Merge pull request #33 from ArchiveBox/claude/magical-pascal-iR9iO

Honor PUPPETEER_CACHE_DIR + PLAYWRIGHT_BROWSERS_PATH env vars

Author: pirate

.github/workflows/deploy-pages.yml

@@ -16,7 +16,8 @@ concurrency:
 
 jobs:
   build:
-    runs-on: ubuntu-latest
+    runs-on:
+      group: Default
     timeout-minutes: 20
     steps:
       - name: Checkout
@@ -48,7 +49,8 @@ jobs:
     environment:
       name: github-pages
       url: ${{ steps.deployment.outputs.page_url }}
-    runs-on: ubuntu-latest
+    runs-on:
+      group: Default
     needs: build
     timeout-minutes: 20
     steps:

.github/workflows/release.yml

@@ -19,7 +19,8 @@ concurrency:
 
 jobs:
   release-state:
-    runs-on: ubuntu-latest
+    runs-on:
+      group: Default
     timeout-minutes: 20
     steps:
       - uses: actions/checkout@v6

.github/workflows/tests.yml

@@ -19,7 +19,8 @@ concurrency:
 
 jobs:
   precheck:
-    runs-on: ubuntu-latest
+    runs-on:
+      group: Default
     timeout-minutes: 20
 
     steps:
@@ -30,6 +31,11 @@ jobs:
       with:
         python-version: '3.12'
 
+    - name: Setup Node
+      uses: actions/setup-node@v6
+      with:
+        node-version: '22'
+
     - name: Install uv
       uses: astral-sh/setup-uv@v8.0.0
       with:
@@ -49,7 +55,8 @@ jobs:
 
   discover-standard-tests:
     needs: precheck
-    runs-on: ubuntu-latest
+    runs-on:
+      group: Default
     timeout-minutes: 20
     outputs:
       test-files: ${{ steps.set-matrix.outputs.test-files }}
@@ -78,7 +85,7 @@ jobs:
         echo "$json_array"
 
   build:
-    name: ${{ matrix.target.os }} py${{ matrix.target.python_version }} ${{ matrix.test.name }}
+    name: ${{ matrix.target.os_name }} py${{ matrix.target.python_version }} ${{ matrix.test.name }}
     needs: [precheck, discover-standard-tests]
     runs-on: ${{ matrix.target.os }}
     timeout-minutes: 20
@@ -88,11 +95,16 @@ jobs:
       max-parallel: 20
       matrix:
         target:
-        - os: ubuntu-latest
+        - os:
+            group: Default
+          os_name: linux
           python_version: '3.11'
-        - os: ubuntu-latest
+        - os:
+            group: Default
+          os_name: linux
           python_version: '3.14'
         - os: macOS-latest
+          os_name: macOS
           python_version: '3.13'
         test: ${{ fromJson(needs.discover-standard-tests.outputs.test-files) }}
 
@@ -219,7 +231,8 @@ jobs:
 
   discover-live-tests:
     needs: precheck
-    runs-on: ubuntu-latest
+    runs-on:
+      group: Default
     timeout-minutes: 20
     outputs:
       live-tests: ${{ steps.set-matrix.outputs.live-tests }}
@@ -239,16 +252,20 @@ jobs:
             needs_docker=true
           fi
 
-          os_targets="ubuntu-latest macOS-latest"
+          os_targets="linux macOS"
           if grep -q "@pytest.mark.docker_required" "$test_file" || grep -q "@pytest.mark.root_required" "$test_file" || grep -q 'skipif("darwin"' "$test_file"; then
-            os_targets="ubuntu-latest"
+            os_targets="linux"
           elif grep -q 'require_tool("brew")' "$test_file" && ! grep -q 'require_tool("apt-get")' "$test_file" && ! grep -q "operations.apt.packages" "$test_file" && ! grep -q "ansible.builtin.apt" "$test_file"; then
-            os_targets="macOS-latest"
+            os_targets="macOS"
           fi
 
