fix: restore step died with exit 141 (SIGPIPE) on every cache hit — act runs step scripts with 'bash -e -o pipefail', so the 'FIRST_ENTRY=$(tar -tf ... | head -1)' detection line killed the step: head -1 closes the pipe after the first line, tar gets SIGPIPE (141), and pipefail propagates it as the script exit status (observed: 'Cache hit' then 'Process completed with exit code 141', step duration 0s, on deploy-leptos after this detection line was added on 2026-08-14)

The previous restores predate the detection line, which is why this only started failing on 2026-08-17. Suppress the pipeline status with '|| true' — FIRST_ENTRY still captures the first archive entry, and a genuinely unreadable archive degrades to the legacy fallback / cache-hit=false path as before.

README.md

@@ -3,13 +3,6 @@ Cache Action
 
 A Gitea Action that caches build files as `tar.xz` archives on the runner filesystem. Simple, fast, no external dependencies.
 
-## Branches
-
-| Branch | Purpose |
-|--------|---------|
-| `restore` | Restores cache archive if it exists |
-| `save` | Saves cache archive (use after build) |
-
 ## Inputs
 
 | Input | Description | Required | Default |
@@ -22,20 +15,43 @@ A Gitea Action that caches build files as `tar.xz` archives on the runner filesy
 
 | Output | Description |
 |--------|-------------|
-| `cache-hit` | `true` if exact key match found, `false` otherwise (restore only) |
+| `cache-hit` | `true` if exact key match found, `false` otherwise |
 
 ## How It Works
 
 1. Hashes `key-file` with SHA-256, combines with `key-prefix` as `<prefix>-<hash>`
 2. Stores archives at `~/.cache/.cache-store/<key>.tar.xz`
-3. `restore` branch extracts the archive, `save` branch creates one
+3. Restores by extracting the archive to `/`, saves by creating one
+4. Touches the archive on every hit, so the retention policy can expire by
+   last use
+5. Archives store the cached directory **relative** (`<basename>/...`), not its
+   absolute path — a later run restores into its own `path` location, so the
+   cache works across runs even though each run gets a fresh workspace
+
+## Cache Retention
+
+Archives are content-addressed, so old keys are never needed again — the store
+is pruned automatically on every restore/save (best-effort, safe under
+concurrent runners):
+
+- The newest `CACHE_KEEP_N` (default `3`) archives per key-prefix family are
+  always kept, so reverting e.g. `Cargo.lock` still hits an older key
+- Anything older than `CACHE_TTL_DAYS` (default `7`) outside the keep set is
+  deleted
+- A hard size cap `CACHE_CAP_GB` (default `2`) deletes oldest-first when
+  exceeded
+- Files younger than 1 hour are never touched (in-flight save protection)
+- Stale run workspaces under `~/.cache/act/*/hostexecutor` older than
+  `CACHE_WORKSPACE_TTL_DAYS` (default `2`) are removed
+
+Override the defaults per-job via `env:` (e.g. `env: { CACHE_KEEP_N: 5 }`).
 
 ## Usage
 
 This action must be used **twice** in your workflow:
 
-1. **Before build** — use `@restore` to restore cache if it exists
-2. **After build** — use `@save` to save the cache archive
+1. **Before build** — restores cache if it exists
+2. **After build** — saves the cache archive (only on cache miss)
 
 ## Rust Example
 
@@ -54,7 +70,7 @@ jobs:
       - uses: actions/checkout@v4
 
       # 1. Restore cache before build
-      - uses: Actions/Cache@restore
+      - uses: Actions/Cache@main
         id: cache
         with:
           key-prefix: 'cargo-registry'
@@ -65,7 +81,7 @@ jobs:
         run: cargo build --release
 
       # 2. Save cache after build
-      - uses: Actions/Cache@save
+      - uses: Actions/Cache@main
         with:
           key-prefix: 'cargo-registry'
           key-file: 'Cargo.lock'
@@ -89,7 +105,7 @@ jobs:
       - uses: actions/checkout@v4
 
       # 1. Restore cache
-      - uses: Actions/Cache@restore
+      - uses: Actions/Cache@main
         id: cache
         with:
           key-prefix: 'rust'
@@ -104,7 +120,7 @@ jobs:
         run: cargo build --release
 
