Repository navigation
fix(ci): merge a stacked PR against the newest commit it holds of the landed PR below it - #422
Conversation
… landed PR below it Every time a PR of a stack landed, the train refused every PR above it: their merge base was an old main commit, so the train's own commits on the landed PR (its merges of main, the renumbered section, the measured baselines) conflicted with the stacked copies of the same lines. After #329 landed, #330 to #341 all refused on examples/probe_dim_assoc.rs, and a lane hand-merged each with `git merge-tree --merge-base=<newest commit of the landed PR it holds>`. refresh.sh now finds that base itself (the head of each PR main squashed since the merge base, its merge base with the branch, the nearest one main does not hold), merges against it with git merge-recursive and the usual conflict handling, and writes it to merge-base.sha for the oracle comparison. A PR that is not stacked merges exactly as before. arch.py renumber also rewrites references in the PR's lines to a section main renumbered when the PR below landed, matched by title.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: bf7e00baaa
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
arch.py renumber printed every section main had renumbered under a stacked PR's base, even when none of the PR's lines cited one; refresh.sh then ran `git commit -qam` with nothing to commit and refused a mergeable PR. It now prints only the moves that rewrote a line, nothing when none did, and refresh.sh commits the renumber only when the tree changed. Checked at a47e1bf against main: the old script printed "§7bh->§7bi §7bi->§7bj" with no change, the new one prints nothing; at #331's head it prints only 7bj->7bk, at #332's 7bj->7bk 7bk->7bl §7bi->§7bj.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9f71d571f3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…branch holds The landed PR's head and the stacked branch can have several merge bases, a commit of the landed PR and a main commit, and `git merge-base` returns one. When the branch merged main after taking the landed PR's commits, merging against that PR's commit drops the main history the branch already holds, and a change main later reverted comes back. refresh.sh now considers every merge base, and takes one only when each of the branch's ordinary merge bases with main is its ancestor; otherwise the branch merges as it always has. Reproduced in a scratch repository: with the landed PR's commit as base the merge kept the reverted f=1, the new rule rejects that base and git merge gives f=0; a plain stack still takes the PR's commit and merges cleanly where git merge conflicts. On the real branches #331, #332 and #341 still take a47e1bf (#330), and #383 takes none.
Summary
refresh.shmerges a PR that is stacked on a landed PR against the newest commit it holds of that landed PR, instead of the usual merge base. A PR that is not stacked merges exactly as before.The problem. A stacked PR's merge base with main is an old main commit. The train's own commits on the landed lower PR then conflict with the stacked copies of the same lines: its merges of main, its renumbered
ARCHITECTURE.mdsection, its measured baselines. When #329 landed on 2026-10-27, the train refused #330 to #341 onexamples/probe_dim_assoc.rs(train.log 23:49). A lane then hand-merged each PR withgit merge-tree --merge-base=<that commit>, as it had after every earlier landing.How the train finds that base now.
headRefOid; fetched fromrefs/pull/N/headif missing).It then merges against that base with
git merge-recursive, which leaves the same index stages asgit merge, so the existing resolvers run unchanged. It writes MERGE_HEAD so the result is a signed two-parent merge, and records the base inmerge-base.shafororacle_base.py.Cross-references.
arch.py renumberalso rewrites references, in the PR's own lines, to a section that main gave a new letter when the PR below landed. Sections are matched by title, when the title is unique on both sides. This is the step the lanes did by hand, for example #332's "§7bi" for #330.Spec reference
Maintainer tooling only.
Test plan
bash -nandpy_compile.refresh.shon real branches, with push, auto-merge and comment calls replaced by no-ops:stacked on #329: merging against 26feabbdc. That is the base the lane chose by hand. The renumber step moves 7bi..7bo to 7bj..7bp and §7bh to §7bi. The merge then needs only the existingARCHITECTURE.mdresolver, and the train used to refuse here onprobe_dim_assoc.rs. Afterwardsarch.py checkandspec_coverage.py --checkpass, the merge commit is signed with parents (PR, main), and no conflict markers remain.Source-provenance declaration
By submitting this PR I certify that my contribution does not incorporate executable code from any source whose license is incompatible with Apache-2.0. Specifically:
CLEANROOM.md).dwg-rsis Apache-2.0 and must remain compatible.My contribution is based on:
CLEANROOM.md.See
CLEANROOM.mdfor the full allowed/forbidden source list and the honest scope of what "clean-room" means for this project.Developer Certificate of Origin
By submitting this PR, I certify that the contribution complies with the
Developer Certificate of Origin 1.1.
Specifically, I affirm that the contribution is my own work (or rightfully
licensed to me) and that it can be distributed under the project's Apache-2.0
license.
I will sign off commits with
git commit -swhen asked, which adds aSigned-off-bytrailer to the commit message asserting the DCO.