feat(cache): add shared S3-compatible storage (#1244)

Store the cache used by `actions/cache` in an S3-compatible bucket such as AWS S3 or RustFS, with no `cache-server`. Runners with the same settings use the same cache.

A cache saved on one runner can be restored on the others within 30 seconds. The runner talks to the bucket over the S3 REST API with SigV4, without an SDK. Lifecycle rules on the bucket expire old caches.

*Description rewritten by Claude.*

---------

Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1244
Reviewed-by: silverwind <[email protected]>
Co-authored-by: bircni <[email protected]>

.gitea/workflows/release-nightly.yml

@@ -74,10 +74,10 @@ jobs:
           fetch-depth: 0 # all history for all branches and tags
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@96fe6ef7f33517b61c61be40b68a1882f3264fb8 # v4
+        uses: docker/setup-qemu-action@99012661954931238ded8c8b007157a8430204e1 # v4
 
       - name: Set up Docker BuildX
-        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4
+        uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4
 
       - name: Login to DockerHub
         uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4
@@ -94,7 +94,7 @@ jobs:
           echo REPO_VERSION=$(git describe --tags --always | sed 's/-/+/' | sed 's/^v//') >> $GITHUB_OUTPUT
 
       - name: Build and push
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7
+        uses: docker/build-push-action@c3c9e263c25d99ce0380d002d59b67737d91b0dc # v7
         with:
           context: .
           file: ./Dockerfile

.gitea/workflows/release-tag.yml

@@ -77,10 +77,10 @@ jobs:
           fetch-depth: 0 # all history for all branches and tags
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@96fe6ef7f33517b61c61be40b68a1882f3264fb8 # v4
+        uses: docker/setup-qemu-action@99012661954931238ded8c8b007157a8430204e1 # v4
 
       - name: Set up Docker BuildX
-        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4
+        uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4
 
       - name: Login to DockerHub
         uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4
@@ -103,7 +103,7 @@ jobs:
             suffix=${{ matrix.variant.tag_suffix }},onlatest=true
 
       - name: Build and push
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7
+        uses: docker/build-push-action@c3c9e263c25d99ce0380d002d59b67737d91b0dc # v7
         with:
           context: .
           file: ./Dockerfile

.gitea/workflows/test.yml

@@ -30,8 +30,8 @@ jobs:
       # daemon retains them between runs, so this is usually a fast manifest re-check.
       - name: pre-pull test images
         env:
-          TEST_JOB_IMAGE: node:24-bookworm-slim@sha256:ba849c60be29959425b8734d57b8b4b7d56f98edd9504c9af091d5281095a71e # renovate: datasource=docker
-          TEST_SERVICE_IMAGE: nginx:alpine@sha256:db35bfc6b2951e7f8a72db5db120288c127ffaeeb4a6d4b95a26fead017d5913 # renovate: datasource=docker
+          TEST_JOB_IMAGE: node:24-bookworm-slim@sha256:0e0ff40c39bc087845bfb27465a0df4ea419520094bc35842ff83dd8cbe6f9b6 # renovate: datasource=docker
+          TEST_SERVICE_IMAGE: nginx:alpine@sha256:62ff2089abf5a9ed33bd232895bef5e22f7bb4b200675cec49a5ebc48e3d4ac8 # renovate: datasource=docker
         run: |
           for image in "$TEST_JOB_IMAGE" "$TEST_SERVICE_IMAGE"; do
             for attempt in 1 2 3; do
@@ -53,6 +53,8 @@ jobs:
       # after `make test` so the images it needs are already present on the host daemon.
       - name: test against dind image
         run: make test-dind
+      - name: test against podman
+        run: make test-dind TARGET=podman
       - name: coverage report
         run: |
           make coverage-report

CLAUDE.md

@@ -1 +0,0 @@
-@AGENTS.md

DEVELOPMENT.md

@@ -31,6 +31,11 @@ depending on the prefix:
 - Command *properties* also escape `%3A` and `%2C`, which the UI never decodes, so the reporter
   decodes exactly those two when folding a location into an annotation.
 
+## Built-in actions
+
+Each lives in `internal/pkg/action/<name>` and is registered in `builtinActions` in
+`act/runner/step_action_builtin.go`.
+
 ## End-to-end compatibility tests
 
 `make test-e2e` runs the runner against `E2E_GITEA_IMAGE`. It defaults to the nightly image.

Dockerfile

@@ -32,6 +32,8 @@ COPY scripts/s6 /etc/s6
 
 VOLUME /data
 
+ENV TINI_SUBREAPER=1
+
 ENTRYPOINT ["s6-svscan","/etc/s6"]
 
 ### DIND-ROOTLESS VARIANT
@@ -56,6 +58,7 @@ VOLUME /data
 RUN mkdir -p /data && chown -R rootless:rootless /etc/s6 /data
 
 ENV DOCKER_HOST=unix:///run/user/1000/docker.sock
+ENV TINI_SUBREAPER=1
 
 USER rootless
 ENTRYPOINT ["s6-svscan","/etc/s6"]

Makefile

@@ -8,6 +8,11 @@ XGO_PACKAGE ?= src.techknowlogick.com/[email protected] # renovate: datasource=go
 XGO_VERSION := go-1.27.x
 GXZ_PACKAGE ?= github.com/ulikunitz/xz/cmd/[email protected] # renovate: datasource=go
 
+PODMAN_TEST_IMAGE ?= quay.io/containers/podman:v5.8.4-immutable@sha256:90405e3c6a86f33c5a6f8dbdfa8e02a98143a8ba4337b84bdaec5336d2ef008a # renovate: datasource=docker
+E2E_JOB_IMAGE ?= node:24-bookworm@sha256:64af3819f9275802414d7cdc38c27e9d82bd564dec4d4da87d008255d36c63b4 # renovate: datasource=docker
+SERVICE_IMAGE ?= nginx:1.31.4-alpine@sha256:db35bfc6b2951e7f8a72db5db120288c127ffaeeb4a6d4b95a26fead017d5913 # renovate: datasource=docker
+E2E_GITEA_IMAGE ?= gitea/gitea:main-nightly
+
 LINUX_ARCHS ?= linux/amd64,linux/arm64
 DARWIN_ARCHS ?= darwin-12/amd64,darwin-12/arm64
 WINDOWS_ARCHS ?= windows/amd64
@@ -18,8 +23,8 @@ DOCKER_TAG ?= nightly
 DOCKER_REF := $(DOCKER_IMAGE):$(DOCKER_TAG)
 DOCKER_ROOTLESS_REF := $(DOCKER_IMAGE):$(DOCKER_TAG)-dind-rootless
 
-GOLANGCI_LINT_PACKAGE ?= github.com/golangci/golangci-lint/v2/cmd/[email protected].1 # renovate: datasource=go
-GOVULNCHECK_PACKAGE ?= golang.org/x/vuln/cmd/govulncheck@v1.7.0 # renovate: datasource=go
+GOLANGCI_LINT_PACKAGE ?= github.com/golangci/golangci-lint/v2/cmd/[email protected].2 # renovate: datasource=go
+GOVULNCHECK_PACKAGE ?= golang.org/x/vuln/cmd/govulncheck@v1.8.0 # renovate: datasource=go
 
 GOTEST_FLAGS ?= -race -timeout 20m -parallel 8
 
@@ -96,13 +101,8 @@ go-check:
 	fi
 
 .PHONY: fmt-check
-fmt-check: fmt
-	@diff=$$(git diff --color=always -- '*.go'); \
-	if [ -n "$$diff" ]; then \
-		echo "Please run 'make fmt' and commit the result:"; \
-		printf "%s" "$${diff}"; \
-		exit 1; \
-	fi
+fmt-check: ## check the Go code formatting
+	$(GO) run $(GOLANGCI_LINT_PACKAGE) fmt --diff-colored
 
 .PHONY: deps-tools
 deps-tools: ## install tool dependencies
@@ -149,13 +149,8 @@ tidy: ## run go mod tidy
 	fi
 
 .PHONY: tidy-check
-tidy-check: tidy
-	@diff=$$(git diff --color=always -- go.mod go.sum); \
-	if [ -n "$$diff" ]; then \
-		echo "Please run 'make tidy' and commit the result:"; \
-		printf "%s" "$${diff}"; \
-		exit 1; \
-	fi
+tidy-check: ## check that go.mod and go.sum are tidy
+	$(GO) mod tidy -diff
 
 .PHONY: test
 test: ## test everything (integration tests self-skip without docker/network)
@@ -168,12 +163,9 @@ coverage-report: ## turn coverage.txt from `make test` into .tmp/coverage.md
 	@echo "Wrote .tmp/coverage.md"
 
 .PHONY: test-dind
-test-dind: ## run the daemon-facing tests against the built dind image (TARGET=dind|dind-rootless)
-	@./scripts/test-dind.sh $(TARGET)
+test-dind: ## run the daemon-facing tests against the built dind image or podman (TARGET=dind|dind-rootless|podman)
+	@PODMAN_TEST_IMAGE=$(PODMAN_TEST_IMAGE) ./scripts/test-dind.sh $(TARGET)
 
-E2E_JOB_IMAGE ?= node:24-bookworm@sha256:be23f54a88d34e8824c741b19b91064094f92c1c97b194144bfc8b50d67258e2 # renovate: datasource=docker
-SERVICE_IMAGE ?= nginx:1.31.4-alpine@sha256:db35bfc6b2951e7f8a72db5db120288c127ffaeeb4a6d4b95a26fead017d5913 # renovate: datasource=docker
-E2E_GITEA_IMAGE ?= gitea/gitea:main-nightly
 E2E_CONCURRENCY ?= 8
 
 .PHONY: test-e2e

README.md

@@ -268,14 +268,14 @@ Unlike GitHub, a job whose steps run on the host (a `host` label without `contai
 
 #### Docker from a job (`GITEA_DOCKER_WORKSPACE`)
 
-A container a job starts through the Docker socket cannot bind-mount the workspace by the job's own path, the daemon does not have it. `GITEA_DOCKER_WORKSPACE` holds the path the daemon sees. Use it as the prefix of workspace binds, with `.` as the fallback for local use, here in a `docker-compose.yaml`:
+Containers a job starts through the Docker socket can bind-mount the workspace by the job's own path, as on a host, for example `docker run -v "$PWD:/src"`. Without the Docker proxy, for example with a remote `DOCKER_HOST`, use `GITEA_DOCKER_WORKSPACE`, the path the daemon sees, as the prefix of workspace binds, with `.` as the fallback for local use:
 
 ```yaml
 volumes:
   - ${GITEA_DOCKER_WORKSPACE:-.}/data:/app/data
 ```
 
-Containers, networks and volumes a job creates through the socket are removed when the job ends. Both apply to jobs in containers, on the host `.` already works.
+Linux container jobs use a Docker proxy when socket sharing and permissions allow it. Other setups use the daemon socket directly. The proxy removes containers, networks and volumes after post steps and the completed hook. Named volumes created through it are job-scoped. To retain resources, mount the daemon socket explicitly in `container.options`. Host jobs use their existing Docker access, where `.` works.
 
 #### Proxy
 
@@ -291,7 +291,7 @@ The runner uses them for its own requests and gives them to every job, in lower
 
 These hosts are added to `no_proxy` for jobs, so they are always reached directly:
 
-- the cache server
+- the built-in cache server, whose address is assigned at startup
 - `localhost`, `127.0.0.1` and `::1`
 - the job's service containers
 - the Docker daemon, when it is reached over `tcp://`
@@ -306,63 +306,68 @@ Dockerfile actions are built with these variables as build arguments, so their `
 
 A password in a proxy URL is hidden in job logs. Any step can still read it, because the step is given the proxy URL in its environment.
 
+#### Built-in actions (`uses: builtin:<name>`)
+
+Use `uses: builtin:checkout` to check out a repository without downloading `actions/checkout` or needing Node in the job image, see **[docs/builtin-actions.md](docs/builtin-actions.md)**.
+
 #### Caching (`actions/cache`)
 
-Each runner starts its own cache server automatically. Cache entries are local to that runner — runners do not share a cache by default.
+Each runner starts its own cache server, so runners do not share cached entries. A runner in a container on the jobs' Docker daemon attaches itself to each job network, so jobs reach its cache without extra setup.
 
-**Eviction**
+**Sharing a cache between runners**
 
-An entry nothing has read or written for `retention` is removed, and a repository past `repo_size_limit` loses its least recently accessed entries until it fits; `size_limit` caps the whole cache the same way. Age alone never retires an entry still in use, and whatever these allow, the cache keeps free space above `health_check.min_free_disk_space_mb` when health checks are enabled.
+Run one `gitea-runner cache-server` and point every runner at it, using the same secret everywhere, for example from `openssl rand -hex 32`:
 
-These apply where the cache server runs, so on a shared server they belong in *its* config, not the runners'. See `retention`, `repo_size_limit`, `size_limit` and `sweep_interval` in [config.example.yaml](internal/pkg/config/config.example.yaml) for units and defaults.
+```yaml
+# cache-server.yaml
+cache:
+  dir: /data/actcache
+  port: 8088
+  external_secret: "<secret>" # or external_secret_file: /path/to/secret
+```
 
-**Cache service v2**
+```bash
+gitea-runner -c cache-server.yaml cache-server
+```
 
-`actions/cache@v4.2` and later can use the *cache service v2* API. The runner serves it from the same store as v1, on by default, and it works with `external_server`. Turn it off with:
+```yaml
+# each runner's config
+cache:
+  external_server: "http://<cache-server-host>:8088/"
+  external_secret: "<secret>"
+```
+
+`--dir`, `--host` and `--port` override the matching `cache` keys. Eviction settings take effect on the cache server, not on the runners.
+
+Runners can also share one `cache.dir` on a file system with working file locks, at the cost of slower cache requests.
+
+**Storing the cache in an S3-compatible bucket**
+
+The cache can also be stored in an S3-compatible bucket such as AWS S3, MinIO or RustFS, with no cache server. Runners with the same settings use the same cache:
 
 ```yaml
 cache:
-  v2: false
+  s3:
+    endpoint: "http://minio:9000"
+    region: "us-east-1"
+    bucket: "runner-cache"
+    prefix: "gitea-runner-cache"
+    path_style: true
+    access_key_file: "/run/secrets/cache-s3-access-key"
+    secret_key_file: "/run/secrets/cache-s3-secret-key"
 ```
 
-Those actions refuse any host they do not take for GitHub. Rather than misreport the server URL, the runner edits that check out of the action's own bundle on its way into the job, undone whenever the action is downloaded again. A bundle it does not recognise is left alone and keeps to v1. The same edit lets the stock `actions/upload-artifact` and `actions/download-artifact` work from `v4.4.0` on, without the `gitea-upload-artifact` fork, so it is made whatever `v2` says: that setting only governs the API the runner advertises. Set `runner.patch_actions: false` to leave every bundle exactly as shipped, an escape hatch for an action the edit breaks. The artifact actions then refuse again and the cache client keeps to v1.
+A cache saved on one runner can be restored on the others within 30 seconds. The eviction settings below do not apply to the bucket, so add lifecycle rules to it that expire old caches and abort incomplete multipart uploads.
 
-**Shared cache across multiple runners**
+**Eviction**
 
-Run one dedicated `gitea-runner cache-server` that all runners point at.
+Entries not read or written within `retention` (default `168h`) are removed. A repository over `repo_size_limit` (default `10GB`) loses its least recently used entries, and `size_limit` (off by default) caps the whole cache the same way. Entries in use are never removed. The cache also keeps 1024 MiB free on its volume, or `health_check.min_free_disk_space_mb` when health checks are enabled. See [config.example.yaml](internal/pkg/config/config.example.yaml) for all options.
 
-1. Create a config file for the cache server host:
+**Cache service v2**
 
-   ```yaml
-   cache:
-     dir: /data/actcache
-     port: 8088
-     external_secret: "replace-with-a-strong-random-secret"
-     # external_secret_file: /path/to/secret # secret can also be passed via a file
-   ```
+`actions/cache` v3.4.0, v4.2.0 and later use the cache service v2 API, which the runner serves by default. These actions fall back to v1 on hosts they do not recognize as GitHub, so the runner removes that check from them while a job runs. This also lets the stock `actions/upload-artifact` v4.4.0 and `actions/download-artifact` v4.1.8 and later work without the `gitea-upload-artifact` fork. With `runner.patch_actions: false`, the cache stays on v1 and the artifact actions fail.
 
-2. Start the server:
-
-   ```bash
-   gitea-runner -c cache-server-config.yaml cache-server
-   ```
-
-3. On every runner:
-
-   ```yaml
-   cache:
-     external_server: "http://<cache-server-host>:8088/"
-     external_secret: "replace-with-a-strong-random-secret"  # must match the server
-     # external_secret_file: /path/to/secret # secret can also be passed via a file
-   ```
-
-Jobs reach the cache server at `external_server`, so when a reverse proxy fronts the server, point `external_server` at the proxy. The cache server itself needs no extra configuration.
-
-Alternatively, mount the same NFS/CIFS share on every runner and point `cache.dir` at it — simpler, but with weaker isolation between repositories.
-
-**S3 / MinIO** — mount object storage as a FUSE filesystem (e.g. [s3fs](https://github.com/s3fs-fuse/s3fs-fuse) or [goofys](https://github.com/kahing/goofys)) and set `cache.dir` to the mount point.
-
-Flags `--dir`, `--host`, and `--port` on `cache-server` override the corresponding `cache.*` YAML keys; all other settings, including `external_secret`, require the config file.
+With v2, artifact calls also go through the cache server, so jobs must be able to reach it. `cache.v2: false` sends them to Gitea directly, unless the Gitea URL has a path or `runner.insecure` is set with HTTPS.
 
 #### Official Docker image
 
@@ -397,6 +402,10 @@ Both hooks are synchronous and block the job while they run. Either one exiting
 
 See **[docs/job-hooks.md](docs/job-hooks.md)** for the execution order, environment, and platform notes.
 
+#### OpenTelemetry (`OTEL_*`)
+
+Set `OTEL_EXPORTER_OTLP_ENDPOINT` in the runner's environment to send a trace of every job to an OpenTelemetry collector. See **[docs/telemetry.md](docs/telemetry.md)**.
+
 #### Local job logs (`log.job.dir`)
 
 Set `log.job.dir` to a path and the runner writes a copy of every task's log there as `<start time>-task-<id>.log`: the rows exactly as Gitea received them, with the same secrets masked and the job's result on the last line. Off by default, and what Gitea shows does not change.

act/artifactcache/handler.go

@@ -10,6 +10,7 @@ import (
 	"crypto/hmac"
 	"crypto/rand"
 	"crypto/sha256"
+	"encoding/binary"
 	"encoding/hex"
 	"encoding/json/v2"
 	"errors"
@@ -17,10 +18,10 @@ import (
 	"io"
 	"net"
 	"net/http"
+	"net/http/httputil"
 	"net/url"
 	"os"
 	"path/filepath"
-	"regexp"
 	"slices"
 	"strconv"
 	"strings"
@@ -58,10 +59,9 @@ type JobCredential struct {
 	Repo string `json:"repo"`
 
 	// Results is the instance whose artifact service this server forwards for the job, and
-	// InsecureTLS how the runner reaches it; see results.go. The tags are the wire format a
-	// remote runner registers with.
-	Results     string `json:"results"`
-	InsecureTLS bool   `json:"insecure_tls"`
+	// InsecureTLS how the runner reaches it; see results.go.
+	Results     string `json:"-"`
+	InsecureTLS bool   `json:"-"`
 
 	// PublicURL is this server as a reverse proxy makes the job reach it, not the listen address.
 	PublicURL string `json:"public_url"`
@@ -79,7 +79,8 @@ type credEntry struct {
 
 type Handler struct {
 	dir      string
-	storage  *Storage
+	storage  cacheStorage
+	s3       *s3Storage
 	router   *httprouter.Router
 	listener net.Listener
 	port     int
@@ -89,6 +90,9 @@ type Handler struct {
 	gcing atomic.Bool
 	gcAt  time.Time
 
+	stopSync context.CancelFunc
+	syncing  sync.WaitGroup
+
 	outboundIP string
 
 	// internalSecret guards /_internal/{register,revoke}. When set, a remote
@@ -118,6 +122,8 @@ type Options struct {
 	Dir        string
 	OutboundIP string
 	Port       uint16
+	S3         *S3Options
+	Upstream   string
 
 	// InternalSecret, when non-empty, enables a control-plane API at
 	// /_internal/{register,revoke} that lets a remote runner pre-register the
@@ -130,7 +136,7 @@ type Options struct {
 	Logger logrus.FieldLogger
 }
 
-// StartHandler opens the on-disk cache store and starts the HTTP server.
+// StartHandler starts an embedded cache or forwards cache requests to Upstream.
 func StartHandler(opts Options) (*Handler, error) {
 	dir, logger := opts.Dir, opts.Logger
 	h := &Handler{
@@ -148,6 +154,25 @@ func StartHandler(opts Options) (*Handler, error) {
 	logger = logger.WithField("module", "artifactcache")
 	h.logger = logger
 
+	if opts.OutboundIP != "" {
+		h.outboundIP = opts.OutboundIP
+	} else if ip := common.GetOutboundIP(); ip == nil {
+		return nil, errors.New("unable to determine outbound IP address")
+	} else {
+		h.outboundIP = ip.String()
+	}
+
+	if opts.Upstream != "" {
+		handler, err := h.cacheForwarder(opts.Upstream)
+		if err != nil {
+			return nil, err
+		}
+		if err := h.serve(opts.Port, handler); err != nil {
+			return nil, err
+		}
+		return h, nil
+	}
+
 	if dir == "" {
 		home, err := os.UserHomeDir()
 		if err != nil {
@@ -166,13 +191,11 @@ func StartHandler(opts Options) (*Handler, error) {
 		return nil, err
 	}
 	h.storage = storage
-
-	if opts.OutboundIP != "" {
-		h.outboundIP = opts.OutboundIP
-	} else if ip := common.GetOutboundIP(); ip == nil {
-		return nil, errors.New("unable to determine outbound IP address")
-	} else {
-		h.outboundIP = ip.String()
+	if opts.S3 != nil {
+		if h.s3, err = newS3Storage(storage, *opts.S3); err != nil {
+			return nil, err
+		}
+		h.storage = h.s3
 	}
 
 	secret, err := loadOrCreateSecret(dir)
@@ -202,39 +225,79 @@ func StartHandler(opts Options) (*Handler, error) {
 
 	h.gcCache()
 
+	if err := h.serve(opts.Port, router); err != nil {
+		return nil, err
+	}
+	if h.s3 != nil {
+		ctx, cancel := context.WithCancel(context.Background())
+		h.stopSync = cancel
+		h.syncing.Go(func() { h.syncMetadata(ctx) })
+	}
+	return h, nil
+}
+
+func (h *Handler) serve(port uint16, handler http.Handler) error {
 	// Listen on all interfaces. Binding to outboundIP only would give no real
 	// security benefit (it is the LAN/internet-facing address either way) and
 	// can break Docker Desktop variants where the host's outbound IP is not
 	// routable from inside the container network. Authentication is enforced
 	// by the bearer middleware and per-repo scoping, not by reachability.
-	listener, err := net.Listen("tcp", fmt.Sprintf(":%d", opts.Port))
+	listener, err := net.Listen("tcp", fmt.Sprintf(":%d", port))
 	if err != nil {
-		return nil, err
+		return err
 	}
 	addr, ok := listener.Addr().(*net.TCPAddr)
 	if !ok {
 		listener.Close()
-		return nil, fmt.Errorf("cache server listens on %T, want a TCP address", listener.Addr())
+		return fmt.Errorf("cache server listens on %T, want a TCP address", listener.Addr())
 	}
 	h.port = addr.Port
 	server := &http.Server{
 		ReadHeaderTimeout: 2 * time.Second,
-		Handler:           router,
+		Handler:           handler,
 	}
 	go func() {
 		if err := server.Serve(listener); err != nil && errors.Is(err, net.ErrClosed) {
-			logger.Errorf("http serve: %v", err)
+			h.logger.Errorf("http serve: %v", err)
 		}
 	}()
 	h.listener = listener
 	h.server = server
 
-	return h, nil
+	return nil
+}
+
+func (h *Handler) cacheForwarder(upstream string) (http.Handler, error) {
+	target, err := url.Parse(strings.TrimRight(upstream, "/"))
+	if err != nil {
+		return nil, fmt.Errorf("parse cache upstream: %w", err)
+	}
+	transport, _ := http.DefaultTransport.(*http.Transport)
+	transport = transport.Clone()
+	transport.MaxIdleConnsPerHost = 64 // parallel block uploads would otherwise redial past the default 2
+	proxy := &httputil.ReverseProxy{
+		Transport: transport,
+		Rewrite: func(r *httputil.ProxyRequest) {
+			r.SetURL(target)
+			r.Out.URL.RawQuery = r.In.URL.RawQuery // Rewrite drops unparsable query params, and signed URLs must pass verbatim.
+		},
+		ErrorHandler: func(w http.ResponseWriter, r *http.Request, err error) {
+			h.logger.Warnf("cache forward %s %s: %v", r.Method, r.URL.Path, err)
+			h.twirpError(w, r, twirpUnavailable, errors.New("cache upstream is unavailable"))
+		},
+	}
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, apiPath+"/") || strings.HasPrefix(r.URL.Path, cacheServiceV2Path+"/") {
+			proxy.ServeHTTP(w, r)
+		} else {
+			h.forwardOrNotFound(w, r)
+		}
+	}), nil
 }
 
 func (h *Handler) ExternalURL() string {
 	// TODO: make the external url configurable if necessary
-	return fmt.Sprintf("http://%s:%d", h.outboundIP, h.port)
+	return "http://" + net.JoinHostPort(h.outboundIP, strconv.Itoa(h.port))
 }
 
 func (h *Handler) baseURL(cred JobCredential) string {
@@ -334,6 +397,10 @@ func (h *Handler) Close() error {
 	if h == nil {
 		return nil
 	}
+	if h.stopSync != nil {
+		h.stopSync()
+		h.syncing.Wait()
+	}
 	var retErr error
 	if h.server != nil {
 		err := h.server.Close()
@@ -367,6 +434,72 @@ func (h *Handler) openDB() (*bolthold.Store, error) {
 	})
 }
 
+// syncMetadata mirrors the bucket into the store until ctx ends, so no request waits on S3 for it.
+func (h *Handler) syncMetadata(ctx context.Context) {
+	ticker := time.NewTicker(metadataSyncInterval)
+	defer ticker.Stop()
+	for {
+		if err := h.importMetadata(ctx); err != nil && ctx.Err() == nil {
+			h.logger.Warnf("sync S3 cache: %v", err)
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+	}
+}
+
+// importMetadata lists after snapshotting the store, so an entry committed meanwhile, published before it is stored, stays.
+func (h *Handler) importMetadata(ctx context.Context) error {
+	db, err := h.openDB()
+	if err != nil {
+		return err
+	}
+	local := h.caches(db, bolthold.Where("Complete").Eq(true).Index("Complete"))
+	db.Close()
+
+	published, err := h.s3.published(ctx)
+	if err != nil {
+		return err
+	}
+	var gone []*Cache
+	for _, cache := range local {
+		if !published[cache.ID] {
+			gone = append(gone, cache)
+		}
+		delete(published, cache.ID)
+	}
+	added := make([]*Cache, 0, len(published))
+	for id := range published {
+		cache, err := h.s3.fetch(ctx, id)
+		if err != nil {
+			return err
+		}
+		added = append(added, cache)
+	}
+
+	db, err = h.openDB()
+	if err != nil {
+		return err
+	}
+	defer db.Close()
+	return db.Bolt().Update(func(tx *bbolt.Tx) error {
+		for _, cache := range added {
+			if err := db.TxUpsert(tx, cache.ID, cache); err != nil {
+				return err
+			}
+		}
+		for _, cache := range gone {
+			_ = h.s3.Storage.Remove(cache.ID) // a blob kept on disk before the runner used S3
+			if err := db.TxDelete(tx, cache.ID, cache); err != nil && !errors.Is(err, bolthold.ErrNotFound) {
+				return err
+			}
+		}
+		return nil
+	})
+}
+
 // GET /_apis/artifactcache/cache
 func (h *Handler) find(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
 	cred := credFromContext(r.Context())
@@ -409,7 +542,7 @@ func (h *Handler) lookupCache(db *bolthold.Store, repo string, keys []string, ve
 		return nil, err
 	}
 	if !ok {
-		_ = db.Delete(cache.ID, cache)
+		h.deleteCache(db, cache)
 		return nil, nil //nolint:nilnil // absence is not an error here
 	}
 	// Handing out a download URL counts as access, or eviction could drop the entry between
@@ -441,7 +574,7 @@ func (h *Handler) reserve(w http.ResponseWriter, r *http.Request, _ httprouter.P
 	now := time.Now().Unix()
 	cache.CreatedAt = now
 	cache.UsedAt = now
-	if err := insertCache(db, cache); err != nil {
+	if err := h.insertCache(db, cache); err != nil {
 		h.responseJSON(w, r, 500, err)
 		return
 	}
@@ -555,6 +688,12 @@ func (h *Handler) commitCache(cache *Cache) error {
 	cache.Size = written
 	cache.Complete = true
 	cache.UsedAt = time.Now().Unix() // a just-written entry counts as accessed, so it cannot be its own eviction victim
+	if h.s3 != nil {
+		if err := h.s3.publish(cache); err != nil {
+			_ = h.storage.Remove(cache.ID)
+			return err
+		}
+	}
 
 	db, err := h.openDB()
 	if err != nil {
@@ -564,6 +703,9 @@ func (h *Handler) commitCache(cache *Cache) error {
 	if err := db.Update(cache.ID, cache); err != nil {
 		return err
 	}
+	if h.s3 != nil {
+		return nil // lifecycle rules on the bucket bound its size
+	}
 	// A commit is the only thing that grows the store, so the only thing that can push the
 	// volume under the floor.
 	h.evictRepo(db, cache.Repo)
@@ -603,12 +745,12 @@ func (h *Handler) bearerAuth(handler httprouter.Handle) httprouter.Handle {
 		h.logger.Debugf("%s %s", r.Method, r.URL.Path)
 		token := bearerToken(r)
 		if token == "" {
-			h.responseJSON(w, r, http.StatusUnauthorized, errors.New("missing bearer token"))
+			h.unauthorized(w, r, errors.New("missing bearer token"))
 			return
 		}
 		cred, ok := h.lookupCredential(token)
 		if !ok {
-			h.responseJSON(w, r, http.StatusUnauthorized, errors.New("unknown bearer token"))
+			h.unauthorized(w, r, errors.New("unknown bearer token"))
 			return
 		}
 		ctx := context.WithValue(r.Context(), credKey{}, cred)
@@ -617,6 +759,14 @@ func (h *Handler) bearerAuth(handler httprouter.Handle) httprouter.Handle {
 	}
 }
 
