fix: resolve symlinked container paths before building tar entries (#1130)

Docker 29.7 extracts copied archives through `os.Root`, which refuses to follow a symlink to an absolute target, so copying into `/var/run/act` fails with `path escapes from parent` on the many images that link `/var/run` to `/run`. The daemon now resolves every path component for us before any tar entry name is built, and the destination is created with one directory entry per missing component, which no daemon version rejects.

Verified against real daemons (29.4.0, 29.5.0, 29.5.1, 29.5.3, 29.6.2, 29.7.0-rc.1) with `debian:bookworm` (absolute symlink) and `alpine:3` (relative symlink), and against `moby/go-archive` v0.2.0 through the pending fix branch.

1. Fixes https://gitea.com/gitea/runner/issues/1128
1. Upstream bug: https://github.com/moby/moby/issues/53258
1. Supersedes the no-op change in https://gitea.com/gitea/runner/pulls/1129, which cannot help since the daemon strips leading slashes itself

Reviewed-on: https://gitea.com/gitea/runner/pulls/1130
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>

.gitea/workflows/pull-pr-title.yml

@@ -18,8 +18,8 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:
-    - uses: actions/checkout@v7
-    - uses: actions/setup-node@v6
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
+    - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7
       with:
         node-version: 24
     - run: make lint-pr-title

.gitea/workflows/release-nightly.yml

@@ -17,14 +17,26 @@ jobs:
   goreleaser:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
         with:
           fetch-depth: 0
-      - uses: actions/setup-go@v6
+      # Custom publishers (the R2 mirror below) run as the very last
+      # step of goreleaser's publish pipeline, after the Gitea release
+      # has already been created and every artifact already uploaded
+      # to S3. Fail here instead, before anything is built or
+      # published, if the R2 secrets are missing.
+      - name: check R2 configuration
+        run: sh scripts/upload-r2.sh --check-config
+        env:
+          R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+          R2_BUCKET: ${{ secrets.R2_BUCKET }}
+          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
         with:
           go-version-file: "go.mod"
       - name: goreleaser
-        uses: goreleaser/goreleaser-action@v7
+        uses: goreleaser/goreleaser-action@f06c13b6b1a9625abc9e6e439d9c05a8f2190e94 # v7
         with:
           distribution: goreleaser-pro
           args: release --nightly
@@ -35,6 +47,10 @@ jobs:
           AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
           S3_REGION: ${{ secrets.AWS_REGION }}
           S3_BUCKET: ${{ secrets.AWS_BUCKET }}
+          R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+          R2_BUCKET: ${{ secrets.R2_BUCKET }}
+          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
           GORELEASER_FORCE_TOKEN: "gitea"
           GITEA_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
@@ -58,18 +74,18 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v7
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
         with:
           fetch-depth: 0 # all history for all branches and tags
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v4
+        uses: docker/setup-qemu-action@96fe6ef7f33517b61c61be40b68a1882f3264fb8 # v4
 
       - name: Set up Docker BuildX
-        uses: docker/setup-buildx-action@v4
+        uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4
 
       - name: Login to DockerHub
-        uses: docker/login-action@v4
+        uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4
         with:
           username: ${{ secrets.DOCKER_USERNAME }}
           password: ${{ secrets.DOCKER_PASSWORD }}
@@ -83,7 +99,7 @@ jobs:
           echo REPO_VERSION=$(git describe --tags --always | sed 's/-/+/' | sed 's/^v//') >> $GITHUB_OUTPUT
 
       - name: Build and push
-        uses: docker/build-push-action@v7
+        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7
         with:
           context: .
           file: ./Dockerfile

.gitea/workflows/release-tag.yml

@@ -9,21 +9,33 @@ jobs:
   goreleaser:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
         with:
           fetch-depth: 0 # all history for all branches and tags
-      - uses: actions/setup-go@v6
+      # Custom publishers (the R2 mirror below) run as the very last
+      # step of goreleaser's publish pipeline, after the Gitea release
+      # has already been created and every artifact already uploaded
+      # to S3. Fail here instead, before anything is built or
+      # published, if the R2 secrets are missing.
+      - name: check R2 configuration
+        run: sh scripts/upload-r2.sh --check-config
+        env:
+          R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+          R2_BUCKET: ${{ secrets.R2_BUCKET }}
+          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
         with:
           go-version-file: "go.mod"
       - name: Import GPG key
         id: import_gpg
-        uses: crazy-max/ghaction-import-gpg@v7
+        uses: crazy-max/ghaction-import-gpg@2dc316deee8e90f13e1a351ab510b4d5bc0c82cd # v7
         with:
           gpg_private_key: ${{ secrets.GPG_PRIVATE_KEY }}
           passphrase: ${{ secrets.PASSPHRASE }}
           fingerprint: CC64B1DB67ABBEECAB24B6455FC346329753F4B0
       - name: goreleaser
-        uses: goreleaser/goreleaser-action@v7
+        uses: goreleaser/goreleaser-action@f06c13b6b1a9625abc9e6e439d9c05a8f2190e94 # v7
         with:
           distribution: goreleaser-pro
           args: release
@@ -34,6 +46,10 @@ jobs:
           AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
           S3_REGION: ${{ secrets.AWS_REGION }}
           S3_BUCKET: ${{ secrets.AWS_BUCKET }}
+          R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+          R2_BUCKET: ${{ secrets.R2_BUCKET }}
+          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
           GORELEASER_FORCE_TOKEN: "gitea"
           GITEA_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           GPG_FINGERPRINT: ${{ steps.import_gpg.outputs.fingerprint }}
@@ -61,25 +77,25 @@ jobs:
       DOCKER_LATEST: latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v7
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
         with:
           fetch-depth: 0 # all history for all branches and tags
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v4
+        uses: docker/setup-qemu-action@96fe6ef7f33517b61c61be40b68a1882f3264fb8 # v4
 
       - name: Set up Docker BuildX
-        uses: docker/setup-buildx-action@v4
+        uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4
 
       - name: Login to DockerHub
-        uses: docker/login-action@v4
+        uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4
         with:
           username: ${{ secrets.DOCKER_USERNAME }}
           password: ${{ secrets.DOCKER_PASSWORD }}
 
       - name: "Docker meta"
         id: docker_meta
-        uses: docker/metadata-action@v6
+        uses: docker/metadata-action@dc802804100637a589fabce1cb79ff13a1411302 # v6
         with:
           images: |
             ${{ env.DOCKER_ORG }}/runner
@@ -92,7 +108,7 @@ jobs:
             suffix=${{ matrix.variant.tag_suffix }},onlatest=true
 
       - name: Build and push
-        uses: docker/build-push-action@v7
+        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7
         with:
           context: .
           file: ./Dockerfile

.gitea/workflows/test.yml

@@ -17,8 +17,8 @@ jobs:
       # to ~/.docker with the stale credentials.
       DOCKER_CONFIG: /tmp/docker-noauth
     steps:
-      - uses: actions/checkout@v7
-      - uses: actions/setup-go@v6
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
+      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
         with:
           go-version-file: 'go.mod'
       - name: prepare anonymous docker config
@@ -33,6 +33,8 @@ jobs:
           done
       - name: lint
         run: make lint
+      - name: checks
+        run: make checks
       - name: build
         run: make build
       - name: test

.gitignore

@@ -2,6 +2,7 @@
 .env
 !/act/runner/testdata/secrets/.env
 .runner
+.runner.lock
 coverage.txt
 .tmp/
 /config.yaml

.golangci.yml

@@ -11,6 +11,7 @@ linters:
     - dupl
     - errcheck
     - forbidigo
+    - forcetypeassert
     - gocheckcompilerdirectives
     - gocritic
     - goheader
@@ -102,6 +103,9 @@ linters:
       - linters:
           - forbidigo
         path: cmd
+      - linters:
+          - forcetypeassert
+        path: _test\.go
 issues:
   max-issues-per-linter: 0
   max-same-issues: 0

.goreleaser.yaml

@@ -93,6 +93,37 @@ blobs:
       - glob: ./**.xz
       - glob: ./**.sha256
 
+# Mirrors the S3 `blobs:` upload above into Cloudflare R2 during the
+# parallel S3+R2 period (S3 will be removed once migration completes).
+# A second `blobs:` entry is impossible here since the blob pipe
+# authenticates from the global AWS_* env with no per-entry
+# credentials; `publishers:` supports per-entry `env:` instead, so
+# it's used to invoke scripts/upload-r2.sh once per artifact. Custom
+# publishers inherit almost nothing from the environment, hence the
+# explicit R2_* forwarding below.
+#
+# This publisher fires 109 times for 73 distinct keys because
+# goreleaser's release pipe already registers `release.extra_files`
+# as UploadableFile artifacts, and `internal/exec`'s filterArtifacts
+# appends this block's own extra_files with no de-duplication. It
+# can't be globbed away, since gobwas/glob (via goreleaser/fileglob)
+# has no substring-exclusion matcher. It's harmless: PUT is
+# idempotent, and the `./**.xz` glob below is kept deliberately so
+# this publisher declares its own complete file set rather than
+# implicitly depending on the `release:` block's globs.
+publishers:
+  - name: cloudflare-r2
+    checksum: true
+    extra_files:
+      - glob: ./**.xz
+      - glob: ./**.sha256
+    cmd: sh scripts/upload-r2.sh {{ abs .ArtifactPath }} gitea-runner/{{ .Version }}/{{ .ArtifactName }}
+    env:
+      - R2_ENDPOINT={{ index .Env "R2_ENDPOINT" }}
+      - R2_BUCKET={{ index .Env "R2_BUCKET" }}
+      - R2_ACCESS_KEY_ID={{ index .Env "R2_ACCESS_KEY_ID" }}
+      - R2_SECRET_ACCESS_KEY={{ index .Env "R2_SECRET_ACCESS_KEY" }}
+
 archives:
   - format: binary
     name_template: "{{ .Binary }}"

Dockerfile

@@ -17,7 +17,7 @@ RUN make clean && make build
 ### DIND VARIANT
 #
 #
-FROM docker:29.6.1-dind AS dind
+FROM docker:29.6.2-dind AS dind
 
 ARG VERSION=dev
 
@@ -37,7 +37,7 @@ ENTRYPOINT ["s6-svscan","/etc/s6"]
 ### DIND-ROOTLESS VARIANT
 #
 #
-FROM docker:29.6.1-dind-rootless AS dind-rootless
+FROM docker:29.6.2-dind-rootless AS dind-rootless
 
 ARG VERSION=dev
 

Makefile

@@ -4,9 +4,9 @@ DIST_DIRS := $(DIST)/binaries $(DIST)/release
 GO ?= go
 SHASUM ?= shasum -a 256
 HAS_GO = $(shell hash $(GO) > /dev/null 2>&1 && echo "GO" || echo "NOGO" )
-XGO_PACKAGE ?= src.techknowlogick.com/xgo@latest
+XGO_PACKAGE ?= src.techknowlogick.com/xgo@v1.9.0 # renovate: datasource=go
 XGO_VERSION := go-1.26.x
-GXZ_PACKAGE ?= github.com/ulikunitz/xz/cmd/[email protected]0
+GXZ_PACKAGE ?= github.com/ulikunitz/xz/cmd/[email protected]5 # renovate: datasource=go
 
 LINUX_ARCHS ?= linux/amd64,linux/arm64
 DARWIN_ARCHS ?= darwin-12/amd64,darwin-12/arm64
@@ -18,8 +18,10 @@ 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]
-GOVULNCHECK_PACKAGE ?= golang.org/x/vuln/cmd/[email protected]
+GOLANGCI_LINT_PACKAGE ?= github.com/golangci/golangci-lint/v2/cmd/[email protected] # renovate: datasource=go
+GOVULNCHECK_PACKAGE ?= golang.org/x/vuln/cmd/[email protected] # renovate: datasource=go
+
+GOTEST_FLAGS ?= -race -timeout 20m -parallel 8
 
 STATIC ?=
 EXTLDFLAGS ?=
@@ -95,7 +97,7 @@ go-check:
 
 .PHONY: fmt-check
 fmt-check: fmt
-	@diff=$$(git diff --color=always); \
+	@diff=$$(git diff --color=always -- '*.go'); \
 	if [ -n "$$diff" ]; then \
 		echo "Please run 'make fmt' and commit the result:"; \
 		printf "%s" "$${diff}"; \
@@ -110,6 +112,9 @@ deps-tools: ## install tool dependencies
 	$(GO) install $(GOVULNCHECK_PACKAGE) & \
 	wait
 
+.PHONY: checks
+checks: tidy-check fmt-check security-check ## run the non-lint source checks
+
 .PHONY: lint
 lint: lint-go lint-go-windows ## lint everything
 
@@ -131,7 +136,7 @@ lint-pr-title: ## lint PR title against Conventional Commits (set PR_TITLE=...)
 	@node ./tools/lint-pr-title.ts
 
 .PHONY: security-check
-security-check: deps-tools
+security-check:
 	GOEXPERIMENT= $(GO) run $(GOVULNCHECK_PACKAGE) -show color ./... || true
 
 .PHONY: tidy
@@ -148,8 +153,8 @@ tidy-check: tidy
 	fi
 
 .PHONY: test
-test: fmt-check security-check ## test everything (integration tests self-skip without docker/network)
-	@$(GO) test -race -timeout 20m -v -cover -coverprofile coverage.txt ./... && echo "\n==>\033[32m Ok\033[m\n" || exit 1
+test: ## test everything (integration tests self-skip without docker/network)
+	@$(GO) test $(GOTEST_FLAGS) -cover -coverprofile coverage.txt ./... && echo "\n==>\033[32m Ok\033[m\n" || exit 1
 
 .PHONY: coverage-report
 coverage-report: ## turn coverage.txt from `make test` into .tmp/coverage.md

README.md

@@ -209,10 +209,50 @@ Whenever the resulting labels differ from the ones in the registration file, the
 
 > **Note:** A runner that only exposes `host` labels still needs access to a Docker daemon (e.g. a mounted `/var/run/docker.sock`) whenever a job uses a `docker://` action or a service container. `host` labels only change where the job's own steps run; container-based steps and actions are still executed with Docker.
 
+#### Proxy
+
+Set these variables in the runner's environment, with systemd `Environment=`, `docker run -e`, or Kubernetes `env:`:
+
+```sh
+http_proxy=http://proxy.example:3128
+https_proxy=http://proxy.example:3128
+no_proxy=gitea.internal,.example.local
+```
+
+The runner uses them for its own requests and gives them to every job, in lower and upper case.
+
+These hosts are added to `no_proxy` for jobs, so they are always reached directly:
+
+- the cache server
+- `localhost`, `127.0.0.1` and `::1`
+- the job's service containers
+- the Docker daemon, when it is reached over `tcp://`
+
+Gitea is not added. Add it to `no_proxy` yourself if it should be reached directly.
+
+To change a value for one job, set it in a step's `env:` or in the job's `container.env`. Setting it at workflow or job level has no effect. To change it for the whole runner, set it in `runner.envs`. A `no_proxy` set there is added to the list above instead of replacing it.
+
+Images are pulled by the Docker daemon, which needs its own proxy setting. In the `dind` images the daemon runs in the same container and reads the variables above. For any other daemon, see [the Docker documentation](https://docs.docker.com/engine/daemon/proxy/). The runner logs a warning at startup if it has a proxy and the daemon does not.
+
+Dockerfile actions are built with these variables as build arguments, so their `RUN` steps can reach the network.
+
+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.
+
 #### 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.
 
+**Cache service v2**
+
+`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
+cache:
+  v2: false
+```
+
+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 and keeps the untouched copy beside it; 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.
+
 **Shared cache across multiple runners**
 
 Run one dedicated `gitea-runner cache-server` that all runners point at.
@@ -224,6 +264,7 @@ Run one dedicated `gitea-runner cache-server` that all runners point at.
      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
    ```
 
 2. Start the server:
@@ -238,6 +279,7 @@ Run one dedicated `gitea-runner cache-server` that all runners point at.
    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
    ```
 
 Alternatively, mount the same NFS/CIFS share on every runner and point `cache.dir` at it — simpler, but with weaker isolation between repositories.
@@ -252,11 +294,12 @@ Besides `GITEA_INSTANCE_URL` and `GITEA_RUNNER_REGISTRATION_TOKEN`, the image en
 
 For a fuller container-oriented walkthrough, see [examples/docker](examples/docker/README.md).
 
-When `container.bind_workdir` is enabled, stale task workspace directories can be cleaned while the runner is idle:
-- directories older than `runner.workdir_cleanup_age` are removed (default: `24h`; set `0` to disable)
-- cleanup runs every `runner.idle_cleanup_interval` (default: `10m`; set `0` to disable)
+While the runner is idle it cleans up after earlier jobs:
+- when `container.bind_workdir` is enabled, stale task workspace directories older than `runner.workdir_cleanup_age` are removed (default: `24h`; set `0` to disable)
 - only purely numeric subdirectories under `container.workdir_parent` are treated as task workspaces and may be removed
 - cleanup assumes `container.workdir_parent` is not shared across multiple runners
+- on runners that use docker, per-job networks left behind by jobs the runner did not live to tear down are removed, identified by the `com.gitea.runner.uuid` label carrying this runner's uuid
+- cleanup runs every `runner.idle_cleanup_interval` (default: `10m`; set `0` to disable), and setting either knob to `0` disables all of the above
 
 #### Post-task script (`runner.post_task_script`)
 
@@ -268,6 +311,16 @@ On Windows, use `.exe`, `.bat`, or `.cmd` paths; **PowerShell (`.ps1`) is not su
 
 See **[docs/post-task-script.md](docs/post-task-script.md)** for lifecycle details, environment variables, timeout interaction, and platform notes.
 
+#### Job hooks (`runner.hooks.job_started`, `runner.hooks.job_completed`)
+
+Optional scripts that run **inside the job environment** (the job container, or the host in host mode), before the job's first step and after its last one. They are the equivalent of GitHub's `ACTIONS_RUNNER_HOOK_JOB_STARTED` / `ACTIONS_RUNNER_HOOK_JOB_COMPLETED`, which are read when the settings are unset.
+
+Because they run where the steps run and see the job's environment, they are the place for per-job setup no workflow should have to carry: registry logins, mirror configuration, or masking runner-wide secrets with `::add-mask::`. Their output is part of the job log and is scanned for workflow commands, and they can export to the job through `$GITHUB_ENV` and `$GITHUB_PATH`.
+
+Both hooks are synchronous and block the job while they run. Either one exiting non-zero fails the job, and there is no per-hook timeout.
+
+See **[docs/job-hooks.md](docs/job-hooks.md)** for the execution order, environment, and platform notes.
+
 ### Example Deployments
 
 Check out the [examples](examples) directory for sample deployment types.

act/artifactcache/handler.go

@@ -70,6 +70,7 @@ type Handler struct {
 	storage  *Storage
 	router   *httprouter.Router
 	listener net.Listener
+	port     int
 	server   *http.Server
 	logger   logrus.FieldLogger
 
@@ -157,12 +158,13 @@ func StartHandler(dir, outboundIP string, port uint16, internalSecret string, lo
 	router.POST(apiPath+"/clean", h.bearerAuth(h.clean))
 	// Artifact GET is signed via query-string HMAC because @actions/cache
 	// does not attach Authorization when downloading archiveLocation.
-	router.GET(apiPath+"/artifacts/:id", h.signedURLAuth(h.get))
+	router.GET(apiPath+"/artifacts/:id", h.signedAuth("", h.get))
 	// Control-plane: a remote runner registers/revokes per-job tokens so the
 	// cache API can authenticate them. Always wired so the routes exist; the
 	// handlers themselves 401 when internalSecret is unset.
 	router.POST(internalPath+"/register", h.internalAuth(h.internalRegister))
 	router.POST(internalPath+"/revoke", h.internalAuth(h.internalRevoke))
+	h.registerV2Routes(router)
 
 	h.router = router
 
@@ -177,6 +179,12 @@ func StartHandler(dir, outboundIP string, port uint16, internalSecret string, lo
 	if err != nil {
 		return nil, 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())
+	}
+	h.port = addr.Port
 	server := &http.Server{
 		ReadHeaderTimeout: 2 * time.Second,
 		Handler:           router,
@@ -194,9 +202,7 @@ func StartHandler(dir, outboundIP string, port uint16, internalSecret string, lo
 
 func (h *Handler) ExternalURL() string {
 	// TODO: make the external url configurable if necessary
-	return fmt.Sprintf("http://%s:%d",
-		h.outboundIP,
-		h.listener.Addr().(*net.TCPAddr).Port)
+	return fmt.Sprintf("http://%s:%d", h.outboundIP, h.port)
 }
 
 // RegisterJob makes token a valid bearer credential for cache requests from
@@ -334,7 +340,7 @@ func (h *Handler) find(w http.ResponseWriter, r *http.Request, _ httprouter.Para
 	}
 	defer db.Close()
 
-	cache, err := findCache(db, cred.Repo, keys, version)
+	cache, err := h.lookupCache(db, cred.Repo, keys, version)
 	if err != nil {
 		h.responseJSON(w, r, 500, err)
 		return
@@ -343,15 +349,6 @@ func (h *Handler) find(w http.ResponseWriter, r *http.Request, _ httprouter.Para
 		h.responseJSON(w, r, 204)
 		return
 	}
-
-	if ok, err := h.storage.Exist(cache.ID); err != nil {
-		h.responseJSON(w, r, 500, err)
-		return
-	} else if !ok {
-		_ = db.Delete(cache.ID, cache)
-		h.responseJSON(w, r, 204)
-		return
-	}
 	h.responseJSON(w, r, 200, map[string]any{
 		"result":          "hit",
 		"archiveLocation": h.signedArtifactURL(cache.ID, time.Now().Add(artifactURLTTL)),
@@ -359,6 +356,25 @@ func (h *Handler) find(w http.ResponseWriter, r *http.Request, _ httprouter.Para
 	})
 }
 
+// lookupCache returns the entry to restore for these keys, or (nil, nil) when there is none:
+// either nothing matched, or the match had lost its blob to a prune, in which case the dangling
+// entry is dropped on the way out.
+func (h *Handler) lookupCache(db *bolthold.Store, repo string, keys []string, version string) (*Cache, error) {
+	cache, err := findCache(db, repo, keys, version)
+	if err != nil || cache == nil {
+		return nil, err
+	}
+	ok, err := h.storage.Exist(cache.ID)
+	if err != nil {
+		return nil, err
+	}
+	if !ok {
+		_ = db.Delete(cache.ID, cache)
+		return nil, nil //nolint:nilnil // absence is not an error here
+	}
+	return cache, nil
+}
+
 // POST /_apis/artifactcache/caches
 func (h *Handler) reserve(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
 	cred := credFromContext(r.Context())
@@ -433,7 +449,7 @@ func (h *Handler) upload(w http.ResponseWriter, r *http.Request, params httprout
 		h.responseJSON(w, r, 500, err)
 		return
 	}
-	h.useCache(id)
+	_ = h.touchCache(uint64(id), false)
 	h.responseJSON(w, r, 200)
 }
 
@@ -474,23 +490,7 @@ func (h *Handler) commit(w http.ResponseWriter, r *http.Request, params httprout
 
 	db.Close()
 
