chore(deps): update dependencies (#1365)

Co-authored-by: Renovate Bot <[email protected]>

.dockerignore

@@ -45,8 +45,13 @@ cpu.out
 
 /bin
 /dist
+/.tmp
+/.git/hooks
+/.git/logs
+/.git/rr-cache
 /.env
 /.runner
+/.runner.lock
 /config.yaml
 /Dockerfile
 .DS_Store

.gitea/actions/go-setup/action.yml

@@ -0,0 +1,29 @@
+name: go-setup
+description: Set up go and restore its caches, which only the cache-seeder workflow saves
+
+runs:
+  using: composite
+  steps:
+    - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
+      id: setup-go
+      with:
+        go-version-file: go.mod
+        check-latest: true
+        cache: false
+    - if: ${{ github.workflow == 'cache-seeder' }}
+      uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+      with:
+        path: |
+          ~/go/pkg/mod
+          ~/.cache/go-build
+          ~/.cache/golangci-lint
+        key: go-${{ runner.os }}-${{ runner.arch }}-${{ steps.setup-go.outputs.go-version }}-${{ hashFiles('go.mod', 'go.sum', '.golangci.yml', 'Makefile', '.gitea/actions/go-setup/action.yml', '.gitea/workflows/cache-seeder.yml') }}
+    - if: ${{ github.workflow != 'cache-seeder' }}
+      uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+      with:
+        path: |
+          ~/go/pkg/mod
+          ~/.cache/go-build
+          ~/.cache/golangci-lint
+        key: go-${{ runner.os }}-${{ runner.arch }}-${{ steps.setup-go.outputs.go-version }}-${{ hashFiles('go.mod', 'go.sum', '.golangci.yml', 'Makefile', '.gitea/actions/go-setup/action.yml', '.gitea/workflows/cache-seeder.yml') }}
+        restore-keys: go-${{ runner.os }}-${{ runner.arch }}-${{ steps.setup-go.outputs.go-version }}-

.gitea/workflows/cache-seeder.yml

@@ -0,0 +1,26 @@
+name: cache-seeder
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+    paths:
+      - go.mod
+      - go.sum
+      - .golangci.yml
+      - Makefile
+      - .gitea/actions/go-setup/action.yml
+      - .gitea/workflows/cache-seeder.yml
+
+jobs:
+  seed:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
+      - uses: ./.gitea/actions/go-setup
+      - run: make deps-tools lint build
+      - name: compile the test binaries without running them
+        run: |
+          make test GOTEST_FLAGS='-race -exec true'
+          go test -tags e2e -exec true ./e2e/...

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

@@ -21,7 +21,7 @@ jobs:
     - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
     - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7
       with:
-        node-version: 24
+        node-version: 26
     - run: make lint-pr-title
       env:
         PR_TITLE: ${{ github.event.pull_request.title }}

.gitea/workflows/release-nightly.yml

@@ -35,7 +35,7 @@ jobs:
           go-version-file: "go.mod"
           check-latest: true
       - name: goreleaser
-        uses: goreleaser/goreleaser-action@f06c13b6b1a9625abc9e6e439d9c05a8f2190e94 # v7
+        uses: goreleaser/goreleaser-action@a4f02ab5355a589de60d917a659ce617ed19b3de # v7
         with:
           distribution: goreleaser-pro
           args: release --nightly
@@ -79,7 +79,7 @@ jobs:
         uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4
 
       - name: Login to DockerHub
-        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4
+        uses: docker/login-action@81c6f93d4bc991e510b342a00b419b3bb4efc5c6 # v4
         with:
           username: ${{ secrets.DOCKER_USERNAME }}
           password: ${{ secrets.DOCKER_PASSWORD }}

.gitea/workflows/release-tag.yml

@@ -35,7 +35,7 @@ jobs:
           passphrase: ${{ secrets.PASSPHRASE }}
           fingerprint: CC64B1DB67ABBEECAB24B6455FC346329753F4B0
       - name: goreleaser
-        uses: goreleaser/goreleaser-action@f06c13b6b1a9625abc9e6e439d9c05a8f2190e94 # v7
+        uses: goreleaser/goreleaser-action@a4f02ab5355a589de60d917a659ce617ed19b3de # v7
         with:
           distribution: goreleaser-pro
           args: release
@@ -82,7 +82,7 @@ jobs:
         uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4
 
       - name: Login to DockerHub
-        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4
+        uses: docker/login-action@81c6f93d4bc991e510b342a00b419b3bb4efc5c6 # v4
         with:
           username: ${{ secrets.DOCKER_USERNAME }}
           password: ${{ secrets.DOCKER_PASSWORD }}

.gitea/workflows/test.yml

@@ -17,25 +17,24 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
-        with:
-          go-version-file: 'go.mod'
-          check-latest: true
+      - uses: ./.gitea/actions/go-setup
       - name: prepare anonymous docker config
         run: mkdir -p "$DOCKER_CONFIG" && echo '{}' > "$DOCKER_CONFIG/config.json"
       # Pre-pull act/runner's two largest base images so a slow pull can't dominate `make test`;
       # the rest (alpine/ubuntu) pull on demand, absorbed by the make-test -timeout. The host
-      # daemon retains them between runs, so this is usually a fast manifest re-check.
+      # daemon retains them between runs, and their digest pins skip the registry when it has them.
       - name: pre-pull test images
         env:
-          TEST_JOB_IMAGE: node:24-bookworm-slim@sha256:d6aa754f16b3197301076f047b5def2f02ea1dbbc2ca920407d46d7ec7f87b20 # renovate: datasource=docker
-          TEST_SERVICE_IMAGE: nginx:alpine@sha256:df221db836e1754089190208cee7eeda94f233197056426eda74a43ab1abeac2 # renovate: datasource=docker
+          TEST_JOB_IMAGE: node:26-bookworm-slim@sha256:bc0a2afd3b74ce12222fc8bab99271e66257f021078f5f4a5d687a77d280c1d9 # renovate: datasource=docker
+          TEST_SERVICE_IMAGE: nginx:alpine@sha256:a3554c0d0b48177b1413c26e01f695334428020d9df01243fbae3e11cfd17417 # renovate: datasource=docker
         run: |
           for image in "$TEST_JOB_IMAGE" "$TEST_SERVICE_IMAGE"; do
+            if [ "$(docker image inspect --format '{{.Os}}/{{.Architecture}}' "$image" 2>/dev/null)" != linux/amd64 ]; then
               for attempt in 1 2 3; do
                 docker pull "$image" && break
                 [ "$attempt" = 3 ] || sleep 5
               done
+            fi
             docker tag "$image" "${image%@*}"
           done
       - name: lint
@@ -68,10 +67,7 @@ jobs:
           - gitea/gitea:main-nightly
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7
-      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7
-        with:
-          go-version-file: 'go.mod'
-          check-latest: true
+      - uses: ./.gitea/actions/go-setup
       - name: prepare anonymous docker config
         run: mkdir -p "$DOCKER_CONFIG" && echo '{}' > "$DOCKER_CONFIG/config.json"
       - name: e2e compatibility

DEVELOPMENT.md

@@ -57,7 +57,15 @@ Each lives in `internal/pkg/action/<name>` and is registered in `builtinActions`
 
 `make test-e2e` runs the runner against `E2E_GITEA_IMAGE`. It defaults to the nightly image.
 It requires Docker and is excluded from `make test`. CI runs stable and nightly variants in
-parallel.
+parallel. Outside a container it runs `go test` in a Linux container on its own network, as CI does,
+so both take the same paths. `E2E_ARGS` passes extra flags, e.g. `make test-e2e E2E_ARGS='-run TestCompatibility/dispatch'`.
 
-The suite shares one Gitea and regular runner. Cache and ephemeral scenarios use isolated
+The `cli_*` scenarios build the binary and drive `register` and `daemon` as a user would, including `--once`,
+`--labels` and shutdown with a running job. `failure_paths` covers failed, skipped and tolerated jobs, `job_timeout`
+a job killed at the runner's deadline. `label_routing` and `capacity` cover scheduling, `reconnect` drops the
+connection to Gitea through a proxy, and `host_mode` and `reusable_workflow` cover host labels and `workflow_call`.
+`dispatch`, `failure_paths` and `reusable_workflow` run in every mode: through the daemon and `exec`, each on
+Docker and Kubernetes.
+
+The suite shares one Gitea, regular runner and k3s cluster. Cache and ephemeral scenarios use isolated
 repository runners. A run path is `<workflow>@<ref>` and a log row is `<timestamp>Z <payload>`.

Dockerfile

@@ -12,12 +12,14 @@ ENV GOPROXY=${GOPROXY:-}
 COPY . /opt/src/runner
 WORKDIR /opt/src/runner
 
-RUN make clean && make build
+RUN --mount=type=cache,target=/go/pkg/mod \
+    --mount=type=cache,target=/root/.cache/go-build \
+    make clean && make build
 
 ### DIND VARIANT
 #
 #
-FROM docker:29.9.1-dind AS dind
+FROM docker:29.9.2-dind AS dind
 
 ARG VERSION=dev
 
@@ -39,7 +41,7 @@ ENTRYPOINT ["s6-svscan","/etc/s6"]
 ### DIND-ROOTLESS VARIANT
 #
 #
-FROM docker:29.9.1-dind-rootless AS dind-rootless
+FROM docker:29.9.2-dind-rootless AS dind-rootless
 
 ARG VERSION=dev
 

Makefile

@@ -4,7 +4,7 @@ GO ?= go
 HAS_GO = $(shell hash $(GO) > /dev/null 2>&1 && echo "GO" || echo "NOGO" )
 
 PODMAN_TEST_IMAGE ?= quay.io/containers/podman:v5.8.9-immutable@sha256:87f42f37fa85b421664c30b8f5f9cfc23920057f5e5fcadf0b6f68d9edc6d415 # renovate: datasource=docker
-E2E_JOB_IMAGE ?= docker.gitea.com/runner-images:ubuntu-latest@sha256:0e62e56b382ebf485bff1e51a05cbe27c708545a2e91425467fd31eb3e249521 # renovate: datasource=docker
+E2E_JOB_IMAGE ?= docker.gitea.com/runner-images:ubuntu-latest@sha256:657378e5cfd898e36f820459c60ef8e69be8fcda9fd354e4381a8b654d091bbc # renovate: datasource=docker
 SERVICE_IMAGE ?= nginx:1.31.7-alpine@sha256:892bb2152464f0c5d8972593a5f5afaf8054552580b699d95f8ff4af08c2d33d # renovate: datasource=docker
 E2E_GITEA_IMAGE ?= gitea/gitea:main-nightly
 E2E_K3S_IMAGE ?= rancher/k3s:v1.37.1-k3s1@sha256:ca7f37d993d82ef0dcdcfecb2e0e2618ea541dbaffc620c8cedebe01a82acd0d # renovate: datasource=docker
@@ -15,8 +15,8 @@ DOCKER_IMAGE ?= gitea/runner
 DOCKER_TAG ?= nightly
 DOCKER_REF := $(DOCKER_IMAGE):$(DOCKER_TAG)
 
-GOLANGCI_LINT_PACKAGE ?= github.com/golangci/golangci-lint/v2/cmd/[email protected]4.0 # renovate: datasource=go
-GOVULNCHECK_PACKAGE ?= golang.org/x/vuln/cmd/govulncheck@v1.8.0 # renovate: datasource=go
+GOLANGCI_LINT_PACKAGE ?= github.com/golangci/golangci-lint/v2/cmd/[email protected]5.0 # renovate: datasource=go
+GOVULNCHECK_PACKAGE ?= golang.org/x/vuln/cmd/govulncheck@v1.9.0 # renovate: datasource=go
 
 GOTEST_FLAGS ?= -race -timeout 20m -parallel 8
 
@@ -154,7 +154,7 @@ E2E_CONCURRENCY ?= 8
 
 .PHONY: test-e2e
 test-e2e: ## run Gitea compatibility tests against E2E_GITEA_IMAGE
-	@E2E_CONCURRENCY=$(E2E_CONCURRENCY) E2E_GITEA_IMAGE=$(E2E_GITEA_IMAGE) E2E_K3S_IMAGE=$(E2E_K3S_IMAGE) E2E_JOB_IMAGE=$(E2E_JOB_IMAGE) GO=$(GO) SERVICE_IMAGE=$(SERVICE_IMAGE) ./tools/test-e2e.sh
+	@E2E_CONCURRENCY=$(E2E_CONCURRENCY) E2E_GITEA_IMAGE=$(E2E_GITEA_IMAGE) E2E_K3S_IMAGE=$(E2E_K3S_IMAGE) E2E_JOB_IMAGE=$(E2E_JOB_IMAGE) GO=$(GO) SERVICE_IMAGE=$(SERVICE_IMAGE) ./tools/test-e2e.sh $(E2E_ARGS)
 
 .PHONY: install
 install: $(GOFILES) ## install the runner binary via `go install`

README.md

@@ -303,6 +303,8 @@ Use `uses: builtin:checkout` to check out a repository without downloading `acti
 
 Each runner starts its own cache server, so runners do not share cached entries. A runner in a container on the jobs' Docker daemon attaches itself to each job network, so jobs reach its cache without extra setup.
 
+As on GitHub, a job restores caches saved for its own ref, its pull request's base branch and the default branch, so a cache saved by a pull request is only restored by that pull request.
+
 **Sharing a cache between runners**
 
 Run one `gitea-runner cache-server` and point every runner at it, using the same secret everywhere, for example from `openssl rand -hex 32`:

act/artifactcache/doc.go

@@ -6,7 +6,5 @@
 //
 // Inspired by https://github.com/sp-ricard-valverde/github-act-cache-server
 //
-// TODO: Authorization
-// TODO: Restrictions for accessing a cache, see https://docs.github.com/en/actions/using-workflows/caching-dependencies-to-speed-up-workflows#restrictions-for-accessing-a-cache
 // TODO: Force deleting cache entries, see https://docs.github.com/en/actions/using-workflows/caching-dependencies-to-speed-up-workflows#force-deleting-cache-entries
 package artifactcache

act/artifactcache/handler.go

@@ -15,6 +15,7 @@ import (
 	"encoding/json/v2"
 	"errors"
 	"fmt"
+	"io"
 	"net"
 	"net/http"
 	"net/http/httputil"
@@ -58,6 +59,10 @@ type credKey struct{}
 type JobCredential struct {
 	Repo string `json:"repo"`
 
+	Ref        string `json:"ref"`
+	BaseRef    string `json:"base_ref"`
+	DefaultRef string `json:"default_ref"`
+
 	// Results is the instance whose artifact service this server forwards for the job, and
 	// InsecureTLS how the runner reaches it; see results.go.
 	Results     string `json:"-"`
@@ -67,6 +72,10 @@ type JobCredential struct {
 	PublicURL string `json:"public_url"`
 }
 
+func (cred JobCredential) owns(cache *Cache) bool {
+	return cache.Repo == cred.Repo && cache.Ref == cred.Ref
+}
+
 // credEntry holds a registered job's credential along with an active
 // registration count. RegisterJob is reference-counted so that if two tasks
 // briefly share an ACTIONS_RUNTIME_TOKEN — e.g. a runner that retries a task
@@ -652,7 +661,7 @@ func (h *Handler) find(w http.ResponseWriter, r *http.Request) {
 	}
 	defer db.Close()
 
-	cache, err := h.lookupCache(db, cred.Repo, keys, version)
+	cache, err := h.lookupCache(db, cred, keys, version)
 	if err != nil {
 		h.responseJSON(w, r, 500, err)
 		return
@@ -668,11 +677,10 @@ func (h *Handler) find(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
-// 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 *cacheDB, repo string, keys []string, version string) (*Cache, error) {
-	cache, err := findCache(db, repo, keys, version)
+// lookupCache returns the entry to restore or (nil, nil), dropping matches whose blob was pruned.
+func (h *Handler) lookupCache(db *cacheDB, cred JobCredential, keys []string, version string) (*Cache, error) {
+	for {
+		cache, err := findCache(db, cred, keys, version)
 		if err != nil || cache == nil {
 			return nil, err
 		}
@@ -680,15 +688,17 @@ func (h *Handler) lookupCache(db *cacheDB, repo string, keys []string, version s
 		if err != nil {
 			return nil, err
 		}
-	if !ok {
-		h.deleteCache(db, cache)
-		return nil, nil //nolint:nilnil // absence is not an error here
-	}
+		if ok {
 			// Handing out a download URL counts as access, or eviction could drop the entry between
 			// this call and the GET that follows it.
 			h.touch(db, cache)
 			return cache, nil
 		}
+		if !h.deleteCache(db, cache) {
+			return nil, nil //nolint:nilnil // absence is not an error here
+		}
+	}
+}
 
 // POST /_apis/artifactcache/caches
 func (h *Handler) reserve(w http.ResponseWriter, r *http.Request) {
@@ -699,7 +709,7 @@ func (h *Handler) reserve(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	cache := &Cache{Repo: cred.Repo, Key: api.Key, Version: api.Version, Size: api.Size}
+	cache := &Cache{Repo: cred.Repo, Ref: cred.Ref, Key: api.Key, Version: api.Version, Size: api.Size}
 	if cache.Size == 0 {
 		cache.Size = -1
 	}
@@ -710,9 +720,10 @@ func (h *Handler) reserve(w http.ResponseWriter, r *http.Request) {
 	}
 	defer db.Close()
 
-	now := time.Now().Unix()
-	cache.CreatedAt = now
-	cache.UsedAt = now
+	now := time.Now()
+	cache.CreatedAt = now.Unix()
+	cache.CreatedAtNanos = now.Nanosecond()
+	cache.UsedAt = now.Unix()
 	if err := h.insertCache(db, cache); err != nil {
 		h.responseJSON(w, r, 500, err)
 		return
@@ -747,7 +758,7 @@ func (h *Handler) upload(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	if cache.Repo != cred.Repo {
+	if !cred.owns(cache) {
 		h.responseJSON(w, r, 403, fmt.Errorf("cache %d: forbidden", id))
 		return
 	}
@@ -778,6 +789,17 @@ func (h *Handler) commit(w http.ResponseWriter, r *http.Request) {
 		h.responseJSON(w, r, 400, err)
 		return
 	}
+	var request struct {
+		Size *int64 `json:"size"`
+	}
+	body, err := io.ReadAll(io.LimitReader(r.Body, 1<<20))
+	if err == nil && len(body) > 0 {
+		err = json.Unmarshal(body, &request)
+	}
+	if err != nil {
+		h.responseJSON(w, r, http.StatusBadRequest, err)
+		return
+	}
 
 	db, err := h.openDB()
 	if err != nil {
@@ -795,7 +817,7 @@ func (h *Handler) commit(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	if cache.Repo != cred.Repo {
+	if !cred.owns(cache) {
 		h.responseJSON(w, r, 403, fmt.Errorf("cache %d: forbidden", id))
 		return
 	}
@@ -805,6 +827,14 @@ func (h *Handler) commit(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	if request.Size != nil {
+		if *request.Size < 0 || cache.Size >= 0 && cache.Size != *request.Size {
+			h.responseJSON(w, r, http.StatusBadRequest, errors.New("invalid cache size"))
+			return
+		}
+		cache.Size = *request.Size
+	}
+
 	db.Close()
 
 	if err := h.commitCache(cache); err != nil {
@@ -1050,27 +1080,37 @@ func (h *Handler) signedArtifactURL(cred JobCredential, cacheID uint64, exp time
 	return h.signedURL(cred, apiPath+"/artifacts", "", cacheID, exp)
 }
 
-// if not found, return (nil, nil) instead of an error.
-func findCache(db *cacheDB, repo string, keys []string, version string) (*Cache, error) {
-	for _, prefix := range keys {
-		// if a key in the list matches exactly, don't return partial matches
-		exact, err := findExactCache(db, repo, prefix, version, true)
-		if err != nil {
-			return nil, err
+// findCache tries every key in one ref before the next ref, and returns (nil, nil) on a miss.
+func findCache(db *cacheDB, cred JobCredential, keys []string, version string) (*Cache, error) {
+	refs := []string{cred.Ref}
+	for _, ref := range []string{cred.BaseRef, cred.DefaultRef} {
+		if ref != "" && !slices.Contains(refs, ref) {
+			refs = append(refs, ref)
 		}
-		if exact != nil {
-			return exact, nil
 	}
 	caches, err := db.find(func(cache *Cache) bool {
-			return cache.Repo == repo && strings.HasPrefix(cache.Key, prefix) && cache.Version == version && cache.Complete
+		return cache.Repo == cred.Repo && cache.Version == version && cache.Complete && slices.Contains(refs, cache.Ref) &&
+			slices.ContainsFunc(keys, func(prefix string) bool { return strings.HasPrefix(cache.Key, prefix) })
 	})
 	if err != nil {
 		return nil, fmt.Errorf("find cache: %w", err)
 	}
-		if cache := latest(caches, func(cache *Cache) int64 { return cache.CreatedAt }); cache != nil {
+	for _, ref := range refs {
+		for _, prefix := range keys {
+			var exact, partial []*Cache
+			for _, cache := range caches {
+				if cache.Ref == ref && cache.Key == prefix {
+					exact = append(exact, cache)
+				} else if cache.Ref == ref && strings.HasPrefix(cache.Key, prefix) {
+					partial = append(partial, cache)
+				}
+			}
+			// if a key in the list matches exactly, don't return partial matches
+			if cache := cmp.Or(latest(exact, (*Cache).createdAtNano), latest(partial, (*Cache).createdAtNano)); cache != nil {
 				return cache, nil
 			}
 		}
+	}
 	return nil, nil //nolint:nilnil // pre-existing issue from nektos/act
 }
 
@@ -1078,7 +1118,9 @@ func latest(caches []*Cache, stamp func(*Cache) int64) *Cache {
 	if len(caches) == 0 {
 		return nil
 	}
-	return slices.MaxFunc(caches, func(left, right *Cache) int { return cmp.Compare(stamp(left), stamp(right)) })
+	return slices.MaxFunc(caches, func(left, right *Cache) int {
+		return cmp.Or(cmp.Compare(stamp(left), stamp(right)), cmp.Compare(left.ID, right.ID))
+	})
 }
 
 // findExactCache returns the entry for exactly this key and version, or (nil, nil) if there is
@@ -1088,16 +1130,16 @@ func latest(caches []*Cache, stamp func(*Cache) int64) *Cache {
 // 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 *cacheDB, repo, key, version string, complete bool) (*Cache, error) {
+func findExactCache(db *cacheDB, cred JobCredential, key, version string, complete bool) (*Cache, error) {
 	caches, err := db.find(func(cache *Cache) bool {
-		return cache.Key == key && cache.Repo == repo && cache.Version == version && cache.Complete == complete
+		return cred.owns(cache) && cache.Key == key && cache.Version == version && cache.Complete == complete
 	})
 	if err != nil {
 		return nil, fmt.Errorf("find cache: %w", err)
 	}
 	stamp := func(cache *Cache) int64 { return cache.UsedAt }
 	if complete {
-		stamp = func(cache *Cache) int64 { return cache.CreatedAt }
+		stamp = (*Cache).createdAtNano
 	}
 	return latest(caches, stamp), nil
 }
@@ -1257,17 +1299,18 @@ func (h *Handler) evictExpired(db *cacheDB) {
 	h.sweep(db, func(cache *Cache) bool { return cache.UsedAt < expired })
 }
 
