Keep the GPUI forks as patches, not as vendored trees

[SDSLeon]

[patch.crates-io] pointed at five in-tree copies of published packages — 207 files and roughly 153,000 lines, of which +1,644/−397 were ours. Those copies were a derived artifact: pinned published source plus the patches this repository already generates and verifies. This keeps only the smaller half.

247 files changed, 1,099 insertions, 158,984 deletions.

How it works

.shots/gpui_patches.py --materialize writes the pinned published packages into the gitignored .vendor/ tree and applies each patch exactly; [patch.crates-io] names those paths. --check no longer compares an in-tree fork against the registry — it proves each patch applies to the published source with no fuzz and reproduces .vendor/ byte for byte, so the patch stays the complete statement of the deviation.

The cost, stated plainly

Cargo resolves [patch] paths at manifest load, before any build script, so nothing can run earlier: a fresh clone must materialize before its first cargo command, rust-analyzer included, or cargo aborts with failed to load source for dependency. Every CI job that touches cargo runs the step first, through .github/actions/rust-env or its own copy. A warm run is a hash comparison finishing in under a second.

Verified, not assumed

Against a git archive extraction of this commit — a genuine fresh clone:

• without the bootstrap, cargo metadata fails with the documented error (exit 101);
• with it, cargo check --workspace --locked exits 0.

The cold-runner path was the real risk: a fresh runner has no gpui-pre in its registry, yet cargo cannot run in this workspace until .vendor/ exists. Tested with an empty CARGO_HOME — cargo info, run outside the workspace, breaks the cycle, and the resulting tree is byte-identical to the warm one.

Two things that had to keep working:

• The gpui-pre-web fork's default = [] is what keeps wasm_thread out of the wasm32 graph and the wasm job on stable. It now arrives as a hunk in docs/upstream/patches/gpui-pre-web-0.3.3.patch; the wasm build succeeds on stable and cargo tree -i wasm_thread reports no such package.
• The patches are build inputs on all three hosts now, so *.patch text eol=lf pins them — * text=auto would have handed the Windows runner CRLF copies that patch rejects.

No git dependency is introduced anywhere, so cargo publish is unaffected.

Local gates, all exit 0: run-tests.sh --workspace --locked (113/113) · clippy -D warnings · rustdoc -D warnings · fmt · hack --each-feature · deny · machete · gpui_patches --check/--self-test · inventory · coverage · all 14 parity audits · wasm.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
herogpui	Ready	Preview	Sep 14, 2026 9:20pm UTC