-	size, err := h.storage.Commit(cache.ID, cache.Size)
-	if err != nil {
-		h.responseJSON(w, r, 500, err)
-		return
-	}
-	// write real size back to cache, it may be different from the current value when the request doesn't specify it.
-	cache.Size = size
-
-	db, err = h.openDB()
-	if err != nil {
-		h.responseJSON(w, r, 500, err)
-		return
-	}
-	defer db.Close()
-
-	cache.Complete = true
-	if err := db.Update(cache.ID, cache); err != nil {
+	if err := h.commitCache(cache); err != nil {
 		h.responseJSON(w, r, 500, err)
 		return
 	}
@@ -498,8 +498,28 @@ func (h *Handler) commit(w http.ResponseWriter, r *http.Request, params httprout
 	h.responseJSON(w, r, 200)
 }
 
+// commitCache assembles the uploaded parts and marks the entry complete. The caller must
+// have closed its store first: Commit concatenates the whole archive and would otherwise
+// hold bolt's exclusive file lock for the duration.
+func (h *Handler) commitCache(cache *Cache) error {
+	written, err := h.storage.Commit(cache.ID, cache.Size)
+	if err != nil {
+		return err
+	}
+	// write real size back to cache, it may be different from the current value when the request doesn't specify it.
+	cache.Size = written
+	cache.Complete = true
+
+	db, err := h.openDB()
+	if err != nil {
+		return err
+	}
+	defer db.Close()
+	return db.Update(cache.ID, cache)
+}
+
 // GET /_apis/artifactcache/artifacts/:id
-// Authenticated via signed URL (see signedURLAuth), not bearer, because the
+// Authenticated via signed URL (see signedAuth), not bearer, because the
 // @actions/cache toolkit downloads archiveLocation without Authorization.
 // Repository scoping is already enforced at find() time; the signature binds
 // the URL to the specific cache ID and an expiry.
@@ -509,7 +529,7 @@ func (h *Handler) get(w http.ResponseWriter, r *http.Request, params httprouter.
 		h.responseJSON(w, r, 400, err)
 		return
 	}
-	h.useCache(id)
+	_ = h.touchCache(uint64(id), false)
 	h.storage.Serve(w, r, uint64(id))
 }
 
@@ -543,7 +563,9 @@ func (h *Handler) bearerAuth(handler httprouter.Handle) httprouter.Handle {
 	}
 }
 
-func (h *Handler) signedURLAuth(handler httprouter.Handle) httprouter.Handle {
+// 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 {
 	return func(w http.ResponseWriter, r *http.Request, params httprouter.Params) {
 		h.logger.Debugf("%s %s", r.Method, r.URL.Path)
 		id, err := strconv.ParseInt(params.ByName("id"), 10, 64)
@@ -566,7 +588,7 @@ func (h *Handler) signedURLAuth(handler httprouter.Handle) httprouter.Handle {
 			h.responseJSON(w, r, http.StatusUnauthorized, errors.New("signature expired"))
 			return
 		}
-		expected := h.computeSignature(id, exp)
+		expected := h.computeSignature(purpose, id, exp)
 		if !hmac.Equal([]byte(sig), []byte(expected)) {
 			h.responseJSON(w, r, http.StatusUnauthorized, errors.New("bad signature"))
 			return
@@ -650,19 +672,26 @@ func credFromContext(ctx context.Context) JobCredential {
 	return JobCredential{}
 }
 
-func (h *Handler) computeSignature(cacheID, exp int64) string {
+// computeSignature signs a URL for one cache entry and expiry. purpose is mixed into the
+// message so a URL handed out for writing an entry cannot be replayed to read one, and the
+// other way round. Downloads use the empty purpose, the message v1 has always signed.
+func (h *Handler) computeSignature(purpose string, cacheID, exp int64) string {
 	mac := hmac.New(sha256.New, h.secret)
-	fmt.Fprintf(mac, "%d:%d", cacheID, exp)
+	fmt.Fprintf(mac, "%s%d:%d", purpose, cacheID, exp)
 	return hex.EncodeToString(mac.Sum(nil))
 }
 
-func (h *Handler) signedArtifactURL(cacheID uint64, exp time.Time) string {
+// signedURL builds a URL under path that signedAuth accepts for the same purpose.
+func (h *Handler) signedURL(path, purpose string, cacheID uint64, exp time.Time) string {
 	expUnix := exp.Unix()
-	sig := h.computeSignature(int64(cacheID), expUnix)
 	q := url.Values{}
 	q.Set("exp", strconv.FormatInt(expUnix, 10))
-	q.Set("sig", sig)
-	return fmt.Sprintf("%s%s/artifacts/%d?%s", h.ExternalURL(), apiPath, cacheID, q.Encode())
+	q.Set("sig", h.computeSignature(purpose, int64(cacheID), expUnix))
+	return fmt.Sprintf("%s%s/%d?%s", h.ExternalURL(), path, cacheID, q.Encode())
+}
+
+func (h *Handler) signedArtifactURL(cacheID uint64, exp time.Time) string {
+	return h.signedURL(apiPath+"/artifacts", "", cacheID, exp)
 }
 
 // if not found, return (nil, nil) instead of an error.
@@ -670,16 +699,12 @@ func findCache(db *bolthold.Store, repo string, keys []string, version string) (
 	cache := &Cache{}
 	for _, prefix := range keys {
 		// if a key in the list matches exactly, don't return partial matches
-		if err := db.FindOne(cache,
-			bolthold.Where("Repo").Eq(repo).
-				And("Key").Eq(prefix).
-				And("Version").Eq(version).
-				And("Complete").Eq(true).
-				SortBy("CreatedAt").Reverse()); err == nil || !errors.Is(err, bolthold.ErrNotFound) {
-			if err != nil {
-				return nil, fmt.Errorf("find cache: %w", err)
-			}
-			return cache, nil
+		exact, err := findExactCache(db, repo, prefix, version, true)
+		if err != nil {
+			return nil, err
+		}
+		if exact != nil {
+			return exact, nil
 		}
 		prefixPattern := "^" + regexp.QuoteMeta(prefix)
 		re, err := regexp.Compile(prefixPattern)
@@ -702,6 +727,34 @@ func findCache(db *bolthold.Store, repo string, keys []string, version string) (
 	return nil, nil //nolint:nilnil // pre-existing issue from nektos/act
 }
 
+// findExactCache returns the entry for exactly this key and version, or (nil, nil) if there is
+// none. Unlike findCache it never falls back to a prefix (restore-key) match, which is what both
+// its callers need: a new key that is only a prefix of an existing key is not the same entry.
+//
+// A completed entry is the one to restore, sorted by when it was written. An incomplete one is a
+// reservation being uploaded to, sorted by when it was last written to, because the upload route
+// touches UsedAt on every part.
+func findExactCache(db *bolthold.Store, repo, key, version string, complete bool) (*Cache, error) {
+	sortBy := "UsedAt"
+	if complete {
+		sortBy = "CreatedAt"
+	}
+	cache := &Cache{}
+	err := db.FindOne(cache,
+		bolthold.Where("Repo").Eq(repo).
+			And("Key").Eq(key).
+			And("Version").Eq(version).
+			And("Complete").Eq(complete).
+			SortBy(sortBy).Reverse())
+	if errors.Is(err, bolthold.ErrNotFound) {
+		return nil, nil //nolint:nilnil // absence is not an error here
+	}
+	if err != nil {
+		return nil, fmt.Errorf("find cache: %w", err)
+	}
+	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)
@@ -713,18 +766,30 @@ func insertCache(db *bolthold.Store, cache *Cache) error {
 	return nil
 }
 
-func (h *Handler) useCache(id int64) {
+// touchCache stamps UsedAt so gcCache does not reap an entry mid-upload. With requireIncomplete
+// it also refuses an entry that is already complete, which is what the v2 blob route needs: its
+// upload URL outlives the finalize call, and overwriting a finished entry would leave the blob
+// other jobs restore no longer matching its recorded size. An entry missing from the store is
+// accepted, since the signature proves the id was handed out.
+func (h *Handler) touchCache(id uint64, requireIncomplete bool) error {
 	db, err := h.openDB()
 	if err != nil {
-		return
+		return err
 	}
 	defer db.Close()
+
 	cache := &Cache{}
 	if err := db.Get(id, cache); err != nil {
-		return
+		if errors.Is(err, bolthold.ErrNotFound) {
+			return nil
+		}
+		return err
+	}
+	if requireIncomplete && cache.Complete {
+		return fmt.Errorf("cache %d: already complete", id)
 	}
 	cache.UsedAt = time.Now().Unix()
-	_ = db.Update(cache.ID, cache)
+	return db.Update(cache.ID, cache)
 }
 
 const (

act/artifactcache/handler_test.go

@@ -445,13 +445,6 @@ func TestHandler(t *testing.T) {
 		require.Equal(t, 404, resp.StatusCode)
 	})
 
-	t.Run("get with not exist id", func(t *testing.T) {
-		resp, err := testClient.Get(signArtifactURL(handler, 100))
-		require.NoError(t, err)
-		defer resp.Body.Close()
-		require.Equal(t, 404, resp.StatusCode)
-	})
-
 	t.Run("get with multiple keys", func(t *testing.T) {
 		version := "c19da02a2bd7e77277f1ac29ab45c09b7d46a4ee758284e26bb3045ad11d9d20"
 		key := strings.ToLower(t.Name())
@@ -469,7 +462,8 @@ func TestHandler(t *testing.T) {
 			_, err := rand.Read(contents[i])
 			require.NoError(t, err)
 			uploadCacheNormally(t, base, keys[i], version, contents[i])
-			time.Sleep(time.Second) // ensure CreatedAt of caches are different
+			// ensure CreatedAt of caches are different, in upload order
+			backdateCache(t, handler, keys[i], time.Duration(len(contents)-i)*time.Second)
 		}
 
 		reqKeys := strings.Join([]string{
@@ -554,7 +548,8 @@ func TestHandler(t *testing.T) {
 			_, err := rand.Read(contents[i])
 			require.NoError(t, err)
 			uploadCacheNormally(t, base, keys[i], version, contents[i])
-			time.Sleep(time.Second) // ensure CreatedAt of caches are different
+			// ensure CreatedAt of caches are different, in upload order
+			backdateCache(t, handler, keys[i], time.Duration(len(contents)-i)*time.Second)
 		}
 
 		reqKeys := strings.Join([]string{
@@ -607,7 +602,8 @@ func TestHandler(t *testing.T) {
 			_, err := rand.Read(contents[i])
 			require.NoError(t, err)
 			uploadCacheNormally(t, base, keys[i], version, contents[i])
-			time.Sleep(time.Second) // ensure CreatedAt of caches are different
+			// ensure CreatedAt of caches are different, in upload order
+			backdateCache(t, handler, keys[i], time.Duration(len(contents)-i)*time.Second)
 		}
 
 		reqKeys := strings.Join([]string{
@@ -646,6 +642,20 @@ func TestHandler(t *testing.T) {
 	})
 }
 
+// backdateCache rewrites a cache's CreatedAt. It has one-second resolution, so age-ordering
+// tests set it directly instead of sleeping a second between uploads.
+func backdateCache(t *testing.T, handler *Handler, key string, age time.Duration) {
+	db, err := handler.openDB()
+	require.NoError(t, err)
+	defer db.Close()
+
+	var caches []*Cache
+	require.NoError(t, db.Find(&caches, bolthold.Where("Key").Eq(key)))
+	require.Len(t, caches, 1)
+	caches[0].CreatedAt = time.Now().Add(-age).Unix()
+	require.NoError(t, db.Update(caches[0].ID, caches[0]))
+}
+
 func uploadCacheNormally(t *testing.T, base, key, version string, content []byte) { //nolint:unparam // pre-existing issue from nektos/act
 	var id uint64
 	{
@@ -1031,14 +1041,14 @@ func TestHandler_SecretPersistsAcrossRestarts(t *testing.T) {
 	first, err := StartHandler(dir, "127.0.0.1", 0, "", nil)
 	require.NoError(t, err)
 	exp := time.Now().Add(artifactURLTTL).Unix()
-	sig := first.computeSignature(42, exp)
+	sig := first.computeSignature("", 42, exp)
 	require.NoError(t, first.Close())
 
 	second, err := StartHandler(dir, "127.0.0.1", 0, "", nil)
 	require.NoError(t, err)
 	defer second.Close()
 
-	assert.Equal(t, sig, second.computeSignature(42, exp))
+	assert.Equal(t, sig, second.computeSignature("", 42, exp))
 }
 
 // TestHandler_ArtifactSignatureDownload is a happy-path round trip that

act/artifactcache/handler_v2.go

@@ -0,0 +1,268 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"cmp"
+	"encoding/json"
+	"encoding/xml"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/julienschmidt/httprouter"
+)
+
+// The cache service v2 API. A client on this version talks twirp to
+// `github.actions.results.api.v1.CacheService` instead of the /_apis/artifactcache
+// endpoints, and uploads the archive to the returned URL with the Azure blob protocol.
+// Both API versions are served from the same store, so a repository keeps its cache
+// when a workflow moves between action versions.
+const (
+	cacheServiceV2Path = "/twirp/github.actions.results.api.v1.CacheService"
+
+	// blobPath authenticates by signature, because the client uploads without an
+	// Authorization header. Downloads are handed the v1 artifact URL instead.
+	blobPath = apiPath + "/blobs"
+
+	// blobUploadPurpose keeps an upload URL from being replayed to read an entry.
+	blobUploadPurpose = "upload:"
+
+	blobUploadURLTTL = time.Hour
+
+	// twirpInternal is the only error code that is not the client's fault.
+	twirpInternal = "internal"
+)
+
+func (h *Handler) registerV2Routes(router *httprouter.Router) {
+	router.POST(cacheServiceV2Path+"/CreateCacheEntry", h.bearerAuth(h.v2CreateCacheEntry))
+	router.POST(cacheServiceV2Path+"/FinalizeCacheEntryUpload", h.bearerAuth(h.v2FinalizeCacheEntryUpload))
+	router.POST(cacheServiceV2Path+"/GetCacheEntryDownloadURL", h.bearerAuth(h.v2GetCacheEntryDownloadURL))
+	router.PUT(blobPath+"/:id", h.signedAuth(blobUploadPurpose, h.v2UploadBlob))
+}
+
+// An entry that already exists is reported as not ok, which is how the client learns to skip
+// the upload.
+func (h *Handler) v2CreateCacheEntry(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
+	cred := credFromContext(r.Context())
+	req, err := decodeTwirpRequest[v2CreateRequest](r)
+	if err != nil {
+		h.twirpError(w, r, "malformed_request", err)
+		return
+	}
+	if req.Key == "" || req.Version == "" {
+		h.twirpError(w, r, "invalid_argument", errors.New("key and version are required"))
+		return
+	}
+
+	db, err := h.openDB()
+	if err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	}
+	defer db.Close()
+
+	// An exact (key, version) match means the entry is already cached; the client then skips
+	// the upload. A prefix match must not count here, or a shorter key would be reported as
+	// existing and silently never saved.
+	if existing, err := findExactCache(db, cred.Repo, req.Key, req.Version, true); err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	} else if existing != nil {
+		h.twirpNotOK(w, r)
+		return
+	}
+
+	now := time.Now().Unix()
+	cache := &Cache{
+		Repo:      cred.Repo,
+		Key:       req.Key,
+		Version:   req.Version,
+		Size:      -1, // the size is only known at finalize time
+		CreatedAt: now,
+		UsedAt:    now,
+	}
+	if err := insertCache(db, cache); err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	}
+
+	h.responseJSON(w, r, http.StatusOK, map[string]any{
+		"ok":              true,
+		"signedUploadUrl": h.signedURL(blobPath, blobUploadPurpose, cache.ID, time.Now().Add(blobUploadURLTTL)),
+	})
+}
+
+func (h *Handler) v2FinalizeCacheEntryUpload(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
+	cred := credFromContext(r.Context())
+	req, err := decodeTwirpRequest[v2FinalizeRequest](r)
+	if err != nil {
+		h.twirpError(w, r, "malformed_request", err)
+		return
+	}
+
+	db, err := h.openDB()
+	if err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	}
+	defer db.Close()
+
+	cache, err := findExactCache(db, cred.Repo, req.Key, req.Version, false)
+	if err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	}
+	if cache == nil {
+		h.twirpNotOK(w, r)
+		return
+	}
+	db.Close() // commitCache needs the store closed
+
+	cache.Size, _ = cmp.Or(req.SizeBytes, req.SizeBytesCamel).Int64()
+	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{
+		"ok": true,
+		// int64 fields travel as strings in the proto JSON mapping.
+		"entryId": strconv.FormatUint(cache.ID, 10),
+	})
+}
+
+func (h *Handler) v2GetCacheEntryDownloadURL(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
+	cred := credFromContext(r.Context())
+	req, err := decodeTwirpRequest[v2DownloadRequest](r)
+	if err != nil {
+		h.twirpError(w, r, "malformed_request", err)
+		return
+	}
+
+	db, err := h.openDB()
+	if err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	}
+	defer db.Close()
+
+	cache, err := h.lookupCache(db, cred.Repo, req.keys(), req.Version)
+	if err != nil {
+		h.twirpError(w, r, twirpInternal, err)
+		return
+	}
+	if cache == nil {
+		h.twirpNotOK(w, r)
+		return
+	}
+
+	h.responseJSON(w, r, http.StatusOK, map[string]any{
+		"ok":                true,
+		"signedDownloadUrl": h.signedArtifactURL(cache.ID, time.Now().Add(artifactURLTTL)),
+		"matchedKey":        cache.Key,
+	})
+}
+
+// The archive arrives over the subset of the Azure blob API the toolkit uses: a small
+// cache is a single PUT, a large one is staged as blocks that a final block list puts
+// in order.
+func (h *Handler) v2UploadBlob(w http.ResponseWriter, r *http.Request, params httprouter.Params) {
+	id, err := strconv.ParseUint(params.ByName("id"), 10, 64)
+	if err != nil {
+		h.responseJSON(w, r, http.StatusBadRequest, err)
+		return
+	}
+
+	if err := h.touchCache(id, true); err != nil {
+		h.responseJSON(w, r, http.StatusBadRequest, err)
+		return
+	}
+
+	query := r.URL.Query()
+	switch strings.ToLower(query.Get("comp")) {
+	case "block":
+		blockID := query.Get("blockid")
+		if blockID == "" {
+			h.responseJSON(w, r, http.StatusBadRequest, errors.New("missing blockid"))
+			return
+		}
+		err = h.storage.WriteBlock(id, blockID, r.Body)
+	case "blocklist":
+		var list struct{ Latest []string }
+		if err := xml.NewDecoder(io.LimitReader(r.Body, 8<<20)).Decode(&list); err != nil {
+			h.responseJSON(w, r, http.StatusBadRequest, fmt.Errorf("malformed block list: %w", err))
+			return
+		}
+		err = h.storage.OrderBlocks(id, list.Latest)
+	default:
+		err = h.storage.Write(id, 0, r.Body)
+	}
+	if err != nil {
+		h.responseJSON(w, r, http.StatusInternalServerError, err)
+		return
+	}
+
+	w.WriteHeader(http.StatusCreated)
+}
+
+// twirpNotOK is the negative answer all three endpoints share: no such entry to restore, no
+// reservation to finalize, or an entry that already exists and need not be uploaded again.
+func (h *Handler) twirpNotOK(w http.ResponseWriter, r *http.Request) {
+	h.responseJSON(w, r, http.StatusOK, map[string]any{"ok": false})
+}
+
+// twirpError reports in the shape a twirp client expects, so the toolkit surfaces the message
+// instead of a parse error.
+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 {
+		status = http.StatusInternalServerError
+	}
+	h.responseJSON(w, r, status, map[string]any{"code": code, "msg": err.Error()})
+}
+
+// The twirp request bodies. The toolkit's client serialises with useProtoFieldName, so the proto
+// names are what arrive; the camelCase spellings of the same mapping are accepted too, as are
+// int64s sent as a bare number rather than the string the mapping prescribes.
+type (
+	v2CreateRequest struct {
+		Key     string `json:"key"`
+		Version string `json:"version"`
+	}
+
+	v2FinalizeRequest struct {
+		Key            string      `json:"key"`
+		Version        string      `json:"version"`
+		SizeBytes      json.Number `json:"size_bytes"`
+		SizeBytesCamel json.Number `json:"sizeBytes"`
+	}
+
+	v2DownloadRequest struct {
+		Key              string   `json:"key"`
+		Version          string   `json:"version"`
+		RestoreKeys      []string `json:"restore_keys"`
+		RestoreKeysCamel []string `json:"restoreKeys"`
+	}
+)
+
+func (d v2DownloadRequest) keys() []string {
+	restoreKeys := d.RestoreKeys
+	if len(restoreKeys) == 0 {
+		restoreKeys = d.RestoreKeysCamel
+	}
+	return append([]string{d.Key}, restoreKeys...)
+}
+
+func decodeTwirpRequest[T any](r *http.Request) (T, error) {
+	var req T
+	err := json.NewDecoder(io.LimitReader(r.Body, 1<<20)).Decode(&req)
+	return req, err
+}