+func (h *Handler) unauthorized(w http.ResponseWriter, r *http.Request, err error) {
+	if strings.HasPrefix(r.URL.Path, cacheServiceV2Path) {
+		h.twirpError(w, r, twirpUnauthenticated, err)
+		return
+	}
+	h.responseJSON(w, r, http.StatusUnauthorized, err)
+}
+
 // signedAuth authenticates a signed URL. purpose separates the flavours of URL the
 // handler hands out, so one cannot be replayed as another; see computeSignature.
 func (h *Handler) signedAuth(purpose string, handler httprouter.Handle) httprouter.Handle {
@@ -701,8 +851,7 @@ func (h *Handler) internalRegister(w http.ResponseWriter, r *http.Request, _ htt
 		return
 	}
 	h.RegisterJob(body.Token, body.JobCredential)
-	// A server too old to forward answers without this, which is how the caller knows.
-	h.responseJSON(w, r, http.StatusOK, map[string]any{"results_url": h.ResultsURL(body.JobCredential)})
+	h.responseJSON(w, r, http.StatusOK)
 }
 
 // POST /_internal/revoke
@@ -720,6 +869,12 @@ func (h *Handler) internalRevoke(w http.ResponseWriter, r *http.Request, _ httpr
 	h.responseJSON(w, r, http.StatusOK)
 }
 
+// hashedToken fingerprints a job's bearer, so a reservation can tell its own retry from another job.
+func hashedToken(token string) string {
+	sum := sha256.Sum256([]byte(token))
+	return hex.EncodeToString(sum[:])
+}
+
 func bearerToken(r *http.Request) string {
 	auth := r.Header.Get("Authorization")
 	const prefix = "Bearer "
@@ -770,14 +925,9 @@ func findCache(db *bolthold.Store, repo string, keys []string, version string) (
 		if exact != nil {
 			return exact, nil
 		}
-		prefixPattern := "^" + regexp.QuoteMeta(prefix)
-		re, err := regexp.Compile(prefixPattern)
-		if err != nil {
-			continue
-		}
 		if err := db.FindOne(cache,
-			bolthold.Where("Repo").Eq(repo).
-				And("Key").RegExp(re).
+			bolthold.Where("Repo").Eq(repo).Index("Repo").
+				And("Key").MatchFunc(func(key string) (bool, error) { return strings.HasPrefix(key, prefix), nil }).
 				And("Version").Eq(version).
 				And("Complete").Eq(true).
 				SortBy("CreatedAt").Reverse()); err != nil {
@@ -805,8 +955,8 @@ func findExactCache(db *bolthold.Store, repo, key, version string, complete bool
 	}
 	cache := &Cache{}
 	err := db.FindOne(cache,
-		bolthold.Where("Repo").Eq(repo).
-			And("Key").Eq(key).
+		bolthold.Where("Key").Eq(key).Index("Key").
+			And("Repo").Eq(repo).
 			And("Version").Eq(version).
 			And("Complete").Eq(complete).
 			SortBy(sortBy).Reverse())
@@ -819,15 +969,23 @@ func findExactCache(db *bolthold.Store, repo, key, version string, complete bool
 	return cache, nil
 }
 
-func insertCache(db *bolthold.Store, cache *Cache) error {
-	if err := db.Insert(bolthold.NextSequence(), cache); err != nil {
-		return fmt.Errorf("insert cache: %w", err)
+func (h *Handler) insertCache(db *bolthold.Store, cache *Cache) error {
+	if h.s3 != nil {
+		var id [8]byte
+		_, _ = rand.Read(id[:])
+		cache.ID = binary.BigEndian.Uint64(id[:]) >> 11 // IDs must be unique across runners, and v1 clients parse an ID as a JS number
+		return db.Insert(cache.ID, cache)
 	}
-	// write back id to db
-	if err := db.Update(cache.ID, cache); err != nil {
-		return fmt.Errorf("write back id to db: %w", err)
-	}
-	return nil
+	return db.Bolt().Update(func(tx *bbolt.Tx) error {
+		if err := db.TxInsert(tx, bolthold.NextSequence(), cache); err != nil {
+			return fmt.Errorf("insert cache: %w", err)
+		}
+		// write back id to db
+		if err := db.TxUpdate(tx, cache.ID, cache); err != nil {
+			return fmt.Errorf("write back id to db: %w", err)
+		}
+		return nil
+	})
 }
 
 // touchCache stamps UsedAt so gcCache does not reap an entry mid-upload. With requireIncomplete
@@ -852,6 +1010,9 @@ func (h *Handler) touchCache(id uint64, requireIncomplete bool) error {
 	if requireIncomplete && cache.Complete {
 		return fmt.Errorf("cache %d: already complete", id)
 	}
+	if !touchNeeded(cache) {
+		return nil
+	}
 	cache.UsedAt = time.Now().Unix()
 	return db.Update(cache.ID, cache)
 }
@@ -861,15 +1022,19 @@ const (
 
 	defaultSweepInterval = time.Hour
 
-	// inUseGrace matches artifactURLTTL so an entry outlives every signed URL still usable
-	// for it, and no sweep cuts off a download in progress.
-	inUseGrace = artifactURLTTL
+	touchStale = time.Minute // how stale a completed entry's UsedAt may get before an access rewrites it, sparing hits the fsync
+
+	// inUseGrace covers artifactURLTTL plus the touchStale lag so an entry outlives every signed URL
+	// still usable for it, and no sweep cuts off a download in progress.
+	inUseGrace = artifactURLTTL + touchStale
 
 	// uploadStallTimeout is how long a reservation may sit without a chunk before it counts
 	// as abandoned. Widening it also widens the window for findExactCache to hand a finalize
 	// a stale reservation.
 	uploadStallTimeout = 5 * time.Minute
 
+	metadataSyncInterval = 30 * time.Second
+
 	defaultMinFreeDisk = 1024 * miB
 )
 
@@ -917,6 +1082,9 @@ func (h *Handler) gcCache() {
 	defer db.Close()
 
 	h.evictIncomplete(db)
+	if h.s3 != nil {
+		return // lifecycle rules on the bucket expire its entries
+	}
 	h.evictExpired(db)
 	h.evictSuperseded(db)
 	h.evictOversized(db)
@@ -957,7 +1125,7 @@ func (h *Handler) evictExpired(db *bolthold.Store) {
 		return
 	}
 	// Never below inUseGrace, or a short retention would outrun a signed URL already issued.
-	window := max(h.policy.Retention, inUseGrace)
+	window := max(h.policy.Retention+touchStale, inUseGrace)
 	h.sweep(db, bolthold.Where("UsedAt").Lt(time.Now().Add(-window).Unix()).Index("UsedAt"))
 }
 
@@ -1058,10 +1226,18 @@ func inUse(cache *Cache) bool {
 	return time.Since(time.Unix(cache.UsedAt, 0)) < inUseGrace
 }
 
+// touchNeeded skips a write that would not change the second, and completed entries fresher than touchStale.
+func touchNeeded(cache *Cache) bool {
+	return cache.UsedAt < time.Now().Unix() && (!cache.Complete || time.Since(time.Unix(cache.UsedAt, 0)) >= touchStale)
+}
+
 // touch stamps UsedAt through the caller's store, a bolt write on the read path. It cannot
 // go through touchCache, which opens its own store and would block on the exclusive lock
 // for as long as the caller holds one.
 func (h *Handler) touch(db *bolthold.Store, cache *Cache) {
+	if !touchNeeded(cache) {
+		return
+	}
 	cache.UsedAt = time.Now().Unix()
 	if err := db.Update(cache.ID, cache); err != nil {
 		h.logger.Warnf("touch cache: %v", err)
@@ -1106,7 +1282,11 @@ func totalSize(caches []*Cache) int64 {
 // deleteCache drops an entry and its bytes, reporting whether it went fully. The blob goes
 // first, so a failed unlink leaves the row for the next sweep instead of orphaning bytes.
 func (h *Handler) deleteCache(db *bolthold.Store, cache *Cache) bool {
-	if err := h.storage.Remove(cache.ID); err != nil {
+	storage := h.storage
+	if h.s3 != nil && !cache.Complete {
+		storage = h.s3.Storage // an upload not yet committed has nothing in the bucket
+	}
+	if err := storage.Remove(cache.ID); err != nil {
 		h.logger.Warnf("remove cache blob: %v", err)
 		return false
 	}

act/artifactcache/handler_test.go

@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"net/http/httptest"
 	"os"
 	"path/filepath"
 	"strings"
@@ -52,6 +53,59 @@ func signArtifactURL(h *Handler, id int64) string {
 	return h.signedArtifactURL(JobCredential{}, uint64(id), time.Now().Add(artifactURLTTL))
 }
 
+func TestHandler_ExternalURL(t *testing.T) {
+	assert.Equal(t, "http://[2001:db8::1]:8080", (&Handler{outboundIP: "2001:db8::1", port: 8080}).ExternalURL())
+}
+
+func TestHandler_Upstream(t *testing.T) {
+	var upstream *httptest.Server
+	upstream = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		assert.Equal(t, http.MethodPut, r.Method)
+		assert.Equal(t, "/cache%20server"+blobPath+"/%34%32", r.URL.EscapedPath())
+		assert.Equal(t, "sig=opaque%2Fvalue&raw=a;b", r.URL.RawQuery)
+		assert.Equal(t, strings.TrimPrefix(upstream.URL, "http://"), r.Host)
+		assert.Equal(t, "Bearer unregistered", r.Header.Get("Authorization"))
+		_, err := io.Copy(w, r.Body)
+		assert.NoError(t, err)
+	}))
+	t.Cleanup(upstream.Close)
+	dir := filepath.Join(t.TempDir(), "artifactcache")
+	handler, err := StartHandler(Options{
+		Dir:        dir,
+		OutboundIP: "127.0.0.1",
+		Upstream:   upstream.URL + "/cache%20server//",
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { require.NoError(t, handler.Close()) })
+
+	req, err := http.NewRequestWithContext(t.Context(), http.MethodPut,
+		handler.ExternalURL()+blobPath+"/%34%32?sig=opaque%2Fvalue&raw=a;b", strings.NewReader("cache content"))
+	require.NoError(t, err)
+	req.Header.Set("Authorization", "Bearer unregistered")
+	resp, err := http.DefaultClient.Do(req)
+	require.NoError(t, err)
+	forwarded, err := io.ReadAll(resp.Body)
+	resp.Body.Close()
+	require.NoError(t, err)
+	assert.Equal(t, "cache content", string(forwarded))
+
+	resp, err = http.Post(handler.ExternalURL()+internalPath+"/register", "application/json", nil)
+	require.NoError(t, err)
+	resp.Body.Close()
+	assert.Equal(t, http.StatusNotFound, resp.StatusCode)
+
+	upstream.Close()
+	resp, err = http.Post(handler.ExternalURL()+cacheServiceV2Path+"/CreateCacheEntry", "application/json", strings.NewReader(`{}`))
+	require.NoError(t, err)
+	var failure map[string]string
+	err = json.UnmarshalRead(resp.Body, &failure)
+	resp.Body.Close()
+	require.NoError(t, err)
+	assert.Equal(t, http.StatusServiceUnavailable, resp.StatusCode)
+	assert.Equal(t, map[string]string{"code": "unavailable", "msg": "cache upstream is unavailable"}, failure)
+	assert.NoDirExists(t, dir)
+}
+
 func TestHandler(t *testing.T) {
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
@@ -767,7 +821,7 @@ func TestHandler_gcCache(t *testing.T) {
 				Key:       "test_key_3",
 				Version:   "test_version",
 				Complete:  true,
-				UsedAt:    now.Add(-(testRetention + time.Second)).Unix(),
+				UsedAt:    now.Add(-(testRetention + touchStale + time.Second)).Unix(),
 				CreatedAt: now.Add(-(testRetention + time.Hour)).Unix(),
 			},
 			Kept: false,
@@ -810,7 +864,7 @@ func TestHandler_gcCache(t *testing.T) {
 	db, err := handler.openDB()
 	require.NoError(t, err)
 	for _, c := range cases {
-		require.NoError(t, insertCache(db, c.Cache))
+		require.NoError(t, handler.insertCache(db, c.Cache))
 	}
 	require.NoError(t, db.Close())
 
@@ -982,7 +1036,7 @@ func TestHandler_SweepKeepsEntryWhenBlobSurvives(t *testing.T) {
 	handler := newTestHandler(t, Policy{Retention: testRetention}, cache)
 
 	// A non-empty directory where the blob belongs makes os.Remove fail on every platform.
-	blob := handler.storage.filename(cache.ID)
+	blob := handler.storage.(*Storage).filename(cache.ID)
 	require.NoError(t, os.MkdirAll(blob, 0o755))
 	require.NoError(t, os.WriteFile(filepath.Join(blob, "held"), []byte("x"), 0o600))
 
@@ -1042,27 +1096,51 @@ func TestHandler_gcCacheInterval(t *testing.T) {
 	assert.Empty(t, keptKeys(t, handler, []*Cache{cache}))
 }
 
+func TestHandler_touchCacheSkipsFreshCompletedEntries(t *testing.T) {
+	fresh := &Cache{Repo: testRepo, Key: "fresh", Version: "v", Complete: true, UsedAt: time.Now().Unix()}
+	pending := &Cache{Repo: testRepo, Key: "pending", Version: "v", UsedAt: time.Now().Add(-time.Second).Unix()}
+	handler := newTestHandler(t, Policy{}, fresh, pending)
+	txID := func() int {
+		db, err := handler.openDB()
+		require.NoError(t, err)
+		defer db.Close()
+		var id int
+		require.NoError(t, db.Bolt().View(func(tx *bbolt.Tx) error { id = tx.ID(); return nil }))
+		return id
+	}
+
+	before := txID()
+	require.NoError(t, handler.touchCache(fresh.ID, false))
+	assert.Equal(t, before, txID())
+	require.NoError(t, handler.touchCache(pending.ID, false))
+	assert.Greater(t, txID(), before)
+}
+
 // newTestHandler starts a handler with testToken registered, seeded with entries.
 func newTestHandler(t *testing.T, policy Policy, entries ...*Cache) *Handler {
 	t.Helper()
-	handler, err := StartHandler(Options{
-		Dir:        filepath.Join(t.TempDir(), "artifactcache"),
-		OutboundIP: "127.0.0.1",
-		Policy:     policy,
-	})
-	require.NoError(t, err)
-	t.Cleanup(func() { require.NoError(t, handler.Close()) })
-	handler.RegisterJob(testToken, JobCredential{Repo: testRepo})
+	handler := startTestHandler(t, Options{Policy: policy})
 
 	db, err := handler.openDB()
 	require.NoError(t, err)
 	for _, e := range entries {
-		require.NoError(t, insertCache(db, e))
+		require.NoError(t, handler.insertCache(db, e))
 	}
 	require.NoError(t, db.Close())
 	return handler
 }
 
+func startTestHandler(t *testing.T, opts Options) *Handler {
+	t.Helper()
+	opts.Dir = filepath.Join(t.TempDir(), "artifactcache")
+	opts.OutboundIP = "127.0.0.1"
+	handler, err := StartHandler(opts)
+	require.NoError(t, err)
+	t.Cleanup(func() { require.NoError(t, handler.Close()) })
+	handler.RegisterJob(testToken, JobCredential{Repo: testRepo})
+	return handler
+}
+
 // keptKeys reports which of entries are still in the store.
 func keptKeys(t *testing.T, handler *Handler, entries []*Cache) []string {
 	t.Helper()
@@ -1403,8 +1481,8 @@ func TestHandler_GC_PerRepoDedup(t *testing.T) {
 	stale := time.Now().Add(-inUseGrace - time.Minute).Unix()
 	a := &Cache{Repo: "owner/repoA", Key: key, Version: version, Complete: true, CreatedAt: stale, UsedAt: stale, Size: 1}
 	b := &Cache{Repo: "owner/repoB", Key: key, Version: version, Complete: true, CreatedAt: now, UsedAt: now, Size: 1}
-	require.NoError(t, insertCache(db, a))
-	require.NoError(t, insertCache(db, b))
+	require.NoError(t, handler.insertCache(db, a))
+	require.NoError(t, handler.insertCache(db, b))
 	// Write the backing blobs so the dedup deletion has something to remove.
 	require.NoError(t, handler.storage.Write(a.ID, 0, strings.NewReader("a")))
 	_, err = handler.storage.Commit(a.ID, 1)
@@ -1498,7 +1576,10 @@ func TestHandler_InternalAPI_AuthAndUsage(t *testing.T) {
 		}
 
 		assert.Equal(t, http.StatusUnauthorized, probe("via-internal-api"))
-		assert.Equal(t, http.StatusOK, post("/_internal/register", secret, `{"token":"via-internal-api","repo":"owner/repo"}`))
+		assert.Equal(t, http.StatusOK, post("/_internal/register", secret, `{"token":"via-internal-api","repo":"owner/repo","public_url":"http://runner.example","results":"https://gitea.example","insecure_tls":true}`))
+		cred, ok := handler.lookupCredential("via-internal-api")
+		require.True(t, ok)
+		assert.Equal(t, JobCredential{Repo: testRepo, PublicURL: "http://runner.example"}, cred)
 		assert.NotEqual(t, http.StatusUnauthorized, probe("via-internal-api"))
 		assert.Equal(t, http.StatusOK, post("/_internal/revoke", secret, `{"token":"via-internal-api"}`))
 		assert.Equal(t, http.StatusUnauthorized, probe("via-internal-api"))

act/artifactcache/handler_v2.go

@@ -39,8 +39,9 @@ const (
 
 	blobUploadURLTTL = time.Hour
 
-	// twirpInternal is the only error code that is not the client's fault.
-	twirpInternal = "internal"
+	twirpInternal        = "internal"
+	twirpUnavailable     = "unavailable"
+	twirpUnauthenticated = "unauthenticated"
 )
 
 func (h *Handler) registerV2Routes(router *httprouter.Router) {
@@ -56,7 +57,7 @@ func (h *Handler) v2CreateCacheEntry(w http.ResponseWriter, r *http.Request, _ h
 	cred := credFromContext(r.Context())
 	req, err := decodeTwirpRequest[v2CreateRequest](r)
 	if err != nil {
-		h.twirpError(w, r, "malformed_request", err)
+		h.twirpError(w, r, "malformed", err)
 		return
 	}
 	if req.Key == "" || req.Version == "" {
@@ -83,16 +84,35 @@ func (h *Handler) v2CreateCacheEntry(w http.ResponseWriter, r *http.Request, _ h
 		return
 	}
 
+	// A second live reservation is finalized by whichever job calls last, against the other's upload.
+	owner := hashedToken(bearerToken(r))
+	if pending, err := findExactCache(db, cred.Repo, req.Key, req.Version, false); err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	} else if pending != nil && pending.UsedAt > time.Now().Add(-uploadStallTimeout).Unix() {
+		if pending.Owner != owner {
+			h.twirpNotOK(w, r)
+			return
+		}
+		h.touch(db, pending)                                // still uploading, so it must not go stale under the sweep
+		h.responseJSON(w, r, http.StatusOK, map[string]any{ // this job retrying its own reservation
+			"ok":                true,
+			"signed_upload_url": h.signedURL(cred, blobPath, blobUploadPurpose, pending.ID, time.Now().Add(blobUploadURLTTL)),
+		})
+		return
+	}
+
 	now := time.Now().Unix()
 	cache := &Cache{
 		Repo:      cred.Repo,
 		Key:       req.Key,
 		Version:   req.Version,
+		Owner:     owner,
 		Size:      -1, // the size is only known at finalize time
 		CreatedAt: now,
 		UsedAt:    now,
 	}
-	if err := insertCache(db, cache); err != nil {
+	if err := h.insertCache(db, cache); err != nil {
 		h.twirpError(w, r, twirpInternal, err)
 		return
 	}
@@ -107,7 +127,7 @@ func (h *Handler) v2FinalizeCacheEntryUpload(w http.ResponseWriter, r *http.Requ
 	cred := credFromContext(r.Context())
 	req, err := decodeTwirpRequest[v2FinalizeRequest](r)
 	if err != nil {
-		h.twirpError(w, r, "malformed_request", err)
+		h.twirpError(w, r, "malformed", err)
 		return
 	}
 
@@ -119,6 +139,13 @@ func (h *Handler) v2FinalizeCacheEntryUpload(w http.ResponseWriter, r *http.Requ
 	defer db.Close()
 
 	cache, err := findExactCache(db, cred.Repo, req.Key, req.Version, false)
+	if err == nil && cache != nil && cache.Owner != hashedToken(bearerToken(r)) {
+		cache = nil // not the reservation this job made, so not this job's to commit
+	}
+	if err == nil && cache == nil {
+		// A retry whose first response was lost finds the entry already committed.
+		cache, err = findExactCache(db, cred.Repo, req.Key, req.Version, true)
+	}
 	if err != nil {
 		h.twirpError(w, r, twirpInternal, err)
 		return
@@ -127,13 +154,15 @@ func (h *Handler) v2FinalizeCacheEntryUpload(w http.ResponseWriter, r *http.Requ
 		h.twirpNotOK(w, r)
 		return
 	}
-	db.Close() // commitCache needs the store closed
 
-	cache.Size = int64(cmp.Or(req.SizeBytes, req.SizeBytesCamel))
-	if err := h.commitCache(cache); err != nil {
-		h.logger.Errorf("finalize cache %d (%s): %v", cache.ID, cache.Key, err)
-		h.twirpNotOK(w, r)
-		return
+	if !cache.Complete {
+		db.Close() // commitCache needs the store closed
+		cache.Size = int64(cmp.Or(req.SizeBytes, req.SizeBytesCamel))
+		if err := h.commitCache(cache); err != nil {
+			h.logger.Errorf("finalize cache %d (%s): %v", cache.ID, cache.Key, err)
+			h.twirpNotOK(w, r)
+			return
+		}
 	}
 
 	h.responseJSON(w, r, http.StatusOK, map[string]any{
@@ -147,7 +176,7 @@ func (h *Handler) v2GetCacheEntryDownloadURL(w http.ResponseWriter, r *http.Requ
 	cred := credFromContext(r.Context())
 	req, err := decodeTwirpRequest[v2DownloadRequest](r)
 	if err != nil {
-		h.twirpError(w, r, "malformed_request", err)
+		h.twirpError(w, r, "malformed", err)
 		return
 	}
 
@@ -230,8 +259,13 @@ func (h *Handler) twirpNotOK(w http.ResponseWriter, r *http.Request) {
 func (h *Handler) twirpError(w http.ResponseWriter, r *http.Request, code string, err error) {
 	h.logger.Debugf("%s %s: %v", r.Method, r.URL.Path, err)
 	status := http.StatusBadRequest
-	if code == twirpInternal {
+	switch code {
+	case twirpInternal:
 		status = http.StatusInternalServerError
+	case twirpUnavailable:
+		status = http.StatusServiceUnavailable
+	case twirpUnauthenticated:
+		status = http.StatusUnauthorized
 	}
 	h.responseJSON(w, r, status, map[string]any{"code": code, "msg": err.Error()})
 }

act/artifactcache/handler_v2_test.go

@@ -10,6 +10,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"net/url"
 	"strconv"
 	"strings"
 	"testing"
@@ -141,6 +142,17 @@ func TestCacheServiceV2BlockUpload(t *testing.T) {
 	}
 	list.WriteString(`</BlockList>`)
 	require.Equal(t, http.StatusCreated, putBlob(t, uploadURL+"&comp=blocklist", list.Bytes()))
+	require.Equal(t, http.StatusCreated, putBlob(t, uploadURL+"&comp=blocklist", list.Bytes()),
+		"a client that lost the first answer retries the list it already sent")
+
+	var reordered bytes.Buffer
+	reordered.WriteString(`<?xml version="1.0" encoding="utf-8"?><BlockList>`)
+	for _, blockID := range []string{order[1], order[0], order[2]} {
+		fmt.Fprintf(&reordered, "<Latest>%s</Latest>", blockID)
+	}
+	reordered.WriteString(`</BlockList>`)
+	assert.Equal(t, http.StatusInternalServerError, putBlob(t, uploadURL+"&comp=blocklist", reordered.Bytes()),
+		"the blocks are already assembled, so a different order would silently disagree with them")
 
 	finalized := v2Call(t, handler, testClient, "FinalizeCacheEntryUpload", map[string]any{
 		"key": "blocks", "version": "v1", "size_bytes": len("hello world!"),
@@ -241,3 +253,101 @@ func TestCacheServiceV2Lookups(t *testing.T) {
 		assert.Equal(t, false, got["ok"])
 	})
 }
+
+func TestCacheServiceV2RefusalsAreTwirp(t *testing.T) {
+	handler := newTestHandler(t, Policy{})
+
+	tests := []struct {
+		name   string
+		token  string
+		body   string
+		status int
+		code   string
+	}{
+		{"no bearer at all", "", `{"key":"k","version":"v"}`, http.StatusUnauthorized, twirpUnauthenticated},
+		{"a bearer nobody registered", "not-a-job", `{"key":"k","version":"v"}`, http.StatusUnauthorized, twirpUnauthenticated},
+		{"a body that is not the request", testToken, `{`, http.StatusBadRequest, "malformed"},
+		{"a request missing its key", testToken, `{"version":"v"}`, http.StatusBadRequest, "invalid_argument"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			req, err := http.NewRequestWithContext(t.Context(), http.MethodPost,
+				handler.ExternalURL()+cacheServiceV2Path+"/CreateCacheEntry", strings.NewReader(tt.body))
+			require.NoError(t, err)
+			if tt.token != "" {
+				req.Header.Set("Authorization", "Bearer "+tt.token)
+			}
+			resp, err := http.DefaultClient.Do(req)
+			require.NoError(t, err)
+			defer resp.Body.Close()
+
+			require.Equal(t, tt.status, resp.StatusCode)
+			got := map[string]string{}
+			require.NoError(t, json.UnmarshalRead(resp.Body, &got))
+			assert.Equal(t, tt.code, got["code"])
+			assert.NotEmpty(t, got["msg"])
+		})
+	}
+}
+
+func TestCacheServiceV2ConcurrentSaveAndRetries(t *testing.T) {
+	handler := newTestHandler(t, Policy{})
+	content := []byte("the cached archive")
+
+	const otherJob = "another-jobs-token"
+	defer handler.RegisterJob(otherJob, JobCredential{Repo: testRepo})()
+	reserve := func(client *http.Client) map[string]any {
+		return v2Call(t, handler, client, "CreateCacheEntry", map[string]any{"key": "shared", "version": "v1"})
+	}
+	first := reserve(testClient)
+	require.Equal(t, true, first["ok"])
+	retry := reserve(testClient)
+	assert.Equal(t, true, retry["ok"], "a job retrying its own reservation is not a conflict")
+	entryOf := func(reserved map[string]any) string {
+		parsed, err := url.Parse(reserved["signed_upload_url"].(string))
+		require.NoError(t, err)
+		return parsed.Path
+	}
+	assert.Equal(t, entryOf(first), entryOf(retry), "the same reservation, however freshly its URL is signed")
+	assert.Equal(t, false, reserve(&http.Client{Transport: &bearerTransport{token: otherJob}})["ok"],
+		"another job's reservation would be finalized against the first upload")
+
+	otherClient := &http.Client{Transport: &bearerTransport{token: otherJob}}
+	reserved := v2Call(t, handler, otherClient, "CreateCacheEntry", map[string]any{"key": "owned", "version": "v1"})
+	require.Equal(t, true, reserved["ok"])
+	require.Equal(t, http.StatusCreated, putBlob(t, reserved["signed_upload_url"].(string), content))
+	assert.Equal(t, false, v2Call(t, handler, testClient, "FinalizeCacheEntryUpload",
+		map[string]any{"key": "owned", "version": "v1", "size_bytes": strconv.Itoa(len(content))})["ok"],
+		"another job's upload is not this job's to finalize")
+
+	base := handler.ExternalURL() + apiPath
+	reserveV1, err := json.Marshal(&Request{Key: "v1-made", Version: "v1", Size: int64(len(content))})
+	require.NoError(t, err)
+	resp, err := testClient.Post(base+"/caches", "application/json", bytes.NewReader(reserveV1))
+	require.NoError(t, err)
+	var madeByV1 struct {
+		CacheID uint64 `json:"cacheId"`
+	}
+	require.NoError(t, json.UnmarshalRead(resp.Body, &madeByV1))
+	resp.Body.Close()
+	upload, err := http.NewRequestWithContext(t.Context(), http.MethodPatch,
+		fmt.Sprintf("%s/caches/%d", base, madeByV1.CacheID), bytes.NewReader(content))
+	require.NoError(t, err)
+	upload.Header.Set("Content-Range", fmt.Sprintf("bytes 0-%d/*", len(content)-1))
+	resp, err = testClient.Do(upload)
+	require.NoError(t, err)
+	resp.Body.Close()
+	require.Equal(t, http.StatusOK, resp.StatusCode)
+
+	assert.Equal(t, false, v2Call(t, handler, testClient, "FinalizeCacheEntryUpload",
+		map[string]any{"key": "v1-made", "version": "v1", "size_bytes": strconv.Itoa(len(content))})["ok"],
+		"a reservation this job did not make through v2 is not its to commit")
+
+	finalized, _ := saveV2(t, handler, "deps", "v1", content)
+	require.Equal(t, true, finalized["ok"])
+	again := v2Call(t, handler, testClient, "FinalizeCacheEntryUpload", map[string]any{
+		"key": "deps", "version": "v1", "size_bytes": strconv.Itoa(len(content)),
+	})
+	assert.Equal(t, true, again["ok"], "a retry whose first answer was lost must not report a failed save")
+	assert.Equal(t, finalized["entry_id"], again["entry_id"])
+}

act/artifactcache/model.go

@@ -16,6 +16,7 @@ type Cache struct {
 	Key       string `json:"key" boltholdIndex:"Key"`
 	Version   string `json:"version" boltholdIndex:"Version"`
 	Size      int64  `json:"cacheSize"`
+	Owner     string `json:"owner"` // hashed bearer, so a job's own retry is not another job's reservation
 	Complete  bool   `json:"complete" boltholdIndex:"Complete"`
 	UsedAt    int64  `json:"usedAt" boltholdIndex:"UsedAt"`
 	CreatedAt int64  `json:"createdAt" boltholdIndex:"CreatedAt"`

act/artifactcache/results.go

@@ -5,6 +5,8 @@ package artifactcache
 
 import (
 	"crypto/tls"
+	"errors"
+	"fmt"
 	"net/http"
 	"net/http/httputil"
 	"net/url"
@@ -15,23 +17,38 @@ import (
 // Gitea implements only the artifact half of it. Forwarding that half from here makes this origin
 // the whole service, so ACTIONS_RESULTS_URL can point at it truthfully, which is what the clients
 // this runner cannot patch need, docker buildx among them.
-//
-// The instance to forward to travels with the job registration rather than with configuration, so
-// a cache server shared between runners serves each of their instances.
 const artifactServicePath = "/twirp/github.actions.results.api.v1.ArtifactService/"
 
