Skip to content

feat(series): Review a series' R side before any landing, with series-land.sh --review --series - #2970

Merged
krlmlr merged 3 commits into
claude/series-tests-no-engine-gatesfrom
claude/series-review-mode
Oct 18, 2026
Merged

krlmlr merged 3 commits into
claude/series-tests-no-engine-gatesfrom
claude/series-review-mode

Conversation

@krlmlr

@krlmlr krlmlr commented Oct 17, 2026 •

Copy link
Copy Markdown
Contributor

Stacked on #2966 (base claude/series-tests-no-engine-gates), which changes scripts/series-land.sh, scripts/series-config.sh and the series-land skill too, so this diff shows only the new work. #2966 stands on #2961 and #2959, and #2959 also edits handbook/branches/model/README.md, a different paragraph from the one this PR adds to. None of #2958, #2960, #2962, #2965, #2967, #2968 or #2969 touches these files.

Why

The draft review #2964 shows the R side of the main series before any landing, and it had to be built by hand, because scripts/series-land.sh --review falls short in three ways. It reads only a landing: a --landing ref with main below it and vendor commits above. The series instead stands on an older main, the flavor pair, the seed and a commit that brings main's changes since the seed in front of the vendor commits. It hides series_makevars with the strand, which assumes an interface change shows as a tools/stimulus/ edit. That holds for duckdb-r, but rigraph's Stimulus also reads the vendored functions.yaml, so the 19 _impl() functions the engine removes and the 12 it adds would not show at all. And it pushes to --canonical, where it also reads main, while #2964's head lives in the fork.

What

--review --series writes <L>-land-review<s> from <S>-green on --remote, without a landing. It runs scripts/series-rc.sh <S> without --push, which mints the unflavored twin of every <S>-dev commit and writes no ref, and takes the strand on the remote instead where that is aligned. The review is built onto main from the twins above the series' base, which is the seed and whatever stands in front of the first vendor commit. A tooling sync or a port of a commit main has, found among the vendor commits, is left out and listed, by the same rule the landing's replay places one with (placed_by_rule in series-forward-build.sh); any other commit there is refused. After the build, the review is held to the series: outside what it hides, it may differ from <S>-green's unflavored twin only in NEWS.md, the READMEs and DESCRIPTION's version. Anything else is a change of main's the series has not ported, and the run stops with no ref written. This is the base check #2964 was built with, generalized so that the mode keeps working once the loop ports main's next changes among the vendor commits. A vendor commit whose diff no longer applies to main is named. Each review commit carries its series commit's message and author and names it in a Series-commit: trailer; a landing's review keeps Landing-commit:. The test report is read off the rc twins, which carry the series' engine and vendor subjects, so the tests removed with their upstream function are classified as removals (#2966).

series_review_hidden in scripts/series-config.sh names what a review hides besides the workflows and DESCRIPTION's version. rigraph sets it to the strand and what is generated for the compiler and the help pages: src/rinterface.c, src/cpp11.cpp, R/cpp11.R, src/sources-glue-{c,cpp}.mk, src/*.dd and man/. It keeps R/aaa-*.R, the Stimulus-generated _impl() wrappers, because they are the one place the engine's interface change shows in R terms, and the hand-written src/Makevars.{in,ucrt,win} and src/vendor/{config.h,igraph_config.h}. The globs work as git exclude pathspecs, as R/aaa-*.R already did in series_makevars. Where series_review_hidden is undefined, as in duckdb-r's copy, a review hides the strand and series_makevars, as before. The comments in both scripts that said an interface change shows as its tools/stimulus/ change are corrected.

--review-remote <name> names where --push writes the branch, --canonical by default, so a landing's review goes where it went before. The printed gh pr create line names the head as <owner>:<branch>, for example krlmlr:main-land-review. Every review run now writes a local branch of the review's name and prints the full SHA, so a review that is not pushed can still be read and compared. The commit that removes the workflows says that a pull_request_target workflow still runs, because it is read from main: format-suggest.yaml ran on #2964.

The series-land skill documents the mode, the remote, the local branch, what is shown and hidden, and where a series review's findings go: each is verified, and those that hold are collected on main-fwd-build with what the revdep checks find, for a fresh rebuild that makes a new generation of the series. Starting that generation is a person's decision, which the review does not make. handbook/branches/model/README.md lists <L>-land-review beside the other landing refs, with a sentence on reviewing a series before any landing.