act/artifactcache/handler_v2_test.go

@@ -0,0 +1,236 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"bytes"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"path/filepath"
+	"strconv"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// v2Call posts a twirp request to the cache service and returns the decoded response.
+// Field names are the proto ones, which is what the toolkit's client sends.
+func v2Call(t *testing.T, handler *Handler, client *http.Client, method string, request any) map[string]any {
+	t.Helper()
+
+	body, err := json.Marshal(request)
+	require.NoError(t, err)
+
+	resp, err := client.Post(handler.ExternalURL()+cacheServiceV2Path+"/"+method, "application/json", bytes.NewReader(body))
+	require.NoError(t, err)
+	defer resp.Body.Close()
+	require.Equal(t, http.StatusOK, resp.StatusCode)
+
+	got := map[string]any{}
+	require.NoError(t, json.NewDecoder(resp.Body).Decode(&got))
+	return got
+}
+
+// putBlob uploads to a signed URL and returns the status, so a test can assert a refusal.
+func putBlob(t *testing.T, url string, content []byte) int {
+	t.Helper()
+
+	req, err := http.NewRequestWithContext(t.Context(), http.MethodPut, url, bytes.NewReader(content))
+	require.NoError(t, err)
+	resp, err := http.DefaultClient.Do(req)
+	require.NoError(t, err)
+	defer resp.Body.Close()
+	return resp.StatusCode
+}
+
+func getURL(t *testing.T, url string) []byte {
+	t.Helper()
+
+	req, err := http.NewRequestWithContext(t.Context(), http.MethodGet, url, nil)
+	require.NoError(t, err)
+	resp, err := http.DefaultClient.Do(req)
+	require.NoError(t, err)
+	defer resp.Body.Close()
+	body, err := io.ReadAll(resp.Body)
+	require.NoError(t, err)
+	return body
+}
+
+func startTestHandler(t *testing.T) *Handler {
+	t.Helper()
+
+	handler, err := StartHandler(filepath.Join(t.TempDir(), "artifactcache"), "127.0.0.1", 0, "", nil)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = handler.Close() })
+	handler.RegisterJob(testToken, testRepo)
+	return handler
+}
+
+// saveV2 runs the reserve/upload/finalize sequence and returns the finalize response along
+// with the upload URL it used.
+func saveV2(t *testing.T, handler *Handler, key, version string, content []byte) (finalized map[string]any, uploadURL string) {
+	t.Helper()
+
+	created := v2Call(t, handler, testClient, "CreateCacheEntry", map[string]any{"key": key, "version": version})
+	require.Equal(t, true, created["ok"])
+	uploadURL, _ = created["signedUploadUrl"].(string)
+	require.NotEmpty(t, uploadURL)
+	require.Equal(t, http.StatusCreated, putBlob(t, uploadURL, content))
+
+	return v2Call(t, handler, testClient, "FinalizeCacheEntryUpload", map[string]any{
+		"key": key, "version": version,
+		"size_bytes": strconv.Itoa(len(content)),
+	}), uploadURL
+}
+
+// The whole round trip an actions/cache v2 client makes, plus the guarantees on the signed
+// URLs it is handed: unsigned requests are refused, an upload URL cannot be replayed to read
+// or to replace a finalized entry.
+func TestCacheServiceV2RoundTrip(t *testing.T) {
+	handler := startTestHandler(t)
+	content := []byte("the cached archive")
+
+	unsigned := fmt.Sprintf("%s%s/1", handler.ExternalURL(), blobPath)
+	assert.Equal(t, http.StatusUnauthorized, putBlob(t, unsigned, content))
+
+	finalized, uploadURL := saveV2(t, handler, "deps-v1", "abc123", content)
+	require.Equal(t, true, finalized["ok"])
+	assert.NotEmpty(t, finalized["entryId"])
+
+	// The upload URL outlives the finalize call, so replaying it must not poison the entry,
+	// and it is an upload URL only: nothing reads a blob back through it.
+	assert.Equal(t, http.StatusBadRequest, putBlob(t, uploadURL, []byte("poisoned")))
+	resp, err := http.Get(uploadURL) //nolint:noctx // the URL is the server under test
+	require.NoError(t, err)
+	defer resp.Body.Close()
+	assert.Equal(t, http.StatusMethodNotAllowed, resp.StatusCode)
+
+	got := v2Call(t, handler, testClient, "GetCacheEntryDownloadURL", map[string]any{"key": "deps-v1", "version": "abc123"})
+	require.Equal(t, true, got["ok"])
+	assert.Equal(t, "deps-v1", got["matchedKey"])
+	downloadURL, _ := got["signedDownloadUrl"].(string)
+	require.NotEmpty(t, downloadURL)
+	assert.Equal(t, content, getURL(t, downloadURL))
+}
+
+// A large archive is staged as blocks and only put in order by the final block list, so
+// blocks that arrive out of order must still be assembled the way the client asked.
+func TestCacheServiceV2BlockUpload(t *testing.T) {
+	handler := startTestHandler(t)
+
+	created := v2Call(t, handler, testClient, "CreateCacheEntry", map[string]any{"key": "blocks", "version": "v1"})
+	uploadURL, _ := created["signedUploadUrl"].(string)
+	require.NotEmpty(t, uploadURL)
+
+	blocks := map[string][]byte{}
+	var order []string
+	for i, part := range []string{"hello ", "world", "!"} {
+		blockID := base64.StdEncoding.EncodeToString(fmt.Appendf(nil, "block-%d", i))
+		blocks[blockID] = []byte(part)
+		order = append(order, blockID)
+	}
+	// Upload in an order that is not the block list order.
+	for _, blockID := range []string{order[2], order[0], order[1]} {
+		require.Equal(t, http.StatusCreated, putBlob(t, uploadURL+"&comp=block&blockid="+blockID, blocks[blockID]))
+	}
+
+	var list bytes.Buffer
+	list.WriteString(`<?xml version="1.0" encoding="utf-8"?><BlockList>`)
+	for _, blockID := range order {
+		fmt.Fprintf(&list, "<Latest>%s</Latest>", blockID)
+	}
+	list.WriteString(`</BlockList>`)
+	require.Equal(t, http.StatusCreated, putBlob(t, uploadURL+"&comp=blocklist", list.Bytes()))
+
+	finalized := v2Call(t, handler, testClient, "FinalizeCacheEntryUpload", map[string]any{
+		"key": "blocks", "version": "v1", "size_bytes": len("hello world!"),
+	})
+	require.Equal(t, true, finalized["ok"])
+
+	got := v2Call(t, handler, testClient, "GetCacheEntryDownloadURL", map[string]any{"key": "blocks", "version": "v1"})
+	require.Equal(t, true, got["ok"])
+	assert.Equal(t, "hello world!", string(getURL(t, got["signedDownloadUrl"].(string))))
+}
+
+func TestCacheServiceV2Lookups(t *testing.T) {
+	handler := startTestHandler(t)
+	saved, _ := saveV2(t, handler, "deps-abc", "v1", []byte("x"))
+	require.Equal(t, true, saved["ok"])
+
+	t.Run("reports a miss for an unknown key", func(t *testing.T) {
+		got := v2Call(t, handler, testClient, "GetCacheEntryDownloadURL", map[string]any{"key": "nothing", "version": "v1"})
+		assert.Equal(t, false, got["ok"])
+	})
+
+	// The toolkit serialises with the proto field names; the camelCase spellings of the same
+	// proto JSON mapping are accepted alongside them.
+	for _, field := range []string{"restore_keys", "restoreKeys"} {
+		t.Run("restore keys match by prefix, spelled "+field, func(t *testing.T) {
+			got := v2Call(t, handler, testClient, "GetCacheEntryDownloadURL", map[string]any{
+				"key": "deps-zzz", field: []string{"deps-"}, "version": "v1",
+			})
+			require.Equal(t, true, got["ok"])
+			assert.Equal(t, "deps-abc", got["matchedKey"])
+		})
+	}
+
+	t.Run("an existing entry is not reserved twice", func(t *testing.T) {
+		again := v2Call(t, handler, testClient, "CreateCacheEntry", map[string]any{"key": "deps-abc", "version": "v1"})
+		assert.Equal(t, false, again["ok"])
+	})
+
+	// A key that is only a prefix of an existing one is a different entry, so the
+	// reservation check must be exact and not a restore-key prefix match, or the shorter
+	// key would be reported as existing and silently never saved.
+	t.Run("a prefix of an existing key is still reserved", func(t *testing.T) {
+		reserved := v2Call(t, handler, testClient, "CreateCacheEntry", map[string]any{"key": "deps", "version": "v1"})
+		require.Equal(t, true, reserved["ok"])
+		assert.NotEmpty(t, reserved["signedUploadUrl"])
+	})
+
+	t.Run("finalizing without a reservation is not ok", func(t *testing.T) {
+		got := v2Call(t, handler, testClient, "FinalizeCacheEntryUpload", map[string]any{
+			"key": "never-reserved", "version": "v1", "size_bytes": 1,
+		})
+		assert.Equal(t, false, got["ok"])
+	})
+
+	// The size the client declares is what Commit validates the assembled archive against.
+	t.Run("finalizing with the wrong size is not ok", func(t *testing.T) {
+		created := v2Call(t, handler, testClient, "CreateCacheEntry", map[string]any{"key": "wrong-size", "version": "v1"})
+		require.Equal(t, http.StatusCreated, putBlob(t, created["signedUploadUrl"].(string), []byte("four")))
+
+		got := v2Call(t, handler, testClient, "FinalizeCacheEntryUpload", map[string]any{
+			"key": "wrong-size", "version": "v1", "size_bytes": 99,
+		})
+		assert.Equal(t, false, got["ok"])
+	})
+
+	// Both API versions are served from one store, so an entry written through v2 is a hit for
+	// a v1 client asking for the same key and version.
+	t.Run("a v1 client sees an entry written through v2", func(t *testing.T) {
+		resp, err := testClient.Get(fmt.Sprintf("%s%s/cache?keys=deps-abc&version=v1", handler.ExternalURL(), apiPath))
+		require.NoError(t, err)
+		defer resp.Body.Close()
+		require.Equal(t, http.StatusOK, resp.StatusCode)
+
+		got := map[string]any{}
+		require.NoError(t, json.NewDecoder(resp.Body).Decode(&got))
+		assert.Equal(t, "deps-abc", got["cacheKey"])
+		assert.NotEmpty(t, got["archiveLocation"])
+	})
+
+	// The cache of one repository must stay invisible to another, as it does for the v1 API.
+	t.Run("another repository sees nothing", func(t *testing.T) {
+		handler.RegisterJob("other-runtime-token", "other/repo")
+		otherClient := &http.Client{Transport: &bearerTransport{token: "other-runtime-token"}}
+
+		got := v2Call(t, handler, otherClient, "GetCacheEntryDownloadURL", map[string]any{"key": "deps-abc", "version": "v1"})
+		assert.Equal(t, false, got["ok"])
+	})
+}

act/artifactcache/storage.go

@@ -5,12 +5,15 @@
 package artifactcache
 
 import (
+	"crypto/sha256"
+	"encoding/hex"
 	"fmt"
 	"io"
 	"net/http"
 	"os"
 	"path/filepath"
 	"strconv"
+	"strings"
 )
 
 type Storage struct {
@@ -37,7 +40,10 @@ func (s *Storage) Exist(id uint64) (bool, error) {
 }
 
 func (s *Storage) Write(id uint64, offset int64, reader io.Reader) error {
-	name := s.tempName(id, offset)
+	return s.writeFile(s.tempName(id, offset), reader)
+}
+
+func (s *Storage) writeFile(name string, reader io.Reader) error {
 	if err := os.MkdirAll(filepath.Dir(name), 0o755); err != nil {
 		return err
 	}
@@ -51,6 +57,26 @@ func (s *Storage) Write(id uint64, offset int64, reader io.Reader) error {
 	return err
 }
 
+func (s *Storage) WriteBlock(id uint64, blockID string, reader io.Reader) error {
+	return s.writeFile(s.blockName(id, blockID), reader)
+}
+
+// OrderBlocks renames the staged blocks into the order the block list gives. A block the list
+// does not name keeps its staged name, which is how Commit leaves it out, as Azure drops it. One
+// 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 {
+	for i, blockID := range blockIDs {
+		if err := os.Rename(s.blockName(id, blockID), s.tempName(id, int64(i))); err != nil {
+			if os.IsNotExist(err) {
+				return fmt.Errorf("block %q of cache %d was never uploaded: %w", blockID, id, err)
+			}
+			return err
+		}
+	}
+	return nil
+}
+
 func (s *Storage) Commit(id uint64, size int64) (int64, error) {
 	defer func() {
 		_ = os.RemoveAll(s.tempDir(id))
@@ -65,6 +91,31 @@ func (s *Storage) Commit(id uint64, size int64) (int64, error) {
 	if err := os.MkdirAll(filepath.Dir(name), 0o755); err != nil {
 		return 0, err
 	}
+	written, err := assemble(name, tempNames)
+	if err != nil {
+		return 0, err
+	}
+	// If size is less than 0, it means the size is unknown.
+	// We can't check the size of the file, just skip the check.
+	// It happens when the request comes from old versions of actions, like `actions/cache@v2`.
+	if size >= 0 && written != size {
+		_ = os.Remove(name)
+		return 0, fmt.Errorf("broken file: %v != %v", written, size)
+	}
+	return written, nil
+}
+
+// assemble concatenates the uploaded parts into name. A single part, which is what the v2 API
+// produces below the client's block threshold, is already the whole archive and is moved.
+func assemble(name string, tempNames []string) (int64, error) {
+	if len(tempNames) == 1 {
+		info, err := os.Stat(tempNames[0])
+		if err != nil {
+			return 0, err
+		}
+		return info.Size(), os.Rename(tempNames[0], name)
+	}
+
 	file, err := os.Create(name)
 	if err != nil {
 		return 0, err
@@ -84,16 +135,6 @@ func (s *Storage) Commit(id uint64, size int64) (int64, error) {
 		}
 		written += n
 	}
-
-	// If size is less than 0, it means the size is unknown.
-	// We can't check the size of the file, just skip the check.
-	// It happens when the request comes from old versions of actions, like `actions/cache@v2`.
-	if size >= 0 && written != size {
-		_ = file.Close()
-		_ = os.Remove(name)
-		return 0, fmt.Errorf("broken file: %v != %v", written, size)
-	}
-
 	return written, nil
 }
 
@@ -119,6 +160,17 @@ func (s *Storage) tempName(id uint64, offset int64) string {
 	return filepath.Join(s.tempDir(id), fmt.Sprintf("%016x", offset))
 }
 
+// blockFilePrefix marks a staged, not yet ordered block, so that tempNames can keep it out of
+// Commit's name-ordered concatenation.
+const blockFilePrefix = "block-"
+
+func (s *Storage) blockName(id uint64, blockID string) string {
+	// The block id is client-chosen (base64), so it is hashed rather than trusted as a
+	// path element.
+	sum := sha256.Sum256([]byte(blockID))
+	return filepath.Join(s.tempDir(id), blockFilePrefix+hex.EncodeToString(sum[:]))
+}
+
 func (s *Storage) tempNames(id uint64) ([]string, error) {
 	dir := s.tempDir(id)
 	files, err := os.ReadDir(dir)
@@ -127,7 +179,7 @@ func (s *Storage) tempNames(id uint64) ([]string, error) {
 	}
 	var names []string
 	for _, v := range files {
-		if !v.IsDir() {
+		if !v.IsDir() && !strings.HasPrefix(v.Name(), blockFilePrefix) {
 			names = append(names, filepath.Join(dir, v.Name()))
 		}
 	}

act/common/executor_max_parallel_test.go

@@ -1,89 +0,0 @@
-// Copyright 2026 The Gitea Authors. All rights reserved.
-// SPDX-License-Identifier: MIT
-
-package common
-
-import (
-	"context"
-	"sync"
-	"sync/atomic"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-)
-
-// Simple fast test that verifies max-parallel: 2 limits concurrency
-func TestMaxParallel2Quick(t *testing.T) {
-	ctx := context.Background()
-
-	var currentRunning atomic.Int32
-	var maxSimultaneous atomic.Int32
-
-	executors := make([]Executor, 4)
-	for i := range 4 {
-		executors[i] = func(ctx context.Context) error {
-			current := currentRunning.Add(1)
-
-			// Update max if needed
-			for {
-				maxValue := maxSimultaneous.Load()
-				if current <= maxValue || maxSimultaneous.CompareAndSwap(maxValue, current) {
-					break
-				}
-			}
-
-			time.Sleep(10 * time.Millisecond)
-			currentRunning.Add(-1)
-			return nil
-		}
-	}
-
-	err := NewParallelExecutor(2, executors...)(ctx)
-
-	assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-	assert.LessOrEqual(t, maxSimultaneous.Load(), int32(2),
-		"Should not exceed max-parallel: 2")
-}
-
-// Test that verifies max-parallel: 1 enforces sequential execution
-func TestMaxParallel1Sequential(t *testing.T) {
-	ctx := context.Background()
-
-	var currentRunning atomic.Int32
-	var maxSimultaneous atomic.Int32
-	var executionOrder []int
-	var orderMutex sync.Mutex
-
-	executors := make([]Executor, 5)
-	for i := range 5 {
-		taskID := i
-		executors[i] = func(ctx context.Context) error {
-			current := currentRunning.Add(1)
-
-			// Track execution order
-			orderMutex.Lock()
-			executionOrder = append(executionOrder, taskID)
-			orderMutex.Unlock()
-
-			// Update max if needed
-			for {
-				maxValue := maxSimultaneous.Load()
-				if current <= maxValue || maxSimultaneous.CompareAndSwap(maxValue, current) {
-					break
-				}
-			}
-
-			time.Sleep(20 * time.Millisecond)
-			currentRunning.Add(-1)
-			return nil
-		}
-	}
-
-	err := NewParallelExecutor(1, executors...)(ctx)
-
-	assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-	assert.Equal(t, int32(1), maxSimultaneous.Load(),
-		"max-parallel: 1 should only run 1 task at a time")
-	assert.Len(t, executionOrder, 5, "All 5 tasks should have executed")
-}

act/common/executor_parallel_advanced_test.go

@@ -1,221 +0,0 @@
-// Copyright 2026 The Gitea Authors. All rights reserved.
-// SPDX-License-Identifier: MIT
-
-package common
-
-import (
-	"context"
-	"sync"
-	"sync/atomic"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-)
-
-// TestMaxParallelJobExecution tests actual job execution with max-parallel
-func TestMaxParallelJobExecution(t *testing.T) {
-	t.Run("MaxParallel=1 Sequential", func(t *testing.T) {
-		var currentRunning atomic.Int32
-		var maxConcurrent int32
-		var executionOrder []int
-		var mu sync.Mutex
-
-		executors := make([]Executor, 5)
-		for i := range 5 {
-			taskID := i
-			executors[i] = func(ctx context.Context) error {
-				current := currentRunning.Add(1)
-
-				// Track max concurrent
-				for {
-					maxValue := atomic.LoadInt32(&maxConcurrent)
-					if current <= maxValue || atomic.CompareAndSwapInt32(&maxConcurrent, maxValue, current) {
-						break
-					}
-				}
-
-				mu.Lock()
-				executionOrder = append(executionOrder, taskID)
-				mu.Unlock()
-
-				time.Sleep(10 * time.Millisecond)
-				currentRunning.Add(-1)
-				return nil
-			}
-		}
-
-		ctx := context.Background()
-		err := NewParallelExecutor(1, executors...)(ctx)
-		assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-
-		assert.Equal(t, int32(1), maxConcurrent, "Should never exceed 1 concurrent execution")
-		assert.Len(t, executionOrder, 5, "All tasks should execute")
-	})
-
-	t.Run("MaxParallel=3 Limited", func(t *testing.T) {
-		var currentRunning atomic.Int32
-		var maxConcurrent int32
-
-		executors := make([]Executor, 10)
-		for i := range 10 {
-			executors[i] = func(ctx context.Context) error {
-				current := currentRunning.Add(1)
-
-				for {
-					maxValue := atomic.LoadInt32(&maxConcurrent)
-					if current <= maxValue || atomic.CompareAndSwapInt32(&maxConcurrent, maxValue, current) {
-						break
-					}
-				}
-
-				time.Sleep(20 * time.Millisecond)
-				currentRunning.Add(-1)
-				return nil
-			}
-		}
-
-		ctx := context.Background()
-		err := NewParallelExecutor(3, executors...)(ctx)
-		assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-
-		assert.LessOrEqual(t, int(maxConcurrent), 3, "Should never exceed 3 concurrent executions")
-		assert.GreaterOrEqual(t, int(maxConcurrent), 1, "Should have at least 1 concurrent execution")
-	})
-
-	t.Run("MaxParallel=0 Uses1Worker", func(t *testing.T) {
-		var maxConcurrent int32
-		var currentRunning atomic.Int32
-
-		executors := make([]Executor, 5)
-		for i := range 5 {
-			executors[i] = func(ctx context.Context) error {
-				current := currentRunning.Add(1)
-
-				for {
-					maxValue := atomic.LoadInt32(&maxConcurrent)
-					if current <= maxValue || atomic.CompareAndSwapInt32(&maxConcurrent, maxValue, current) {
-						break
-					}
-				}
-
-				time.Sleep(10 * time.Millisecond)
-				currentRunning.Add(-1)
-				return nil
-			}
-		}
-
-		ctx := context.Background()
-		// When maxParallel is 0 or negative, it defaults to 1
-		err := NewParallelExecutor(0, executors...)(ctx)
-		assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-
-		assert.Equal(t, int32(1), maxConcurrent, "Should use 1 worker when max-parallel is 0")
-	})
-}
-
-// TestMaxParallelWithErrors tests error handling with max-parallel
-func TestMaxParallelWithErrors(t *testing.T) {
-	t.Run("OneTaskFailsOthersContinue", func(t *testing.T) {
-		var successCount int32
-
-		executors := make([]Executor, 5)
-		for i := range 5 {
-			taskID := i
-			executors[i] = func(ctx context.Context) error {
-				if taskID == 2 {
-					return assert.AnError
-				}
-				atomic.AddInt32(&successCount, 1)
-				return nil
-			}
-		}
-
-		ctx := context.Background()
-		err := NewParallelExecutor(2, executors...)(ctx)
-
-		// Should return the error from task 2
-		assert.Error(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-
-		// Other tasks should still execute
-		assert.Equal(t, int32(4), successCount, "4 tasks should succeed")
-	})
-
-	t.Run("ContextCancellation", func(t *testing.T) {
-		ctx, cancel := context.WithCancel(context.Background())
-
-		var startedCount int32
-		executors := make([]Executor, 10)
-		for i := range 10 {
-			executors[i] = func(ctx context.Context) error {
-				atomic.AddInt32(&startedCount, 1)
-				time.Sleep(100 * time.Millisecond)
-				return nil
-			}
-		}
-
-		// Cancel after a short delay
-		go func() {
-			time.Sleep(30 * time.Millisecond)
-			cancel()
-		}()
-
-		err := NewParallelExecutor(3, executors...)(ctx)
-		assert.Error(t, err)                     //nolint:testifylint // pre-existing issue from nektos/act
-		assert.ErrorIs(t, err, context.Canceled) //nolint:testifylint // pre-existing issue from nektos/act
-
-		// Not all tasks should start due to cancellation (but timing may vary)
-		// Just verify cancellation occurred
-		t.Logf("Started %d tasks before cancellation", startedCount)
-	})
-}
-
-// TestMaxParallelResourceSharing tests resource sharing scenarios
-func TestMaxParallelResourceSharing(t *testing.T) {
-	t.Run("SharedResourceWithMutex", func(t *testing.T) {
-		var sharedCounter int
-		var mu sync.Mutex
-
-		executors := make([]Executor, 100)
-		for i := range 100 {
-			executors[i] = func(ctx context.Context) error {
-				mu.Lock()
-				sharedCounter++
-				mu.Unlock()
-				return nil
-			}
-		}
-
-		ctx := context.Background()
-		err := NewParallelExecutor(10, executors...)(ctx)
-		assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-
-		assert.Equal(t, 100, sharedCounter, "All tasks should increment counter")
-	})
-
-	t.Run("ChannelCommunication", func(t *testing.T) {
-		resultChan := make(chan int, 50)
-
-		executors := make([]Executor, 50)
-		for i := range 50 {
-			taskID := i
-			executors[i] = func(ctx context.Context) error {
-				resultChan <- taskID
-				return nil
-			}
-		}
-
-		ctx := context.Background()
-		err := NewParallelExecutor(5, executors...)(ctx)
-		assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
-
-		close(resultChan)
-
-		results := make(map[int]bool)
-		for result := range resultChan {
-			results[result] = true
-		}
-
-		assert.Len(t, results, 50, "All task IDs should be received")
-	})
-}

act/common/executor_test.go

@@ -9,9 +9,9 @@ import (
 	"errors"
 	"reflect"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"testing"
-	"time"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -82,44 +82,45 @@ func TestNewConditionalExecutor(t *testing.T) {
 	assert.Equal(1, falseCount)
 }
 
-func TestNewParallelExecutor(t *testing.T) {
-	assert := assert.New(t)
+// concurrencyProbe returns an executor recording the peak number of concurrent copies. Copies
+// block until wantActive are in flight so the peak is exact without sleeping, and later copies
+// find the gate already open so the last one still finishes with no partner left.
+func concurrencyProbe(wantActive int32) (exec Executor, count, maxActive *atomic.Int32) {
+	var counted, active, peak atomic.Int32
+	var once sync.Once
+	reached := make(chan struct{})
 