 // forwardOrNotFound is the router's fallback: the artifact service of the instance the job
 // registered with, and the 404 the router would have written otherwise.
 func (h *Handler) forwardOrNotFound(w http.ResponseWriter, r *http.Request) {
-	cred, ok := h.lookupCredential(bearerToken(r))
-	if !ok || cred.Results == "" || !strings.HasPrefix(r.URL.Path, artifactServicePath) {
+	if !strings.HasPrefix(r.URL.Path, artifactServicePath) {
 		http.NotFound(w, r)
 		return
 	}
+	cred, ok := h.lookupCredential(bearerToken(r))
+	if !ok {
+		h.twirpError(w, r, twirpUnauthenticated, errors.New("unknown bearer token"))
+		return
+	}
+	if cred.Results == "" {
+		h.twirpError(w, r, twirpInternal, errors.New("no instance is registered for this job"))
+		return
+	}
 	target, err := url.Parse(strings.TrimSuffix(cred.Results, "/"))
+	if err == nil && (target.Hostname() == "" || (target.Scheme != "http" && target.Scheme != "https")) {
+		err = errors.New("not an absolute http or https URL the job could use")
+	}
 	if err != nil {
 		h.logger.Errorf("artifact service forward to %q: %v", cred.Results, err)
-		w.WriteHeader(http.StatusBadGateway)
+		h.twirpError(w, r, twirpInternal, fmt.Errorf("artifact service address %q is unusable: %w", cred.Results, err))
+		return
+	}
+	if !h.mustProxy(r, cred, target) {
+		redirect := target.JoinPath(r.URL.EscapedPath())
+		redirect.RawQuery = mergeQuery(target.RawQuery, r.URL.RawQuery)
+		h.logger.Debugf("%s %s: redirecting to %s", r.Method, r.URL.Path, target)
+		http.Redirect(w, r, redirect.String(), http.StatusTemporaryRedirect)
 		return
 	}
 	h.logger.Debugf("%s %s: forwarding to %s", r.Method, r.URL.Path, target)
@@ -44,9 +61,9 @@ func (h *Handler) forwardOrNotFound(w http.ResponseWriter, r *http.Request) {
 			// directly and never through here.
 			r.Out.Host = target.Host
 		},
-		ErrorHandler: func(w http.ResponseWriter, _ *http.Request, err error) {
+		ErrorHandler: func(w http.ResponseWriter, r *http.Request, err error) {
 			h.logger.Warnf("artifact service forward to %s: %v", target, err)
-			w.WriteHeader(http.StatusBadGateway)
+			h.twirpError(w, r, twirpInternal, fmt.Errorf("cache server cannot reach the artifact service at %s: %w", target, err))
 		},
 	}
 	if cred.InsecureTLS {
@@ -55,5 +72,24 @@ func (h *Handler) forwardOrNotFound(w http.ResponseWriter, r *http.Request) {
 	proxy.ServeHTTP(w, r)
 }
 
+// A protocol switch loses the bearer or the agent, and an untrusted instance needs skipped verification.
+func (h *Handler) mustProxy(r *http.Request, cred JobCredential, target *url.URL) bool {
+	if cred.InsecureTLS && target.Scheme == "https" {
+		return true
+	}
+	scheme := r.Header.Get("X-Forwarded-Proto")
+	if scheme == "" {
+		scheme = "http"
+	}
+	return scheme != target.Scheme
+}
+
+func mergeQuery(target, request string) string {
+	if target == "" || request == "" {
+		return target + request
+	}
+	return target + "&" + request
+}
+
 // insecureTransport is shared, because a transport per request would pool no connections.
-var insecureTransport = &http.Transport{TLSClientConfig: &tls.Config{InsecureSkipVerify: true}} //nolint:gosec // the runner reaches its instance on the operator's say-so
+var insecureTransport = &http.Transport{Proxy: http.ProxyFromEnvironment, TLSClientConfig: &tls.Config{InsecureSkipVerify: true}} //nolint:gosec // the runner reaches its instance on the operator's say-so

act/artifactcache/results_test.go

@@ -4,6 +4,8 @@
 package artifactcache
 
 import (
+	"cmp"
+	"encoding/json/v2"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -14,8 +16,6 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-// The artifact half is forwarded under the Host Gitea knows itself by, so the URLs it hands back
-// still point at Gitea, and nothing else is proxied.
 func TestFrontResultsService(t *testing.T) {
 	var gotHost, gotPath, gotProto string
 	gitea := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
@@ -23,34 +23,112 @@ func TestFrontResultsService(t *testing.T) {
 		_, _ = io.WriteString(w, `{"ok":true}`)
 	}))
 	defer gitea.Close()
+	selfSigned := httptest.NewTLSServer(gitea.Config.Handler)
+	defer selfSigned.Close()
 
 	handler, err := StartHandler(Options{Dir: t.TempDir(), OutboundIP: "127.0.0.1"})
 	require.NoError(t, err)
 	defer handler.Close()
-	const token = "forward-token"
+	const token, rpc = "forward-token", artifactServicePath + "CreateArtifact"
 
-	client := &http.Client{Transport: &bearerTransport{token: token}}
-	post := func(path string) int {
+	noRedirect := func(token string) *http.Client {
+		return &http.Client{
+			Transport:     &bearerTransport{token: token},
+			CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse },
+		}
+	}
+	post := func(path string) (int, map[string]string) {
 		req, err := http.NewRequestWithContext(t.Context(), http.MethodPost, handler.ExternalURL()+path, nil)
 		require.NoError(t, err)
-		resp, err := client.Do(req)
+		resp, err := noRedirect(token).Do(req)
 		require.NoError(t, err)
-		resp.Body.Close()
-		return resp.StatusCode
+		defer resp.Body.Close()
+		body := map[string]string{}
+		_ = json.UnmarshalRead(resp.Body, &body)
+		return resp.StatusCode, body
 	}
 
-	assert.Equal(t, http.StatusNotFound, post(artifactServicePath+"CreateArtifact"),
-		"an unregistered token is forwarded nowhere")
+	status, body := post(rpc)
+	assert.Equal(t, http.StatusUnauthorized, status, "an unregistered token is told so, not left to parse a 404 page")
+	assert.Equal(t, twirpUnauthenticated, body["code"])
+	assert.Contains(t, body["msg"], "bearer")
+	assert.NotEmpty(t, body["msg"])
 
-	defer handler.RegisterJob(token, JobCredential{Repo: "owner/repo", Results: gitea.URL})()
+	defer handler.RegisterJob(token, JobCredential{Repo: "owner/repo", Results: selfSigned.URL, InsecureTLS: true})()
 
-	assert.Equal(t, http.StatusOK, post(artifactServicePath+"CreateArtifact"))
-	assert.Equal(t, strings.TrimPrefix(gitea.URL, "http://"), gotHost, "Gitea must see the host it mints its URLs from")
+	status, _ = post(rpc)
+	require.Equal(t, http.StatusOK, status, "an http job reaching an https instance must proxy")
+	assert.Equal(t, strings.TrimPrefix(selfSigned.URL, "https://"), gotHost, "Gitea must see the host it mints its URLs from")
 	assert.Empty(t, gotProto, "a forwarded scheme would make an https Gitea mint http URLs")
-	assert.Equal(t, artifactServicePath+"CreateArtifact", gotPath)
+	assert.Equal(t, rpc, gotPath)
 
 	gotPath = ""
-	assert.Equal(t, http.StatusNotFound, post("/twirp/github.actions.results.api.v1.OtherService/Do"))
-	assert.Equal(t, http.StatusNotFound, post("/api/v1/repos/owner/repo"))
+	status, _ = post("/twirp/github.actions.results.api.v1.OtherService/Do")
+	assert.Equal(t, http.StatusNotFound, status)
+	status, _ = post("/api/v1/repos/owner/repo")
+	assert.Equal(t, http.StatusNotFound, status)
 	assert.Empty(t, gotPath, "only the artifact service is forwarded")
+
+	tests := []struct {
+		name           string
+		cred           JobCredential
+		forwardedProto string
+		reqQuery       string
+		want           int
+		location       string
+		wantMsg        string
+	}{
+		{name: "an instance the job reaches itself", cred: JobCredential{Results: gitea.URL}, want: http.StatusTemporaryRedirect},
+		{name: "a sub-path instance keeps its prefix", cred: JobCredential{Results: gitea.URL + "/sub"}, want: http.StatusTemporaryRedirect},
+		{name: "an instance with no certificate to distrust", cred: JobCredential{Results: gitea.URL, InsecureTLS: true}, want: http.StatusTemporaryRedirect},
+		{
+			name: "both queries survive", cred: JobCredential{Results: gitea.URL + "?tenant=one"}, reqQuery: "?run=7",
+			want: http.StatusTemporaryRedirect, location: gitea.URL + rpc + "?tenant=one&run=7",
+		},
+		{
+			name: "a trusted https instance is reached by the job itself",
+			cred: JobCredential{Results: selfSigned.URL}, forwardedProto: "https", want: http.StatusTemporaryRedirect,
+		},
+		{
+			name: "the job does not share this server's disregard for the certificate",
+			cred: JobCredential{Results: selfSigned.URL, InsecureTLS: true}, forwardedProto: "https", want: http.StatusOK,
+		},
+		{name: "an http job is not redirected to an https instance", cred: JobCredential{Results: selfSigned.URL}, want: http.StatusInternalServerError},
+		{
+			name: "a forwarded https scheme onto an http instance is proxied",
+			cred: JobCredential{Results: gitea.URL}, forwardedProto: "https", want: http.StatusOK,
+		},
+		{
+			name: "a scheme the job could not use is refused before it is sent one",
+			cred: JobCredential{Results: "ftp://gitea.example"}, want: http.StatusInternalServerError, wantMsg: "unusable",
+		},
+		{
+			name: "an address with no host is refused before it is sent one",
+			cred: JobCredential{Results: "http://:3000"}, want: http.StatusInternalServerError, wantMsg: "unusable",
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			defer handler.RegisterJob(tt.name, tt.cred)()
+			req, err := http.NewRequestWithContext(t.Context(), http.MethodPost, handler.ExternalURL()+rpc+tt.reqQuery, nil)
+			require.NoError(t, err)
+			if tt.forwardedProto != "" {
+				req.Header.Set("X-Forwarded-Proto", tt.forwardedProto)
+			}
+			resp, err := noRedirect(tt.name).Do(req)
+			require.NoError(t, err)
+			defer resp.Body.Close()
+
+			require.Equal(t, tt.want, resp.StatusCode)
+			switch tt.want {
+			case http.StatusTemporaryRedirect:
+				assert.Equal(t, cmp.Or(tt.location, tt.cred.Results+rpc), resp.Header.Get("Location"))
+			case http.StatusInternalServerError:
+				body := map[string]string{}
+				require.NoError(t, json.UnmarshalRead(resp.Body, &body))
+				assert.Equal(t, twirpInternal, body["code"])
+				assert.Contains(t, body["msg"], cmp.Or(tt.wantMsg, tt.cred.Results))
+			}
+		})
+	}
 }

act/artifactcache/s3_storage.go

@@ -0,0 +1,326 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"bytes"
+	"cmp"
+	"context"
+	"crypto/hmac"
+	"crypto/sha256"
+	"encoding/json/v2"
+	"encoding/xml"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"maps"
+	"net/http"
+	"net/url"
+	"os"
+	"path"
+	"slices"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// S3Options configures the bucket that stores completed cache entries.
+type S3Options struct {
+	Endpoint  string
+	Region    string
+	Bucket    string
+	Prefix    string
+	PathStyle *bool // nil uses virtual-hosted requests on AWS and path-style ones elsewhere
+	AccessKey string
+	SecretKey string
+}
+
+type cacheStorage interface {
+	Exist(uint64) (bool, error)
+	Write(uint64, int64, io.Reader) error
+	WriteBlock(uint64, string, io.Reader) error
+	OrderBlocks(uint64, []string) error
+	Commit(uint64, int64) (int64, error)
+	Serve(http.ResponseWriter, *http.Request, uint64)
+	Remove(uint64) error
+}
+
+const s3PartSize = 5 << 30 // the largest part S3 accepts
+
+// s3Storage stages uploads on disk and keeps completed entries and their metadata in a bucket.
+type s3Storage struct {
+	*Storage
+	client    *http.Client
+	bucketURL *url.URL
+	opts      S3Options
+}
+
+type s3Part struct {
+	PartNumber int
+	ETag       string
+}
+
+func newS3Storage(staging *Storage, opts S3Options) (*s3Storage, error) {
+	endpoint, err := url.Parse(opts.Endpoint)
+	if err != nil || endpoint.Host == "" || (endpoint.Scheme != "http" && endpoint.Scheme != "https") {
+		return nil, fmt.Errorf("invalid S3 endpoint %q", opts.Endpoint)
+	}
+	transport, _ := http.DefaultTransport.(*http.Transport)
+	transport = transport.Clone()
+	transport.ResponseHeaderTimeout = time.Minute // a stalled endpoint must not hold the store forever
+	pathStyle := !strings.HasSuffix(endpoint.Hostname(), ".amazonaws.com")
+	if opts.PathStyle != nil {
+		pathStyle = *opts.PathStyle
+	}
+	if pathStyle {
+		endpoint.Path = "/" + opts.Bucket
+	} else {
+		endpoint.Host = opts.Bucket + "." + endpoint.Host
+		endpoint.Path = "/"
+	}
+	opts.Region = cmp.Or(opts.Region, "us-east-1")
+	return &s3Storage{Storage: staging, client: &http.Client{Transport: transport}, bucketURL: endpoint, opts: opts}, nil
+}
+
+func (s *s3Storage) blobKey(id uint64) string {
+	return path.Join(s.opts.Prefix, "blobs", strconv.FormatUint(id, 10))
+}
+
+func (s *s3Storage) metadataDir() string {
+	return path.Join(s.opts.Prefix, "metadata") + "/"
+}
+
+func (s *s3Storage) metadataKey(id uint64) string {
+	return s.metadataDir() + strconv.FormatUint(id, 10) + ".json"
+}
+
+func (s *s3Storage) url(key string, query url.Values) *url.URL {
+	u := *s.bucketURL
+	u.Path = path.Join(u.Path, key)
+	u.RawPath = strings.NewReplacer("+", "%20", "%2F", "/").Replace(url.QueryEscape(u.Path)) // SigV4 encoding
+	u.RawQuery = strings.ReplaceAll(query.Encode(), "+", "%20")
+	return &u
+}
+
+// sign authorizes req with AWS Signature Version 4, leaving the payload unsigned so bodies can stream.
+// https://docs.aws.amazon.com/AmazonS3/latest/API/sig-v4-header-based-auth.html
+func (s *s3Storage) sign(req *http.Request, now time.Time) {
+	date := now.UTC().Format("20060102T150405Z")
+	req.Header.Set("X-Amz-Date", date)
+	if req.Header.Get("X-Amz-Content-Sha256") == "" {
+		req.Header.Set("X-Amz-Content-Sha256", "UNSIGNED-PAYLOAD")
+	}
+	headers := map[string]string{"host": req.URL.Host}
+	for name, values := range req.Header {
+		if name = strings.ToLower(name); strings.HasPrefix(name, "x-amz-") {
+			headers[name] = strings.Join(values, ",")
+		}
+	}
+	names := slices.Sorted(maps.Keys(headers))
+	canonical := []string{req.Method, req.URL.EscapedPath(), req.URL.RawQuery}
+	for _, name := range names {
+		canonical = append(canonical, name+":"+headers[name])
+	}
+	canonical = append(canonical, "", strings.Join(names, ";"), req.Header.Get("X-Amz-Content-Sha256"))
+	scope := date[:8] + "/" + s.opts.Region + "/s3/aws4_request"
+	signature := []byte("AWS4" + s.opts.SecretKey)
+	for _, data := range []string{date[:8], s.opts.Region, "s3", "aws4_request", fmt.Sprintf("AWS4-HMAC-SHA256\n%s\n%s\n%x", date, scope, sha256.Sum256([]byte(strings.Join(canonical, "\n"))))} { // key derivation, the last round signs
+		mac := hmac.New(sha256.New, signature)
+		_, _ = mac.Write([]byte(data))
+		signature = mac.Sum(nil)
+	}
+	req.Header.Set("Authorization", fmt.Sprintf("AWS4-HMAC-SHA256 Credential=%s/%s,SignedHeaders=%s,Signature=%x", s.opts.AccessKey, scope, strings.Join(names, ";"), signature))
+}
+
+// do returns the response to a signed request when it succeeded, fs.ErrNotExist for a missing key.
+func (s *s3Storage) do(ctx context.Context, method, key string, query url.Values, header http.Header, body io.Reader) (*http.Response, error) {
+	var size int64
+	if sized, ok := body.(interface{ Size() int64 }); ok {
+		size = sized.Size()
+	}
+	if size == 0 {
+		body = nil
+	}
+	req, err := http.NewRequestWithContext(ctx, method, s.url(key, query).String(), body)
+	if err != nil {
+		return nil, err
+	}
+	req.ContentLength = size
+	maps.Copy(req.Header, header)
+	s.sign(req, time.Now())
+	resp, err := s.client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	if resp.StatusCode < http.StatusMultipleChoices {
+		return resp, nil
+	}
+	defer drain(resp)
+	var failure struct{ Code, Message string }
+	_ = xml.NewDecoder(resp.Body).Decode(&failure)
+	if resp.StatusCode == http.StatusNotFound && (failure.Code == "NoSuchKey" || failure.Code == "" && method == http.MethodHead) {
+		return nil, fs.ErrNotExist
+	}
+	return nil, fmt.Errorf("S3 %s %s: %s", method, key, cmp.Or(strings.TrimSpace(failure.Code+" "+failure.Message), resp.Status))
+}
+
+func (s *s3Storage) call(ctx context.Context, method, key string, query url.Values, body io.Reader, result any) (http.Header, error) {
+	resp, err := s.do(ctx, method, key, query, nil, body)
+	if err != nil {
+		return nil, err
+	}
+	defer drain(resp)
+	if result == nil {
+		return resp.Header, nil
+	}
+	return resp.Header, xml.NewDecoder(resp.Body).Decode(result)
+}
+
+// drain reads a response to its end before closing it, so its connection gets reused.
+func drain(resp *http.Response) {
+	_, _ = io.Copy(io.Discard, resp.Body)
+	_ = resp.Body.Close()
+}
+
+func (s *s3Storage) Exist(id uint64) (bool, error) {
+	_, err := s.call(context.Background(), http.MethodHead, s.blobKey(id), nil, nil, nil)
+	if errors.Is(err, fs.ErrNotExist) {
+		return false, nil
+	}
+	return err == nil, err
+}
+
+func (s *s3Storage) Commit(id uint64, size int64) (int64, error) {
+	written, err := s.Storage.Commit(id, size)
+	if err != nil {
+		return 0, err
+	}
+	defer func() { _ = s.Storage.Remove(id) }()
+	file, err := os.Open(s.filename(id))
+	if err != nil {
+		return 0, err
+	}
+	defer file.Close()
+	if err := s.upload(context.Background(), s.blobKey(id), file, written); err != nil {
+		return 0, fmt.Errorf("upload cache to S3: %w", err)
+	}
+	return written, nil
+}
+
+// upload always goes in parts, the only way S3 takes an object over 5 GiB.
+func (s *s3Storage) upload(ctx context.Context, key string, file io.ReaderAt, size int64) error {
+	var created struct {
+		UploadID string `xml:"UploadId"`
+	}
+	if _, err := s.call(ctx, http.MethodPost, key, url.Values{"uploads": {""}}, nil, &created); err != nil {
+		return err
+	}
+	var completion struct {
+		XMLName xml.Name `xml:"CompleteMultipartUpload"`
+		Parts   []s3Part `xml:"Part"`
+	}
+	for offset := int64(0); offset == 0 || offset < size; offset += s3PartSize {
+		number, length := len(completion.Parts)+1, min(s3PartSize, size-offset)
+		query := url.Values{"partNumber": {strconv.Itoa(number)}, "uploadId": {created.UploadID}}
+		header, err := s.call(ctx, http.MethodPut, key, query, io.NewSectionReader(file, offset, length), nil)
+		if err != nil {
+			return err
+		}
+		completion.Parts = append(completion.Parts, s3Part{number, header.Get("ETag")})
+	}
+	body, err := xml.Marshal(completion)
+	if err != nil {
+		return err
+	}
+	var result struct{ Code, Message string } // a completion can fail after its 200 status
+	if _, err := s.call(ctx, http.MethodPost, key, url.Values{"uploadId": {created.UploadID}}, bytes.NewReader(body), &result); err != nil {
+		return err
+	}
+	if result.Code != "" {
+		return fmt.Errorf("S3 complete upload %s: %s %s", key, result.Code, result.Message)
+	}
+	return nil
+}
+
+func (s *s3Storage) Serve(w http.ResponseWriter, r *http.Request, id uint64) {
+	header := http.Header{}
+	if byteRange := r.Header.Get("Range"); byteRange != "" {
+		header.Set("Range", byteRange)
+	}
+	resp, err := s.do(r.Context(), http.MethodGet, s.blobKey(id), nil, header, nil)
+	if errors.Is(err, fs.ErrNotExist) {
+		http.NotFound(w, r)
+		return
+	} else if err != nil {
+		http.Error(w, "read cache from S3", http.StatusBadGateway)
+		return
+	}
+	defer resp.Body.Close()
+	for _, name := range []string{"Content-Length", "Content-Range"} {
+		if value := resp.Header.Get(name); value != "" {
+			w.Header().Set(name, value)
+		}
+	}
+	w.WriteHeader(resp.StatusCode)
+	_, _ = io.Copy(w, resp.Body)
+}
+
+// Remove unpublishes an entry before deleting its blob, so no runner imports it without its bytes.
+func (s *s3Storage) Remove(id uint64) error {
+	for _, key := range []string{s.metadataKey(id), s.blobKey(id)} {
+		if _, err := s.call(context.Background(), http.MethodDelete, key, nil, nil, nil); err != nil && !errors.Is(err, fs.ErrNotExist) {
+			return fmt.Errorf("remove %s from S3: %w", key, err)
+		}
+	}
+	return s.Storage.Remove(id)
+}
+
+func (s *s3Storage) publish(cache *Cache) error {
+	data, err := json.Marshal(cache)
+	if err != nil {
+		return err
+	}
+	if _, err := s.call(context.Background(), http.MethodPut, s.metadataKey(cache.ID), nil, bytes.NewReader(data), nil); err != nil {
+		return fmt.Errorf("publish cache to S3: %w", err)
+	}
+	return nil
+}
+
+func (s *s3Storage) published(ctx context.Context) (map[uint64]bool, error) {
+	ids := make(map[uint64]bool)
+	query := url.Values{"list-type": {"2"}, "prefix": {s.metadataDir()}}
+	for {
+		var page struct {
+			Contents              []struct{ Key string }
+			IsTruncated           bool
+			NextContinuationToken string
+		}
+		if _, err := s.call(ctx, http.MethodGet, "", query, nil, &page); err != nil {
+			return nil, fmt.Errorf("list S3 cache metadata: %w", err)
+		}
+		for _, object := range page.Contents {
+			if id, err := strconv.ParseUint(strings.TrimSuffix(strings.TrimPrefix(object.Key, s.metadataDir()), ".json"), 10, 64); err == nil {
+				ids[id] = true
+			}
+		}
+		if !page.IsTruncated {
+			return ids, nil
+		}
+		query.Set("continuation-token", page.NextContinuationToken)
+	}
+}
+
+func (s *s3Storage) fetch(ctx context.Context, id uint64) (*Cache, error) {
+	resp, err := s.do(ctx, http.MethodGet, s.metadataKey(id), nil, nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+	cache := &Cache{}
+	if err := json.UnmarshalRead(resp.Body, cache); err != nil {
+		return nil, fmt.Errorf("read S3 cache metadata %d: %w", id, err)
+	}
+	return cache, nil
+}

act/artifactcache/s3_storage_test.go

@@ -0,0 +1,46 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/johannesboyne/gofakes3"
+	"github.com/johannesboyne/gofakes3/backend/s3mem"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestHandler_S3SharesEntriesAndDropsExpiredOnes(t *testing.T) {
+	backend := s3mem.New()
+	require.NoError(t, backend.CreateBucket("cache"))
+	server := httptest.NewServer(gofakes3.New(backend).Server())
+	t.Cleanup(server.Close)
+	opts := Options{S3: &S3Options{Endpoint: server.URL, Bucket: "cache", Prefix: "runners", AccessKey: "key", SecretKey: "secret"}}
+	saver, restorer := startTestHandler(t, opts), startTestHandler(t, opts)
+	entry := map[string]any{"key": "deps", "version": "v1"}
+
+	finalized, _ := saveV2(t, saver, "deps", "v1", []byte("archive"))
+	require.Equal(t, true, finalized["ok"])
+	require.NoError(t, restorer.importMetadata(t.Context()))
+	found := v2Call(t, restorer, testClient, "GetCacheEntryDownloadURL", entry)
+	require.Equal(t, true, found["ok"])
+	assert.Equal(t, []byte("archive"), getURL(t, found["signed_download_url"].(string)))
+
+	require.NoError(t, backend.ForceDeleteBucket("cache"))
+	require.NoError(t, backend.CreateBucket("cache"))
+	require.NoError(t, restorer.importMetadata(t.Context()))
+	assert.Equal(t, true, v2Call(t, restorer, testClient, "CreateCacheEntry", entry)["ok"])
+}
+
+func TestS3StorageSignsLikeAWS(t *testing.T) {
+	storage := &s3Storage{opts: S3Options{Region: "us-east-1", AccessKey: "AKIAIOSFODNN7EXAMPLE", SecretKey: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY"}}
+	req := httptest.NewRequest(http.MethodGet, "https://examplebucket.s3.amazonaws.com/?max-keys=2&prefix=J", nil)
+	req.Header.Set("X-Amz-Content-Sha256", "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855")
+	storage.sign(req, time.Date(2013, 5, 24, 0, 0, 0, 0, time.UTC))
+	assert.Equal(t, "AWS4-HMAC-SHA256 Credential=AKIAIOSFODNN7EXAMPLE/20130524/us-east-1/s3/aws4_request,SignedHeaders=host;x-amz-content-sha256;x-amz-date,Signature=34b48302e7b5fa45bde8084f4b7868a86f0a534bc59db6670ed5711ef69dc6f7", req.Header.Get("Authorization"))
+}

act/artifactcache/storage.go

@@ -5,8 +5,10 @@
 package artifactcache
 
 import (
+	"bytes"
 	"crypto/sha256"
 	"encoding/hex"
+	"encoding/json/v2"
 	"fmt"
 	"io"
 	"net/http"
@@ -66,6 +68,18 @@ func (s *Storage) WriteBlock(id uint64, blockID string, reader io.Reader) error
 // rename pass is safe because a staged name always carries blockFilePrefix and a target name
 // never does, so no rename can collide with a block not yet moved.
 func (s *Storage) OrderBlocks(id uint64, blockIDs []string) error {
+	wanted, err := json.Marshal(blockIDs) // a block id is client-chosen, so it cannot be a separator
+	if err != nil {
+		return err
+	}
+	list := filepath.Join(s.tempDir(id), blockFilePrefix+"list")
+	if recorded, err := os.ReadFile(list); err == nil {
+		// A retry repeats the list it already sent; another one would reorder what is now staged.
+		if !bytes.Equal(recorded, wanted) {
+			return fmt.Errorf("cache %d was already assembled from a different block list", id)
+		}
+		return nil
+	}
 	for i, blockID := range blockIDs {
 		if err := os.Rename(s.blockName(id, blockID), s.tempName(id, int64(i))); err != nil {
 			if os.IsNotExist(err) {
@@ -74,7 +88,7 @@ func (s *Storage) OrderBlocks(id uint64, blockIDs []string) error {
 			return err
 		}
 	}
-	return nil
+	return os.WriteFile(list, wanted, 0o600)
 }
 
 func (s *Storage) Commit(id uint64, size int64) (int64, error) {

act/artifacts/server.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"io"
 	"io/fs"
+	"net"
 	"net/http"
 	"os"
 	"path/filepath"
@@ -220,14 +221,14 @@ func Serve(ctx context.Context, artifactPath, addr, port string) context.CancelF
 	downloads(router, artifactPath)
 
 	server := &http.Server{
-		Addr:              fmt.Sprintf("%s:%s", addr, port),
+		Addr:              net.JoinHostPort(addr, port),
 		ReadHeaderTimeout: 2 * time.Second,
 		Handler:           router,
 	}
 
