​

UAT Batch Trains

When agents land merge-ready features faster than you can review and merge them one at a time, the bottleneck becomes the last-mile UAT. UAT batch trains fix that: Overdeck keeps one assembled, testable batch of every ready feature available at all times, so you can exercise the combined result and then land exactly what you tested — in a single click.

​

What you see

On the Flywheel page, the UAT batches card shows, newest first:

• A ready batch — a uat/<codename> branch bundling every ready feature, with the conflicts between them already resolved. It offers Open UAT frontend (a live dashboard serving that exact branch), a per-feature What to UAT checklist, and Merge batch (N) to main.
• A batch assembling in the background when a new feature becomes ready — the current batch stays fully testable until the new one is ready and supersedes it.
• Held-out features, with the reason, when a conflict couldn’t be resolved automatically. Nothing blocks the batch; the held-out feature is retried in a later generation.

​

How it works

1. Assemble. A 60-second reconciler builds a generation — a throwaway uat/<label>-<codename>-<MMDD> branch off main — merging each ready feature in order. When two features conflict, a timeboxed assembly agent resolves the conflict inside the batch; if it can’t, that feature is held out (never the whole batch).
2. Accumulate. New ready features trigger a new, larger generation in the background. Generations accumulate — you can test and promote an earlier, smaller batch if the newest isn’t what you want.
3. Test. Click Open UAT frontend to exercise the combined branch on a real running stack. The checklist pulls each feature’s acceptance criteria, plus an explicit “verify the touchpoint” item wherever a conflict was resolved.
4. Promote. Merge batch (N) to main lands exactly the tree you tested — conflict resolutions included — in one merge. Each feature then closes out through the normal post-merge flow.

Merging a single feature on its own is still available as an escape hatch; doing so simply restales the live batches and a fresh one reassembles automatically.

​

Projects that span several repos

A polyrepo project — one where a single issue’s work lands across several member repositories — gets batch trains too, on the same terms as a single-repo project. There is no separate setting: the merge-train flag is the only gate. A batch there spans one uat/… branch per member repo that a ready feature actually touches. Repos no feature touches are left alone entirely. The batch still behaves as one unit:

• A feature that conflicts anywhere is held out everywhere. You never test a feature that landed in the API repo but not the frontend one — if it can’t be merged cleanly in every repo it contributes to, it waits for the next batch.
• Merging is all-or-nothing. Overdeck first merges every repo locally, without publishing anything. Only once all of them succeed does it push. If any repo would fail, nothing is published at all and the batch stays exactly as it was.
• An interrupted merge resumes instead of unwinding. If a push fails after some repos already landed, Overdeck tells you which landed and which are still pending, and never rewrites a repo’s main to undo one. Clicking merge again skips what already landed and finishes the rest.

The live UAT stack works the same way as for a single-repo project — one stack per batch, serving all of its repos together.

​

Live UAT stacks

Each generation can serve a live dashboard at uat-<codename>.overdeck.localhost, spun on demand from the batch’s worktree. Overdeck runs at most two UAT stacks at once (Docker’s network pool is limited) and tears down the oldest when you start a third; promoting or invalidating a batch always tears its stack down.

​

Honest controls

Every action names its exact effect and confirms before anything happens — “Merge batch (3) to main” lists the three issues and states that it lands the tree you tested. There is no opaque “Ship batch” button: a batch builds itself, and nothing reaches main until you say so.