-	ctx := context.Background()
-
-	var count, activeCount, maxCount atomic.Int32
-	emptyWorkflow := NewPipelineExecutor(func(ctx context.Context) error {
-		count.Add(1)
-
-		active := activeCount.Add(1)
+	return func(ctx context.Context) error {
+		counted.Add(1)
+		running := active.Add(1)
 		for {
-			m := maxCount.Load()
-			if active <= m || maxCount.CompareAndSwap(m, active) {
+			seen := peak.Load()
+			if running <= seen || peak.CompareAndSwap(seen, running) {
 				break
 			}
 		}
-		time.Sleep(2 * time.Second)
-		activeCount.Add(-1)
-
+		if running >= wantActive {
+			once.Do(func() { close(reached) })
+		}
+		<-reached
+		active.Add(-1)
 		return nil
-	})
+	}, &counted, &peak
+}
 
-	err := NewParallelExecutor(2, emptyWorkflow, emptyWorkflow, emptyWorkflow)(ctx)
+func TestNewParallelExecutor(t *testing.T) {
+	ctx := context.Background()
 
-	assert.Equal(int32(3), count.Load(), "should run all 3 executors")
-	assert.Equal(int32(2), maxCount.Load(), "should run at most 2 executors in parallel")
-	assert.NoError(err) //nolint:testifylint // pre-existing issue from nektos/act
+	exec, count, maxActive := concurrencyProbe(2)
+	require.NoError(t, NewParallelExecutor(2, exec, exec, exec)(ctx))
+	assert.Equal(t, int32(3), count.Load(), "should run all 3 executors")
+	assert.Equal(t, int32(2), maxActive.Load(), "should run at most 2 executors in parallel")
 
-	// Reset to test running the executor with 0 parallelism
-	count.Store(0)
-	activeCount.Store(0)
-	maxCount.Store(0)
-
-	errSingle := NewParallelExecutor(0, emptyWorkflow, emptyWorkflow, emptyWorkflow)(ctx)
-
-	assert.Equal(int32(3), count.Load(), "should run all 3 executors")
-	assert.Equal(int32(1), maxCount.Load(), "should run at most 1 executors in parallel")
-	assert.NoError(errSingle)
+	// parallelism below 1 falls back to a single worker
+	exec, count, maxActive = concurrencyProbe(1)
+	require.NoError(t, NewParallelExecutor(0, exec, exec, exec)(ctx))
+	assert.Equal(t, int32(3), count.Load(), "should run all 3 executors")
+	assert.Equal(t, int32(1), maxActive.Load(), "should run at most 1 executor in parallel")
 }
 
 func TestNewParallelExecutorEmpty(t *testing.T) {
@@ -173,6 +174,23 @@ func TestNewParallelExecutorCanceled(t *testing.T) {
 	assert.Error(errExpected, err) //nolint:testifylint // pre-existing issue from nektos/act
 }
 
+func TestNewParallelExecutorRunsRemainingAfterFailure(t *testing.T) {
+	var successCount atomic.Int32
+	executors := make([]Executor, 5)
+	for i := range executors {
+		executors[i] = func(ctx context.Context) error {
+			if i == 2 {
+				return errors.New("fake error")
+			}
+			successCount.Add(1)
+			return nil
+		}
+	}
+
+	require.Error(t, NewParallelExecutor(2, executors...)(context.Background()))
+	assert.Equal(t, int32(4), successCount.Load(), "a failing executor must not stop the others")
+}
+
 func TestExecutorConditionalsAndFinally(t *testing.T) {
 	ctx := context.Background()
 	var calls []string

act/common/git/git.go

@@ -16,6 +16,7 @@ import (
 	"sync"
 
 	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/internal/pkg/lock"
 
 	"github.com/go-git/go-git/v5"
 	"github.com/go-git/go-git/v5/config"
@@ -32,7 +33,7 @@ var (
 	githubHTTPRegex     = regexp.MustCompile(`^https?://.*github.com.*/(.+)/(.+?)(?:.git)?$`)
 	githubSSHRegex      = regexp.MustCompile(`github.com[:/](.+)/(.+?)(?:.git)?$`)
 
-	cloneLocks sync.Map // key: clone target directory; value: *sync.Mutex
+	cloneLocks lock.Keyed[string] // key: clone target directory
 
 	ErrShortRef = errors.New("short SHA references are not supported")
 	ErrNoRepo   = errors.New("unable to find git repo")
@@ -43,10 +44,7 @@ var (
 // Callers reading files inside dir (e.g. tarring a checked-out action into a job container) must hold this lock too,
 // otherwise a concurrent NewGitCloneExecutor on the same dir can mutate the worktree mid-read.
 func AcquireCloneLock(dir string) func() {
-	v, _ := cloneLocks.LoadOrStore(dir, &sync.Mutex{})
-	mu := v.(*sync.Mutex)
-	mu.Lock()
-	return mu.Unlock
+	return cloneLocks.Lock(dir)
 }
 
 type Error struct {
@@ -261,6 +259,10 @@ type NewGitCloneExecutorInput struct {
 	// 0 for full clone.
 	Depth int
 
+	// Quiet drops the informational clone line to debug level, for callers that log their own
+	// download summary (the setup section's action report).
+	Quiet bool
+
 	// For Gitea
 	InsecureSkipTLS bool
 }
@@ -347,7 +349,11 @@ func gitOptions(token string) (fetchOptions git.FetchOptions, pullOptions git.Pu
 func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 	return func(ctx context.Context) error {
 		logger := common.Logger(ctx)
-		logger.Infof("git clone '%s' # ref=%s", input.URL, input.Ref)
+		if input.Quiet {
+			logger.Debugf("git clone '%s' # ref=%s", input.URL, input.Ref)
+		} else {
+			logger.Infof("git clone '%s' # ref=%s", input.URL, input.Ref)
+		}
 		logger.Debugf("  cloning %s to %s", input.URL, input.Dir)
 
 		defer AcquireCloneLock(input.Dir)()

act/common/git/git_test.go

@@ -12,11 +12,14 @@ import (
 	"path/filepath"
 	"strconv"
 	"strings"
-	"sync"
 	"syscall"
 	"testing"
 	"time"
 
+	"gitea.com/gitea/runner/act/common"
+
+	log "github.com/sirupsen/logrus"
+	logrustest "github.com/sirupsen/logrus/hooks/test"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -404,6 +407,44 @@ func TestGitCloneExecutorOfflineMode(t *testing.T) {
 	})
 }
 
+func TestGitCloneExecutorQuietDemotesCloneLine(t *testing.T) {
+	remoteDir := t.TempDir()
+	require.NoError(t, gitCmd("init", "--bare", "--initial-branch=main", remoteDir))
+	workDir := t.TempDir()
+	require.NoError(t, gitCmd("clone", remoteDir, workDir))
+	require.NoError(t, gitCmd("-C", workDir, "checkout", "-b", "main"))
+	require.NoError(t, gitCmd("-C", workDir, "commit", "--allow-empty", "-m", "initial"))
+	require.NoError(t, gitCmd("-C", workDir, "push", "-u", "origin", "main"))
+
+	// Quiet callers report the download themselves, so the clone line must not reach the job log.
+	for name, quiet := range map[string]bool{"quiet": true, "not quiet": false} {
+		t.Run(name, func(t *testing.T) {
+			logger, hook := logrustest.NewNullLogger()
+			logger.SetLevel(log.InfoLevel)
+			ctx := common.WithLogger(context.Background(), logger.WithField("job", "j1"))
+
+			require.NoError(t, NewGitCloneExecutor(NewGitCloneExecutorInput{
+				URL:   remoteDir,
+				Ref:   "main",
+				Dir:   t.TempDir(),
+				Quiet: quiet,
+			})(ctx))
+
+			var cloneLines int
+			for _, entry := range hook.AllEntries() {
+				if strings.HasPrefix(entry.Message, "git clone ") {
+					cloneLines++
+				}
+			}
+			if quiet {
+				assert.Zero(t, cloneLines)
+			} else {
+				assert.Equal(t, 1, cloneLines)
+			}
+		})
+	}
+}
+
 func TestGitCloneExecutorShallow(t *testing.T) {
 	// Build a local "remote" with several commits on main plus a tag, so a full clone would pull noticeably more history than a shallow one.
 	remoteDir := t.TempDir()
@@ -568,12 +609,4 @@ func TestAcquireCloneLock(t *testing.T) {
 			t.Fatal("acquire on a different directory must not block")
 		}
 	})
-
-	t.Run("same directory reuses the same mutex", func(t *testing.T) {
-		dir := t.TempDir()
-
-		v1, _ := cloneLocks.LoadOrStore(dir, &sync.Mutex{})
-		v2, _ := cloneLocks.LoadOrStore(dir, &sync.Mutex{})
-		require.Same(t, v1, v2)
-	})
 }

act/common/outbound_ip.go

@@ -19,7 +19,9 @@ func GetOutboundIP() net.IP {
 	conn, err := net.Dial("udp", "8.8.8.8:80")
 	if err == nil {
 		defer conn.Close()
-		return conn.LocalAddr().(*net.UDPAddr).IP
+		if addr, ok := conn.LocalAddr().(*net.UDPAddr); ok {
+			return addr.IP
+		}
 	}
 
 	// So the machine cannot access the internet. Pick an IP address from network interfaces.

act/container/container_types.go

@@ -82,12 +82,14 @@ type NewDockerBuildExecutorInput struct {
 	BuildContext io.Reader
 	ImageTag     string
 	Platform     string
+	BuildArgs    map[string]*string
 }
 
 // NewDockerNetworkCreateExecutorInput the input for the NewDockerNetworkCreateExecutor function
 type NewDockerNetworkCreateExecutorInput struct {
 	EnableIPv4 *bool
 	EnableIPv6 *bool
+	RunnerUUID string
 }
 
 // NewDockerPullExecutorInput the input for the NewDockerPullExecutor function

act/container/docker_build.go

@@ -49,6 +49,7 @@ func NewDockerBuildExecutor(input NewDockerBuildExecutorInput) common.Executor {
 			Remove:      true,
 			AuthConfigs: LoadDockerAuthConfigs(ctx),
 			Dockerfile:  input.Dockerfile,
+			BuildArgs:   input.BuildArgs,
 		}
 		platform, err := parsePlatform(input.Platform)
 		if err != nil {

act/container/docker_create_flags.go

@@ -0,0 +1,86 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build !(WITHOUT_DOCKER || !(linux || darwin || windows || netbsd))
+
+package container
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"slices"
+
+	"github.com/kballard/go-shellquote"
+	"github.com/spf13/pflag"
+)
+
+const (
+	pullPolicyAlways  = "always"
+	pullPolicyMissing = "missing"
+	pullPolicyNever   = "never"
+)
+
+var pullPolicies = []string{pullPolicyAlways, pullPolicyMissing, pullPolicyNever}
+
+// createFlags are the flags docker/cli registers on the `create` and `run` commands
+// instead of in addFlags, so they are not part of containerOptions.
+type createFlags struct {
+	platform     string
+	pull         string
+	name         string
+	useAPISocket bool
+}
+
+func registerCreateFlags(flags *pflag.FlagSet) *createFlags {
+	cf := new(createFlags)
+	flags.StringVar(&cf.platform, "platform", "", "Set platform if server is multi-platform capable")
+	flags.StringVar(&cf.pull, "pull", pullPolicyMissing, `Pull image before creating ("always", "missing", "never")`)
+	flags.StringVar(&cf.name, "name", "", "Assign a name to the container")
+	flags.BoolVar(&cf.useAPISocket, "use-api-socket", false, "Bind mount Docker API socket and required auth")
+	// Accepted without effect: pull progress is only logged at debug level, and docker
+	// no longer implements content trust.
+	flags.BoolP("quiet", "q", false, "Suppress the pull output")
+	flags.Bool("disable-content-trust", true, "Skip image verification (deprecated)")
+	return cf
+}
+
+// parseContainerOptions parses a container options string. The flags are returned even
+// on error, holding whatever was read before the failure.
+func parseContainerOptions(options string) (*pflag.FlagSet, *containerOptions, *createFlags, error) {
+	flags := pflag.NewFlagSet("container_flags", pflag.ContinueOnError)
+	flags.SetOutput(io.Discard)
+	copts := addFlags(flags)
+	cf := registerCreateFlags(flags)
+
+	args, err := shellquote.Split(options)
+	if err != nil {
+		return flags, copts, cf, fmt.Errorf("Cannot split container options: '%s': '%w'", options, err)
+	}
+
+	if err := flags.Parse(args); err != nil {
+		return flags, copts, cf, fmt.Errorf("Cannot parse container options: '%s': '%w'", options, err)
+	}
+
+	return flags, copts, cf, nil
+}
+
+// createFlagsFromOptions reads the create-level flags that have to be known before the
+// container is created. Malformed options keep the defaults here and are reported by
+// mergeContainerConfigs at create time.
+func createFlagsFromOptions(options string) *createFlags {
+	_, _, cf, _ := parseContainerOptions(options)
+	return cf
+}
+
+func (cf *createFlags) validate() error {
+	if !slices.Contains(pullPolicies, cf.pull) {
+		return fmt.Errorf("invalid --pull option %q: must be one of %q", cf.pull, pullPolicies)
+	}
+
+	if cf.useAPISocket {
+		return errors.New("--use-api-socket is not supported, use the runner's container.docker_host setting to expose a docker socket")
+	}
+
+	return nil
+}

act/container/docker_create_flags_test.go

@@ -0,0 +1,62 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package container
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCreateFlagsFromOptions(t *testing.T) {
+	for _, tc := range []struct {
+		options  string
+		platform string
+		pull     string
+	}{
+		{"", "", pullPolicyMissing},
+		{"-v /a:/b --platform=linux/arm64 --pull always", "linux/arm64", pullPolicyAlways},
+		{"--platform linux/arm/v7 --pull never", "linux/arm/v7", pullPolicyNever},
+		{`--platform "linux/amd64`, "", pullPolicyMissing}, // malformed, defaults kept
+	} {
+		t.Run(tc.options, func(t *testing.T) {
+			cf := createFlagsFromOptions(tc.options)
+			assert.Equal(t, tc.platform, cf.platform)
+			assert.Equal(t, tc.pull, cf.pull)
+		})
+	}
+}
+
+func TestCreateFlagsValidate(t *testing.T) {
+	for _, tc := range []struct {
+		options string
+		wantErr string
+	}{
+		{"--quiet --disable-content-trust --name mine", ""},
+		{"--pull sometimes", `invalid --pull option "sometimes"`},
+		{"--use-api-socket", "--use-api-socket is not supported"},
+	} {
+		t.Run(tc.options, func(t *testing.T) {
+			err := createFlagsFromOptions(tc.options).validate()
+			if tc.wantErr == "" {
+				require.NoError(t, err)
+				return
+			}
+			require.ErrorContains(t, err, tc.wantErr)
+		})
+	}
+}
+
+func TestNewContainerAppliesCreateFlags(t *testing.T) {
+	input := &NewContainerInput{Platform: "linux/amd64", Options: "--platform linux/arm64 --pull never"}
+	cr, ok := NewContainer(input).(*containerReference)
+	require.True(t, ok)
+	assert.Equal(t, "linux/arm64", input.Platform)
+	assert.Equal(t, pullPolicyNever, cr.pullPolicy)
+
+	kept := &NewContainerInput{Platform: "linux/amd64", Options: "--privileged"}
+	NewContainer(kept)
+	assert.Equal(t, "linux/amd64", kept.Platform)
+}

act/container/docker_logger.go

@@ -57,7 +57,7 @@ func logDockerResponse(logger logrus.FieldLogger, dockerResponse io.ReadCloser,
 
 		if msg.ErrorDetail.Message != "" {
 			writeLog(logger, isError, "%s", msg.ErrorDetail.Message)
-			return errors.New(msg.Error)
+			return errors.New(msg.ErrorDetail.Message)
 		}
 
 		if msg.Status != "" {

act/container/docker_network.go

@@ -8,12 +8,69 @@ package container
 
 import (
 	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
 
 	"gitea.com/gitea/runner/act/common"
 
 	"github.com/moby/moby/client"
 )
 
+const (
+	networkCreateAttempts   = 3
+	networkCreateRetryDelay = time.Second
+
+	// marks the networks a runner creates for its jobs, so it can tell its own leftovers from
+	// those of another runner sharing the daemon
+	runnerUUIDLabel = "com.gitea.runner.uuid"
+)
+
+// RemoveOrphanNetworks removes the networks this runner created for jobs whose teardown did
+// not get to them: the runner died with the job, the teardown timed out, or the network still
+// 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 {
+	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)
+}
+
+func removeOrphanNetworks(ctx context.Context, cli client.APIClient, runnerUUID string, createdBefore time.Time) error {
+	networks, err := cli.NetworkList(ctx, client.NetworkListOptions{
+		Filters: make(client.Filters).Add("label", runnerUUIDLabel+"="+runnerUUID),
+	})
+	if err != nil {
+		return err
+	}
+
+	var errs []error
+	for _, n := range networks.Items {
+		result, err := cli.NetworkInspect(ctx, n.ID, client.NetworkInspectOptions{})
+		if err != nil {
+			errs = append(errs, fmt.Errorf("failed to inspect network %s: %w", n.Name, err))
+			continue
+		}
+		// 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) {
+			continue
+		}
+		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
+		}
+		common.Logger(ctx).Infof("removed docker network %s left behind by an earlier job", n.Name)
+	}
+	return errors.Join(errs...)
+}
+
 func NewDockerNetworkCreateExecutor(name string, opts NewDockerNetworkCreateExecutorInput) common.Executor {
 	return func(ctx context.Context) error {
 		cli, err := GetDockerClient(ctx)
@@ -36,18 +93,45 @@ func NewDockerNetworkCreateExecutor(name string, opts NewDockerNetworkCreateExec
 			}
 		}
 
-		_, err = cli.NetworkCreate(ctx, name, client.NetworkCreateOptions{
-			Driver:     "bridge",
-			Scope:      "local",
-			EnableIPv4: opts.EnableIPv4,
-			EnableIPv6: opts.EnableIPv6,
-		})
-		if err != nil {
-			return err
+		for i := range networkCreateAttempts {
+			if i > 0 {
+				common.Logger(ctx).Infof("Waiting for a free docker address pool to create network %s", name)
+				select {
+				case <-ctx.Done():
+					return ctx.Err()
+				case <-time.After(time.Duration(i) * networkCreateRetryDelay):
+				}
+			}
+			if _, err = cli.NetworkCreate(ctx, name, client.NetworkCreateOptions{
+				Driver:     "bridge",
+				Scope:      "local",
+				EnableIPv4: opts.EnableIPv4,
+				EnableIPv6: opts.EnableIPv6,
+				Labels:     runnerLabels(opts.RunnerUUID),
+			}); err == nil {
+				return nil
+			}
+			if !isAddressPoolExhausted(err) {
+				return err
+			}
 		}
+		return fmt.Errorf("docker has no address pool left for this job's network, lower runner.capacity or widen default-address-pools in the docker daemon config: %w", err)
+	}
+}
 
+func runnerLabels(runnerUUID string) map[string]string {
+	if runnerUUID == "" {
 		return nil
 	}
+	return map[string]string{runnerUUIDLabel: runnerUUID}
+}
+
+// The daemon reports this as a plain invalid-parameter error, the same kind it uses for every
+// malformed request, so the message is the only discriminator.
+func isAddressPoolExhausted(err error) bool {
+	msg := err.Error()
+	return strings.Contains(msg, "all predefined address pools have been fully subnetted") ||
+		strings.Contains(msg, "could not find an available, non-overlapping IPv4 address pool among the defaults") // docker 24 and older
 }
 
 func NewDockerNetworkRemoveExecutor(name string) common.Executor {
@@ -66,6 +150,7 @@ func NewDockerNetworkRemoveExecutor(name string) common.Executor {
 		}
 		// For Gitea, reduce log noise
 		// common.Logger(ctx).Debugf("%v", networks)
+		var errs []error
 		for _, n := range networks.Items {
 			if n.Name == name {
 				result, err := cli.NetworkInspect(ctx, n.ID, client.NetworkInspectOptions{})
@@ -73,16 +158,17 @@ func NewDockerNetworkRemoveExecutor(name string) common.Executor {
 					return err
 				}
 
-				if len(result.Network.Containers) == 0 {
-					if _, err = cli.NetworkRemove(ctx, n.ID, client.NetworkRemoveOptions{}); err != nil {
-						common.Logger(ctx).Debugf("%v", err)
-					}
-				} else {
-					common.Logger(ctx).Debugf("Refusing to remove network %v because it still has active endpoints", name)
+				// it holds a subnet out of the daemon's pool until something reclaims it
+				if len(result.Network.Containers) != 0 {
+					common.Logger(ctx).Warnf("Refusing to remove network %s because it still has active endpoints, the idle cleanup reclaims it once they are gone", name)
+					continue
+				}
+				if _, err = cli.NetworkRemove(ctx, n.ID, client.NetworkRemoveOptions{}); err != nil {
+					errs = append(errs, fmt.Errorf("failed to remove network %s: %w", name, err))
 				}
 			}
 		}
 
-		return err
+		return errors.Join(errs...)
 	}
 }

act/container/docker_network_test.go

@@ -0,0 +1,50 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package container
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	cerrdefs "github.com/containerd/errdefs"
+	"github.com/moby/moby/api/types/network"
+	mobyclient "github.com/moby/moby/client"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+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 TestRemoveOrphanNetworks(t *testing.T) {
+	ctx := context.Background()
+	cutoff := time.Date(2026, time.April, 29, 20, 0, 0, 0, time.UTC)
+	client := &mockDockerClient{}
+	client.On("NetworkList", ctx, mobyclient.NetworkListOptions{
+		Filters: make(mobyclient.Filters).Add("label", runnerUUIDLabel+"=runner-1"),
+	}).Return(mobyclient.NetworkListResult{Items: []network.Summary{
+		{Network: network.Network{ID: "orphan"}},
+		{Network: network.Network{ID: "busy"}},
+		{Network: network.Network{ID: "starting"}},
+	}}, 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)
+	client.On("NetworkRemove", ctx, "orphan", mobyclient.NetworkRemoveOptions{}).
+		Return(mobyclient.NetworkRemoveResult{}, nil)
+
+	require.NoError(t, removeOrphanNetworks(ctx, client, "runner-1", cutoff))
+	client.AssertExpectations(t)
+}

act/container/docker_pull.go

@@ -30,23 +30,19 @@ func NewDockerPullExecutor(input NewDockerPullExecutorInput) common.Executor {
 			return nil
 		}
 
-		pull := input.ForcePull
-		if !pull {
+		// skip the pull when the image is already here: either none was forced, or a digest
+		// pins the content so a forced pull could only fetch the same bytes again
+		if !input.ForcePull || isPinnedImage(input.Image) {
 			imageExists, err := ImageExistsLocally(ctx, input.Image, input.Platform)
 			logger.Debugf("Image exists? %v", imageExists)
 			if err != nil {
 				return fmt.Errorf("unable to determine if image already exists for image '%s' (%s): %w", input.Image, input.Platform, err)
 			}
-
-			if !imageExists {
-				pull = true
+			if imageExists {
+				return nil
 			}
 		}
 
-		if !pull {
-			return nil
-		}
-
 		imageRef := cleanImage(ctx, input.Image)
 		logger.Debugf("pulling image '%v' (%s)", imageRef, input.Platform)
 
