fix(publish): drop macos-13, cross-build darwin-x64 on macos-latest (v0.2.2) (#25)

## Why

GitHub's `macos-13` (Intel) runner pool is being phased out — allocation
queue routinely sits at 1–4+ hours during peak load. The `v0.2.1`
publish run from #24 was held there indefinitely and had to be
cancelled, so `0.2.1` never reached the Marketplace despite the rest of
the matrix succeeding cleanly.

## What

`vsce package --target X` only stamps the .vsix manifest with the target
name — the .vsix's actual native `.node` bindings are whatever happens
to be in `node_modules` at package time. So `darwin-x64` doesn't need an
Intel runner: we can install darwin-x64 optional bindings on
`macos-latest` (M-series, arm64) via pnpm's `supportedArchitectures`
filter and then package as `--target darwin-x64`.

The build step (`pnpm build` runs `tsdown` which uses rolldown) still
needs a runnable native binding for the host arch, so the workflow now
installs twice:

1. `pnpm install --frozen-lockfile` — default `supportedArchitectures`
matches the host; `pnpm build` runs against host bindings.
2. `pnpm install --frozen-lockfile
--config.supportedArchitectures.cpu=${{ matrix.pkg_cpu }}` — swaps
`node_modules` to carry the target arch's bindings before `vsce
package`.

For host==target entries (`win32-x64` on windows, `linux-x64` on ubuntu,
`darwin-arm64` on macos-latest) the second install is effectively a
no-op. For `darwin-x64` on macos-latest it prunes the arm64 bindings and
installs x64 ones, producing a .vsix Intel Mac users can actually run.

## Matrix change

| Target | Old runner | New runner |
|---|---|---|
| win32-x64 | windows-latest | windows-latest *(unchanged)* |
| linux-x64 | ubuntu-latest | ubuntu-latest *(unchanged)* |
| darwin-x64 | **macos-13 (Intel)** | **macos-latest (cross-install x64
bindings)** |
| darwin-arm64 | macos-latest | macos-latest *(unchanged)* |

Same four targets ship. Allocation latency is gone.

## Version

Bumped to `0.2.2`. CHANGELOG notes that `0.2.1` was tagged but never
reached the Marketplace — `0.2.2` is the first version that actually
completes the platform-specific publish flow introduced in `0.2.1`.

## Test plan

- [x] YAML parses; matrix expands to 4 entries with `pkg_cpu` keys
- [ ] Workflow dry-run via `gh workflow run publish.yml --ref
fix/cross-build-darwin-x64 -f dry_run=true` — should complete in ~10–15
min total instead of stalling on Intel Mac allocation
- [ ] After merge: real publish completes; verify all four platform
builds appear on Marketplace and Open VSX under `0.2.2`

## Risks

- **`pnpm install --config.supportedArchitectures.cpu=x64` on arm64
macOS** — assumes pnpm prunes the wrong-arch bindings on the second
install. If it instead leaves both arch bindings present, the
darwin-arm64 .vsix would be ~3–5 MB larger than necessary (extra unused
x64 .node files inside) but would still function. Worth verifying on the
dry run.
- **Wrong-arch `.node` files in node_modules** during the second install
shouldn't trigger postinstall scripts because the napi-rs binding
packages are pure data (just `.node` + `package.json`), not executable
installs. If a future native dep adds a postinstall that
platform-checks, this swap would need to be more surgical.

## Stuck runs

I cancelled the two queued runs from #24:
- `25370742474` (push to main, real publish for 0.2.1)
- `25369816967` (workflow_dispatch dry-run on the feature branch)

Author: AlemTuzlak

.github/workflows/publish.yml

@@ -107,15 +107,27 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
+        # `pkg_cpu` selects which optionalDependency native bindings get
+        # installed in node_modules before `vsce package`. For host==target
+        # (most entries) this matches the runner's arch and is a no-op. For
+        # darwin-x64 we run on macos-latest (M-series, arm64) and cross-
+        # install x64 bindings — `vsce package --target X` only stamps the
+        # manifest, so the .vsix's native bindings come from whatever's in
+        # node_modules. This avoids the macos-13 (Intel) runner, whose
+        # allocation queue can sit at hours during peak load.
         include:
           - target: win32-x64
             runner: windows-latest
+            pkg_cpu: x64
           - target: linux-x64
             runner: ubuntu-latest
+            pkg_cpu: x64
           - target: darwin-x64
-            runner: macos-13
+            runner: macos-latest
+            pkg_cpu: x64
           - target: darwin-arm64
             runner: macos-latest
+            pkg_cpu: arm64
     runs-on: ${{ matrix.runner }}
     steps:
       - name: Checkout
@@ -131,12 +143,24 @@ jobs:
         with:
           node-version: 20.x
 
-      - name: Install dependencies
+      - name: Install dependencies (host arch, for build)
+        # pnpm defaults supportedArchitectures to the runner's host. The
+        # build step needs a runnable native binding (rolldown, oxc-parser,
+        # @tailwindcss/oxide), so always install host bindings first.
         run: pnpm install --frozen-lockfile
 
       - name: Build
         run: pnpm build
 
+      - name: Re-install for target arch (${{ matrix.pkg_cpu }})
+        # Swap node_modules' optional native bindings to match the publish
+        # target. For host==target this is a no-op; for darwin-x64 on the
+        # macos-latest (arm64) runner this prunes the arm64 bindings and
+        # installs the x64 ones, so the resulting .vsix carries exactly
+        # the bindings users on that target need.
+        shell: bash
+        run: pnpm install --frozen-lockfile --config.supportedArchitectures.cpu=${{ matrix.pkg_cpu }}
+
       - name: Strip [Unreleased] from CHANGELOG for packaging
         shell: bash
         run: python3 -c "import re,pathlib;p=pathlib.Path('CHANGELOG.md');t=p.read_text();p.write_text(re.sub(r'## \[Unreleased\]\n.*?(?=\n## \d)','',t,flags=re.DOTALL) if '## [Unreleased]' in t else t)"

CHANGELOG.md

@@ -2,6 +2,16 @@
 
 ## [Unreleased]
 
+## 0.2.2 — 2026-05-05
+
+### Fixed
+
+- Publish pipeline: dropped the dedicated `macos-13` (Intel) runner; GitHub's Intel Mac queue is being wound down and routinely held publish runs for 1–4+ hours waiting for allocation. The `darwin-x64` matrix entry now runs on `macos-latest` (M-series) and cross-installs x64 native bindings via `pnpm --config.supportedArchitectures.cpu=x64` before packaging — `vsce package --target` only stamps the manifest, so the .vsix's bindings come from whatever's in `node_modules`. Same four targets ship; allocation latency is gone.
+
+### Note
+
+- `0.2.1` was tagged but never reached the Marketplace — its publish run was held in queue waiting for `macos-13` allocation when it was cancelled. `0.2.2` is the first version that actually completes the platform-specific publish flow added in `0.2.1`.
+
 ## 0.2.1 — 2026-05-05
 
 ### Fixed

package.json

@@ -1,7 +1,7 @@
 {
   "name": "copilotkit-vscode-extension",
   "displayName": "CopilotKit",
-  "version": "0.2.1",
+  "version": "0.2.2",
   "description": "Preview generative-UI components, explore CopilotKit hooks, and inspect AG-UI agent runs — all without leaving your editor.",
   "categories": [
     "Other",