-// evictSuperseded removes entries a newer one with the same key and version replaced. The
-// grouping includes Repo so two repos sharing a (key, version) do not evict each other.
+// evictSuperseded removes entries a newer one with the same repo, ref, key and version replaced.
 func (h *Handler) evictSuperseded(db *cacheDB) {
-	type entry struct{ repo, key, version string }
+	type entry struct{ repo, ref, key, version string }
 	groups := make(map[entry][]*Cache)
 	for _, cache := range h.caches(db, func(cache *Cache) bool { return cache.Complete }) {
-		group := entry{cache.Repo, cache.Key, cache.Version}
+		group := entry{cache.Repo, cache.Ref, cache.Key, cache.Version}
 		groups[group] = append(groups[group], cache)
 	}
 	for _, caches := range groups {
-		slices.SortFunc(caches, func(left, right *Cache) int { return cmp.Compare(left.CreatedAt, right.CreatedAt) })
+		slices.SortFunc(caches, func(left, right *Cache) int {
+			return cmp.Or(cmp.Compare(left.createdAtNano(), right.createdAtNano()), cmp.Compare(left.ID, right.ID))
+		})
 		for _, cache := range caches[:len(caches)-1] {
 			if inUse(cache) {
 				continue

act/artifactcache/handler_test.go

@@ -15,6 +15,7 @@ import (
 	"net/http/httptest"
 	"os"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"testing"
 	"testing/synctest"
@@ -47,10 +48,12 @@ func signArtifactURL(h *Handler, id int64) string {
 }
 
 func TestHandler_ExternalURL(t *testing.T) {
+	t.Parallel()
 	assert.Equal(t, "http://[2001:db8::1]:8080", (&Handler{outboundIP: "2001:db8::1", port: 8080}).ExternalURL())
 }
 
 func TestHandler_Upstream(t *testing.T) {
+	t.Parallel()
 	var upstream *httptest.Server
 	upstream = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		assert.Equal(t, http.MethodPut, r.Method)
@@ -100,6 +103,7 @@ func TestHandler_Upstream(t *testing.T) {
 }
 
 func TestHandler(t *testing.T) {
+	t.Parallel()
 	const version = "c19da02a2bd7e77277f1ac29ab45c09b7d46a4ee758284e26bb3045ad11d9d20"
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
@@ -302,6 +306,37 @@ func TestHandler(t *testing.T) {
 		}
 	})
 
+	t.Run("chunked commit validates size omitted at reservation", func(t *testing.T) {
+		for _, size := range []int64{5, 10} {
+			t.Run(strconv.FormatInt(size, 10), func(t *testing.T) {
+				id := reserveCache(t, testClient, base, strings.ToLower(t.Name()), version, 0)
+				req, err := http.NewRequest(http.MethodPatch, fmt.Sprintf("%s/caches/%d", base, id), strings.NewReader("cache"))
+				require.NoError(t, err)
+				req.Header.Set("Content-Range", "bytes 0-4/*")
+				resp, err := testClient.Do(req)
+				require.NoError(t, err)
+				resp.Body.Close()
+				require.Equal(t, http.StatusOK, resp.StatusCode)
+
+				resp, err = testClient.Post(fmt.Sprintf("%s/caches/%d", base, id), "application/json", io.MultiReader(strings.NewReader(fmt.Sprintf(`{"size":%d}`, size))))
+				require.NoError(t, err)
+				resp.Body.Close()
+				if size == 5 {
+					require.Equal(t, http.StatusOK, resp.StatusCode)
+				} else {
+					require.Equal(t, http.StatusInternalServerError, resp.StatusCode)
+				}
+
+				db, err := handler.openDB()
+				require.NoError(t, err)
+				defer db.Close()
+				cache, err := db.get(id)
+				require.NoError(t, err)
+				require.Equal(t, size == 5, cache.Complete)
+			})
+		}
+	})
+
 	t.Run("get with bad id", func(t *testing.T) {
 		resp, err := testClient.Get(base + "/artifacts/invalid_id")
 		require.NoError(t, err)
@@ -438,6 +473,7 @@ func uploadCacheNormally(t *testing.T, base, key, version string, content []byte
 }
 
 func TestHandler_gcCache(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir, Policy: Policy{Retention: testRetention}})
 	require.NoError(t, err)
@@ -540,6 +576,7 @@ func TestHandler_gcCache(t *testing.T) {
 }
 
 func TestHandler_evictPolicy(t *testing.T) {
+	t.Parallel()
 	now := time.Now()
 	stale := func(d time.Duration) int64 { return now.Add(-d).Unix() }
 	mib := func(n int64) int64 { return n * miB }
@@ -637,6 +674,7 @@ func TestHandler_evictPolicy(t *testing.T) {
 }
 
 func TestHandler_evictForFreeSpace(t *testing.T) {
+	t.Parallel()
 	free := func(n int64) func(string) (uint64, error) {
 		return func(string) (uint64, error) { return uint64(n), nil }
 	}
@@ -676,6 +714,7 @@ func TestHandler_evictForFreeSpace(t *testing.T) {
 }
 
 func TestHandler_SweepKeepsEntryWhenBlobSurvives(t *testing.T) {
+	t.Parallel()
 	cache := &Cache{Key: "stuck", Complete: true, UsedAt: time.Now().Add(-(testRetention + time.Hour)).Unix()}
 	handler := newTestHandler(t, Policy{Retention: testRetention}, cache)
 
@@ -690,6 +729,7 @@ func TestHandler_SweepKeepsEntryWhenBlobSurvives(t *testing.T) {
 }
 
 func TestHandler_FindProtectsFromEviction(t *testing.T) {
+	t.Parallel()
 	wanted := &Cache{Repo: testRepo, Key: "wanted", Version: "v", Complete: true, Size: 4 * miB, UsedAt: time.Now().Add(-3 * time.Hour).Unix()}
 	other := &Cache{Repo: testRepo, Key: "other", Version: "v", Complete: true, Size: 4 * miB, UsedAt: time.Now().Add(-2 * time.Hour).Unix()}
 	newest := &Cache{Repo: testRepo, Key: "newest", Version: "v", Complete: true, Size: 4 * miB, UsedAt: time.Now().Add(-time.Hour).Unix()}
@@ -709,7 +749,67 @@ func TestHandler_FindProtectsFromEviction(t *testing.T) {
 	assert.Equal(t, []string{wanted.Key}, keptKeys(t, handler, []*Cache{wanted, other}), "the entry just promised to a job survives, the next least recently used goes instead")
 }
 
+func TestHandler_PrefixUsesNewestReservationWithinSecond(t *testing.T) {
+	t.Parallel()
+	created := time.Now().Add(-time.Hour).Unix()
+	older := &Cache{Repo: testRepo, Key: "deps-old", Version: "v", Complete: true, CreatedAt: created}
+	newer := &Cache{Repo: testRepo, Key: "deps-new", Version: "v", Complete: true, CreatedAt: created}
+	handler := newTestHandler(t, Policy{}, older, newer)
+
+	db, err := handler.openDB()
+	require.NoError(t, err)
+	defer db.Close()
+	cache, err := findCache(db, JobCredential{Repo: testRepo}, []string{"deps-"}, "v")
+	require.NoError(t, err)
+	require.Equal(t, newer, cache)
+}
+
+func TestHandler_RestoreRefs(t *testing.T) {
+	t.Parallel()
+	const main, release, pull = "refs/heads/main", "refs/heads/release", "refs/pull/1/head"
+	now := time.Now().Unix()
+	entry := func(ref, key string, age int64) *Cache {
+		return &Cache{Repo: testRepo, Ref: ref, Key: key, Version: "v", Complete: true, CreatedAt: now - 3600 + age, UsedAt: now}
+	}
+	exactOnMain := entry(main, "deps-exact", -1)
+	onRelease := entry(release, "deps-release", 0)
+	onMain := entry(main, "deps-main", 1)
+	onPull := entry(pull, "deps-pull", 2)
+	unscoped := entry("", "deps-unscoped", 3)
+	handler := newTestHandler(t, Policy{}, exactOnMain, onRelease, onMain, onPull, unscoped, entry("refs/pull/2/head", "deps-lost", 4))
+	for _, cache := range []*Cache{exactOnMain, onRelease, onMain, onPull, unscoped} {
+		writeBlob(t, handler, cache.ID)
+	}
+
+	db, err := handler.openDB()
+	require.NoError(t, err)
+	defer db.Close()
+
+	for _, tc := range []struct {
+		name string
+		cred JobCredential
+		keys []string
+		want *Cache
+	}{
+		{"the default branch never restores what a pull request wrote", JobCredential{Ref: main, DefaultRef: main}, []string{"deps-"}, onMain},
+		{"a pull request restores its own entry first", JobCredential{Ref: pull, BaseRef: release, DefaultRef: main}, []string{"deps-"}, onPull},
+		{"a pull request falls back past a lost blob to its base branch before the default branch", JobCredential{Ref: "refs/pull/2/head", BaseRef: release, DefaultRef: main}, []string{"deps-"}, onRelease},
+		{"a branch does not restore a sibling branch", JobCredential{Ref: "refs/heads/other", DefaultRef: main}, []string{"deps-"}, onMain},
+		{"a prefix in the own ref wins over an exact key in the default branch", JobCredential{Ref: release, DefaultRef: main}, []string{"deps-exact", "deps-"}, onRelease},
+		{"a job with a ref never restores entries saved without one", JobCredential{Ref: "refs/heads/other"}, []string{"deps-"}, nil},
+		{"a job without a ref only restores entries saved without one", JobCredential{}, []string{"deps-"}, unscoped},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			tc.cred.Repo = testRepo
+			cache, err := handler.lookupCache(db, tc.cred, tc.keys, "v")
+			require.NoError(t, err)
+			assert.Equal(t, tc.want, cache)
+		})
+	}
+}
+
 func TestHandler_evictOnCommit(t *testing.T) {
+	t.Parallel()
 	full := &Cache{Repo: testRepo, Key: "full", Version: "v", Complete: true, Size: 4 * miB, UsedAt: time.Now().Add(-time.Hour).Unix()}
 	handler := newTestHandler(t, Policy{RepoSizeLimit: 4 * miB}, full)
 
@@ -719,6 +819,7 @@ func TestHandler_evictOnCommit(t *testing.T) {
 }
 
 func TestHandler_gcCacheInterval(t *testing.T) {
+	t.Parallel()
 	cache := &Cache{Key: "temp", UsedAt: time.Now().Add(-time.Hour).Unix()}
 	handler := newTestHandler(t, Policy{SweepInterval: 30 * time.Minute}, cache)
 
@@ -728,6 +829,7 @@ func TestHandler_gcCacheInterval(t *testing.T) {
 }
 
 func TestHandler_touchCacheSkipsFreshCompletedEntries(t *testing.T) {
+	t.Parallel()
 	fresh := &Cache{Repo: testRepo, Key: "fresh", Version: "v", Complete: true, UsedAt: time.Now().Unix()}
 	pending := &Cache{Repo: testRepo, Key: "pending", Version: "v", UsedAt: time.Now().Add(-time.Second).Unix()}
 	handler := newTestHandler(t, Policy{}, fresh, pending)
@@ -794,6 +896,7 @@ func writeBlob(t *testing.T, handler *Handler, id uint64) {
 }
 
 func TestHandler_RejectsBearer(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
 	require.NoError(t, err)
@@ -830,6 +933,7 @@ func TestHandler_RejectsBearer(t *testing.T) {
 }
 
 func TestHandler_UnregisterRevokes(t *testing.T) {
+	t.Parallel()
 	for _, tc := range []struct {
 		name          string
 		token         string
@@ -871,12 +975,14 @@ func TestHandler_UnregisterRevokes(t *testing.T) {
 }
 
 func TestHandler_CrossRepoIsolation(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
 	require.NoError(t, err)
 	defer handler.Close()
 	handler.RegisterJob("token-a", JobCredential{Repo: "owner/repoA"})
 	handler.RegisterJob("token-b", JobCredential{Repo: "owner/repoB"})
+	handler.RegisterJob("token-a-pull", JobCredential{Repo: "owner/repoA", Ref: "refs/pull/1/head"})
 
 	base := handler.ExternalURL() + apiPath
 	key := "shared-key"
@@ -911,16 +1017,24 @@ func TestHandler_CrossRepoIsolation(t *testing.T) {
 	resp.Body.Close()
 	assert.Equal(t, http.StatusOK, resp.StatusCode)
 
+	for _, client := range []*http.Client{clientB, {Transport: &bearerTransport{token: "token-a-pull"}}} {
 		req, err = http.NewRequest(http.MethodPatch, fmt.Sprintf("%s/caches/%d", base, id), bytes.NewReader([]byte("poison")))
 		require.NoError(t, err)
 		req.Header.Set("Content-Range", "bytes 0-5/*")
-	resp, err = clientB.Do(req)
+		resp, err = client.Do(req)
+		require.NoError(t, err)
+		resp.Body.Close()
+		assert.Equal(t, http.StatusForbidden, resp.StatusCode)
+
+		resp, err = client.Post(fmt.Sprintf("%s/caches/%d", base, id), "", nil)
 		require.NoError(t, err)
 		resp.Body.Close()
 		assert.Equal(t, http.StatusForbidden, resp.StatusCode)
 	}
+}
 
 func TestHandler_ArtifactSignature(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
 	require.NoError(t, err)
@@ -952,6 +1066,7 @@ func TestHandler_ArtifactSignature(t *testing.T) {
 }
 
 func TestHandler_SecretPersistsAcrossRestarts(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 
 	first, err := StartHandler(Options{Dir: dir, OutboundIP: "127.0.0.1"})
@@ -968,6 +1083,7 @@ func TestHandler_SecretPersistsAcrossRestarts(t *testing.T) {
 }
 
 func TestHandler_SeesWritesOfOtherProcesses(t *testing.T) {
+	t.Parallel()
 	handler := newTestHandler(t, Policy{})
 	insert := func(by *Handler, cache *Cache) {
 		db, err := by.openDB()
@@ -985,6 +1101,7 @@ func TestHandler_SeesWritesOfOtherProcesses(t *testing.T) {
 }
 
 func TestHandler_ArtifactSignatureDownload(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
 	require.NoError(t, err)
@@ -1018,51 +1135,24 @@ func TestHandler_ArtifactSignatureDownload(t *testing.T) {
 	assert.Equal(t, content, body)
 }
 
-func TestHandler_GC_PerRepoDedup(t *testing.T) {
-	dir := filepath.Join(t.TempDir(), "artifactcache")
-	handler, err := StartHandler(Options{Dir: dir})
-	require.NoError(t, err)
-	defer handler.Close()
-	handler.RegisterJob("tok-a", JobCredential{Repo: "owner/repoA"})
-	handler.RegisterJob("tok-b", JobCredential{Repo: "owner/repoB"})
-
-	key := "shared-dedup-key"
-	version := "c19da02a2bd7e77277f1ac29ab45c09b7d46a4ee758284e26bb3045ad11d9d20"
-
-	db, err := handler.openDB()
-	require.NoError(t, err)
+func TestHandler_GC_DedupPerRepoAndRef(t *testing.T) {
+	t.Parallel()
 	now := time.Now().Unix()
 	stale := time.Now().Add(-inUseGrace - time.Minute).Unix()
-	a := &Cache{Repo: "owner/repoA", Key: key, Version: version, Complete: true, CreatedAt: stale, UsedAt: stale, Size: 1}
-	b := &Cache{Repo: "owner/repoB", Key: key, Version: version, Complete: true, CreatedAt: now, UsedAt: now, Size: 1}
-	require.NoError(t, handler.insertCache(db, a))
-	require.NoError(t, handler.insertCache(db, b))
-	require.NoError(t, handler.storage.Write(a.ID, 0, strings.NewReader("a")))
-	_, err = handler.storage.Commit(a.ID, 1)
-	require.NoError(t, err)
-	require.NoError(t, handler.storage.Write(b.ID, 0, strings.NewReader("b")))
-	_, err = handler.storage.Commit(b.ID, 1)
-	require.NoError(t, err)
-	require.NoError(t, db.Close())
-
+	entries := []*Cache{
+		{Repo: "owner/repoA", Key: "shared", Version: "v", Complete: true, CreatedAt: stale, UsedAt: stale, Size: 1},
+		{Repo: "owner/repoB", Key: "shared", Version: "v", Complete: true, CreatedAt: now, UsedAt: now, Size: 1},
+		{Repo: "owner/repoA", Ref: "refs/pull/1/head", Key: "shared", Version: "v", Complete: true, CreatedAt: now, UsedAt: now, Size: 1},
+	}
+	handler := newTestHandler(t, Policy{}, entries...)
 	handler.gcAt = time.Time{}
 	handler.gcCache()
 
-	db, err = handler.openDB()
-	require.NoError(t, err)
-	defer db.Close()
-	after, err := db.find(func(cache *Cache) bool { return cache.Key == key && cache.Version == version })
-	require.NoError(t, err)
-
-	repos := make(map[string]bool)
-	for _, c := range after {
-		repos[c.Repo] = true
-	}
-	assert.True(t, repos["owner/repoA"], "repoA's cache must survive dedup against repoB")
-	assert.True(t, repos["owner/repoB"], "repoB's cache must survive dedup against repoA")
+	assert.Len(t, keptKeys(t, handler, entries), len(entries))
 }
 
 func TestHandler_InternalAPI_Disabled(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
 	require.NoError(t, err)
@@ -1077,6 +1167,7 @@ func TestHandler_InternalAPI_Disabled(t *testing.T) {
 }
 
 func TestHandler_InternalAPI_AuthAndUsage(t *testing.T) {
+	t.Parallel()
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	const secret = "internal-secret"
 	handler, err := StartHandler(Options{Dir: dir, InternalSecret: secret})
@@ -1122,10 +1213,10 @@ func TestHandler_InternalAPI_AuthAndUsage(t *testing.T) {
 		}
 
 		assert.Equal(t, http.StatusUnauthorized, probe("via-internal-api"))
-		assert.Equal(t, http.StatusOK, post("/_internal/register", secret, `{"token":"via-internal-api","repo":"owner/repo","public_url":"http://runner.example","results":"https://gitea.example","insecure_tls":true}`))
+		assert.Equal(t, http.StatusOK, post("/_internal/register", secret, `{"token":"via-internal-api","repo":"owner/repo","ref":"refs/pull/1/head","base_ref":"refs/heads/release","default_ref":"refs/heads/main","public_url":"http://runner.example","results":"https://gitea.example","insecure_tls":true}`))
 		cred, ok := handler.lookupCredential("via-internal-api")
 		require.True(t, ok)
-		assert.Equal(t, JobCredential{Repo: testRepo, PublicURL: "http://runner.example"}, cred)
+		assert.Equal(t, JobCredential{Repo: testRepo, Ref: "refs/pull/1/head", BaseRef: "refs/heads/release", DefaultRef: "refs/heads/main", PublicURL: "http://runner.example"}, cred)
 		assert.NotEqual(t, http.StatusUnauthorized, probe("via-internal-api"))
 		assert.Equal(t, http.StatusOK, post("/_internal/revoke", secret, `{"token":"via-internal-api"}`))
 		assert.Equal(t, http.StatusUnauthorized, probe("via-internal-api"))

act/artifactcache/handler_v2.go

@@ -73,7 +73,7 @@ func (h *Handler) v2CreateCacheEntry(w http.ResponseWriter, r *http.Request) {
 	// 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 {
+	if existing, err := findExactCache(db, cred, req.Key, req.Version, true); err != nil {
 		h.twirpError(w, r, twirpInternal, err)
 		return
 	} else if existing != nil {
@@ -84,7 +84,7 @@ func (h *Handler) v2CreateCacheEntry(w http.ResponseWriter, r *http.Request) {
 
 	// A second live reservation is finalized by whichever job calls last, against the other's upload.
 	owner := hashedToken(bearerToken(r))
-	if pending, err := findExactCache(db, cred.Repo, req.Key, req.Version, false); err != nil {
+	if pending, err := findExactCache(db, cred, req.Key, req.Version, false); err != nil {
 		h.twirpError(w, r, twirpInternal, err)
 		return
 	} else if pending != nil && pending.UsedAt > time.Now().Add(-uploadStallTimeout).Unix() {
@@ -100,15 +100,17 @@ func (h *Handler) v2CreateCacheEntry(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	now := time.Now().Unix()
+	now := time.Now()
 	cache := &Cache{
 		Repo:           cred.Repo,
+		Ref:            cred.Ref,
 		Key:            req.Key,
 		Version:        req.Version,
 		Owner:          owner,
 		Size:           -1, // the size is only known at finalize time
-		CreatedAt: now,
-		UsedAt:    now,
+		CreatedAt:      now.Unix(),
+		CreatedAtNanos: now.Nanosecond(),
+		UsedAt:         now.Unix(),
 	}
 	if err := h.insertCache(db, cache); err != nil {
 		h.twirpError(w, r, twirpInternal, err)
@@ -136,13 +138,13 @@ func (h *Handler) v2FinalizeCacheEntryUpload(w http.ResponseWriter, r *http.Requ
 	}
 	defer db.Close()
 
-	cache, err := findExactCache(db, cred.Repo, req.Key, req.Version, false)
+	cache, err := findExactCache(db, cred, req.Key, req.Version, false)
 	if err == nil && cache != nil && cache.Owner != hashedToken(bearerToken(r)) {
 		cache = nil // not the reservation this job made, so not this job's to commit
 	}
 	if err == nil && cache == nil {
 		// A retry whose first response was lost finds the entry already committed.
-		cache, err = findExactCache(db, cred.Repo, req.Key, req.Version, true)
+		cache, err = findExactCache(db, cred, req.Key, req.Version, true)
 	}
 	if err != nil {
 		h.twirpError(w, r, twirpInternal, err)
@@ -185,7 +187,7 @@ func (h *Handler) v2GetCacheEntryDownloadURL(w http.ResponseWriter, r *http.Requ
 	}
 	defer db.Close()
 
-	cache, err := h.lookupCache(db, cred.Repo, req.keys(), req.Version)
+	cache, err := h.lookupCache(db, cred, req.keys(), req.Version)
 	if err != nil {
 		h.twirpError(w, r, twirpInternal, err)
 		return

act/artifactcache/handler_v2_test.go

@@ -78,6 +78,7 @@ func saveV2(t *testing.T, handler *Handler, key, version string, content []byte)
 }
 
 func TestCacheServiceV2RoundTrip(t *testing.T) {
+	t.Parallel()
 	handler := newTestHandler(t, Policy{})
 	content := []byte("the cached archive")
 
@@ -103,6 +104,7 @@ func TestCacheServiceV2RoundTrip(t *testing.T) {
 }
 
 func TestCacheServiceV2BlockUpload(t *testing.T) {
+	t.Parallel()
 	handler := newTestHandler(t, Policy{})
 
 	created := v2Call(t, handler, testClient, "CreateCacheEntry", map[string]any{"key": "blocks", "version": "v1"})
@@ -150,6 +152,7 @@ func TestCacheServiceV2BlockUpload(t *testing.T) {
 }
 
 func TestCacheServiceV2Lookups(t *testing.T) {
+	t.Parallel()
 	handler := newTestHandler(t, Policy{})
 	saved, _ := saveV2(t, handler, "deps-abc", "v1", []byte("x"))
 	require.Equal(t, true, saved["ok"])
@@ -228,9 +231,28 @@ func TestCacheServiceV2Lookups(t *testing.T) {
 		got := v2Call(t, handler, otherClient, "GetCacheEntryDownloadURL", map[string]any{"key": "deps-abc", "version": "v1"})
 		assert.Equal(t, false, got["ok"])
 	})
+
+	t.Run("each ref saves and restores its own entry for the same key", func(t *testing.T) {
+		clients := map[string]*http.Client{"x": testClient}
+		for _, ref := range []string{"y", "z"} {
+			handler.RegisterJob(ref, JobCredential{Repo: testRepo, Ref: "refs/pull/" + ref + "/head"})
+			clients[ref] = &http.Client{Transport: &bearerTransport{token: ref}}
+			created := v2Call(t, handler, clients[ref], "CreateCacheEntry", map[string]any{"key": "deps-abc", "version": "v1"})
+			require.Equal(t, true, created["ok"])
+			require.Equal(t, http.StatusCreated, putBlob(t, created["signed_upload_url"].(string), []byte(ref)))
+		}
+		for _, ref := range []string{"y", "z"} {
+			require.Equal(t, true, v2Call(t, handler, clients[ref], "FinalizeCacheEntryUpload", map[string]any{"key": "deps-abc", "version": "v1", "size_bytes": 1})["ok"])
+		}
+		for content, client := range clients {
+			got := v2Call(t, handler, client, "GetCacheEntryDownloadURL", map[string]any{"key": "deps-abc", "version": "v1"})
+			assert.Equal(t, content, string(getURL(t, got["signed_download_url"].(string))))
+		}
+	})
 }
 
 func TestCacheServiceV2RefusalsAreTwirp(t *testing.T) {
+	t.Parallel()
 	handler := newTestHandler(t, Policy{})
 
 	tests := []struct {
@@ -267,6 +289,7 @@ func TestCacheServiceV2RefusalsAreTwirp(t *testing.T) {
 }
 
 func TestCacheServiceV2ConcurrentSaveAndRetries(t *testing.T) {
+	t.Parallel()
 	handler := newTestHandler(t, Policy{})
 	content := []byte("the cached archive")
 

act/artifactcache/model.go

@@ -4,6 +4,8 @@
 
 package artifactcache
 
+import "time"
+
 type Request struct {
 	Key     string `json:"key" `
 	Version string `json:"version"`
@@ -13,6 +15,7 @@ type Request struct {
 type Cache struct {
 	ID             uint64 `json:"id"`
 	Repo           string `json:"repo"`
+	Ref            string `json:"ref,omitzero"`
 	Key            string `json:"key"`
 	Version        string `json:"version"`
 	Size           int64  `json:"cacheSize"`
@@ -20,4 +23,9 @@ type Cache struct {
 	Complete       bool   `json:"complete"`
 	UsedAt         int64  `json:"usedAt"`
 	CreatedAt      int64  `json:"createdAt"`
+	CreatedAtNanos int    `json:"createdAtNanos,omitzero"`
+}
+
+func (cache *Cache) createdAtNano() int64 {
+	return cache.CreatedAt*int64(time.Second) + int64(cache.CreatedAtNanos)
 }

act/artifactcache/results_test.go

@@ -17,6 +17,7 @@ import (
 )
 
 func TestFrontResultsService(t *testing.T) {
+	t.Parallel()
 	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")

act/artifactcache/s3_storage_test.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"net/http"
 	"net/http/httptest"
+	"strconv"
 	"strings"
 	"sync"
 	"sync/atomic"
@@ -19,6 +20,7 @@ import (
 )
 
 func TestHandler_S3SharesEntriesAndDropsExpiredOnes(t *testing.T) {
+	t.Parallel()
 	var objects sync.Map
 	var uploads atomic.Int32
 	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
@@ -51,14 +53,33 @@ func TestHandler_S3SharesEntriesAndDropsExpiredOnes(t *testing.T) {
 	saver, restorer := startTestHandler(t, opts), startTestHandler(t, opts)
 	entry := map[string]any{"key": "deps", "version": "v1"}
 
