quill and merge-steward.
Review interview evidence and turn it into prioritized work.
Counts from GitHub and from review-quill’s and merge-steward’s own records, October 17, 2026.