       # 2. Save cache
-      - uses: Actions/Cache@save
+      - uses: Actions/Cache@main
         with:
           key-prefix: 'rust'
           key-file: 'Cargo.lock'

action.yaml

@@ -0,0 +1,74 @@
+name: 'Cache Restore'
+description: 'Restore cache from tar.xz archive on the runner filesystem'
+inputs:
+  path:
+    description: 'Directory to cache'
+    required: false
+    default: ${{ github.workspace }}
+  key-prefix:
+    description: 'Prefix for the cache key'
+    required: true
+  key-file:
+    description: 'File to hash for the cache key'
+    required: true
+outputs:
+  cache-hit:
+    description: 'true if exact key match found, false otherwise'
+    value: ${{ steps.restore.outputs.cache-hit }}
+runs:
+  using: "composite"
+  steps:
+    - name: Restore cache
+      id: restore
+      shell: bash
+      run: |
+        CACHE_DIR="$HOME/.cache/.cache-store"
+        mkdir -p "$CACHE_DIR"
+
+        KEY_HASH=$(sha256sum "${{ inputs.key-file }}" | awk '{print $1}')
+        CACHE_KEY="${{ inputs.key-prefix }}-${KEY_HASH}"
+        ARCHIVE="${CACHE_DIR}/${CACHE_KEY}.tar.xz"
+
+        EXPANDED_PATH=$(eval echo "${{ inputs.path }}")
+
+        echo "Key file: ${{ inputs.key-file }}"
+        echo "Path input: ${{ inputs.path }}"
+        echo "Expanded path: ${EXPANDED_PATH}"
+        echo "Cache key: ${CACHE_KEY}"
+        echo "Archive path: ${ARCHIVE}"
+
+        if [ -f "$ARCHIVE" ]; then
+          echo "Cache hit: ${CACHE_KEY}"
+          mkdir -p "$EXPANDED_PATH"
+          # New-format archives hold entries relative to the cached directory
+          # (<basename>/...), so extract into the parent of the requested
+          # path. Legacy archives embed the absolute path of the run that
+          # saved them; fall back to extracting them to "/" (status quo,
+          # they are TTL-pruned within the retention window anyway).
+          # NOTE: `|| true` is mandatory — act runs step scripts with
+          # `bash -e -o pipefail`, and `head -1` closes the pipe after the
+          # first line, so tar dies with SIGPIPE (141) and pipefail turns
+          # that into a step failure (restore only ever "worked" before
+          # this detection line existed).
+          FIRST_ENTRY=$(tar -tf "$ARCHIVE" 2>/dev/null | head -1 || true)
+          if [ "${FIRST_ENTRY%%/*}" = "$(basename "$EXPANDED_PATH")" ]; then
+            echo "Restoring to: $(dirname "$EXPANDED_PATH")"
+            tar -xf "$ARCHIVE" -C "$(dirname "$EXPANDED_PATH")" 2>/dev/null
+          else
+            tar -xf "$ARCHIVE" -C "/" 2>/dev/null
+          fi
+          if [ -d "$EXPANDED_PATH" ] && [ -n "$(ls -A "$EXPANDED_PATH" 2>/dev/null)" ]; then
+            # record last use so the retention policy can expire by access time
+            touch "$ARCHIVE"
+            echo "cache-hit=true" >> $GITHUB_OUTPUT
+          else
+            echo "Cache restore failed"
+            echo "cache-hit=false" >> $GITHUB_OUTPUT
+          fi
+        else
+          echo "Cache miss: ${CACHE_KEY}"
+          echo "cache-hit=false" >> $GITHUB_OUTPUT
+        fi
+
+        # Trim the shared cache store and stale workspaces (best-effort)
+        bash "${{ github.action_path }}/prune.sh" || true