 	// run server
 	go func() {
-		logger.Infof("Start server on http://%s:%s", addr, port)
+		logger.Infof("Start server on http://%s", server.Addr)
 		if err := server.ListenAndServe(); err != nil && err != http.ErrServerClosed {
 			logger.Fatal(err)
 		}

act/common/git/git.go

@@ -325,33 +325,6 @@ func CloneIfRequired(ctx context.Context, refName plumbing.ReferenceName, input
 	return r, false, nil
 }
 
-func gitOptions(token string) (fetchOptions git.FetchOptions, pullOptions git.PullOptions) {
-	fetchOptions.RefSpecs = []config.RefSpec{"refs/*:refs/*", "HEAD:refs/heads/HEAD"}
-	fetchOptions.Force = true
-	pullOptions.Force = true
-
-	if token != "" {
-		auth := &http.BasicAuth{
-			Username: "token",
-			Password: token,
-		}
-		fetchOptions.Auth = auth
-		pullOptions.Auth = auth
-	}
-
-	return fetchOptions, pullOptions
-}
-
-// staleRefreshErr reports why a failed refresh must abort: the resolve and
-// checkout that follow are local and succeed on a cancelled context, which
-// would hand back the cached revision as if it were fresh.
-func staleRefreshErr(ctx context.Context, err error) error {
-	if err == nil || errors.Is(err, git.NoErrAlreadyUpToDate) {
-		return nil
-	}
-	return ctx.Err()
-}
-
 // NewGitCloneExecutor creates an executor to clone git repos
 func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 	return func(ctx context.Context) error {
@@ -373,18 +346,21 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 
 		isOfflineMode := input.OfflineMode
 
-		// fetch latest changes
-		fetchOptions, pullOptions := gitOptions(input.Token)
-
-		if input.InsecureSkipTLS { // For Gitea
-			fetchOptions.InsecureSkipTLS = true
-			pullOptions.InsecureSkipTLS = true
+		fetchOptions := git.FetchOptions{
+			RefSpecs:        []config.RefSpec{"refs/*:refs/*", "HEAD:refs/heads/HEAD", "refs/heads/*:refs/remotes/origin/*"},
+			Force:           true,
+			InsecureSkipTLS: input.InsecureSkipTLS,
+		}
+		if input.Token != "" {
+			fetchOptions.Auth = &http.BasicAuth{
+				Username: "token",
+				Password: input.Token,
+			}
 		}
 
 		// Action clones only ever need the tip commit, so keep a shallow cache cheap on update at depth 1 regardless of its original depth
 		// Turning action_shallow_clone off does not convert an existing shallow cache; evict it for a full clone.
-		shallow := isShallow(r)
-		if shallow {
+		if isShallow(r) {
 			fetchOptions.Depth = 1
 			if spec, ok := shallowFetchRefSpec(r, input.Ref); ok {
 				fetchOptions.RefSpecs = []config.RefSpec{spec}
@@ -392,7 +368,9 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 		}
 
 		// A just-cloned ref is as current as a fetch would make it, and a commit hash never moves.
-		_, _, present := refRevision(r, input.Ref)
+		// TODO: revalidate a mutable ref with a conditional archive request instead, once every
+		// supported Gitea sends an ETag for them: https://github.com/go-gitea/gitea/pull/39289
+		_, present := refRevision(r, input.Ref)
 		refresh := !isOfflineMode && (!present || (reused && !plumbing.IsHash(input.Ref)))
 
 		if refresh {
@@ -415,7 +393,7 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 			}
 		}
 
-		rev, refType, _ := refRevision(r, input.Ref)
+		rev, _ = refRevision(r, input.Ref)
 
 		if hash, err = r.ResolveRevision(rev); err != nil {
 			logger.Errorf("Unable to resolve %s: %v", input.Ref, err)
@@ -427,47 +405,13 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 			return err
 		}
 
-		// If the hash resolved doesn't match the ref provided in a workflow then we're
-		// using a branch or tag ref, not a sha
-		//
-		// Repos on disk point to commit hashes, and need to checkout input.Ref before
-		// we try and pull down any changes
-		if hash.String() != input.Ref && refType == "branch" {
-			logger.Debugf("Provided ref is not a sha. Checking out branch before pulling changes")
-			sourceRef := plumbing.ReferenceName(path.Join("refs", "remotes", "origin", input.Ref))
-			if err = w.Checkout(&git.CheckoutOptions{
-				Branch: sourceRef,
-				Force:  true,
-			}); err != nil {
-				logger.Errorf("Unable to checkout %s: %v", sourceRef, err)
-				return err
-			}
-		}
-
 		reusedMsg := ""
-
-		switch {
-		case refresh && !shallow:
-			// In shallow mode the depth-limited fetch above already advanced the ref.
-			if err = w.PullContext(ctx, &pullOptions); err != nil && !errors.Is(err, git.NoErrAlreadyUpToDate) {
-				logger.Debugf("Unable to pull %s: %v", refName, err)
-			}
-			if err := staleRefreshErr(ctx, err); err != nil {
-				return err
-			}
-		case isOfflineMode && reused:
+		if isOfflineMode && reused {
 			reusedMsg = " (reused in offline mode)"
 		}
 
 		logger.Debugf("Cloned %s to %s%s", input.URL, input.Dir, reusedMsg)
 
-		if hash.String() != input.Ref && refType == "branch" {
-			logger.Debugf("Provided ref is not a sha. Updating branch ref after pull")
-			if hash, err = r.ResolveRevision(rev); err != nil {
-				logger.Errorf("Unable to resolve %s: %v", input.Ref, err)
-				return err
-			}
-		}
 		if err = w.Checkout(&git.CheckoutOptions{
 			Hash:  *hash,
 			Force: true,
@@ -476,14 +420,6 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 			return err
 		}
 
-		if err = w.Reset(&git.ResetOptions{
-			Mode:   git.HardReset,
-			Commit: *hash,
-		}); err != nil {
-			logger.Errorf("Unable to reset to %s: %v", hash.String(), err)
-			return err
-		}
-
 		logger.Debugf("Checked out %s", input.Ref)
 		return nil
 	}
@@ -502,10 +438,11 @@ func cloneAtDepth(ctx context.Context, input NewGitCloneExecutorInput, opts git.
 				return nil, err
 			}
 		} else {
-			for _, refName := range []plumbing.ReferenceName{
-				plumbing.NewBranchReferenceName(input.Ref),
-				plumbing.NewTagReferenceName(input.Ref),
-			} {
+			refNames := []plumbing.ReferenceName{plumbing.NewBranchReferenceName(input.Ref), plumbing.NewTagReferenceName(input.Ref)}
+			if input.Ref == plumbing.HEAD.String() || strings.HasPrefix(input.Ref, "refs/") {
+				refNames = []plumbing.ReferenceName{plumbing.ReferenceName(input.Ref)} // HEAD is the remote's default branch
+			}
+			for _, refName := range refNames {
 				shallowOpts := opts
 				shallowOpts.Depth = input.Depth
 				shallowOpts.SingleBranch = true
@@ -562,22 +499,22 @@ func pinnedRefSpec(sha string) config.RefSpec {
 }
 
 // refRevision picks the revision to check out and reports whether it resolves locally.
-func refRevision(r *git.Repository, ref string) (plumbing.Revision, string, bool) {
+func refRevision(r *git.Repository, ref string) (plumbing.Revision, bool) {
 	if plumbing.IsHash(ref) {
 		// git ignores a ref named as 40 hex digits, so a full hash always denotes the commit itself.
 		_, err := r.CommitObject(plumbing.NewHash(ref))
-		return plumbing.Revision(ref), "sha", err == nil
+		return plumbing.Revision(ref), err == nil
 	}
 	if _, err := r.Tag(ref); err == nil {
-		return plumbing.Revision(path.Join("refs", "tags", ref)), "tag", true
+		return plumbing.Revision(path.Join("refs", "tags", ref)), true
 	}
 	remoteRef := plumbing.ReferenceName(path.Join("refs", "remotes", "origin", ref))
 	if _, err := r.Reference(remoteRef, false); err == nil {
-		return plumbing.Revision(remoteRef), "branch", true
+		return plumbing.Revision(remoteRef), true
 	}
 	rev := plumbing.Revision(ref)
 	_, err := r.ResolveRevision(rev)
-	return rev, "sha", err == nil
+	return rev, err == nil
 }
 
 // isShallow reports whether the local repository was cloned with a limited depth.

act/common/git/git_test.go

@@ -6,7 +6,6 @@ package git
 
 import (
 	"context"
-	"errors"
 	"fmt"
 	"net/http"
 	"net/http/httptest"
@@ -524,30 +523,51 @@ func TestGitCloneExecutorShallow(t *testing.T) {
 	})
 }
 
-func TestGitCloneExecutorColdCloneSkipsRefresh(t *testing.T) {
-	remoteDir := t.TempDir()
-	require.NoError(t, gitCmd("init", "--bare", "--initial-branch=main", remoteDir))
+func TestGitCloneExecutorTransportSessions(t *testing.T) {
 	workDir := t.TempDir()
-	require.NoError(t, gitCmd("clone", remoteDir, workDir))
-	require.NoError(t, gitCmd("-C", workDir, "checkout", "-b", "main"))
+	require.NoError(t, gitCmd("init", "--initial-branch=main", workDir))
 	require.NoError(t, gitCmd("-C", workDir, "commit", "--allow-empty", "-m", "c1"))
 	require.NoError(t, gitCmd("-C", workDir, "tag", "v1"))
-	require.NoError(t, gitCmd("-C", workDir, "push", "-u", "origin", "main"))
-	require.NoError(t, gitCmd("-C", workDir, "push", "origin", "v1"))
+	require.NoError(t, gitCmd("-C", workDir, "tag", "-a", "v2", "-m", "v2"))
 
 	for name, tt := range map[string]struct {
 		Ref   string
 		Depth int
 	}{
-		"shallow branch": {"main", 1},
-		"full clone tag": {"v1", 0},
+		"shallow branch":       {"main", 1},
+		"full clone branch":    {"main", 0},
+		"full lightweight tag": {"v1", 0},
+		"full annotated tag":   {"v2", 0},
 	} {
 		t.Run(name, func(t *testing.T) {
 			counter := installCountingTransport(t)
-			require.NoError(t, NewGitCloneExecutor(NewGitCloneExecutorInput{
-				URL: remoteDir, Ref: tt.Ref, Dir: t.TempDir(), Depth: tt.Depth,
-			})(t.Context()))
-			assert.Equal(t, int64(1), counter.sessions.Load())
+			dir := t.TempDir()
+			clone := NewGitCloneExecutor(NewGitCloneExecutorInput{
+				URL: workDir, Ref: tt.Ref, Dir: dir, Depth: tt.Depth,
+			})
+			require.NoError(t, clone(t.Context()))
+			assert.Equal(t, int64(1), counter.sessions.Swap(0), "cold clone")
+			assert.Equal(t, gitRevParse(t, workDir, tt.Ref+"^{commit}"), gitRevParse(t, dir, "HEAD"))
+
+			require.NoError(t, clone(t.Context()))
+			assert.Equal(t, int64(1), counter.sessions.Swap(0), "unchanged warm cache")
+
+			require.NoError(t, os.WriteFile(filepath.Join(workDir, "action.yml"), []byte(name), 0o644))
+			require.NoError(t, gitCmd("-C", workDir, "add", "action.yml"))
+			require.NoError(t, gitCmd("-C", workDir, "commit", "-m", name))
+			require.NoError(t, gitCmd("-C", workDir, "tag", "--force", "v1"))
+			require.NoError(t, gitCmd("-C", workDir, "tag", "--force", "-a", "v2", "-m", "v2"))
+
+			require.NoError(t, os.WriteFile(filepath.Join(dir, "action.yml"), []byte("staged"), 0o644))
+			require.NoError(t, gitCmd("-C", dir, "add", "action.yml"))
+			require.NoError(t, os.WriteFile(filepath.Join(dir, "action.yml"), []byte("unstaged"), 0o644))
+
+			require.NoError(t, clone(t.Context()))
+			assert.Equal(t, int64(1), counter.sessions.Load(), "updated warm cache")
+			assert.Equal(t, gitRevParse(t, workDir, "HEAD"), gitRevParse(t, dir, "HEAD"))
+			status, err := exec.Command("git", "-C", dir, "status", "--porcelain").Output()
+			require.NoError(t, err)
+			assert.Empty(t, string(status))
 		})
 	}
 }
@@ -736,12 +756,3 @@ func TestNewGitCloneExecutorFetchHonoursContext(t *testing.T) {
 		t.Fatal("fetch ignored context cancellation")
 	}
 }
-
-func TestStaleRefreshErr(t *testing.T) {
-	ctx, cancel := context.WithCancel(t.Context())
-	require.NoError(t, staleRefreshErr(ctx, errors.New("remote hung up")))
-
-	cancel()
-	require.ErrorIs(t, staleRefreshErr(ctx, errors.New("remote hung up")), context.Canceled)
-	require.NoError(t, staleRefreshErr(ctx, gogit.NoErrAlreadyUpToDate))
-}

act/container/container_types.go

@@ -9,6 +9,10 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"path"
+	"strings"
+	"sync"
+	"sync/atomic"
 
 	"gitea.com/gitea/runner/act/common"
 
@@ -77,13 +81,20 @@ var ErrContainerNotFound = errors.New("does not exist")
 
 // DockerProxy is a job's docker socket, fronting the daemon's for the job's lifetime.
 type DockerProxy struct {
-	Socket string
-	close  func(context.Context) error
+	Socket    string
+	close     func(context.Context) error
+	closeOnce sync.Once
+	closeErr  error
+	mounts    atomic.Value
+}
+
+func (p *DockerProxy) SetMounts(mounts map[string]string) {
+	p.mounts.Store(mounts)
 }
 
-// Close removes what the job created through the socket, then stops serving it.
 func (p *DockerProxy) Close(ctx context.Context) error {
-	return p.close(ctx)
+	p.closeOnce.Do(func() { p.closeErr = p.close(ctx) })
+	return p.closeErr
 }
 
 // Info is a snapshot of a container, as of one inspect.
@@ -98,6 +109,16 @@ type Info struct {
 	Mounts       map[string]string // container path to its source on the daemon
 }
 
+// DaemonPath maps target to its daemon path through the nearest mount with a source, empty without one.
+func (info *Info) DaemonPath(target string) string {
+	for dir := target; dir != "/" && dir != "."; dir = path.Dir(dir) {
+		if source := info.Mounts[dir]; source != "" {
+			return path.Join(source, strings.TrimPrefix(target, dir))
+		}
+	}
+	return ""
+}
+
 // Container for managing docker run containers
 type Container interface {
 	Create(capAdd, capDrop []string) common.Executor

act/container/docker_cli.go

@@ -4,9 +4,10 @@
 
 //go:build !(WITHOUT_DOCKER || !(linux || darwin || windows || netbsd))
 
-// This file is exact copy of https://github.com/docker/cli/blob/dfc4efb1e2ab8c06d70d2a1366ad448d2f917e90/cli/command/container/opts.go with:
+// This file is exact copy of https://github.com/docker/cli/blob/4a63305d74332de5ceba7fcbccbc3cbb7412f5ba/cli/command/container/opts.go with:
 // * appended with license information
 // * regexp and loader.ParseVolume in place of the import-restricted internal/lazyregexp and internal/volumespec
+// * jsontext in place of json.Compact for the seccomp profile
 // * invalidParameter from the package's errors.go, and convertPortSet/convertPortMap for the callers in docker_run.go
 //
 // docker/cli is licensed under the Apache License, Version 2.0.
@@ -152,6 +153,7 @@ type containerOptions struct {
 	runtime             string
 	autoRemove          bool
 	init                bool
+	umask               opts.UmaskOpt
 	annotations         *opts.MapOpts
 
 	Image string
@@ -326,6 +328,9 @@ func addFlags(flags *pflag.FlagSet) *containerOptions {
 	flags.Var(&copts.shmSize, "shm-size", "Size of /dev/shm")
 	flags.StringVar(&copts.utsMode, "uts", "", "UTS namespace to use")
 	flags.StringVar(&copts.runtime, "runtime", "", "Runtime to use for this container")
+	flags.Var(&copts.umask, "umask", "Set umask for the container")
+	flags.SetAnnotation("umask", "version", []string{"1.56"})
+	flags.SetAnnotation("umask", "ostype", []string{"linux"})
 
 	flags.BoolVar(&copts.init, "init", false, "Run an init inside the container that forwards signals and reaps processes")
 	flags.SetAnnotation("init", "version", []string{"1.25"})
@@ -721,6 +726,7 @@ func parse(flags *pflag.FlagSet, copts *containerOptions, serverOS string) (*con
 		MaskedPaths:    maskedPaths,
 		ReadonlyPaths:  readonlyPaths,
 		Annotations:    copts.annotations.GetAll(),
+		Umask:          copts.umask.Value(),
 	}
 
 	if copts.autoRemove && !hostConfig.RestartPolicy.IsNone() {

act/container/docker_cli_test.go

@@ -2,7 +2,7 @@
 // Copyright 2022 The nektos/act Authors. All rights reserved.
 // SPDX-License-Identifier: MIT
 
-// This file is exact copy of https://github.com/docker/cli/blob/dfc4efb1e2ab8c06d70d2a1366ad448d2f917e90/cli/command/container/opts_test.go with:
+// This file is exact copy of https://github.com/docker/cli/blob/4a63305d74332de5ceba7fcbccbc3cbb7412f5ba/cli/command/container/opts_test.go with:
 // * appended with license information
 // * added tests for the locally changed parseDevice, validateDevice and invalidParameter
 //
@@ -167,6 +167,19 @@ func TestParseRunLinks(t *testing.T) {
 	}
 }
 
+func TestParseRunWithoutUmask(t *testing.T) {
+	_, hostConfig, _, err := parseRun([]string{"ubuntu", "bash"})
+	assert.NilError(t, err)
+	assert.Assert(t, hostConfig.Umask == nil)
+}
+
+func TestParseRunUmask(t *testing.T) {
+	_, hostConfig, _, err := parseRun([]string{"--umask", "0022", "ubuntu", "bash"})
+	assert.NilError(t, err)
+	assert.Assert(t, hostConfig.Umask != nil)
+	assert.Equal(t, uint32(0o22), *hostConfig.Umask)
+}
+
 func TestParseRunAttach(t *testing.T) {
 	tests := []struct {
 		input    string

act/container/docker_network.go

@@ -10,11 +10,14 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"net/netip"
 	"strings"
 	"time"
 
 	"gitea.com/gitea/runner/act/common"
 
+	cerrdefs "github.com/containerd/errdefs"
+	"github.com/moby/moby/api/types/network"
 	"github.com/moby/moby/client"
 )
 
@@ -32,17 +35,17 @@ const (
 // had an endpoint on it at the time. Each one holds a subnet of the daemon's address pool
 // until it is removed. Networks created after createdBefore are left alone, so a job starting
 // while this runs cannot lose the network it has created but not yet attached a container to.
-func RemoveOrphanNetworks(ctx context.Context, runnerUUID string, createdBefore time.Time) error {
+func RemoveOrphanNetworks(ctx context.Context, runnerUUID, detach string, createdBefore time.Time) error {
 	cli, err := GetDockerClient(ctx)
 	if err != nil {
 		return fmt.Errorf("failed to connect to the docker daemon: %w", err)
 	}
 	defer cli.Close()
 
-	return removeOrphanNetworks(ctx, cli, runnerUUID, createdBefore)
+	return removeOrphanNetworks(ctx, cli, runnerUUID, detach, createdBefore)
 }
 
-func removeOrphanNetworks(ctx context.Context, cli client.APIClient, runnerUUID string, createdBefore time.Time) error {
+func removeOrphanNetworks(ctx context.Context, cli client.APIClient, runnerUUID, detach string, createdBefore time.Time) error {
 	networks, err := cli.NetworkList(ctx, client.NetworkListOptions{
 		Filters: make(client.Filters).Add("label", runnerUUIDLabel+"="+runnerUUID),
 	})
@@ -51,6 +54,7 @@ func removeOrphanNetworks(ctx context.Context, cli client.APIClient, runnerUUID
 	}
 
 	var errs []error
+networks:
 	for _, n := range networks.Items {
 		result, err := cli.NetworkInspect(ctx, n.ID, client.NetworkInspectOptions{})
 		if err != nil {
@@ -59,9 +63,18 @@ func removeOrphanNetworks(ctx context.Context, cli client.APIClient, runnerUUID
 		}
 		// the emptiness check, not the label, is what keeps a live job of another process
 		// sharing this registration safe
-		if len(result.Network.Containers) != 0 || result.Network.Created.After(createdBefore) {
+		if len(result.Network.Containers) > 1 || result.Network.Created.After(createdBefore) {
 			continue
 		}
+		for containerID := range result.Network.Containers {
+			if detach == "" || !strings.HasPrefix(containerID, detach) {
+				continue networks
+			}
+			if err := disconnectNetwork(ctx, cli, n.ID, containerID); err != nil {
+				errs = append(errs, fmt.Errorf("failed to disconnect container from network %s: %w", n.Name, err))
+				continue networks
+			}
+		}
 		if _, err := cli.NetworkRemove(ctx, n.ID, client.NetworkRemoveOptions{}); err != nil {
 			errs = append(errs, fmt.Errorf("failed to remove network %s: %w", n.Name, err))
 			continue
@@ -134,6 +147,50 @@ func isAddressPoolExhausted(err error) bool {
 		strings.Contains(msg, "could not find an available, non-overlapping IPv4 address pool among the defaults") // docker 24 and older
 }
 
+func NewDockerNetworkConnectExecutor(networkName, container string) common.Executor {
+	return func(ctx context.Context) error {
+		cli, err := GetDockerClient(ctx)
+		if err != nil {
+			return err
+		}
+		defer cli.Close()
+		if err := connectNetwork(ctx, cli, networkName, container); err != nil {
+			common.Logger(ctx).Warnf("cannot connect %s to network %s, so jobs on it cannot reach the cache: %v", container, networkName, err)
+		}
+		return nil
+	}
+}
+
+func connectNetwork(ctx context.Context, cli client.APIClient, networkName, container string) error {
+	_, err := cli.NetworkConnect(ctx, networkName, client.NetworkConnectOptions{
+		Container:      container,
+		EndpointConfig: &network.EndpointSettings{Aliases: []string{container}},
+	})
+	if err != nil && (strings.Contains(err.Error(), "already exists in network") || strings.Contains(err.Error(), "already connected")) {
+		return nil
+	}
+	return err
+}
+
+func NewDockerNetworkDisconnectExecutor(networkName, container string) common.Executor {
+	return func(ctx context.Context) error {
+		cli, err := GetDockerClient(ctx)
+		if err != nil {
+			return err
+		}
+		defer cli.Close()
+		return disconnectNetwork(ctx, cli, networkName, container)
+	}
+}
+
+func disconnectNetwork(ctx context.Context, cli client.APIClient, networkName, container string) error {
+	_, err := cli.NetworkDisconnect(ctx, networkName, client.NetworkDisconnectOptions{Container: container, Force: true})
+	if cerrdefs.IsNotFound(err) || (err != nil && strings.Contains(err.Error(), "is not connected")) {
+		return nil
+	}
+	return err
+}
+
 func NewDockerNetworkRemoveExecutor(name string) common.Executor {
 	return func(ctx context.Context) error {
 		cli, err := GetDockerClient(ctx)
@@ -172,3 +229,49 @@ func NewDockerNetworkRemoveExecutor(name string) common.Executor {
 		return errors.Join(errs...)
 	}
 }
+
+// IsolatedCacheContainer returns the container holding addr, when jobs on jobNetwork cannot reach it there.
+func IsolatedCacheContainer(ctx context.Context, addr netip.Addr, jobNetwork string) (string, error) {
+	cli, err := GetDockerClient(ctx)
+	if err != nil {
+		return "", err
+	}
+	defer cli.Close()
+	return isolatedCacheContainer(ctx, cli, addr, jobNetwork)
+}
+
+func isolatedCacheContainer(ctx context.Context, cli client.APIClient, addr netip.Addr, jobNetwork string) (string, error) {
+	if jobNetwork == "host" || strings.HasPrefix(jobNetwork, "container:") {
+		return "", nil
+	}
+	containers, err := cli.ContainerList(ctx, client.ContainerListOptions{})
+	if err != nil {
+		return "", err
+	}
+	var holderID, holderContainer string
+	for _, summary := range containers.Items {
+		if summary.NetworkSettings == nil {
+			continue
+		}
+		for _, endpoint := range summary.NetworkSettings.Networks {
+			if endpoint != nil && (endpoint.IPAddress == addr || endpoint.GlobalIPv6Address == addr) {
+				holderID = endpoint.NetworkID
+				holderContainer = summary.ID[:12]
+			}
+		}
+	}
+	if holderID == "" {
+		return "", nil
+	}
+	holder, err := cli.NetworkInspect(ctx, holderID, client.NetworkInspectOptions{})
+	if err != nil || holder.Network.Driver != "bridge" {
+		return "", err
+	}
+	if jobNetwork != "" {
+		job, err := cli.NetworkInspect(ctx, jobNetwork, client.NetworkInspectOptions{})
+		if err != nil || job.Network.ID == holderID {
+			return "", err
+		}
+	}
+	return holderContainer, nil
+}

act/container/docker_network_test.go

@@ -6,25 +6,95 @@ package container
 import (
 	"context"
 	"errors"
+	"net/netip"
 	"testing"
 	"time"
 
 	cerrdefs "github.com/containerd/errdefs"
+	"github.com/moby/moby/api/types/container"
 	"github.com/moby/moby/api/types/network"
 	mobyclient "github.com/moby/moby/client"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
 