+	started := time.Now()
 	finalized, _ := saveV2(t, saver, "deps", "v1", []byte("archive"))
 	require.Equal(t, true, finalized["ok"])
 	assert.EqualValues(t, 1, uploads.Load())
+	id, err := strconv.ParseUint(finalized["entry_id"].(string), 10, 64)
+	require.NoError(t, err)
+	cache, err := saver.s3.fetch(t.Context(), id)
+	require.NoError(t, err)
+	assert.WithinRange(t, time.Unix(cache.CreatedAt, int64(cache.CreatedAtNanos)), started, time.Now())
 	require.NoError(t, restorer.importMetadata(t.Context()))
 	found := v2Call(t, restorer, testClient, "GetCacheEntryDownloadURL", entry)
 	require.Equal(t, true, found["ok"])
 	assert.Equal(t, []byte("archive"), getURL(t, found["signed_download_url"].(string)))
 
+	for _, cache := range []*Cache{
+		{ID: 2, Repo: testRepo, Key: "deps-old", Version: "v1", Complete: true, CreatedAt: cache.CreatedAt},
+		{ID: 1, Repo: testRepo, Key: "deps-new", Version: "v1", Complete: true, CreatedAt: cache.CreatedAt, CreatedAtNanos: 1},
+	} {
+		require.NoError(t, saver.s3.publish(cache))
+		objects.Store(saver.s3.blobKey(cache.ID), []byte(cache.Key))
+	}
+	require.NoError(t, restorer.importMetadata(t.Context()))
+	found = v2Call(t, restorer, testClient, "GetCacheEntryDownloadURL", map[string]any{"key": "deps-", "version": "v1"})
+	require.Equal(t, true, found["ok"])
+	assert.Equal(t, "deps-new", found["matched_key"])
+	assert.Equal(t, []byte("deps-new"), getURL(t, found["signed_download_url"].(string)))
+
 	objects.Clear()
 	require.NoError(t, restorer.importMetadata(t.Context()))
 	assert.Equal(t, true, v2Call(t, restorer, testClient, "CreateCacheEntry", entry)["ok"])
@@ -71,6 +92,7 @@ func (transport s3TestTransport) RoundTrip(req *http.Request) (*http.Response, e
 }
 
 func TestS3StorageUploadsLargeObjectsInPartsAndAbortsOnFailure(t *testing.T) {
+	t.Parallel()
 	for _, failingPart := range []string{"", "2"} {
 		var parts, aborts atomic.Int32
 		var completion struct {
@@ -110,6 +132,7 @@ func TestS3StorageUploadsLargeObjectsInPartsAndAbortsOnFailure(t *testing.T) {
 }
 
 func TestS3StorageSignsLikeAWS(t *testing.T) {
+	t.Parallel()
 	storage := &s3Storage{opts: S3Options{Region: "us-east-1", AccessKey: "AKIAIOSFODNN7EXAMPLE", SecretKey: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY"}}
 	req := httptest.NewRequest(http.MethodGet, "https://examplebucket.s3.amazonaws.com/?max-keys=2&prefix=J", nil)
 	req.Header.Set("X-Amz-Content-Sha256", "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855")

act/artifacts/server.go

@@ -13,6 +13,7 @@ import (
 	"io/fs"
 	"net"
 	"net/http"
+	"net/url"
 	"os"
 	"path/filepath"
 	"strings"
@@ -84,16 +85,25 @@ func uploads(router *http.ServeMux, baseDir string) {
 		if err := os.MkdirAll(filepath.Dir(safePath), os.ModePerm); err != nil {
 			panic(err)
 		}
-		flags := os.O_CREATE | os.O_WRONLY | os.O_TRUNC
-		appendUpload := req.Header.Get("Content-Range")
-		if appendUpload != "" && !strings.HasPrefix(appendUpload, "bytes 0-") {
-			flags = os.O_CREATE | os.O_WRONLY | os.O_APPEND
+		var start int64
+		if contentRange := req.Header.Get("Content-Range"); contentRange != "" {
+			if _, err := fmt.Sscanf(contentRange, "bytes %d-", &start); err != nil || start < 0 {
+				http.Error(w, "invalid content range", http.StatusBadRequest)
+				return
+			}
+		}
+		flags := os.O_CREATE | os.O_WRONLY
+		if start == 0 {
+			flags |= os.O_TRUNC
 		}
 		file, err := os.OpenFile(safePath, flags, 0o644)
 		if err != nil {
 			panic(err)
 		}
 		defer file.Close()
+		if _, err := file.Seek(start, io.SeekStart); err != nil {
+			panic(err)
+		}
 		if req.Body == nil {
 			panic(errors.New("no body given"))
 		}
@@ -120,7 +130,7 @@ func downloads(router *http.ServeMux, baseDir string) {
 		runID := req.PathValue("runId")
 
 		entries, err := os.ReadDir(safeResolve(baseDir, runID))
-		if err != nil {
+		if err != nil && !os.IsNotExist(err) {
 			panic(err)
 		}
 
@@ -145,6 +155,9 @@ func downloads(router *http.ServeMux, baseDir string) {
 
 		var files []ContainerItem
 		err := filepath.WalkDir(safePath, func(path string, entry fs.DirEntry, err error) error {
+			if err != nil {
+				return err
+			}
 			if !entry.IsDir() {
 				rel, err := filepath.Rel(safePath, path)
 				if err != nil {
@@ -157,12 +170,12 @@ func downloads(router *http.ServeMux, baseDir string) {
 				files = append(files, ContainerItem{
 					Path:            filepath.ToSlash(filepath.Join(itemPath, rel)),
 					ItemType:        "file",
-					ContentLocation: fmt.Sprintf("http://%s/artifact/%s/%s/%s", req.Host, container, itemPath, filepath.ToSlash(rel)),
+					ContentLocation: (&url.URL{Scheme: "http", Host: req.Host, Path: "/artifact/" + container + "/" + itemPath + "/" + filepath.ToSlash(rel)}).String(),
 				})
 			}
 			return nil
 		})
-		if err != nil {
+		if err != nil && !os.IsNotExist(err) {
 			panic(err)
 		}
 
@@ -205,6 +218,7 @@ func Serve(ctx context.Context, artifactPath, addr, port string) context.CancelF
 	logger.Debugf("Artifacts base path '%s'", artifactPath)
 	uploads(router, artifactPath)
 	downloads(router, artifactPath)
+	registerV4(router, artifactPath)
 
 	server := &http.Server{
 		Addr:              net.JoinHostPort(addr, port),

act/artifacts/server_test.go

@@ -28,6 +28,7 @@ func TestArtifactFlow(t *testing.T) {
 	router := http.NewServeMux()
 	uploads(router, artifactPath)
 	downloads(router, artifactPath)
+	registerV4(router, artifactPath)
 	server := httptest.NewServer(router)
 	defer server.Close()
 
@@ -48,8 +49,17 @@ func TestArtifactFlow(t *testing.T) {
 		return resp.StatusCode, data
 	}
 
+	t.Run("no-artifacts", func(t *testing.T) {
+		status, data := request(t, http.MethodGet, baseURL+"/_apis/pipelines/workflows/0/artifacts", nil, nil)
+		require.Equal(t, http.StatusOK, status, string(data))
+		var list NamedFileContainerResourceURLResponse
+		require.NoError(t, json.Unmarshal(data, &list))
+		require.Zero(t, list.Count)
+		require.Empty(t, list.Value)
+	})
+
 	t.Run("upload-and-download", func(t *testing.T) {
-		const runID, item, content = "1", "my-artifact/data.txt", "hello artifact\n"
+		const runID, item, content = "1", "my-artifact/data#1%25.txt", "hello artifact\n"
 
 		status, data := request(t, http.MethodPost, baseURL+"/_apis/pipelines/workflows/"+runID+"/artifacts", nil, nil)
 		require.Equal(t, http.StatusOK, status, string(data))
@@ -81,13 +91,13 @@ func TestArtifactFlow(t *testing.T) {
 		require.NoError(t, json.Unmarshal(data, &items))
 		require.Len(t, items.Value, 1)
 		require.Equal(t, "file", items.Value[0].ItemType)
-		require.Equal(t, "my-artifact/data.txt", items.Value[0].Path)
+		require.Equal(t, item, items.Value[0].Path)
 
 		status, data = request(t, http.MethodGet, items.Value[0].ContentLocation, nil, nil)
 		require.Equal(t, http.StatusOK, status)
 		require.Equal(t, content, string(data))
 
-		stored, err := os.ReadFile(filepath.Join(artifactPath, runID, "my-artifact", "data.txt"))
+		stored, err := os.ReadFile(filepath.Join(artifactPath, runID, filepath.FromSlash(item)))
 		require.NoError(t, err)
 		require.Equal(t, content, string(stored))
 	})
@@ -95,16 +105,25 @@ func TestArtifactFlow(t *testing.T) {
 	t.Run("content-range", func(t *testing.T) {
 		const rawURL = "/upload/4?itemPath=chunks.txt"
 		status, data := request(t, http.MethodPut, baseURL+rawURL, strings.NewReader("first"),
-			http.Header{"Content-Range": []string{"bytes 0-4/11"}})
+			http.Header{"Content-Range": []string{"bytes 0-4/12"}})
 		require.Equal(t, http.StatusOK, status, string(data))
 
+		for range 2 {
 			status, data = request(t, http.MethodPut, baseURL+rawURL, strings.NewReader("-second"),
-			http.Header{"Content-Range": []string{"bytes 5-11/11"}})
+				http.Header{"Content-Range": []string{"bytes 5-11/12"}})
 			require.Equal(t, http.StatusOK, status, string(data))
+		}
 
 		stored, err := os.ReadFile(filepath.Join(artifactPath, "4", "chunks.txt"))
 		require.NoError(t, err)
 		require.Equal(t, "first-second", string(stored))
+
+		status, data = request(t, http.MethodPut, baseURL+rawURL, strings.NewReader(""),
+			http.Header{"Content-Range": []string{"bytes 0--1/0"}})
+		require.Equal(t, http.StatusOK, status, string(data))
+		stored, err = os.ReadFile(filepath.Join(artifactPath, "4", "chunks.txt"))
+		require.NoError(t, err)
+		require.Empty(t, stored)
 	})
 
 	t.Run("gzip-roundtrip", func(t *testing.T) {
@@ -156,6 +175,92 @@ func TestArtifactFlow(t *testing.T) {
 		require.Equal(t, http.StatusOK, status)
 		require.Equal(t, content, string(data))
 	})
+
+	t.Run("v4", func(t *testing.T) {
+		const artifactRef = `"workflow_run_backend_id":"7","name":"50% art@v1"`
+		call := func(method, body string) (int, map[string]any) {
+			status, data := request(t, http.MethodPost, baseURL+artifactServicePath+"/"+method, strings.NewReader(body), nil)
+			var out map[string]any
+			require.NoError(t, json.Unmarshal(data, &out), string(data))
+			return status, out
+		}
+		put := func(rawURL, body string) int {
+			status, _ := request(t, http.MethodPut, rawURL, strings.NewReader(body), nil)
+			return status
+		}
+		download := func() (string, http.Header) {
+			status, signed := call("GetSignedArtifactURL", "{"+artifactRef+"}")
+			require.Equal(t, http.StatusOK, status, signed)
+			resp, err := client.Get(signed["signedUrl"].(string))
+			require.NoError(t, err)
+			defer resp.Body.Close()
+			require.Equal(t, http.StatusOK, resp.StatusCode)
+			body, err := io.ReadAll(resp.Body)
+			require.NoError(t, err)
+			return string(body), resp.Header
+		}
+
+		status, created := call("CreateArtifact", "{"+artifactRef+"}")
+		require.Equal(t, http.StatusOK, status, created)
+		upload := created["signedUploadUrl"].(string)
+		require.Equal(t, http.StatusCreated, put(upload+"&comp=block&blockid=Yg", "world"))
+		require.Equal(t, http.StatusCreated, put(upload+"&comp=block&blockid=YQ", "hello "))
+		for attempt := range 2 {
+			require.Equal(t, http.StatusCreated, put(upload+"&comp=blocklist", `<?xml version="1.0"?><BlockList><Latest>YQ</Latest><Latest>Yg</Latest></BlockList>`), "commit attempt %d", attempt)
+		}
+
+		digest := "sha256:b94d27b9934d3e08a52e52d7da7dabfac484efe37a5380ee9088f7ace2efcde9"
+		status, _ = call("FinalizeArtifact", "{"+artifactRef+`,"size":"10"}`)
+		require.Equal(t, http.StatusBadRequest, status)
+		status, finalized := call("FinalizeArtifact", "{"+artifactRef+`,"size":"11","hash":"`+digest+`"}`)
+		require.Equal(t, http.StatusOK, status, finalized)
+		id := finalized["artifactId"].(string)
+		status, refinalized := call("FinalizeArtifact", "{"+artifactRef+`,"size":"10"}`)
+		require.Equal(t, http.StatusOK, status)
+		require.Equal(t, id, refinalized["artifactId"])
+		status, conflict := call("CreateArtifact", "{"+artifactRef+"}")
+		require.Equal(t, http.StatusConflict, status)
+		require.Equal(t, "already_exists", conflict["code"])
+
+		status, listed := call("ListArtifacts", `{"workflowRunBackendId":"7","idFilter":"`+id+`"}`)
+		require.Equal(t, http.StatusOK, status)
+		artifacts := listed["artifacts"].([]any)
+		require.Len(t, artifacts, 1)
+		artifact := artifacts[0].(map[string]any)
+		require.Regexp(t, `^\d{4}-\d\d-\d\dT\d\d:\d\d:\d\dZ$`, artifact["createdAt"])
+		delete(artifact, "createdAt")
+		require.Equal(t, map[string]any{"workflowRunBackendId": "7", "name": "50% art@v1", "size": "11", "digest": digest, "databaseId": id}, artifact)
+		_, other := call("ListArtifacts", `{"workflow_run_backend_id":"8"}`)
+		require.Empty(t, other["artifacts"])
+
+		body, header := download()
+		require.Equal(t, "hello world", body)
+		require.Equal(t, "application/zip", header.Get("Content-Type"))
+		require.Equal(t, "attachment; filename*=UTF-8''50%25%20art%40v1.zip", header.Get("Content-Disposition"))
+
+		status, _ = call("DeleteArtifact", "{"+artifactRef+"}")
+		require.Equal(t, http.StatusOK, status)
+		status, missing := call("GetSignedArtifactURL", "{"+artifactRef+"}")
+		require.Equal(t, http.StatusNotFound, status)
+		require.Equal(t, "not_found", missing["code"])
+
+		status, created = call("CreateArtifact", "{"+artifactRef+"}")
+		require.Equal(t, http.StatusOK, status, created)
+		status, _ = request(t, http.MethodPut, created["signedUploadUrl"].(string), strings.NewReader("gz"), http.Header{"X-Ms-Blob-Content-Type": {"application/gzip"}})
+		require.Equal(t, http.StatusCreated, status)
+		status, _ = call("FinalizeArtifact", "{"+artifactRef+`,"size":"2"}`)
+		require.Equal(t, http.StatusOK, status)
+		body, header = download()
+		require.Equal(t, "gz", body)
+		require.Equal(t, "application/gzip", header.Get("Content-Type"))
+		require.Equal(t, "attachment; filename*=UTF-8''50%25%20art%40v1", header.Get("Content-Disposition"))
+
+		rerun := http.NewServeMux()
+		registerV4(rerun, artifactPath)
+		recorder := httptest.NewRecorder()
+		rerun.ServeHTTP(recorder, httptest.NewRequest(http.MethodPost, artifactServicePath+"/CreateArtifact", strings.NewReader("{"+artifactRef+"}")))
+		require.Equal(t, http.StatusOK, recorder.Code, recorder.Body.String())
+	})
 }
 