prune.sh

@@ -0,0 +1,90 @@
+#!/usr/bin/env bash
+# Prune the shared runner build-cache store (~/.cache/.cache-store) and stale
+# run workspaces (~/.cache/act). Best-effort: safe to run from any job on any
+# runner, since all runners on a host share the same store.
+#
+# Policy (per family, where a family is the key-prefix, e.g. `leptos-target`):
+#   - the newest CACHE_KEEP_N archives per family are always kept (rollback
+#     safety: reverting Cargo.lock must still hit an older key)
+#   - any archive older than CACHE_TTL_DAYS outside the keep set is deleted
+#   - a hard CACHE_CAP_GB size cap deletes oldest-first when exceeded
+#   - files younger than 1 hour are never touched (in-flight save protection;
+#     saves are atomic, see the save action: write .tmp then mv)
+#
+# Overrides via env: CACHE_KEEP_N (default 3), CACHE_TTL_DAYS (default 7),
+#                    CACHE_CAP_GB (default 2), CACHE_WORKSPACE_TTL_DAYS
+#                    (default 2)
+
+set -u
+
+CACHE_DIR="$HOME/.cache/.cache-store"
+KEEP_N="${CACHE_KEEP_N:-3}"
+TTL_SECS=$(( ${CACHE_TTL_DAYS:-7} * 86400 ))
+CAP_BYTES=$(( ${CACHE_CAP_GB:-2} * 1024 * 1024 * 1024 ))
+SAFETY_SECS=3600
+WORKSPACE_TTL_DAYS="${CACHE_WORKSPACE_TTL_DAYS:-2}"
+
+[ -d "$CACHE_DIR" ] || exit 0
+NOW=$(date +%s)
+
+# stale tmp files left by interrupted saves
+find "$CACHE_DIR" -maxdepth 1 -name '*.tar.xz.tmp' -mmin +60 -delete 2>/dev/null
+
+# newest KEEP_N archives per family are protected from deletion
+declare -A protected=()
+declare -A fam_count=()
+while IFS=$'\t' read -r mt name; do
+  fam=$(sed -E 's/-[0-9a-f]{64}\.tar\.xz$//' <<< "$name")
+  if [ -n "$fam" ] && [ "${fam_count[$fam]:-0}" -lt "$KEEP_N" ]; then
+    protected[$name]=1
+    fam_count[$fam]=$(( ${fam_count[$fam]:-0} + 1 ))
+  fi
+done < <(find "$CACHE_DIR" -maxdepth 1 -name '*.tar.xz' -printf '%T@\t%f\n' | sort -rn)
+
+# deletion candidates: not protected, older than 1h, oldest first
+declare -a candidates=()
+while IFS=$'\t' read -r mt name; do
+  [ "${protected[$name]:-0}" = "1" ] && continue
+  mtime=$(printf '%.0f' "$mt")
+  [ $(( NOW - mtime )) -lt "$SAFETY_SECS" ] && continue
+  candidates+=("$mtime|$name")
+done < <(find "$CACHE_DIR" -maxdepth 1 -name '*.tar.xz' -printf '%T@\t%f\n' | sort -n)
+
+# TTL pass: drop archives past their retention window
+deleted=0
+for c in "${candidates[@]:-}"; do
+  [ -n "$c" ] || continue
+  mtime="${c%%|*}"; name="${c#*|}"
+  if [ $(( NOW - mtime )) -gt "$TTL_SECS" ]; then
+    rm -f "$CACHE_DIR/$name" && deleted=$((deleted + 1))
+  fi
+done
+[ "$deleted" -gt 0 ] && echo "prune: deleted $deleted archive(s) older than ${CACHE_TTL_DAYS:-7} day(s)"
+
+# size-cap pass: keep the store under CAP_BYTES, oldest first. The cap is a
+# hard backstop and overrides keep-N protection (only the 1h safety floor is
+# absolute), so the size bound always holds even for small families.
+total=$(du -sb "$CACHE_DIR" 2>/dev/null | awk '{print $1}')
+if [ "${total:-0}" -gt "$CAP_BYTES" ]; then
+  while IFS=$'\t' read -r mt name; do
+    [ "${total:-0}" -le "$CAP_BYTES" ] && break
+    mtime=$(printf '%.0f' "$mt")
+    [ $(( NOW - mtime )) -lt "$SAFETY_SECS" ] && continue
+    if [ -f "$CACHE_DIR/$name" ]; then
+      size=$(stat -c %s "$CACHE_DIR/$name")
+      rm -f "$CACHE_DIR/$name" && total=$(( total - size ))
+    fi
+  done < <(find "$CACHE_DIR" -maxdepth 1 -name '*.tar.xz' -printf '%T@\t%f\n' | sort -n)
+  echo "prune: store size now $(du -sh "$CACHE_DIR" 2>/dev/null | cut -f1) (cap ${CACHE_CAP_GB:-2}G)"
+fi
+
+# stale run workspaces: every run gets a fresh ~/.cache/act/<id>/hostexecutor
+# dir that is never reused, so old ones are dead weight (they can hold entire
+# target/ dirs); nothing younger than the floor is touched
+ws_deleted=0
+while IFS= read -r ws; do
+  [ -n "$ws" ] || continue
+  rm -rf "$ws" && ws_deleted=$((ws_deleted + 1))
+done < <(find "$HOME/.cache/act" -maxdepth 2 -type d -name hostexecutor -mtime +"$WORKSPACE_TTL_DAYS" 2>/dev/null)
+find "$HOME/.cache/act" -maxdepth 1 -type d -empty -delete 2>/dev/null || true
+[ "$ws_deleted" -gt 0 ] && echo "prune: removed $ws_deleted stale workspace(s) older than ${WORKSPACE_TTL_DAYS} day(s)"