           for os_target in $os_targets; do
-            os_name=$(printf '%s' "$os_target" | tr '[:upper:]' '[:lower:]' | sed 's/-latest//')
-            entry="{\"name\":\"${test_name}-${os_name}\",\"path\":\"$test_file\",\"os\":\"$os_target\",\"needs_docker\":$needs_docker}"
+            if [ "$os_target" = "linux" ]; then
+              os_json='{"group":"Default"}'
+            else
+              os_json='"macOS-latest"'
+            fi
+            entry="{\"name\":\"${test_name}-${os_target}\",\"path\":\"$test_file\",\"os\":${os_json},\"os_name\":\"${os_target}\",\"needs_docker\":$needs_docker}"
             if [ "$first" = true ]; then first=false; else json_array+=","; fi
             json_array+="$entry"
           done

README.md

@@ -1077,7 +1077,8 @@ install_root = $ABXPKG_PUPPETEER_ROOT or $ABXPKG_LIB_DIR/puppeteer
 bin_dir = <install_root>/bin
 ```
 
-- Install root: set `install_root` for the root dir and `bin_dir` for symlinked executables. Downloaded browser artifacts live under `<install_root>/cache` when an install root is pinned. Leave it unset for ambient/global mode, where cache ownership stays with the host and `INSTALLER_BINARY` must already be resolvable from the ambient provider set.
+- Install root: set `install_root` for the root dir and `bin_dir` for symlinked executables. Leave it unset for ambient/global mode, where cache ownership stays with the host and `INSTALLER_BINARY` must already be resolvable from the ambient provider set.
+- Browser cache: when `install_root` is pinned, abxpkg manages `<install_root>/cache` end-to-end — it's exported as `PUPPETEER_CACHE_DIR` to every subprocess, used for `--path=` on `puppeteer-browsers install` / `list`, and `uninstall()` resolves the real browser directory via `load()` then rmtrees it. When `install_root` is unset the provider is in pure passthrough mode: the caller's ambient `$PUPPETEER_CACHE_DIR` (or the CLI's `~/.cache/puppeteer` default) flows through to subprocesses unchanged, `load()` trusts whatever path `puppeteer-browsers list` reports, and `uninstall()` still rmtrees the real browser directory returned by `load()` — leaving any unrelated browsers in the shared cache alone.
 - Auto-switching: bootstraps `@puppeteer/browsers` through `NpmProvider` and then uses that CLI for browser installs.
 - `dry_run`: shared behavior.
 - Security: `min_release_age` is unsupported for browser installs and is ignored with a warning if explicitly requested. `postinstall_scripts=False` is supported for the underlying npm bootstrap path, and `ABXPKG_POSTINSTALL_SCRIPTS` hydrates the provider default here.
@@ -1094,18 +1095,19 @@ Source: [`abxpkg/binprovider_playwright.py`](./abxpkg/binprovider_playwright.py)
 ```python
 INSTALLER_BIN = "playwright"
 PATH = ""
-install_root = None              # when set, doubles as PLAYWRIGHT_BROWSERS_PATH
+install_root = None              # abxpkg-managed root dir for bin_dir / nested npm prefix
 bin_dir = <install_root>/bin     # symlink dir for resolved browsers
 euid = 0                         # routes exec() through sudo-first-then-fallback
 ```
 
-- Install root: set `install_root` to pin both the abxpkg root dir AND `PLAYWRIGHT_BROWSERS_PATH` to the same directory. Leave it unset to let playwright use its own OS-default browsers path (`~/.cache/ms-playwright` on Linux etc.) — in that case abxpkg maintains no symlink dir or npm prefix at all, the `playwright` npm CLI bootstraps against the host's npm default, and `load()` returns the resolved `executablePath()` directly. `bin_dir` overrides the symlink directory when `install_root` is pinned.
+- Install root: set `install_root` to pin the abxpkg-managed root dir (where `bin_dir` symlinks and the nested npm prefix live). Leave it unset to let playwright use its own OS-default browsers path (`~/.cache/ms-playwright` on Linux etc.) — in that case abxpkg maintains no symlink dir or npm prefix at all, the `playwright` npm CLI bootstraps against the host's npm default, and `load()` returns the resolved `executablePath()` directly. `bin_dir` overrides the symlink directory when `install_root` is pinned.