 func TestSafeResolve(t *testing.T) {

act/artifacts/server_v4.go

@@ -0,0 +1,329 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifacts
+
+import (
+	"cmp"
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json/v2"
+	"encoding/xml"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"net/http"
+	"net/url"
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"sync"
+	"time"
+)
+
+type serverV4 struct { // the subset of Gitea's v4 artifact service that the toolkit uses, unauthenticated like the v3 routes beside it
+	dir     string
+	mu      sync.Mutex
+	lastID  int64
+	created map[string]bool
+}
+
+const (
+	zipMimeType   = "application/zip"
+	maxTwirpBytes = 1 << 20
+)
+
+func registerV4(router *http.ServeMux, baseDir string) {
+	server := &serverV4{dir: filepath.Join(baseDir, "v4"), lastID: time.Now().UnixMilli(), created: map[string]bool{}} // IDs stay unique across invocations sharing baseDir
+	router.HandleFunc("POST "+artifactServicePath+"/CreateArtifact", twirp(server.createArtifact))
+	router.HandleFunc("POST "+artifactServicePath+"/FinalizeArtifact", twirp(server.finalizeArtifact))
+	router.HandleFunc("POST "+artifactServicePath+"/ListArtifacts", twirp(server.listArtifacts))
+	router.HandleFunc("POST "+artifactServicePath+"/GetSignedArtifactURL", twirp(server.getSignedArtifactURL))
+	router.HandleFunc("POST "+artifactServicePath+"/DeleteArtifact", twirp(server.deleteArtifact))
+	router.HandleFunc("PUT "+artifactServicePath+"/UploadArtifact", server.uploadArtifact)
+	router.HandleFunc("GET "+artifactServicePath+"/DownloadArtifact", server.downloadArtifact)
+}
+
+type twirpError struct {
+	status int
+	code   string
+	msg    string
+}
+
+func (e *twirpError) Error() string { return e.msg }
+
+func notFound(name string) error {
+	return &twirpError{http.StatusNotFound, "not_found", fmt.Sprintf("artifact %q not found", name)}
+}
+
+func badRequest(format string, args ...any) error {
+	return &twirpError{http.StatusBadRequest, "invalid_argument", fmt.Sprintf(format, args...)}
+}
+
+func twirp[Req, Resp any](handle func(*http.Request, Req) (Resp, error)) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		var req Req
+		var resp Resp
+		err := json.UnmarshalRead(io.LimitReader(r.Body, maxTwirpBytes), &req, json.MatchCaseInsensitiveNames(true)) // also matches the toolkit's proto field names
+		if err != nil {
+			err = badRequest("malformed request: %v", err)
+		} else {
+			resp, err = handle(r, req)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		if err != nil {
+			terr := &twirpError{http.StatusInternalServerError, "internal", err.Error()}
+			errors.As(err, &terr)
+			w.WriteHeader(terr.status)
+			_ = json.MarshalWrite(w, map[string]string{"code": terr.code, "msg": terr.msg})
+			return
+		}
+		_ = json.MarshalWrite(w, resp)
+	}
+}
+
+func fileName(name string) string { // safe on any filesystem for any client-chosen name, case-insensitive ones included
+	sum := sha256.Sum256([]byte(name))
+	return hex.EncodeToString(sum[:])
+}
+
+func (s *serverV4) runDir(runID string) string {
+	return filepath.Join(s.dir, fileName(runID))
+}
+
+func (s *serverV4) artifactDir(runID, name string) string {
+	return filepath.Join(s.runDir(runID), fileName(name))
+}
+
+func (s *serverV4) urlDir(r *http.Request) string {
+	return s.artifactDir(r.URL.Query().Get("run"), r.URL.Query().Get("name"))
+}
+
+func readMeta(dir string) (*monolithArtifact, error) {
+	data, err := os.ReadFile(filepath.Join(dir, "meta.json"))
+	if err != nil {
+		return nil, err
+	}
+	var meta monolithArtifact
+	return &meta, json.Unmarshal(data, &meta)
+}
+
+func writeMeta(dir string, meta *monolithArtifact) error {
+	data, err := json.Marshal(meta)
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(filepath.Join(dir, "meta.json"), data, 0o644)
+}
+
+func blobURL(r *http.Request, endpoint, runID, name string) string {
+	return "http://" + r.Host + artifactServicePath + "/" + endpoint + "?" + url.Values{"run": {runID}, "name": {name}}.Encode()
+}
+
+func (s *serverV4) createArtifact(r *http.Request, req createArtifactRequest) (createArtifactResponse, error) { // an exec invocation re-runs its workflows, so it replaces earlier invocations' artifacts but conflicts with its own
+	dir := s.artifactDir(req.WorkflowRunBackendID, req.Name)
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.created[dir] {
+		return createArtifactResponse{}, &twirpError{http.StatusConflict, "already_exists", "an artifact with this name already exists on the workflow run"}
+	}
+	if err := os.RemoveAll(dir); err != nil {
+		return createArtifactResponse{}, err
+	}
+	if err := os.MkdirAll(filepath.Join(dir, "blocks"), 0o755); err != nil {
+		return createArtifactResponse{}, err
+	}
+	s.lastID++
+	if err := writeMeta(dir, &monolithArtifact{
+		WorkflowRunBackendID: req.WorkflowRunBackendID,
+		DatabaseID:           s.lastID,
+		Name:                 req.Name,
+		CreatedAt:            time.Now().UTC().Truncate(time.Second),
+	}); err != nil {
+		return createArtifactResponse{}, err
+	}
+	s.created[dir] = true
+	return createArtifactResponse{Ok: true, SignedUploadURL: blobURL(r, "UploadArtifact", req.WorkflowRunBackendID, req.Name)}, nil
+}
+
+type blockList struct {
+	Blocks []string `xml:",any"` // Latest, Committed and Uncommitted all name a staged block
+}
+
+func (s *serverV4) uploadArtifact(w http.ResponseWriter, r *http.Request) {
+	dir := s.urlDir(r)
+	if meta, err := readMeta(dir); err != nil || meta.Digest != "" {
+		http.Error(w, "no artifact awaits an upload", http.StatusNotFound)
+		return
+	}
+	contentType := r.Header.Get("x-ms-blob-content-type")
+	var err error
+	switch comp := r.URL.Query().Get("comp"); comp {
+	case "block":
+		err = writeFile(filepath.Join(dir, "blocks", fileName(r.URL.Query().Get("blockid"))), r.Body)
+	case "blocklist":
+		err = commitBlob(dir, contentType, func(out io.Writer) error { return joinBlocks(out, dir, r.Body) })
+	case "": // Put Blob, which docker/build-push-action uses
+		err = commitBlob(dir, contentType, func(out io.Writer) error {
+			_, err := io.Copy(out, r.Body)
+			return err
+		})
+	default:
+		http.Error(w, "unsupported comp "+comp, http.StatusBadRequest)
+		return
+	}
+	if err != nil {
+		http.Error(w, err.Error(), http.StatusInternalServerError)
+		return
+	}
+	w.WriteHeader(http.StatusCreated)
+}
+
+func writeFile(name string, body io.Reader) error {
+	file, err := os.Create(name)
+	if err != nil {
+		return err
+	}
+	_, err = io.Copy(file, body)
+	return errors.Join(err, file.Close())
+}
+
+func commitBlob(dir, contentType string, write func(io.Writer) error) error { // replaces the blob whole and keeps its content type, as Azure does
+	out, err := os.CreateTemp(dir, "blob")
+	if err != nil {
+		return err
+	}
+	if err = errors.Join(write(out), out.Close()); err == nil {
+		err = os.Rename(out.Name(), filepath.Join(dir, "blob"))
+	}
+	if err != nil {
+		return errors.Join(err, os.Remove(out.Name()))
+	}
+	return os.WriteFile(filepath.Join(dir, "content-type"), []byte(contentType), 0o644)
+}
+
+func joinBlocks(out io.Writer, dir string, body io.Reader) error { // keeps the blocks for finalizeArtifact, so that a retried commit can join them again
+	var list blockList
+	if err := xml.NewDecoder(io.LimitReader(body, maxTwirpBytes)).Decode(&list); err != nil {
+		return err
+	}
+	for _, id := range list.Blocks {
+		if err := appendBlock(out, filepath.Join(dir, "blocks", fileName(id))); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func appendBlock(out io.Writer, name string) error {
+	in, err := os.Open(name)
+	if err != nil {
+		return err
+	}
+	defer in.Close()
+	_, err = io.Copy(out, in)
+	return err
+}
+
+func (s *serverV4) finalizeArtifact(_ *http.Request, req finalizeArtifactRequest) (finalizeArtifactResponse, error) {
+	dir := s.artifactDir(req.WorkflowRunBackendID, req.Name)
+	meta, err := readMeta(dir)
+	if err != nil {
+		return finalizeArtifactResponse{}, notFound(req.Name)
+	} else if meta.Digest != "" {
+		return finalizeArtifactResponse{Ok: true, ArtifactID: meta.DatabaseID}, nil
+	}
+	blob, err := os.Open(filepath.Join(dir, "blob"))
+	if err != nil {
+		return finalizeArtifactResponse{}, badRequest("artifact %q has no upload", req.Name)
+	}
+	defer blob.Close()
+	hash := sha256.New()
+	size, err := io.Copy(hash, blob)
+	if err != nil {
+		return finalizeArtifactResponse{}, err
+	}
+	digest := "sha256:" + hex.EncodeToString(hash.Sum(nil))
+	if size != req.Size {
+		return finalizeArtifactResponse{}, badRequest("size %d does not match the %d bytes uploaded", req.Size, size)
+	}
+	if req.Hash != "" && req.Hash != digest {
+		return finalizeArtifactResponse{}, badRequest("hash %s does not match the uploaded %s", req.Hash, digest)
+	}
+	s.mu.Lock() // only after hashing, so finalizing one large artifact does not stall the others
+	defer s.mu.Unlock()
+	if current, err := readMeta(dir); err != nil || current.DatabaseID != meta.DatabaseID {
+		return finalizeArtifactResponse{}, notFound(req.Name)
+	}
+	meta.Size, meta.Digest = size, digest
+	if err := writeMeta(dir, meta); err != nil {
+		return finalizeArtifactResponse{}, err
+	}
+	return finalizeArtifactResponse{Ok: true, ArtifactID: meta.DatabaseID}, os.RemoveAll(filepath.Join(dir, "blocks"))
+}
+
+func (s *serverV4) listArtifacts(_ *http.Request, req listArtifactsRequest) (listArtifactsResponse, error) {
+	run := s.runDir(req.WorkflowRunBackendID)
+	entries, err := os.ReadDir(run)
+	if err != nil && !errors.Is(err, fs.ErrNotExist) {
+		return listArtifactsResponse{}, err
+	}
+	var resp listArtifactsResponse
+	for _, entry := range entries {
+		meta, err := readMeta(filepath.Join(run, entry.Name()))
+		if err != nil || meta.Digest == "" ||
+			req.NameFilter != "" && req.NameFilter != meta.Name || req.IDFilter != 0 && req.IDFilter != meta.DatabaseID {
+			continue
+		}
+		resp.Artifacts = append(resp.Artifacts, *meta)
+	}
+	slices.SortFunc(resp.Artifacts, func(left, right monolithArtifact) int { return cmp.Compare(left.DatabaseID, right.DatabaseID) })
+	return resp, nil
+}
+
+func (s *serverV4) getSignedArtifactURL(r *http.Request, req getSignedArtifactURLRequest) (getSignedArtifactURLResponse, error) {
+	if meta, err := readMeta(s.artifactDir(req.WorkflowRunBackendID, req.Name)); err != nil || meta.Digest == "" {
+		return getSignedArtifactURLResponse{}, notFound(req.Name)
+	}
+	return getSignedArtifactURLResponse{SignedURL: blobURL(r, "DownloadArtifact", req.WorkflowRunBackendID, req.Name)}, nil
+}
+
+func (s *serverV4) downloadArtifact(w http.ResponseWriter, r *http.Request) {
+	dir := s.urlDir(r)
+	meta, err := readMeta(dir)
+	if err != nil || meta.Digest == "" {
+		http.Error(w, "artifact not found", http.StatusNotFound)
+		return
+	}
+	blob, err := os.Open(filepath.Join(dir, "blob"))
+	if err != nil {
+		http.Error(w, err.Error(), http.StatusInternalServerError)
+		return
+	}
+	defer blob.Close()
+	data, _ := os.ReadFile(filepath.Join(dir, "content-type"))
+	contentType, filename := string(data), meta.Name
+	if !strings.Contains(contentType, "/") {
+		contentType = zipMimeType
+	}
+	if contentType == zipMimeType && !strings.HasSuffix(filename, ".zip") {
+		filename += ".zip"
+	}
+	w.Header().Set("Content-Type", contentType)
+	w.Header().Set("Content-Disposition", "attachment; filename*=UTF-8''"+strings.ReplaceAll(url.QueryEscape(filename), "+", "%20")) // RFC 5987 attr-chars only, which the toolkit and mime.ParseMediaType both accept
+	http.ServeContent(w, r, "", meta.CreatedAt, blob)
+}
+
+func (s *serverV4) deleteArtifact(_ *http.Request, req deleteArtifactRequest) (deleteArtifactResponse, error) {
+	dir := s.artifactDir(req.WorkflowRunBackendID, req.Name)
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	meta, err := readMeta(dir)
+	if err != nil {
+		return deleteArtifactResponse{}, notFound(req.Name)
+	}
+	delete(s.created, dir)
+	return deleteArtifactResponse{Ok: true, ArtifactID: meta.DatabaseID}, os.RemoveAll(dir)
+}

act/artifacts/token.go

@@ -0,0 +1,23 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifacts
+
+import (
+	"crypto/rand"
+	"encoding/base64"
+	"encoding/json/v2"
+	"time"
+)
+
+func LocalRuntimeToken(run string, lifetime time.Duration) string { // an unsecured JWT with the claims the toolkit and buildx read, its run hashed for scp and its jti an unguessable bearer
+	now := time.Now()
+	claims, _ := json.Marshal(map[string]any{
+		"scp": "Actions.Results:" + fileName(run) + ":1",
+		"ac":  `[{"Scope":"","Permission":3}]`,
+		"nbf": now.Unix(),
+		"exp": now.Add(lifetime).Unix(),
+		"jti": rand.Text(),
+	})
+	return base64.RawURLEncoding.EncodeToString([]byte(`{"alg":"none","typ":"JWT"}`)) + "." + base64.RawURLEncoding.EncodeToString(claims) + "."
+}

act/artifacts/twirp.go

@@ -0,0 +1,55 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifacts
+
+import "time"
+
+const artifactServicePath = "/twirp/github.actions.results.api.v1.ArtifactService"
+
+type ( // the ArtifactService messages in the protobuf JSON mapping, which writes int64 as strings
+	createArtifactResponse struct {
+		Ok              bool   `json:"ok"`
+		SignedUploadURL string `json:"signedUploadUrl"`
+	}
+
+	finalizeArtifactRequest struct {
+		WorkflowRunBackendID string
+		Name                 string
+		Size                 int64 `json:",string"`
+		Hash                 string
+	}
+	finalizeArtifactResponse struct {
+		Ok         bool  `json:"ok"`
+		ArtifactID int64 `json:"artifactId,string"`
+	}
+
+	listArtifactsRequest struct {
+		WorkflowRunBackendID string
+		NameFilter           string
+		IDFilter             int64 `json:",string"`
+	}
+	listArtifactsResponse struct {
+		Artifacts []monolithArtifact `json:"artifacts"`
+	}
+	monolithArtifact struct {
+		WorkflowRunBackendID string    `json:"workflowRunBackendId"` // the toolkit sends it back to GetSignedArtifactURL
+		DatabaseID           int64     `json:"databaseId,string"`
+		Name                 string    `json:"name"`
+		Size                 int64     `json:"size,string"`
+		CreatedAt            time.Time `json:"createdAt"` // whole seconds, the toolkit rejects the trimmed fractions json writes
+		Digest               string    `json:"digest"`    // empty until the upload is finalized
+	}
+
+	getSignedArtifactURLRequest struct {
+		WorkflowRunBackendID string
+		Name                 string
+	}
+	getSignedArtifactURLResponse struct {
+		SignedURL string `json:"signedUrl"`
+	}
+
+	createArtifactRequest  = getSignedArtifactURLRequest
+	deleteArtifactRequest  = getSignedArtifactURLRequest
+	deleteArtifactResponse = finalizeArtifactResponse
+)

act/common/executor.go

@@ -164,7 +164,10 @@ func (e Executor) Finally(finally Executor) Executor {
 		err := e(ctx)
 		err2 := finally(ctx)
 		if err2 != nil {
-			return fmt.Errorf("error occurred running finally: %v (original error: %v)", err2, err)
+			if err != nil {
+				return fmt.Errorf("error occurred running finally: %w (original error: %w)", err2, err)
+			}
+			return fmt.Errorf("error occurred running finally: %w (original error: %v)", err2, err)
 		}
 		return err
 	}

act/common/executor_test.go

@@ -7,6 +7,7 @@ package common
 import (
 	"context"
 	"errors"
+	"fmt"
 	"sync"
 	"sync/atomic"
 	"testing"
@@ -134,9 +135,13 @@ func TestExecutorConditionalsAndFinally(t *testing.T) {
 }
 
 func TestExecutorFinallyReturnsFinallyErrorWithOriginal(t *testing.T) {
-	mainErr := errors.New("main failed")
 	finalErr := errors.New("cleanup failed")
-
+	for _, mainErr := range []error{context.Canceled, nil} {
 		err := NewErrorExecutor(mainErr).Finally(NewErrorExecutor(finalErr))(t.Context())
-	require.EqualError(t, err, "error occurred running finally: cleanup failed (original error: main failed)")
+		require.EqualError(t, err, fmt.Sprintf("error occurred running finally: cleanup failed (original error: %v)", mainErr))
+		require.ErrorIs(t, err, finalErr)
+		if mainErr != nil {
+			require.ErrorIs(t, err, mainErr)
+		}
+	}
 }

act/common/git/fetch_action.go

@@ -8,6 +8,7 @@ import (
 	"bytes"
 	"compress/gzip"
 	"context"
+	"encoding/json"
 	"errors"
 	"io"
 	"math/rand/v2"
@@ -28,6 +29,13 @@ import (
 
 var actionFetches sync.Map
 
+const treeSizes = ".sizes" // beside the tree rather than in it, which is copied into job containers as it is
+
+type treeFile struct {
+	Name []byte // not a string, which JSON would turn into UTF-8, as git allows any bytes but NUL
+	Size int64
+}
+
 type actionFetch struct {
 	done     chan struct{}
 	dir, sha string
@@ -86,9 +94,14 @@ func fetchAction(ctx context.Context, input NewGitCloneExecutorInput) (string, s
 	record := filepath.Join(input.Dir, "ref")
 	data, _ := os.ReadFile(record)
 	sha, etag, _ := strings.Cut(string(data), "\n")
-	if !isFullObjectID(sha) || !exists(tree(sha)) {
+	switch {
+	case !isFullObjectID(sha):
 		sha, etag = "", ""
-	} else if input.OfflineMode || isFullObjectID(input.Ref) {
+	case input.OfflineMode && exists(tree(sha)) && !exists(tree(sha)+treeSizes):
+		return tree(sha), sha, nil // from an older runner, with no sizes to check, so offline mode uses it as before
+	case !treeIntact(input.Dir, sha):
+		sha, etag = "", ""
+	case input.OfflineMode || isFullObjectID(input.Ref):
 		return tree(sha), sha, nil
 	}
 
@@ -173,7 +186,7 @@ func downloadArchiveOnce(ctx context.Context, client *http.Client, archive, toke
 		newEtag := response.Header.Get("ETag")
 		gz, err := gzip.NewReader(response.Body)
 		if err == nil {
-			sha, err = extractTree(gz, dir, !strings.HasSuffix(newEtag, `-false"`)) // Gitea's mark of an unprefixed archive
+			sha, err = extractTree(ctx, gz, dir, !strings.HasSuffix(newEtag, `-false"`)) // Gitea's mark of an unprefixed archive
 		}
 		var netErr net.Error
 		if errors.Is(err, io.ErrUnexpectedEOF) || errors.As(err, &netErr) {
@@ -191,18 +204,18 @@ func exportClone(ctx context.Context, input NewGitCloneExecutorInput) (string, e
 	clone := input
 	clone.Dir = filepath.Join(input.Dir, "git")
 	sha, _, err := resolveClone(ctx, clone)
-	if err != nil || exists(filepath.Join(input.Dir, sha)) {
+	if err != nil || treeIntact(input.Dir, sha) {
 		return sha, err
 	}
 	archive, err := runGit(ctx, clone.Dir, nil, "archive", "--format=tar", sha)
 	if err != nil {
 		return "", err
 	}
-	_, err = extractTree(bytes.NewReader(archive), input.Dir, false)
+	_, err = extractTree(ctx, bytes.NewReader(archive), input.Dir, false)
 	return sha, err
 }
 
-func extractTree(archive io.Reader, dir string, strip bool) (string, error) { // into dir/<commit of the pax header>, strip dropping the directory wrapping the files
+func extractTree(ctx context.Context, archive io.Reader, dir string, strip bool) (string, error) { // into dir/<commit of the pax header>, strip dropping the directory wrapping the files
 	temp, err := os.MkdirTemp(dir, "tmp-")
 	if err != nil {
 		return "", err
@@ -215,6 +228,7 @@ func extractTree(archive io.Reader, dir string, strip bool) (string, error) { //
 	defer root.Close()
 
 	sha := ""
+	sizes := map[string]int64{}
 	reader := tar.NewReader(archive)
 	for {
 		header, err := reader.Next()
@@ -225,7 +239,7 @@ func extractTree(archive io.Reader, dir string, strip bool) (string, error) { //
 			return "", err
 		}
 		if header.Typeflag == tar.TypeXGlobalHeader {
-			if sha = header.PAXRecords["comment"]; isFullObjectID(sha) && exists(filepath.Join(dir, sha)) {
+			if sha = header.PAXRecords["comment"]; isFullObjectID(sha) && treeIntact(dir, sha) {
 				return sha, nil
 			}
 			continue
@@ -241,10 +255,10 @@ func extractTree(archive io.Reader, dir string, strip bool) (string, error) { //
 		case tar.TypeDir:
 			err = root.MkdirAll(name, 0o755)
 		case tar.TypeReg:
-			err = writeTreeFile(root, name, header.Mode, reader)
+			sizes[name], err = writeTreeFile(root, name, header.Mode, reader)
 		case tar.TypeSymlink:
 			if err = root.Symlink(header.Linkname, name); err != nil && runtime.GOOS == "windows" {
-				err = writeTreeFile(root, name, 0, strings.NewReader(header.Linkname))
+				sizes[name], err = writeTreeFile(root, name, 0, strings.NewReader(header.Linkname))
 			}
 		}
 		if err != nil {
@@ -257,20 +271,76 @@ func extractTree(archive io.Reader, dir string, strip bool) (string, error) { //
 	if err := root.Close(); err != nil {
 		return "", err
 	}
-	return sha, os.Rename(temp, filepath.Join(dir, sha))
+	return sha, installTree(ctx, dir, sha, temp, sizes)
 }
 
-func writeTreeFile(root *os.Root, name string, mode int64, content io.Reader) error {
+var lockTree = AcquireCloneLock // replaced in tests to see installTree wait for a job holding the tree
+
+// installTree moves temp into place as dir/sha, replacing a tree that failed treeIntact.
+func installTree(ctx context.Context, dir, sha, temp string, sizes map[string]int64) error {
+	files := make([]treeFile, 0, len(sizes))
+	for name, size := range sizes {
+		files = append(files, treeFile{[]byte(name), size})
+	}
+	data, err := json.Marshal(files)
+	if err != nil {
+		return err
+	}
+	// Written before the swap: it describes the commit, not a copy of it, so it holds for whichever tree is left.
+	if err := os.WriteFile(filepath.Join(dir, sha+treeSizes), data, 0o600); err != nil { // a torn record fails treeIntact, so it needs no rename
+		return err
+	}
+	tree := filepath.Join(dir, sha)
+	if !exists(tree) {
+		return os.Rename(temp, tree)
+	}
+	stale, err := os.MkdirTemp(dir, "tmp-")
+	if err != nil {
+		return err
+	}
+	defer os.RemoveAll(stale)          // after the unlock, as no job reads the old tree once it is moved
+	unlock, err := lockTree(ctx, tree) // jobs copying the old tree hold it
+	if err != nil {
+		return err
+	}
+	defer unlock()
+	if err := os.Rename(tree, filepath.Join(stale, sha)); err != nil {
+		return err
+	}
+	return os.Rename(temp, tree)
+}
+
+// treeIntact reports whether dir/sha still holds each file extracted into it at the size it was written.
+// Only sizes are compared, not content: a stat per file is cheap enough for every job, hashing is not.
+func treeIntact(dir, sha string) bool {
+	data, err := os.ReadFile(filepath.Join(dir, sha+treeSizes))
+	if err != nil || !exists(filepath.Join(dir, sha)) {
+		return false
+	}
+	var files []treeFile
+	if json.Unmarshal(data, &files) != nil {
+		return false
+	}
+	for _, file := range files {
+		info, err := os.Lstat(filepath.Join(dir, sha, filepath.FromSlash(string(file.Name))))
+		if err != nil || !info.Mode().IsRegular() || info.Size() != file.Size {
+			return false
+		}
+	}
+	return true
+}
+
+func writeTreeFile(root *os.Root, name string, mode int64, content io.Reader) (int64, error) {
 	perm := os.FileMode(0o644)
 	if mode&0o111 != 0 {
 		perm = 0o755
 	}
 	file, err := root.OpenFile(name, os.O_CREATE|os.O_EXCL|os.O_WRONLY, perm)
 	if err != nil {
-		return err
+		return 0, err
 	}
-	_, err = io.Copy(file, content)
-	return errors.Join(err, file.Close())
+	n, err := io.Copy(file, content)
+	return n, errors.Join(err, file.Close())
 }
 
 func exists(name string) bool {

act/common/git/fetch_action_test.go

@@ -15,6 +15,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"sync"
 	"sync/atomic"
@@ -41,6 +42,7 @@ func TestFetchAction(t *testing.T) {
 	require.NoError(t, gitCmd("-C", remote, "add", "."))
 	require.NoError(t, gitCmd("-C", remote, "commit", "-m", "initial"))
 	require.NoError(t, gitCmd("-C", remote, "tag", "v1"))
+	expectedSha := gitRevParse(t, remote, "main")
 
 	gitPath, err := exec.LookPath("git")
 	require.NoError(t, err)
@@ -48,6 +50,7 @@ func TestFetchAction(t *testing.T) {
 	var mu sync.Mutex
 	var archiveRequests []http.Header
 	var mirror *httptest.Server
+	archiveCache := map[string][]byte{}
 	requests, archives, failures := 0, "", 0
 	handler := http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
 		mu.Lock()
@@ -84,7 +87,12 @@ func TestFetchAction(t *testing.T) {
 		if archives != "unprefixed" {
 			args = append(args, "--prefix=action/")
 		}
-		archive, _ := runGit(request.Context(), remote, nil, args...)
+		archiveKey := strings.Join(args, "\x00")
+		archive, ok := archiveCache[archiveKey]
+		if !ok {
+			archive, _ = runGit(request.Context(), remote, nil, args...)
+			archiveCache[archiveKey] = archive
+		}
 		writer.Header().Set("ETag", etag)
 		_, _ = writer.Write(archive)
 	})
@@ -107,7 +115,6 @@ func TestFetchAction(t *testing.T) {
 	}
 
 	t.Run("shares identical fetches and isolates cancellation, panics and differing inputs", func(t *testing.T) {
-		expectedSha := gitRevParse(t, remote, "main")
 		for _, scenario := range []string{"shared", "leader cancelled", "leader panics", "different token"} {
 			t.Run(scenario, func(t *testing.T) {
 				serve("prefixed", 0)
@@ -180,7 +187,7 @@ func TestFetchAction(t *testing.T) {
 		input := newInput("v1")
 		tree, sha, err := FetchAction(t.Context(), input)
 		require.NoError(t, err)
-		assert.Equal(t, gitRevParse(t, remote, "v1"), sha)
+		assert.Equal(t, expectedSha, sha)
 		assert.Equal(t, filepath.Join(input.Dir, sha), tree)
 		assert.NoDirExists(t, filepath.Join(tree, ".git"))
 		info, err := os.Stat(filepath.Join(tree, "run.sh"))
@@ -196,9 +203,10 @@ func TestFetchAction(t *testing.T) {
 
 		require.NoError(t, gitCmd("-C", remote, "commit", "--allow-empty", "-m", "moved"))
 		require.NoError(t, gitCmd("-C", remote, "tag", "--force", "v1"))
+		expectedSha = gitRevParse(t, remote, "v1")
 		moved, movedSha, err := FetchAction(t.Context(), input)
 		require.NoError(t, err)
-		assert.Equal(t, gitRevParse(t, remote, "v1"), movedSha)
+		assert.Equal(t, expectedSha, movedSha)
 		assert.NotEqual(t, tree, moved)
 		assert.FileExists(t, filepath.Join(tree, "action.yml"))
 
@@ -206,10 +214,17 @@ func TestFetchAction(t *testing.T) {
 		_, offlineSha, err := FetchAction(t.Context(), input)
 		require.NoError(t, err)
 		assert.Equal(t, movedSha, offlineSha)
+		require.NoError(t, os.Truncate(filepath.Join(moved, "action.yml"), 4)) // fetched again, offline or not
+		_, _, err = FetchAction(t.Context(), input)
+		require.NoError(t, err)
+		require.NoError(t, os.Remove(moved+treeSizes)) // from an older runner, used as it is offline
+		reused, _, err := FetchAction(t.Context(), input)
+		require.NoError(t, err)
+		assert.Equal(t, moved, reused)
 
 		requests, headers := served()
-		assert.Equal(t, 3, requests)
-		require.Len(t, headers, 3)
+		assert.Equal(t, 4, requests)
+		require.Len(t, headers, 4)
 		assert.Equal(t, "Basic "+base64.StdEncoding.EncodeToString([]byte("token:secret")), headers[0].Get("Authorization"))
 		assert.Empty(t, headers[0].Get("If-None-Match"))
 		assert.Equal(t, fmt.Sprintf(`W/"%s-tar.gz-true"`, sha), headers[1].Get("If-None-Match"))
@@ -218,7 +233,7 @@ func TestFetchAction(t *testing.T) {
 
 	t.Run("downloads a pinned commit once and rejects a short one", func(t *testing.T) {
 		serve("prefixed", 0)
-		sha := gitRevParse(t, remote, "main")
+		sha := expectedSha
 		input := newInput(sha)
 		for range 2 {
 			tree, got, err := FetchAction(t.Context(), input)
@@ -233,12 +248,74 @@ func TestFetchAction(t *testing.T) {
 		assert.NoDirExists(t, filepath.Join(input.Dir, "git"))
 	})
 
