BiocJobs moves the wrapper inside the package, and the drift problem with it

The design constraint the write-up keeps returning to is that a generated wrapper is only worth having if it is as good as a careful hand-written one, which is why the post rejects a lowest-common-denominator scrape of function signatures in favour of a deliberate declaration by the author ^[7]. That distinction is the entire bet. Automation reading an R function cannot know which of its arguments a batch user should ever be allowed to set; the person who does know is the author, who also knows which entry points make sense with no human in the loop and what the inputs actually mean ^[5].

The DESeq2 example gives the size of the ask. Two inputs, three outputs and nine options ^[13], so fourteen declared items in total ^[15], alongside a job name, a package, a version, a script path and two soft dependencies ^[14]. That is a page of YAML in `inst/biocjobs/`, added without new imports, code changes or build-system requirements ^[12]. It is small enough that a diff against it is a reviewable event at release time, which is the mechanism by which drift is supposed to die: the interface and the code it wraps travel in the same tarball, rather than sitting in a downstream repository maintained by someone else and going stale at every release ^[3].

What the author is not asked to learn is the target systems. A GA4GH TES task is inputs staged in, a short sequence of container-plus-command executors, resource requirements, and outputs collected out, and that sequence is the only structure available ^[10]. One declaration in that shape is claimed to project onto a Galaxy tool, a Nextflow process, a WDL task or a cloud batch submission without rewriting ^[11], which is how a Galaxy conversation ended up producing something broader than a Galaxy generator ^[8].

The minimalism cuts both ways. TES has no branching and no data flow between tasks, and the post is explicit that orchestration belongs to somebody else ^[10]. So what a package owns is a leaf, and the pipeline stays exactly where it was, with whoever maintains it now. The wrapper text stops drifting, but the code that turns a declaration into a Galaxy tool or a Nextflow process lives outside the package, and the material as published says nothing about who maintains those projectors or how they are tested. The declaration format is itself versioned as `biocjobs: "1.0"` ^[14], which concedes that the contract can move under the packages that adopted it.

Opt-in is doing more work here than it looks. Interactive packages simply omit the directory ^[12], which keeps the scheme honest and also means the long tail depends on authors who currently ship no wrapper and therefore receive no complaints about one ^[4]. Nothing described adds a build-system hook ^[12], so nothing described makes an author notice that their declaration has gone wrong. The real test is not DESeq2, whose parameters are well understood by many hands. It is whether a projected tool from a package nobody ever wrapped is good enough that a Galaxy administrator installs it without reading the generator's output line by line.