@@ -61,22 +57,32 @@ func NewDockerPullExecutor(input NewDockerPullExecutorInput) common.Executor {
 			return err
 		}
 
-		reader, err := cli.ImagePull(ctx, imageRef, imagePullOptions)
-
-		_ = logDockerResponse(logger, reader, err != nil)
-		if err != nil {
-			if imagePullOptions.RegistryAuth != "" && strings.Contains(err.Error(), "unauthorized") {
-				logger.Errorf("pulling image '%v' (%s) failed with credentials %s retrying without them, please check for stale docker config files", imageRef, input.Platform, err.Error())
-				imagePullOptions.RegistryAuth = ""
-				reader, err = cli.ImagePull(ctx, imageRef, imagePullOptions)
-
-				_ = logDockerResponse(logger, reader, err != nil)
-			}
+		// the daemon reports a failure that happens after the first progress line in the
+		// stream rather than on the call itself, so both have to be checked
+		pullOnce := func(opts client.ImagePullOptions) error {
+			reader, err := cli.ImagePull(ctx, imageRef, opts)
+			streamErr := logDockerResponse(logger, reader, err != nil)
 			if err != nil {
-				return fmt.Errorf("failed to pull image '%s' (%s): %w", imageRef, input.Platform, err)
+				return err
 			}
+			return streamErr
 		}
-		return nil
+
+		err = pullOnce(imagePullOptions)
+		if err != nil && imagePullOptions.RegistryAuth != "" && strings.Contains(err.Error(), "unauthorized") {
+			logger.Errorf("pulling image '%v' (%s) failed with credentials %s retrying without them, please check for stale docker config files", imageRef, input.Platform, err.Error())
+			imagePullOptions.RegistryAuth = ""
+			err = pullOnce(imagePullOptions)
+		}
+		if err == nil {
+			return nil
+		}
+		// a registry that is down should not fail a job whose image is already here
+		if exists, existsErr := ImageExistsLocally(ctx, input.Image, input.Platform); existsErr == nil && exists {
+			logger.Warnf("could not update image '%s' (%s), continuing with the local copy: %v", imageRef, input.Platform, err)
+			return nil
+		}
+		return fmt.Errorf("failed to pull image '%s' (%s): %w", imageRef, input.Platform, err)
 	}
 }
 
@@ -122,6 +128,15 @@ func getImagePullOptions(ctx context.Context, input NewDockerPullExecutorInput)
 	return imagePullOptions, nil
 }
 
+func isPinnedImage(image string) bool {
+	ref, err := reference.ParseAnyReference(image)
+	if err != nil {
+		return false
+	}
+	_, pinned := ref.(reference.Canonical)
+	return pinned
+}
+
 func cleanImage(ctx context.Context, imageName string) string {
 	ref, err := reference.ParseAnyReference(imageName)
 	if err != nil {

act/container/docker_pull_test.go

@@ -6,11 +6,15 @@ package container
 
 import (
 	"context"
+	"io"
+	"strings"
 	"testing"
 
 	"github.com/docker/cli/cli/config"
 	log "github.com/sirupsen/logrus"
+	"github.com/sirupsen/logrus/hooks/test"
 	assert "github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func init() {
@@ -65,3 +69,21 @@ func TestGetImagePullOptions(t *testing.T) {
 	assert.NoError(t, err, "Failed to create ImagePullOptions") //nolint:testifylint // pre-existing issue from nektos/act
 	assert.Equal(t, "eyJ1c2VybmFtZSI6InVzZXJuYW1lIiwicGFzc3dvcmQiOiJwYXNzd29yZFxuIiwic2VydmVyYWRkcmVzcyI6Imh0dHBzOi8vaW5kZXguZG9ja2VyLmlvL3YxLyJ9", options.RegistryAuth, "RegistryAuth should be taken from local docker config")
 }
+
+// A digest-pinned image is immutable, so its local copy is always current.
+func TestIsPinnedImage(t *testing.T) {
+	assert.True(t, isPinnedImage("alpine@sha256:28bd5fe8b56d1bd048e5babf5b10710ebe0bae67db86916198a6eec434943f8b"))
+	assert.False(t, isPinnedImage("alpine:latest"))
+}
+
+// The pull path reports a failure the daemon sent mid-stream, so it must carry the reason
+// whichever of the two shapes the daemon used.
+func TestLogDockerResponseError(t *testing.T) {
+	logger, _ := test.NewNullLogger()
+	streamErr := func(line string) error {
+		return logDockerResponse(logger, io.NopCloser(strings.NewReader(line)), false)
+	}
+	require.EqualError(t, streamErr(`{"error":"toomanyrequests: rate limit exceeded"}`), "toomanyrequests: rate limit exceeded")
+	require.EqualError(t, streamErr(`{"errorDetail":{"message":"unexpected EOF"}}`), "unexpected EOF")
+	require.NoError(t, streamErr(`{"status":"Downloading"}`))
+}

act/container/docker_run.go

@@ -14,6 +14,7 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"path"
 	"path/filepath"
 	"regexp"
 	"runtime"
@@ -35,7 +36,6 @@ import (
 	"github.com/go-git/go-git/v5/plumbing/format/gitignore"
 	"github.com/gobwas/glob"
 	"github.com/joho/godotenv"
-	"github.com/kballard/go-shellquote"
 	"github.com/moby/moby/api/pkg/stdcopy"
 	"github.com/moby/moby/api/types/container"
 	"github.com/moby/moby/api/types/mount"
@@ -43,7 +43,7 @@ import (
 	"github.com/moby/moby/api/types/system"
 	"github.com/moby/moby/client"
 	specs "github.com/opencontainers/image-spec/specs-go/v1"
-	"github.com/spf13/pflag"
+	"github.com/sirupsen/logrus"
 )
 
 // drainGracePeriod bounds how long we wait for an output-copy goroutine to
@@ -57,6 +57,12 @@ const drainGracePeriod = 2 * time.Second
 func NewContainer(input *NewContainerInput) ExecutionsEnvironment {
 	cr := new(containerReference)
 	cr.input = input
+	// Resolved up front because the image pull runs before the container is created.
+	cf := createFlagsFromOptions(input.Options)
+	if cf.platform != "" {
+		cr.input.Platform = cf.platform
+	}
+	cr.pullPolicy = cf.pull
 	return cr
 }
 
@@ -137,6 +143,11 @@ func (cr *containerReference) Start(attach bool) common.Executor {
 }
 
 func (cr *containerReference) Pull(forcePull bool) common.Executor {
+	if cr.pullPolicy == pullPolicyNever {
+		return common.NewInfoExecutor("docker pull skipped image=%s, --pull=never in the options", cr.input.Image)
+	}
+	forcePull = forcePull || cr.pullPolicy == pullPolicyAlways
+
 	return common.
 		NewInfoExecutor("docker pull image=%s platform=%s username=%s forcePull=%t", cr.input.Image, cr.input.Platform, cr.input.Username, forcePull).
 		Then(
@@ -232,11 +243,12 @@ func (cr *containerReference) ReplaceLogWriter(stdout, stderr io.Writer) (io.Wri
 }
 
 type containerReference struct {
-	cli   client.APIClient
-	id    string
-	input *NewContainerInput
-	UID   int
-	GID   int
+	cli        client.APIClient
+	id         string
+	input      *NewContainerInput
+	pullPolicy string
+	UID        int
+	GID        int
 	// attachDone is closed by the attach() streaming goroutine once it has
 	// drained and flushed the container's output. wait() blocks on it so the
 	// tail of the log lands before the step proceeds.
@@ -378,25 +390,69 @@ func (cr *containerReference) find() common.Executor {
 
 func (cr *containerReference) remove() common.Executor {
 	return func(ctx context.Context) error {
-		if cr.id == "" {
+		idOrName := cr.id
+		if idOrName == "" && cr.input != nil {
+			idOrName = cr.input.Name
+		}
+		if idOrName == "" {
 			return nil
 		}
 
 		logger := common.Logger(ctx)
-		_, err := cr.cli.ContainerRemove(ctx, cr.id, client.ContainerRemoveOptions{
+		// Kill first so removal never waits out a daemon's stop timeout: Docker kills outright
+		// on a forced remove, Podman sends SIGTERM and waits. Only worth it for a container
+		// this started, and removal can still deal with one it could not kill.
+		if cr.id != "" {
+			_, err := cr.cli.ContainerKill(ctx, cr.id, client.ContainerKillOptions{Signal: "SIGKILL"})
+			if err != nil && !cerrdefs.IsConflict(err) && !cerrdefs.IsNotFound(err) {
+				logger.Debugf("Container %s could not be killed: %v", cr.id, err)
+			}
+		}
+		_, err := cr.cli.ContainerRemove(ctx, idOrName, client.ContainerRemoveOptions{
 			RemoveVolumes: true,
 			Force:         true,
 		})
-		if err != nil {
-			logger.Error(fmt.Errorf("failed to remove container: %w", err))
+		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()
 		}
 
-		logger.Debugf("Removed container: %v", cr.id)
+		logger.Debugf("Removed container: %v", idOrName)
 		cr.id = ""
 		return nil
 	}
 }
 
+func (cr *containerReference) waitForRemoval(ctx context.Context, idOrName string) {
+	// 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)
+	defer cancel()
+
+	waitResult := cr.cli.ContainerWait(ctx, idOrName, client.ContainerWaitOptions{
+		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)
+	}
+}
+
 func (cr *containerReference) mergeContainerConfigs(ctx context.Context, config *container.Config, hostConfig *container.HostConfig) (*container.Config, *container.HostConfig, error) {
 	logger := common.Logger(ctx)
 	input := cr.input
@@ -406,17 +462,13 @@ func (cr *containerReference) mergeContainerConfigs(ctx context.Context, config
 	}
 
 	// parse configuration from CLI container.options
-	flags := pflag.NewFlagSet("container_flags", pflag.ContinueOnError)
-	copts := addFlags(flags)
-
-	optionsArgs, err := shellquote.Split(input.Options)
+	flags, copts, cf, err := parseContainerOptions(input.Options)
 	if err != nil {
-		return nil, nil, fmt.Errorf("Cannot split container options: '%s': '%w'", input.Options, err)
+		return nil, nil, err
 	}
 
-	err = flags.Parse(optionsArgs)
-	if err != nil {
-		return nil, nil, fmt.Errorf("Cannot parse container options: '%s': '%w'", input.Options, err)
+	if err := cf.validate(); err != nil {
+		return nil, nil, fmt.Errorf("Cannot process container options: '%s': '%w'", input.Options, err)
 	}
 
 	// FIXME: If everything is fine after gitea/act v0.260.0, remove the following comment.
@@ -450,6 +502,16 @@ func (cr *containerReference) mergeContainerConfigs(ctx context.Context, config
 		return nil, nil, fmt.Errorf("Cannot process container options: '%s': '%w'", input.Options, err)
 	}
 
+	// For Gitea
+	// When privileged mode is disabled, container.options is workflow-controlled
+	// untrusted input. Strip the HostConfig fields that would let a workflow break
+	// out of the container (host namespaces, capability expansion, security profile
+	// overrides, device and runtime access). Otherwise these survive into the final
+	// HostConfig even though --privileged is forced off.
+	if !hostConfig.Privileged {
+		sanitizeOptionsHostConfig(logger, containerConfig.HostConfig)
+	}
+
 	logger.Debugf("Custom container.Config from options ==> %+v", containerConfig.Config)
 
 	err = mergo.Merge(config, containerConfig.Config, mergo.WithOverride, mergo.WithAppendSlice)
@@ -460,8 +522,7 @@ func (cr *containerReference) mergeContainerConfigs(ctx context.Context, config
 
 	logger.Debugf("Custom container.HostConfig from options ==> %+v", containerConfig.HostConfig)
 
-	hostConfig.Binds = append(hostConfig.Binds, containerConfig.HostConfig.Binds...)
-	hostConfig.Mounts = append(hostConfig.Mounts, containerConfig.HostConfig.Mounts...)
+	overlayVolumes(hostConfig, containerConfig.HostConfig)
 	binds := hostConfig.Binds
 	mounts := hostConfig.Mounts
 	networkMode := hostConfig.NetworkMode
@@ -471,6 +532,9 @@ func (cr *containerReference) mergeContainerConfigs(ctx context.Context, config
 	}
 	hostConfig.Binds = binds
 	hostConfig.Mounts = mounts
+	if cf.name != "" {
+		logger.Warn("--name in the options will be ignored.")
+	}
 	if len(copts.netMode.Value()) > 0 {
 		logger.Warn("--network and --net in the options will be ignored.")
 	}
@@ -801,23 +865,59 @@ func (cr *containerReference) waitForCommand(ctx context.Context, isTerminal boo
 	}
 }
 
+// mkdirInContainer creates containerPath and returns it with the symlinked components
+// replaced by the targets the daemon reports for them. Docker 29.7 rejects tar entries
+// traversing a symlink to an absolute target, like the "/var/run" of most images, with
+// "path escapes from parent", and not every daemon creates the implied parents of a
+// directory entry, so one entry per missing component is extracted at the deepest
+// existing ancestor.
+// WORKAROUND: https://github.com/moby/moby/issues/53258
+func (cr *containerReference) mkdirInContainer(ctx context.Context, containerPath string) (string, error) {
+	parts := strings.Split(strings.Trim(path.Clean(containerPath), "/"), "/")
+	existing := "/"
+	for i, part := range parts {
+		if part == "" {
+			return existing, nil
+		}
+		stat, err := cr.cli.ContainerStatPath(ctx, cr.id, client.ContainerStatPathOptions{Path: path.Join(existing, part)})
+		if err != nil {
+			// nothing below exists either, so create the remaining components
+			return path.Join(existing, path.Join(parts[i:]...)), cr.mkdirEntries(ctx, existing, parts[i:])
+		}
+		existing = path.Join(existing, part)
+		if target := stat.Stat.LinkTarget; target != "" {
+			if !path.IsAbs(target) {
+				target = path.Join(path.Dir(existing), target)
+			}
+			existing = target
+		}
+	}
+	return existing, nil
+}
+
+func (cr *containerReference) mkdirEntries(ctx context.Context, destPath string, missing []string) error {
+	buf := &bytes.Buffer{}
+	tw := tar.NewWriter(buf)
+	for i := range missing {
+		_ = tw.WriteHeader(&tar.Header{
+			Name:     path.Join(missing[:i+1]...),
+			Mode:     0o777,
+			Typeflag: tar.TypeDir,
+		})
+	}
+	tw.Close()
+	_, err := cr.cli.CopyToContainer(ctx, cr.id, client.CopyToContainerOptions{
+		DestinationPath: destPath,
+		Content:         buf,
+	})
+	return err
+}
+
 func (cr *containerReference) CopyTarStream(ctx context.Context, destPath string, tarStream io.Reader) error {
 	if cr.id == "" {
 		return cr.missingContainerError("copy to %s", destPath)
 	}
-	// Mkdir
-	buf := &bytes.Buffer{}
-	tw := tar.NewWriter(buf)
-	_ = tw.WriteHeader(&tar.Header{
-		Name:     destPath,
-		Mode:     0o777,
-		Typeflag: tar.TypeDir,
-	})
-	tw.Close()
-	_, err := cr.cli.CopyToContainer(ctx, cr.id, client.CopyToContainerOptions{
-		DestinationPath: "/",
-		Content:         buf,
-	})
+	destPath, err := cr.mkdirInContainer(ctx, destPath)
 	if err != nil {
 		return fmt.Errorf("failed to mkdir to copy content to container: %w", err)
 	}
@@ -842,6 +942,10 @@ func (cr *containerReference) copyDir(dstPath, srcPath string, useGitIgnore bool
 			return cr.missingContainerError("copy directory to %s", dstPath)
 		}
 		logger := common.Logger(ctx)
+		dstPath, err := cr.mkdirInContainer(ctx, dstPath)
+		if err != nil {
+			return fmt.Errorf("failed to mkdir to copy directory to container: %w", err)
+		}
 		tarFile, err := os.CreateTemp("", "act")
 		if err != nil {
 			return err
@@ -1049,6 +1153,77 @@ func (cr *containerReference) wait() common.Executor {
 	}
 }
 
+// For Gitea
+// sanitizeOptionsHostConfig clears the HostConfig fields parsed from a
+// workflow-controlled container.options string that could be used to escape the
+// container when privileged mode is disabled. It must only be called when the
+// runner has privileged mode turned off; with privileged mode enabled the
+// administrator has already opted into host access.
+func sanitizeOptionsHostConfig(logger logrus.FieldLogger, hostConfig *container.HostConfig) {
+	warn := func(option string) {
+		logger.Warnf("container option %q is not allowed when privileged mode is disabled and will be ignored", option)
+	}
+
+	if hostConfig.PidMode != "" {
+		warn("--pid")
+		hostConfig.PidMode = ""
+	}
+	if hostConfig.IpcMode != "" {
+		warn("--ipc")
+		hostConfig.IpcMode = ""
+	}
+	if hostConfig.UTSMode != "" {
+		warn("--uts")
+		hostConfig.UTSMode = ""
+	}
+	if hostConfig.CgroupnsMode != "" {
+		warn("--cgroupns")
+		hostConfig.CgroupnsMode = ""
+	}
+	// UsernsMode is set from the runner-controlled input; never let options
+	// override it (e.g. --userns=host disables user namespace remapping).
+	if hostConfig.UsernsMode != "" {
+		warn("--userns")
+		hostConfig.UsernsMode = ""
+	}
+	if len(hostConfig.CapAdd) > 0 {
+		warn("--cap-add")
+		hostConfig.CapAdd = nil
+	}
+	if len(hostConfig.SecurityOpt) > 0 {
+		warn("--security-opt")
+		hostConfig.SecurityOpt = nil
+	}
+	if len(hostConfig.Devices) > 0 {
+		warn("--device")
+		hostConfig.Devices = nil
+	}
+	if len(hostConfig.DeviceCgroupRules) > 0 {
+		warn("--device-cgroup-rule")
+		hostConfig.DeviceCgroupRules = nil
+	}
+	if len(hostConfig.DeviceRequests) > 0 {
+		warn("--gpus")
+		hostConfig.DeviceRequests = nil
+	}
+	if len(hostConfig.VolumesFrom) > 0 {
+		warn("--volumes-from")
+		hostConfig.VolumesFrom = nil
+	}
+	if hostConfig.Runtime != "" {
+		warn("--runtime")
+		hostConfig.Runtime = ""
+	}
+	if hostConfig.CgroupParent != "" {
+		warn("--cgroup-parent")
+		hostConfig.CgroupParent = ""
+	}
+	if len(hostConfig.Sysctls) > 0 {
+		warn("--sysctl")
+		hostConfig.Sysctls = nil
+	}
+}
+
 // For Gitea
 // sanitizeConfig remove the invalid configurations from `config` and `hostConfig`
 func (cr *containerReference) sanitizeConfig(ctx context.Context, config *container.Config, hostConfig *container.HostConfig) (*container.Config, *container.HostConfig) {
@@ -1094,6 +1269,34 @@ func (cr *containerReference) sanitizeConfig(ctx context.Context, config *contai
 	return config, hostConfig
 }
 
+// bindTarget returns the container path a bind mounts onto, empty if it cannot be parsed.
+func bindTarget(bind string) string {
+	parsed, err := loader.ParseVolume(bind)
+	if err != nil {
+		return ""
+	}
+	return parsed.Target
+}
+
+// overlayVolumes appends src's volumes to dst, dropping the dst ones they mount over. Docker
+// rejects two mounts on one target, so the volumes declared last have to win.
+func overlayVolumes(dst, src *container.HostConfig) {
+	claimed := map[string]bool{}
+	for _, bind := range src.Binds {
+		if target := bindTarget(bind); target != "" {
+			claimed[target] = true
+		}
+	}
+	for _, mt := range src.Mounts {
+		claimed[mt.Target] = true
+	}
+
+	dst.Binds = append(slices.DeleteFunc(slices.Clone(dst.Binds),
+		func(bind string) bool { return claimed[bindTarget(bind)] }), src.Binds...)
+	dst.Mounts = append(slices.DeleteFunc(slices.Clone(dst.Mounts),
+		func(mt mount.Mount) bool { return claimed[mt.Target] }), src.Mounts...)
+}
+
 type validVolumeMatcher struct {
 	allowAll bool
 	named    []glob.Glob

act/container/docker_run_test.go

@@ -5,6 +5,7 @@
 package container
 
 import (
+	"archive/tar"
 	"bufio"
 	"bytes"
 	"context"
@@ -23,6 +24,7 @@ import (
 	cerrdefs "github.com/containerd/errdefs"
 	"github.com/moby/moby/api/pkg/stdcopy"
 	"github.com/moby/moby/api/types/container"
+	"github.com/moby/moby/api/types/mount"
 	mobyclient "github.com/moby/moby/client"
 	"github.com/sirupsen/logrus/hooks/test"
 	"github.com/stretchr/testify/assert"
@@ -91,6 +93,11 @@ func (m *mockDockerClient) ExecInspect(ctx context.Context, execID string, opts
 	return args.Get(0).(mobyclient.ExecInspectResult), args.Error(1)
 }
 
+func (m *mockDockerClient) ContainerStatPath(ctx context.Context, containerID string, opts mobyclient.ContainerStatPathOptions) (mobyclient.ContainerStatPathResult, error) {
+	args := m.Called(ctx, containerID, opts)
+	return args.Get(0).(mobyclient.ContainerStatPathResult), args.Error(1)
+}
+
 func (m *mockDockerClient) ContainerAttach(ctx context.Context, containerID string, opts mobyclient.ContainerAttachOptions) (mobyclient.ContainerAttachResult, error) {
 	args := m.Called(ctx, containerID, opts)
 	return args.Get(0).(mobyclient.ContainerAttachResult), args.Error(1)
@@ -116,6 +123,31 @@ func (m *mockDockerClient) ContainerList(ctx context.Context, opts mobyclient.Co
 	return args.Get(0).(mobyclient.ContainerListResult), args.Error(1)
 }
 
+func (m *mockDockerClient) ContainerRemove(ctx context.Context, id string, opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
+	args := m.Called(ctx, id, opts)
+	return args.Get(0).(mobyclient.ContainerRemoveResult), args.Error(1)
+}
+
+func (m *mockDockerClient) ContainerKill(ctx context.Context, id string, opts mobyclient.ContainerKillOptions) (mobyclient.ContainerKillResult, error) {
+	args := m.Called(ctx, id, opts)
+	return args.Get(0).(mobyclient.ContainerKillResult), args.Error(1)
+}
+
+func (m *mockDockerClient) NetworkList(ctx context.Context, opts mobyclient.NetworkListOptions) (mobyclient.NetworkListResult, error) {
+	args := m.Called(ctx, opts)
+	return args.Get(0).(mobyclient.NetworkListResult), args.Error(1)
+}
+
+func (m *mockDockerClient) NetworkInspect(ctx context.Context, id string, opts mobyclient.NetworkInspectOptions) (mobyclient.NetworkInspectResult, error) {
+	args := m.Called(ctx, id, opts)
+	return args.Get(0).(mobyclient.NetworkInspectResult), args.Error(1)
+}
+
+func (m *mockDockerClient) NetworkRemove(ctx context.Context, id string, opts mobyclient.NetworkRemoveOptions) (mobyclient.NetworkRemoveResult, error) {
+	args := m.Called(ctx, id, opts)
+	return args.Get(0).(mobyclient.NetworkRemoveResult), args.Error(1)
+}
+
 type endlessReader struct {
 	io.Reader
 }
@@ -309,15 +341,37 @@ func TestDockerWaitFailure(t *testing.T) {
 	client.AssertExpectations(t)
 }
 