+func TestIsolatedCacheContainer(t *testing.T) {
+	ctx := context.Background()
+	cache, lan := netip.MustParseAddr("172.18.0.3"), netip.MustParseAddr("192.168.1.20")
+	client := &mockDockerClient{}
+	client.On("ContainerList", ctx, mobyclient.ContainerListOptions{}).Return(mobyclient.ContainerListResult{Items: []container.Summary{{}, {
+		ID: "0123456789abcdef",
+		NetworkSettings: &container.NetworkSettingsSummary{Networks: map[string]*network.EndpointSettings{
+			"compose": {NetworkID: "c0ffee", IPAddress: cache},
+			"lan":     {NetworkID: "beef", IPAddress: lan},
+		}},
+	}}}, nil)
+	for ref, inspected := range map[string]network.Network{
+		"c0ffee": {ID: "c0ffee", Name: "compose", Driver: "bridge"},
+		"beef":   {ID: "beef", Name: "lan", Driver: "macvlan"},
+		"c0f":    {ID: "c0ffee"},
+		"jobs":   {ID: "f00d"},
+	} {
+		client.On("NetworkInspect", ctx, ref, mobyclient.NetworkInspectOptions{}).
+			Return(mobyclient.NetworkInspectResult{Network: network.Inspect{Network: inspected}}, nil)
+	}
+	check := func(addr netip.Addr, jobNetwork, want string) {
+		got, err := isolatedCacheContainer(ctx, client, addr, jobNetwork)
+		require.NoError(t, err)
+		assert.Equal(t, want, got, jobNetwork)
+	}
+	check(cache, "", "0123456789ab")
+	check(cache, "jobs", "0123456789ab")
+	check(cache, "c0f", "")
+	check(cache, "host", "")
+	check(lan, "", "")
+	check(netip.MustParseAddr("10.0.0.5"), "", "")
+}
+
 func TestIsAddressPoolExhausted(t *testing.T) {
 	assert.True(t, isAddressPoolExhausted(cerrdefs.ErrInvalidArgument.WithMessage("Error response from daemon: all predefined address pools have been fully subnetted")))
 	assert.True(t, isAddressPoolExhausted(errors.New("could not find an available, non-overlapping IPv4 address pool among the defaults to assign to the network")))
 	assert.False(t, isAddressPoolExhausted(cerrdefs.ErrInvalidArgument.WithMessage("invalid subnet 10.0.0.0/8: it overlaps with an existing network")))
 }
 
-// Of this runner's networks, only the ones nothing is attached to and old enough to predate
-// any job now starting are the runner's to reclaim. An unexpected NetworkRemove fails the
-// test on its own, since testify has no expectation to match it against.
+func (m *mockDockerClient) NetworkConnect(ctx context.Context, id string, opts mobyclient.NetworkConnectOptions) (mobyclient.NetworkConnectResult, error) {
+	return mobyclient.NetworkConnectResult{}, m.Called(ctx, id, opts).Error(0)
+}
+
+func (m *mockDockerClient) NetworkDisconnect(ctx context.Context, id string, opts mobyclient.NetworkDisconnectOptions) (mobyclient.NetworkDisconnectResult, error) {
+	return mobyclient.NetworkDisconnectResult{}, m.Called(ctx, id, opts).Error(0)
+}
+
+func TestConnectAndDisconnectNetwork(t *testing.T) {
+	ctx := context.Background()
+	connect := func(answer error) error {
+		client := &mockDockerClient{}
+		client.On("NetworkConnect", ctx, "jobs", mobyclient.NetworkConnectOptions{
+			Container:      "0123456789ab",
+			EndpointConfig: &network.EndpointSettings{Aliases: []string{"0123456789ab"}},
+		}).Return(answer).Once()
+		return connectNetwork(ctx, client, "jobs", "0123456789ab")
+	}
+	require.NoError(t, connect(nil))
+	require.NoError(t, connect(errors.New("endpoint with name runner already exists in network jobs")))
+	require.NoError(t, connect(errors.New("container runner already connected to network jobs")))
+	require.ErrorIs(t, connect(context.Canceled), context.Canceled)
+
+	disconnect := func(answer error) error {
+		client := &mockDockerClient{}
+		client.On("NetworkDisconnect", ctx, "jobs", mobyclient.NetworkDisconnectOptions{
+			Container: "0123456789ab", Force: true,
+		}).Return(answer).Once()
+		return disconnectNetwork(ctx, client, "jobs", "0123456789ab")
+	}
+	require.NoError(t, disconnect(cerrdefs.ErrNotFound))
+	require.NoError(t, disconnect(errors.New("container runner is not connected to the network jobs")))
+	require.ErrorIs(t, disconnect(context.Canceled), context.Canceled)
+}
+
+// Of this runner's networks, only the ones nothing but its own cache container is attached to
+// and old enough to predate any job now starting are the runner's to reclaim. An unexpected
+// NetworkRemove fails the test on its own, since testify has no expectation to match it against.
 func TestRemoveOrphanNetworks(t *testing.T) {
 	ctx := context.Background()
 	cutoff := time.Date(2026, time.April, 29, 20, 0, 0, 0, time.UTC)
@@ -35,16 +105,29 @@ func TestRemoveOrphanNetworks(t *testing.T) {
 		{ID: "orphan"},
 		{ID: "busy"},
 		{ID: "starting"},
+		{ID: "cached"},
+		{ID: "shared"},
 	}}, nil)
 	client.On("NetworkInspect", ctx, "orphan", mobyclient.NetworkInspectOptions{}).
 		Return(mobyclient.NetworkInspectResult{}, nil)
 	client.On("NetworkInspect", ctx, "busy", mobyclient.NetworkInspectOptions{}).
 		Return(mobyclient.NetworkInspectResult{Network: network.Inspect{Containers: map[string]network.EndpointResource{"c": {}}}}, nil)
 	client.On("NetworkInspect", ctx, "starting", mobyclient.NetworkInspectOptions{}).
-		Return(mobyclient.NetworkInspectResult{Network: network.Inspect{Network: network.Network{Created: cutoff.Add(time.Second)}}}, nil)
+		Return(mobyclient.NetworkInspectResult{Network: network.Inspect{
+			Network:    network.Network{Created: cutoff.Add(time.Second)},
+			Containers: map[string]network.EndpointResource{"0123456789abcdef": {}},
+		}}, nil)
+	client.On("NetworkInspect", ctx, "cached", mobyclient.NetworkInspectOptions{}).
+		Return(mobyclient.NetworkInspectResult{Network: network.Inspect{Containers: map[string]network.EndpointResource{"0123456789abcdef": {}}}}, nil)
+	client.On("NetworkInspect", ctx, "shared", mobyclient.NetworkInspectOptions{}).
+		Return(mobyclient.NetworkInspectResult{Network: network.Inspect{Containers: map[string]network.EndpointResource{"0123456789abcdef": {}, "job": {}}}}, nil)
 	client.On("NetworkRemove", ctx, "orphan", mobyclient.NetworkRemoveOptions{}).
 		Return(mobyclient.NetworkRemoveResult{}, nil)
+	client.On("NetworkDisconnect", ctx, "cached", mobyclient.NetworkDisconnectOptions{Container: "0123456789abcdef", Force: true}).
+		Return(nil).Once()
+	client.On("NetworkRemove", ctx, "cached", mobyclient.NetworkRemoveOptions{}).
+		Return(mobyclient.NetworkRemoveResult{}, nil).Once()
 
-	require.NoError(t, removeOrphanNetworks(ctx, client, "runner-1", cutoff))
+	require.NoError(t, removeOrphanNetworks(ctx, client, "runner-1", "0123456789ab", cutoff))
 	client.AssertExpectations(t)
 }

act/container/docker_proxy_test.go

@@ -10,17 +10,23 @@ import (
 	"bufio"
 	"bytes"
 	"context"
-	"encoding/json"
+	"errors"
+	"fmt"
 	"io"
 	"net"
 	"net/http"
+	"os"
 	"path/filepath"
+	"runtime"
+	"strconv"
 	"strings"
 	"testing"
 	"time"
 
 	cerrdefs "github.com/containerd/errdefs"
 	"github.com/moby/moby/api/types/container"
+	"github.com/moby/moby/api/types/image"
+	"github.com/moby/moby/api/types/mount"
 	"github.com/moby/moby/api/types/network"
 	"github.com/moby/moby/api/types/volume"
 	mobyclient "github.com/moby/moby/client"
@@ -56,7 +62,12 @@ func daemonSocketPath(t testing.TB, cli mobyclient.APIClient) string {
 }
 
 func TestDockerProxy(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("Unix socket ownership is unavailable on Windows")
+	}
 	bodies := make(chan []byte, 8)
+	createStarted, releaseCreate := make(chan struct{}), make(chan struct{})
+	streamsDone := make(chan struct{}, 2)
 	daemonSocket := filepath.Join(shortTempDir(t), "d.sock")
 	listener, err := net.Listen("unix", daemonSocket)
 	require.NoError(t, err)
@@ -64,16 +75,33 @@ func TestDockerProxy(t *testing.T) {
 		switch {
 		case strings.HasSuffix(r.URL.Path, "/create"):
 			body, _ := io.ReadAll(r.Body)
+			if r.URL.RawQuery == "wait" {
+				close(createStarted)
+				<-releaseCreate
+			}
 			bodies <- body
 			w.WriteHeader(http.StatusCreated)
-		case strings.HasSuffix(r.URL.Path, "/start"):
+		case r.URL.Path == "/exec/detached/start" || r.URL.Path == "/exec/error/start":
+			if r.URL.Path == "/exec/error/start" {
+				w.Header().Set("Content-Type", "application/vnd.docker.raw-stream")
+				w.WriteHeader(http.StatusBadRequest)
+			}
+			_, _ = w.Write([]byte("ordinary"))
+		case rawStreamPath.MatchString(r.URL.Path) || r.URL.Path == "/session":
+			_, _ = io.Copy(io.Discard, r.Body)
 			conn, buffered, err := w.(http.Hijacker).Hijack()
 			if err != nil {
 				return
 			}
 			defer conn.Close()
-			_, _ = conn.Write([]byte("HTTP/1.1 200 OK\r\nContent-Type: application/vnd.docker.raw-stream\r\n\r\n"))
-			_, _ = io.Copy(conn, buffered)
+			if r.Header.Get("Upgrade") != "" {
+				_, _ = conn.Write([]byte("HTTP/1.1 101 Switching Protocols\r\nConnection: Upgrade\r\nUpgrade: tcp\r\n\r\nready\n"))
+			} else {
+				_, _ = conn.Write([]byte("HTTP/1.1 200 OK\r\nContent-Type: application/vnd.docker.raw-stream\r\n\r\nready\n"))
+			}
+			input, _ := io.ReadAll(buffered)
+			_, _ = conn.Write(append([]byte("final:"), input...))
+			streamsDone <- struct{}{}
 		default:
 			w.Header().Set("Api-Version", "1.47")
 			_, _ = w.Write([]byte("OK " + r.Method + " " + r.URL.Path))
@@ -83,34 +111,94 @@ func TestDockerProxy(t *testing.T) {
 	t.Cleanup(func() { _ = daemon.Close() })
 	proxy, err := StartDockerProxy(daemonSocket, shortTempDir(t), "job-1")
 	require.NoError(t, err)
-	client := &http.Client{Transport: &http.Transport{DialContext: func(ctx context.Context, _, _ string) (net.Conn, error) {
+	t.Cleanup(func() { _ = proxy.Close(context.Background()) })
+	proxy.SetMounts(map[string]string{"/workspace/o/r": "/volumes/job/_data", "/workspace/o/r/tmp": "", "/var/run/docker.sock": "/tmp/p/docker.sock", "/volumes": "/daemon/volumes"})
+	client := &http.Client{Timeout: 5 * time.Second, Transport: &http.Transport{DialContext: func(ctx context.Context, _, _ string) (net.Conn, error) {
 		return (&net.Dialer{}).DialContext(ctx, "unix", proxy.Socket)
 	}}}
+	t.Cleanup(client.CloseIdleConnections)
+	dialProxy := func(t *testing.T) *net.UnixConn {
+		t.Helper()
+		conn, err := net.DialUnix("unix", nil, &net.UnixAddr{Name: proxy.Socket, Net: "unix"})
+		require.NoError(t, err)
+		t.Cleanup(func() { _ = conn.Close() })
+		require.NoError(t, conn.SetDeadline(time.Now().Add(5*time.Second)))
+		return conn
+	}
+	startStream := func(t *testing.T, path, upgrade string) (*net.UnixConn, *bufio.Reader) {
+		t.Helper()
+		conn := dialProxy(t)
+		_, err := fmt.Fprintf(conn, "POST %s HTTP/1.1\r\nHost: docker\r\n%sContent-Length: 2\r\n\r\n{}input", path, upgrade)
+		require.NoError(t, err)
+		reader := bufio.NewReader(conn)
+		resp, err := http.ReadResponse(reader, nil)
+		require.NoError(t, err)
+		t.Cleanup(func() {
+			_ = conn.Close()
+			_ = resp.Body.Close()
+		})
+		prefix, err := reader.ReadString('\n')
+		require.NoError(t, err)
+		require.Equal(t, "ready\n", prefix)
+		return conn, reader
+	}
 
 	t.Run("labels creates", func(t *testing.T) {
-		for path, body := range map[string]string{
-			"/v1.47/containers/create": `{"Image":"alpine","Labels":{"own":"1"}}`,
-			"/networks/create":         `{"Name":"n"}`,
-			"/volumes/create":          `{"Name":"v","labels":null}`,
+		for _, testCase := range []struct{ path, body, want string }{
+			{
+				"/v1.47/containers/create", `{"Image":"alpine","Unknown":{"enabled":true},"Labels":{"own":"1","com.gitea.runner.job":"other"}}`,
+				`{"Image":"alpine","Unknown":{"enabled":true},"Labels":{"own":"1","com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/containers/create", `{"HostConfig":{"Binds":["/workspace/o/r:/src:ro","/workspace/o/r/../r/sub/:/sub","workspace/o/r:/relative","/workspace/o/rest:/rest","/workspace/o/r/tmp/x:/tmp","named:/named","/anonymous","/workspace/o/r:ro","/volumes/job/_data/x:/daemon-path"]},"DriverOpts":{"o":"bind","device":"/workspace/o/r"},"Labels":{"Foo":"1","foo":"2"}}`,
+				`{"HostConfig":{"Binds":["/volumes/job/_data:/src:ro","/volumes/job/_data/sub:/sub","workspace/o/r:/relative","/workspace/o/rest:/rest","/workspace/o/r/tmp/x:/tmp","named:/named","/anonymous","/workspace/o/r:ro","/volumes/job/_data/x:/daemon-path"]},"DriverOpts":{"o":"bind","device":"/workspace/o/r"},"Labels":{"Foo":"1","foo":"2","com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/containers/create", `{"HostConfig":{"Mounts":[{"Type":"bind","Source":"/var/run/docker.sock","Target":"/var/run/docker.sock"},{"type":"bind","source":"/workspace/o/r/sub/","target":"/m"},{"Type":"bind","Source":"/workspace/o/r/sub/../data","Target":"/dotted"},{"Type":"bind","Source":"/workspace/o/r/../r","Target":"/escaping"},{"Type":"BIND","Source":"/workspace/o/r"},{"Type":"volume","Source":"/workspace/o/r"}]}}`,
+				`{"HostConfig":{"Mounts":[{"Type":"bind","Source":"/tmp/p/docker.sock","Target":"/var/run/docker.sock"},{"Type":"bind","Source":"/volumes/job/_data/sub/","target":"/m"},{"Type":"bind","Source":"/volumes/job/_data/sub/../data","Target":"/dotted"},{"Type":"bind","Source":"/volumes/job/_data","Target":"/escaping"},{"Type":"BIND","Source":"/workspace/o/r"},{"Type":"volume","Source":"/workspace/o/r"}]},"Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/containers/create", `{"HostConfig":{"Mounts":[{"Type":"volume","VolumeOptions":{"DriverConfig":{"Options":{"o":"bind","device":"/workspace/o/r/data"}}}},{"Type":"volume","VolumeOptions":{"DriverConfig":{"Options":{"o":"unbindable","device":"/workspace/o/r"}}}},{"Type":"volume","VolumeOptions":{"DriverConfig":{"Options":{"o":"bind,remount","device":"/workspace/o/r"}}}},{"Type":"volume","VolumeOptions":{"DriverConfig":{"Name":"plugin","Options":{"o":"bind","device":"/workspace/o/r"}}}}]}}`,
+				`{"HostConfig":{"Mounts":[{"Type":"volume","VolumeOptions":{"DriverConfig":{"Options":{"o":"bind","device":"/volumes/job/_data/data"}}}},{"Type":"volume","VolumeOptions":{"DriverConfig":{"Options":{"o":"unbindable","device":"/workspace/o/r"}}}},{"Type":"volume","VolumeOptions":{"DriverConfig":{"Options":{"o":"bind,remount","device":"/workspace/o/r"}}}},{"Type":"volume","VolumeOptions":{"DriverConfig":{"Name":"plugin","Options":{"o":"bind","device":"/workspace/o/r"}}}}]},"Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/containers/create", `{"HostConfig":{"privileged":true,"Privileged":false,"Binds":["/workspace/o/r:/src"]},"hoſtconfig":null}`,
+				`{"HostConfig":{"privileged":true,"Privileged":false,"Binds":["/workspace/o/r:/src"]},"hoſtconfig":null,"Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/networks/create", `{"Name":"n"}`,
+				`{"Name":"n","Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/volumes/create", `{"Name":"v","labels":null,"DriverOpts":{"type":"none","o":"rbind,ro","device":"/workspace/o/r/"},"HostConfig":{"Binds":["/workspace/o/r:/src"]}}`,
+				`{"Name":"v","Labels":{"com.gitea.runner.job":"job-1"},"DriverOpts":{"type":"none","o":"rbind,ro","device":"/volumes/job/_data/"},"HostConfig":{"Binds":["/workspace/o/r:/src"]}}`,
+			},
+			{
+				"/volumes/create", `{"Name":"d","DriverOpts":{"o":"bind","device":1,"device":"/workspace/o/r"}}`,
+				`{"Name":"d","DriverOpts":{"o":"bind","device":1,"device":"/workspace/o/r"},"Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/volumes/create", `{"Name":"p","Driver":"plugin","DriverOpts":{"o":"bind","device":"/workspace/o/r"}}`,
+				`{"Name":"p","Driver":"plugin","DriverOpts":{"o":"bind","device":"/workspace/o/r"},"Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/volumes/create", "",
+				`{"Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/volumes/create", "null",
+				`{"Labels":{"com.gitea.runner.job":"job-1"}}`,
+			},
+			{
+				"/volumes/create", `{"Labels":{"discard":"1"},"labels":null,"LABELS":{"own":"1"},"LABELS":{"extra":"2"}}`,
+				`{"Labels":{"own":"1","extra":"2","com.gitea.runner.job":"job-1"}}`,
+			},
 		} {
-			resp, err := client.Post("http://docker"+path, "application/json", strings.NewReader(body))
+			resp, err := client.Post("http://docker"+testCase.path, "application/json", strings.NewReader(testCase.body))
 			require.NoError(t, err)
 			resp.Body.Close()
-			assert.Equal(t, http.StatusCreated, resp.StatusCode)
-
-			var got map[string]json.RawMessage
-			require.NoError(t, json.Unmarshal(<-bodies, &got))
-			key := "Labels"
-			if _, ok := got["labels"]; ok {
-				key = "labels"
-			}
-			var labels map[string]string
-			require.NoError(t, json.Unmarshal(got[key], &labels))
-			assert.Equal(t, "job-1", labels[jobLabel], path)
-			if strings.Contains(body, "own") {
-				assert.Equal(t, "1", labels["own"])
-				assert.JSONEq(t, `"alpine"`, string(got["Image"]))
-			}
+			require.Equal(t, http.StatusCreated, resp.StatusCode)
+			assert.JSONEq(t, testCase.want, string(<-bodies), testCase.body)
 		}
 	})
 
@@ -123,40 +211,102 @@ func TestDockerProxy(t *testing.T) {
 		assert.Equal(t, "OK GET /v1.47/_ping", string(body))
 	})
 
-	t.Run("tunnels raw streams", func(t *testing.T) {
-		conn, err := net.Dial("unix", proxy.Socket)
-		require.NoError(t, err)
-		defer conn.Close()
-		_, err = conn.Write([]byte("POST /v1.47/exec/abc/start HTTP/1.1\r\nHost: docker\r\nContent-Length: 2\r\n\r\n{}ping\n"))
-		require.NoError(t, err)
+	t.Run("tunnels raw streams through stdin EOF", func(t *testing.T) {
+		for _, stream := range []struct{ path, upgrade string }{
+			{path: "/v1.47/exec/abc/start"},
+			{path: "/containers/abc/attach", upgrade: "Connection: Upgrade\r\nUpgrade: tcp\r\n"},
+		} {
+			conn, reader := startStream(t, stream.path, stream.upgrade)
+			require.NoError(t, conn.CloseWrite())
+			output, err := io.ReadAll(reader)
+			require.NoError(t, err)
+			assert.Equal(t, "final:input", string(output))
+			<-streamsDone
+		}
+	})
+
+	t.Run("detached and error responses retain keepalive label injection", func(t *testing.T) {
+		conn := dialProxy(t)
 		reader := bufio.NewReader(conn)
-		resp, err := http.ReadResponse(reader, nil)
+		for endpoint, status := range map[string]int{"detached": http.StatusOK, "error": http.StatusBadRequest} {
+			_, err := fmt.Fprintf(conn, "POST /exec/%s/start HTTP/1.1\r\nHost: docker\r\nContent-Length: 2\r\n\r\n{}", endpoint)
+			require.NoError(t, err)
+			resp, err := http.ReadResponse(reader, nil)
+			require.NoError(t, err)
+			body, err := io.ReadAll(resp.Body)
+			require.NoError(t, err)
+			require.NoError(t, resp.Body.Close())
+			assert.Equal(t, status, resp.StatusCode)
+			assert.Equal(t, "ordinary", string(body))
+			_, err = io.WriteString(conn, "POST /volumes/create HTTP/1.1\r\nHost: docker\r\nContent-Length: 0\r\n\r\n")
+			require.NoError(t, err)
+			resp, err = http.ReadResponse(reader, nil)
+			require.NoError(t, err)
+			require.NoError(t, resp.Body.Close())
+			require.Equal(t, http.StatusCreated, resp.StatusCode)
+			assert.JSONEq(t, `{"Labels":{"com.gitea.runner.job":"job-1"}}`, string(<-bodies))
+		}
+	})
+
+	t.Run("close joins streams and admitted create", func(t *testing.T) {
+		t.Cleanup(func() { close(releaseCreate) })
+		_, raw := startStream(t, "/exec/live/start", "")
+		_, upgraded := startStream(t, "/session", "Connection: Upgrade\r\nUpgrade: tcp\r\n")
+		creator := dialProxy(t)
+		_, err := io.WriteString(creator, "POST /volumes/create?wait HTTP/1.1\r\nHost: docker\r\nContent-Length: 2\r\n\r\n{}")
 		require.NoError(t, err)
-		defer resp.Body.Close()
-		assert.Equal(t, "application/vnd.docker.raw-stream", resp.Header.Get("Content-Type"))
-		echoed, err := reader.ReadString('\n')
+		<-createStarted
+		require.NoError(t, creator.CloseWrite())
+		closed := make(chan error, 1)
+		ctx, cancel := context.WithTimeout(t.Context(), 5*time.Second)
+		defer cancel()
+		go func() { closed <- proxy.Close(ctx) }()
+		for _, reader := range []*bufio.Reader{raw, upgraded} {
+			_, err := reader.ReadByte()
+			require.ErrorIs(t, err, io.EOF)
+			<-streamsDone
+		}
+		select {
+		case err := <-closed:
+			t.Fatalf("Close returned before admitted create settled: %v", err)
+		default:
+		}
+		releaseCreate <- struct{}{}
+		resp, err := http.ReadResponse(bufio.NewReader(creator), nil)
 		require.NoError(t, err)
-		assert.Equal(t, "{}ping\n", echoed)
+		require.NoError(t, resp.Body.Close())
+		require.Equal(t, http.StatusCreated, resp.StatusCode)
+		assert.JSONEq(t, `{"Labels":{"com.gitea.runner.job":"job-1"}}`, string(<-bodies))
+		require.NoError(t, <-closed)
 	})
 }
 
 func TestRemoveLabelledRemovesContainersNetworksAndVolumes(t *testing.T) {
+	containerFailure := errors.New("container removal failed")
+	listFailure := errors.New("listing failed")
+	volumeFailure := errors.New("volume removal failed")
 	ctx := context.Background()
 	filters := make(mobyclient.Filters).Add("label", jobLabel+"=job-1")
 	cli := &mockDockerClient{}
 	cli.On("ContainerList", ctx, mobyclient.ContainerListOptions{All: true, Filters: filters}).
-		Return(mobyclient.ContainerListResult{Items: []container.Summary{{ID: "c1", Names: []string{"/app"}}}}, nil)
-	cli.On("ContainerKill", ctx, "c1", mock.Anything).Return(mobyclient.ContainerKillResult{}, nil)
+		Return(mobyclient.ContainerListResult{Items: []container.Summary{{ID: "c1", Names: []string{"/app"}}}}, nil).Once()
+	cli.On("ContainerKill", ctx, "c1", mock.Anything).Return(mobyclient.ContainerKillResult{}, nil).Once()
 	cli.On("ContainerRemove", ctx, "c1", mobyclient.ContainerRemoveOptions{RemoveVolumes: true, Force: true}).
-		Return(mobyclient.ContainerRemoveResult{}, nil)
+		Return(mobyclient.ContainerRemoveResult{}, containerFailure).Once()
 	cli.On("NetworkList", ctx, mobyclient.NetworkListOptions{Filters: filters}).
-		Return(mobyclient.NetworkListResult{Items: []network.Summary{{ID: "n1", Name: "app_default"}}}, nil)
-	cli.On("NetworkRemove", ctx, "n1", mock.Anything).Return(mobyclient.NetworkRemoveResult{}, nil)
+		Return(mobyclient.NetworkListResult{Items: []network.Summary{{ID: "n1", Name: "stack_default", Scope: "swarm"}}}, listFailure).Once()
+	cli.On("NetworkRemove", ctx, "n1", mobyclient.NetworkRemoveOptions{}).Return(mobyclient.NetworkRemoveResult{}, cerrdefs.ErrInvalidArgument).Once()
 	cli.On("VolumeList", ctx, mobyclient.VolumeListOptions{Filters: filters}).
-		Return(mobyclient.VolumeListResult{Items: []volume.Volume{{Name: "app_data"}}}, nil)
-	cli.On("VolumeRemove", ctx, "app_data", mobyclient.VolumeRemoveOptions{}).Return(mobyclient.VolumeRemoveResult{}, nil)
+		Return(mobyclient.VolumeListResult{Items: []volume.Volume{{Name: "app_data"}}}, nil).Once()
+	cli.On("VolumeRemove", ctx, "app_data", mobyclient.VolumeRemoveOptions{}).Return(mobyclient.VolumeRemoveResult{}, volumeFailure).Once()
 
-	require.NoError(t, removeLabelled(ctx, cli, "job-1"))
+	err := removeLabelled(ctx, cli, "job-1")
+	require.ErrorIs(t, err, containerFailure)
+	require.ErrorIs(t, err, listFailure)
+	require.ErrorIs(t, err, volumeFailure)
+	require.NotErrorIs(t, err, cerrdefs.ErrInvalidArgument)
+	require.ErrorContains(t, err, "failed to remove container c1")
+	require.ErrorContains(t, err, "failed to remove volume app_data")
 	cli.AssertExpectations(t)
 }
 
