cli: anchor aube install at workspace root from member subdir

crates/aube/src/commands/dlx.rs

@@ -150,9 +150,15 @@ pub async fn run(args: DlxArgs) -> miette::Result<()> {
     // restores the original on drop — so the exec path below, any error
     // diagnostic rendering, and even a panic unwinding past this frame all
     // observe the user's real cwd instead of a dir that's about to vanish.
+    // Pin `project_dir` to the scratch path explicitly so install::run
+    // does not walk upward looking for a workspace root — the scratch
+    // dir lives under TMPDIR, which inside a parent workspace would
+    // otherwise resolve to that workspace and install the wrong tree.
     let prev_cwd = {
         let _cwd_guard = CwdGuard::switch_to(&project_dir)?;
-        let install_result = super::install::run(dlx_install_options()).await;
+        let mut opts = dlx_install_options();
+        opts.project_dir = Some(project_dir.clone());
+        let install_result = super::install::run(opts).await;
         let prev = _cwd_guard.original.clone();
         install_result.wrap_err("dlx install failed")?;
         prev

crates/aube/src/commands/install/mod.rs

@@ -1709,11 +1709,22 @@ pub async fn run(opts: InstallOptions) -> miette::Result<()> {
         project_dir.clone()
     } else {
         let initial_cwd = crate::dirs::cwd()?;
-        // Walk upward to the nearest `package.json` so `aube install` run
-        // from a subdirectory (e.g. `repo/docs`) installs against the
-        // project root instead of erroring with "package.json not found".
-        // Matches pnpm's behavior.
-        match crate::dirs::find_project_root(&initial_cwd) {
+        // Prefer the workspace root so `aube install` from inside a
+        // workspace member installs against the workspace, not the
+        // member as a standalone project — otherwise the member gets
+        // its own `aube-lock.yaml`, its own `.aube/` virtual store,
+        // and re-downloads anything not already in the global cache.
+        // The workspace root must have its own `package.json` —
+        // `find_workspace_root` returns yaml-only roots too, but
+        // `install::run` reads the root manifest unconditionally and
+        // would error on a yaml-only workspace. Fall through to the
+        // nearest `package.json` (the member itself, or the cwd for
+        // non-workspace subdirectory installs like `repo/docs`).
+        // Mirrors the anchor logic in `ensure_installed`.
+        match crate::dirs::find_workspace_root(&initial_cwd)
+            .filter(|root| root.join("package.json").is_file())
+            .or_else(|| crate::dirs::find_project_root(&initial_cwd))
+        {
             Some(root) => root,
             None => {
                 return Err(miette!(

mise-tasks/ci/bats/nonserial

@@ -25,13 +25,23 @@ printf "Running non-serial BATS shard %s/%s with %s files\n" "$((shard_index + 1
 printf "  %s\n" "${test_files[@]}"
 mkdir -p "$report_dir"
 status=0
-BATS_REPORT_FILENAME="$report_name" ./test/bats/bin/bats \
-	--report-formatter junit \
-	--output "$report_dir" \
-	--jobs "$jobs" \
-	--parallel-binary-name parallel \
-	--filter-tags "!serial" \
-	"${test_files[@]}" || status=$?
+# DEBUG (TEMP): run files one at a time with a 90s per-file timeout so
+# any hanging test surfaces as a SIGTERM with the file name visible in
+# the log. Revert this change once the hang is identified.
+for tf in "${test_files[@]}"; do
+	echo "::group::bats $tf"
+	if ! timeout --signal=KILL --verbose 90 ./test/bats/bin/bats \
+		--filter-tags "!serial" \
+		--timing \
+		"$tf"; then
+		rc=$?
+		echo "::endgroup::"
+		echo "BATS file failed or timed out: $tf (exit $rc)"
+		status=$rc
+		break
+	fi
+	echo "::endgroup::"
+done
 
 upload_status=0
 mise run ci:test-engine:upload -- "$report_path" || upload_status=$?

test/workspace_member_install_walks_up.bats

@@ -0,0 +1,97 @@
+#!/usr/bin/env bats
+#
+# Regression: `aube install` run from inside a workspace member must
+# resolve up to the workspace root (pnpm parity) instead of treating the
+# member as a standalone project. The user-visible bug is that a member
+# install runs a fresh resolve/fetch/link cycle and writes a private
+# lockfile, virtual store, and install state next to the member's own
+# package.json — duplicating work the prior workspace-root install
+# already did, and re-downloading anything not already in the global
+# cache.
+#
+# Root cause: install/mod.rs's `find_project_root` walks up to the
+# *nearest* package.json, which inside a workspace is the member's own.
+# A workspace-aware walk should resolve to the workspace root when the
+# member's ancestor is a workspace declaration.
+#
+# bats file_tags=serial
+
+# shellcheck disable=SC2034
+BATS_NO_PARALLELIZE_WITHIN_FILE=1
+
+setup() {
+	load 'test_helper/common_setup'
+	_common_setup
+}
+
+teardown() {
+	_common_teardown
+}
+
+_setup_minimal_workspace() {
+	cat >package.json <<-'JSON'
+		{
+		  "name": "ws-root",
+		  "version": "0.0.0",
+		  "private": true
+		}
+	JSON
+	cat >pnpm-workspace.yaml <<-'YAML'
+		packages:
+		  - packages/*
+	YAML
+	mkdir -p packages/a
+	cat >packages/a/package.json <<-'JSON'
+		{
+		  "name": "@ws/a",
+		  "version": "1.0.0",
+		  "dependencies": {
+		    "is-odd": "3.0.1"
+		  }
+		}
+	JSON
+}
+
+@test "aube install from a workspace member resolves up to the workspace root" {
+	_setup_minimal_workspace
+
+	# Prime the workspace at the root. After this, the workspace root
+	# owns the lockfile, the virtual store, and the install state.
+	run aube install
+	assert_success
+	assert_file_exists aube-lock.yaml
+	assert_dir_exists node_modules/.aube
+	assert_file_exists node_modules/.aube-state
+	assert_link_exists packages/a/node_modules/is-odd
+
+	# Snapshot the root install state. A correct member install must
+	# resolve back up to this same root and short-circuit on the warm
+	# fast path — the root state must be byte-identical afterward.
+	local root_state_before
+	root_state_before="$(cat node_modules/.aube-state)"
+
+	cd packages/a
+	run aube install
+	assert_success
+	cd ../..
+
+	# No standalone artifacts inside the workspace member. Each
+	# assertion below proves a distinct symptom of the bug:
+	#   - own lockfile  -> member ran its own resolve
+	#   - own .aube     -> member ran its own linker pass
+	#   - own .aube-state -> member ran the full install pipeline
+	# Any single failure means the member install ignored the
+	# workspace root.
+	run test -e packages/a/aube-lock.yaml
+	assert_failure
+	run test -d packages/a/node_modules/.aube
+	assert_failure
+	run test -e packages/a/node_modules/.aube-state
+	assert_failure
+
+	# The root install state is byte-identical — the only correct
+	# warm path is "do nothing", and "do nothing" can't write any
+	# state file. `assert_equal` (not `[`) so a regression prints
+	# both snapshots and we can diff what changed.
+	assert_equal "$(cat node_modules/.aube-state)" "$root_state_before"
+}