+	t.Run("downloads again over a tree whose files changed size or that has no record, not under a job copying it", func(t *testing.T) {
+		sha := gitRevParse(t, remote, "main")
+		for _, mode := range []string{"prefixed", "none"} {
+			serve(mode, 0)
+			input := newInput(sha)
+			tree, _, err := FetchAction(t.Context(), input)
+			require.NoError(t, err, mode)
+			yml := filepath.Join(tree, "action.yml")
+			require.NoError(t, os.WriteFile(filepath.Join(tree, "trampoline.js"), nil, 0o644)) // the runner's own, no damage
+			_, _, err = FetchAction(t.Context(), input)
+			require.NoError(t, err, mode)
+			for _, damage := range []func() error{
+				func() error { return os.Truncate(yml, 4) },
+				func() error { return os.Remove(tree + treeSizes) },
+			} {
+				require.NoError(t, damage(), mode)
+				_, _, err := FetchAction(t.Context(), input)
+				require.NoError(t, err, mode)
+				body, err := os.ReadFile(yml)
+				require.NoError(t, err, mode)
+				assert.Equal(t, "name: action\n", string(body), mode)
+			}
+			_, headers := served()
+			assert.Len(t, headers, 3, mode)
+		}
+
+		serve("prefixed", 0)
+		input := newInput(sha)
+		tree, _, err := FetchAction(t.Context(), input)
+		require.NoError(t, err)
+		yml := filepath.Join(tree, "action.yml")
+		require.NoError(t, os.Truncate(yml, 4))
+		unlock, err := AcquireCloneLock(t.Context(), tree)
+		require.NoError(t, err)
+		release := sync.OnceFunc(unlock)
+		defer release()
+		waiting := make(chan struct{})
+		lockTree = func(ctx context.Context, dir string) (func(), error) {
+			close(waiting)
+			return AcquireCloneLock(ctx, dir)
+		}
+		t.Cleanup(func() { lockTree = AcquireCloneLock })
+		done := make(chan error, 1)
+		go func() {
+			_, _, err := FetchAction(t.Context(), input)
+			done <- err
+		}()
+		select {
+		case err := <-done:
+			t.Fatalf("FetchAction replaced the tree without waiting for the job holding it: %v", err)
+		case <-waiting:
+		}
+		info, err := os.Stat(yml)
+		require.NoError(t, err)
+		assert.Equal(t, int64(4), info.Size(), "the tree a job holds is left alone")
+		release()
+		require.NoError(t, <-done)
+		body, err := os.ReadFile(yml)
+		require.NoError(t, err)
+		assert.Equal(t, "name: action\n", string(body))
+	})
+
 	t.Run("retries a failing download and drops the token on a redirect to another origin", func(t *testing.T) {
 		serve("redirect", 2)
 		input := newInput("main")
 		_, sha, err := FetchAction(t.Context(), input)
 		require.NoError(t, err)
-		assert.Equal(t, gitRevParse(t, remote, "main"), sha)
+		assert.Equal(t, expectedSha, sha)
 		assert.NoDirExists(t, filepath.Join(input.Dir, "git"))
 		_, headers := served()
 		require.Len(t, headers, 4)
@@ -251,7 +328,7 @@ func TestFetchAction(t *testing.T) {
 			serve(mode, 0)
 			tree, sha, err := FetchAction(t.Context(), newInput("main"))
 			require.NoError(t, err, mode)
-			assert.Equal(t, gitRevParse(t, remote, "main"), sha, mode)
+			assert.Equal(t, expectedSha, sha, mode)
 			assert.FileExists(t, filepath.Join(tree, "action.yml"), mode)
 			assert.NoDirExists(t, filepath.Join(tree, ".git"), mode)
 			_, headers := served()
@@ -266,7 +343,7 @@ func TestFetchAction(t *testing.T) {
 		input.URL = "https://example.invalid/org/action"
 		_, sha, err := FetchAction(t.Context(), input)
 		require.NoError(t, err)
-		assert.Equal(t, gitRevParse(t, remote, "main"), sha)
+		assert.Equal(t, expectedSha, sha)
 		_, headers := served()
 		assert.Len(t, headers, 1)
 		assert.NoDirExists(t, filepath.Join(input.Dir, "git"))
@@ -306,10 +383,30 @@ func TestExtractTreeConfinesEntries(t *testing.T) {
 			require.NoError(t, writer.Close())
 
 			dir := t.TempDir()
-			_, err := extractTree(&archive, dir, false)
+			_, err := extractTree(t.Context(), &archive, dir, false)
 			require.Error(t, err)
 			assert.NoFileExists(t, filepath.Join(dir, "escape"))
 			assert.NoFileExists(t, filepath.Join(outside, "escape"))
 		})
 	}
 }
+
+func TestTreeIntactFindsNamesThatAreNotUTF8(t *testing.T) {
+	if runtime.GOOS != "linux" {
+		t.Skip("macOS and Windows store no name that is not UTF-8")
+	}
+	sha := strings.Repeat("a", 40)
+	var archive bytes.Buffer
+	writer := tar.NewWriter(&archive)
+	require.NoError(t, writer.WriteHeader(&tar.Header{Typeflag: tar.TypeXGlobalHeader, PAXRecords: map[string]string{"comment": sha}}))
+	require.NoError(t, writer.WriteHeader(&tar.Header{Name: "caf\xe9.js", Typeflag: tar.TypeReg, Mode: 0o644, Size: 1, Format: tar.FormatGNU}))
+	_, err := writer.Write([]byte("x"))
+	require.NoError(t, err)
+	require.NoError(t, writer.Close())
+
+	dir := t.TempDir()
+	_, err = extractTree(t.Context(), &archive, dir, false)
+	require.NoError(t, err)
+	require.FileExists(t, filepath.Join(dir, sha, "caf\xe9.js"))
+	assert.True(t, treeIntact(dir, sha))
+}

act/common/git/git.go

@@ -10,6 +10,7 @@ import (
 	"encoding/base64"
 	"errors"
 	"fmt"
+	"io"
 	"net/http"
 	"net/url"
 	"os"
@@ -320,11 +321,20 @@ func cachedRef(ref string) string {
 
 func resolveRef(ctx context.Context, dir, ref string) (string, error) {
 	if !isFullObjectID(ref) {
-		for _, name := range []string{cachedRef(ref), "refs/tags/" + ref, "refs/remotes/origin/" + ref} {
-			if sha, err := gitOutput(ctx, dir, "rev-parse", "--verify", name+"^{commit}"); err == nil {
+		revisions := []string{cachedRef(ref) + "^{commit}", "refs/tags/" + ref + "^{commit}", "refs/remotes/origin/" + ref + "^{commit}"}
+		if strings.ContainsAny(ref, "\x00\n") {
+			for _, revision := range revisions {
+				if sha, err := gitOutput(ctx, dir, "rev-parse", "--verify", revision); err == nil {
 					return sha, nil
 				}
 			}
+		} else if output, err := runGitWithStdin(ctx, dir, nil, strings.NewReader(strings.Join(revisions, "\n")+"\n"), "cat-file", "--batch-check=%(objectname)"); err == nil {
+			for sha := range strings.SplitSeq(string(output), "\n") {
+				if isFullObjectID(sha) {
+					return sha, nil
+				}
+			}
+		}
 	}
 	sha, err := gitOutput(ctx, dir, "rev-parse", "--verify", ref+"^{commit}")
 	if err == nil && hexRefRegex.MatchString(ref) && len(sha) != len(ref) {
@@ -427,6 +437,10 @@ func ListFiles(ctx context.Context, dir string) ([]string, error) {
 }
 
 func runGit(ctx context.Context, dir string, input *NewGitCloneExecutorInput, args ...string) ([]byte, error) {
+	return runGitWithStdin(ctx, dir, input, nil, args...)
+}
+
+func runGitWithStdin(ctx context.Context, dir string, input *NewGitCloneExecutorInput, stdin io.Reader, args ...string) ([]byte, error) {
 	config := []string{"core.autocrlf=false", "core.eol=lf"} // keep action files LF
 	if input != nil {
 		config = append(config, RemoteConfig(input.Token, input.InsecureSkipTLS)...)
@@ -438,6 +452,7 @@ func runGit(ctx context.Context, dir string, input *NewGitCloneExecutorInput, ar
 		args = append([]string{"-C", dir}, args...)
 	}
 	cmd := exec.CommandContext(ctx, "git", args...)
+	cmd.Stdin = stdin
 	cmd.Env = slices.DeleteFunc(os.Environ(), func(entry string) bool {
 		name, _, _ := strings.Cut(entry, "=")
 		return slices.Contains(LocalEnvVars, name)

act/common/git/git_test.go

@@ -29,6 +29,31 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+var gitFixtureDir string
+
+func TestMain(m *testing.M) {
+	var err error
+	gitFixtureDir, err = os.MkdirTemp("", "runner-git-fixture-")
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+	code := 1
+	if err = gitCmd("init", "--quiet", "--template=", "--bare", "--initial-branch=main", filepath.Join(gitFixtureDir, "remote")); err == nil {
+		err = gitCmd("init", "--quiet", "--template=", "--initial-branch=main", filepath.Join(gitFixtureDir, "work"))
+	}
+	if err == nil {
+		code = m.Run()
+	} else {
+		fmt.Fprintln(os.Stderr, err)
+	}
+	if err := os.RemoveAll(gitFixtureDir); err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		code = 1
+	}
+	os.Exit(code)
+}
+
 func TestFindGitSlug(t *testing.T) {
 	for _, testcase := range []struct {
 		remoteURL string
@@ -141,9 +166,8 @@ func TestGitFindRef(t *testing.T) {
 	} {
 		t.Run(name, func(t *testing.T) {
 			dir := filepath.Join(basedir, name)
-			require.NoError(t, os.MkdirAll(dir, 0o755))
-			require.NoError(t, gitCmd("-C", dir, "init", "--initial-branch=master"))
-			require.NoError(t, os.RemoveAll(filepath.Join(dir, ".git", "hooks")))
+			require.NoError(t, os.CopyFS(dir, os.DirFS(filepath.Join(gitFixtureDir, "work"))))
+			require.NoError(t, os.WriteFile(filepath.Join(dir, ".git", "HEAD"), []byte("ref: refs/heads/master\n"), 0o644))
 			for _, command := range tt.commands {
 				require.NoError(t, gitCmd(append([]string{"-C", dir}, command...)...))
 			}
@@ -180,6 +204,7 @@ func TestGitCloneExecutor(t *testing.T) {
 	t.Setenv("GIT_CONFIG_GLOBAL", hostConfig)
 	t.Setenv("GIT_DIR", t.TempDir())
 
+	var clones sync.WaitGroup
 	for name, tt := range map[string]struct {
 		Err error
 		Ref string
@@ -207,6 +232,7 @@ func TestGitCloneExecutor(t *testing.T) {
 			Ref: fullSha[:4],
 		},
 	} {
+		clones.Go(func() {
 			t.Run(name, func(t *testing.T) {
 				dir := t.TempDir()
 				clone := NewGitCloneExecutor(NewGitCloneExecutorInput{
@@ -231,7 +257,9 @@ func TestGitCloneExecutor(t *testing.T) {
 				require.NoError(t, err)
 				assert.Equal(t, "#!/bin/sh\necho hi\n", string(script), "host core.autocrlf must not convert action files")
 			})
+		})
 	}
+	clones.Wait()
 }
 
 func TestGitCloneExecutorReclonesWhenOriginURLChanges(t *testing.T) {
@@ -476,9 +504,12 @@ func TestGitCloneExecutorPinnedHashIgnoresShadowingRef(t *testing.T) {
 func newGitRemote(t *testing.T, messages ...string) (string, string) {
 	t.Helper()
 	remoteDir, workDir := t.TempDir(), t.TempDir()
-	require.NoError(t, gitCmd("init", "--bare", "--initial-branch=main", remoteDir))
-	require.NoError(t, gitCmd("clone", remoteDir, workDir))
-	require.NoError(t, gitCmd("-C", workDir, "checkout", "-b", "main"))
+	require.NoError(t, os.CopyFS(remoteDir, os.DirFS(filepath.Join(gitFixtureDir, "remote"))))
+	require.NoError(t, os.CopyFS(workDir, os.DirFS(filepath.Join(gitFixtureDir, "work"))))
+	config, err := os.ReadFile(filepath.Join(workDir, ".git", "config"))
+	require.NoError(t, err)
+	require.NoError(t, os.WriteFile(filepath.Join(workDir, ".git", "config"),
+		fmt.Appendf(config, "\n[remote \"origin\"]\n\turl = %s\n\tfetch = +refs/heads/*:refs/remotes/origin/*\n", strconv.Quote(remoteDir)), 0o644))
 	for _, message := range messages {
 		require.NoError(t, gitCmd("-C", workDir, "commit", "--allow-empty", "-m", message))
 	}

act/container/docker_build.go

@@ -74,8 +74,10 @@ func NewDockerBuildExecutor(input NewDockerBuildExecutorInput) common.Executor {
 
 		logger.Debugf("Creating image from context dir '%s' with tag '%s' and platform '%s'", input.ContextDir, input.ImageTag, input.Platform)
 		resp, err := cli.ImageBuild(ctx, buildContext, options)
-
-		return logDockerResponse(logger, resp.Body, err != nil)
+		if err != nil {
+			return err
+		}
+		return logDockerResponse(logger, resp.Body, false)
 	}
 }
 

act/container/docker_build_test.go

@@ -8,13 +8,35 @@ package container
 import (
 	"archive/tar"
 	"io"
+	"net/http"
+	"net/http/httptest"
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"
 
 	"github.com/stretchr/testify/require"
 )
 
+func TestDockerBuildReportsDaemonError(t *testing.T) {
+	daemon := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		writer.Header().Set("API-Version", "1.47")
+		if strings.HasSuffix(request.URL.Path, "/build") {
+			writer.Header().Set("Content-Type", "application/json")
+			writer.WriteHeader(http.StatusInternalServerError)
+			_, _ = io.WriteString(writer, `{"message":"build rejected"}`)
+		}
+	}))
+	t.Cleanup(daemon.Close)
+	t.Setenv("DOCKER_HOST", daemon.URL)
+	t.Setenv("DOCKER_CONFIG", t.TempDir())
+	t.Setenv("DOCKER_AUTH_CONFIG", "")
+
+	require.ErrorContains(t, NewDockerBuildExecutor(NewDockerBuildExecutorInput{
+		ImageTag: "action:test", Dockerfile: "Dockerfile", BuildContext: strings.NewReader(""),
+	})(t.Context()), "build rejected")
+}
+
 func TestCreateBuildContextFollowsDockerignoreAndSymlinkedRootButKeepsDockerfile(t *testing.T) {
 	dir := t.TempDir()
 	for name, content := range map[string]string{

act/container/docker_images_test.go

@@ -5,14 +5,17 @@
 package container
 
 import (
+	"bytes"
+	"context"
 	"fmt"
-	"os"
-	"os/exec"
 	"strings"
 	"testing"
+	"time"
 
+	"gitea.com/gitea/runner/act/common"
 	"gitea.com/gitea/runner/act/common/logging"
 
+	mobyclient "github.com/moby/moby/client"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -21,48 +24,47 @@ func init() {
 	logging.SetLevel(logging.LevelDebug)
 }
 
-// buildScratchImage builds a tiny empty image for the given platform locally (FROM scratch, no
-// network or emulation since there is nothing to run) and returns its tag, removing it after
-// the test.
-func buildScratchImage(t *testing.T, platform string) string {
+func loadScratchImage(t *testing.T, platform string) string {
 	t.Helper()
-	tag := fmt.Sprintf("act-test-exists-%s:latest", strings.TrimPrefix(platform, "linux/"))
-	cmd := exec.Command("docker", "build", "--platform", platform, "-t", tag, "-")
-	cmd.Stdin = strings.NewReader("FROM scratch\nLABEL act-test=1\n")
-	// Force BuildKit: it records the requested architecture in the image config for a
-	// FROM-scratch build, whereas the classic builder ignores --platform and tags it with the
-	// host arch, which would break the per-platform existence assertions below.
-	cmd.Env = append(os.Environ(), "DOCKER_BUILDKIT=1")
-	out, err := cmd.CombinedOutput()
-	require.NoError(t, err, string(out))
-	t.Cleanup(func() { _ = exec.Command("docker", "rmi", "-f", tag).Run() })
+	tag := fmt.Sprintf("act-test-exists-%s-%d:latest", strings.TrimPrefix(platform, "linux/"), time.Now().UnixNano())
+	var archive bytes.Buffer
+	require.NoError(t, writeFilesTar(t.Context(), &archive, 0, 0,
+		&FileEntry{Name: "config.json", Mode: 0o600, Body: fmt.Sprintf(`{"architecture":%q,"os":"linux","rootfs":{"type":"layers","diff_ids":[]}}`, strings.TrimPrefix(platform, "linux/"))},
+		&FileEntry{Name: "manifest.json", Mode: 0o600, Body: fmt.Sprintf(`[{"Config":"config.json","RepoTags":[%q],"Layers":[]}]`, tag)},
+	))
+	cli, err := GetDockerClient(t.Context())
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = cli.Close() })
+	response, err := cli.ImageLoad(t.Context(), &archive)
+	require.NoError(t, err)
+	defer response.Close()
+	require.NoError(t, logDockerResponse(common.Logger(t.Context()), response, false))
+	t.Cleanup(func() {
+		_, _ = cli.ImageRemove(context.WithoutCancel(t.Context()), tag, mobyclient.ImageRemoveOptions{Force: true})
+	})
 	return tag
 }
 
 func TestImageExistsLocally(t *testing.T) {
+	t.Parallel()
 	requireDocker(t)
 	ctx := t.Context()
 
-	// a non-existent image is reported absent
 	missing, err := ImageExistsLocally(ctx, "library/alpine:this-random-tag-will-never-exist", "linux/amd64")
 	assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
 	assert.False(t, missing)
 
-	// Build tiny images for two architectures locally so per-platform existence can be checked
-	// offline (formerly pulled node:24-bookworm-slim for amd64 and arm64 over the network).
-	amd64Ref := buildScratchImage(t, "linux/amd64")
-	arm64Ref := buildScratchImage(t, "linux/arm64")
+	amd64Ref := loadScratchImage(t, "linux/amd64")
+	arm64Ref := loadScratchImage(t, "linux/arm64")
 
 	amd64Exists, err := ImageExistsLocally(ctx, amd64Ref, "linux/amd64")
 	assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
 	assert.True(t, amd64Exists)
 
-	// a non-host architecture image is detected for its own architecture
 	arm64Exists, err := ImageExistsLocally(ctx, arm64Ref, "linux/arm64")
 	assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
 	assert.True(t, arm64Exists)
 
-	// a present image is reported absent for a different platform
 	wrongPlatform, err := ImageExistsLocally(ctx, amd64Ref, "linux/arm64")
 	assert.NoError(t, err) //nolint:testifylint // pre-existing issue from nektos/act
 	assert.False(t, wrongPlatform)

act/container/docker_logger.go

@@ -33,6 +33,7 @@ func logDockerResponse(logger *logging.Logger, dockerResponse io.ReadCloser, isE
 	defer dockerResponse.Close()
 
 	scanner := bufio.NewScanner(dockerResponse)
+	scanner.Buffer(nil, 64*1024*1024) // a build echoes long instructions in one record
 	for scanner.Scan() {
 		line := scanner.Bytes()
 		var msg dockerMessage
@@ -65,7 +66,7 @@ func logDockerResponse(logger *logging.Logger, dockerResponse io.ReadCloser, isE
 		}
 	}
 
-	return nil
+	return scanner.Err()
 }
 
 func writeLog(logger *logging.Logger, isError bool, format string, args ...any) {

act/container/docker_network.go

@@ -85,14 +85,14 @@ networks:
 }
 
 func NewDockerNetworkCreateExecutor(name string, opts NewDockerNetworkCreateExecutorInput) common.Executor {
-	return func(ctx context.Context) error {
+	return common.Executor(func(ctx context.Context) error {
 		cli, err := GetDockerClient(ctx)
 		if err != nil {
 			return err
 		}
 		defer cli.Close()
 		return createDockerNetwork(ctx, cli, name, opts)
-	}
+	}).IfNot(common.Dryrun)
 }
 
 func createDockerNetwork(ctx context.Context, cli client.APIClient, name string, opts NewDockerNetworkCreateExecutorInput) error {
@@ -157,7 +157,7 @@ func isAddressPoolExhausted(err error) bool {
 }
 
 func NewDockerNetworkConnectExecutor(networkName, container string) common.Executor {
-	return func(ctx context.Context) error {
+	return common.Executor(func(ctx context.Context) error {
 		cli, err := GetDockerClient(ctx)
 		if err != nil {
 			return err
@@ -167,7 +167,7 @@ func NewDockerNetworkConnectExecutor(networkName, container string) common.Execu
 			common.Logger(ctx).Warnf("cannot connect %s to network %s, so jobs on it cannot reach the cache: %v", container, networkName, err)
 		}
 		return nil
-	}
+	}).IfNot(common.Dryrun)
 }
 
 func connectNetwork(ctx context.Context, cli client.APIClient, networkName, container string) error {
@@ -182,14 +182,14 @@ func connectNetwork(ctx context.Context, cli client.APIClient, networkName, cont
 }
 
 func NewDockerNetworkDisconnectExecutor(networkName, container string) common.Executor {
-	return func(ctx context.Context) error {
+	return common.Executor(func(ctx context.Context) error {
 		cli, err := GetDockerClient(ctx)
 		if err != nil {
 			return err
 		}
 		defer cli.Close()
 		return disconnectNetwork(ctx, cli, networkName, container)
-	}
+	}).IfNot(common.Dryrun)
 }
 
 func disconnectNetwork(ctx context.Context, cli client.APIClient, networkName, container string) error {
@@ -201,14 +201,14 @@ func disconnectNetwork(ctx context.Context, cli client.APIClient, networkName, c
 }
 
 func NewDockerNetworkRemoveExecutor(name string) common.Executor {
-	return func(ctx context.Context) error {
+	return common.Executor(func(ctx context.Context) error {
 		cli, err := GetDockerClient(ctx)
 		if err != nil {
 			return err
 		}
 		defer cli.Close()
 		return removeNetwork(ctx, cli, name)
-	}
+	}).IfNot(common.Dryrun)
 }
 
 func removeNetwork(ctx context.Context, cli client.APIClient, name string) error {

act/container/docker_proxy.go

@@ -41,6 +41,8 @@ const (
 	maxCreateBody = 8 << 20
 
 	dockerProxyProbeTimeout       = 5 * time.Second
+	dockerProxyProbeWorkerTimeout = time.Minute
+	dockerProxyProbeTarget        = "/gitea-runner-probe"
 )
 
 var (
@@ -48,20 +50,27 @@ var (
 	rawStreamPath = regexp.MustCompile(`^(/v[0-9.]+)?/(containers/[^/]+/attach|exec/[^/]+/start)$`)
 )
 
-var dockerProxyDirectories sync.Map
+var dockerProxyProbes sync.Map
 
 type dockerProxyDirectory struct {
 	dir, daemonDir string
 	socket         os.FileInfo
 }
 
+type dockerProxyProbe struct {
+	directory dockerProxyDirectory
+	done      chan struct{}
+	seen      bool
+	err       error
+}
+
 func NewDockerProxy(ctx context.Context, job string) *DockerProxy {
 	if host := os.Getenv("DOCKER_HOST"); runtime.GOOS != "linux" || host != "" && !strings.HasPrefix(host, "unix://") {
 		return nil
 	}
 	probeCtx, cancel := context.WithTimeout(ctx, dockerProxyProbeTimeout)
 	defer cancel()
-	cli, err := GetDockerClient(probeCtx)
+	cli, err := client.New(client.FromEnv) // negotiates on first request, so a cached probe costs no round trip
 	if err != nil {
 		return nil
 	}
@@ -76,42 +85,72 @@ func NewDockerProxy(ctx context.Context, job string) *DockerProxy {
 	}
 	workdir, _ := os.Getwd()
 	key := [3]string{cli.DaemonHost(), workdir, os.TempDir()}
-	cached, _ := dockerProxyDirectories.Load(key)
-	directory, seen := cached.(dockerProxyDirectory)
-	seen = seen && os.SameFile(directory.socket, socket) // daemon replacement invalidates cached paths
-	dir, daemonDir := directory.dir, directory.daemonDir
-	if !seen {
-		dir, daemonDir = runnerContainerWorkdir(probeCtx, cli)
-		seen = daemonDir != ""
-		if !seen {
-			if dir, err = filepath.Abs(os.TempDir()); err == nil {
-				daemonDir = dir
-				seen, err = daemonSeesDir(probeCtx, cli, dir, daemonDir)
-			}
-		}
-	}
+	probe, err := probeDockerProxyDirectory(probeCtx, cli, key, socket)
 	if err != nil {
 		common.Logger(ctx).Infof("docker proxy probe failed, jobs get the daemon socket directly: %v", err)
 		return nil
 	}
-	if !seen {
+	if !probe.seen {
 		common.Logger(ctx).Infof("the docker daemon cannot reach the runner's temporary or working directory, jobs get the daemon socket directly")
 		return nil
 	}
 	if ctx.Err() != nil {
 		return nil
 	}
-	proxy, err := StartDockerProxy(daemonSocket, dir, job)
+	proxy, err := StartDockerProxy(daemonSocket, probe.directory.dir, job)
 	if err != nil {
-		dockerProxyDirectories.Delete(key)
+		dockerProxyProbes.CompareAndDelete(key, probe)
 		common.Logger(ctx).Warnf("docker proxy not started, the job gets the daemon socket directly: %v", err)
 		return nil
 	}
-	dockerProxyDirectories.Store(key, dockerProxyDirectory{dir, daemonDir, socket})
-	proxy.Socket = daemonDir + strings.TrimPrefix(proxy.Socket, dir)
+	proxy.Socket = probe.directory.daemonDir + strings.TrimPrefix(proxy.Socket, probe.directory.dir)
 	return proxy
 }
 
+func probeDockerProxyDirectory(ctx context.Context, cli client.APIClient, key [3]string, socket os.FileInfo) (*dockerProxyProbe, error) {
+	for {
+		probe := &dockerProxyProbe{directory: dockerProxyDirectory{socket: socket}, done: make(chan struct{})}
+		actual, loaded := dockerProxyProbes.LoadOrStore(key, probe)
+		if loaded {
+			probe, _ = actual.(*dockerProxyProbe)
+			if !os.SameFile(probe.directory.socket, socket) { // daemon replacement invalidates cached paths
+				dockerProxyProbes.CompareAndDelete(key, probe)
+				continue
+			}
+		} else {
+			go func() { // outlives callers' deadlines so a slow daemon still gets its result cached
+				defer cli.Close()
+				ctx, cancel := context.WithTimeout(context.Background(), dockerProxyProbeWorkerTimeout)
+				defer cancel()
+				probe.directory.dir, probe.directory.daemonDir = runnerContainerWorkdir(ctx, cli)
+				probe.seen = probe.directory.daemonDir != ""
+				if !probe.seen {
+					if probe.directory.dir, probe.err = filepath.Abs(key[2]); probe.err == nil {
+						probe.directory.daemonDir = probe.directory.dir
+						probe.seen, probe.err = daemonSeesDir(ctx, cli, probe.directory.dir, probe.directory.daemonDir)
+					}
+				}
+				if probe.seen {
+					dockerProxyProbes.LoadOrStore(key, probe)
+				} else {
+					dockerProxyProbes.CompareAndDelete(key, probe)
+				}
+				close(probe.done)
+				removeStaleProbeContainers(cli)
+			}()
+		}
+		select {
+		case <-ctx.Done():
+			return &dockerProxyProbe{}, ctx.Err()
+		case <-probe.done:
+		}
+		if ctx.Err() == nil && errors.Is(probe.err, context.DeadlineExceeded) {
+			continue
+		}
+		return probe, probe.err
+	}
+}
+
 // runnerContainerWorkdir looks the runner's container up by hostname to find the daemon's path to its working directory.
 func runnerContainerWorkdir(ctx context.Context, cli client.APIClient) (workdir, daemonDir string) {
 	workdir, err := os.Getwd()
@@ -162,19 +201,25 @@ func daemonSeesDir(ctx context.Context, cli client.APIClient, dir, daemonDir str
 	if len(images.Items) == 0 {
 		return false, errors.New("no image available for the docker proxy probe")
 	}
-	created, err := cli.ContainerCreate(ctx, client.ContainerCreateOptions{
-		Config: &container.Config{Image: images.Items[0].ID, Cmd: []string{"true"}},
+	var created client.ContainerCreateResult
+	for _, image := range images.Items {
+		created, err = cli.ContainerCreate(ctx, client.ContainerCreateOptions{
+			Config: &container.Config{Image: image.ID, Cmd: []string{"true"}},
 			HostConfig: &container.HostConfig{Mounts: []mount.Mount{
-			{Type: mount.TypeBind, Source: daemonDir, Target: "/gitea-runner-probe", ReadOnly: true}, // not the marker itself, podman creates a missing bind source where docker rejects it
+				{Type: mount.TypeBind, Source: daemonDir, Target: dockerProxyProbeTarget, ReadOnly: true}, // not the marker itself, podman creates a missing bind source where docker rejects it
 			}},
 		})
+		if !cerrdefs.IsNotFound(err) {
+			break
+		}
+	}
 	if cerrdefs.IsInvalidArgument(err) {
 		return false, nil
 	}
 	if err != nil {
 		return false, err
 	}
-	seen, err := containerSeesMarker(ctx, cli, dir, created.ID, "/gitea-runner-probe")
+	seen, err := containerSeesMarker(ctx, cli, dir, created.ID, dockerProxyProbeTarget)
 	cleanupCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), dockerProxyProbeTimeout)
 	defer cancel()
 	if _, removeErr := cli.ContainerRemove(cleanupCtx, created.ID, client.ContainerRemoveOptions{Force: true, RemoveVolumes: true}); removeErr != nil {
@@ -183,6 +228,36 @@ func daemonSeesDir(ctx context.Context, cli client.APIClient, dir, daemonDir str
 	return seen, err
 }
 
+// removeStaleProbeContainers removes probe containers left by abandoned creates once any probe in flight when listed has ended.
+func removeStaleProbeContainers(cli client.APIClient) {
+	ctx, cancel := context.WithTimeout(context.Background(), dockerProxyProbeTimeout)
+	defer cancel()
+	listed, err := cli.ContainerList(ctx, client.ContainerListOptions{
+		All:     true,
+		Filters: make(client.Filters).Add("status", string(container.StateCreated)).Add("volume", dockerProxyProbeTarget),
+	})
+	if err != nil {
+		common.Logger(ctx).Warnf("listing stale docker proxy probe containers failed: %v", err)
+		return
+	}
+	listed.Items = slices.DeleteFunc(listed.Items, func(stale container.Summary) bool {
+		return stale.State != container.StateCreated || !slices.ContainsFunc(stale.Mounts, func(mountPoint container.MountPoint) bool {
+			return mountPoint.Type == mount.TypeBind && !mountPoint.RW && mountPoint.Destination == dockerProxyProbeTarget
+		})
+	})
+	if len(listed.Items) == 0 {
+		return
+	}
+	time.Sleep(dockerProxyProbeWorkerTimeout + dockerProxyProbeTimeout)
+	for _, stale := range listed.Items {
+		ctx, cancel := context.WithTimeout(context.Background(), dockerProxyProbeTimeout)
+		if err := (&containerReference{cli: cli, id: stale.ID}).remove()(ctx); err != nil {
+			common.Logger(ctx).Warnf("removing stale docker proxy probe container failed: %v", err)
+		}
+		cancel()
+	}
+}
+
 // StartDockerProxy serves a job's docker socket in dir, labelling what the job creates through it.
 func StartDockerProxy(daemonSocket, dir, job string) (*DockerProxy, error) {
 	info, err := os.Stat(daemonSocket)

act/container/docker_proxy_test.go

@@ -22,6 +22,7 @@ import (
 	"strings"
 	"sync/atomic"
 	"testing"
+	"testing/synctest"
 	"time"
 
 	cerrdefs "github.com/containerd/errdefs"
@@ -93,6 +94,8 @@ func TestDockerProxy(t *testing.T) {
 				return
 			}
 			_, _ = w.Write([]byte(`[{"Id":"probe-image"}]`))
+		case strings.HasSuffix(r.URL.Path, "/containers/json"):
+			_, _ = w.Write([]byte(`[]`))
 		case strings.HasSuffix(r.URL.Path, "/json"):
 			http.NotFound(w, r)
 		case strings.HasSuffix(r.URL.Path, "/archive"):
@@ -234,7 +237,8 @@ func TestDockerProxy(t *testing.T) {
 			t.Skip("automatic proxy discovery requires Linux")
 		}
 		t.Setenv("DOCKER_HOST", "unix://"+daemonSocket)
-		t.Setenv("TMPDIR", shortTempDir(t))
+		tempDir := shortTempDir(t)
+		t.Setenv("TMPDIR", tempDir)
 		require.Nil(t, NewDockerProxy(t.Context(), "failed"))
 		for _, job := range []string{"first", "second"} {
 			proxy := NewDockerProxy(t.Context(), job)
@@ -242,12 +246,10 @@ func TestDockerProxy(t *testing.T) {
 			require.NoError(t, proxy.Close(context.Background()))
 		}
 		assert.EqualValues(t, 2, probes.Load())
-		dockerProxyDirectories.Range(func(key, value any) bool {
-			directory, _ := value.(dockerProxyDirectory)
-			directory.socket, _ = os.Stat(t.TempDir())
-			dockerProxyDirectories.Store(key, directory)
-			return true
-		})
+		workdir, err := os.Getwd()
+		require.NoError(t, err)
+		cached, _ := dockerProxyProbes.Load([3]string{"unix://" + daemonSocket, workdir, tempDir})
+		cached.(*dockerProxyProbe).directory.socket, _ = os.Stat(t.TempDir())
 		proxy := NewDockerProxy(t.Context(), "replaced daemon")
 		require.NotNil(t, proxy)
 		require.NoError(t, proxy.Close(context.Background()))
@@ -371,6 +373,9 @@ func TestDockerProxyWithDaemon(t *testing.T) {
 
 	job := "proxy-test-" + strconv.FormatInt(time.Now().UnixNano(), 36)
 	proxy, err := StartDockerProxy(daemonSocketPath(t, direct), dir, job)
+	if os.Geteuid() != 0 && errors.Is(err, os.ErrPermission) {
+		t.Skipf("non-root runners that cannot take the daemon socket's ownership get it directly: %v", err)
+	}
 	require.NoError(t, err)
 	t.Cleanup(func() { _ = proxy.Close(context.Background()) })
 	viaProxy, err := mobyclient.New(mobyclient.WithHost("unix://" + proxy.Socket))
@@ -472,10 +477,23 @@ func BenchmarkDockerProxy(b *testing.B) {
 
 type probeClient struct {
 	mobyclient.APIClient
-	create func(mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error)
+	create     func(context.Context, mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error)
 	remove     func(context.Context, string, mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error)
 	stat       func(string) error
 	mounts     []container.MountPoint
+	containers []container.Summary
+}
+
+func (c *probeClient) Close() error {
+	return nil
+}
+
+func (c *probeClient) ContainerList(context.Context, mobyclient.ContainerListOptions) (mobyclient.ContainerListResult, error) {
+	return mobyclient.ContainerListResult{Items: c.containers}, nil
+}
+
+func (c *probeClient) ContainerKill(context.Context, string, mobyclient.ContainerKillOptions) (mobyclient.ContainerKillResult, error) {
+	return mobyclient.ContainerKillResult{}, cerrdefs.ErrConflict
 }
 
 func (c *probeClient) ContainerInspect(context.Context, string, mobyclient.ContainerInspectOptions) (mobyclient.ContainerInspectResult, error) {
@@ -483,11 +501,11 @@ func (c *probeClient) ContainerInspect(context.Context, string, mobyclient.Conta
 }
 
 func (c *probeClient) ImageList(context.Context, mobyclient.ImageListOptions) (mobyclient.ImageListResult, error) {
-	return mobyclient.ImageListResult{Items: []image.Summary{{ID: "probe-image"}}}, nil
+	return mobyclient.ImageListResult{Items: []image.Summary{{ID: "removed-image"}, {ID: "probe-image"}}}, nil
 }
 
-func (c *probeClient) ContainerCreate(_ context.Context, opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
-	return c.create(opts)
+func (c *probeClient) ContainerCreate(ctx context.Context, opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+	return c.create(ctx, opts)
 }
 
 func (c *probeClient) ContainerRemove(ctx context.Context, id string, opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
@@ -498,9 +516,193 @@ func (c *probeClient) ContainerStatPath(_ context.Context, _ string, opts mobycl
 	return mobyclient.ContainerStatPathResult{}, c.stat(opts.Path)
 }
 
+func TestDockerProxyDirectoryProbes(t *testing.T) {
+	type result struct {
+		seen bool
+		err  error
+	}
+	for name, createErr := range map[string]error{"unavailable directory": nil, "daemon error": errors.New("probe failed")} {
+		t.Run(name, func(t *testing.T) {
+			synctest.Test(t, func(t *testing.T) {
+				dir := t.TempDir()
+				socket, err := os.Stat(t.TempDir())
+				require.NoError(t, err)
+				key := [3]string{"daemon", dir, dir}
+				release := make(chan struct{})
+				var calls, removals atomic.Int32
+				visible := false
+				cli := &probeClient{
+					create: func(_ctx context.Context, _opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+						calls.Add(1)
+						<-release
+						return mobyclient.ContainerCreateResult{ID: "probe"}, createErr
+					},
+					stat: func(_path string) error {
+						if visible {
+							return nil
+						} else {
+							return cerrdefs.ErrNotFound
+						}
+					},
+					remove: func(_ctx context.Context, _id string, _opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
+						removals.Add(1)
+						return mobyclient.ContainerRemoveResult{}, nil
+					},
+				}
+				results := make(chan result, 8)
+				for range cap(results) {
+					go func() {
+						probe, err := probeDockerProxyDirectory(t.Context(), cli, key, socket)
+						results <- result{probe.seen, err}
+					}()
+				}
+				synctest.Wait()
+				assert.EqualValues(t, 1, calls.Load())
+				replacement, err := os.Stat(t.TempDir())
+				require.NoError(t, err)
+				replaced := make(chan result, 1)
+				go func() {
+					probe, err := probeDockerProxyDirectory(t.Context(), cli, key, replacement)
+					replaced <- result{probe.seen, err}
+				}()
+				synctest.Wait()
+				assert.EqualValues(t, 2, calls.Load())
+				close(release)
+				for range cap(results) {
+					result := <-results
+					require.ErrorIs(t, result.err, createErr)
+					assert.False(t, result.seen)
+				}
+				replacementResult := <-replaced
+				require.ErrorIs(t, replacementResult.err, createErr)
+				assert.False(t, replacementResult.seen)
+				assert.EqualValues(t, 2, calls.Load())
+				if createErr == nil {
+					assert.EqualValues(t, 2, removals.Load())
+				} else {
+					assert.Zero(t, removals.Load())
+				}
+				createErr, visible = nil, true
+				probe, err := probeDockerProxyDirectory(t.Context(), cli, key, socket)
+				require.NoError(t, err)
+				assert.True(t, probe.seen)
+				assert.Equal(t, dir, probe.directory.dir)
+				assert.Equal(t, dir, probe.directory.daemonDir)
+				assert.EqualValues(t, 3, calls.Load())
+				entries, err := os.ReadDir(dir)
+				require.NoError(t, err)
+				assert.Empty(t, entries)
+			})
+		})
+	}
+	t.Run("cancelled caller leaves the probe to finish for the others", func(t *testing.T) {
+		synctest.Test(t, func(t *testing.T) {
+			dir := t.TempDir()
+			socket, err := os.Stat(t.TempDir())
+			require.NoError(t, err)
+			key := [3]string{"daemon", dir, dir}
+			ctx, cancel := context.WithCancel(t.Context())
+			defer cancel()
+			release := make(chan struct{})
+			var calls, removals atomic.Int32
+			cli := &probeClient{
+				create: func(ctx context.Context, _opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+					calls.Add(1)
+					<-release
+					return mobyclient.ContainerCreateResult{ID: "probe"}, ctx.Err()
+				},
+				stat: func(_path string) error { return nil },
+				remove: func(_ctx context.Context, _id string, _opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
+					removals.Add(1)
+					return mobyclient.ContainerRemoveResult{}, nil
+				},
+			}
+			leader, follower, late := make(chan result, 1), make(chan result, 1), make(chan result, 1)
+			go func() {
+				probe, err := probeDockerProxyDirectory(ctx, cli, key, socket)
+				leader <- result{probe.seen, err}
+			}()
+			synctest.Wait()
+			go func() {
+				probe, err := probeDockerProxyDirectory(t.Context(), cli, key, socket)
+				follower <- result{probe.seen, err}
+			}()
+			synctest.Wait()
+			cancel()
+			leaderResult := <-leader
+			require.ErrorIs(t, leaderResult.err, context.Canceled)
+			assert.False(t, leaderResult.seen)
+			go func() {
+				probe, err := probeDockerProxyDirectory(t.Context(), cli, key, socket)
+				late <- result{probe.seen, err}
+			}()
+			synctest.Wait()
+			close(release)
+			for _, results := range []chan result{follower, late} {
+				result := <-results
+				require.NoError(t, result.err)
+				assert.True(t, result.seen)
+			}
+			assert.EqualValues(t, 1, calls.Load())
+			assert.EqualValues(t, 1, removals.Load())
+			_, cached := dockerProxyProbes.Load(key)
+			assert.True(t, cached)
+		})
+	})
+	t.Run("caller joining a probe that ran out of time retries", func(t *testing.T) {
+		synctest.Test(t, func(t *testing.T) {
+			dir := t.TempDir()
+			socket, err := os.Stat(t.TempDir())
+			require.NoError(t, err)
+			var calls atomic.Int32
+			cli := &probeClient{
+				create: func(ctx context.Context, _opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+					if calls.Add(1) == 1 {
+						<-ctx.Done()
+						return mobyclient.ContainerCreateResult{}, ctx.Err()
+					}
+					return mobyclient.ContainerCreateResult{ID: "probe"}, nil
+				},
+				stat: func(_path string) error { return nil },
+				remove: func(_ctx context.Context, _id string, _opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
+					return mobyclient.ContainerRemoveResult{}, nil
+				},
+			}
+			probe, err := probeDockerProxyDirectory(t.Context(), cli, [3]string{"daemon", dir, dir}, socket)
+			require.NoError(t, err)
+			assert.True(t, probe.seen)
+			assert.EqualValues(t, 2, calls.Load())
+		})
+	})
+}
+
+func TestRemoveStaleProbeContainers(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		probeMount := container.MountPoint{Type: mount.TypeBind, Destination: "/gitea-runner-probe"}
+		writableMount := container.MountPoint{Type: mount.TypeBind, Destination: "/gitea-runner-probe", RW: true}
+		var removed []string
+		cli := &probeClient{
+			containers: []container.Summary{
+				{ID: "abandoned", State: container.StateCreated, Mounts: []container.MountPoint{probeMount}},
+				{ID: "foreign", State: container.StateCreated, Mounts: []container.MountPoint{writableMount}},
+				{ID: "started", State: container.StateExited, Mounts: []container.MountPoint{probeMount}},
+			},
+			remove: func(_ctx context.Context, id string, _opts mobyclient.ContainerRemoveOptions) (mobyclient.ContainerRemoveResult, error) {
+				removed = append(removed, id)
+				return mobyclient.ContainerRemoveResult{}, nil
+			},
+		}
+		start := time.Now()
+		removeStaleProbeContainers(cli)
+		assert.Equal(t, []string{"abandoned"}, removed)
+		assert.GreaterOrEqual(t, time.Since(start), dockerProxyProbeWorkerTimeout+dockerProxyProbeTimeout)
+	})
+}
+
 func TestDaemonSeesDir(t *testing.T) {
 	dir := t.TempDir()
 	markers := make(map[string]bool)
+
 	for _, testCase := range []struct {
 		name      string
 		daemonDir string
@@ -517,9 +719,14 @@ func TestDaemonSeesDir(t *testing.T) {
 			ctx, cancel := context.WithCancel(t.Context())
 			defer cancel()
 			removed := false
+			var attemptedImages []string
 			cli := &probeClient{
-				create: func(opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+				create: func(_ctx context.Context, opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
 					assert.Equal(t, []mount.Mount{{Type: mount.TypeBind, Source: testCase.daemonDir, Target: "/gitea-runner-probe", ReadOnly: true}}, opts.HostConfig.Mounts)
+					attemptedImages = append(attemptedImages, opts.Config.Image)
+					if opts.Config.Image == "removed-image" {
+						return mobyclient.ContainerCreateResult{}, cerrdefs.ErrNotFound
+					}
 					cancel()
 					if _, err := os.Stat(testCase.daemonDir); err != nil {
 						return mobyclient.ContainerCreateResult{}, cerrdefs.ErrInvalidArgument
@@ -547,8 +754,10 @@ func TestDaemonSeesDir(t *testing.T) {
 			}
 			seen, err := daemonSeesDir(ctx, cli, dir, testCase.daemonDir)
 			require.ErrorIs(t, err, testCase.removeErr)
+			assert.Equal(t, []string{"removed-image", "probe-image"}, attemptedImages)
 			assert.Equal(t, testCase.seen, seen)
 			assert.Equal(t, testCase.removed, removed)
+
 			entries, err := os.ReadDir(dir)
 			require.NoError(t, err)
 			assert.Empty(t, entries)

act/container/docker_pull.go

@@ -114,7 +114,7 @@ func getImagePullOptions(ctx context.Context, input NewDockerPullExecutorInput)
 		if err != nil {
 			return imagePullOptions, err
 		}
-		if authConfig.Username == "" && authConfig.Password == "" {
+		if authConfig.Username == "" && authConfig.Password == "" && authConfig.IdentityToken == "" && authConfig.RegistryToken == "" {
 			return imagePullOptions, nil
 		}
 		logger.Info("using DockerAuthConfig authentication for docker pull")

act/container/docker_pull_test.go

@@ -6,12 +6,17 @@ package container
 
 import (
 	"io"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
+	"testing/iotest"
 
 	"gitea.com/gitea/runner/act/common/logging"
 	"gitea.com/gitea/runner/act/common/logging/logtest"
 
+	"github.com/moby/moby/api/pkg/authconfig"
+	"github.com/moby/moby/api/types/registry"
 	assert "github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -64,6 +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")
+
+	dir := t.TempDir()
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"auths":{"identity.example.com":{"identitytoken":"identity-token"},"registry.example.com":{"registrytoken":"registry-token"}}}`), 0o600))
+	t.Setenv("DOCKER_CONFIG", dir)
+	for host, want := range map[string]registry.AuthConfig{
+		"identity.example.com": {IdentityToken: "identity-token", ServerAddress: "identity.example.com"},
+		"registry.example.com": {RegistryToken: "registry-token", ServerAddress: "registry.example.com"},
+	} {
+		options, err := getImagePullOptions(ctx, NewDockerPullExecutorInput{Image: host + "/app"})
+		require.NoError(t, err)
+		require.NotEmpty(t, options.RegistryAuth, host)
+		got, err := authconfig.Decode(options.RegistryAuth)
+		require.NoError(t, err)
+		assert.Equal(t, want, *got)
+	}
 }
 
 // A digest-pinned image is immutable, so its local copy is always current.
@@ -82,4 +102,8 @@ func TestLogDockerResponseError(t *testing.T) {
 	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"}`))
+	require.NoError(t, streamErr(`{"stream":"`+strings.Repeat("x", 70000)+`"}`))
+	require.ErrorIs(t, logDockerResponse(logger, io.NopCloser(io.MultiReader(
+		strings.NewReader(`{"status":"Downloading"}`+"\n"), iotest.ErrReader(io.ErrUnexpectedEOF),
+	)), false), io.ErrUnexpectedEOF)
 }

act/container/docker_run.go

@@ -205,18 +205,30 @@ func (cr *containerReference) DumpLogs(ctx context.Context) error {
 	if cr.id == "" {
 		return cr.missingContainerError("read its logs")
 	}
+	tty, err := cr.containerTTY(ctx)
+	if err != nil {
+		return err
+	}
 
 	logs, err := cr.cli.ContainerLogs(ctx, cr.id, client.ContainerLogsOptions{ShowStdout: true, ShowStderr: true})
 	if err != nil {
 		return err
 	}
 	defer logs.Close()
-	return cr.copyOutput(logs)
+	return cr.copyOutput(logs, tty)
+}
+
+func (cr *containerReference) containerTTY(ctx context.Context) (bool, error) {
+	inspect, err := cr.cli.ContainerInspect(ctx, cr.id, client.ContainerInspectOptions{})
+	if err != nil {
+		return false, fmt.Errorf("failed to inspect container: %w", err)
+	}
+	return inspect.Container.Config != nil && inspect.Container.Config.Tty, nil
 }
 