@@ -168,13 +318,17 @@ func TestDockerProxyWithDaemon(t *testing.T) {
 	require.NoError(t, err)
 	defer direct.Close()
 	dir := shortTempDir(t)
-	seen, err := daemonSeesDir(ctx, dir)
+	seen, err := daemonSeesDir(ctx, direct, dir, dir)
 	require.NoError(t, err)
 	t.Logf("daemon sees the runner's filesystem: %v", seen)
+	seen, err = daemonSeesDir(ctx, direct, dir, shortTempDir(t))
+	require.NoError(t, err)
+	assert.False(t, seen)
 
-	job := "proxy-test-" + t.Name()
+	job := "proxy-test-" + strconv.FormatInt(time.Now().UnixNano(), 36)
 	proxy, err := StartDockerProxy(daemonSocketPath(t, direct), dir, job)
 	require.NoError(t, err)
+	t.Cleanup(func() { _ = proxy.Close(context.Background()) })
 	viaProxy, err := mobyclient.New(mobyclient.WithHost("unix://" + proxy.Socket))
 	require.NoError(t, err)
 	defer viaProxy.Close()
@@ -212,6 +366,7 @@ func TestDockerProxyWithDaemon(t *testing.T) {
 	attached.Close()
 
 	require.NoError(t, proxy.Close(ctx))
+	require.NoError(t, RemoveDockerJobResources(ctx, job))
 	_, err = direct.ContainerInspect(ctx, created.ID, mobyclient.ContainerInspectOptions{})
 	assert.True(t, cerrdefs.IsNotFound(err))
 	_, err = direct.NetworkInspect(ctx, net.ID, mobyclient.NetworkInspectOptions{})
@@ -268,3 +423,121 @@ func BenchmarkDockerProxy(b *testing.B) {
 		})
 	}
 }
+
+type probeClient struct {
+	mobyclient.APIClient
+	create func(mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error)
+	remove func(context.Context, string, mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error)
+	stat   func(string) error
+	mounts []container.MountPoint
+}
+
+func (c *probeClient) ContainerInspect(context.Context, string, mobyclient.ContainerInspectOptions) (mobyclient.ContainerInspectResult, error) {
+	return mobyclient.ContainerInspectResult{Container: container.InspectResponse{Mounts: c.mounts}}, nil
+}
+
+func (c *probeClient) ImageList(context.Context, mobyclient.ImageListOptions) (mobyclient.ImageListResult, error) {
+	return mobyclient.ImageListResult{Items: []image.Summary{{ID: "probe-image"}}}, nil
+}
+
+func (c *probeClient) ContainerCreate(_ context.Context, opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+	return c.create(opts)
+}
+
+func (c *probeClient) ContainerRemove(ctx context.Context, id string, opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
+	return c.remove(ctx, id, opts)
+}
+
+func (c *probeClient) ContainerStatPath(_ context.Context, _ string, opts mobyclient.ContainerStatPathOptions) (mobyclient.ContainerStatPathResult, error) {
+	return mobyclient.ContainerStatPathResult{}, c.stat(opts.Path)
+}
+
+func TestDaemonSeesDir(t *testing.T) {
+	dir := t.TempDir()
+	markers := make(map[string]bool)
+	for _, testCase := range []struct {
+		name      string
+		daemonDir string
+		removeErr error
+		seen      bool
+		removed   bool
+	}{
+		{name: "cleanup after cancellation", daemonDir: dir, seen: true, removed: true},
+		{name: "daemon without the directory", daemonDir: filepath.Join(dir, "missing")},
+		{name: "daemon with a different directory at that path", daemonDir: t.TempDir(), removed: true},
+		{name: "cleanup error after cancellation", daemonDir: dir, removeErr: errors.New("cleanup failed"), removed: true},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			ctx, cancel := context.WithCancel(t.Context())
+			defer cancel()
+			removed := false
+			cli := &probeClient{
+				create: func(opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+					assert.Equal(t, []mount.Mount{{Type: mount.TypeBind, Source: testCase.daemonDir, Target: "/gitea-runner-probe", ReadOnly: true}}, opts.HostConfig.Mounts)
+					cancel()
+					if _, err := os.Stat(testCase.daemonDir); err != nil {
+						return mobyclient.ContainerCreateResult{}, cerrdefs.ErrInvalidArgument
+					}
+					return mobyclient.ContainerCreateResult{ID: "probe"}, nil
+				},
+				stat: func(path string) error {
+					marker := strings.TrimPrefix(path, "/gitea-runner-probe/")
+					assert.False(t, markers[marker])
+					markers[marker] = true
+					if _, err := os.Stat(filepath.Join(testCase.daemonDir, marker)); err != nil {
+						return cerrdefs.ErrNotFound
+					}
+					return nil
+				},
+				remove: func(ctx context.Context, id string, opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
+					removed = true
+					assert.Equal(t, "probe", id)
+					assert.Equal(t, mobyclient.ContainerRemoveOptions{Force: true, RemoveVolumes: true}, opts)
+					require.NoError(t, ctx.Err())
+					_, bounded := ctx.Deadline()
+					assert.True(t, bounded)
+					return mobyclient.ContainerRemoveResult{}, testCase.removeErr
+				},
+			}
+			seen, err := daemonSeesDir(ctx, cli, dir, testCase.daemonDir)
+			require.ErrorIs(t, err, testCase.removeErr)
+			assert.Equal(t, testCase.seen, seen)
+			assert.Equal(t, testCase.removed, removed)
+			entries, err := os.ReadDir(dir)
+			require.NoError(t, err)
+			assert.Empty(t, entries)
+		})
+	}
+}
+
+func TestRunnerContainerWorkdir(t *testing.T) {
+	workdir := t.TempDir()
+	t.Chdir(workdir)
+	volume := container.MountPoint{Type: mount.TypeVolume, Source: "/daemon/_data", Destination: filepath.Dir(workdir)}
+	for _, testCase := range []struct {
+		name      string
+		mount     container.MountPoint
+		foreign   bool
+		daemonDir string
+	}{
+		{name: "volume above the working directory", mount: volume, daemonDir: filepath.Join(volume.Source, filepath.Base(workdir))},
+		{name: "tmpfs", mount: container.MountPoint{Type: mount.TypeTmpfs, Destination: volume.Destination}},
+		{name: "hostname of another container", mount: volume, foreign: true},
+	} {
+		cli := &probeClient{
+			mounts: []container.MountPoint{testCase.mount},
+			stat: func(path string) error {
+				if _, err := os.Stat(path); err != nil || testCase.foreign {
+					return cerrdefs.ErrNotFound
+				}
+				return nil
+			},
+		}
+		dir, daemonDir := runnerContainerWorkdir(t.Context(), cli)
+		assert.Equal(t, testCase.daemonDir, daemonDir, testCase.name)
+		assert.Equal(t, testCase.daemonDir != "", dir == workdir, testCase.name)
+	}
+	entries, err := os.ReadDir(workdir)
+	require.NoError(t, err)
+	assert.Empty(t, entries)
+}

act/container/docker_proxy_unix.go

@@ -0,0 +1,48 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build !WITHOUT_DOCKER && (linux || darwin || netbsd)
+
+package container
+
+import (
+	"encoding/binary"
+	"errors"
+	"io/fs"
+	"math"
+	"os"
+	"path/filepath"
+	"syscall"
+
+	"golang.org/x/sys/unix"
+)
+
+func copyDockerSocketPermissions(daemonSocket, socket string, info os.FileInfo) error {
+	stat, ok := info.Sys().(*syscall.Stat_t)
+	if !ok {
+		return errors.New("docker socket ownership is unavailable")
+	}
+	groupErr := os.Chown(socket, int(stat.Uid), int(stat.Gid))
+	if groupErr != nil && (!errors.Is(groupErr, fs.ErrPermission) || int(stat.Uid) != os.Geteuid()) {
+		return groupErr
+	}
+	if err := os.Chown(filepath.Dir(socket), int(stat.Uid), -1); err != nil {
+		return err
+	}
+	mode := uint16(info.Mode().Perm())
+	if err := os.Chmod(socket, fs.FileMode(mode)); err != nil || groupErr == nil {
+		return err
+	}
+	if size, err := unix.Getxattr(daemonSocket, "system.posix_acl_access", nil); err == nil && size > 0 {
+		return groupErr
+	}
+	owner, group, other := mode>>6, mode>>3&7, mode&7
+	acl := binary.LittleEndian.AppendUint32(nil, 2)
+	for _, entry := range []struct {
+		tag, perm uint16
+		id        uint32
+	}{{1, owner, math.MaxUint32}, {4, other, math.MaxUint32}, {8, group, stat.Gid}, {16, group | other, math.MaxUint32}, {32, other, math.MaxUint32}} {
+		acl = binary.LittleEndian.AppendUint32(binary.LittleEndian.AppendUint16(binary.LittleEndian.AppendUint16(acl, entry.tag), entry.perm), entry.id)
+	}
+	return unix.Setxattr(socket, "system.posix_acl_access", acl, 0) // names the group a rootless runner cannot chown to, its own group keeps what others had
+}

act/container/docker_proxy_windows.go

@@ -0,0 +1,15 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build !WITHOUT_DOCKER
+
+package container
+
+import (
+	"errors"
+	"os"
+)
+
+func copyDockerSocketPermissions(_, _ string, _ os.FileInfo) error {
+	return errors.New("docker socket ownership cannot be preserved on Windows")
+}

act/container/docker_run.go

@@ -14,6 +14,7 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"path"
 	"path/filepath"
 	"reflect"
 	"regexp"
@@ -27,13 +28,13 @@ import (
 	"gitea.com/gitea/runner/act/filecollector"
 
 	"dario.cat/mergo"
+	"github.com/bmatcuk/doublestar/v4"
 	cerrdefs "github.com/containerd/errdefs"
 	"github.com/docker/cli/cli/compose/loader"
 	"github.com/docker/cli/cli/connhelper"
 	"github.com/go-git/go-billy/v5/helper/polyfill"
 	"github.com/go-git/go-billy/v5/osfs"
 	"github.com/go-git/go-git/v5/plumbing/format/gitignore"
-	"github.com/gobwas/glob"
 	"github.com/joho/godotenv"
 	"github.com/moby/moby/api/pkg/stdcopy"
 	"github.com/moby/moby/api/types/container"
@@ -257,6 +258,27 @@ func containerInfoFromInspect(inspect container.InspectResponse) *Info {
 	for _, mountPoint := range inspect.Mounts {
 		info.Mounts[mountPoint.Destination] = mountPoint.Source
 	}
+	if hostConfig := inspect.HostConfig; hostConfig != nil { // Mounts omits --tmpfs targets and subpaths
+		for target := range hostConfig.Tmpfs {
+			info.Mounts[path.Clean(target)] = ""
+		}
+		for _, spec := range hostConfig.Mounts {
+			subpath := ""
+			if spec.VolumeOptions != nil {
+				subpath = spec.VolumeOptions.Subpath
+			} else if spec.ImageOptions != nil {
+				subpath = spec.ImageOptions.Subpath
+			}
+			if target := path.Clean(spec.Target); subpath != "" && info.Mounts[target] != "" {
+				info.Mounts[target] = path.Join(info.Mounts[target], subpath)
+			}
+		}
+	}
+	for _, target := range []string{"/etc/hosts", "/etc/hostname", "/etc/resolv.conf"} { // specific to the job's network namespace
+		if _, mounted := info.Mounts[target]; !mounted {
+			info.Mounts[target] = ""
+		}
+	}
 
 	if state := inspect.State; state != nil {
 		info.State = string(state.Status)
@@ -488,14 +510,11 @@ func (cr *containerReference) remove() common.Executor {
 			RemoveVolumes: true,
 			Force:         true,
 		})
-		switch {
-		case cerrdefs.IsConflict(err):
-			// the daemon's own AutoRemove teardown is running, and it releases the volume
-			// references and the network endpoint only once it finishes
-			cr.waitForRemoval(ctx, idOrName)
-		case err != nil && !cerrdefs.IsNotFound(err):
-			logger.Error(fmt.Errorf("failed to remove container %s: %w", idOrName, err))
-			return nil // keep the id, the container is still there for a later Remove()
+		if cerrdefs.IsConflict(err) {
+			err = cr.waitForRemoval(ctx, idOrName)
+		}
+		if err != nil && !cerrdefs.IsNotFound(err) {
+			return fmt.Errorf("failed to remove container %s: %w", idOrName, err)
 		}
 
 		logger.Debugf("Removed container: %v", idOrName)
@@ -504,7 +523,7 @@ func (cr *containerReference) remove() common.Executor {
 	}
 }
 
-func (cr *containerReference) waitForRemoval(ctx context.Context, idOrName string) {
+func (cr *containerReference) waitForRemoval(ctx context.Context, idOrName string) error {
 	// per container, against the one minute the post-job executor allows for the whole
 	// cleanup, so a job with several services can spend most of that budget here
 	ctx, cancel := context.WithTimeout(ctx, 15*time.Second)
@@ -514,18 +533,13 @@ func (cr *containerReference) waitForRemoval(ctx context.Context, idOrName strin
 		Condition: container.WaitConditionRemoved,
 	})
 	select {
-	case <-waitResult.Result:
-	case <-waitResult.Error:
-	case <-ctx.Done():
-		// the client delivers the result over an unbuffered channel, so leave a receiver
-		// behind or its goroutine parks on the send for the lifetime of the process
-		go func() {
-			select {
-			case <-waitResult.Result:
-			case <-waitResult.Error:
-			}
-		}()
-		common.Logger(ctx).Warnf("Timed out waiting for the daemon to remove container %s, its volumes and network may be left behind", idOrName)
+	case result := <-waitResult.Result:
+		if result.Error != nil {
+			return errors.New(result.Error.Message)
+		}
+		return nil
+	case err := <-waitResult.Error:
+		return err
 	}
 }
 
@@ -1277,15 +1291,15 @@ func overlayVolumes(dst, src *container.HostConfig) {
 
 type validVolumeMatcher struct {
 	allowAll bool
-	named    []glob.Glob
-	host     []glob.Glob
+	named    []string
+	host     []string
 }
 
 func newValidVolumeMatcher(ctx context.Context, validVolumes []string) validVolumeMatcher {
 	logger := common.Logger(ctx)
 	ret := validVolumeMatcher{
-		named: make([]glob.Glob, 0, len(validVolumes)),
-		host:  make([]glob.Glob, 0, len(validVolumes)),
+		named: make([]string, 0, len(validVolumes)),
+		host:  make([]string, 0, len(validVolumes)),
 	}
 
 	for _, v := range validVolumes {
@@ -1294,10 +1308,10 @@ func newValidVolumeMatcher(ctx context.Context, validVolumes []string) validVolu
 			continue
 		}
 		if !isHostVolumePattern(v) {
-			if g, err := glob.Compile(v); err != nil {
-				logger.Errorf("create glob from %s error: %v", v, err)
+			if doublestar.ValidatePattern(v) {
+				ret.named = append(ret.named, v)
 			} else {
-				ret.named = append(ret.named, g)
+				logger.Errorf("invalid volume pattern %s", v)
 			}
 			continue
 		}
@@ -1306,10 +1320,10 @@ func newValidVolumeMatcher(ctx context.Context, validVolumes []string) validVolu
 			logger.Errorf("normalize volume pattern %s error: %v", v, err)
 			continue
 		}
-		if g, err := glob.Compile(normalized); err != nil {
-			logger.Errorf("create glob from %s error: %v", normalized, err)
+		if doublestar.ValidatePathPattern(normalized) {
+			ret.host = append(ret.host, normalized)
 		} else {
-			ret.host = append(ret.host, g)
+			logger.Errorf("invalid volume pattern %s", normalized)
 		}
 	}
 
@@ -1325,15 +1339,15 @@ func (m validVolumeMatcher) isValid(source string, sourceType mount.Type) bool {
 		if err != nil {
 			return false
 		}
-		for _, g := range m.host {
-			if g.Match(normalized) {
+		for _, pattern := range m.host {
+			if doublestar.PathMatchUnvalidated(pattern, normalized) {
 				return true
 			}
 		}
 		return false
 	}
-	for _, g := range m.named {
-		if g.Match(source) {
+	for _, pattern := range m.named {
+		if doublestar.MatchUnvalidated(pattern, source) {
 			return true
 		}
 	}

act/container/docker_run_test.go

@@ -345,18 +345,22 @@ func TestDockerWaitFailure(t *testing.T) {
 // A remove that raced the daemon's AutoRemove teardown is not a failure and must not
 // be logged as one.
 func TestRemoveIgnoresAutoRemoveRace(t *testing.T) {
+	removeFailure := errors.New("driver failed to remove root filesystem")
 	removeOpts := mobyclient.ContainerRemoveOptions{RemoveVolumes: true, Force: true}
 	killOpts := mobyclient.ContainerKillOptions{Signal: "SIGKILL"}
 	for _, tc := range []struct {
-		name        string
-		err         error
-		wantWait    bool
-		wantFailure bool
+		name     string
+		err      error
+		wantWait bool
+		waitErr  error
+		wantErr  error
 	}{
 		{name: "removal in progress", err: cerrdefs.ErrConflict.WithMessage("removal of container abc is already in progress"), wantWait: true},
+		{name: "wait canceled", err: cerrdefs.ErrConflict, wantWait: true, waitErr: context.Canceled, wantErr: context.Canceled},
+		{name: "removed during wait", err: cerrdefs.ErrConflict, wantWait: true, waitErr: cerrdefs.ErrNotFound},
 		{name: "already removed", err: cerrdefs.ErrNotFound.WithMessage("No such container: abc")},
 		{name: "removed cleanly", err: nil},
-		{name: "real failure", err: errors.New("driver failed to remove root filesystem"), wantFailure: true},
+		{name: "real failure", err: removeFailure, wantErr: removeFailure},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			logger, hook := test.NewNullLogger()
@@ -366,21 +370,24 @@ func TestRemoveIgnoresAutoRemoveRace(t *testing.T) {
 			client.On("ContainerRemove", ctx, "abc", removeOpts).Return(mobyclient.ContainerRemoveResult{}, tc.err)
 			if tc.wantWait {
 				removed := make(chan container.WaitResponse, 1)
-				removed <- container.WaitResponse{}
+				waitErrors := make(chan error, 1)
+				if tc.waitErr == nil {
+					removed <- container.WaitResponse{}
+				} else {
+					waitErrors <- tc.waitErr
+				}
 				client.On("ContainerWait", mock.Anything, "abc", mobyclient.ContainerWaitOptions{Condition: container.WaitConditionRemoved}).
-					Return(mobyclient.ContainerWaitResult{Result: removed})
+					Return(mobyclient.ContainerWaitResult{Result: removed, Error: waitErrors})
 			}
 			cr := &containerReference{id: "abc", cli: client}
 
-			require.NoError(t, cr.remove()(ctx))
-			// a failure keeps the id, so a later Remove() can retry it
-			if tc.wantFailure {
+			require.ErrorIs(t, cr.remove()(ctx), tc.wantErr)
+			if tc.wantErr != nil {
 				assert.Equal(t, "abc", cr.id)
-				assert.Len(t, hook.AllEntries(), 1)
 			} else {
 				assert.Empty(t, cr.id)
-				assert.Empty(t, hook.AllEntries())
 			}
+			assert.Empty(t, hook.AllEntries())
 			client.AssertExpectations(t)
 		})
 	}
@@ -553,11 +560,14 @@ func TestCheckVolumes(t *testing.T) {
 				"shared_volume",
 				"/home/test/data",
 				"/etc/conf.d/*.json",
+				"/opt/**",
 			},
 			binds: []string{
 				"shared_volume:/shared_volume",
 				"/home/test/data:/test_data",
 				"/etc/conf.d/base.json:/config/base.json",
+				"/etc/conf.d/sub/nested.json:/config/nested.json",
+				"/opt/app/cache:/cache",
 				"sql_data:/sql_data",
 				"/secrets/keys:/keys",
 			},
@@ -565,6 +575,7 @@ func TestCheckVolumes(t *testing.T) {
 				"shared_volume:/shared_volume",
 				"/home/test/data:/test_data",
 				"/etc/conf.d/base.json:/config/base.json",
+				"/opt/app/cache:/cache",
 			},
 		},
 	}
@@ -807,12 +818,23 @@ func TestContainerInfoFromInspect(t *testing.T) {
 			Mounts: []container.MountPoint{
 				{Type: mount.TypeVolume, Name: "job", Source: "/var/lib/docker/volumes/job/_data", Destination: "/workspace/owner/repo"},
 				{Type: mount.TypeBind, Source: "/var/run/docker.sock", Destination: "/var/run/docker.sock"},
+				{Type: mount.TypeVolume, Name: "cache", Source: "/var/lib/docker/volumes/cache/_data", Destination: "/cache"},
+				{Type: mount.TypeBind, Source: "/custom/resolv.conf", Destination: "/etc/resolv.conf"},
+			},
+			HostConfig: &container.HostConfig{
+				Tmpfs:  map[string]string{"/workspace/owner/repo//tmp/": "size=1m"},
+				Mounts: []mount.Mount{{Type: mount.TypeVolume, Source: "cache", Target: "/cache/", VolumeOptions: &mount.VolumeOptions{Subpath: "project"}}},
 			},
 		})
 
 		assert.Equal(t, map[string]string{
-			"/workspace/owner/repo": "/var/lib/docker/volumes/job/_data",
-			"/var/run/docker.sock":  "/var/run/docker.sock",
+			"/workspace/owner/repo":     "/var/lib/docker/volumes/job/_data",
+			"/workspace/owner/repo/tmp": "",
+			"/var/run/docker.sock":      "/var/run/docker.sock",
+			"/cache":                    "/var/lib/docker/volumes/cache/_data/project",
+			"/etc/resolv.conf":          "/custom/resolv.conf",
+			"/etc/hosts":                "",
+			"/etc/hostname":             "",
 		}, info.Mounts)
 	})
 }

act/container/docker_stub.go

@@ -9,6 +9,7 @@ package container
 import (
 	"context"
 	"errors"
+	"net/netip"
 	"runtime"
 	"time"
 
@@ -17,6 +18,10 @@ import (
 	"github.com/moby/moby/api/types/system"
 )
 
+func IsolatedCacheContainer(context.Context, netip.Addr, string) (string, error) {
+	return "", nil
+}
+
 // ImageExistsLocally returns a boolean indicating if an image with the
 // requested name, tag and architecture exists in the local docker image store
 func ImageExistsLocally(ctx context.Context, imageName, platform string) (bool, error) {
@@ -29,8 +34,12 @@ func RemoveImage(ctx context.Context, imageName string, force, pruneChildren boo
 	return false, errors.New("Unsupported Operation")
 }
 
-func DockerProxyDir(ctx context.Context) string {
-	return ""
+func RemoveDockerJobResources(_ context.Context, _ string) error {
+	return nil
+}
+
+func NewDockerProxy(_ context.Context, _ string) *DockerProxy {
+	return nil
 }
 
 func StartDockerProxy(daemonSocket, dir, job string) (*DockerProxy, error) {
@@ -82,6 +91,26 @@ func NewDockerNetworkRemoveExecutor(name string) common.Executor {
 	}
 }
 
-func RemoveOrphanNetworks(ctx context.Context, runnerUUID string, createdBefore time.Time) error {
+func NewDockerNetworkConnectExecutor(_, _ string) common.Executor {
+	return func(_ context.Context) error {
+		return nil
+	}
+}
+
+func NewDockerNetworkDisconnectExecutor(_, _ string) common.Executor {
+	return func(_ context.Context) error {
+		return nil
+	}
+}
+
+func RemoveOrphanNetworks(_ context.Context, _, _ string, _ time.Time) error {
+	return nil
+}
+
+func CreateJobVolumes(ctx context.Context, runnerUUID string, volumeNames []string) error {
+	return nil
+}
+
+func RemoveOrphanJobVolumes(ctx context.Context, runnerUUID string, createdBefore time.Time) error {
 	return nil
 }

act/container/docker_volume.go

@@ -8,40 +8,67 @@ package container
 
 import (
 	"context"
+	"errors"
+	"fmt"
+	"time"
 
 	"gitea.com/gitea/runner/act/common"
 
+	cerrdefs "github.com/containerd/errdefs"
 	"github.com/moby/moby/client"
 )
 
-func NewDockerVolumeRemoveExecutor(volumeName string, force bool) common.Executor {
-	return func(ctx context.Context) error {
-		cli, err := GetDockerClient(ctx)
-		if err != nil {
-			return err
-		}
-		defer cli.Close()
-
-		list, err := cli.VolumeList(ctx, client.VolumeListOptions{})
-		if err != nil {
-			return err
-		}
-
-		for _, vol := range list.Items {
-			if vol.Name == volumeName {
-				return removeExecutor(volumeName, force)(ctx)
-			}
-		}
-
-		// Volume not found - do nothing
-		return nil
+func CreateJobVolumes(ctx context.Context, runnerUUID string, volumeNames []string) error {
+	cli, err := GetDockerClient(ctx)
+	if err != nil {
+		return err
 	}
+	defer cli.Close()
+
+	for _, volumeName := range volumeNames {
+		if _, err := cli.VolumeCreate(ctx, client.VolumeCreateOptions{
+			Name:   volumeName,
+			Labels: runnerLabels(runnerUUID),
+		}); err != nil {
+			return err
+		}
+	}
+	return nil
 }
 
-func removeExecutor(volume string, force bool) common.Executor {
+func RemoveOrphanJobVolumes(ctx context.Context, runnerUUID string, createdBefore time.Time) error {
+	cli, err := GetDockerClient(ctx)
+	if err != nil {
+		return fmt.Errorf("failed to connect to the docker daemon: %w", err)
+	}
+	defer cli.Close()
+
+	volumes, err := cli.VolumeList(ctx, client.VolumeListOptions{
+		Filters: make(client.Filters).Add("label", runnerUUIDLabel+"="+runnerUUID).Add("dangling", "true"),
+	})
+	if err != nil {
+		return err
+	}
+
+	var errs []error
+	for _, item := range volumes.Items {
+		created, err := time.Parse(time.RFC3339, item.CreatedAt)
+		// an unreadable or recent timestamp may belong to a job that is still starting up
+		if err != nil || created.After(createdBefore) || item.Labels[runnerUUIDLabel] != runnerUUID {
+			continue
+		}
+		if _, err := cli.VolumeRemove(ctx, item.Name, client.VolumeRemoveOptions{}); err != nil && !cerrdefs.IsNotFound(err) {
+			errs = append(errs, fmt.Errorf("failed to remove volume %s: %w", item.Name, err))
+			continue
+		}
+		common.Logger(ctx).Infof("removed docker volume %s left behind by an earlier job", item.Name)
+	}
+	return errors.Join(errs...)
+}
+
+func NewDockerVolumeRemoveExecutor(volumeName string, force bool) common.Executor {
 	return func(ctx context.Context) error {
-		logger := common.Logger(ctx)
-		logger.Debugf("docker volume rm %s", volume)
+		common.Logger(ctx).Debugf("docker volume rm %s", volumeName)
 
 		if common.Dryrun(ctx) {
 			return nil
@@ -53,7 +80,10 @@ func removeExecutor(volume string, force bool) common.Executor {
 		}
 		defer cli.Close()
 
-		_, err = cli.VolumeRemove(ctx, volume, client.VolumeRemoveOptions{Force: force})
+		_, err = cli.VolumeRemove(ctx, volumeName, client.VolumeRemoveOptions{Force: force})
+		if cerrdefs.IsNotFound(err) { // already gone is the outcome we wanted
+			return nil
+		}
 		return err
 	}
 }

act/ghcontext/github_context.go

@@ -12,6 +12,7 @@ import (
 	"context"
 	"fmt"
 	"strings"
+	"sync"
 
 	"gitea.com/gitea/runner/act/common"
 	"gitea.com/gitea/runner/act/common/git"
@@ -25,6 +26,26 @@ var (
 	findGithubRepo  = git.FindGithubRepo
 )
 
