[renatoathaydes]

Summary

devpod up <folder> streamed the whole folder (minus .devpodignore) into a remote machine or Kubernetes pod, including .git, node_modules, build output and nested git worktrees. On a checkout of this repository that was 8.8 GB; it is now 120 MB (35 s → 2.6 s to build the tar).

When the workspace folder is in a git work tree:

• Files: only what git ls-files --cached --others --exclude-standard lists is uploaded: tracked files and untracked files no .gitignore, .git/info/exclude or core.excludesFile ignores. Files of checked out submodules are included; nested repositories and worktrees are not. Folders git ignores are never read. Names are matched in Unicode NFC (and case-insensitively when core.ignorecase is set), and listed files are archived with git's spelling, so macOS decomposed names and case-only renames match the index.
• .devpodignore takes precedence (like ripgrep's .ignore over .gitignore): its patterns leave out listed files, and its negation patterns upload files git ignores (e.g. !.env). It keeps its .dockerignore syntax.
• .git: at the top of the work tree, the repository's .git is replaced by a shallow one built with git fetch --depth from the local repository: the checked out commit, the branch, its upstream branch (fetched deep enough to reach the merge base, so ahead/behind counts are right), the remotes and info/exclude. It has no user.* config, so it doesn't override the identity Give a workspace the git identity of its repository #28 puts in the container's global config. Enough to commit and push from the workspace; staged changes show as unstaged. Each checked out submodule gets its own shallow .git (embedded in its folder), and the submodule config is copied. Tracked files .devpodignore leaves out are marked skip-worktree, so they don't show as deleted (their blobs are still in .git). The snapshot uses the local repository's object and ref formats, not the user's init.* defaults. This also fixes uploads from a git worktree, whose .git file pointed at a path missing in the container. !.git in .devpodignore uploads the whole .git, .git uploads none.
• Without git, outside a git work tree, or when git lists nothing (e.g. a dotfiles repo that ignores everything), behaviour is unchanged. Local Docker bind-mounts the folder and is unaffected.

Changes

• pkg/extract: ExcludeRules (tri-state verdict: unmatched / excluded / re-included), FileList, and TarOptions.Files / TarOptions.Overlays (folders replaced at any path).
• pkg/git/worktree.go: ListFiles, IsTopLevel, IgnoresCase, Snapshot.
• go.mod: golang.org/x/text (already vendored) is now a direct dependency.
• pkg/agent/tunnelserver: workspaceTarOptions used by StreamWorkspace and StreamMount of the workspace mount.
• Docs: "Which files are uploaded" in Create a Workspace; CHANGELOG entry.

Testing

• Unit tests for the archiver (including NFD and case-only spellings), for git listing and snapshots against real repositories (nested submodules, nested repos, worktrees, detached HEAD, unborn branch, ahead/behind upstream including a merged upstream, skip-worktree, ref format, config copying, commit and push from the snapshot), and for the tunnel server streaming a git workspace, extracting it and checking git status is clean.
• An independent review found the issues fixed in the last commit; each has a regression test.
• go test ./pkg/... ./cmd/..., go vet, golangci-lint v1.60.1: pass.
• Not yet tried with a real devpod up against a remote machine or Kubernetes cluster.

🤖 Generated with Claude Code