+// stubStatPath answers path resolution: the given paths exist, mapped to their target
+// when they are a symlink, everything else does not exist.
+func stubStatPath(client *mockDockerClient, existing map[string]string) {
+	for containerPath, target := range existing {
+		client.On("ContainerStatPath", mock.Anything, "123", mobyclient.ContainerStatPathOptions{Path: containerPath}).
+			Return(mobyclient.ContainerStatPathResult{Stat: container.PathStat{LinkTarget: target}}, nil).Maybe()
+	}
+	client.On("ContainerStatPath", mock.Anything, "123", mock.Anything).
+		Return(mobyclient.ContainerStatPathResult{}, cerrdefs.ErrNotFound).Maybe()
+}
+
+// The mkdir tarball is extracted at the deepest existing ancestor, with entries relative
+// to it that never traverse the "/var/run" symlink, see moby/moby#53258.
 func TestDockerCopyTarStream(t *testing.T) {
 	ctx := context.Background()
 
+	var mkdirNames []string
 	client := &mockDockerClient{}
+	stubStatPath(client, map[string]string{"/var": "", "/var/run": "/run", "/run": ""})
 	client.On("CopyToContainer", ctx, "123", mock.MatchedBy(func(opts mobyclient.CopyToContainerOptions) bool {
-		return opts.DestinationPath == "/" && opts.Content != nil
+		if opts.DestinationPath != "/run" || opts.Content == nil {
+			return false
+		}
+		tr := tar.NewReader(opts.Content)
+		for hdr, err := tr.Next(); err == nil; hdr, err = tr.Next() {
+			mkdirNames = append(mkdirNames, hdr.Name)
+		}
+		return true
 	})).Return(mobyclient.CopyToContainerResult{}, nil)
 	client.On("CopyToContainer", ctx, "123", mock.MatchedBy(func(opts mobyclient.CopyToContainerOptions) bool {
-		return opts.DestinationPath == "/var/run/act" && opts.Content != nil
+		return opts.DestinationPath == "/run/act" && opts.Content != nil
 	})).Return(mobyclient.CopyToContainerResult{}, nil)
 	cr := &containerReference{
 		id:  "123",
@@ -327,57 +381,103 @@ func TestDockerCopyTarStream(t *testing.T) {
 		},
 	}
 
-	_ = cr.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{})
+	require.NoError(t, cr.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{}))
+	assert.Equal(t, []string{"act"}, mkdirNames)
 
 	client.AssertExpectations(t)
 }
 
-func TestDockerCopyTarStreamErrorInCopyFiles(t *testing.T) {
-	ctx := context.Background()
-
+func TestDockerCopyTarStreamErrors(t *testing.T) {
 	merr := errors.New("Failure")
+	for _, testCase := range []struct {
+		name     string
+		mkdirErr error
+		copyErr  error
+	}{
+		{"mkdir", merr, nil},
+		{"copy content", nil, merr},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			ctx := context.Background()
 
-	client := &mockDockerClient{}
-	client.On("CopyToContainer", ctx, "123", mock.MatchedBy(func(opts mobyclient.CopyToContainerOptions) bool {
-		return opts.DestinationPath == "/" && opts.Content != nil
-	})).Return(mobyclient.CopyToContainerResult{}, merr)
-	cr := &containerReference{
-		id:  "123",
-		cli: client,
-		input: &NewContainerInput{
-			Image: "image",
-		},
+			client := &mockDockerClient{}
+			stubStatPath(client, map[string]string{"/var": "", "/var/run": ""})
+			client.On("CopyToContainer", ctx, "123", mock.MatchedBy(func(opts mobyclient.CopyToContainerOptions) bool {
+				return opts.DestinationPath == "/var/run" && opts.Content != nil
+			})).Return(mobyclient.CopyToContainerResult{}, testCase.mkdirErr)
+			client.On("CopyToContainer", ctx, "123", mock.MatchedBy(func(opts mobyclient.CopyToContainerOptions) bool {
+				return opts.DestinationPath == "/var/run/act" && opts.Content != nil
+			})).Return(mobyclient.CopyToContainerResult{}, testCase.copyErr).Maybe()
+			cr := &containerReference{
+				id:  "123",
+				cli: client,
+				input: &NewContainerInput{
+					Image: "image",
+				},
+			}
+
+			require.ErrorIs(t, cr.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{}), merr)
+
+			client.AssertExpectations(t)
+		})
 	}
-
-	err := cr.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{})
-	assert.ErrorIs(t, err, merr) //nolint:testifylint // pre-existing issue from nektos/act
-
-	client.AssertExpectations(t)
 }
 
-func TestDockerCopyTarStreamErrorInMkdir(t *testing.T) {
-	ctx := context.Background()
+// 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) {
+	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: "removal in progress", err: cerrdefs.ErrConflict.WithMessage("removal of container abc is already in progress"), wantWait: true},
+		{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},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			logger, hook := test.NewNullLogger()
+			ctx := common.WithLogger(context.Background(), logger)
+			client := &mockDockerClient{}
+			client.On("ContainerKill", ctx, "abc", killOpts).Return(mobyclient.ContainerKillResult{}, nil)
+			client.On("ContainerRemove", ctx, "abc", removeOpts).Return(mobyclient.ContainerRemoveResult{}, tc.err)
+			if tc.wantWait {
+				removed := make(chan container.WaitResponse, 1)
+				removed <- container.WaitResponse{}
+				client.On("ContainerWait", mock.Anything, "abc", mobyclient.ContainerWaitOptions{Condition: container.WaitConditionRemoved}).
+					Return(mobyclient.ContainerWaitResult{Result: removed})
+			}
+			cr := &containerReference{id: "abc", cli: client}
 
-	merr := errors.New("Failure")
-
-	client := &mockDockerClient{}
-	client.On("CopyToContainer", ctx, "123", mock.MatchedBy(func(opts mobyclient.CopyToContainerOptions) bool {
-		return opts.DestinationPath == "/" && opts.Content != nil
-	})).Return(mobyclient.CopyToContainerResult{}, nil)
-	client.On("CopyToContainer", ctx, "123", mock.MatchedBy(func(opts mobyclient.CopyToContainerOptions) bool {
-		return opts.DestinationPath == "/var/run/act" && opts.Content != nil
-	})).Return(mobyclient.CopyToContainerResult{}, merr)
-	cr := &containerReference{
-		id:  "123",
-		cli: client,
-		input: &NewContainerInput{
-			Image: "image",
-		},
+			require.NoError(t, cr.remove()(ctx))
+			// a failure keeps the id, so a later Remove() can retry it
+			if tc.wantFailure {
+				assert.Equal(t, "abc", cr.id)
+				assert.Len(t, hook.AllEntries(), 1)
+			} else {
+				assert.Empty(t, cr.id)
+				assert.Empty(t, hook.AllEntries())
+			}
+			client.AssertExpectations(t)
+		})
 	}
+}
 
-	err := cr.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{})
-	assert.ErrorIs(t, err, merr) //nolint:testifylint // pre-existing issue from nektos/act
+// A container whose id was never learned, because find() could not reach the daemon or
+// create() lost its reply, must still be removed rather than leaking with its network. It
+// was never started here, so it is not worth a kill of its own.
+func TestRemoveWithoutIDUsesName(t *testing.T) {
+	ctx := context.Background()
+	client := &mockDockerClient{}
+	client.On("ContainerRemove", ctx, "job-1", mobyclient.ContainerRemoveOptions{RemoveVolumes: true, Force: true}).
+		Return(mobyclient.ContainerRemoveResult{}, nil)
+	cr := &containerReference{cli: client, input: &NewContainerInput{Name: "job-1"}}
 
+	require.NoError(t, cr.remove()(ctx))
 	client.AssertExpectations(t)
 }
 
@@ -502,9 +602,8 @@ func TestDockerCopyToSymlinkPath(t *testing.T) {
 		_ = rc.Close()(ctx)
 	})
 
-	// CopyTarStream first creates the destination directory by extracting a tar at "/",
-	// which makes the daemon mkdir var, then var/run (the symlink), then act — the exact
-	// step that fails on the broken daemon.
+	// CopyTarStream resolves the var/run symlink and creates act below its target, the
+	// exact step that fails on a broken daemon.
 	err := rc.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{})
 	require.NoError(t, err)
 }
@@ -585,6 +684,110 @@ func TestCheckVolumes(t *testing.T) {
 	}
 }
 
+func TestSanitizeOptionsHostConfig(t *testing.T) {
+	logger, _ := test.NewNullLogger()
+
+	dangerous := func() *container.HostConfig {
+		return &container.HostConfig{
+			PidMode:      "host",
+			IpcMode:      "host",
+			UTSMode:      "host",
+			CgroupnsMode: "host",
+			UsernsMode:   "host",
+			CapAdd:       []string{"ALL"},
+			SecurityOpt:  []string{"seccomp=unconfined", "apparmor=unconfined"},
+			VolumesFrom:  []string{"other"},
+			Runtime:      "runc",
+			Resources: container.Resources{
+				CgroupParent:      "/custom",
+				Devices:           []container.DeviceMapping{{PathOnHost: "/dev/sda", PathInContainer: "/dev/sda", CgroupPermissions: "rwm"}},
+				DeviceCgroupRules: []string{"a *:* rwm"},
+			},
+			Sysctls: map[string]string{"net.ipv4.ip_forward": "1"},
+		}
+	}
+
+	hostConfig := dangerous()
+	sanitizeOptionsHostConfig(logger, hostConfig)
+
+	assert.Empty(t, string(hostConfig.PidMode))
+	assert.Empty(t, string(hostConfig.IpcMode))
+	assert.Empty(t, string(hostConfig.UTSMode))
+	assert.Empty(t, string(hostConfig.CgroupnsMode))
+	assert.Empty(t, string(hostConfig.UsernsMode))
+	assert.Empty(t, hostConfig.CapAdd)
+	assert.Empty(t, hostConfig.SecurityOpt)
+	assert.Empty(t, hostConfig.Devices)
+	assert.Empty(t, hostConfig.DeviceCgroupRules)
+	assert.Empty(t, hostConfig.VolumesFrom)
+	assert.Empty(t, hostConfig.Runtime)
+	assert.Empty(t, hostConfig.CgroupParent)
+	assert.Empty(t, hostConfig.Sysctls)
+}
+
+func TestMergeContainerConfigsStripsDangerousOptionsWhenUnprivileged(t *testing.T) {
+	// OS-independent options only: --device parsing requires a linux/windows
+	// server OS, which is not guaranteed for the test host.
+	const dangerousOptions = "--pid=host --ipc=host --uts=host --cgroupns=host " +
+		"--userns=host --cap-add=ALL --security-opt seccomp=unconfined " +
+		"--security-opt apparmor=unconfined --volumes-from other " +
+		"--runtime runc --cgroup-parent /custom --sysctl net.ipv4.ip_forward=1"
+
+	t.Run("unprivileged strips host-escape options", func(t *testing.T) {
+		logger, _ := test.NewNullLogger()
+		ctx := common.WithLogger(context.Background(), logger)
+		cr := &containerReference{
+			input: &NewContainerInput{
+				Options:     dangerousOptions,
+				NetworkMode: "bridge",
+				UsernsMode:  "private",
+			},
+		}
+
+		_, hostConfig, err := cr.mergeContainerConfigs(ctx, &container.Config{}, &container.HostConfig{
+			Privileged:  false,
+			UsernsMode:  container.UsernsMode("private"),
+			NetworkMode: container.NetworkMode("bridge"),
+		})
+		require.NoError(t, err)
+
+		assert.False(t, hostConfig.Privileged)
+		assert.Empty(t, string(hostConfig.PidMode))
+		assert.Empty(t, string(hostConfig.IpcMode))
+		assert.Empty(t, string(hostConfig.UTSMode))
+		assert.Empty(t, string(hostConfig.CgroupnsMode))
+		// UsernsMode must keep the runner-controlled value, not the one from options.
+		assert.Equal(t, "private", string(hostConfig.UsernsMode))
+		assert.Empty(t, hostConfig.CapAdd)
+		assert.Empty(t, hostConfig.SecurityOpt)
+		assert.Empty(t, hostConfig.VolumesFrom)
+		assert.Empty(t, hostConfig.Runtime)
+		assert.Empty(t, hostConfig.CgroupParent)
+		assert.Empty(t, hostConfig.Sysctls)
+	})
+
+	t.Run("privileged preserves options", func(t *testing.T) {
+		logger, _ := test.NewNullLogger()
+		ctx := common.WithLogger(context.Background(), logger)
+		cr := &containerReference{
+			input: &NewContainerInput{
+				Options:     "--pid=host --cap-add=ALL --security-opt seccomp=unconfined",
+				NetworkMode: "bridge",
+			},
+		}
+
+		_, hostConfig, err := cr.mergeContainerConfigs(ctx, &container.Config{}, &container.HostConfig{
+			Privileged:  true,
+			NetworkMode: container.NetworkMode("bridge"),
+		})
+		require.NoError(t, err)
+
+		assert.Equal(t, "host", string(hostConfig.PidMode))
+		assert.Equal(t, []string{"ALL"}, hostConfig.CapAdd)
+		assert.Equal(t, []string{"seccomp=unconfined"}, hostConfig.SecurityOpt)
+	})
+}
+
 func TestCheckVolumesRejectsEscapingHostPaths(t *testing.T) {
 	logger, _ := test.NewNullLogger()
 	ctx := common.WithLogger(context.Background(), logger)
@@ -621,3 +824,22 @@ func TestCheckVolumesRejectsEscapingHostPaths(t *testing.T) {
 	})
 	assert.Empty(t, hostConf.Binds)
 }
+
+func TestMergeContainerConfigsVolumesReplaceRunnerMounts(t *testing.T) {
+	logger, _ := test.NewNullLogger()
+	ctx := common.WithLogger(context.Background(), logger)
+	cr := &containerReference{
+		input: &NewContainerInput{
+			NetworkMode: "bridge",
+			Options:     "--volume /host/tools:/opt/hostedtoolcache",
+		},
+	}
+
+	_, hostConf, err := cr.mergeContainerConfigs(ctx, &container.Config{}, &container.HostConfig{
+		Binds:  []string{"/var/run/docker.sock:/var/run/docker.sock"},
+		Mounts: []mount.Mount{{Type: mount.TypeVolume, Source: "act-toolcache", Target: "/opt/hostedtoolcache"}},
+	})
+	require.NoError(t, err)
+	assert.Equal(t, []string{"/var/run/docker.sock:/var/run/docker.sock", "/host/tools:/opt/hostedtoolcache"}, hostConf.Binds)
+	assert.Empty(t, hostConf.Mounts)
+}

act/container/docker_stub.go

@@ -9,6 +9,7 @@ package container
 import (
 	"context"
 	"runtime"
+	"time"
 
 	"gitea.com/gitea/runner/act/common"
 
@@ -72,3 +73,7 @@ func NewDockerNetworkRemoveExecutor(name string) common.Executor {
 		return nil
 	}
 }
+
+func RemoveOrphanNetworks(ctx context.Context, runnerUUID string, createdBefore time.Time) error {
+	return nil
+}

act/container/host_environment.go

@@ -17,6 +17,7 @@ import (
 	"path/filepath"
 	"runtime"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"time"
 
@@ -43,6 +44,25 @@ type HostEnvironment struct {
 	CleanUp      func()
 	StdOut       io.Writer
 	AllocatePTY  bool // allocate a pseudo-TTY for each step's process
+
+	// procGroup owns every process the job's steps start. Atomic: Remove may read
+	// it while a step is still starting.
+	procGroupOnce sync.Once
+	procGroup     atomic.Pointer[process.Group]
+}
+
+// processGroup returns the job-scoped process group, creating it on first use.
+// Returns nil if the job object could not be created; Group is nil-safe.
+func (e *HostEnvironment) processGroup(ctx context.Context) *process.Group {
+	e.procGroupOnce.Do(func() {
+		group, err := process.NewGroup()
+		if err != nil {
+			common.Logger(ctx).Warnf("could not create the job's process group; processes a step leaves behind can only be reclaimed by the workspace scan: %v", err)
+			return
+		}
+		e.procGroup.Store(group)
+	})
+	return e.procGroup.Load()
 }
 
 func (e *HostEnvironment) Create(_, _ []string) common.Executor {
@@ -310,6 +330,10 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
 	} else {
 		wd = e.Path
 	}
+	// Flush any buffered, not-yet-newline-terminated trailing line, as the docker backend
+	// does in waitForCommand, so the final line of a command's output is not lost.
+	defer common.FlushWriter(e.StdOut)
+
 	f, err := lookupPathHost(command[0], env, e.StdOut)
 	if err != nil {
 		return err
@@ -324,11 +348,8 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
 	cmd.Dir = wd
 	cmd.SysProcAttr = process.SysProcAttr(cmdline, false)
 
-	// Kill the step's whole process tree on cancellation (a step often launches a
-	// shell that spawns further background or GUI children) and bound the post-exit
-	// I/O wait, so an orphan inheriting cmd's stdout/stderr pipe can never hang
-	// cmd.Wait() and the runner. See process.TreeKill. The PTY path below may
-	// override SysProcAttr, but never touches Cancel/WaitDelay.
+	// Kills the step's whole tree on cancellation and bounds the post-exit I/O
+	// wait, so an orphan holding cmd's stdout pipe cannot hang cmd.Wait().
 	treeKill := process.NewTreeKill(cmd)
 
 	var ppty *os.File
@@ -360,6 +381,11 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
 	if err := cmd.Start(); err != nil {
 		return err
 	}
+	// Assign before the step's Killer so the step's job nests inside the group's;
+	// cancellation still scopes to this step's tree.
+	if err := e.processGroup(ctx).Assign(cmd.Process); err != nil {
+		common.Logger(ctx).Warnf("could not assign the step's process to the job's process group; a process it leaves behind may outlive the job: %v", err)
+	}
 	if k, kerr := treeKill.Capture(cmd.Process); kerr != nil {
 		common.Logger(ctx).Warnf("process tree kill setup failed, falling back to single-process kill: %v", kerr)
 	} else {
@@ -407,13 +433,12 @@ func (e *HostEnvironment) UpdateFromEnv(srcPath string, env *map[string]string)
 	return parseEnvFile(e, srcPath, env)
 }
 
-// removeAll is the filesystem delete used by removeAllWithContext. A package
-// var so tests can substitute a blocking stub without patching os.RemoveAll.
+// removeAll is a var so tests can substitute a blocking stub.
 var removeAll = os.RemoveAll
 
-// removeAllWithContext runs removeAll in a goroutine and returns once it
-// finishes or ctx is cancelled. On cancellation the goroutine is left running —
-// a delete blocked inside a syscall cannot be interrupted (see runWithTimeout).
+// removeAllWithContext returns once the delete finishes or ctx is cancelled. On
+// cancellation the goroutine leaks: a delete inside a syscall cannot be
+// interrupted (see runWithTimeout).
 func removeAllWithContext(ctx context.Context, path string) error {
 	done := make(chan error, 1)
 	go func() { done <- removeAll(path) }()
@@ -455,17 +480,12 @@ func removePathWithRetry(ctx context.Context, path string) error {
 	return lastErr
 }
 
-// buildWindowsWorkspaceKillScript builds a PowerShell command that `taskkill
-// /T /F`s every process tree whose ExecutablePath or CommandLine references one
-// of the given absolute workspace dirs, releasing file handles for cleanup.
-//
-// Win32_Process is used because it exposes both ExecutablePath and CommandLine
-// (Get-Process doesn't, wmic is deprecated). Both match the dir+separator
-// prefix, so a sibling dir sharing a name prefix (job1 vs job10) is spared.
-// Ordinal String methods, not -like, so path metacharacters ([ ] ? *) stay
-// literal.
-//
-// Pure function so the quote-escaping can be unit-tested without PowerShell.
+// buildWindowsWorkspaceKillScript builds a PowerShell command that taskkills
+// every process tree whose ExecutablePath or CommandLine references one of the
+// given workspace dirs, releasing file handles for cleanup. Win32_Process
+// exposes both fields (Get-Process doesn't, wmic is deprecated); matching is on
+// the dir+separator prefix via ordinal String methods, so a name-prefix sibling
+// (job1 vs job10) is spared and path metacharacters stay literal.
 func buildWindowsWorkspaceKillScript(dirs []string) string {
 	quoted := make([]string, len(dirs))
 	for i, d := range dirs {
@@ -501,9 +521,8 @@ func (e *HostEnvironment) terminateRunningProcesses(ctx context.Context) {
 
 	logger := common.Logger(ctx)
 
-	// Workspace dirs we own. Any process running from or referencing one is a
-	// leftover job process. ToolCache is shared across jobs; Workdir only when
-	// we own it (else it's a caller-provided checkout, e.g. act local mode).
+	// Dirs we own; a process referencing one is a leftover. ToolCache is shared
+	// across jobs, and Workdir may be a caller-owned checkout.
 	owned := []string{e.Path, e.TmpDir}
 	if e.CleanWorkdir {
 		owned = append(owned, e.Workdir)
@@ -530,21 +549,24 @@ func (e *HostEnvironment) terminateRunningProcesses(ctx context.Context) {
 	if err != nil {
 		logger.Debugf("workspace process-tree kill via PowerShell failed: %v output=%s", err, strings.TrimSpace(string(out)))
 	}
+
+	// Win32_Process exposes no working directory, so the scan above misses a
+	// process that merely runs in a workspace dir while pinning a handle on it.
+	if killed, err := process.KillProcessesWithCWDUnder(killCtx, dirs); err != nil {
+		logger.Debugf("workspace process kill by working directory reported errors: %v", err)
+	} else if killed > 0 {
+		logger.Debugf("terminated %d leftover process(es) by workspace working directory", killed)
+	}
 }
 
-// hostCleanupTimeout bounds each filesystem-teardown phase of the host
-// environment so a single stalled delete cannot wedge the runner slot forever.
-// A var (not const) so tests can shrink it.
+// hostCleanupTimeout bounds each teardown phase so one stalled delete cannot
+// wedge the runner slot. A var so tests can shrink it.
 var hostCleanupTimeout = 30 * time.Second
 
-// runWithTimeout runs fn in a goroutine and returns once it finishes or timeout
-// elapses, whichever comes first. On timeout the goroutine is left running — an
-// os.RemoveAll blocked inside a delete syscall (AV/EDR filter drivers, an
-// unresponsive network mount, a dying disk) cannot be interrupted — and
-// context.DeadlineExceeded is returned. Leaking the goroutine and the scratch
-// state it was deleting is strictly better than blocking the caller forever and
-// permanently losing the runner's capacity slot; the leaked scratch dir is
-// reclaimed later by the runner's idle stale-dir sweep.
+// runWithTimeout returns context.DeadlineExceeded once timeout elapses, leaking
+// the goroutine: a delete blocked in a syscall (AV filter driver, dead network
+// mount) cannot be interrupted, and leaking scratch state beats losing the
+// runner's capacity slot forever. The idle stale-dir sweep reclaims it later.
 func runWithTimeout(fn func(), timeout time.Duration) error {
 	done := make(chan struct{})
 	go func() {
@@ -565,14 +587,15 @@ func (e *HostEnvironment) Remove() common.Executor {
 	return func(ctx context.Context) error {
 		logger := common.Logger(ctx)
 
-		// Ensure any lingering child processes are ended before attempting
-		// to remove the workspace (Windows file locks otherwise prevent cleanup).
+		// End lingering processes before removing the workspace; on Windows their
+		// file locks block cleanup. Closing the group is deterministic, the scan a net.
+		if err := e.procGroup.Load().Close(); err != nil {
+			logger.Debugf("closing the job's process group failed: %v", err)
+		}
 		e.terminateRunningProcesses(ctx)
 
-		// Only removes per-job misc state. Must not remove the cache/toolcache root.
-		// Bound it: CleanUp is a caller-supplied, typically unbounded os.RemoveAll,
-		// and a delete stalled by a filesystem filter driver would otherwise hang
-		// the job forever at "Cleaning up container" and hold the capacity slot.
+		// Removes per-job misc state only, never the toolcache root. Bounded because
+		// CleanUp is a caller-supplied, typically unbounded os.RemoveAll.
 		if e.CleanUp != nil {
 			logger.Debugf("running host environment cleanup callback")
 			if err := runWithTimeout(e.CleanUp, hostCleanupTimeout); err != nil {
@@ -603,8 +626,7 @@ func (e *HostEnvironment) Remove() common.Executor {
 				return errors.Join(errs...)
 			}
 		}
-		// Bounded teardown timed out; warnings already logged above. Do not
-		// fail job completion — leaked scratch is reclaimed by the idle sweep.
+		// Teardown timed out; warned above. Do not fail job completion over it.
 		return nil
 	}
 }

act/container/linux_container_environment_extensions.go

@@ -66,12 +66,15 @@ func (*LinuxContainerEnvironmentExtensions) JoinPathVariable(paths ...string) st
 	return strings.Join(paths, ":")
 }
 
+// DefaultToolCache is where the runner mounts the tool cache inside job containers.
+const DefaultToolCache = "/opt/hostedtoolcache"
+
 func (*LinuxContainerEnvironmentExtensions) GetRunnerContext(ctx context.Context) map[string]any {
 	return map[string]any{
 		"os":         "Linux",
 		"arch":       RunnerArch(ctx),
 		"temp":       "/tmp",
-		"tool_cache": "/opt/hostedtoolcache",
+		"tool_cache": DefaultToolCache,
 	}
 }
 

act/exprparser/functions.go

@@ -274,8 +274,17 @@ func (impl *interperterImpl) jobSuccess() (bool, error) { //nolint:unparam // pr
 	return true, nil
 }
 
+// jobStatus returns the current job status, treating a nil Job context as an
+// empty status so status-check functions never panic on a nil dereference.
+func (impl *interperterImpl) jobStatus() string {
+	if impl.env.Job == nil {
+		return ""
+	}
+	return impl.env.Job.Status
+}
+
 func (impl *interperterImpl) stepSuccess() (bool, error) { //nolint:unparam // pre-existing issue from nektos/act
-	return impl.env.Job.Status == "success", nil
+	return impl.jobStatus() == "success", nil
 }
 
 func (impl *interperterImpl) jobFailure() (bool, error) { //nolint:unparam // pre-existing issue from nektos/act
@@ -292,9 +301,9 @@ func (impl *interperterImpl) jobFailure() (bool, error) { //nolint:unparam // pr
 }
 
 func (impl *interperterImpl) stepFailure() (bool, error) { //nolint:unparam // pre-existing issue from nektos/act
-	return impl.env.Job.Status == "failure", nil
+	return impl.jobStatus() == "failure", nil
 }
 
 func (impl *interperterImpl) cancelled() (bool, error) { //nolint:unparam // pre-existing issue from nektos/act
-	return impl.env.Job.Status == "cancelled", nil
+	return impl.jobStatus() == "cancelled", nil
 }

act/exprparser/functions_test.go

@@ -254,3 +254,27 @@ func TestFunctionFormat(t *testing.T) {
 		})
 	}
 }
+
+func TestStatusFunctionsNilJob(t *testing.T) {
+	// A nil Job context must not panic: the status-check functions should treat
+	// it as an empty status and return false rather than dereferencing nil.
+	env := &EvaluationEnvironment{}
+
+	table := []struct {
+		input   string
+		context string
+		name    string
+	}{
+		{"cancelled()", "job", "cancelled-nil-job"},
+		{"success()", "step", "step-success-nil-job"},
+		{"failure()", "step", "step-failure-nil-job"},
+	}
+
+	for _, tt := range table {
+		t.Run(tt.name, func(t *testing.T) {
+			output, err := NewInterpeter(env, Config{Context: tt.context}).Evaluate(tt.input, DefaultStatusCheckNone)
+			assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
+			assert.Equal(t, false, output)
+		})
+	}
+}

act/exprparser/interpreter_test.go

@@ -11,6 +11,7 @@ import (
 	"gitea.com/gitea/runner/act/model"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func TestLiterals(t *testing.T) {
@@ -523,7 +524,9 @@ func TestOperatorsBooleanEvaluation(t *testing.T) {
 			assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
 
 			if expected, ok := tt.expected.(float64); ok && math.IsNaN(expected) {
-				assert.True(t, math.IsNaN(output.(float64)))
+				number, ok := output.(float64)
+				require.True(t, ok, "want a number, got %T", output)
+				assert.True(t, math.IsNaN(number))
 			} else {
 				assert.Equal(t, tt.expected, output)
 			}

act/model/action.go

@@ -76,17 +76,19 @@ func (a ActionRunsUsing) IsComposite() bool {
 
 // ActionRuns are a field in Action
 type ActionRuns struct {
-	Using      ActionRunsUsing   `yaml:"using"`
-	Env        map[string]string `yaml:"env"`
-	Main       string            `yaml:"main"`
-	Pre        string            `yaml:"pre"`
-	PreIf      string            `yaml:"pre-if"`
-	Post       string            `yaml:"post"`
-	PostIf     string            `yaml:"post-if"`
-	Image      string            `yaml:"image"`
-	Entrypoint string            `yaml:"entrypoint"`
-	Args       []string          `yaml:"args"`
-	Steps      []Step            `yaml:"steps"`
+	Using          ActionRunsUsing   `yaml:"using"`
+	Env            map[string]string `yaml:"env"`
+	Main           string            `yaml:"main"`
+	Pre            string            `yaml:"pre"`
+	PreIf          string            `yaml:"pre-if"`
+	Post           string            `yaml:"post"`
+	PostIf         string            `yaml:"post-if"`
+	Image          string            `yaml:"image"`
+	PreEntrypoint  string            `yaml:"pre-entrypoint"`
+	Entrypoint     string            `yaml:"entrypoint"`
+	PostEntrypoint string            `yaml:"post-entrypoint"`
+	Args           []string          `yaml:"args"`
+	Steps          []Step            `yaml:"steps"`
 }
 
 // Action describes a metadata file for GitHub actions. The metadata filename must be either action.yml or action.yaml. The data in the metadata file defines the inputs, outputs and main entrypoint for your action.

act/model/action_test.go

@@ -61,3 +61,22 @@ runs:
 		t.Fatalf("error = %q, want invalid value", err)
 	}
 }
+
+func TestReadActionDockerEntrypoints(t *testing.T) {
+	action, err := ReadAction(strings.NewReader(`
+runs:
+  using: docker
+  image: Dockerfile
+  pre-entrypoint: pre.sh
+  post-entrypoint: post.sh
+`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if action.Runs.PreEntrypoint != "pre.sh" {
+		t.Fatalf("pre-entrypoint = %q, want pre.sh", action.Runs.PreEntrypoint)
+	}
+	if action.Runs.PostEntrypoint != "post.sh" {
+		t.Fatalf("post-entrypoint = %q, want post.sh", action.Runs.PostEntrypoint)
+	}
+}

act/model/workflow.go

@@ -86,11 +86,12 @@ func (w *Workflow) OnSchedule() []string {
 	case []any:
 		allSchedules := []string{}
 		for _, v := range val {
-			for k, cron := range v.(map[string]any) {
-				if k != "cron" {
-					continue
-				}
-				allSchedules = append(allSchedules, cron.(string))
+			entry, ok := v.(map[string]any)
+			if !ok {
+				continue
+			}
+			if cron, ok := entry["cron"].(string); ok {
+				allSchedules = append(allSchedules, cron)
 			}
 		}
 		return allSchedules
@@ -443,9 +444,9 @@ func normalizeMatrixValue(key string, val any) ([]any, error) {
 // Scalar values are wrapped into single-element arrays automatically.
 // Template expressions are resolved by EvaluateYamlNode before this method is
 // called; if unresolved, the literal string is wrapped as a one-element fallback.
-func (j *Job) Matrix() map[string][]any {
+func (j *Job) Matrix() (map[string][]any, error) {
 	if j.Strategy == nil || j.Strategy.RawMatrix.Kind != yaml.MappingNode {
-		return nil
+		return map[string][]any{}, nil
 	}
 
 	// Decode to flexible map first so that scalar values don't cause a type error.
@@ -455,9 +456,9 @@ func (j *Job) Matrix() map[string][]any {
 		// Fall back to the strict array-only format for backward compatibility.
 		var val map[string][]any
 		if !decodeNode(j.Strategy.RawMatrix, &val) {
-			return nil
+			return map[string][]any{}, nil
 		}
-		return val
+		return val, nil
 	}
 
 	// Convert flexible format to expected format with validation
@@ -465,12 +466,11 @@ func (j *Job) Matrix() map[string][]any {
 	for k, v := range flexVal {
 		normalized, err := normalizeMatrixValue(k, v)
 		if err != nil {
-			log.Errorf("matrix validation error: %v", err)
-			return nil
+			return nil, err
 		}
 		val[k] = normalized
 	}
-	return val
+	return val, nil
 }
 
 // GetMatrixes returns the matrix cross product
@@ -482,38 +482,38 @@ func (j *Job) GetMatrixes() ([]map[string]any, error) {
 		j.Strategy.FailFast = j.Strategy.GetFailFast()
 		j.Strategy.MaxParallel = j.Strategy.GetMaxParallel()
 
-		if m := j.Matrix(); m != nil {
+		m, err := j.Matrix()
+		if err != nil {
+			return nil, err
+		}
+		if len(m) > 0 {
 			includes := make([]map[string]any, 0)
 			extraIncludes := make([]map[string]any, 0)
+			addInclude := func(raw any) error {
+				include, ok := raw.(map[string]any)
+				if !ok {
+					return fmt.Errorf("the workflow is not valid. Matrix include %v is not a map of matrix keys to values", raw)
+				}
+				for k := range include {
+					if _, ok := m[k]; ok {
+						includes = append(includes, include)
+						return nil
+					}
+				}
+				extraIncludes = append(extraIncludes, include)
+				return nil
+			}
 			for _, v := range m["include"] {
 				switch t := v.(type) {
 				case []any:
 					for _, i := range t {
-						i := i.(map[string]any)
-						extraInclude := true
-						for k := range i {
-							if _, ok := m[k]; ok {
-								includes = append(includes, i)
-								extraInclude = false
-								break
-							}
-						}
-						if extraInclude {
-							extraIncludes = append(extraIncludes, i)
+						if err := addInclude(i); err != nil {
+							return nil, err
 						}
 					}
 				case any:
-					v := v.(map[string]any)
-					extraInclude := true
-					for k := range v {
-						if _, ok := m[k]; ok {
-							includes = append(includes, v)
-							extraInclude = false
-							break
-						}
-					}
-					if extraInclude {
-						extraIncludes = append(extraIncludes, v)
+					if err := addInclude(t); err != nil {
+						return nil, err
 					}
 				}
 			}
@@ -521,10 +521,13 @@ func (j *Job) GetMatrixes() ([]map[string]any, error) {
 
 			excludes := make([]map[string]any, 0)
 			for _, e := range m["exclude"] {
-				e := e.(map[string]any)
-				for k := range e {
+				exclude, ok := e.(map[string]any)
+				if !ok {
+					return nil, fmt.Errorf("the workflow is not valid. Matrix exclude %v is not a map of matrix keys to values", e)
+				}
+				for k := range exclude {
 					if _, ok := m[k]; ok {
-						excludes = append(excludes, e)
+						excludes = append(excludes, exclude)
 					} else {
 						// We fail completely here because that's what GitHub does for non-existing matrix keys, fail on exclude, silent skip on include
 						return nil, fmt.Errorf("the workflow is not valid. Matrix exclude key %q does not match any key within the matrix", k)

act/model/workflow_test.go

@@ -667,7 +667,9 @@ func TestReadWorkflow_Strategy(t *testing.T) {
 	matrixes, err := job.GetMatrixes()
 	assert.NoError(t, err)                          //nolint:testifylint // pre-existing issue from nektos/act
 	assert.Equal(t, matrixes, []map[string]any{{}}) //nolint:testifylint // pre-existing issue from nektos/act
-	assert.Equal(t, job.Matrix(), map[string][]any(nil))
+	matrix, err := job.Matrix()
+	require.NoError(t, err)
+	assert.Empty(t, matrix)
 	assert.Equal(t, job.Strategy.MaxParallel, 2) //nolint:testifylint // pre-existing issue from nektos/act
 	assert.Equal(t, job.Strategy.FailFast, true) //nolint:testifylint // pre-existing issue from nektos/act
 
@@ -675,7 +677,9 @@ func TestReadWorkflow_Strategy(t *testing.T) {
 	matrixes, err = job.GetMatrixes()
 	assert.NoError(t, err)                          //nolint:testifylint // pre-existing issue from nektos/act
 	assert.Equal(t, matrixes, []map[string]any{{}}) //nolint:testifylint // pre-existing issue from nektos/act
-	assert.Equal(t, job.Matrix(), map[string][]any(nil))
+	matrix, err = job.Matrix()
+	require.NoError(t, err)
+	assert.Empty(t, matrix)
 	assert.Equal(t, job.Strategy.MaxParallel, 4)  //nolint:testifylint // pre-existing issue from nektos/act
 	assert.Equal(t, job.Strategy.FailFast, false) //nolint:testifylint // pre-existing issue from nektos/act
 
@@ -683,7 +687,9 @@ func TestReadWorkflow_Strategy(t *testing.T) {
 	matrixes, err = job.GetMatrixes()
 	assert.NoError(t, err)                          //nolint:testifylint // pre-existing issue from nektos/act
 	assert.Equal(t, matrixes, []map[string]any{{}}) //nolint:testifylint // pre-existing issue from nektos/act
-	assert.Equal(t, job.Matrix(), map[string][]any(nil))
+	matrix, err = job.Matrix()
+	require.NoError(t, err)
+	assert.Empty(t, matrix)
 	assert.Equal(t, job.Strategy.MaxParallel, 2)  //nolint:testifylint // pre-existing issue from nektos/act
 	assert.Equal(t, job.Strategy.FailFast, false) //nolint:testifylint // pre-existing issue from nektos/act
 
@@ -700,7 +706,9 @@ func TestReadWorkflow_Strategy(t *testing.T) {
 			{"datacenter": "site-b", "node-version": "12.x", "site": "dev"},
 		},
 	)
-	assert.Equal(t, job.Matrix(), //nolint:testifylint // pre-existing issue from nektos/act
+	matrix, err = job.Matrix()
+	require.NoError(t, err)
+	assert.Equal(t, matrix, //nolint:testifylint // pre-existing issue from nektos/act
 		map[string][]any{
 			"datacenter": {"site-c", "site-d"},
 			"exclude": {
@@ -1092,13 +1100,15 @@ jobs:
 				t.Fatal("job not found")
 			}
 
-			matrix := job.Matrix()
+			matrix, err := job.Matrix()
 
 			if tt.wantErr {
+				require.Error(t, err)
 				assert.Nil(t, matrix, "matrix should be nil on error")
 			} else {
+				require.NoError(t, err)
 				if tt.wantLen == 0 {
-					assert.Nil(t, matrix, "matrix should be nil for jobs without strategy")
+					assert.Empty(t, matrix, "no matrix for jobs without strategy")
 				} else {
 					assert.NotNil(t, matrix, "matrix should not be nil")
 					assert.Len(t, matrix, tt.wantLen, "matrix should have expected number of keys")
@@ -1130,11 +1140,9 @@ func TestJobMatrixValidation(t *testing.T) {
 			},
 		}
 
-		// Attempt to get matrix
-		matrix := job.Matrix()
-
-		// Should return nil due to validation error
-		assert.Nil(t, matrix, "matrix with nested map should return nil")
+		matrix, err := job.Matrix()
+		require.ErrorContains(t, err, `matrix key "config" has invalid nested object value`)
+		assert.Nil(t, matrix)
 	})
 }
 

act/runner/action.go

@@ -129,6 +129,16 @@ func readActionImpl(ctx context.Context, step *model.Step, actionDir, actionPath
 	return action, err
 }
 
+// cachedActionTar returns the action's tree from the action cache, which only a remote action
+// has an entry in.
+func cachedActionTar(ctx context.Context, step actionStep, name, includePrefix string) (io.ReadCloser, error) {
+	remote, ok := step.(*stepActionRemote)
+	if !ok {
+		return nil, fmt.Errorf("action %q is a remote action but runs as %T", name, step)
+	}
+	return step.getRunContext().Config.ActionCache.GetTarArchive(ctx, remote.cacheDir, remote.resolvedSha, includePrefix)
+}
+
 func maybeCopyToActionDir(ctx context.Context, step actionStep, actionDir, actionPath, containerActionDir string) error {
 	logger := common.Logger(ctx)
 	rc := step.getRunContext()
@@ -147,8 +157,7 @@ func maybeCopyToActionDir(ctx context.Context, step actionStep, actionDir, actio
 	}
 
 	if rc.Config != nil && rc.Config.ActionCache != nil {
-		raction := step.(*stepActionRemote)
-		ta, err := rc.Config.ActionCache.GetTarArchive(ctx, raction.cacheDir, raction.resolvedSha, "")
+		ta, err := cachedActionTar(ctx, step, stepModel.Uses, "")
 		if err != nil {
 			return err
 		}
@@ -207,7 +216,7 @@ func runActionImpl(step actionStep, actionDir string, remoteAction *remoteAction
 			if remoteAction == nil {
 				location = containerActionDir
 			}
-			return execAsDocker(ctx, step, actionName, actionDir, location, remoteAction == nil)
+			return execAsDocker(ctx, step, actionName, actionDir, location, remoteAction == nil, stepStageMain)
 		case x.IsComposite():
 			if err := maybeCopyToActionDir(ctx, step, actionDir, actionPath, containerActionDir); err != nil {
 				return err
@@ -305,7 +314,7 @@ func dockerActionImageTag(repository, actionName string, localAction bool) strin
 }
 
 // TODO: break out parts of function to reduce complexicity
-func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, basedir string, localAction bool) error {
+func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, basedir string, localAction bool, stage stepStage) error {
 	logger := common.Logger(ctx)
 	rc := step.getRunContext()
 	action := step.getActionModel()
@@ -351,8 +360,7 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 				}
 				defer buildContext.Close()
 			} else if rc.Config.ActionCache != nil {
-				rstep := step.(*stepActionRemote)
-				buildContext, err = rc.Config.ActionCache.GetTarArchive(ctx, rstep.cacheDir, rstep.resolvedSha, contextDir)
+				buildContext, err = cachedActionTar(ctx, step, actionName, contextDir)
 				if err != nil {
 					return err
 				}
@@ -364,6 +372,7 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 				ImageTag:     image,
 				BuildContext: buildContext,
 				Platform:     rc.Config.ContainerArchitecture,
+				BuildArgs:    rc.proxyBuildArgs(),
 			})
 			if buildContext == nil {
 				// Held across the whole build: the daemon drains contextDir lazily.
@@ -386,16 +395,9 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 		cmd = action.Runs.Args
 		evalDockerArgs(ctx, step, action, &cmd)
 	}
-	entrypoint := strings.Fields(eval.Interpolate(ctx, step.getStepModel().With["entrypoint"]))
-	if len(entrypoint) == 0 {
-		if action.Runs.Entrypoint != "" {
-			entrypoint, err = shellquote.Split(action.Runs.Entrypoint)
-			if err != nil {
-				return err
-			}
-		} else {
-			entrypoint = nil
-		}
+	entrypoint, err := dockerEntrypoint(ctx, step, eval, stage)
+	if err != nil {
+		return err
 	}
 	stepContainer := newStepContainer(ctx, step, image, cmd, entrypoint)
 	return common.NewPipelineExecutor(
@@ -405,10 +407,34 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 		stepContainer.Create(rc.Config.ContainerCapAdd, rc.Config.ContainerCapDrop),
 		stepContainer.Start(true),
 	).Finally(
-		stepContainer.Remove().IfBool(!rc.Config.ReuseContainers),
+		stepContainer.Remove().IfBool(!rc.Config.ReuseContainers && !rc.Config.AutoRemove),
 	).Finally(stepContainer.Close())(ctx)
 }
 
+// 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) {
+	runs := step.getActionModel().Runs
+
+	var entrypoint string
+	switch stage {
+	case stepStagePre:
+		entrypoint = runs.PreEntrypoint
+	case stepStagePost:
+		entrypoint = runs.PostEntrypoint
+	default:
+		if fields := strings.Fields(eval.Interpolate(ctx, step.getStepModel().With["entrypoint"])); len(fields) > 0 {
+			return fields, nil
+		}
+		entrypoint = runs.Entrypoint
+	}
+
+	if entrypoint == "" {
+		return nil, nil
+	}
+	return shellquote.Split(entrypoint)
+}
+
 func evalDockerArgs(ctx context.Context, step step, action *model.Action, cmd *[]string) {
 	rc := step.getRunContext()
 	stepModel := step.getStepModel()
@@ -455,10 +481,7 @@ func newStepContainer(ctx context.Context, step step, image string, cmd, entrypo
 		envList = append(envList, fmt.Sprintf("%s=%s", k, v))
 	}
 
-	envList = append(envList, fmt.Sprintf("%s=%s", "RUNNER_TOOL_CACHE", "/opt/hostedtoolcache"))
-	envList = append(envList, fmt.Sprintf("%s=%s", "RUNNER_OS", "Linux"))
-	envList = append(envList, fmt.Sprintf("%s=%s", "RUNNER_ARCH", container.RunnerArch(ctx)))
-	envList = append(envList, fmt.Sprintf("%s=%s", "RUNNER_TEMP", "/tmp"))
+	envList = append(envList, rc.runnerEnv(ctx)...)
 
 	binds, mounts := rc.GetBindsAndMounts()
 	networkMode := "container:" + rc.jobContainerName()
@@ -559,44 +582,57 @@ func hasPreStep(step actionStep) common.Conditional {
 		return action.Runs.Using.IsComposite() ||
 			(action.Runs.Using.IsNode() &&
 				action.Runs.Pre != "") ||
+			(action.Runs.Using.IsDocker() &&
+				action.Runs.PreEntrypoint != "") ||
 			(action.Runs.Using == model.ActionRunsUsingGo &&
 				action.Runs.Pre != "")
 	}
 }
 