+type pinnedKey struct{}
+
+// WithPinnedCheckout keeps a run's sha and ref fixed when HEAD moves.
+func WithPinnedCheckout(ctx context.Context) context.Context {
+	if ctx.Value(pinnedKey{}) != nil {
+		return ctx
+	}
+	return context.WithValue(ctx, pinnedKey{}, &sync.Map{})
+}
+
+func pinned(ctx context.Context, key string, lookup func() (string, error)) (string, error) {
+	cache, ok := ctx.Value(pinnedKey{}).(*sync.Map)
+	if !ok {
+		return lookup()
+	}
+	cached, _ := cache.LoadOrStore(key, sync.OnceValues(lookup))
+	lookupOnce, _ := cached.(func() (string, error))
+	return lookupOnce()
+}
+
 // SetRef resolves the ref of the context from its event payload, falling back
 // to the ref checked out in repoPath.
 func SetRef(ctx context.Context, ghc *model.GithubContext, repoPath string) {
@@ -51,7 +72,7 @@ func SetRef(ctx context.Context, ghc *model.GithubContext, repoPath string) {
 	}
 
 	if ghc.Ref == "" {
-		ref, err := findGitRef(ctx, repoPath)
+		ref, err := pinned(ctx, "ref:"+repoPath, func() (string, error) { return findGitRef(ctx, repoPath) })
 		if err != nil {
 			logger.Warningf("unable to get git ref: %v", err)
 		} else {
@@ -95,7 +116,10 @@ func SetSha(ctx context.Context, ghc *model.GithubContext, repoPath string) {
 	}
 
 	if ghc.Sha == "" {
-		_, sha, err := findGitRevision(ctx, repoPath)
+		sha, err := pinned(ctx, "sha:"+repoPath, func() (string, error) {
+			_, revision, err := findGitRevision(ctx, repoPath)
+			return revision, err
+		})
 		if err != nil {
 			logger.Warningf("unable to get git revision: %v", err)
 		} else {

act/ghcontext/github_context_test.go

@@ -214,4 +214,17 @@ func TestSetSha(t *testing.T) {
 			assert.Equal(t, table.sha, ghc.Sha)
 		})
 	}
+
+	t.Run("pinned checkout keeps the first sha and ref", func(t *testing.T) {
+		ctx := WithPinnedCheckout(context.Background())
+		for _, head := range []string{"first", "moved"} {
+			findGitRevision = func(context.Context, string) (string, string, error) { return "", head, nil }
+			findGitRef = func(context.Context, string) (string, error) { return "refs/heads/" + head, nil }
+			ghc := &model.GithubContext{Event: map[string]any{}}
+			SetSha(ctx, ghc, "/some/dir")
+			SetRef(ctx, ghc, "/some/dir")
+			assert.Equal(t, "first", ghc.Sha)
+			assert.Equal(t, "refs/heads/first", ghc.Ref)
+		}
+	})
 }

act/runner/action.go

@@ -13,6 +13,7 @@ import (
 	"fmt"
 	"io"
 	"io/fs"
+	"maps"
 	"os"
 	"path"
 	"path/filepath"
@@ -147,7 +148,7 @@ func maybeCopyToActionDir(ctx context.Context, step actionStep, actionDir, actio
 
 	if !rc.Config.NoActionPatch {
 		// A concurrent job's prepare resets this directory, so patch under the copy's lock.
-		patchActions(ctx, actionScriptPaths(filepath.Join(actionDir, actionPath), step.getActionModel()))
+		defer patchActions(ctx, actionScriptPaths(filepath.Join(actionDir, actionPath), step.getActionModel()))()
 	}
 
 	if err := removeGitIgnore(ctx, actionDir); err != nil {
@@ -353,12 +354,7 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 			logger.Debugf("image '%s' for architecture '%s' already exists", image, rc.Config.ContainerArchitecture)
 		}
 	}
-	eval := rc.NewActionInputsExpressionEvaluator(ctx, step)
-	args, err := eval.Interpolate(ctx, step.getStepModel().With["args"])
-	if err != nil {
-		return fmt.Errorf("unable to interpolate with.args: %w", err)
-	}
-	cmd, err := shellquote.Split(args)
+	cmd, err := shellquote.Split(step.getStepModel().With["args"])
 	if err != nil {
 		return err
 	}
@@ -375,14 +371,11 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 			}
 		}
 	}
-	entrypoint, err := dockerEntrypoint(ctx, step, eval, stage)
-	if err != nil {
-		return err
-	}
-	stepContainer, err := newStepContainer(ctx, step, image, cmd, entrypoint, rc.Config.ContainerOptions)
+	entrypoint, err := dockerEntrypoint(step, stage)
 	if err != nil {
 		return err
 	}
+	stepContainer := newStepContainer(ctx, step, image, cmd, entrypoint, rc.Config.ContainerOptions)
 	return common.NewPipelineExecutor(
 		prepImage,
 		stepContainer.Pull(forcePull),
@@ -394,7 +387,7 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 
 // dockerEntrypoint returns the entrypoint the action's image runs with for the given
 // stage. Only the main stage honours the `entrypoint` input.
-func dockerEntrypoint(ctx context.Context, step actionStep, eval *expressionEvaluator, stage stepStage) ([]string, error) {
+func dockerEntrypoint(step actionStep, stage stepStage) ([]string, error) {
 	runs := step.getActionModel().Runs
 
 	var entrypoint string
@@ -406,11 +399,7 @@ func dockerEntrypoint(ctx context.Context, step actionStep, eval *expressionEval
 	default:
 		entrypoint = runs.Entrypoint
 		if entrypoint == "" {
-			withEntrypoint, err := eval.Interpolate(ctx, step.getStepModel().With["entrypoint"])
-			if err != nil {
-				return nil, fmt.Errorf("unable to interpolate with.entrypoint: %w", err)
-			}
-			if fields := strings.Fields(withEntrypoint); len(fields) > 0 {
+			if fields := strings.Fields(step.getStepModel().With["entrypoint"]); len(fields) > 0 {
 				return fields, nil
 			}
 		}
@@ -425,7 +414,6 @@ func dockerEntrypoint(ctx context.Context, step actionStep, eval *expressionEval
 // evalDockerEnv returns an evaluator bound to the environment it installed.
 func evalDockerEnv(ctx context.Context, step step, action *model.Action) (*expressionEvaluator, error) {
 	rc := step.getRunContext()
-	stepModel := step.getStepModel()
 
 	var err error
 	inputs := make(map[string]string)
@@ -436,27 +424,20 @@ func evalDockerEnv(ctx context.Context, step step, action *model.Action) (*expre
 			return nil, fmt.Errorf("unable to interpolate the default of input %s: %w", k, err)
 		}
 	}
-	for k, v := range stepModel.With {
-		if inputs[k], err = eval.Interpolate(ctx, v); err != nil {
-			return nil, fmt.Errorf("unable to interpolate with.%s: %w", k, err)
-		}
-	}
-	mergeIntoMap(step, step.getEnv(), inputs)
-
-	env := make(map[string]string, len(action.Runs.Env)+len(*step.getEnv()))
-	mergeIntoMap(step, &env, action.Runs.Env, *step.getEnv())
-	*step.getEnv() = env
+	mergeIntoMap(step, step.getEnv(), inputs, step.getStepModel().With)
 
 	ee := rc.NewActionInputsExpressionEvaluator(ctx, step)
-	for k, v := range *step.getEnv() {
-		if (*step.getEnv())[k], err = ee.Interpolate(ctx, v); err != nil {
+	runsEnv := make(map[string]string, len(action.Runs.Env))
+	for k, v := range action.Runs.Env {
+		if runsEnv[k], err = ee.Interpolate(ctx, v); err != nil {
 			return nil, fmt.Errorf("unable to interpolate env %s: %w", k, err)
 		}
 	}
+	mergeIntoMap(step, step.getEnv(), runsEnv, maps.Clone(*step.getEnv()))
 	return ee, nil
 }
 
-func newStepContainer(ctx context.Context, step step, image string, cmd, entrypoint []string, runnerOptions string) (container.Container, error) {
+func newStepContainer(ctx context.Context, step step, image string, cmd, entrypoint []string, runnerOptions string) container.Container {
 	rc := step.getRunContext()
 	logWriter := rc.commandLogWriter(ctx)
 	envList := make([]string, 0)
@@ -466,10 +447,7 @@ func newStepContainer(ctx context.Context, step step, image string, cmd, entrypo
 
 	envList = append(envList, rc.runnerEnv(ctx)...)
 
-	binds, mounts, err := rc.GetBindsAndMounts()
-	if err != nil {
-		return nil, err
-	}
+	binds, mounts := rc.GetBindsAndMounts()
 	networkMode := "container:" + rc.jobContainerName()
 	if rc.IsHostEnv() {
 		networkMode = "default"
@@ -493,7 +471,7 @@ func newStepContainer(ctx context.Context, step step, image string, cmd, entrypo
 		AutoRemove:    true,
 		ValidVolumes:  rc.validVolumes(),
 		AllocatePTY:   rc.Config.AllocatePTY,
-	}), nil
+	})
 }
 
 func populateEnvsFromSavedState(env *map[string]string, step actionStep, rc *RunContext) {

act/runner/action_composite.go

@@ -104,6 +104,8 @@ func newCompositeRunContext(ctx context.Context, parent *RunContext, step action
 		Parent:        parent,
 		EventJSON:     parent.EventJSON,
 		platformImage: parent.platformImage,
+		jobIndex:      parent.jobIndex,
+		jobTotal:      parent.jobTotal,
 	}
 	compositerc.setCompositeActionEnv(env)
 	compositerc.ExprEval = compositerc.NewExpressionEvaluator(ctx)

act/runner/action_composite_test.go

@@ -12,18 +12,21 @@ import (
 	"gitea.dev/actionslib/pkg/model"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+	"go.yaml.in/yaml/v4"
 )
 
 func TestCompositeActionParity(t *testing.T) {
 	t.Run("inherits contexts without leaking inputs", func(t *testing.T) {
 		ctx := t.Context()
-		strategy := &model.Strategy{MaxParallel: 3}
+		strategy := &model.Strategy{MaxParallelString: "3", RawMatrix: yaml.Node{Kind: yaml.MappingNode}}
 		parent := &RunContext{
 			Config:        &Config{},
 			Matrix:        map[string]any{"os": "linux"},
 			Run:           &model.Run{JobID: "job", Workflow: &model.Workflow{Name: "workflow", Jobs: map[string]*model.Job{"job": {Strategy: strategy}}}},
 			JobContainer:  &jobContainerMock{},
 			platformImage: "-self-hosted",
+			jobIndex:      1,
+			jobTotal:      4,
 		}
 		composite, err := newCompositeRunContext(ctx, parent, &stepActionRemote{
 			Step:       &model.Step{With: map[string]string{"SHARED": "outer"}},
@@ -36,9 +39,9 @@ func TestCompositeActionParity(t *testing.T) {
 		assert.Same(t, strategy, composite.Run.Job().Strategy)
 		assert.True(t, composite.IsHostEnv())
 		interpolated, err := composite.NewExpressionEvaluator(ctx).Interpolate(ctx,
-			"${{ matrix.os }}|${{ strategy.max-parallel }}|${{ inputs.shared }}")
+			"${{ matrix.os }}|${{ strategy.max-parallel }}|${{ strategy.job-index }}|${{ inputs.shared }}")
 		require.NoError(t, err)
-		assert.Equal(t, "linux|3|outer", interpolated)
+		assert.Equal(t, "linux|3|1|outer", interpolated)
 		assert.NotContains(t, composite.Env, "INPUT_SHARED")
 
 		nestedEnv := composite.GetEnv()
@@ -117,3 +120,51 @@ func TestAppendUniqueMasksNoExponentialGrowth(t *testing.T) {
 
 	assert.Equal(t, []string{"secret"}, parentMasks)
 }
+
+func TestCompositeStepIfReadsJobStatusExceptInItsOwnMainSteps(t *testing.T) {
+	failStep := func(rc *RunContext) {
+		rc.StepResults["failed"] = &model.StepResult{Conclusion: model.StepStatusFailure}
+	}
+	tests := []struct {
+		name   string
+		setup  func(job, composite *RunContext)
+		nested bool
+		stage  stepStage
+		want   map[string]bool
+	}{
+		{"post-if reads the failed job", func(job, _ *RunContext) { failStep(job) }, false, stepStagePost, map[string]bool{"success()": false, "failure()": true, "job.status == 'failure'": true}},
+		{"main if ignores the failed job", func(job, _ *RunContext) { failStep(job) }, false, stepStageMain, map[string]bool{"success()": true, "failure()": false}},
+		{"main if reads the composite's own failure", func(_, composite *RunContext) { failStep(composite) }, false, stepStageMain, map[string]bool{"success()": false, "failure()": true}},
+		{"nested main if ignores the outer composite's failure", func(_, composite *RunContext) { failStep(composite) }, true, stepStageMain, map[string]bool{"success()": true}},
+		{"main if sees the cancelled job", func(job, _ *RunContext) { job.markCancelled() }, false, stepStageMain, map[string]bool{"cancelled()": true, "success()": false}},
+		{"post-if sees the cancelled job", func(job, _ *RunContext) { job.markCancelled() }, false, stepStagePost, map[string]bool{"cancelled()": true, "success()": false}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			job := makeTestRC(t, "runs-on: ubuntu-latest")
+			composite := newTestCompositeRunContext(t, job)
+			rc := composite
+			if tt.nested {
+				rc = newTestCompositeRunContext(t, composite)
+			}
+			tt.setup(job, composite)
+			step := &stepRun{RunContext: rc}
+			for expr, want := range tt.want {
+				enabled, err := isStepEnabled(t.Context(), expr, step, tt.stage)
+				require.NoError(t, err)
+				assert.Equal(t, want, enabled, expr)
+			}
+		})
+	}
+}
+
+func newTestCompositeRunContext(t *testing.T, parent *RunContext) *RunContext {
+	composite, err := newCompositeRunContext(t.Context(), parent, &stepActionRemote{
+		Step:       &model.Step{},
+		RunContext: parent,
+		action:     &model.Action{},
+		env:        map[string]string{},
+	}, "/action")
+	require.NoError(t, err)
+	return composite
+}

act/runner/action_test.go

@@ -242,7 +242,7 @@ func TestActionRunner(t *testing.T) {
 
 			tt.step.getRunContext().JobContainer = cm
 
-			err := runActionImpl(tt.step, "dir", newRemoteAction("org/repo/path@ref"))(ctx)
+			err := runActionImpl(tt.step, "dir", mustNewRemoteAction(t, "org/repo/path@ref"))(ctx)
 
 			assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
 			cm.AssertExpectations(t)
@@ -316,7 +316,7 @@ func TestNewStepContainerDoesNotUseDockerSecrets(t *testing.T) {
 	step.On("getStepModel").Return(&model.Step{ID: "action"})
 	step.On("getEnv").Return(&env)
 
-	_, _ = newStepContainer(ctx, step, "registry.example.com/action:tag", nil, nil, "")
+	newStepContainer(ctx, step, "registry.example.com/action:tag", nil, nil, "")
 
 	// DOCKER_USERNAME/DOCKER_PASSWORD should not be injected as pull credentials for docker action containers.
 	assert.Empty(t, captured.Username)
@@ -461,18 +461,11 @@ func TestExecAsDockerHoldsCloneLockForRemoteUncached(t *testing.T) {
 	case <-innerEntered:
 	case err := <-done:
 		t.Fatalf("execAsDocker returned without entering inner: %v", err)
-	case <-time.After(time.Second):
-		t.Fatal("inner build executor not entered after lock released")
 	}
 
 	cancel()
 	releaseOnce()
-
-	select {
-	case <-done:
-	case <-time.After(time.Second):
-		t.Fatal("execAsDocker did not return after inner was released and ctx was canceled")
-	}
+	<-done
 }
 
 func TestDockerActionImageTag(t *testing.T) {

act/runner/cancellation_test.go

@@ -14,6 +14,7 @@ import (
 	"gitea.dev/actionslib/pkg/model"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+	"go.opentelemetry.io/otel/trace"
 	"go.yaml.in/yaml/v4"
 )
 
@@ -249,7 +250,8 @@ func TestPreStepFailureAffectsMainStepIfStatus(t *testing.T) {
 // SetJobError dereferences a nil map and panics. The post context must therefore be detached
 // from cancellation (so the steps run) yet still carry a usable error container.
 func TestPostStepsContextCancelledIsUsableForFailingStep(t *testing.T) {
-	cancelled, cancel := context.WithCancel(common.WithJobErrorContainer(context.Background()))
+	jobSpan := trace.NewSpanContext(trace.SpanContextConfig{TraceID: trace.TraceID{1}, SpanID: trace.SpanID{1}})
+	cancelled, cancel := context.WithCancel(trace.ContextWithSpanContext(common.WithJobErrorContainer(context.Background()), jobSpan))
 	cancel()
 	require.ErrorIs(t, cancelled.Err(), context.Canceled)
 
@@ -258,6 +260,7 @@ func TestPostStepsContextCancelledIsUsableForFailingStep(t *testing.T) {
 
 	// Detached from cancellation, so the post steps actually run.
 	require.NoError(t, postCtx.Err(), "post context must not be cancelled")
+	assert.Equal(t, jobSpan, trace.SpanContextFromContext(postCtx), "post steps must stay in the job's trace")
 
 	// A failing post step records its error instead of panicking.
 	require.NotPanics(t, func() {

act/runner/docker_proxy_integration_test.go

@@ -0,0 +1,110 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build !(WITHOUT_DOCKER || !(linux || darwin || windows || netbsd))
+
+package runner
+
+import (
+	"context"
+	"crypto/rand"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/container"
+
+	"gitea.dev/actionslib/pkg/model"
+	cerrdefs "github.com/containerd/errdefs"
+	"github.com/moby/moby/client"
+	"github.com/sirupsen/logrus/hooks/test"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestDockerProxyMountedJob(t *testing.T) {
+	mode := os.Getenv("ACT_TEST_DOCKER_PROXY")
+	if mode == "" {
+		t.Skip("set ACT_TEST_DOCKER_PROXY=proxy or direct to verify mounted Docker access")
+	}
+	require.Contains(t, []string{"proxy", "direct"}, mode)
+	ctx, cancel := context.WithTimeout(t.Context(), 2*time.Minute)
+	defer cancel()
+	dockerClient, err := container.GetDockerClient(ctx)
+	require.NoError(t, err)
+	t.Cleanup(func() { assert.NoError(t, dockerClient.Close()) })
+	require.True(t, strings.HasPrefix(dockerClient.DaemonHost(), "unix://"), "mounted Docker coverage requires a Unix daemon socket")
+
+	fixtureDir, err := filepath.Abs("testdata/docker-proxy")
+	require.NoError(t, err)
+	resourceName := "gitea-proxy-test-" + strings.ToLower(rand.Text())
+	runner, err := New(&Config{
+		Workdir:               fixtureDir,
+		ActionCacheDir:        t.TempDir(),
+		EventName:             "push",
+		PlatformPicker:        mapPlatformPicker(platforms),
+		ContainerNamePrefix:   resourceName,
+		ContainerDaemonSocket: dockerClient.DaemonHost(),
+		ContainerMaxLifetime:  2 * time.Minute,
+		ContainerOptions:      "-v /usr/local/bin/docker:/usr/local/bin/docker:ro -v /usr/local/libexec/docker/cli-plugins:/usr/local/libexec/docker/cli-plugins:ro",
+		ValidVolumes:          []string{"/usr/local/bin/docker", "/usr/local/libexec/docker/cli-plugins"},
+		Env:                   map[string]string{"PROXY_TEST_RESOURCE": resourceName, "PROXY_TEST_MODE": mode, "PROXY_TEST_IMAGE": baseImage, "COMPOSE_PROJECT_NAME": resourceName},
+	})
+	require.NoError(t, err)
+	planner, err := model.NewWorkflowPlanner(filepath.Join(fixtureDir, "push.yml"), true)
+	require.NoError(t, err)
+	plan, err := planner.PlanEvent("push")
+	if mode == "direct" {
+		plan, err = planner.PlanJob("proxy")
+	}
+	require.NoError(t, err)
+	runContext, err := runner.newRunContext(ctx, plan.Stages[0].Runs[0], nil)
+	require.NoError(t, err)
+	jobName := runContext.jobContainerName()
+	jobNetwork, _ := runContext.networkNameForGitea()
+	t.Cleanup(func() {
+		cleanCtx, cleanCancel := context.WithTimeout(context.WithoutCancel(ctx), 30*time.Second)
+		defer cleanCancel()
+		if _, err := dockerClient.ContainerRemove(cleanCtx, jobName, client.ContainerRemoveOptions{Force: true, RemoveVolumes: true}); !cerrdefs.IsNotFound(err) {
+			assert.NoError(t, err)
+		}
+		for _, name := range []string{resourceName, jobNetwork} {
+			if _, err := dockerClient.NetworkRemove(cleanCtx, name, client.NetworkRemoveOptions{}); !cerrdefs.IsNotFound(err) {
+				assert.NoError(t, err)
+			}
+		}
+		for _, name := range []string{resourceName, jobName, jobName + "-env"} {
+			if _, err := dockerClient.VolumeRemove(cleanCtx, name, client.VolumeRemoveOptions{}); !cerrdefs.IsNotFound(err) {
+				assert.NoError(t, err)
+			}
+		}
+	})
+
+	hook := &test.Hook{}
+	require.NoError(t, runner.NewPlanExecutor(plan)(common.WithLoggerHook(ctx, hook)))
+	var messages []string
+	for _, entry := range hook.AllEntries() {
+		messages = append(messages, strings.TrimSpace(entry.Message))
+	}
+	require.Contains(t, messages, "docker proxy post verified")
+	if mode == "proxy" {
+		require.Contains(t, messages, "docker binds verified")
+	}
+	_, err = dockerClient.ContainerInspect(ctx, jobName, client.ContainerInspectOptions{})
+	assert.True(t, cerrdefs.IsNotFound(err), "job container survived cleanup: %v", err)
+	_, err = dockerClient.NetworkInspect(ctx, resourceName, client.NetworkInspectOptions{})
+	if mode == "proxy" {
+		assert.True(t, cerrdefs.IsNotFound(err), "labelled network survived cleanup: %v", err)
+	} else {
+		require.NoError(t, err)
+	}
+	_, err = dockerClient.VolumeInspect(ctx, resourceName, client.VolumeInspectOptions{})
+	if mode == "proxy" {
+		assert.True(t, cerrdefs.IsNotFound(err), "labelled volume survived cleanup: %v", err)
+	} else {
+		require.NoError(t, err)
+	}
+}

act/runner/expression.go

@@ -13,7 +13,6 @@ import (
 	"reflect"
 	"regexp"
 	"strings"
-	"sync"
 	"time"
 
 	"gitea.com/gitea/runner/act/common"
@@ -36,24 +35,9 @@ func (rc *RunContext) NewExpressionEvaluatorWithEnv(ctx context.Context, env map
 	var workflowCallResult map[string]*model.WorkflowCallResult
 
 	// todo: cleanup EvaluationEnvironment creation
-	using := make(map[string]exprparser.Needs)
 	strategy := make(map[string]any)
 	if rc.Run != nil {
-		job := rc.Run.Job()
-		if job != nil && job.Strategy != nil {
-			strategy["fail-fast"] = job.Strategy.FailFast
-			strategy["max-parallel"] = job.Strategy.MaxParallel
-		}
-
-		jobs := rc.Run.Workflow.Jobs
-		jobNeeds := rc.Run.Job().Needs()
-
-		for _, needs := range jobNeeds {
-			using[needs] = exprparser.Needs{
-				Outputs: jobs[needs].Outputs,
-				Result:  jobs[needs].NeedsResult(),
-			}
-		}
+		strategy = exprparser.StrategyContext(rc.Run.Job().Strategy, rc.jobIndex, rc.jobTotal)
 
 		// only setup jobs context in case of workflow_call
 		// and existing expression evaluator (this means, jobs are at
@@ -61,7 +45,7 @@ func (rc *RunContext) NewExpressionEvaluatorWithEnv(ctx context.Context, env map
 		if rc.caller != nil && rc.ExprEval != nil {
 			workflowCallResult = map[string]*model.WorkflowCallResult{}
 
-			for jobName, job := range jobs {
+			for jobName, job := range rc.Run.Workflow.Jobs {
 				result := model.WorkflowCallResult{
 					Outputs: map[string]string{},
 				}
@@ -86,7 +70,7 @@ func (rc *RunContext) NewExpressionEvaluatorWithEnv(ctx context.Context, env map
 		Vars:      getWorkflowVars(ctx, rc),
 		Strategy:  strategy,
 		Matrix:    rc.Matrix,
-		Needs:     using,
+		Needs:     exprparser.NeedsContext(rc.Run),
 		Inputs:    inputs,
 		HashFiles: getHashFilesFunction(ctx, rc),
 	}
@@ -105,44 +89,26 @@ var hashfiles string
 
 // NewStepExpressionEvaluator creates a new evaluator with the `inputs` of the enclosing workflow or composite action
 func (rc *RunContext) NewStepExpressionEvaluator(ctx context.Context, step step) *ExpressionEvaluator {
-	return rc.newStepExpressionEvaluator(ctx, step, rc.actionInputs)
+	return rc.newStepExpressionEvaluator(ctx, step, rc.actionInputs, rc.getJobContext())
 }
 
 // NewActionInputsExpressionEvaluator creates a new evaluator with the step's own with: values as `inputs`
 func (rc *RunContext) NewActionInputsExpressionEvaluator(ctx context.Context, step step) *ExpressionEvaluator {
-	return rc.newStepExpressionEvaluator(ctx, step, inputsFromEnv(*step.getEnv()))
+	return rc.newStepExpressionEvaluator(ctx, step, inputsFromEnv(*step.getEnv()), rc.getJobContext())
 }
 
-func (rc *RunContext) newStepExpressionEvaluator(ctx context.Context, step step, stepInputs map[string]any) *ExpressionEvaluator {
+func (rc *RunContext) newStepExpressionEvaluator(ctx context.Context, step step, stepInputs map[string]any, jobContext *model.JobContext) *ExpressionEvaluator {
 	// todo: cleanup EvaluationEnvironment creation
-	job := rc.Run.Job()
-	strategy := make(map[string]any)
-	if job.Strategy != nil {
-		strategy["fail-fast"] = job.Strategy.FailFast
-		strategy["max-parallel"] = job.Strategy.MaxParallel
-	}
-
-	jobs := rc.Run.Workflow.Jobs
-	jobNeeds := rc.Run.Job().Needs()
-
-	using := make(map[string]exprparser.Needs)
-	for _, needs := range jobNeeds {
-		using[needs] = exprparser.Needs{
-			Outputs: jobs[needs].Outputs,
-			Result:  jobs[needs].NeedsResult(),
-		}
-	}
-
 	ee := &exprparser.EvaluationEnvironment{
 		Github:   step.getGithubContext(ctx),
 		Env:      *step.getEnv(),
-		Job:      rc.getJobContext(),
+		Job:      jobContext,
 		Steps:    rc.getStepsContext(),
 		Secrets:  getWorkflowSecrets(rc),
 		Vars:     getWorkflowVars(ctx, rc),
-		Strategy: strategy,
+		Strategy: exprparser.StrategyContext(rc.Run.Job().Strategy, rc.jobIndex, rc.jobTotal),
 		Matrix:   rc.Matrix,
-		Needs:    using,
+		Needs:    exprparser.NeedsContext(rc.Run),
 		// todo: should be unavailable
 		// but required to interpolate/evaluate the inputs in actions/composite
 		Inputs:    getEvaluatorInputs(rc, stepInputs, rc.getGithubContext(ctx)),
@@ -285,45 +251,30 @@ func getEvaluatorInputs(rc *RunContext, stepInputs map[string]any, ghc *model.Gi
 	maps.Copy(inputs, rc.workflowCallInputs)
 	maps.Copy(inputs, stepInputs)
 
-	if ghc.EventName == "workflow_dispatch" {
-		config := rc.Run.Workflow.WorkflowDispatchConfig()
-		if config != nil && config.Inputs != nil {
-			for k, v := range config.Inputs {
-				value := nestedMapLookup(ghc.Event, "inputs", k)
-				if value == nil {
-					value = v.Default
+	switch ghc.EventName {
+	case "workflow_dispatch":
+		if config := rc.Run.Workflow.WorkflowDispatchConfig(); config != nil {
+			for name, input := range config.Inputs {
+				if input.Type == "number" {
+					input.Type = "string" // boolean is the only typed dispatch input, as in Gitea
 				}
-				inputs[k] = coerceInputValue(value, v.Type)
+				setEvaluatorInput(inputs, ghc, name, model.WorkflowCallInput{Type: input.Type, Default: input.Default})
 			}
 		}
-	}
-
-	if ghc.EventName == "workflow_call" {
-		config := rc.Run.Workflow.WorkflowCallConfig()
-		if config != nil && config.Inputs != nil {
-			for k, v := range config.Inputs {
-				value := nestedMapLookup(ghc.Event, "inputs", k)
-				if value == nil {
-					value = v.Default
-				}
-				inputs[k] = coerceInputValue(value, v.Type)
-			}
+	case "workflow_call":
+		for name, input := range rc.Run.Workflow.WorkflowCallConfig().Inputs {
+			setEvaluatorInput(inputs, ghc, name, input)
 		}
 	}
 	return inputs
 }
 
-// coerceInputValue converts an input value to the type declared by the workflow.
-// The event payload carries natively typed JSON values on newer Gitea versions,
-// while defaults and older servers provide strings.
-func coerceInputValue(value any, inputType string) any {
-	if inputType != "boolean" {
-		return value
+func setEvaluatorInput(inputs map[string]any, ghc *model.GithubContext, name string, input model.WorkflowCallInput) {
+	value := model.NestedMapLookup(ghc.Event, "inputs", name)
+	if value == nil {
+		value, _ = input.DefaultValue()
 	}
-	if b, ok := value.(bool); ok {
-		return b
-	}
-	return value == "true"
+	inputs[name] = model.CoerceInputValue(value, input.Type)
 }
 