+- Browser cache: when `install_root` is pinned, abxpkg manages `<install_root>/cache` end-to-end — exported as `PLAYWRIGHT_BROWSERS_PATH` to every subprocess (including the `env KEY=VAL -- ...` wrapper used when we go through sudo), used to scope `executablePath()` hits on `load()`, and `uninstall()` resolves the real browser directory via `load()` then rmtrees it. When `install_root` is unset the provider is in pure passthrough mode: the caller's ambient `$PLAYWRIGHT_BROWSERS_PATH` (or playwright's `~/.cache/ms-playwright` default on Linux) flows through to subprocesses unchanged, `load()` trusts whatever path `executablePath()` reports, and `uninstall()` still rmtrees the real browser directory returned by `load()`.
 - Auto-switching: bootstraps the `playwright` npm package through `NpmProvider`, then runs `playwright install --with-deps <install_args>` against it. Resolves each installed browser's real executable via the `playwright-core` Node.js API (`chromium.executablePath()` etc.) and writes a symlink into `bin_dir` when one is configured.
 - `dry_run`: shared behavior — the install handler short-circuits to a placeholder without touching the host.
 - Privilege handling: `--with-deps` installs system packages and requires root on Linux. ``euid`` defaults to ``0``, which routes every ``exec()`` call through the base ``BinProvider.exec`` sudo-first-then-fallback path — it tries ``sudo -n -- playwright install --with-deps ...`` first on non-root hosts, falls back to running the command directly if sudo fails or isn't available, and merges both stderr outputs into the final error if both attempts fail.
 - Security: `min_release_age` and `postinstall_scripts=False` are unsupported for browser installs and are ignored with a warning if explicitly requested.
 - Overrides: `install_args` are appended onto `playwright install` after `playwright_install_args` (defaults to `["--with-deps"]`) and passed through verbatim — use whatever browser names / flags the `playwright install` CLI accepts (`chromium`, `firefox`, `webkit`, `--no-shell`, `--only-shell`, `--force`, etc.).
-- Notes: `update()` bumps the `playwright` npm package in `install_root` first (via `NpmProvider.update`) so its pinned browser versions refresh, then re-runs `playwright install --force <install_args>` to pull any new browser builds. `uninstall()` removes the relevant `<bin_name>-*/` directories from `install_root` alongside the bin-dir symlink, since `playwright uninstall` only drops *unused* browsers on its own. Both `update()` and `uninstall()` leave playwright's OS-default cache untouched when `install_root` is unset.
+- Notes: `update()` bumps the `playwright` npm package in `install_root` first (via `NpmProvider.update`) so its pinned browser versions refresh, then re-runs `playwright install --force <install_args>` to pull any new browser builds. `uninstall()` resolves the browser's real install directory via `playwright-core`'s `executablePath()`, walks up to the containing `<bin_name>-<buildId>/` dir, and rmtrees that dir — in both managed and passthrough modes — because `playwright uninstall` itself has no per-browser argument and only drops *unused* browsers wholesale.
 
 </details>
 

abxpkg/binprovider_apt.py

@@ -28,8 +28,14 @@ class AptProvider(BinProvider):
 
     def setup_PATH(self, no_cache: bool = False) -> None:
         """Populate PATH on first use from dpkg-discovered package runtime bin dirs, not from apt-get itself."""
-        if no_cache or (
-            self._INSTALLER_BINARY is None
+        # Rebuild PATH on first use, when the caller forces no_cache, or when
+        # PATH is still empty — the last case covers the "INSTALLER_BINARY was
+        # resolved out-of-band (hook preflight etc.), so _INSTALLER_BINARY is
+        # non-None but self.PATH was never populated" race.
+        if (
+            no_cache
+            or not self.PATH
+            or self._INSTALLER_BINARY is None
             or self._INSTALLER_BINARY.loaded_abspath is None
         ):
             dpkg_binary = EnvProvider().load("dpkg")