Verification

scripts/series-land-test.sh passes 129 checks (84 before). Its new section 14 builds a series seeded on an older main, with main's fix brought in front of its vendor commits as one commit, a tooling sync and a port among them, an interface-only vendor commit, a vendor commit that changes DESCRIPTION beyond its version, and a commit above green. It checks the review's base, the workflow commit and its message, the engine and version kept from main, the unflavored package, the interface's R half and the hand-written build files shown, its C half, the registration, the source and dependency lists and man/ hidden, a vendor commit that changes generated C alone left out, nothing above green, the subjects, the Series-commit: trailer and message, the test report, the local branch without --push, the push to --review-remote with the <owner>: head, the same SHA from a rerun and from an aligned rc strand, the hidden Makevars where series_review_hidden is undefined, and the refusals: a change of main's the series lacks, a vendor commit that does not apply to main, a commit among the vendor commits that is not main's, and --series without --review or with --landing. series-args-test.sh (87), series-lib-test.sh (41), series-tests-test.sh (49), series-forward-build-test.sh (50), series-port-test.sh (17), series-patch-sync-test.sh (32), series-fold-test.sh (26), series-advance-test.sh (149), series-vendor-test.sh (51), vendor-one-test.sh (78) and rcc-store-test.sh pass, shellcheck is clean on the changed scripts, and check-handbook.sh is clean.

The review was rebuilt locally for real, without pushing, with scripts/series-land.sh main --review --series --suffix -hidden --remote origin --canonical upstream --review-remote origin, from main-green at krlmlr/rigraph@2242c53ee2 onto main at 267cc39. It reports 104 of 301 series commits changing the R side, 51 files, +1,778 −1,653, and its test report ends in TESTS KEPT ... but 29 removed with their function upstream. Its diff shows 16 R/aaa-*.R files, the Makevars and the two config headers, and no file the configuration hides.

Limits

series-rc.sh is not changed, because #2958 edits it and this PR is stacked elsewhere, so the mode reads the strand's tip off the line series-rc.sh prints when it is not aligned, and checks it by its Unflavored-from: trailer. series-rc.sh unflavors all of <S>-dev, so a commit above green that it cannot unflavor stops a review of green too. series-args-test.sh is not extended for the two new options, since #2959 to #2961 edit it; both follow the contract, which its existing checks cover.

LLM use: this PR was prepared with Claude Code (Claude Opus 5.5), at the maintainer's request.

  • By submitting this pull request, I assign the copyright of my contribution to The igraph development team.

Generated by Claude Code

…s-land.sh --review --series`

`--review` wrote `<L>-land-review` only from a landing,
hid the generated R/C interface,
and pushed to the repository it reads `main` from.
#2964 was built by hand around all three.

`--review --series` writes the branch from `<S>-green`,
unflavored by `series-rc.sh` without a push,
taking the vendor commits above the series' base,
leaving out the tooling syncs and ports `main` has,
and refusing while `main` has a change the series lacks.
Each commit names its series commit in a `Series-commit:` trailer.

`series_review_hidden` in `series-config.sh` says what a review hides:
the strand alone for rigraph, whose Stimulus reads the vendored `functions.yaml`,
and the strand with its Makevars where it is undefined.

`--review-remote` names where the branch is pushed, `--canonical` by default,
and the run prints the pull request's head as `<owner>:<branch>`.
Every review is kept as a local branch, and named by its full SHA.
The first commit says that a `pull_request_target` workflow still runs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JXnA76zWTu4ZD8BQCSYRpk
@krlmlr
krlmlr added this pull request to stack #2971 October 17, 2026 06:05
…a review, and keep `R/aaa-*.R`

A review showed the whole generated interface,
so the glue a reviewer has to read sat among `src/rinterface.c`,
the cpp11 registration, the source and dependency lists and the help pages.

`series_review_hidden` now names the strand and those files.
`R/aaa-*.R` stays, because Stimulus reads the vendored `functions.yaml`,
and it is the one place an `_impl()` function the engine adds or removes shows in R terms.
The Makevars and `src/vendor/config.h`, which are written by hand, stay too.
`src/*.dd` works as an exclude pathspec as `R/aaa-*.R` already did.

The test pins what is shown and what is hidden,
including a vendor commit that changes generated C alone and is left out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JXnA76zWTu4ZD8BQCSYRpk
…, not into the running series

A review of the series before any landing said nothing about where its findings go.
They are verified one by one, then collected on `main-fwd-build` with what the revdep checks find,
and a fresh rebuild from there makes a new generation of the series.
Starting it is a person's decision, which the review does not make.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JXnA76zWTu4ZD8BQCSYRpk
@krlmlr
krlmlr removed this pull request from stack #2971 October 17, 2026 07:29
@krlmlr
krlmlr added this pull request to stack #2979 October 17, 2026 07:34
@krlmlr
krlmlr merged commit 2472f78 into main Oct 18, 2026
13 checks passed
@krlmlr
krlmlr deleted the claude/series-review-mode branch October 18, 2026 09:07
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants