fix: serve the whole results service from the cache server (#1141)

`ACTIONS_RESULTS_URL` names one origin serving every `github.actions.results.api.v1` service. Gitea serves the artifact half and this runner the cache half, so announcing `ACTIONS_CACHE_SERVICE_V2` while that URL pointed at Gitea was a promise the environment could not keep, and `docker buildx` posted its cache calls at Gitea and got a 404.

The cache server now forwards the artifact half to the instance each job registers with, so it is the whole results service and jobs are pointed at it. The announcement follows, and the bundle patch follows the cache URL instead.

Also fixes three things no JavaScript client reached: camelCase in the v2 responses where the Go clients read proto names, the missing `x-ms-request-id` on blob uploads that panics buildkit, and `cache.external_server` passed through without the trailing slash the v1 client concatenates onto.

Tests run the real actions against the services they look for: `actions/cache` over both API versions, the artifact actions up and back down through the forwarding, and `setup-node`. The regression itself is covered by asserting that whatever a job is handed as `ACTIONS_RESULTS_URL` answers a cache service call.

Fixes https://gitea.com/gitea/runner/issues/1139

Reviewed-on: https://gitea.com/gitea/runner/pulls/1141
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>

.gitea/workflows/release-nightly.yml

@@ -85,7 +85,7 @@ jobs:
         uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4
 
       - name: Login to DockerHub
-        uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4
+        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4
         with:
           username: ${{ secrets.DOCKER_USERNAME }}
           password: ${{ secrets.DOCKER_PASSWORD }}

.gitea/workflows/release-tag.yml

@@ -88,7 +88,7 @@ jobs:
         uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4
 
       - name: Login to DockerHub
-        uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4
+        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4
         with:
           username: ${{ secrets.DOCKER_USERNAME }}
           password: ${{ secrets.DOCKER_PASSWORD }}

.gitea/workflows/test.yml

@@ -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
@@ -36,12 +37,8 @@ linters:
       rules:
         main:
           deny:
-            - pkg: io/ioutil
-              desc: use os or io instead
             - pkg: golang.org/x/exp
               desc: it's experimental and unreliable
-            - pkg: github.com/pkg/errors
-              desc: use builtin errors package instead
     nolintlint:
       allow-unused: false
       require-explanation: true
@@ -102,6 +99,9 @@ linters:
       - linters:
           - forbidigo
         path: cmd
+      - linters:
+          - forcetypeassert
+        path: _test\.go
 issues:
   max-issues-per-linter: 0
   max-same-issues: 0

AGENTS.md

@@ -1,10 +1,18 @@
+- Never assume, verify before claiming
 - Use `make help` to find available development targets
-- Run `make fmt` to format `.go` files, and run `make lint-go` to lint them
-- Run `make tidy` after any `go.mod` changes
-- Run single go unit tests with `go test -run '^TestName$' ./modulepath/`
+- PR descriptions: minimal, only what and why, no task lists or file listings
+- Reference issues and PRs by full URL, not by number
+- Use Conventional Commits for commit messages and PR titles, plus the `enhance` type for user-facing enhancements
+- Add an `Assisted-by: AGENT_NAME:MODEL_VERSION` trailer to commit messages, never `Co-Authored-By` or `Signed-off-by`
+- Attribute agent authorship on one trailing line in issue and pull request comments, never as a PR description section
+- Never force-push, amend, or squash unless asked. Use new commits and normal push for pull request updates
+- Comments: write almost none, short and preferably same-line, explaining why for a future reader. Never narrate code, the change or the prompt. Preserve existing ones that still apply
 - Add the current year into the copyright header of new `.go` files
 - Ensure no trailing whitespace in edited files
-- Never force-push, amend, or squash unless asked. Use new commits and normal push for pull request updates
-- Preserve existing code comments, do not remove or rewrite comments that are still relevant
-- Include authorship attribution in issue and pull request comments
-- Add `Co-Authored-By` lines to all commits, indicating name and model used
+- Run `make fmt` after `.go` edits, `make tidy` after `go.mod` edits, and `make checks` for the non-lint source checks
+- Lint what changed with `make lint-go`, and `make lint-go-windows` for Windows and platform-split files
+- Fix the cause rather than disabling a linter or weakening a test. Where unavoidable, use the narrowest scope with a trailing comment giving the reason
+- Run single go tests with `go test -run '^TestName$' ./modulepath/`. `make test` self-skips the integration tests without docker or network, `make test-dind` runs the daemon-facing tests against the built dind image
+- Write the fewest, fastest tests covering the behavior, extending an existing one where possible. Prefer unit tests where logic is testable in isolation
+- Wait on a deterministic condition rather than `sleep`
+- Update the files under `docs/` when behavior documented there changes

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]6 # 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/govulncheck@v1.3.0
+GOLANGCI_LINT_PACKAGE ?= github.com/golangci/golangci-lint/v2/cmd/[email protected] # renovate: datasource=go
+GOVULNCHECK_PACKAGE ?= golang.org/x/vuln/cmd/govulncheck@v1.6.0 # 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.
@@ -254,11 +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`)
 
@@ -270,6 +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

@@ -52,7 +52,13 @@ type credKey struct{}
 // poison another repo's cache, even from inside a container that reaches the
 // cache server over the docker bridge network.
 type JobCredential struct {
-	Repo string
+	Repo string `json:"repo"`
+
+	// Results is the instance whose artifact service this server forwards for the job, and
+	// InsecureTLS how the runner reaches it; see results.go. The tags are the wire format a
+	// remote runner registers with.
+	Results     string `json:"results"`
+	InsecureTLS bool   `json:"insecure_tls"`
 }
 
 // credEntry holds a registered job's credential along with an active
@@ -70,6 +76,7 @@ type Handler struct {
 	storage  *Storage
 	router   *httprouter.Router
 	listener net.Listener
+	port     int
 	server   *http.Server
 	logger   logrus.FieldLogger
 
@@ -157,12 +164,14 @@ 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)
+	router.NotFound = http.HandlerFunc(h.forwardOrNotFound)
 
 	h.router = router
 
@@ -177,6 +186,12 @@ 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 +209,7 @@ 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
@@ -205,10 +218,11 @@ func (h *Handler) ExternalURL() string {
 // is only accepted while the job is running.
 //
 // Registrations are reference-counted: if a token is already registered, the
-// existing repo is kept and the refcount is incremented. The entry is
-// removed only when every revoker returned by RegisterJob has been called.
+// credential it was registered with is kept and the refcount is incremented.
+// The entry is removed only when every revoker returned by RegisterJob has
+// been called.
 // This keeps a stray re-registration from silently revoking a live job.
-func (h *Handler) RegisterJob(token, repo string) func() {
+func (h *Handler) RegisterJob(token string, cred JobCredential) func() {
 	if h == nil || token == "" {
 		return func() {}
 	}
@@ -217,7 +231,7 @@ func (h *Handler) RegisterJob(token, repo string) func() {
 		existing.refs++
 	} else {
 		h.creds[token] = &credEntry{
-			cred: JobCredential{Repo: repo},
+			cred: cred,
 			refs: 1,
 		}
 	}
@@ -334,7 +348,7 @@ 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 +357,6 @@ 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 +364,25 @@ 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 +457,7 @@ 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 +498,7 @@ 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 +506,28 @@ 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 +537,7 @@ 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 +571,9 @@ 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 +596,7 @@ 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
@@ -597,7 +627,7 @@ func (h *Handler) internalAuth(handler httprouter.Handle) httprouter.Handle {
 
 type internalRegisterBody struct {
 	Token string `json:"token"`
-	Repo  string `json:"repo"`
+	JobCredential
 }
 
 type internalRevokeBody struct {
@@ -605,6 +635,15 @@ type internalRevokeBody struct {
 }
 
 // POST /_internal/register
+// ResultsURL is what a job registered with cred should be given as ACTIONS_RESULTS_URL, or "" when
+// the credential names no instance to forward the artifact half to.
+func (h *Handler) ResultsURL(cred JobCredential) string {
+	if h == nil || cred.Results == "" {
+		return ""
+	}
+	return h.ExternalURL()
+}
+
 func (h *Handler) internalRegister(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
 	var body internalRegisterBody
 	if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
@@ -615,8 +654,9 @@ func (h *Handler) internalRegister(w http.ResponseWriter, r *http.Request, _ htt
 		h.responseJSON(w, r, http.StatusBadRequest, errors.New("token is required"))
 		return
 	}
-	h.RegisterJob(body.Token, body.Repo)
-	h.responseJSON(w, r, http.StatusOK)
+	h.RegisterJob(body.Token, body.JobCredential)
+	// A server too old to forward answers without this, which is how the caller knows.
+	h.responseJSON(w, r, http.StatusOK, map[string]any{"results_url": h.ResultsURL(body.JobCredential)})
 }
 
 // POST /_internal/revoke
@@ -650,19 +690,26 @@ 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 +717,12 @@ 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) {
+		exact, err := findExactCache(db, repo, prefix, version, true)
 		if err != nil {
-				return nil, fmt.Errorf("find cache: %w", err)
+			return nil, err
 		}
-			return cache, nil
+		if exact != nil {
+			return exact, nil
 		}
 		prefixPattern := "^" + regexp.QuoteMeta(prefix)
 		re, err := regexp.Compile(prefixPattern)
@@ -702,6 +745,34 @@ 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 +784,30 @@ 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

@@ -52,7 +52,7 @@ func TestHandler(t *testing.T) {
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(dir, "", 0, "", nil)
 	require.NoError(t, err)
-	handler.RegisterJob(testToken, testRepo)
+	handler.RegisterJob(testToken, JobCredential{Repo: testRepo})
 
 	base := fmt.Sprintf("%s%s", handler.ExternalURL(), apiPath)
 
@@ -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
 	{
@@ -880,7 +890,7 @@ func TestHandler_UnregisterRevokes(t *testing.T) {
 	require.NoError(t, err)
 	defer handler.Close()
 
-	unregister := handler.RegisterJob("tmp-token", testRepo)
+	unregister := handler.RegisterJob("tmp-token", JobCredential{Repo: testRepo})
 
 	base := handler.ExternalURL() + apiPath
 	req, err := http.NewRequest(http.MethodGet, base+"/cache?keys=x&version=y", nil)
@@ -910,8 +920,8 @@ func TestHandler_CrossRepoIsolation(t *testing.T) {
 	handler, err := StartHandler(dir, "", 0, "", nil)
 	require.NoError(t, err)
 	defer handler.Close()
-	handler.RegisterJob("token-a", "owner/repoA")
-	handler.RegisterJob("token-b", "owner/repoB")
+	handler.RegisterJob("token-a", JobCredential{Repo: "owner/repoA"})
+	handler.RegisterJob("token-b", JobCredential{Repo: "owner/repoB"})
 
 	base := handler.ExternalURL() + apiPath
 	key := "shared-key"
@@ -976,7 +986,7 @@ func TestHandler_ArtifactSignature(t *testing.T) {
 	handler, err := StartHandler(dir, "", 0, "", nil)
 	require.NoError(t, err)
 	defer handler.Close()
-	handler.RegisterJob(testToken, testRepo)
+	handler.RegisterJob(testToken, JobCredential{Repo: testRepo})
 
 	base := handler.ExternalURL() + apiPath
 
@@ -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
@@ -1049,7 +1059,7 @@ func TestHandler_ArtifactSignatureDownload(t *testing.T) {
 	handler, err := StartHandler(dir, "", 0, "", nil)
 	require.NoError(t, err)
 	defer handler.Close()
-	handler.RegisterJob(testToken, testRepo)
+	handler.RegisterJob(testToken, JobCredential{Repo: testRepo})
 
 	base := handler.ExternalURL() + apiPath
 	key := "download-key"
@@ -1090,8 +1100,8 @@ func TestHandler_RegisterJob_RefCounted(t *testing.T) {
 	require.NoError(t, err)
 	defer handler.Close()
 
-	first := handler.RegisterJob("shared", testRepo)
-	second := handler.RegisterJob("shared", testRepo)
+	first := handler.RegisterJob("shared", JobCredential{Repo: testRepo})
+	second := handler.RegisterJob("shared", JobCredential{Repo: testRepo})
 
 	base := handler.ExternalURL() + apiPath
 	probe := func() int {
@@ -1121,8 +1131,8 @@ func TestHandler_GC_PerRepoDedup(t *testing.T) {
 	handler, err := StartHandler(dir, "", 0, "", nil)
 	require.NoError(t, err)
 	defer handler.Close()
-	handler.RegisterJob("tok-a", "owner/repoA")
-	handler.RegisterJob("tok-b", "owner/repoB")
+	handler.RegisterJob("tok-a", JobCredential{Repo: "owner/repoA"})
+	handler.RegisterJob("tok-b", JobCredential{Repo: "owner/repoB"})
 
 	key := "shared-dedup-key"
 	version := "c19da02a2bd7e77277f1ac29ab45c09b7d46a4ee758284e26bb3045ad11d9d20"

act/artifactcache/handler_v2.go

@@ -0,0 +1,273 @@
+// 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.
+//
+// Responses carry the proto field names, which is what Gitea's own results API emits and the only
+// spelling the Go clients parse. The JavaScript toolkit accepts either.
+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,
+		"signed_upload_url": 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.
+		"entry_id": 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,
+		"signed_download_url": h.signedArtifactURL(cache.ID, time.Now().Add(artifactURLTTL)),
+		"matched_key":         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
+	}
+
+	// The Azure SDK client dereferences this without checking, so its absence panics the caller.
+	w.Header().Set("x-ms-request-id", strconv.FormatInt(time.Now().UnixNano(), 10))
+	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 +1,241 @@
+// 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()
+	if resp.StatusCode == http.StatusCreated {
+		// The Azure SDK client dereferences this header without checking, so a blob upload that
+		// omits it panics the caller rather than failing it.
+		require.NotEmpty(t, resp.Header.Get("x-ms-request-id"))
+	}
+	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, JobCredential{Repo: 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["signed_upload_url"].(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["entry_id"])
+
+	// 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["matched_key"])
+	downloadURL, _ := got["signed_download_url"].(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["signed_upload_url"].(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["signed_download_url"].(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["matched_key"])
+		})
+	}
+
+	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["signed_upload_url"])
+	})
+
+	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["signed_upload_url"].(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", JobCredential{Repo: "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/results.go

@@ -0,0 +1,59 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"crypto/tls"
+	"net/http"
+	"net/http/httputil"
+	"net/url"
+	"strings"
+)
+
+// The results service is one origin serving every github.actions.results.api.v1 service, and
+// Gitea implements only the artifact half of it. Forwarding that half from here makes this origin
+// the whole service, so ACTIONS_RESULTS_URL can point at it truthfully, which is what the clients
+// this runner cannot patch need, docker buildx among them.
+//
+// The instance to forward to travels with the job registration rather than with configuration, so
+// a cache server shared between runners serves each of their instances.
+const artifactServicePath = "/twirp/github.actions.results.api.v1.ArtifactService/"
+
+// forwardOrNotFound is the router's fallback: the artifact service of the instance the job
+// registered with, and the 404 the router would have written otherwise.
+func (h *Handler) forwardOrNotFound(w http.ResponseWriter, r *http.Request) {
+	cred, ok := h.lookupCredential(bearerToken(r))
+	if !ok || cred.Results == "" || !strings.HasPrefix(r.URL.Path, artifactServicePath) {
+		http.NotFound(w, r)
+		return
+	}
+	target, err := url.Parse(strings.TrimSuffix(cred.Results, "/"))
+	if err != nil {
+		h.logger.Errorf("artifact service forward to %q: %v", cred.Results, err)
+		w.WriteHeader(http.StatusBadGateway)
+		return
+	}
+	h.logger.Debugf("%s %s: forwarding to %s", r.Method, r.URL.Path, target)
+
+	proxy := &httputil.ReverseProxy{
+		Rewrite: func(r *httputil.ProxyRequest) {
+			r.SetURL(target)
+			// Gitea builds the URLs it hands back from this Host, and their scheme from the
+			// connection unless a forwarded header overrides it, so artifact bodies go to Gitea
+			// directly and never through here.
+			r.Out.Host = target.Host
+		},
+		ErrorHandler: func(w http.ResponseWriter, _ *http.Request, err error) {
+			h.logger.Warnf("artifact service forward to %s: %v", target, err)
+			w.WriteHeader(http.StatusBadGateway)
+		},
+	}
+	if cred.InsecureTLS {
+		proxy.Transport = insecureTransport
+	}
+	proxy.ServeHTTP(w, r)
+}
+
+// insecureTransport is shared, because a transport per request would pool no connections.
+var insecureTransport = &http.Transport{TLSClientConfig: &tls.Config{InsecureSkipVerify: true}} //nolint:gosec // the runner reaches its instance on the operator's say-so

act/artifactcache/results_test.go

@@ -0,0 +1,56 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// The artifact half is forwarded under the Host Gitea knows itself by, so the URLs it hands back
+// still point at Gitea, and nothing else is proxied.
+func TestFrontResultsService(t *testing.T) {
+	var gotHost, gotPath, gotProto string
+	gitea := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotHost, gotPath, gotProto = r.Host, r.URL.Path, r.Header.Get("X-Forwarded-Proto")
+		_, _ = io.WriteString(w, `{"ok":true}`)
+	}))
+	defer gitea.Close()
+
+	handler, err := StartHandler(t.TempDir(), "127.0.0.1", 0, "", nil)
+	require.NoError(t, err)
+	defer handler.Close()
+	const token = "forward-token"
+
+	client := &http.Client{Transport: &bearerTransport{token: token}}
+	post := func(path string) int {
+		req, err := http.NewRequestWithContext(t.Context(), http.MethodPost, handler.ExternalURL()+path, nil)
+		require.NoError(t, err)
+		resp, err := client.Do(req)
+		require.NoError(t, err)
+		resp.Body.Close()
+		return resp.StatusCode
+	}
+
+	assert.Equal(t, http.StatusNotFound, post(artifactServicePath+"CreateArtifact"),
+		"an unregistered token is forwarded nowhere")
+
+	defer handler.RegisterJob(token, JobCredential{Repo: "owner/repo", Results: gitea.URL})()
+
+	assert.Equal(t, http.StatusOK, post(artifactServicePath+"CreateArtifact"))
+	assert.Equal(t, strings.TrimPrefix(gitea.URL, "http://"), gotHost, "Gitea must see the host it mints its URLs from")
+	assert.Empty(t, gotProto, "a forwarded scheme would make an https Gitea mint http URLs")
+	assert.Equal(t, artifactServicePath+"CreateArtifact", gotPath)
+
+	gotPath = ""
+	assert.Equal(t, http.StatusNotFound, post("/twirp/github.actions.results.api.v1.OtherService/Do"))
+	assert.Equal(t, http.StatusNotFound, post("/api/v1/repos/owner/repo"))
+	assert.Empty(t, gotPath, "only the artifact service is forwarded")
+}

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 {

act/common/git/git_test.go

@@ -12,7 +12,6 @@ import (
 	"path/filepath"
 	"strconv"
 	"strings"
-	"sync"
 	"syscall"
 	"testing"
 	"time"
@@ -610,12 +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_cli.go

@@ -4,8 +4,10 @@
 
 //go:build !(WITHOUT_DOCKER || !(linux || darwin || windows || netbsd))
 
-// This file is exact copy of https://github.com/docker/cli/blob/9a471180cb7d39c236d090399a9d362c3f5a8ebd/cli/command/container/opts.go
-// appended with license information.
+// This file is exact copy of https://github.com/docker/cli/blob/dfc4efb1e2ab8c06d70d2a1366ad448d2f917e90/cli/command/container/opts.go with:
+// * appended with license information
+// * regexp and loader.ParseVolume in place of the import-restricted internal/lazyregexp and internal/volumespec
+// * invalidParameter from the package's errors.go, and convertPortSet/convertPortMap for the callers in docker_run.go
 //
 // docker/cli is licensed under the Apache License, Version 2.0.
 // See DOCKER_LICENSE for the full license text.
@@ -30,6 +32,7 @@ import (
 	"strings"
 	"time"
 
+	cerrdefs "github.com/containerd/errdefs"
 	"github.com/docker/cli/cli/compose/loader"
 	"github.com/docker/cli/opts"
 	"github.com/docker/go-connections/nat"
@@ -380,7 +383,7 @@ func parse(flags *pflag.FlagSet, copts *containerOptions, serverOS string) (*con
 	var binds []string
 	volumes := copts.volumes.GetMap()
 	// add any bind targets to the list of container volumes
-	for bind := range copts.volumes.GetMap() {
+	for bind := range volumes {
 		parsed, err := loader.ParseVolume(bind)
 		if err != nil {
 			return nil, err
@@ -515,13 +518,13 @@ func parse(flags *pflag.FlagSet, copts *containerOptions, serverOS string) (*con
 	// collect all the environment variables for the container
 	envVariables, err := opts.ReadKVEnvStrings(copts.envFile.GetSlice(), copts.env.GetSlice())
 	if err != nil {
-		return nil, err
+		return nil, fmt.Errorf("--env-file: %w", err)
 	}
 
 	// collect all the labels for the container
 	labels, err := opts.ReadKVStrings(copts.labelsFile.GetSlice(), copts.labels.GetSlice())
 	if err != nil {
-		return nil, err
+		return nil, fmt.Errorf("--label-file: %w", err)
 	}
 
 	pidMode := container.PidMode(copts.pidMode)
@@ -1164,16 +1167,17 @@ func toNetipAddrSlice(ips []string) []netip.Addr {
 }
 
 // invalidParameter wraps an error to indicate it was caused by invalid input.
-// This is a local replacement for docker/docker/errdefs.InvalidParameter.
-type invalidParameterError struct{ error }
+// This is a copy of docker/cli's cli/command/container/errors.go, which is not importable.
+type invalidParameterErr struct{ error }
 
-func (e invalidParameterError) InvalidParameter() {}
+func (invalidParameterErr) InvalidParameter() {}
+func (e invalidParameterErr) Unwrap() error   { return e.error }
 
 func invalidParameter(err error) error {
-	if err == nil {
-		return nil
+	if err == nil || cerrdefs.IsInvalidArgument(err) {
+		return err
 	}
-	return invalidParameterError{err}
+	return invalidParameterErr{err}
 }
 
 func convertPortSet(ports nat.PortSet) (network.PortSet, error) {

act/container/docker_create_flags_test.go

@@ -51,7 +51,8 @@ func TestCreateFlagsValidate(t *testing.T) {
 
 func TestNewContainerAppliesCreateFlags(t *testing.T) {
 	input := &NewContainerInput{Platform: "linux/amd64", Options: "--platform linux/arm64 --pull never"}
-	cr := NewContainer(input).(*containerReference)
+	cr, ok := NewContainer(input).(*containerReference)
+	require.True(t, ok)
 	assert.Equal(t, "linux/arm64", input.Platform)
 	assert.Equal(t, pullPolicyNever, cr.pullPolicy)
 

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{
+		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,
-		})
-		if err != nil {
-			return err
-		}
-
+				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)
+				// 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
 				}
-				} else {
-					common.Logger(ctx).Debugf("Refusing to remove network %v because it still has active endpoints", name)
+				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

@@ -30,22 +30,18 @@ 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 !pull {
+			if imageExists {
 				return nil
 			}
+		}
 
 		imageRef := cleanImage(ctx, input.Image)
 		logger.Debugf("pulling image '%v' (%s)", imageRef, input.Platform)
@@ -61,23 +57,33 @@ func NewDockerPullExecutor(input NewDockerPullExecutorInput) common.Executor {
 			return err
 		}
 
-		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 {
-			if imagePullOptions.RegistryAuth != "" && strings.Contains(err.Error(), "unauthorized") {
+				return err
+			}
+			return streamErr
+		}
+
+		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 = ""
-				reader, err = cli.ImagePull(ctx, imageRef, imagePullOptions)
-
-				_ = logDockerResponse(logger, reader, err != nil)
-			}
-			if err != nil {
-				return fmt.Errorf("failed to pull image '%s' (%s): %w", imageRef, input.Platform, err)
-			}
+			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)
+	}
 }
 
 func getImagePullOptions(ctx context.Context, input NewDockerPullExecutorInput) (client.ImagePullOptions, error) {
@@ -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"
@@ -26,7 +27,6 @@ import (
 	"gitea.com/gitea/runner/act/filecollector"
 
 	"dario.cat/mergo"
-	"github.com/Masterminds/semver"
 	cerrdefs "github.com/containerd/errdefs"
 	"github.com/docker/cli/cli/compose/loader"
 	"github.com/docker/cli/cli/connhelper"
@@ -41,7 +41,9 @@ import (
 	"github.com/moby/moby/api/types/network"
 	"github.com/moby/moby/api/types/system"
 	"github.com/moby/moby/client"
+	"github.com/moby/moby/client/pkg/versions"
 	specs "github.com/opencontainers/image-spec/specs-go/v1"
+	"github.com/sirupsen/logrus"
 )
 
 // drainGracePeriod bounds how long we wait for an output-copy goroutine to
@@ -90,19 +92,11 @@ func (cr *containerReference) connectToNetwork(name string, aliases []string) co
 // supportsContainerImagePlatform returns true if the underlying Docker server
 // API version is 1.41 and beyond
 func supportsContainerImagePlatform(ctx context.Context, cli client.APIClient) bool {
-	logger := common.Logger(ctx)
 	ver, err := cli.ServerVersion(ctx, client.ServerVersionOptions{})
 	if err != nil {
-		logger.Panicf("Failed to get Docker API Version: %s", err)
-		return false
+		common.Logger(ctx).Panicf("Failed to get Docker API Version: %s", err)
 	}
-	sv, err := semver.NewVersion(ver.APIVersion)
-	if err != nil {
-		logger.Panicf("Failed to unmarshal Docker Version: %s", err)
-		return false
-	}
-	constraint, _ := semver.NewConstraint(">= 1.41")
-	return constraint.Check(sv)
+	return versions.GreaterThanOrEqualTo(ver.APIVersion, "1.41")
 }
 
 func (cr *containerReference) Create(capAdd, capDrop []string) common.Executor {
@@ -386,32 +380,71 @@ func (cr *containerReference) find() common.Executor {
 	}
 }
 
-// isContainerGone reports whether a failed remove still left the container gone (NotFound or Conflict).
-func isContainerGone(err error) bool {
-	return cerrdefs.IsNotFound(err) || cerrdefs.IsConflict(err)
-}
-
 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 && !isContainerGone(err) {
-			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
@@ -461,6 +494,16 @@ 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)
@@ -471,8 +514,7 @@ 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
@@ -538,7 +580,7 @@ func (cr *containerReference) create(capAdd, capDrop []string) common.Executor {
 		}
 
 		var platSpecs *specs.Platform
-		if supportsContainerImagePlatform(ctx, cr.cli) && cr.input.Platform != "" {
+		if cr.input.Platform != "" && supportsContainerImagePlatform(ctx, cr.cli) {
 			platSpecs, err = parsePlatform(cr.input.Platform)
 			if err != nil {
 				return err
@@ -815,23 +857,59 @@ 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)
 	}
@@ -856,6 +934,10 @@ 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
@@ -1063,6 +1145,77 @@ 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) {
@@ -1108,6 +1261,34 @@ 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)
@@ -121,6 +128,26 @@ func (m *mockDockerClient) ContainerRemove(ctx context.Context, id string, 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
 }
@@ -314,15 +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",
@@ -332,46 +381,33 @@ 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) {
+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()
 
-	merr := errors.New("Failure")
-
 			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 == "/" && opts.Content != nil
-	})).Return(mobyclient.CopyToContainerResult{}, merr)
-	cr := &containerReference{
-		id:  "123",
-		cli: client,
-		input: &NewContainerInput{
-			Image: "image",
-		},
-	}
-
-	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()
-
-	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)
+				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{}, merr)
+			})).Return(mobyclient.CopyToContainerResult{}, testCase.copyErr).Maybe()
 			cr := &containerReference{
 				id:  "123",
 				cli: client,
@@ -380,39 +416,50 @@ func TestDockerCopyTarStreamErrorInMkdir(t *testing.T) {
 				},
 			}
 
-	err := cr.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{})
-	assert.ErrorIs(t, err, merr) //nolint:testifylint // pre-existing issue from nektos/act
+			require.ErrorIs(t, cr.CopyTarStream(ctx, "/var/run/act", &bytes.Buffer{}), merr)
 
 			client.AssertExpectations(t)
+		})
+	}
 }
 
 // A remove that raced the daemon's AutoRemove teardown is not a failure and must not
 // be logged as one.
 func TestRemoveIgnoresAutoRemoveRace(t *testing.T) {
 	removeOpts := mobyclient.ContainerRemoveOptions{RemoveVolumes: true, Force: true}
+	killOpts := mobyclient.ContainerKillOptions{Signal: "SIGKILL"}
 	for _, tc := range []struct {
 		name        string
 		err         error
-		wantLogs bool
+		wantWait    bool
+		wantFailure bool
 	}{
-		{name: "removal in progress", err: cerrdefs.ErrConflict.WithMessage("removal of container abc is already in progress")},
+		{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"), wantLogs: true},
+		{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}
 
 			require.NoError(t, cr.remove()(ctx))
-			assert.Empty(t, cr.id)
-
-			if tc.wantLogs {
+			// 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)
@@ -420,6 +467,20 @@ func TestRemoveIgnoresAutoRemoveRace(t *testing.T) {
 	}
 }
 
+// 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)
+}
+
 // find() must drop a stale cached id so later Copy/Exec don't hit the
 // daemon with a torn-down container.
 func TestFindRevalidatesStaleID(t *testing.T) {
@@ -541,9 +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)
 }
@@ -624,6 +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)
@@ -660,3 +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

@@ -8,12 +8,13 @@ package container
 
 import (
 	"context"
+	"errors"
 	"runtime"
+	"time"
 
 	"gitea.com/gitea/runner/act/common"
 
 	"github.com/moby/moby/api/types/system"
-	"github.com/pkg/errors"
 )
 
 // ImageExistsLocally returns a boolean indicating if an image with the
@@ -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

@@ -330,6 +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

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

@@ -28,8 +28,8 @@ func (impl *interperterImpl) contains(search, item reflect.Value) (bool, error)
 	switch search.Kind() {
 	case reflect.String, reflect.Int, reflect.Float64, reflect.Bool, reflect.Invalid:
 		return strings.Contains(
-			strings.ToLower(impl.coerceToString(search).String()),
-			strings.ToLower(impl.coerceToString(item).String()),
+			strings.ToLower(CoerceToString(search)),
+			strings.ToLower(CoerceToString(item)),
 		), nil
 
 	case reflect.Slice:
@@ -51,15 +51,15 @@ func (impl *interperterImpl) contains(search, item reflect.Value) (bool, error)
 
 func (impl *interperterImpl) startsWith(searchString, searchValue reflect.Value) (bool, error) { //nolint:unparam // pre-existing issue from nektos/act
 	return strings.HasPrefix(
-		strings.ToLower(impl.coerceToString(searchString).String()),
-		strings.ToLower(impl.coerceToString(searchValue).String()),
+		strings.ToLower(CoerceToString(searchString)),
+		strings.ToLower(CoerceToString(searchValue)),
 	), nil
 }
 
 func (impl *interperterImpl) endsWith(searchString, searchValue reflect.Value) (bool, error) { //nolint:unparam // pre-existing issue from nektos/act
 	return strings.HasSuffix(
-		strings.ToLower(impl.coerceToString(searchString).String()),
-		strings.ToLower(impl.coerceToString(searchValue).String()),
+		strings.ToLower(CoerceToString(searchString)),
+		strings.ToLower(CoerceToString(searchValue)),
 	), nil
 }
 
@@ -70,7 +70,7 @@ const (
 )
 
 func (impl *interperterImpl) format(str reflect.Value, replaceValue ...reflect.Value) (string, error) {
-	input := impl.coerceToString(str).String()
+	input := CoerceToString(str)
 	var output strings.Builder
 	replacementIndex := ""
 
@@ -108,7 +108,7 @@ func (impl *interperterImpl) format(str reflect.Value, replaceValue ...reflect.V
 					return "", fmt.Errorf("The following format string references more arguments than were supplied: '%s'", input)
 				}
 
-				output.WriteString(impl.coerceToString(replaceValue[index]).String())
+				output.WriteString(CoerceToString(replaceValue[index]))
 
 				state = passThrough
 
@@ -124,7 +124,7 @@ func (impl *interperterImpl) format(str reflect.Value, replaceValue ...reflect.V
 				state = passThrough
 
 			default:
-				panic("Invalid format parser state")
+				return "", fmt.Errorf("Closing bracket without opening one. The following format string is invalid: '%s'", input)
 			}
 		}
 	}
@@ -143,17 +143,17 @@ func (impl *interperterImpl) format(str reflect.Value, replaceValue ...reflect.V
 }
 
 func (impl *interperterImpl) join(array, sep reflect.Value) (string, error) { //nolint:unparam // pre-existing issue from nektos/act
-	separator := impl.coerceToString(sep).String()
+	separator := CoerceToString(sep)
 	switch array.Kind() {
 	case reflect.Slice:
 		var items []string
 		for i := 0; i < array.Len(); i++ {
-			items = append(items, impl.coerceToString(array.Index(i).Elem()).String())
+			items = append(items, CoerceToString(array.Index(i)))
 		}
 
 		return strings.Join(items, separator), nil
 	default:
-		return strings.Join([]string{impl.coerceToString(array).String()}, separator), nil
+		return strings.Join([]string{CoerceToString(array)}, separator), nil
 	}
 }
 

act/exprparser/functions_test.go

@@ -121,6 +121,7 @@ func TestFunctionJoin(t *testing.T) {
 		{"join(fromJSON('[\"a\", \"b\", null]'), null)", "ab", "join-number"},
 		{"join(fromJSON('[\"a\", \"b\"]'))", "a,b", "join-number"},
 		{"join(fromJSON('[\"a\", \"b\", null]'), 1)", "a1b1", "join-number"},
+		{"join(fromJSON('[1, true, null]'), '-')", "1-true-", "join-mixed-types"},
 	}
 
 	env := &EvaluationEnvironment{}
@@ -230,8 +231,10 @@ func TestFunctionFormat(t *testing.T) {
 		{`format('Hello "{0}" {1} {2} {3} {4}', null, true, -3.14, NaN, Infinity)`, `Hello "" true -3.14 NaN Infinity`, nil, "format-with-primitives"},
 		{`format('Hello "{0}" {1} {2}', fromJSON('[0, true, "abc"]'), fromJSON('[{"a":1}]'), fromJSON('{"a":{"b":1}}'))`, `Hello "Array" Array Object`, nil, "format-with-complex-types"},
 		{"format(true)", "true", nil, "format-with-primitive-args"},
+		{"format('{0}', github)", "Object", nil, "format-with-context"},
 		{"format('echo Hello {0} ${{Test}}', github.undefined_property)", "echo Hello  ${Test}", nil, "format-with-undefined-value"},
 		{"format('{0}}', '{1}', 'World')", nil, "Closing bracket without opening one. The following format string is invalid: '{0}}'", "format-invalid-format-string"},
+		{"format('a}b')", nil, "Closing bracket without opening one. The following format string is invalid: 'a}b'", "format-unmatched-closing-brace"},
 		{"format('{0', '{1}', 'World')", nil, "Unclosed brackets. The following format string is invalid: '{0'", "format-invalid-format-string"},
 		{"format('{2}', '{1}', 'World')", "", "The following format string references more arguments than were supplied: '{2}'", "format-invalid-replacement-reference"},
 		{"format('{2147483648}')", "", "The following format string is invalid: '{2147483648}'", "format-invalid-replacement-reference"},

act/exprparser/interpreter.go

@@ -10,6 +10,7 @@ import (
 	"fmt"
 	"math"
 	"reflect"
+	"strconv"
 	"strings"
 
 	"gitea.com/gitea/runner/act/model"
@@ -429,41 +430,54 @@ func (impl *interperterImpl) coerceToNumber(value reflect.Value) reflect.Value {
 	return reflect.ValueOf(math.NaN())
 }
 
-func (impl *interperterImpl) coerceToString(value reflect.Value) reflect.Value {
+// CoerceToString converts an evaluated expression value to a string the way GitHub does,
+// see https://docs.github.com/en/actions/reference/workflows-and-actions/expressions#operators
+// An already reflected value is accepted as-is, since Interface() would panic on an invalid one.
+func CoerceToString(v any) string {
+	value, ok := v.(reflect.Value)
+	if !ok {
+		value = reflect.ValueOf(v)
+	}
+
 	switch value.Kind() {
 	case reflect.Invalid:
-		return reflect.ValueOf("")
+		return ""
 
 	case reflect.Bool:
-		switch value.Bool() {
-		case true:
-			return reflect.ValueOf("true")
-		case false:
-			return reflect.ValueOf("false")
-		}
+		return strconv.FormatBool(value.Bool())
 
 	case reflect.String:
-		return value
+		return value.String()
 
-	case reflect.Int:
-		return reflect.ValueOf(fmt.Sprint(value))
+	case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
+		return strconv.FormatInt(value.Int(), 10)
 
-	case reflect.Float64:
+	case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64:
+		return strconv.FormatUint(value.Uint(), 10)
+
+	case reflect.Float32, reflect.Float64:
 		if math.IsInf(value.Float(), 1) {
-			return reflect.ValueOf("Infinity")
+			return "Infinity"
 		} else if math.IsInf(value.Float(), -1) {
-			return reflect.ValueOf("-Infinity")
+			return "-Infinity"
 		}
-		return reflect.ValueOf(fmt.Sprintf("%.15G", value.Float()))
+		return fmt.Sprintf("%.15G", value.Float())
 
-	case reflect.Slice:
-		return reflect.ValueOf("Array")
+	case reflect.Slice, reflect.Array:
+		return "Array"
 
-	case reflect.Map:
-		return reflect.ValueOf("Object")
+	// contexts such as `github` are pointers to structs, so they stringify as objects too
+	case reflect.Map, reflect.Struct:
+		return "Object"
+
+	case reflect.Interface, reflect.Pointer:
+		if value.IsNil() {
+			return ""
+		}
+		return CoerceToString(value.Elem())
 	}
 
-	return value
+	return fmt.Sprintf("%v", value)
 }
 
 func (impl *interperterImpl) compareString(left, right string, kind actionlint.CompareOpNodeKind) (bool, error) {

act/exprparser/interpreter_test.go

@@ -6,11 +6,13 @@ package exprparser
 
 import (
 	"math"
+	"reflect"
 	"testing"
 
 	"gitea.com/gitea/runner/act/model"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func TestLiterals(t *testing.T) {
@@ -523,7 +525,9 @@ 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)
 			}
@@ -630,3 +634,51 @@ func TestContexts(t *testing.T) {
 		})
 	}
 }
+
+func TestCoerceToString(t *testing.T) {
+	type object struct{ Name string }
+	obj := object{Name: "x"}
+	var nilPointer *object
+	var nilMap map[string]any
+	var nilSlice []any
+
+	table := []struct {
+		input    any
+		expected string
+		name     string
+	}{
+		{nil, "", "null"},
+		{true, "true", "true"},
+		{false, "false", "false"},
+		{"foo", "foo", "string"},
+		{"", "", "empty-string"},
+		{123, "123", "int"},
+		{int64(-9), "-9", "int64"},
+		{uint8(7), "7", "uint8"},
+		{1.0, "1", "float-integral"},
+		{-9.7, "-9.7", "float"},
+		{2.99e-2, "0.0299", "float-exponential"},
+		{1e21, "1E+21", "float-large"},
+		{float32(1.5), "1.5", "float32"},
+		{math.NaN(), "NaN", "nan"},
+		{math.Inf(1), "Infinity", "positive-infinity"},
+		{math.Inf(-1), "-Infinity", "negative-infinity"},
+		{[]any{1, 2}, "Array", "slice"},
+		{nilSlice, "Array", "nil-slice"},
+		{[2]int{1, 2}, "Array", "fixed-size-array"},
+		{map[string]any{"a": 1}, "Object", "map"},
+		{nilMap, "Object", "nil-map"},
+		{obj, "Object", "struct"},
+		{&obj, "Object", "pointer-to-struct"},
+		{nilPointer, "", "nil-pointer"},
+		{&model.GithubContext{Action: "push"}, "Object", "github-context"},
+		{reflect.ValueOf(42), "42", "reflected-value"},
+		{reflect.Value{}, "", "invalid-reflected-value"},
+	}
+
+	for _, tt := range table {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.expected, CoerceToString(tt.input))
+		})
+	}
+}

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" {
+			entry, ok := v.(map[string]any)
+			if !ok {
 				continue
 			}
-				allSchedules = append(allSchedules, cron.(string))
+			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
 		}
@@ -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.
@@ -472,10 +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()
@@ -589,8 +595,8 @@ func actionStagePaths(step actionStep) (actionDir, actionPath, actionName, conta
 	rc := step.getRunContext()
 	stepModel := step.getStepModel()
 
-	if _, ok := step.(*stepActionRemote); ok {
-		actionDir = fmt.Sprintf("%s/%s", rc.ActionCacheDir(), stepModel.UsesHash())
+	if sar, ok := step.(*stepActionRemote); ok {
+		actionDir = sar.actionDir()
 		actionPath = newRemoteAction(stepModel.Uses).Path
 	} else {
 		actionDir = filepath.Join(rc.Config.Workdir, stepModel.Uses)

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

@@ -226,11 +226,16 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 		}
 	}
 
-	// The setup section of the job log: download the actions, run the pre steps, then name the job.
+	// 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

act/runner/job_executor_test.go

@@ -33,6 +33,7 @@ import (
 )
 
 func TestJobExecutor(t *testing.T) {
+	t.Parallel()
 	// Dryrun only checks syntax/planning; all cases resolve locally, so this runs offline.
 	tables := []TestJobFileInfo{
 		{workdir, "uses-and-run-in-one-step", "push", "Invalid run/uses syntax for job:test step:Test", platforms, secrets},
@@ -46,6 +47,7 @@ func TestJobExecutor(t *testing.T) {
 	ctx := common.WithDryrun(context.Background(), true)
 	for _, table := range tables {
 		t.Run(table.workflowPath, func(t *testing.T) {
+			t.Parallel()
 			table.runTest(ctx, t, &Config{})
 		})
 	}

act/runner/job_hooks.go

@@ -0,0 +1,115 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package runner
+
+import (
+	"cmp"
+	"context"
+	"fmt"
+	"maps"
+	"path"
+	"strings"
+
+	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/container"
+)
+
+// GitHub's job-hook variables, read as a fallback when the settings are unset.
+const (
+	jobStartedHookEnv   = "ACTIONS_RUNNER_HOOK_JOB_STARTED"
+	jobCompletedHookEnv = "ACTIONS_RUNNER_HOOK_JOB_COMPLETED"
+)
+
+// Kept apart from the per-step file-command files, which are truncated on every step.
+const (
+	hookEnvFileCommand  = "workflow/hook-envs.txt"
+	hookPathFileCommand = "workflow/hook-path.txt"
+)
+
+func (rc *RunContext) runJobStartedHook(ctx context.Context) error {
+	return rc.runJobHook(ctx, cmp.Or(rc.Config.JobStartedHook, rc.Config.Env[jobStartedHookEnv]), "job started")
+}
+
+func (rc *RunContext) runJobCompletedHook(ctx context.Context) error {
+	return rc.runJobHook(ctx, cmp.Or(rc.Config.JobCompletedHook, rc.Config.Env[jobCompletedHookEnv]), "job completed")
+}
+
+// runJobHook runs one hook in the job environment. Either hook failing fails the job, as
+// on GitHub, where the operator is responsible for the hook's own resilience.
+func (rc *RunContext) runJobHook(ctx context.Context, hookPath, name string) error {
+	if hookPath == "" {
+		return nil
+	}
+
+	cmd, shell := hookCommand(hookPath)
+	rawLogger := common.Logger(ctx).WithField(rawOutputField, true)
+	defer rawLogger.Infof("::endgroup::")
+	rawLogger.Infof("::group::Run '%s'", escapeCommandData(hookPath))
+	rawLogger.Infof("A %s hook has been configured by the runner administrator", name)
+	if shell != "" {
+		rawLogger.Infof("shell: %s", shell)
+	}
+
+	env := maps.Clone(rc.GetEnv())
+	if jobContainer := rc.Run.Job().Container(); jobContainer != nil {
+		maps.Copy(env, jobContainer.Env)
+	}
+	rc.withGithubEnv(ctx, rc.getGithubContext(ctx), env)
+	rc.ApplyExtraPath(ctx, &env)
+
+	err := rc.setupHookFileCommands(ctx, env)
+	if err == nil {
+		err = rc.JobContainer.Exec(cmd, env, "", "")(ctx)
+	}
+	// Processed even on failure, so a hook that exports what it managed to set up before
+	// failing still hands it to the job.
+	err = cmp.Or(err, rc.processHookFileCommands(ctx))
+	if err == nil {
+		return nil
+	}
+
+	err = fmt.Errorf("the %s hook %q failed: %w", name, hookPath, err)
+	// Flip the job status the way a failing pre step does, so success()-default main steps
+	// skip and the task is reported failed.
+	reportStepError(ctx, rc, err)
+	return err
+}
+
+// setupHookFileCommands points the hook at its GITHUB_ENV and GITHUB_PATH files, so it can
+// export to the job's steps, and truncates them so the second hook does not re-read what
+// the first one wrote.
+func (rc *RunContext) setupHookFileCommands(ctx context.Context, env map[string]string) error {
+	actPath := rc.JobContainer.GetActPath()
+	env["GITHUB_ENV"] = path.Join(actPath, hookEnvFileCommand)
+	env["GITHUB_PATH"] = path.Join(actPath, hookPathFileCommand)
+	env["GITEA_ENV"] = env["GITHUB_ENV"]
+	env["GITEA_PATH"] = env["GITHUB_PATH"]
+
+	return rc.JobContainer.Copy(actPath,
+		&container.FileEntry{Name: hookEnvFileCommand, Mode: 0o666},
+		&container.FileEntry{Name: hookPathFileCommand, Mode: 0o666},
+	)(ctx)
+}
+
+func (rc *RunContext) processHookFileCommands(ctx context.Context) error {
+	if err := processRunnerEnvFileCommand(ctx, hookEnvFileCommand, rc, rc.setEnv); err != nil {
+		return err
+	}
+	return rc.UpdateExtraPath(ctx, path.Join(rc.JobContainer.GetActPath(), hookPathFileCommand))
+}
+
+// hookCommand mirrors actions/runner, which deliberately does not apply the shell flags it
+// gives `run:` steps — a hook sets its own. See docs/adrs/1751-runner-job-hooks.md there.
+// The second return value is how the invocation is shown in the log, empty when the file is
+// executed directly.
+func hookCommand(hookPath string) (cmd []string, shell string) {
+	switch strings.ToLower(path.Ext(hookPath)) {
+	case ".sh":
+		return []string{"bash", "-e", hookPath}, "bash -e {0}"
+	case ".ps1":
+		return []string{"pwsh", "-command", ". '" + hookPath + "'"}, `pwsh -command ". '{0}'"`
+	default:
+		return []string{hookPath}, ""
+	}
+}

act/runner/job_hooks_test.go

@@ -0,0 +1,162 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package runner
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"io"
+	"maps"
+	"testing"
+
+	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/model"
+
+	"github.com/sirupsen/logrus/hooks/test"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// hookContainer records the command a hook was run with and answers with what the hook
+// wrote to its GITHUB_ENV and GITHUB_PATH files.
+type hookContainer struct {
+	fakeContainer
+	cmd     []string
+	env     map[string]string
+	err     error
+	envFile map[string]string
+	pathTar []byte
+}
+
+func (c *hookContainer) ToContainerPath(path string) string { return path }
+func (c *hookContainer) IsEnvironmentCaseInsensitive() bool { return false }
+
+func (c *hookContainer) GetRunnerContext(context.Context) map[string]any {
+	return map[string]any{"os": "Linux"}
+}
+
+func (c *hookContainer) Exec(command []string, env map[string]string, _, _ string) common.Executor {
+	return func(context.Context) error {
+		c.cmd, c.env = command, env
+		return c.err
+	}
+}
+
+func (c *hookContainer) UpdateFromEnv(_ string, env *map[string]string) common.Executor {
+	return func(context.Context) error {
+		maps.Copy(*env, c.envFile)
+		return nil
+	}
+}
+
+func (c *hookContainer) GetContainerArchive(context.Context, string) (io.ReadCloser, error) {
+	return io.NopCloser(bytes.NewReader(c.pathTar)), nil
+}
+
+// newHookRunContext returns a RunContext and the context to run a hook with, whose logger is
+// silenced so the hook's job-log output does not reach the test output.
+func newHookRunContext(jobContainer *hookContainer, config *Config) (*RunContext, context.Context) {
+	// Env is left nil so that it is built from the config, as it is for a real job.
+	rc := &RunContext{
+		Config:       config,
+		Run:          &model.Run{JobID: "job", Workflow: &model.Workflow{Jobs: map[string]*model.Job{"job": {}}}},
+		JobContainer: jobContainer,
+	}
+	logger, _ := test.NewNullLogger()
+	ctx := common.WithJobErrorContainer(common.WithLogger(context.Background(), logger.WithField("test", true)))
+	rc.ExprEval = rc.NewExpressionEvaluator(ctx)
+	return rc, ctx
+}
+
+func TestRunJobHook(t *testing.T) {
+	t.Run("runs the hook with the job environment", func(t *testing.T) {
+		jobContainer := &hookContainer{}
+		rc, ctx := newHookRunContext(jobContainer, &Config{
+			JobStartedHook: "/hooks/started.sh",
+			Env:            map[string]string{"A_VAR": "value", jobStartedHookEnv: "/from/env.sh"},
+		})
+
+		require.NoError(t, rc.runJobStartedHook(ctx))
+
+		// The setting wins over the environment variable.
+		assert.Equal(t, []string{"bash", "-e", "/hooks/started.sh"}, jobContainer.cmd)
+		assert.Equal(t, "value", jobContainer.env["A_VAR"])
+		// The github environment is there too, so a hook can tell which job it runs for.
+		assert.Equal(t, "job", jobContainer.env["GITHUB_JOB"])
+		assert.Equal(t, "/var/run/act/workflow/hook-envs.txt", jobContainer.env["GITHUB_ENV"])
+		assert.Equal(t, "/var/run/act/workflow/hook-path.txt", jobContainer.env["GITHUB_PATH"])
+	})
+
+	// Each hook reads its own variable, so a swapped constant cannot pass.
+	t.Run("falls back to the GitHub environment variables", func(t *testing.T) {
+		for name, hook := range map[string]struct {
+			env string
+			run func(*RunContext, context.Context) error
+		}{
+			"started":   {jobStartedHookEnv, (*RunContext).runJobStartedHook},
+			"completed": {jobCompletedHookEnv, (*RunContext).runJobCompletedHook},
+		} {
+			t.Run(name, func(t *testing.T) {
+				jobContainer := &hookContainer{}
+				rc, ctx := newHookRunContext(jobContainer, &Config{Env: map[string]string{hook.env: "/from/env.sh"}})
+
+				require.NoError(t, hook.run(rc, ctx))
+				assert.Equal(t, []string{"bash", "-e", "/from/env.sh"}, jobContainer.cmd)
+			})
+		}
+	})
+
+	t.Run("exports what the hook wrote to GITHUB_ENV and GITHUB_PATH", func(t *testing.T) {
+		jobContainer := &hookContainer{
+			envFile: map[string]string{"FROM_HOOK": "1"},
+			pathTar: tarArchive(t, tarEntry{name: "hook-path.txt", body: "/opt/tool/bin\n"}),
+		}
+		rc, ctx := newHookRunContext(jobContainer, &Config{JobStartedHook: "/hooks/started.sh"})
+
+		require.NoError(t, rc.runJobStartedHook(ctx))
+
+		assert.Equal(t, "1", rc.Env["FROM_HOOK"])
+		assert.Equal(t, []string{"/opt/tool/bin"}, rc.ExtraPath)
+	})
+
+	t.Run("a failing hook fails the job", func(t *testing.T) {
+		rc, ctx := newHookRunContext(&hookContainer{err: errors.New("boom")}, &Config{JobStartedHook: "/hooks/started.sh"})
+
+		err := rc.runJobStartedHook(ctx)
+
+		require.ErrorContains(t, err, `the job started hook "/hooks/started.sh" failed`)
+		require.ErrorContains(t, err, "boom")
+		// The failure has to flip the job status, or success()-default steps would still
+		// run and the task would be reported successful despite the missing setup.
+		assert.Equal(t, "failure", rc.getJobContext().Status)
+		require.ErrorContains(t, common.JobError(ctx), "boom")
+	})
+
+	t.Run("is a no-op without a hook", func(t *testing.T) {
+		jobContainer := &hookContainer{}
+		rc, ctx := newHookRunContext(jobContainer, &Config{})
+
+		require.NoError(t, rc.runJobStartedHook(ctx))
+		require.NoError(t, rc.runJobCompletedHook(ctx))
+		assert.Nil(t, jobContainer.cmd)
+	})
+}
+
+// actions/runner deliberately runs a hook without the flags it gives `run:` steps, and an
+// executable without a known extension speaks for itself through its shebang.
+func TestHookCommand(t *testing.T) {
+	for hookPath, want := range map[string]struct {
+		cmd   []string
+		shell string
+	}{
+		"/hooks/started.sh":  {[]string{"bash", "-e", "/hooks/started.sh"}, "bash -e {0}"},
+		"/hooks/started.PS1": {[]string{"pwsh", "-command", ". '/hooks/started.PS1'"}, `pwsh -command ". '{0}'"`},
+		"/hooks/started":     {[]string{"/hooks/started"}, ""},
+	} {
+		cmd, shell := hookCommand(hookPath)
+		assert.Equal(t, want.cmd, cmd, hookPath)
+		assert.Equal(t, want.shell, shell, hookPath)
+	}
+}

act/runner/logger.go

@@ -7,8 +7,11 @@ package runner
 import (
 	"bytes"
 	"context"
+	"encoding/base64"
+	"encoding/json"
 	"fmt"
 	"io"
+	"net/url"
 	"os"
 	"slices"
 	"strings"
@@ -167,6 +170,76 @@ func withStepLogger(ctx context.Context, stepNumber int, stepID, stepName, stage
 
 type entryProcessor func(entry *logrus.Entry) *logrus.Entry
 
+// secretValueEncoders are the shapes a secret takes on its way into a log: a base64
+// payload, a JSON string, or a URL component. An action that serializes a secret leaks
+// it in one of these forms, which a mask of the verbatim value alone does not catch, so
+// every form is masked as well. This mirrors the value encoders of GitHub's runner.
+var secretValueEncoders = []func(string) string{
+	func(v string) string { return base64.StdEncoding.EncodeToString([]byte(v)) },
+	base64ShiftEncoder(1),
+	base64ShiftEncoder(2),
+	jsonStringEscape,
+	jsonStringEscapeNoHTML,
+	url.QueryEscape,
+	url.PathEscape,
+}
+
+// minShiftedBase64Len is the shortest shifted base64 fragment worth masking. A shorter
+// one carries too few bytes of the secret to identify it and would mask unrelated output.
+const minShiftedBase64Len = 8
+
+// base64ShiftEncoder returns the part of a secret's base64 form that survives when the
+// secret does not start on a 3-byte boundary of the payload it is embedded in. base64
+// encodes three bytes at a time, so `Authorization: Basic base64("user:token")` contains
+// the base64 of the token alone only when the prefix length happens to be a multiple of
+// three; at the other two alignments the encoding of the whole value differs. Encoding
+// the secret behind shift filler bytes reproduces those alignments, which is what the
+// Base64StringEscapeShift1/2 encoders of GitHub's runner do.
+//
+// The leading group (filler mixed with the secret's first bytes) and the trailing group
+// (padded here, but continuing into whatever follows the secret) are dropped, leaving the
+// group-aligned middle that does appear verbatim in the log.
+func base64ShiftEncoder(shift int) func(string) string {
+	return func(v string) string {
+		buf := make([]byte, shift+len(v))
+		copy(buf[shift:], v)
+		encoded := base64.StdEncoding.EncodeToString(buf)
+		// Keep only the aligned middle, and only when enough of it is left to be a
+		// distinctive pattern rather than a fragment that matches unrelated output.
+		if len(encoded) < 8+minShiftedBase64Len {
+			return ""
+		}
+		return encoded[4 : len(encoded)-4]
+	}
+}
+
+// jsonStringEscape returns v as it appears inside a JSON string, without the quotes,
+// which is what `toJSON(secrets)` or any action logging a JSON body produces. Go's encoder
+// escapes <, >, & (as act's own toJSON does); the non-HTML variant below covers the runtimes
+// that do not. When v has none of those characters both forms are equal and deduplicated.
+func jsonStringEscape(v string) string {
+	encoded, err := json.Marshal(v)
+	if err != nil {
+		return v
+	}
+	return string(encoded[1 : len(encoded)-1])
+}
+
+// jsonStringEscapeNoHTML is jsonStringEscape without HTML escaping, matching the JSON a
+// JavaScript (JSON.stringify) or .NET action emits, so a secret containing < > or & is
+// masked in that form too.
+func jsonStringEscapeNoHTML(v string) string {
+	var buf bytes.Buffer
+	enc := json.NewEncoder(&buf)
+	enc.SetEscapeHTML(false)
+	if err := enc.Encode(v); err != nil {
+		return v
+	}
+	// Encode appends a newline; drop it along with the surrounding quotes.
+	encoded := strings.TrimRight(buf.String(), "\n")
+	return encoded[1 : len(encoded)-1]
+}
+
 func AppendSecretMasker(oldnew []string, v string) []string {
 	ret := oldnew
 
@@ -182,6 +255,21 @@ func AppendSecretMasker(oldnew []string, v string) []string {
 		}
 	}
 
+	// The encoded forms are derived from the whole value: a multi-line secret is
+	// encoded as one string, not line by line.
+	trimmed := strings.TrimSpace(v)
+	if len(trimmed) <= 1 {
+		return ret
+	}
+	for _, encode := range secretValueEncoders {
+		encoded := encode(trimmed)
+		// An encoding that leaves the value unchanged is already masked above.
+		if encoded == trimmed || len(encoded) <= 1 || slices.Contains(ret, encoded) {
+			continue
+		}
+		ret = append(ret, encoded, "***")
+	}
+
 	return ret
 }
 