abxpkg/binprovider_brew.py

@@ -183,15 +183,29 @@ def _refresh_bin_link(
                 except Exception:
                     break
                 walk_path = walk_path.parent
+        # Idempotent refresh: skip when shim already points at target.
+        # Rewriting on every load() bumps mtime and churns the inode,
+        # which invalidates fingerprint caches unnecessarily.
+        if link_path.is_symlink():
+            try:
+                if link_path.readlink() == Path(target):
+                    return TypeAdapter(HostBinPath).validate_python(link_path)
+            except OSError:
+                pass
         if link_path.exists() or link_path.is_symlink():
             link_path.unlink(missing_ok=True)
         link_path.symlink_to(target)
         return TypeAdapter(HostBinPath).validate_python(link_path)
 
     def setup_PATH(self, no_cache: bool = False) -> None:
         """Populate PATH on first use from the resolved brew prefix and known runtime brew bin dirs."""
-        if no_cache or (
-            self._INSTALLER_BINARY is None
+        # Rebuild PATH on first use, when the caller forces no_cache, or when
+        # PATH is still empty — the last case covers provider copies that
+        # inherited a resolved ``_INSTALLER_BINARY`` but an unset ``PATH``.
+        if (
+            no_cache
+            or not self.PATH
+            or self._INSTALLER_BINARY is None
             or self._INSTALLER_BINARY.loaded_abspath is None
         ):
             install_root = self.install_root
@@ -391,12 +405,12 @@ def default_abspath_handler(
         if not self.PATH:
             return None
 
+        # Authoritative lookup: search brew's own Cellar / opt / PATH
+        # entries for the real formula binary. The managed ``bin_dir``
+        # shim is a convenience side-effect of install — never a source
+        # of truth — so we always consult brew's paths first and only
+        # refresh the shim to match the freshly-resolved target.
         linked_bin = self._linked_bin_path(bin_name)
-        if linked_bin is not None:
-            linked_abspath = bin_abspath(bin_name, PATH=str(self.bin_dir))
-            if linked_abspath:
-                return linked_abspath
-
         search_paths = self._brew_search_paths(bin_name, no_cache=no_cache)
         abspath = bin_abspath(bin_name, PATH=search_paths)
         if abspath:

abxpkg/binprovider_goget.py

@@ -294,6 +294,15 @@ def default_abspath_handler(
         assert bin_dir is not None
         link_path = bin_dir / str(bin_name)
         link_path.parent.mkdir(parents=True, exist_ok=True)
+        # Idempotent refresh: skip when shim already points at target.
+        # Rewriting on every load() bumps mtime and churns the inode,
+        # which invalidates fingerprint caches unnecessarily.
+        if link_path.is_symlink():
+            try:
+                if link_path.readlink() == Path(direct_abspath):
+                    return TypeAdapter(HostBinPath).validate_python(link_path)
+            except OSError:
+                pass
         if link_path.exists() or link_path.is_symlink():
             link_path.unlink(missing_ok=True)
         link_path.symlink_to(direct_abspath)

abxpkg/binprovider_npm.py

@@ -377,6 +377,15 @@ def _refresh_bin_link(
         link_path = self._linked_bin_path(bin_name)
         assert link_path is not None, "_refresh_bin_link requires bin_dir to be set"
         link_path.parent.mkdir(parents=True, exist_ok=True)
+        # Idempotent refresh: skip when shim already points at target.
+        # Rewriting on every load() bumps mtime and churns the inode,
+        # which invalidates fingerprint caches unnecessarily.
+        if link_path.is_symlink():
+            try:
+                if link_path.readlink() == Path(target):
+                    return TypeAdapter(HostBinPath).validate_python(link_path)
+            except OSError:
+                pass
         if link_path.exists() or link_path.is_symlink():
             link_path.unlink(missing_ok=True)
         link_path.symlink_to(target)