mhoennigandClaude Opus 5
e61adfbbaa
Die offene Frage im PR-Dokument sagte „löst sich beim Merge". Das Log sagt es genauer: Am Merge-Commit wurde `werkdock` noch eingereiht, erst am nächsten nicht mehr. Die Projektschicht ist die Datei im Arbeitsbaum des beobachteten Klons — die Definition überlebt, bis dieser Baum an der Entfernung vorbei ist, der Merge selbst bekommt also einen letzten Geister-Build. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
5.1 KiB
WARNING: This document describes only the change applied in this PR. It may already be outdated once the next PR is merged. Historic PR-documentation is not maintained along with new PRs — treat it as a snapshot, not as current documentation.
The Problem
Werkdock was grown inside this repository on purpose — plan step 21, sessions B and C — and the plan said from the first line that it "moves to its own repository later".
Everything that move needs has been true since session C: the tool has no imports from Werkator code, no Gradle coupling, and Werkator reaches it the way it reaches git — as a binary on the PATH, named by the pinned key bwrap.werkdock.
What kept it here was three references, not a dependency:
.werkator.yml carried a second build definition for a Go module that has nothing to do with the Kotlin build around it,
tools/remote built the binary out of the subdirectory,
and the documentation described a directory that was about to stop existing.
The instance makes the cost visible: every Werkator branch built a <branch>@werkdock artifact, so Werkdock was rebuilt for changes that could not possibly affect it, and its own history was invisible under Werkator's branch names.
Non-Goals
- Changing Werkdock itself: the extracted repository holds the same nine commits, byte for byte, only without the path prefix.
- Changing how Werkator uses Werkdock:
bwrap.enabled,bwrap.rootfsandbwrap.werkdockstay as they are, still pinned, still resolved viaPATHat run time. - Publishing releases of the binary:
tools/remotekeeps installing the locally built one (ADR 0006 — never built on the target). - Registering the new repository on the instance: that is one
tools/remote werkator repo-addplus a registry entry, done in the rollout, not in this branch.
The Solution
The history moves with the files.
git subtree split -P werkdock lifts the nine commits that touched the subdirectory into a root-level history; the new repository starts from that, so git log and git blame keep working across the move.
A plain copy would have made this PR cheaper and the tool's past unreadable.
The build definition follows, and becomes the only one.
What was builds.werkdock here is builds.default in the new repository — same commands, minus the cd werkdock prefix, same gofmt-twice idiom (the first call names the unformatted files, the second fails the build on them).
This repository's .werkator.yml is back to one definition, and a Werkator branch no longer rebuilds a Go module it cannot influence.
tools/remote asks for a checkout instead of a subdirectory.
WERKDOCK_REPO names it (default: a sibling of this repository — the usual layout when you work on both), WERKDOCK_BINARY the built binary within it, and both are overridable from the environment like every other transport value.
A missing checkout now fails with the clone URL in the message rather than a bare "file not found": the script cannot fix the situation itself any more, so it says what would.
The documentation stops describing a subdirectory.
docs/deployment.md links the repository instead of ../werkdock/README.md, the plan index and the architecture skill say the extraction happened, and plan step 21 gains session E with what was decided and what it left alone.
Verification
./gradlew ktlintFormat buildgreen in this repository withoutwerkdock/.- The extracted repository builds and tests green on its own —
gofmt,go vet,go test ./...(four packages),go build— which is the acceptance criterion the plan set for a self-contained tool. bash -n tools/remote, and the sibling-checkout default resolved against a real checkout.
Open Questions
A branch cannot remove a build definition — only add or override one.
Observed on this very branch: the instance built 21e-werkdock-own-repo@werkdock and it failed, although the branch's committed .werkator.yml no longer defines werkdock.
The cause is not a bug in the watcher — it reads the branch layer, and no fallback warning was logged — but the merge itself:
withBranchLayer does deepMerge(loadRaw(workingDir), stripPinned(branchLayer)), so a definition present in the project layer and absent on the branch survives the merge.
AGENTS.md and the architecture skill describe the branch layer as winning "including the whole builds section", which reads as replacement.
The mismatch resolves itself on merge, but one commit later than one would guess, and the watcher log shows exactly where:
at the merge commit ac53b5f both default and werkdock were still enqueued, at the next commit only default.
The project layer is the .werkator.yml in the watched clone's working tree, so the definition survives until that tree has advanced past its removal — the merge itself still gets one last stray build.
It does not block this PR.
It is a decision, not an oversight to fix in passing: either builds is replaced as a whole (then a branch can retire a build, and a branch that only adds one must repeat the others), or the merge stays and the two documents are corrected to say that removal is not expressible on a branch.