+// actionStagePaths resolves where a step's action lives and where the job container sees
+// it, for the pre and post stage.
+func actionStagePaths(step actionStep) (actionDir, actionPath, actionName, containerActionDir string) {
+	rc := step.getRunContext()
+	stepModel := step.getStepModel()
+
+	if sar, ok := step.(*stepActionRemote); ok {
+		actionDir = sar.actionDir()
+		actionPath = newRemoteAction(stepModel.Uses).Path
+	} else {
+		actionDir = filepath.Join(rc.Config.Workdir, stepModel.Uses)
+	}
+
+	actionName, containerActionDir = getContainerActionPaths(stepModel, path.Join(actionDir, actionPath), rc)
+	return actionDir, actionPath, actionName, containerActionDir
+}
+
+// execDockerActionStage runs a docker action's image for its pre or post stage.
+func execDockerActionStage(ctx context.Context, step actionStep, stage stepStage) error {
+	actionDir, actionPath, actionName, containerActionDir := actionStagePaths(step)
+
+	_, remote := step.(*stepActionRemote)
+	location := containerActionDir
+	if remote {
+		location = path.Join(actionDir, actionPath)
+	}
+	return execAsDocker(ctx, step, actionName, actionDir, location, !remote, stage)
+}
+
 func runPreStep(step actionStep) common.Executor {
 	return func(ctx context.Context) error {
 		logger := common.Logger(ctx)
 		logger.Debugf("run pre step for '%s'", step.getStepModel())
 
 		rc := step.getRunContext()
-		stepModel := step.getStepModel()
 		action := step.getActionModel()
 
+		actionDir, actionPath, _, containerActionDir := actionStagePaths(step)
+
 		x := action.Runs.Using
 		switch {
 		case x.IsNode():
 			// defaults in pre steps were missing, however provided inputs are available
 			populateEnvsFromInput(ctx, step.getEnv(), action, rc)
-			// todo: refactor into step
-			var actionDir string
-			var actionPath string
-			if _, ok := step.(*stepActionRemote); ok {
-				actionPath = newRemoteAction(stepModel.Uses).Path
-				actionDir = fmt.Sprintf("%s/%s", rc.ActionCacheDir(), stepModel.UsesHash())
-			} else {
-				actionDir = filepath.Join(rc.Config.Workdir, stepModel.Uses)
-				actionPath = ""
-			}
-
-			var actionLocation string
-			if actionPath != "" {
-				actionLocation = path.Join(actionDir, actionPath)
-			} else {
-				actionLocation = actionDir
-			}
-
-			_, containerActionDir := getContainerActionPaths(stepModel, actionLocation, rc)
 
 			if err := maybeCopyToActionDir(ctx, step, actionDir, actionPath, containerActionDir); err != nil {
 				return err
@@ -609,6 +645,12 @@ func runPreStep(step actionStep) common.Executor {
 
 			return rc.execJobContainer(containerArgs, *step.getEnv(), "", "")(ctx)
 
+		case x.IsDocker():
+			// defaults in pre steps were missing, however provided inputs are available
+			populateEnvsFromInput(ctx, step.getEnv(), action, rc)
+
+			return execDockerActionStage(ctx, step, stepStagePre)
+
 		case x.IsComposite():
 			if step.getCompositeSteps() == nil {
 				step.getCompositeRunContext(ctx)
@@ -622,25 +664,6 @@ func runPreStep(step actionStep) common.Executor {
 		case x == model.ActionRunsUsingGo:
 			// defaults in pre steps were missing, however provided inputs are available
 			populateEnvsFromInput(ctx, step.getEnv(), action, rc)
-			// todo: refactor into step
-			var actionDir string
-			var actionPath string
-			if _, ok := step.(*stepActionRemote); ok {
-				actionPath = newRemoteAction(stepModel.Uses).Path
-				actionDir = fmt.Sprintf("%s/%s", rc.ActionCacheDir(), stepModel.UsesHash())
-			} else {
-				actionDir = filepath.Join(rc.Config.Workdir, stepModel.Uses)
-				actionPath = ""
-			}
-
-			var actionLocation string
-			if actionPath != "" {
-				actionLocation = path.Join(actionDir, actionPath)
-			} else {
-				actionLocation = actionDir
-			}
-
-			_, containerActionDir := getContainerActionPaths(stepModel, actionLocation, rc)
 
 			if err := maybeCopyToActionDir(ctx, step, actionDir, actionPath, containerActionDir); err != nil {
 				return err
@@ -693,6 +716,8 @@ func hasPostStep(step actionStep) common.Conditional {
 		return action.Runs.Using.IsComposite() ||
 			(action.Runs.Using.IsNode() &&
 				action.Runs.Post != "") ||
+			(action.Runs.Using.IsDocker() &&
+				action.Runs.PostEntrypoint != "") ||
 			(action.Runs.Using == model.ActionRunsUsingGo &&
 				action.Runs.Post != "")
 	}
@@ -704,28 +729,9 @@ func runPostStep(step actionStep) common.Executor {
 		logger.Debugf("run post step for '%s'", step.getStepModel())
 
 		rc := step.getRunContext()
-		stepModel := step.getStepModel()
 		action := step.getActionModel()
 
-		// todo: refactor into step
-		var actionDir string
-		var actionPath string
-		if _, ok := step.(*stepActionRemote); ok {
-			actionPath = newRemoteAction(stepModel.Uses).Path
-			actionDir = fmt.Sprintf("%s/%s", rc.ActionCacheDir(), stepModel.UsesHash())
-		} else {
-			actionDir = filepath.Join(rc.Config.Workdir, stepModel.Uses)
-			actionPath = ""
-		}
-
-		var actionLocation string
-		if actionPath != "" {
-			actionLocation = path.Join(actionDir, actionPath)
-		} else {
-			actionLocation = actionDir
-		}
-
-		_, containerActionDir := getContainerActionPaths(stepModel, actionLocation, rc)
+		actionDir, actionPath, _, containerActionDir := actionStagePaths(step)
 
 		x := action.Runs.Using
 		switch {
@@ -740,6 +746,11 @@ func runPostStep(step actionStep) common.Executor {
 
 			return rc.execJobContainer(containerArgs, *step.getEnv(), "", "")(ctx)
 
+		case x.IsDocker():
+			populateEnvsFromSavedState(step.getEnv(), step, rc)
+
+			return execDockerActionStage(ctx, step, stepStagePost)
+
 		case x.IsComposite():
 			if err := maybeCopyToActionDir(ctx, step, actionDir, actionPath, containerActionDir); err != nil {
 				return err

act/runner/action_test.go

@@ -20,6 +20,7 @@ import (
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/mock"
+	"github.com/stretchr/testify/require"
 )
 
 type closerMock struct {
@@ -150,6 +151,44 @@ runs:
 	}
 }
 
+// With AutoRemove the daemon reaps the container on exit, so act must not remove it afterwards.
+func TestExecAsDockerAutoRemove(t *testing.T) {
+	orig := ContainerNewContainer
+	defer func() { ContainerNewContainer = orig }()
+
+	for _, tc := range []struct {
+		autoRemove bool
+		removes    int
+	}{
+		{false, 2}, // stale + post-run
+		{true, 1},  // post-run skipped
+	} {
+		cm := &containerMock{}
+		ContainerNewContainer = func(*container.NewContainerInput) container.ExecutionsEnvironment { return cm }
+
+		step := &stepActionRemote{
+			Step: &model.Step{ID: "1", Uses: "org/action@v1"},
+			RunContext: &RunContext{
+				Config:       &Config{AutoRemove: tc.autoRemove},
+				Run:          &model.Run{JobID: "1", Workflow: &model.Workflow{Jobs: map[string]*model.Job{"1": {}}}},
+				JobContainer: cm,
+			},
+			action: &model.Action{Runs: model.ActionRuns{Using: "docker", Image: "docker://node:14"}},
+		}
+
+		removes := 0
+		cm.On("Pull", false).Return(func(context.Context) error { return nil })
+		cm.On("Remove").Return(func(context.Context) error { removes++; return nil })
+		cm.On("Create", []string(nil), []string(nil)).Return(func(context.Context) error { return nil })
+		cm.On("Start", true).Return(func(context.Context) error { return nil })
+		cm.On("Close").Return(func(context.Context) error { return nil })
+
+		require.NoError(t, execAsDocker(context.Background(), step, "action", t.TempDir(), t.TempDir(), false, stepStageMain))
+		cm.AssertExpectations(t)
+		assert.Equal(t, tc.removes, removes)
+	}
+}
+
 func TestActionRunner(t *testing.T) {
 	table := []struct {
 		name        string
@@ -426,7 +465,7 @@ func TestExecAsDockerHoldsCloneLockForRemoteUncached(t *testing.T) {
 	defer cancel()
 
 	done := make(chan error, 1)
-	go func() { done <- execAsDocker(ctx, step, "test-action", actionDir, actionDir, false) }()
+	go func() { done <- execAsDocker(ctx, step, "test-action", actionDir, actionDir, false, stepStageMain) }()
 
 	select {
 	case <-innerEntered:
@@ -502,3 +541,86 @@ func TestDockerActionImageTag(t *testing.T) {
 		dockerActionImageTag("owner/repo", "./sub", true),
 	)
 }
+
+// Only the entrypoint is stage specific: every stage of a docker action receives runs.args
+// and runs.env, and the `entrypoint` input applies to the main stage alone.
+func TestExecAsDockerStageEntrypoint(t *testing.T) {
+	orig := ContainerNewContainer
+	defer func() { ContainerNewContainer = orig }()
+
+	for _, tc := range []struct {
+		name           string
+		stage          stepStage
+		wantEntrypoint []string
+	}{
+		{
+			name:           "main stage prefers the entrypoint input",
+			stage:          stepStageMain,
+			wantEntrypoint: []string{"input.sh"},
+		},
+		{
+			name:           "pre stage uses runs.pre-entrypoint",
+			stage:          stepStagePre,
+			wantEntrypoint: []string{"pre.sh", "--verbose"},
+		},
+		{
+			name:           "post stage uses runs.post-entrypoint",
+			stage:          stepStagePost,
+			wantEntrypoint: []string{"post.sh"},
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			cm := &containerMock{}
+			var input *container.NewContainerInput
+			ContainerNewContainer = func(in *container.NewContainerInput) container.ExecutionsEnvironment {
+				input = in
+				return cm
+			}
+
+			step := &stepActionRemote{
+				Step: &model.Step{ID: "1", Uses: "org/action@v1", With: map[string]string{"entrypoint": "input.sh"}},
+				RunContext: &RunContext{
+					Config:       &Config{},
+					Run:          &model.Run{JobID: "1", Workflow: &model.Workflow{Jobs: map[string]*model.Job{"1": {}}}},
+					JobContainer: cm,
+				},
+				action: &model.Action{Runs: model.ActionRuns{
+					Using:          "docker",
+					Image:          "docker://node:14",
+					PreEntrypoint:  "pre.sh --verbose",
+					Entrypoint:     "main.sh",
+					PostEntrypoint: "post.sh",
+					Args:           []string{"hello"},
+					Env:            map[string]string{"MY_VAR": "world"},
+				}},
+				env: map[string]string{},
+			}
+
+			cm.On("Pull", false).Return(func(context.Context) error { return nil })
+			cm.On("Remove").Return(func(context.Context) error { return nil })
+			cm.On("Create", []string(nil), []string(nil)).Return(func(context.Context) error { return nil })
+			cm.On("Start", true).Return(func(context.Context) error { return nil })
+			cm.On("Close").Return(func(context.Context) error { return nil })
+
+			require.NoError(t, execAsDocker(context.Background(), step, "action", t.TempDir(), t.TempDir(), false, tc.stage))
+			require.NotNil(t, input)
+			assert.Equal(t, tc.wantEntrypoint, input.Entrypoint)
+			assert.Equal(t, []string{"hello"}, input.Cmd)
+			assert.Contains(t, input.Env, "MY_VAR=world")
+		})
+	}
+}
+
+func TestDockerActionHasPreAndPostStep(t *testing.T) {
+	newStep := func(runs model.ActionRuns) actionStep {
+		return &stepActionRemote{action: &model.Action{Runs: runs}}
+	}
+	ctx := context.Background()
+
+	assert.False(t, hasPreStep(newStep(model.ActionRuns{Using: "docker", Image: "Dockerfile"}))(ctx))
+	assert.False(t, hasPostStep(newStep(model.ActionRuns{Using: "docker", Image: "Dockerfile"}))(ctx))
+
+	withStages := model.ActionRuns{Using: "docker", Image: "Dockerfile", PreEntrypoint: "pre.sh", PostEntrypoint: "post.sh"}
+	assert.True(t, hasPreStep(newStep(withStages))(ctx))
+	assert.True(t, hasPostStep(newStep(withStages))(ctx))
+}

act/runner/cancellation_test.go

@@ -283,3 +283,30 @@ func TestPostStepsContextDeadlinePreservesJobError(t *testing.T) {
 	require.NoError(t, postCtx.Err(), "post context must not carry the expired deadline")
 	assert.ErrorIs(t, common.JobError(postCtx), assert.AnError, "the timeout job error must be preserved")
 }
+
+// reportStepError must treat a context.Canceled (e.g. a teardown-cancelled read) as an
+// interruption, never a job failure.
+func TestReportStepErrorTreatsCancelAsInterruption(t *testing.T) {
+	rc := &RunContext{}
+
+	// stray read cancellation while the job context is live: ignored, not a failure
+	live := common.WithJobErrorContainer(context.Background())
+	reportStepError(live, rc, context.Canceled)
+	require.NoError(t, common.JobError(live))
+	assert.False(t, rc.jobFailed)
+	assert.False(t, rc.jobCancelled)
+
+	// genuine job cancellation: recorded as cancelled, still not a failure
+	cancelled, cancel := context.WithCancel(common.WithJobErrorContainer(context.Background()))
+	cancel()
+	reportStepError(cancelled, rc, context.Canceled)
+	require.NoError(t, common.JobError(cancelled))
+	assert.False(t, rc.jobFailed)
+	assert.True(t, rc.jobCancelled)
+
+	// a real error still fails the job
+	failed := common.WithJobErrorContainer(context.Background())
+	reportStepError(failed, rc, assert.AnError)
+	require.ErrorIs(t, common.JobError(failed), assert.AnError)
+	assert.True(t, rc.jobFailed)
+}

act/runner/command.go

@@ -154,30 +154,25 @@ func parseKeyValuePairs(kvPairs, separator string) map[string]string {
 	return rtn
 }
 
+// A Replacer never rescans what it wrote, so "%250A" stays a literal "%0A".
+var (
+	commandDataEscaper       = strings.NewReplacer("%", "%25", "\r", "%0D", "\n", "%0A")
+	commandDataUnescaper     = strings.NewReplacer("%25", "%", "%0D", "\r", "%0A", "\n")
+	commandPropertyUnescaper = strings.NewReplacer("%25", "%", "%0D", "\r", "%0A", "\n", "%3A", ":", "%2C", ",")
+)
+
+// escapeCommandData encodes the data part of a "::cmd::" or "##[cmd]" line the runner writes itself,
+// so the log renderer decodes it back. Lines forwarded from step output are already escaped.
+func escapeCommandData(arg string) string {
+	return commandDataEscaper.Replace(arg)
+}
+
 func UnescapeCommandData(arg string) string {
-	escapeMap := map[string]string{
-		"%25": "%",
-		"%0D": "\r",
-		"%0A": "\n",
-	}
-	for k, v := range escapeMap {
-		arg = strings.ReplaceAll(arg, k, v)
-	}
-	return arg
+	return commandDataUnescaper.Replace(arg)
 }
 
 func unescapeCommandProperty(arg string) string {
-	escapeMap := map[string]string{
-		"%25": "%",
-		"%0D": "\r",
-		"%0A": "\n",
-		"%3A": ":",
-		"%2C": ",",
-	}
-	for k, v := range escapeMap {
-		arg = strings.ReplaceAll(arg, k, v)
-	}
-	return arg
+	return commandPropertyUnescaper.Replace(arg)
 }
 
 func unescapeKvPairs(kvPairs map[string]string) map[string]string {

act/runner/command_test.go

@@ -214,3 +214,10 @@ func TestSaveState(t *testing.T) {
 
 	assert.Equal(t, "state-value", rc.IntraActionState["step"]["state-name"])
 }
+
+func TestEscapeCommandData(t *testing.T) {
+	a := assert.New(t)
+
+	a.Equal("a%25b%0Dc%0Ad%250A", escapeCommandData("a%b\rc\nd%0A"))
+	a.Equal("a%b\rc\nd%0A", UnescapeCommandData("a%25b%0Dc%0Ad%250A"))
+}

act/runner/expression.go

@@ -95,9 +95,7 @@ func (rc *RunContext) NewExpressionEvaluatorWithEnv(ctx context.Context, env map
 		Inputs:    inputs,
 		HashFiles: getHashFilesFunction(ctx, rc),
 	}
-	if rc.JobContainer != nil {
-		ee.Runner = rc.JobContainer.GetRunnerContext(ctx)
-	}
+	ee.Runner = rc.getRunnerContext(ctx)
 	return expressionEvaluator{
 		interpreter: exprparser.NewInterpeter(ee, exprparser.Config{
 			Run:        rc.Run,
@@ -149,9 +147,7 @@ func (rc *RunContext) NewStepExpressionEvaluator(ctx context.Context, step step)
 		Inputs:    inputs,
 		HashFiles: getHashFilesFunction(ctx, rc),
 	}
-	if rc.JobContainer != nil {
-		ee.Runner = rc.JobContainer.GetRunnerContext(ctx)
-	}
+	ee.Runner = rc.getRunnerContext(ctx)
 	return expressionEvaluator{
 		interpreter: exprparser.NewInterpeter(ee, exprparser.Config{
 			Run:        rc.Run,

act/runner/helpers_test.go

@@ -5,11 +5,9 @@ package runner
 
 import (
 	"context"
-	"net"
 	"os/exec"
 	"runtime"
 	"testing"
-	"time"
 
 	"gitea.com/gitea/runner/act/container"
 
@@ -42,18 +40,6 @@ func requireDocker(t *testing.T) {
 	}
 }
 
-// requireNetwork skips the test unless github.com is reachable. A few tests exercise behaviour
-// that inherently needs the network (force-pulling an image, resolving a remote short-sha ref);
-// gating lets the rest of the suite run offline without these failing.
-func requireNetwork(t *testing.T) {
-	t.Helper()
-	conn, err := net.DialTimeout("tcp", "github.com:443", 3*time.Second)
-	if err != nil {
-		t.Skipf("skipping: network unavailable: %v", err)
-	}
-	_ = conn.Close()
-}
-
 // requireHostTools skips the test unless every named executable is on PATH. Used by the
 // self-hosted (host environment) suite, which runs steps directly on the host.
 func requireHostTools(t *testing.T, tools ...string) {

act/runner/job_executor.go

@@ -10,6 +10,7 @@ import (
 	"context"
 	"encoding/base64"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -56,17 +57,81 @@ type jobInfo interface {
 	result(result string)
 }
 
-// reportStepError emits the GitHub Actions ##[error] annotation and records
-// the error against the job so the job is reported as failed.
+// reportStepError records a step error so the job is reported failed — except a
+// cancellation, which is an interruption, not a failure.
 func reportStepError(ctx context.Context, rc *RunContext, err error) {
-	common.Logger(ctx).Errorf("##[error]%v", err)
+	if errors.Is(err, context.Canceled) {
+		// Defer to the job context: a genuine cancel reports cancelled, a stray teardown
+		// cancellation on a live ctx is ignored — never a step FAILURE.
+		rc.markInterrupted(ctx.Err())
+		return
+	}
+	common.Logger(ctx).Errorf("##[error]%s", escapeCommandData(err.Error()))
 	common.SetJobError(ctx, err)
 	rc.markFailed()
 }
 
+// actionPreparer is implemented by steps that download an action before they run, so the job
+// executor can fetch all of them up front.
+type actionPreparer interface {
+	prepareActionExecutor() common.Executor
+	actionDownloadInfo() (reference, sha string, ok bool)
+}
+
+// printPrepareActions downloads every action the job uses before its first step runs and reports
+// them as actions/runner's "Prepare all required actions" section does. The steps still call
+// prepareActionExecutor themselves; it is a no-op once the action is resolved here.
+func printPrepareActions(rc *RunContext, preparers []actionPreparer) common.Executor {
+	return func(ctx context.Context) error {
+		if len(preparers) == 0 {
+			return nil
+		}
+
+		rawLogger := common.Logger(ctx).WithField(rawOutputField, true)
+		rawLogger.Infof("Prepare all required actions")
+
+		for _, preparer := range preparers {
+			if err := preparer.prepareActionExecutor()(ctx); err != nil {
+				// No step has run yet, so the failure belongs to the job.
+				reportStepError(ctx, rc, err)
+				return err
+			}
+			reference, sha, ok := preparer.actionDownloadInfo()
+			if !ok {
+				continue
+			}
+			if sha == "" {
+				rawLogger.Infof("Download action repository '%s'", reference)
+			} else {
+				rawLogger.Infof("Download action repository '%s' (SHA:%s)", reference, sha)
+			}
+		}
+		return nil
+	}
+}
+
+// printCompleteJobName closes the setup section the way actions/runner ends its "Set up job" step.
+func printCompleteJobName(rc *RunContext) common.Executor {
+	return func(ctx context.Context) error {
+		// Name holds a matrix combination; JobName is the shared name GitHub reports.
+		name := rc.JobName
+		if name == "" {
+			name = rc.Name
+		}
+		if name == "" && rc.Run != nil {
+			name = rc.Run.JobID
+		}
+		common.Logger(ctx).WithField(rawOutputField, true).Infof("Complete job name: %s", name)
+		return nil
+	}
+}
+
 func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executor {
 	steps := make([]common.Executor, 0)
 	preSteps := make([]common.Executor, 0)
+	// Collected separately: every action is downloaded before the first pre step runs.
+	stepPreSteps := make([]common.Executor, 0)
+	preparers := make([]actionPreparer, 0)
 	var postExecutor common.Executor
 
 	steps = append(steps, func(ctx context.Context) error {
@@ -113,9 +178,13 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 			return common.NewErrorExecutor(err)
 		}
 
+		if preparer, ok := step.(actionPreparer); ok {
+			preparers = append(preparers, preparer)
+		}
+
 		stepIdx := stepModel.Number
 		preExec := step.pre()
-		preSteps = append(preSteps, useStepLogger(rc, stepModel, stepStagePre, func(ctx context.Context) error {
+		stepPreSteps = append(stepPreSteps, useStepLogger(rc, stepModel, stepStagePre, func(ctx context.Context) error {
 			rc.CurrentStepIndex = stepIdx
 			preErr := preExec(ctx)
 			if preErr != nil {
@@ -157,6 +226,16 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 		}
 	}
 
+	// The setup section of the job log. The started hook goes first, so what it sets up is
+	// in place for the first action download and the first step.
+	preSteps = append(preSteps, rc.runJobStartedHook)
+	preSteps = append(preSteps, printPrepareActions(rc, preparers))
+	preSteps = append(preSteps, stepPreSteps...)
+	preSteps = append(preSteps, printCompleteJobName(rc))
+
+	// Ahead of the teardown below, while the job environment is still up.
+	postExecutor = postExecutor.Finally(rc.runJobCompletedHook)
+
 	postExecutor = postExecutor.Finally(func(ctx context.Context) error {
 		jobError := common.JobError(ctx)
 		var err error