 // resolveWorkflowCall evaluates the caller's with: and secrets: once, as the server does before dispatching a called workflow.
@@ -331,36 +282,32 @@ func (rc *RunContext) resolveWorkflowCall(ctx context.Context) error {
 	if rc.caller == nil {
 		return nil
 	}
+	logger := common.Logger(ctx)
 	callerJob := rc.caller.runContext.Run.Job()
-	callerEval := rc.caller.runContext.ExprEval
-	calleeEval := sync.OnceValue(func() *expressionEvaluator { return rc.NewExpressionEvaluator(ctx) })
+	callerEval := rc.caller.runContext.ExprEval.shared(ctx)
+	var callerWith map[string]any
+	if err := model.DecodeEvaluated("workflow inputs", callerJob.RawWith, callerEval.EvaluateYamlNode, &callerWith); err != nil {
+		return err
+	}
+	if !rc.Run.Workflow.IsWorkflowCall() {
+		logger.Warn("workflow_call key is not defined in the referenced workflow")
+	}
 	config := rc.Run.Workflow.WorkflowCallConfig()
-
-	rc.workflowCallInputs = make(map[string]any, len(config.Inputs))
-	for name, input := range config.Inputs {
-		value, eval, label := callerJob.With[name], callerEval, "input"
-		if value == nil {
-			value, eval, label = input.Default, calleeEval(), "the default of input"
-		}
-		if str, ok := value.(string); ok {
-			var err error
-			if value, err = eval.Interpolate(ctx, str); err != nil {
-				return fmt.Errorf("unable to interpolate %s %s: %w", label, name, err)
-			}
-		}
-		rc.workflowCallInputs[name] = coerceInputValue(value, input.Type)
+	var err error
+	if rc.workflowCallInputs, err = callerEval.ResolveWorkflowCallInputs(config, callerWith); err != nil {
+		logger.Warn(err)
 	}
 
+	if callerJob.InheritSecrets() {
+		rc.workflowCallSecrets = getWorkflowSecrets(rc.caller.runContext)
+		return nil
+	}
 	secrets := callerJob.Secrets()
-	if secrets == nil && callerJob.InheritSecrets() {
-		secrets = rc.caller.runContext.Config.Secrets
+	if err := config.ValidateSecrets(secrets); err != nil {
+		logger.Warn(err)
 	}
-	rc.workflowCallSecrets = make(map[string]string, len(secrets))
-	for k, v := range secrets {
-		var err error
-		if rc.workflowCallSecrets[k], err = callerEval.Interpolate(ctx, v); err != nil {
-			return fmt.Errorf("unable to interpolate secret %s: %w", k, err)
-		}
+	if rc.workflowCallSecrets, err = callerEval.EvaluateCallerSecrets(secrets); err != nil {
+		return fmt.Errorf("unable to interpolate %w", err)
 	}
 	return nil
 }

act/runner/expression_test.go

@@ -89,7 +89,7 @@ func TestEvaluateRunContext(t *testing.T) {
 		out     any
 		errMesg string
 	}{
-		{" 1 ", 1, ""},
+		{" 1 ", 1.0, ""},
 		// {"1 + 3", "4", ""},
 		// {"(1 + 3) * -2", "-8", ""},
 		{"'my text'", "my text", ""},
@@ -266,7 +266,6 @@ func TestInterpolate(t *testing.T) {
 		{"${{ null }}", ""},
 		{"${{ fromJSON('[1,2]') }}", "Array"},
 		{"${{ fromJSON('{\"a\":1}') }}", "Object"},
-		{"${{ 1", "${{ 1"},
 	}
 
 	for _, table := range tables {
@@ -277,7 +276,7 @@ func TestInterpolate(t *testing.T) {
 		})
 	}
 
-	for _, in := range []string{"${{ 1) && (2 }}", "run ${{ 1) && (2 }} now"} {
+	for _, in := range []string{"${{ 1) && (2 }}", "run ${{ 1) && (2 }} now", "${{ 1"} {
 		_, err := ee.Interpolate(context.Background(), in)
 		assert.Error(t, err, in)
 	}
@@ -295,6 +294,8 @@ on:
       name:
         type: string
         default: gitea
+      unset:
+        type: boolean
 `,
 		"workflow_dispatch": `
 on:
@@ -306,6 +307,8 @@ on:
       name:
         type: string
         default: gitea
+      unset:
+        type: boolean
 `,
 	}
 
@@ -357,9 +360,76 @@ on:
 				inputs := getEvaluatorInputs(rc, nil, ghc)
 				assert.Equal(t, table.flag, inputs["flag"])
 				assert.Equal(t, "gitea", inputs["name"])
+				assert.Equal(t, false, inputs["unset"])
 			})
 		}
 	}
+
+	t.Run("deferred workflow call inputs", func(t *testing.T) {
+		workflow, err := model.ReadWorkflow(strings.NewReader(workflows["workflow_call"] + `
+jobs:
+  test: {}
+  call:
+    uses: ./reuse.yml
+    with: ${{ fromJSON(matrix.args) }}
+`))
+		require.NoError(t, err)
+		workflow.GetJob("test")
+		job := workflow.GetJob("call")
+		rawWith := model.CloneYamlNode(job.RawWith)
+		for _, value := range []string{"true", "false"} {
+			t.Run(value, func(t *testing.T) {
+				t.Parallel()
+				parent, err := (&runnerImpl{config: &Config{}}).newRunContext(t.Context(), &model.Run{Workflow: workflow, JobID: "call"}, map[string]any{"args": `{"flag":` + value + `,"name":"${{ 'runner' }}"}`})
+				require.NoError(t, err)
+				child, err := (&runnerImpl{config: &Config{}, caller: &caller{runContext: parent}}).newRunContext(t.Context(), &model.Run{Workflow: workflow, JobID: "test"}, nil)
+				require.NoError(t, err)
+				assert.Equal(t, map[string]any{"flag": value == "true", "name": "${{ 'runner' }}", "unset": false}, child.workflowCallInputs)
+				assert.Equal(t, rawWith, job.RawWith)
+				assert.Nil(t, job.With)
+			})
+		}
+	})
+}
+
+func TestResolveWorkflowCallToleratesMismatchesAndPassesInheritedSecretsVerbatim(t *testing.T) {
+	callerWorkflow, err := model.ReadWorkflow(strings.NewReader(`
+jobs:
+  call:
+    uses: ./reuse.yml
+    with: {flag: not-a-boolean}
+    secrets: {undeclared: "${{ secrets.token }}"}
+  inherit:
+    uses: ./reuse.yml
+    secrets: inherit
+`))
+	require.NoError(t, err)
+	callee, err := model.ReadWorkflow(strings.NewReader(`
+on:
+  workflow_call:
+    inputs:
+      flag: {type: boolean, default: true}
+jobs:
+  test: {}
+`))
+	require.NoError(t, err)
+	notCallable, err := model.ReadWorkflow(strings.NewReader("on: push\njobs:\n  test: {}\n"))
+	require.NoError(t, err)
+	config := &Config{Secrets: map[string]string{"token": "${{ github.token }}"}}
+	resolve := func(callerJobID string, workflow *model.Workflow) *RunContext {
+		parent, err := (&runnerImpl{config: config}).newRunContext(t.Context(), &model.Run{Workflow: callerWorkflow, JobID: callerJobID}, nil)
+		require.NoError(t, err)
+		child, err := (&runnerImpl{config: config, caller: &caller{runContext: parent}}).newRunContext(t.Context(), &model.Run{Workflow: workflow, JobID: "test"}, nil)
+		require.NoError(t, err)
+		return child
+	}
+
+	called := resolve("call", callee)
+	assert.Equal(t, map[string]any{"flag": true}, called.workflowCallInputs)
+	assert.Equal(t, map[string]string{"undeclared": "${{ github.token }}"}, called.workflowCallSecrets)
+	inherited := resolve("inherit", notCallable)
+	assert.Empty(t, inherited.workflowCallInputs)
+	assert.Equal(t, config.Secrets, inherited.workflowCallSecrets)
 }
 
 func TestJobNameMasksSecrets(t *testing.T) {

act/runner/job_executor.go

@@ -18,7 +18,6 @@ import (
 	"slices"
 	"strconv"
 	"strings"
-	"sync"
 	"time"
 	"unicode"
 
@@ -263,26 +262,27 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 		return err
 	})
 
-	stepsExecutor := newStepsExecutor(rc, preSteps, steps)
-
-	return common.NewPipelineExecutor(info.startContainer(), stepsExecutor.
-		Finally(func(ctx context.Context) error {
+	return common.Executor(func(ctx context.Context) error {
+		if err := info.startContainer()(ctx); err != nil {
+			return err
+		}
+		return newStepsExecutor(rc, preSteps, steps).Finally(func(ctx context.Context) error {
 			// Record an interrupt (backstop for interrupts that land outside the main
 			// step loop) so the post steps observe the cancelled/failed job status.
 			rc.markInterrupted(ctx.Err())
 			postCtx, cancel := postStepsContext(ctx)
 			defer cancel()
 			return postExecutor(postCtx)
-		}).
-		Finally(info.closeContainer()))
+		})(ctx)
+	}).Finally(info.closeContainer())
 }
 
 // postStepsContext derives the context used to run the job's post/cleanup steps from the
 // finished main-pipeline context. Cleanup has to run even when the run was interrupted, so the
 // returned context always carries a fresh bounded deadline and is never itself cancelled.
 //
-//   - context.Canceled (server cancel): detach from the cancelled context via a fresh root so
-//     the post steps can run.
+//   - context.Canceled (server cancel): detach with WithoutCancel under a fresh job-error
+//     container, so a failing post step cannot turn the cancellation into a failure.
 //   - context.DeadlineExceeded (job timeout): detach the deadline with WithoutCancel, which
 //     keeps the original values — including the job-error container — so the timeout failure and
 //     any post-step error are preserved and the job is still reported as failed.
@@ -290,10 +290,7 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 func postStepsContext(ctx context.Context) (context.Context, context.CancelFunc) {
 	switch ctx.Err() {
 	case context.Canceled:
-		// The cancelled context is abandoned for a fresh root, which drops the job-error
-		// container installed at the job root. Re-attach a fresh one so a failing post step
-		// records its error via SetJobError instead of panicking on a nil container.
-		return context.WithTimeout(common.WithJobErrorContainer(common.WithLogger(context.Background(), common.Logger(ctx))), 5*time.Minute)
+		return context.WithTimeout(common.WithJobErrorContainer(context.WithoutCancel(ctx)), 5*time.Minute)
 	case context.DeadlineExceeded:
 		return context.WithTimeout(context.WithoutCancel(ctx), 5*time.Minute)
 	default:
@@ -413,41 +410,45 @@ func setJobResult(ctx context.Context, info jobInfo, rc *RunContext, success boo
 	logger.WithField("jobResult", jobResult).Infof("Job %s", jobResultMessage)
 }
 
-// evaluateJobEnvAndDefaults resolves the job's env and defaults.run once, as GitHub does at job setup.
+// evaluateJobEnvAndDefaults resolves the job's env, defaults.run and container env once, as GitHub does at job setup.
 func evaluateJobEnvAndDefaults(ctx context.Context, rc *RunContext) error {
 	rc.ExprEval = rc.NewExpressionEvaluator(ctx)
+	env := rc.GetEnv()
+	var workflowEnv map[string]string
+	if rc.Run.Workflow.Env == nil {
+		if err := model.DecodeEvaluated("workflow env", rc.Run.Workflow.RawEnv, rc.ExprEval.shared(ctx).EvaluateYamlNode, &workflowEnv); err != nil {
+			return err
+		}
+	}
+	if workflowEnv != nil {
+		rc.Env = mergeMaps(workflowEnv, env)
+		rc.ExprEval = rc.NewExpressionEvaluator(ctx)
+	}
 	var err error
-	for k, v := range rc.GetEnv() {
+	for k, v := range env {
 		if rc.Env[k], err = rc.ExprEval.Interpolate(ctx, v); err != nil {
 			return fmt.Errorf("unable to interpolate env %s: %w", k, err)
 		}
 	}
-	defaults := rc.Run.Job().Defaults.Run
-	if rc.jobRunDefaults.Shell, err = rc.ExprEval.Interpolate(ctx, defaults.Shell); err != nil {
-		return fmt.Errorf("unable to interpolate defaults.run.shell: %w", err)
+	var defaults model.Defaults
+	if err := model.DecodeEvaluated("defaults", rc.Run.Job().RawDefaults, rc.ExprEval.shared(ctx).EvaluateYamlNode, &defaults); err != nil {
+		return err
 	}
-	if rc.jobRunDefaults.WorkingDirectory, err = rc.ExprEval.Interpolate(ctx, defaults.WorkingDirectory); err != nil {
-		return fmt.Errorf("unable to interpolate defaults.run.working-directory: %w", err)
+	rc.jobRunDefaults = defaults.Run
+	if rc.containerSpec.Image == "" {
+		return nil
 	}
-	return nil
+	_, containerEnv := splitContainerEnv(rc.Run.Job().RawContainer)
+	return model.DecodeEvaluated("container env", containerEnv, rc.ExprEval.shared(ctx).EvaluateYamlNode, &rc.containerSpec.Env)
 }
 
 func setJobOutputs(ctx context.Context, rc *RunContext) error {
 	if rc.caller == nil {
 		return nil
 	}
-	outputs := rc.Run.Workflow.WorkflowCallConfig().Outputs
-	callerOutputs := make(map[string]string, len(outputs))
-	ee := sync.OnceValue(func() *expressionEvaluator { return rc.NewExpressionEvaluator(ctx) })
-	for k, v := range outputs {
-		value := v.Value
-		for range 2 { // two passes, the value resolves through a job output
-			var err error
-			if value, err = ee().Interpolate(ctx, value); err != nil {
-				return fmt.Errorf("unable to interpolate workflow output %s: %w", k, err)
-			}
-		}
-		callerOutputs[k] = value
+	callerOutputs, err := rc.NewExpressionEvaluator(ctx).shared(ctx).EvaluateWorkflowCallOutputs(rc.Run.Workflow.WorkflowCallConfig())
+	if err != nil {
+		return fmt.Errorf("unable to interpolate workflow %w", err)
 	}
 
 	// Matrix combinations of a reusable-workflow caller share the caller's *model.Job;
@@ -465,7 +466,7 @@ func applyJobTimeout(ctx context.Context, rc *RunContext, job *model.Job) (conte
 	if err != nil {
 		common.Logger(ctx).Errorf("An error occurred when attempting to determine the job timeout: %s", err)
 	} else if timeout != "" {
-		if timeoutMinutes, err := strconv.ParseInt(timeout, 10, 64); err == nil {
+		if timeoutMinutes, err := strconv.ParseInt(timeout, 10, 64); err == nil && timeoutMinutes > 0 {
 			return context.WithTimeout(ctx, time.Duration(timeoutMinutes)*time.Minute)
 		}
 	}

act/runner/job_executor_test.go

@@ -300,6 +300,8 @@ func TestNewJobExecutor(t *testing.T) {
 		result        string
 		hasError      bool
 		output        string
+		startError    error
+		cancelOnStart bool
 	}{
 		{
 			name:          "zeroSteps",
@@ -435,6 +437,21 @@ func TestNewJobExecutor(t *testing.T) {
 			result:        "failure",
 			output:        "${{ 'test' != test }}",
 		},
+		{
+			name:          "start failure",
+			steps:         []*model.Step{{ID: "1"}},
+			executedSteps: []string{"startContainer", "closeContainer"},
+			startError:    errors.New("start failed"),
+		},
+		{
+			name:          "cancelled at startup boundary",
+			steps:         []*model.Step{{ID: "1"}},
+			preSteps:      []bool{false},
+			postSteps:     []bool{true},
+			executedSteps: []string{"startContainer", "step1", "post1", "interpolateOutputs", "stopContainer", "closeContainer"},
+			result:        "cancelled",
+			cancelOnStart: true,
+		},
 	}
 
 	contains := func(needle string, haystack []string) bool {
@@ -445,7 +462,8 @@ func TestNewJobExecutor(t *testing.T) {
 		t.Run(tt.name, func(t *testing.T) {
 			fmt.Printf("::group::%s\n", tt.name) //nolint:forbidigo // pre-existing issue from nektos/act
 
-			ctx := common.WithJobErrorContainer(context.Background())
+			ctx, cancel := context.WithCancel(common.WithJobErrorContainer(context.Background()))
+			defer cancel()
 			jim := &jobInfoMock{}
 			sfm := &stepFactoryMock{}
 			rc := &RunContext{
@@ -461,8 +479,7 @@ func TestNewJobExecutor(t *testing.T) {
 				Config: &Config{},
 			}
 			if tt.output != "" {
-				rc.Run.Job().Outputs = map[string]string{"bad": tt.output}
-				rc.outputTemplate = map[string]string{"bad": tt.output}
+				require.NoError(t, rc.Run.Job().RawOutputs.Encode(map[string]string{"bad": tt.output}))
 			}
 			rc.ExprEval = rc.NewExpressionEvaluator(ctx)
 			executorOrder := make([]string, 0)
@@ -470,9 +487,12 @@ func TestNewJobExecutor(t *testing.T) {
 			jim.On("steps").Return(tt.steps)
 
 			if len(tt.steps) > 0 {
-				jim.On("startContainer").Return(func(ctx context.Context) error {
+				jim.On("startContainer").Return(func(_ context.Context) error {
 					executorOrder = append(executorOrder, "startContainer")
-					return nil
+					if tt.cancelOnStart {
+						cancel()
+					}
+					return tt.startError
 				})
 			}
 
@@ -506,7 +526,7 @@ func TestNewJobExecutor(t *testing.T) {
 				defer sm.AssertExpectations(t)
 			}
 
-			if len(tt.steps) > 0 {
+			if len(tt.steps) > 0 && tt.startError == nil {
 				jim.On("matrix").Return(map[string]any{})
 
 				jim.On("interpolateOutputs").Return(func(ctx context.Context) error {
@@ -520,12 +540,17 @@ func TestNewJobExecutor(t *testing.T) {
 				if contains("stopContainer", tt.executedSteps) {
 					jim.On("stopContainer").Return(func(ctx context.Context) error {
 						executorOrder = append(executorOrder, "stopContainer")
+						require.NoError(t, ctx.Err())
+						_, bounded := ctx.Deadline()
+						require.True(t, bounded)
 						return nil
 					})
 				}
 
 				jim.On("result", tt.result)
+			}
 
+			if len(tt.steps) > 0 {
 				jim.On("closeContainer").Return(func(ctx context.Context) error {
 					executorOrder = append(executorOrder, "closeContainer")
 					return nil
@@ -534,7 +559,14 @@ func TestNewJobExecutor(t *testing.T) {
 
 			executor := newJobExecutor(jim, sfm, rc)
 			err := executor(ctx)
-			assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
+			switch {
+			case tt.startError != nil:
+				require.ErrorIs(t, err, tt.startError)
+			case tt.cancelOnStart:
+				require.ErrorIs(t, err, context.Canceled)
+			default:
+				require.NoError(t, err)
+			}
 			assert.Empty(t, rc.Run.Job().Outputs["bad"])
 			assert.Equal(t, tt.executedSteps, executorOrder)
 
@@ -1049,6 +1081,7 @@ func TestApplyJobTimeout(t *testing.T) {
 	}{
 		{"empty", "runs-on: ubuntu-latest", false},
 		{"integer", "timeout-minutes: 5\nruns-on: ubuntu-latest", true},
+		{"zero ignored", "timeout-minutes: 0\nruns-on: ubuntu-latest", false},
 		{"non-numeric ignored", "timeout-minutes: abc\nruns-on: ubuntu-latest", false},
 	}
 	for _, tc := range cases {

act/runner/job_hooks.go

@@ -7,7 +7,6 @@ import (
 	"cmp"
 	"context"
 	"fmt"
-	"maps"
 	"path"
 	"strings"
 
@@ -51,11 +50,7 @@ func (rc *RunContext) runJobHook(ctx context.Context, hookPath, name string) err
 		rawLogger.Infof("shell: %s", shell)
 	}
 
-	env := map[string]string{}
-	if jobContainer := rc.Run.Job().Container(); jobContainer != nil {
-		maps.Copy(env, jobContainer.Env)
-	}
-	maps.Copy(env, rc.GetEnv())
+	env := mergeMaps(rc.containerSpec.Env, rc.GetEnv())
 	rc.withGithubEnv(ctx, rc.getGithubContext(ctx), env)
 	rc.ApplyExtraPath(ctx, &env)
 

act/runner/max_parallel_test.go

@@ -4,6 +4,7 @@
 package runner
 
 import (
+	"strings"
 	"testing"
 
 	"gitea.dev/actionslib/pkg/model"
@@ -34,9 +35,14 @@ func TestMaxParallelStrategy(t *testing.T) {
 			expectedMaxParallel: 4,
 		},
 		{
-			name:                "max-parallel-10",
+			name:                "max-parallel-10-clamped-to-combinations",
 			maxParallelString:   "10",
-			expectedMaxParallel: 10,
+			expectedMaxParallel: 5,
+		},
+		{
+			name:                "max-parallel-invalid-falls-back-to-default",
+			maxParallelString:   "tow",
+			expectedMaxParallel: 4,
 		},
 	}
 
@@ -61,9 +67,29 @@ func TestMaxParallelStrategy(t *testing.T) {
 			assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
 			assert.NotNil(t, matrixes)
 			assert.Len(t, matrixes, 5)
-			assert.Equal(t, tt.expectedMaxParallel, job.Strategy.MaxParallel)
+			assert.Equal(t, tt.expectedMaxParallel, maxParallelFor(job.Strategy, len(matrixes)))
 		})
 	}
+
+	t.Run("deferred strategy", func(t *testing.T) {
+		workflow, err := model.ReadWorkflow(strings.NewReader(`
+jobs:
+  test:
+    if: false
+    strategy: ${{ fromJSON('{"max-parallel":2,"fail-fast":false,"matrix":{"os":["ubuntu","windows"]}}') }}
+`))
+		require.NoError(t, err)
+		job := workflow.Jobs["test"]
+		rawStrategy := model.CloneYamlNode(job.RawStrategy)
+		require.NoError(t, (&runnerImpl{config: &Config{}}).NewPlanExecutor(&model.Plan{Stages: []*model.Stage{{Runs: []*model.Run{{Workflow: workflow, JobID: "test"}}}}})(t.Context()))
+		require.NotNil(t, job.Strategy)
+		assert.False(t, job.Strategy.GetFailFast())
+		assert.Equal(t, 2, maxParallelFor(job.Strategy, 5))
+		matrixes, err := job.GetMatrixes()
+		require.NoError(t, err)
+		assert.Equal(t, []map[string]any{{"os": "ubuntu"}, {"os": "windows"}}, matrixes)
+		assert.Equal(t, rawStrategy, job.RawStrategy)
+	})
 }
 
 func TestNewPlanExecutorInvalidMatrix(t *testing.T) {
@@ -81,4 +107,15 @@ func TestNewPlanExecutorInvalidMatrix(t *testing.T) {
 	runner := &runnerImpl{config: &Config{}}
 
 	require.ErrorContains(t, runner.NewPlanExecutor(plan)(t.Context()), "could not get job matrix:")
+
+	for _, strategy := range []string{
+		`${{ fromJSON('invalid') }}`,
+		`${{ '${{ inputs.unresolved }}' }}`,
+	} {
+		t.Run(strategy, func(t *testing.T) {
+			workflow, err := model.ReadWorkflow(strings.NewReader("jobs:\n  test:\n    strategy: " + strategy))
+			require.NoError(t, err)
+			require.Error(t, runner.NewPlanExecutor(&model.Plan{Stages: []*model.Stage{{Runs: []*model.Run{{Workflow: workflow, JobID: "test"}}}}})(t.Context()))
+		})
+	}
 }

act/runner/patch_actions.go

@@ -28,8 +28,8 @@ import (
 // of upload-artifact.
 //
 // getCacheServiceURL() then resolves the cache service from ACTIONS_RESULTS_URL alone, where v1
-// reads ACTIONS_CACHE_URL first. Both reads there are given the same preference, which is what
-// keeps the runner out of the artifact path: the results URL still points at Gitea.
+// reads ACTIONS_CACHE_URL first. Both reads there are given the same preference, so a patched cache
+// client reaches the cache server by its own address rather than through the results origin.
 //
 // Either of these landing upstream makes this file deletable:
 //
@@ -105,37 +105,45 @@ func actionScriptPaths(dir string, action *model.Action) []string {
 	return paths
 }
 
-// patchActions edits the toolkit in an action's bundles. The caller holds the action directory's
-// clone lock, which is what keeps another job's checkout from resetting them before the copy.
-func patchActions(ctx context.Context, scripts []string) {
+// patchActions returns a restore: the tree is shared, so a job with patching off gets no edits.
+func patchActions(ctx context.Context, scripts []string) func() {
+	restore := map[string][]byte{}
 	for _, script := range scripts {
-		switch patched, err := patchBundle(script); {
-		case err != nil:
+		original, err := patchBundle(script)
+		if original != nil {
+			restore[script] = original // also when the write failed part way through
+		}
+		if err != nil {
 			common.Logger(ctx).Warnf("actions toolkit: %s left unpatched: %v", script, err)
-		case patched:
-			common.Logger(ctx).Debugf("actions toolkit: patched %s", script)
+		}
+	}
+	return func() {
+		for script, original := range restore {
+			if err := os.WriteFile(script, original, 0o644); err != nil { //nolint:gosec // as the checkout wrote it
+				common.Logger(ctx).Warnf("actions toolkit: %s left patched in the shared copy: %v", script, err)
+			}
 		}
 	}
 }
 
-func patchBundle(script string) (bool, error) {
+// patchBundle returns the bytes it replaced, or nil when it left the bundle alone.
+func patchBundle(script string) ([]byte, error) {
 	info, err := os.Stat(script)
 	if err != nil {
-		return false, err
+		return nil, err
 	}
 	if info.Size() > maxBundleSize {
-		return false, nil
+		return nil, nil
 	}
 	data, err := os.ReadFile(script)
 	if err != nil {
-		return false, err
+		return nil, err
 	}
 	patched, ok := patchedBundle(data)
 	if !ok {
-		return false, nil
+		return nil, nil
 	}
-	// No atomic write needed: every prepare checks the action out and hard resets it.
-	return true, os.WriteFile(script, patched, info.Mode().Perm())
+	return data, os.WriteFile(script, patched, info.Mode().Perm())
 }
 
 // patchedBundle opens the GHES gate, and where the cache toolkit is present, points the cache

act/runner/patch_actions_test.go

@@ -176,16 +176,16 @@ func TestPatchedBundleSurvivesInsideAStringLiteral(t *testing.T) {
 func TestPatchBundleIsIdempotent(t *testing.T) {
 	script := bundleFile(t, gateTSC)
 
-	done, err := patchBundle(script)
+	original, err := patchBundle(script)
 	require.NoError(t, err)
-	require.True(t, done)
+	require.NotNil(t, original)
 	patched, err := os.ReadFile(script)
 	require.NoError(t, err)
 	require.True(t, gateOpened(string(patched)))
 
-	done, err = patchBundle(script)
+	original, err = patchBundle(script)
 	require.NoError(t, err)
-	assert.False(t, done, "a patched bundle is not patched again")
+	assert.Nil(t, original, "a patched bundle is not patched again")
 	again, err := os.ReadFile(script)
 	require.NoError(t, err)
 	assert.Equal(t, string(patched), string(again))
@@ -194,9 +194,9 @@ func TestPatchBundleIsIdempotent(t *testing.T) {
 func TestPatchBundleLeavesOtherActionsAlone(t *testing.T) {
 	script := bundleFile(t, `console.log("checkout")`)
 
-	done, err := patchBundle(script)
+	original, err := patchBundle(script)
 	require.NoError(t, err)
-	assert.False(t, done)
+	assert.Nil(t, original)
 	body, err := os.ReadFile(script)
 	require.NoError(t, err)
 	assert.Equal(t, `console.log("checkout")`, string(body))
@@ -251,6 +251,9 @@ func TestPatchActionsAtTheContainerCopy(t *testing.T) {
 			return nil
 		})
 		require.NoError(t, maybeCopyToActionDir(t.Context(), sar, sar.actionDir(), "sub", "/var/run/act/actions/repo/sub"))
+		source, err := os.ReadFile(script)
+		require.NoError(t, err)
+		assert.Equal(t, gateTSC, string(source), "the shared copy is left as it was found")
 		return copied
 	}