 // copyOutput writes a container stream to the writers the container was created with,
 // demultiplexing it unless the container has a TTY, which sends a single raw stream.
-func (cr *containerReference) copyOutput(stream io.Reader) error {
+func (cr *containerReference) copyOutput(stream io.Reader, tty bool) error {
 	outWriter := cr.input.Stdout
 	if outWriter == nil {
 		outWriter = os.Stdout
@@ -227,7 +239,7 @@ func (cr *containerReference) copyOutput(stream io.Reader) error {
 	}
 
 	var err error
-	if !cr.input.AllocatePTY || os.Getenv("NORAW") != "" {
+	if !tty || os.Getenv("NORAW") != "" {
 		_, err = stdcopy.StdCopy(outWriter, errWriter, stream)
 	} else {
 		_, err = io.Copy(outWriter, stream)
@@ -296,7 +308,7 @@ func containerInfoFromInspect(inspect container.InspectResponse) *Info {
 }
 
 func (cr *containerReference) UpdateFromEnv(srcPath string, env *map[string]string) common.Executor {
-	return parseEnvFile(cr, srcPath, env).IfNot(common.Dryrun)
+	return parseEnvFile(cr, srcPath, env, false).IfNot(common.Dryrun)
 }
 
 func (cr *containerReference) UpdateFromImageEnv(env *map[string]string) common.Executor {
@@ -499,7 +511,7 @@ func (cr *containerReference) remove() common.Executor {
 		if idOrName == "" && cr.input != nil {
 			idOrName = cr.input.Name
 		}
-		if idOrName == "" {
+		if idOrName == "" || cr.cli == nil {
 			return nil
 		}
 
@@ -851,7 +863,7 @@ func (cr *containerReference) waitForCommand(ctx context.Context, resp client.Hi
 	cmdResponse := make(chan error, 1)
 
 	go func() {
-		cmdResponse <- cr.copyOutput(resp.Reader)
+		cmdResponse <- cr.copyOutput(resp.Reader, cr.input.AllocatePTY)
 	}()
 
 	select {
@@ -867,7 +879,10 @@ func (cr *containerReference) waitForCommand(ctx context.Context, resp client.Hi
 		// truncate the tail of the log. The goroutine exits once the hijacked
 		// stream is closed by resp.Close() in the caller's defer.
 		select {
-		case <-cmdResponse:
+		case err := <-cmdResponse:
+			if err != nil {
+				logger.Debugf("Failed to drain command output after cancellation: %v", err)
+			}
 		case <-time.After(drainGracePeriod):
 			logger.Warn("Timed out draining command output after cancellation")
 		}
@@ -876,11 +891,7 @@ func (cr *containerReference) waitForCommand(ctx context.Context, resp client.Hi
 		// from executing
 		return ctx.Err()
 	case err := <-cmdResponse:
-		if err != nil {
-			logger.Error(err)
-		}
-
-		return nil
+		return err
 	}
 }
 
@@ -897,7 +908,7 @@ func (cr *containerReference) copyDir(dstPath, srcPath string, useGitIgnore, ski
 		}
 		logger.Debugf("Writing tarball %s from %s", tarFile.Name(), srcPath)
 		defer func() {
-			if err := tarFile.Close(); !errors.Is(err, os.ErrClosed) {
+			if err := tarFile.Close(); err != nil && !errors.Is(err, os.ErrClosed) {
 				logger.Error(err)
 			}
 			if err := os.Remove(tarFile.Name()); err != nil {
@@ -944,6 +955,10 @@ func (cr *containerReference) upload(ctx context.Context, dstPath string, conten
 
 func (cr *containerReference) attach() common.Executor {
 	return func(ctx context.Context) error {
+		tty, err := cr.containerTTY(ctx)
+		if err != nil {
+			return err
+		}
 		out, err := cr.cli.ContainerAttach(ctx, cr.id, client.ContainerAttachOptions{
 			Stream: true,
 			Stdout: true,
@@ -954,9 +969,12 @@ func (cr *containerReference) attach() common.Executor {
 		}
 		done := make(chan struct{})
 		cr.attachDone = done
+		stopClose := context.AfterFunc(ctx, out.Close)
 		go func() {
 			defer close(done)
-			if copyErr := cr.copyOutput(out.Reader); copyErr != nil {
+			defer out.Close()
+			defer stopClose()
+			if copyErr := cr.copyOutput(out.Reader, tty); copyErr != nil && ctx.Err() == nil {
 				common.Logger(ctx).Error(copyErr)
 			}
 		}()
@@ -991,6 +1009,9 @@ func (cr *containerReference) wait() common.Executor {
 				return fmt.Errorf("failed to wait for container: %w", err)
 			}
 		case status := <-waitResult.Result:
+			if status.Error != nil {
+				return fmt.Errorf("failed to wait for container: %s", status.Error.Message)
+			}
 			statusCode = status.StatusCode
 		}
 

act/container/docker_run_test.go

@@ -92,6 +92,16 @@ func (m *mockDockerClient) ContainerAttach(ctx context.Context, containerID stri
 	return args.Get(0).(mobyclient.ContainerAttachResult), args.Error(1)
 }
 
+func (m *mockDockerClient) ContainerLogs(ctx context.Context, containerID string, opts mobyclient.ContainerLogsOptions) (mobyclient.ContainerLogsResult, error) {
+	args := m.Called(ctx, containerID, opts)
+	return args.Get(0).(mobyclient.ContainerLogsResult), args.Error(1)
+}
+
+func (m *mockDockerClient) ContainerStart(ctx context.Context, containerID string, opts mobyclient.ContainerStartOptions) (mobyclient.ContainerStartResult, error) {
+	args := m.Called(ctx, containerID, opts)
+	return args.Get(0).(mobyclient.ContainerStartResult), args.Error(1)
+}
+
 func (m *mockDockerClient) ContainerWait(ctx context.Context, containerID string, opts mobyclient.ContainerWaitOptions) mobyclient.ContainerWaitResult {
 	args := m.Called(ctx, containerID, opts)
 	return args.Get(0).(mobyclient.ContainerWaitResult)
@@ -158,6 +168,7 @@ func (r *interruptReader) Read(_ []byte) (int, error) {
 type mockConn struct {
 	net.Conn
 	mock.Mock
+	closed atomic.Bool
 }
 
 func (m *mockConn) Write(b []byte) (n int, err error) {
@@ -166,6 +177,7 @@ func (m *mockConn) Write(b []byte) (n int, err error) {
 }
 
 func (m *mockConn) Close() (err error) {
+	m.closed.Store(true)
 	return nil
 }
 
@@ -236,6 +248,21 @@ func TestDockerExecFailure(t *testing.T) {
 	client.AssertExpectations(t)
 }
 
+func TestDockerExecStreamFailure(t *testing.T) {
+	ctx := t.Context()
+	client := &mockDockerClient{}
+	client.On("ExecCreate", ctx, "123", mock.AnythingOfType("client.ExecCreateOptions")).
+		Return(mobyclient.ExecCreateResult{ID: "id"}, nil)
+	client.On("ExecAttach", ctx, "id", mock.AnythingOfType("client.ExecAttachOptions")).
+		Return(mobyclient.ExecAttachResult{
+			Conn:   &mockConn{},
+			Reader: bufio.NewReader(bytes.NewReader(stdcopyFrame(stdcopy.Systemerr, "unable to start exec"))),
+		}, nil)
+	require.ErrorContains(t, (&containerReference{id: "123", cli: client, input: &NewContainerInput{}}).
+		exec([]string{"sh"}, nil, "", "")(ctx), "unable to start exec")
+	client.AssertExpectations(t)
+}
+
 // stdcopyFrame wraps payload in a single Docker multiplexed-stream frame, the
 // format StdCopy expects: an 8-byte header (stream type + 4-byte big-endian
 // length) followed by the payload.
@@ -248,60 +275,88 @@ func stdcopyFrame(stream stdcopy.StdType, payload string) []byte {
 }
 
 func TestDockerAttachFlushesTrailingLine(t *testing.T) {
+	for _, testcase := range []struct {
+		name string
+		tty  bool
+	}{
+		{name: "multiplexed"},
+		{name: "tty from container options", tty: true},
+	} {
+		t.Run(testcase.name, func(t *testing.T) {
 			ctx := t.Context()
-
-	framed := bytes.NewBuffer(stdcopyFrame(stdcopy.Stdout, "line one\nlast line without newline"))
-
+			output := []byte("line one\nlast line without newline")
+			if !testcase.tty {
+				output = stdcopyFrame(stdcopy.Stdout, string(output))
+			}
 			var lines []string
-	logWriter := common.NewLineWriter(func(s string) bool {
-		lines = append(lines, s)
+			logWriter := common.NewLineWriter(func(line string) bool {
+				lines = append(lines, line)
 				return true
 			})
-
 			client := &mockDockerClient{}
+			client.On("ContainerInspect", ctx, "123", mobyclient.ContainerInspectOptions{}).
+				Return(mobyclient.ContainerInspectResult{Container: container.InspectResponse{Config: &container.Config{Tty: testcase.tty}}}, nil)
+			conn := &mockConn{}
 			client.On("ContainerAttach", ctx, "123", mock.AnythingOfType("client.ContainerAttachOptions")).
 				Return(mobyclient.ContainerAttachResult{
-			Conn:   &mockConn{},
-			Reader: bufio.NewReader(framed),
+					Conn:   conn,
+					Reader: bufio.NewReader(bytes.NewReader(output)),
 				}, nil)
-
+			client.On("ContainerLogs", ctx, "123", mobyclient.ContainerLogsOptions{ShowStdout: true, ShowStderr: true}).
+				Return(io.NopCloser(bytes.NewReader(output)), nil)
 			statusCh := make(chan container.WaitResponse, 1)
 			statusCh <- container.WaitResponse{StatusCode: 0}
 			client.On("ContainerWait", ctx, "123", mobyclient.ContainerWaitOptions{Condition: container.WaitConditionNotRunning}).
 				Return(mobyclient.ContainerWaitResult{Result: statusCh})
-
-	cr := &containerReference{
-		id:  "123",
-		cli: client,
-		input: &NewContainerInput{
-			Stdout: logWriter,
-			Stderr: logWriter,
-		},
-	}
+			cr := &containerReference{id: "123", cli: client, input: &NewContainerInput{Stdout: logWriter, Stderr: logWriter}}
 
 			require.NoError(t, cr.attach()(ctx))
 			require.NoError(t, cr.wait()(ctx))
-
 			assert.Equal(t, []string{"line one\n", "last line without newline"}, lines)
+			assert.True(t, conn.closed.Load())
+			require.NoError(t, cr.DumpLogs(ctx))
+			assert.Equal(t, []string{"line one\n", "last line without newline", "line one\n", "last line without newline"}, lines)
+			client.AssertExpectations(t)
+		})
+	}
+}
 
+func TestDockerAttachStreamClosesWithItsContext(t *testing.T) {
+	conn, peer := net.Pipe()
+	t.Cleanup(func() { _ = conn.Close() })
+	t.Cleanup(func() { _ = peer.Close() })
+	client := &mockDockerClient{}
+	client.On("ContainerInspect", mock.Anything, "123", mobyclient.ContainerInspectOptions{}).
+		Return(mobyclient.ContainerInspectResult{Container: container.InspectResponse{Config: &container.Config{}}}, nil)
+	client.On("ContainerAttach", mock.Anything, "123", mock.AnythingOfType("client.ContainerAttachOptions")).
+		Return(mobyclient.ContainerAttachResult{HijackedResponse: mobyclient.NewHijackedResponse(conn, "")}, nil)
+	client.On("ContainerStart", mock.Anything, "123", mobyclient.ContainerStartOptions{}).
+		Return(mobyclient.ContainerStartResult{}, errors.New("container start failed"))
+	cr := &containerReference{id: "123", cli: client, input: &NewContainerInput{}}
+
+	ctx, cancel := context.WithCancel(t.Context())
+	require.ErrorContains(t, cr.Start(true)(ctx), "container start failed")
+	cancel()
+	<-cr.attachDone
 	client.AssertExpectations(t)
 }
 
 func TestDockerWaitFailure(t *testing.T) {
 	ctx := t.Context()
-
 	statusCh := make(chan container.WaitResponse, 1)
 	statusCh <- container.WaitResponse{StatusCode: 2}
 	client := &mockDockerClient{}
 	client.On("ContainerWait", ctx, "123", mobyclient.ContainerWaitOptions{Condition: container.WaitConditionNotRunning}).
 		Return(mobyclient.ContainerWaitResult{Result: statusCh})
+	cr := &containerReference{id: "123", cli: client}
 
-	err := (&containerReference{id: "123", cli: client}).wait()(ctx)
+	err := cr.wait()(ctx)
 	var exitErr ExitCodeError
 	require.ErrorAs(t, err, &exitErr)
 	assert.Equal(t, ExitCodeError(2), exitErr)
 	assert.Equal(t, "Process completed with exit code 2.", err.Error())
-
+	statusCh <- container.WaitResponse{Error: &container.WaitExitError{Message: "container wait failed"}}
+	require.EqualError(t, cr.wait()(ctx), "failed to wait for container: container wait failed")
 	client.AssertExpectations(t)
 }
 
@@ -370,6 +425,13 @@ func TestRemoveWithoutIDUsesName(t *testing.T) {
 	client.AssertExpectations(t)
 }
 
+func TestRemovePreservesClientCreationError(t *testing.T) {
+	t.Setenv("DOCKER_HOST", "://")
+
+	require.ErrorContains(t, (&containerReference{input: &NewContainerInput{Name: "job-1"}}).
+		Remove()(t.Context()), "failed to connect to docker daemon")
+}
+
 func TestFindRevalidatesStaleID(t *testing.T) {
 	ctx := t.Context()
 	inspectOpts := mobyclient.ContainerInspectOptions{}
@@ -477,7 +539,8 @@ func TestPublicCopyPipelineHandlesStaleID(t *testing.T) {
 		{name: "sibling prefix", directory: true, destination: "/var/run/actor/action", extractionRoot: "/", entry: "var/run/actor/action/x"},
 	} {
 		t.Run(testcase.name, func(t *testing.T) {
-			ctx := t.Context()
+			hook := logtest.NewWithLevel(logging.LevelError)
+			ctx := common.WithLogger(t.Context(), logging.New(hook))
 			client := &mockDockerClient{}
 			client.On("ContainerInspect", ctx, "stale", mobyclient.ContainerInspectOptions{}).
 				Return(mobyclient.ContainerInspectResult{}, cerrdefs.ErrNotFound.WithMessage("gone"))
@@ -501,6 +564,7 @@ func TestPublicCopyPipelineHandlesStaleID(t *testing.T) {
 				require.NoError(t, cr.Copy(testcase.destination, &FileEntry{Name: "x", Mode: 0o644})(ctx))
 			}
 			assert.Equal(t, "fresh", cr.id)
+			assert.Empty(t, hook.AllRecords())
 			client.AssertExpectations(t)
 		})
 	}
@@ -916,7 +980,7 @@ func TestDockerCreateKeepsInputEntrypoint(t *testing.T) {
 					RunnerOptions:   "--entrypoint=runner-entrypoint",
 					WorkflowOptions: "--entrypoint=" + optionEntrypoint + " --init",
 				},
-				cli: &probeClient{create: func(opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
+				cli: &probeClient{create: func(_ctx context.Context, opts mobyclient.ContainerCreateOptions) (mobyclient.ContainerCreateResult, error) {
 					if len(entrypoint) > 0 {
 						assert.Equal(t, entrypoint, opts.Config.Entrypoint)
 					} else {

act/container/docker_ssh_test.go

@@ -30,6 +30,7 @@ func TestSSHArgs(t *testing.T) {
 }
 
 func TestDialCommandStreamsStdioAndReportsExit(t *testing.T) {
+	t.Parallel()
 	if _, err := exec.LookPath("sh"); err != nil {
 		t.Skip("needs a POSIX shell")
 	}

act/container/host_environment.go

@@ -225,12 +225,16 @@ func setupPty(cmd *exec.Cmd, cmdline string) (*os.File, *os.File, error) {
 	return ppty, tty, nil
 }
 
-func writeKeepAlive(ppty io.Writer) {
-	c := 1
-	var err error
-	for c == 1 && err == nil {
-		c, err = ppty.Write([]byte{4})
-		<-time.After(time.Second)
+func writeKeepAlive(ctx context.Context, ppty io.Writer) {
+	for ctx.Err() == nil {
+		if count, err := ppty.Write([]byte{4}); count != 1 || err != nil {
+			return
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(time.Second):
+		}
 	}
 }
 
@@ -302,7 +306,7 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
 			defer finishLog()
 			_, _ = io.Copy(writer, ppty)
 		}()
-		go writeKeepAlive(ppty)
+		go writeKeepAlive(logctx, ppty)
 	}
 	if err := cmd.Start(); err != nil {
 		return err
@@ -318,21 +322,21 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
 		defer k.Close()
 	}
 	err = treeKill.Wait(ctx, cmd)
-	if err != nil {
-		if exitErr, ok := errors.AsType[*exec.ExitError](err); ok {
-			return ExitCodeError(exitErr.ExitCode())
-		}
-		return err
-	}
 	if tty != nil {
 		writer.AutoStop.Store(true)
 		if _, err := tty.WriteString("\x04"); err != nil {
 			common.Logger(ctx).Debug("Failed to write EOT")
 		}
-		<-logctx.Done()
+		select {
+		case <-logctx.Done():
+		case <-ctx.Done(): // a cancelled step does not wait on output nobody consumes
+		}
 		ppty.Close()
 		ppty = nil
 	}
+	if exitErr, ok := errors.AsType[*exec.ExitError](err); ok {
+		return ExitCodeError(exitErr.ExitCode())
+	}
 	return err
 }
 
@@ -355,7 +359,7 @@ func (e *HostEnvironment) ExecWithCmdLine(command []string, cmdline string, env
 }
 
 func (e *HostEnvironment) UpdateFromEnv(srcPath string, env *map[string]string) common.Executor {
-	return parseEnvFile(e, srcPath, env)
+	return parseEnvFile(e, srcPath, env, runtime.GOOS == "windows")
 }
 
 // removeAll is a var so tests can substitute a blocking stub.

act/container/host_environment_test.go

@@ -14,6 +14,7 @@ import (
 	"runtime"
 	"strings"
 	"testing"
+	"testing/synctest"
 	"time"
 
 	"gitea.com/gitea/runner/act/common"
@@ -79,20 +80,48 @@ func TestGetContainerArchive(t *testing.T) {
 }
 
 func TestHostEnvironmentExecExitCode(t *testing.T) {
+	t.Parallel()
 	if runtime.GOOS == "windows" {
 		t.Skip("uses POSIX shell")
 	}
-	ctx := t.Context()
-	e := &HostEnvironment{Path: t.TempDir(), StdOut: io.Discard}
+	for name, allocatePTY := range map[string]bool{"without PTY": false, "with PTY": true} {
+		t.Run(name, func(t *testing.T) {
+			t.Parallel()
+			synctest.Test(t, func(t *testing.T) {
+				reader, writer := io.Pipe()
+				defer reader.Close()
+				e := &HostEnvironment{Path: t.TempDir(), StdOut: writer, AllocatePTY: allocatePTY}
+				t.Cleanup(func() { assert.NoError(t, e.Remove()(t.Context())) })
 
-	err := e.Exec([]string{"sh", "-c", "exit 3"}, map[string]string{"PATH": os.Getenv("PATH")}, "", "")(ctx)
+				result := make(chan error, 1)
+				go func() {
+					defer writer.Close()
+					result <- e.Exec([]string{"sh", "-c", "printf 'failed\\n'; exit 3"}, map[string]string{"PATH": os.Getenv("PATH")}, "", "")(t.Context())
+				}()
+				first := make([]byte, 1)
+				_, err := io.ReadFull(reader, first)
+				require.NoError(t, err)
+				synctest.Wait()
+				select {
+				case err := <-result:
+					t.Fatalf("Exec returned before its output was consumed: %v", err)
+				default:
+				}
+				remaining, err := io.ReadAll(reader)
+				require.NoError(t, err)
+				err = <-result
 				var exitErr ExitCodeError
 				require.ErrorAs(t, err, &exitErr)
 				assert.Equal(t, ExitCodeError(3), exitErr)
 				assert.Equal(t, "Process completed with exit code 3.", err.Error())
+				assert.Equal(t, "failed\n", string(first)+string(remaining))
+			})
+		})
+	}
 }
 
 func TestHostEnvironmentAllocatePTY(t *testing.T) {
+	t.Parallel()
 	if runtime.GOOS == "windows" {
 		t.Skip("uses POSIX shell")
 	}
@@ -105,6 +134,7 @@ func TestHostEnvironmentAllocatePTY(t *testing.T) {
 		{name: "on", allocPTY: true, expect: "TTY"},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
 			buf := &bytes.Buffer{}
 			e := &HostEnvironment{
 				Path:        t.TempDir(),
@@ -146,6 +176,7 @@ func TestHostEnvironmentRemoveWorkdir(t *testing.T) {
 }
 
 func TestRemoveAllWithContextDoesNotHangOnStuckDelete(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
 		release := make(chan struct{})
 		stubDone := make(chan struct{})
 
@@ -166,11 +197,13 @@ func TestRemoveAllWithContextDoesNotHangOnStuckDelete(t *testing.T) {
 
 		err := removeAllWithContext(ctx, t.TempDir())
 		require.ErrorIs(t, err, context.DeadlineExceeded)
+	})
 }
 
 func TestHostEnvironmentRemoveDoesNotHang(t *testing.T) {
 	for _, phase := range []string{"CleanUp callback", "path removal"} {
 		t.Run(phase, func(t *testing.T) {
+			synctest.Test(t, func(t *testing.T) {
 				origTimeout := hostCleanupTimeout
 				hostCleanupTimeout = 100 * time.Millisecond
 				t.Cleanup(func() { hostCleanupTimeout = origTimeout })
@@ -204,6 +237,7 @@ func TestHostEnvironmentRemoveDoesNotHang(t *testing.T) {
 					t.Fatalf("Remove() hung on a stuck %s", phase)
 				}
 			})
+		})
 	}
 }
 

act/container/kubernetes.go

@@ -7,12 +7,15 @@ import (
 	"bufio"
 	"bytes"
 	"cmp"
+	"compress/gzip"
 	"context"
 	"crypto/rand"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
 	"maps"
+	"net"
 	"net/http"
 	"net/url"
 	"os"
@@ -23,6 +26,7 @@ import (
 	"time"
 
 	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/common/logging"
 
 	"github.com/kballard/go-shellquote"
 	"github.com/spf13/pflag"
@@ -36,6 +40,7 @@ type KubernetesOptions struct {
 	MaxLifetime  time.Duration
 	ForcePull    bool
 	RunnerUUID   string
+	Transfer     *KubernetesTransfer
 }
 