@@ -194,6 +282,18 @@ func valueMasker(insecureSecrets bool, secrets map[string]string) entryProcessor
 	}
 	oldnew = slices.Clip(oldnew)
 	defReplacer := strings.NewReplacer(oldnew...)
+
+	// A ::add-mask:: only ever appends to the job's mask slice, so the replacer built for
+	// it stays valid until the slice grows. Cache it, keyed by the slice itself and its
+	// length, instead of encoding every secret and mask again for each log line.
+	var (
+		mu       sync.Mutex
+		masksRef *[]string
+		pairs    []string
+		masked   int
+		replacer *strings.Replacer
+	)
+
 	return func(entry *logrus.Entry) *logrus.Entry {
 		if insecureSecrets {
 			return entry
@@ -203,15 +303,26 @@ func valueMasker(insecureSecrets bool, secrets map[string]string) entryProcessor
 
 		if len(*masks) == 0 {
 			entry.Message = defReplacer.Replace(entry.Message)
-		} else {
-			cmasker := oldnew
-
-			for _, v := range *masks {
-				cmasker = AppendSecretMasker(cmasker, v)
+			return entry
 		}
 
-			entry.Message = strings.NewReplacer(cmasker...).Replace(entry.Message)
+		mu.Lock()
+		// A composite action logs through the same masker with its own mask slice, so a
+		// different slice starts the cache over.
+		if masksRef != masks {
+			masksRef, pairs, masked, replacer = masks, oldnew, 0, nil
 		}
+		if replacer == nil || masked != len(*masks) {
+			for _, v := range (*masks)[masked:] {
+				pairs = AppendSecretMasker(pairs, v)
+			}
+			masked = len(*masks)
+			replacer = strings.NewReplacer(pairs...)
+		}
+		cmasker := replacer
+		mu.Unlock()
+
+		entry.Message = cmasker.Replace(entry.Message)
 
 		return entry
 	}

act/runner/logger_test.go

@@ -4,7 +4,9 @@
 package runner
 
 import (
+	"encoding/base64"
 	"io"
+	"net/url"
 	"strings"
 	"testing"
 
@@ -59,6 +61,136 @@ func TestValueMasker(t *testing.T) {
 	}
 }
 
+// A secret that reaches the log through an encoding — a base64 payload, a JSON body, a
+// URL — must be masked as well: masking only the verbatim value leaks it.
+func TestValueMaskerEncodedSecrets(t *testing.T) {
+	secret := `p@ss w"rd/1`
+	masker := valueMasker(false, map[string]string{"TOKEN": secret})
+
+	for _, tc := range []struct {
+		name string
+		line string
+	}{
+		{"verbatim", "the token is " + secret},
+		{"base64", "Authorization: Basic " + base64.StdEncoding.EncodeToString([]byte(secret))},
+		{"json", `{"token":"` + jsonStringEscape(secret) + `"}`},
+		{"query escaped", "https://example.com/?token=" + url.QueryEscape(secret)},
+		{"path escaped", "https://example.com/" + url.PathEscape(secret) + "/x"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			entry := masker(&logrus.Entry{Context: t.Context(), Message: tc.line})
+
+			assert.Contains(t, entry.Message, "***")
+			assert.NotContains(t, entry.Message, secret)
+			assert.NotContains(t, entry.Message, base64.StdEncoding.EncodeToString([]byte(secret)))
+			assert.NotContains(t, entry.Message, url.QueryEscape(secret))
+		})
+	}
+}
+
+// A secret containing " together with <, > or & serializes to JSON differently depending
+// on the runtime: act's own toJSON (and Go) HTML-escape <>&, while a JavaScript
+// (JSON.stringify) or .NET action leaves them literal. The secret must be masked in either
+// form, so a JS-serialized JSON body does not leak it.
+func TestValueMaskerJSONEscapesBothWays(t *testing.T) {
+	secret := `a"<b>&c`
+	masker := valueMasker(false, map[string]string{"TOKEN": secret})
+
+	for _, tc := range []struct {
+		name string
+		form string
+	}{
+		{"html escaped (act toJSON / Go)", jsonStringEscape(secret)},
+		{"literal (JS JSON.stringify / .NET)", jsonStringEscapeNoHTML(secret)},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			entry := masker(&logrus.Entry{Context: t.Context(), Message: `{"t":"` + tc.form + `"}`})
+
+			assert.Contains(t, entry.Message, "***")
+			assert.NotContains(t, entry.Message, tc.form)
+		})
+	}
+}
+
+// ::add-mask:: values go through the same masker, so they get the same treatment.
+func TestValueMaskerEncodedMasks(t *testing.T) {
+	masks := []string{"s3cr3t value"}
+	masker := valueMasker(false, nil)
+
+	entry := masker(&logrus.Entry{
+		Context: WithMasks(t.Context(), &masks),
+		Message: "encoded: " + base64.StdEncoding.EncodeToString([]byte("s3cr3t value")),
+	})
+
+	assert.Equal(t, "encoded: ***", entry.Message)
+}
+
+// A token in a Basic auth header is base64'd together with the user name, so the token's
+// own base64 only appears when the prefix length is a multiple of three. The other two
+// alignments must be masked as well, or `Authorization: Basic base64("user:token")` leaks
+// the token to anyone who can decode the log.
+func TestValueMaskerBase64Alignments(t *testing.T) {
+	secret := "s3cr3t-token-value"
+	masker := valueMasker(false, map[string]string{"TOKEN": secret})
+
+	// One prefix per alignment: len%3 of 0, 1 and 2.
+	for _, prefix := range []string{"x-access-token:", "user:", "ab:"} {
+		t.Run(prefix, func(t *testing.T) {
+			encoded := base64.StdEncoding.EncodeToString([]byte(prefix + secret))
+			entry := masker(&logrus.Entry{Context: t.Context(), Message: "Authorization: Basic " + encoded})
+
+			assert.Contains(t, entry.Message, "***")
+			// The aligned middle of the secret must be gone, so the payload can no longer be
+			// decoded back into the token.
+			assert.NotEqual(t, "Authorization: Basic "+encoded, entry.Message)
+			decodable := strings.TrimPrefix(entry.Message, "Authorization: Basic ")
+			decoded, err := base64.StdEncoding.DecodeString(decodable)
+			if err == nil {
+				assert.NotContains(t, string(decoded), secret)
+			}
+		})
+	}
+}
+
+// The masker caches its replacer, so it has to notice both a mask appended to the same
+// slice and a composite action logging with a slice of its own.
+func TestValueMaskerCachedReplacerSeesNewMasks(t *testing.T) {
+	masker := valueMasker(false, map[string]string{"TOKEN": "secret-token"})
+	mask := func(masks *[]string, message string) string {
+		return masker(&logrus.Entry{Context: WithMasks(t.Context(), masks), Message: message}).Message
+	}
+
+	job := []string{"first mask"}
+	assert.Equal(t, "a *** and ***", mask(&job, "a first mask and secret-token"))
+
+	// ::add-mask:: appends to the same slice
+	job = append(job, "second mask")
+	assert.Equal(t, "*** and ***", mask(&job, "first mask and second mask"))
+
+	// a composite action brings its own slice
+	composite := []string{"composite mask"}
+	assert.Equal(t, "*** but first mask", mask(&composite, "composite mask but first mask"))
+
+	// and the job's masks still apply once it is back
+	assert.Equal(t, "*** and *** but composite mask", mask(&job, "first mask and second mask but composite mask"))
+}
+
+func TestAppendSecretMaskerSkipsUselessEncodings(t *testing.T) {
+	// A token with no character an escape would touch only gains its base64 forms:
+	// JSON, query and path escaping all leave it unchanged.
+	pairs := AppendSecretMasker(nil, "plaintoken")
+	assert.Equal(t, []string{
+		"plaintoken", "***",
+		base64.StdEncoding.EncodeToString([]byte("plaintoken")), "***",
+		// The two shifted alignments, each without its leading and trailing group.
+		"YWludG9r", "***",
+		"bGFpbnRv", "***",
+	}, pairs)
+
+	// Too short to mask.
+	assert.Empty(t, AppendSecretMasker(nil, "x"))
+}
+
 func TestJobLogFormatterDecodesCommandData(t *testing.T) {
 	logger := logrus.New()
 	logger.Out = io.Discard