 // KubernetesPod runs a job container and its service containers as one pod.
@@ -46,6 +51,10 @@ type KubernetesPod struct {
 	job              *kubernetesContainer
 	services         []*kubernetesContainer
 	arch, path       string
+	ips              []net.IP
+	transfer         *transferServer
+	transferImage    string
+	transferHelper   string // empty when copies go through the API server
 }
 
 type kubernetesContainer struct {
@@ -55,6 +64,7 @@ type kubernetesContainer struct {
 	name  string
 	input *NewContainerInput
 	probe map[string]any
+	gzip  bool // the container can gunzip, so copies into it go compressed
 }
 
 type containerStatus struct {
@@ -74,6 +84,7 @@ type podStatus struct {
 	Spec   struct{ Containers []struct{} }
 	Status struct {
 		Phase, Reason, Message                   string
+		PodIPs                                   []struct{ IP string }
 		Conditions                               []struct{ Type, Status, Message string }
 		InitContainerStatuses, ContainerStatuses []containerStatus
 	}
@@ -129,14 +140,24 @@ func (p *KubernetesPod) create(ctx context.Context) error {
 		labels[runnerUUIDLabel] = p.options.RunnerUUID
 	}
 	metadata := map[string]any{"name": name, "labels": labels}
+	var runnerImage string
 	if hostname, err := os.Hostname(); p.client.inCluster && err == nil {
-		var runner struct{ Metadata struct{ Name, UID string } }
+		var runner struct {
+			Metadata struct{ Name, UID string }
+			Spec     struct{ Containers []struct{ Image string } }
+		}
 		if err := p.client.do(ctx, http.MethodGet, p.client.path("pods/"+hostname), nil, &runner); err == nil {
 			metadata["ownerReferences"] = []any{map[string]any{"apiVersion": "v1", "kind": "Pod", "name": runner.Metadata.Name, "uid": runner.Metadata.UID}}
+			if len(runner.Spec.Containers) == 1 {
+				runnerImage = runner.Spec.Containers[0].Image
+			}
 		} else {
 			common.Logger(ctx).Debugf("The runner is not pod %s, job pods get no owner: %v", hostname, err)
 		}
 	}
+	if err := p.setupTransfer(runnerImage); err != nil {
+		return err
+	}
 	secretData := map[string]string{}
 	manifest, err := p.manifest(name, secretData, metadata)
 	if err != nil {
@@ -205,6 +226,18 @@ func (p *KubernetesPod) manifest(secret string, secretData map[string]string, me
 		}
 	}
 	spec := map[string]any{"restartPolicy": "Never", "containers": containers, "volumes": volumes}
+	if p.transferHelper != "" {
+		transfer := map[string]any{
+			"name": "gitea-runner-transfer", "image": p.transferImage, "terminationMessagePolicy": "FallbackToLogsOnError",
+			"command":         []string{"gitea-runner", KubernetesTransferCommand, "install", p.transferHelper},
+			"volumeMounts":    []any{map[string]any{"name": "act", "mountPath": p.job.GetActPath()}},
+			"securityContext": map[string]any{"allowPrivilegeEscalation": false, "capabilities": map[string]any{"drop": []string{"ALL"}}},
+		}
+		if p.options.ForcePull {
+			transfer["imagePullPolicy"] = "Always"
+		}
+		spec["initContainers"] = []any{transfer}
+	}
 	if len(hostnames) > 0 {
 		spec["hostAliases"] = []any{map[string]any{"ip": "127.0.0.1", "hostnames": hostnames}}
 	}
@@ -341,7 +374,13 @@ func (p *KubernetesPod) waitStarted(ctx context.Context) error {
 			return err
 		}
 		if pod.Status.Phase == "Failed" {
-			return fmt.Errorf("pod %s failed: %s", p.name, cmp.Or(pod.Status.Message, pod.Status.Reason))
+			reason := cmp.Or(pod.Status.Message, pod.Status.Reason)
+			for _, status := range append(pod.Status.InitContainerStatuses, pod.Status.ContainerStatuses...) {
+				if terminated := status.State.Terminated; reason == "" && terminated != nil && terminated.ExitCode != 0 {
+					reason = fmt.Sprintf("container %s exited with %d: %s", status.Name, terminated.ExitCode, strings.TrimSpace(terminated.Message))
+				}
+			}
+			return fmt.Errorf("pod %s failed: %s", p.name, reason)
 		}
 		started := len(pod.Status.ContainerStatuses) == len(pod.Spec.Containers)
 		for _, status := range append(pod.Status.InitContainerStatuses, pod.Status.ContainerStatuses...) {
@@ -354,6 +393,9 @@ func (p *KubernetesPod) waitStarted(ctx context.Context) error {
 			}
 		}
 		if started {
+			for _, ip := range pod.Status.PodIPs {
+				p.ips = append(p.ips, net.ParseIP(ip.IP))
+			}
 			return nil
 		}
 		for _, condition := range pod.Status.Conditions {
@@ -391,8 +433,11 @@ func (c *kubernetesContainer) Start(_ bool) common.Executor {
 		if err != nil {
 			return err
 		}
-		machine, path, _ := strings.Cut(output, "\n")
-		c.pod.arch, c.pod.path = goArchToActionArch(machine), path
+		probed := strings.SplitN(output, "\n", 3)
+		for len(probed) < 3 {
+			probed = append(probed, "")
+		}
+		c.pod.arch, c.gzip, c.pod.path = goArchToActionArch(probed[0]), probed[1] == "gzip", probed[2]
 		return nil
 	}
 }
@@ -403,7 +448,7 @@ func (p *KubernetesPod) probe(ctx context.Context, container string) (string, er
 	deadline := time.Now().Add(time.Minute)
 	for delay := 100 * time.Millisecond; ; delay = min(2*delay, 2*time.Second) {
 		var stdout bytes.Buffer
-		err := p.client.exec(ctx, p.name, container, []string{"sh", "-c", `printf '%s\n%s' "$(uname -m)" "$PATH"`}, nil, &stdout, io.Discard)
+		err := p.client.exec(ctx, p.name, container, []string{"sh", "-c", `printf '%s\n%s\n%s' "$(uname -m)" "$(command -v gzip >/dev/null && echo gzip)" "$PATH"`}, nil, &stdout, io.Discard)
 		var exitErr ExitCodeError
 		var apiErr *apiError
 		if err == nil || ctx.Err() != nil || errors.As(err, &exitErr) || time.Now().After(deadline) ||
@@ -452,26 +497,66 @@ func (c *kubernetesContainer) Exec(command []string, env map[string]string, _, w
 	}
 }
 
-func (c *kubernetesContainer) extract(ctx context.Context, destPath string, write func(io.Writer) error) error {
+// compressCopyMin is the size from which Copy compresses and uses the transfer helper: below it that saves next to nothing on the wire and costs a process in the pod.
+const compressCopyMin = 64 << 10
+
+// extract streams the tar write produces into destPath, gzipped with compress, as copied actions shrink about fivefold and API server traffic often crosses zones.
+func (c *kubernetesContainer) extract(ctx context.Context, destPath string, compress bool, write func(io.Writer) error) error {
+	if compress && c.pod.transferHelper != "" {
+		request, done := c.pod.transfer.offer(c.pod.ips, write)
+		stdin, _ := json.Marshal(request)
+		var stderr bytes.Buffer
+		err := c.pod.client.exec(ctx, c.pod.name, c.name, []string{c.pod.transferHelper, KubernetesTransferCommand, "fetch", destPath}, bytes.NewReader(stdin), io.Discard, &stderr)
+		taken, writeErr := done()
+		if err == nil || ctx.Err() != nil {
+			return err
+		}
+		if taken || !c.pod.options.Transfer.Fallback {
+			return fmt.Errorf("copy to %s over the pod network: %w: %s", destPath, errors.Join(writeErr, err), strings.TrimSpace(stderr.String()))
+		}
+		c.pod.transferHelper = "" // a pod that cannot reach the runner once will not later
+		logging.Warnf("Copying into pod %s over the pod network failed, copying through the API server instead: %v: %s", c.pod.name, err, strings.TrimSpace(stderr.String()))
+	}
+	compress = compress && c.gzip
 	reader, writer := io.Pipe()
 	defer reader.Close()
-	go func() { writer.CloseWithError(write(writer)) }()
+	go func() {
+		if compress {
+			writer.CloseWithError(writeGzip(writer, write))
+		} else {
+			writer.CloseWithError(write(writer))
+		}
+	}()
+	script := `mkdir -p "$1" && tar -xf - -C "$1"`
+	if compress {
+		script = `mkdir -p "$1" && gzip -dc | tar -xf - -C "$1"` // gzip -dc rather than tar -z, which GNU tar runs gzip for anyway
+	}
 	var stderr bytes.Buffer
-	if err := c.pod.client.exec(ctx, c.pod.name, c.name, []string{"sh", "-c", `mkdir -p "$1" && tar -xf - -C "$1"`, "sh", destPath}, reader, io.Discard, &stderr); err != nil {
+	if err := c.pod.client.exec(ctx, c.pod.name, c.name, []string{"sh", "-c", script, "sh", destPath}, reader, io.Discard, &stderr); err != nil {
 		return fmt.Errorf("extract to %s: %w: %s", destPath, err, strings.TrimSpace(stderr.String()))
 	}
 	return nil
 }
 
+// writeGzip gzips what write produces into w, at a level fast enough to not slow copies down and nearly as small as the default.
+func writeGzip(w io.Writer, write func(io.Writer) error) error {
+	compressed, _ := gzip.NewWriterLevel(w, gzip.BestSpeed)
+	return errors.Join(write(compressed), compressed.Close())
+}
+
 func (c *kubernetesContainer) Copy(destPath string, files ...*FileEntry) common.Executor {
 	return func(ctx context.Context) error {
-		return c.extract(ctx, destPath, func(writer io.Writer) error { return writeFilesTar(ctx, writer, 0, 0, files...) })
+		size := 0
+		for _, file := range files {
+			size += len(file.Body)
+		}
+		return c.extract(ctx, destPath, size >= compressCopyMin, func(writer io.Writer) error { return writeFilesTar(ctx, writer, 0, 0, files...) })
 	}
 }
 
 func (c *kubernetesContainer) CopyDir(destPath, srcPath string, useGitIgnore, skipGitDir bool) common.Executor {
 	return func(ctx context.Context) error {
-		return c.extract(ctx, destPath, func(writer io.Writer) error {
+		return c.extract(ctx, destPath, true, func(writer io.Writer) error { // directories are actions and workspaces, megabytes of compressible sources
 			return writeDirTar(ctx, writer, "/", srcPath, useGitIgnore, skipGitDir, 0, 0)
 		})
 	}
@@ -500,7 +585,7 @@ func (c *kubernetesContainer) GetContainerArchive(ctx context.Context, srcPath s
 }
 
 func (c *kubernetesContainer) UpdateFromEnv(srcPath string, env *map[string]string) common.Executor {
-	return parseEnvFile(c, srcPath, env)
+	return parseEnvFile(c, srcPath, env, false)
 }
 
 func (c *kubernetesContainer) UpdateFromImageEnv(env *map[string]string) common.Executor {

act/container/kubernetes_test.go

@@ -89,7 +89,7 @@ func execFailure(t *testing.T, code int, status string) http.HandlerFunc {
 	}
 }
 
-func fakeKubernetes(t *testing.T, podStatus string) (string, *fakeCluster) {
+func fakeKubernetes(t *testing.T, podStatus string, env ...string) (string, *fakeCluster) {
 	cluster := &fakeCluster{requests: map[string][]string{}}
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if r.Header.Get("Authorization") != "Bearer test-token" {
@@ -109,7 +109,7 @@ func fakeKubernetes(t *testing.T, podStatus string) (string, *fakeCluster) {
 			if failure := cluster.nextExecFailure(); failure != nil {
 				failure(w, r)
 			} else {
-				serveLocalExec(t, w, r)
+				serveLocalExec(t, w, r, env)
 			}
 		case strings.HasSuffix(r.URL.Path, "/log"):
 			_, _ = io.WriteString(w, "log of "+r.URL.Query().Get("container"))
@@ -132,7 +132,7 @@ users: [{name: test, user: {tokenFile: token}}]
 	return kubeconfig, cluster
 }
 
-func serveLocalExec(t *testing.T, w http.ResponseWriter, r *http.Request) {
+func serveLocalExec(t *testing.T, w http.ResponseWriter, r *http.Request, env []string) {
 	conn, err := websocket.Accept(w, r, &websocket.AcceptOptions{Subprotocols: []string{execProtocol}})
 	if err != nil {
 		t.Error(err)
@@ -154,7 +154,7 @@ func serveLocalExec(t *testing.T, w http.ResponseWriter, r *http.Request) {
 		stdin.Write(frame[1:])
 	}
 	cmd := exec.Command(command[0], command[1:]...)
-	cmd.Env = append(os.Environ(), "COPYFILE_DISABLE=1")
+	cmd.Env = append(append(os.Environ(), "COPYFILE_DISABLE=1"), env...)
 	var stdout, stderr bytes.Buffer
 	cmd.Stdin, cmd.Stdout, cmd.Stderr = &stdin, &stdout, &stderr
 	status := `{"status":"Success"}`
@@ -183,15 +183,17 @@ func TestKubernetesServiceHostnames(t *testing.T) {
 }
 
 func TestKubernetesPodRunsJobThroughAPI(t *testing.T) {
+	t.Parallel()
 	if runtime.GOOS == "windows" {
 		t.Skip("uses POSIX shell")
 	}
 	ctx := t.Context()
-	kubeconfig, cluster := fakeKubernetes(t, `{"spec":{"containers":[{},{}]},"status":{"phase":"Running","containerStatuses":[
-		{"name":"job","ready":true,"state":{"running":{}}},
-		{"name":"my-db","ready":true,"state":{"running":{}}}]}}`)
 	workdir, err := filepath.EvalSymlinks(t.TempDir())
 	require.NoError(t, err)
+	path := workdir + ":" + os.Getenv("PATH")
+	kubeconfig, cluster := fakeKubernetes(t, `{"spec":{"containers":[{},{}]},"status":{"phase":"Running","containerStatuses":[
+		{"name":"job","ready":true,"state":{"running":{}}},
+		{"name":"my-db","ready":true,"state":{"running":{}}}]}}`, "PATH="+path)
 	options := KubernetesOptions{Kubeconfig: kubeconfig, RunnerUUID: "runner-uuid", PodTemplates: []map[string]any{
 		{"spec": map[string]any{
 			"automountServiceAccountToken": true,
@@ -252,13 +254,13 @@ func TestKubernetesPodRunsJobThroughAPI(t *testing.T) {
 	assert.Contains(t, cluster.requests["POST /api/v1/namespaces/ci/pods"][0], `"args":["postgres","-D","$$(PGDATA)"]`)
 
 	require.NoError(t, os.WriteFile(filepath.Join(workdir, "env"), []byte("#!/bin/sh\necho \"$@\" > \"$0-args\"\nexec /usr/bin/env \"$@\"\n"), 0o755))
-	t.Setenv("PATH", workdir+":"+os.Getenv("PATH"))
 	cluster.failExec(
 		execFailure(t, http.StatusInternalServerError, `{"message":"error dialing backend: remote error: tls: internal error"}`),
 		execFailure(t, 0, `{"status":"Failure","reason":"InternalError","message":"Internal error occurred: error sending request: remote error: tls: internal error"}`),
 	)
 	require.NoError(t, job.Start(false)(ctx))
 	assert.Len(t, cluster.requests["exec"], 3)
+	assert.True(t, pod.job.gzip)
 
 	stepEnv := map[string]string{"INPUT_WHO-TO-GREET": "it's\nme", "SECRET": "hunter2", "UID": "5"}
 	require.NoError(t, job.Exec([]string{"printenv", "SECRET"}, stepEnv, "", "")(ctx))
@@ -275,7 +277,9 @@ func TestKubernetesPodRunsJobThroughAPI(t *testing.T) {
 		assert.NotContains(t, query, "hunter2")
 	}
 
+	lastExec := func() string { return cluster.requests["exec"][len(cluster.requests["exec"])-1] }
 	require.NoError(t, job.Copy(workdir+"/act/", &FileEntry{Name: "workflow/event.json", Mode: 0o644, Body: `{"ref":"main"}`})(ctx))
+	assert.NotContains(t, lastExec(), "gzip")
 	archive, err := job.GetContainerArchive(ctx, workdir+"/act/workflow/event.json")
 	require.NoError(t, err)
 	reader := tar.NewReader(archive)
@@ -288,6 +292,12 @@ func TestKubernetesPodRunsJobThroughAPI(t *testing.T) {
 	require.NoError(t, archive.Close())
 	_, err = job.GetContainerArchive(ctx, workdir+"/missing")
 	require.Error(t, err)
+	large := strings.Repeat("a", compressCopyMin)
+	require.NoError(t, job.Copy(workdir+"/act/", &FileEntry{Name: "workflow/large", Mode: 0o644, Body: large})(ctx))
+	assert.Contains(t, lastExec(), url.QueryEscape("gzip -dc |"))
+	content, err = os.ReadFile(workdir + "/act/workflow/large")
+	require.NoError(t, err)
+	assert.Equal(t, large, string(content))
 
 	actionDir := filepath.Join(t.TempDir(), "action")
 	require.NoError(t, os.MkdirAll(actionDir, 0o755))
@@ -296,11 +306,12 @@ func TestKubernetesPodRunsJobThroughAPI(t *testing.T) {
 	script, err := os.ReadFile(workdir + "/act/actions/sha/index.js")
 	require.NoError(t, err)
 	assert.Equal(t, "main()", string(script))
-	assert.Contains(t, cluster.requests["exec"][len(cluster.requests["exec"])-1], "command="+url.QueryEscape(workdir+"/act/actions/sha/"))
+	assert.Contains(t, lastExec(), "command="+url.QueryEscape(workdir+"/act/actions/sha/"))
+	assert.Contains(t, lastExec(), url.QueryEscape("gzip -dc |"))
 
 	pathEnv := map[string]string{}
 	require.NoError(t, job.UpdateFromImageEnv(&pathEnv)(ctx))
-	assert.Equal(t, os.Getenv("PATH"), pathEnv["PATH"])
+	assert.Equal(t, path, pathEnv["PATH"])
 
 	info, err := service.Inspect(ctx)
 	require.NoError(t, err)
@@ -320,6 +331,38 @@ func TestKubernetesPodRunsJobThroughAPI(t *testing.T) {
 	assert.Empty(t, cluster.requests["DELETE /api/v1/namespaces/ci/secrets/new"])
 }
 
+func TestKubernetesCopiesUncompressedWithoutGzip(t *testing.T) {
+	t.Parallel()
+	if runtime.GOOS == "windows" {
+		t.Skip("uses POSIX shell")
+	}
+	ctx := t.Context()
+	workdir, bin := t.TempDir(), t.TempDir()
+	for _, tool := range []string{"uname", "tar", "mkdir"} { // all the job container needs but gzip, as on a minimal image
+		found, err := exec.LookPath(tool)
+		require.NoError(t, err)
+		require.NoError(t, os.Symlink(found, filepath.Join(bin, tool)))
+	}
+	kubeconfig, cluster := fakeKubernetes(t, `{"spec":{"containers":[{}]},"status":{"phase":"Running","containerStatuses":[
+		{"name":"job","ready":true,"state":{"running":{}}}]}}`, "PATH="+bin)
+	pod, err := NewKubernetesPod(KubernetesOptions{Kubeconfig: kubeconfig})
+	require.NoError(t, err)
+	job := pod.JobContainer(&NewContainerInput{Image: "node", WorkingDir: workdir})
+	require.NoError(t, job.Create(nil, nil)(ctx))
+	require.NoError(t, job.Start(false)(ctx))
+	assert.False(t, pod.job.gzip)
+	assert.Equal(t, bin, pod.path)
+
+	actionDir := filepath.Join(t.TempDir(), "action")
+	require.NoError(t, os.MkdirAll(actionDir, 0o755))
+	require.NoError(t, os.WriteFile(filepath.Join(actionDir, "index.js"), []byte("main()"), 0o644))
+	require.NoError(t, job.CopyDir(workdir+"/act/actions/sha/", actionDir+"/", false, false)(ctx))
+	script, err := os.ReadFile(workdir + "/act/actions/sha/index.js")
+	require.NoError(t, err)
+	assert.Equal(t, "main()", string(script))
+	assert.NotContains(t, cluster.requests["exec"][len(cluster.requests["exec"])-1], "gzip")
+}
+
 func TestKubernetesPodFailsBeforeSteps(t *testing.T) {
 	kubeconfig, _ := fakeKubernetes(t, `{"spec":{"containers":[{}]},"status":{"containerStatuses":[
 		{"name":"job","state":{"waiting":{"reason":"ErrImagePull","message":"no such image"}}}]}}`)
@@ -360,6 +403,13 @@ func TestKubernetesPodFailsBeforeSteps(t *testing.T) {
 	job := pod.JobContainer(&NewContainerInput{Image: "missing"})
 	require.NoError(t, job.Create(nil, nil)(t.Context()))
 	require.ErrorContains(t, job.Start(false)(t.Context()), "ErrImagePull: no such image")
+	failed, _ := fakeKubernetes(t, `{"spec":{"containers":[{}]},"status":{"phase":"Failed","initContainerStatuses":[
+		{"name":"gitea-runner-transfer","state":{"terminated":{"exitCode":1,"message":"exec format error\n"}}}]}}`)
+	pod, err = NewKubernetesPod(KubernetesOptions{Kubeconfig: failed})
+	require.NoError(t, err)
+	job = pod.JobContainer(&NewContainerInput{Image: "node"})
+	require.NoError(t, job.Create(nil, nil)(t.Context()))
+	require.ErrorContains(t, job.Start(false)(t.Context()), "container gitea-runner-transfer exited with 1: exec format error")
 
 	plugin := filepath.Join(t.TempDir(), "kubeconfig")
 	require.NoError(t, os.WriteFile(plugin, []byte(`

act/container/kubernetes_transfer.go

@@ -0,0 +1,248 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package container
+
+import (
+	"cmp"
+	"context"
+	"crypto/ecdsa"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/sha256"
+	"crypto/tls"
+	"crypto/x509"
+	"encoding/hex"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"math/big"
+	"net"
+	"net/http"
+	"os"
+	"os/exec"
+	"path"
+	"path/filepath"
+	"slices"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/common/logging"
+)
+
+// KubernetesTransfer has job pods fetch large copies from the runner over the pod network instead of through the API server.
+type KubernetesTransfer struct {
+	Host     string // empty uses the outbound IP
+	Port     uint16 // 0 picks a free port
+	Image    string // has gitea-runner on its PATH, for an init container to copy into each job pod
+	Fallback bool   // copy through the API server when a job pod cannot reach the runner, else fail
+}
+
+// KubernetesTransferCommand is the hidden runner subcommand job pods run the transfer helper with.
+const KubernetesTransferCommand = "kubernetes-transfer"
+
+// transferRequest goes to the helper on the exec stream's stdin, keeping the token out of the command line and the audit log.
+type transferRequest struct {
+	URL string `json:"url"`
+	Pin string `json:"pin"` // hex SHA-256 of the runner's certificate
+}
+
+type transferServer struct {
+	base    string
+	pin     string
+	mu      sync.Mutex
+	pending map[string]transferOffer
+}
+
+type transferOffer struct {
+	clients []net.IP
+	write   func(io.Writer) error
+	result  chan error
+}
+
+var (
+	transferServersMu sync.Mutex
+	transferServers   = map[string]*transferServer{}
+)
+
+// startTransferServer returns the runner's server for these settings, started on first use.
+func startTransferServer(host string, port uint16) (*transferServer, error) {
+	transferServersMu.Lock()
+	defer transferServersMu.Unlock()
+	configured := net.JoinHostPort(host, strconv.Itoa(int(port)))
+	if server, ok := transferServers[configured]; ok {
+		return server, nil
+	}
+	if host == "" {
+		ip := common.GetOutboundIP()
+		if ip == nil {
+			return nil, errors.New("no outbound IP for job pods to reach the runner at, set kubernetes.transfer.host")
+		}
+		host = ip.String()
+	}
+	privateKey, err := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	if err != nil {
+		return nil, err
+	}
+	template := &x509.Certificate{SerialNumber: big.NewInt(1)} // trusted by its pin alone, so no CA, names or validity
+	certificate, err := x509.CreateCertificate(rand.Reader, template, template, &privateKey.PublicKey, privateKey)
+	if err != nil {
+		return nil, err
+	}
+	listener, err := net.Listen("tcp", net.JoinHostPort(host, strconv.Itoa(int(port))))
+	if err != nil {
+		return nil, fmt.Errorf("listen for job pod transfers: %w", err)
+	}
+	pin := sha256.Sum256(certificate)
+	server := &transferServer{base: "https://" + listener.Addr().String() + "/", pin: hex.EncodeToString(pin[:]), pending: map[string]transferOffer{}}
+	httpServer := &http.Server{
+		Handler:           server,
+		TLSConfig:         &tls.Config{Certificates: []tls.Certificate{{Certificate: [][]byte{certificate}, PrivateKey: privateKey}}, MinVersion: tls.VersionTLS13},
+		ReadHeaderTimeout: 10 * time.Second,
+		IdleTimeout:       30 * time.Second,
+	}
+	go func() { _ = httpServer.ServeTLS(listener, "", "") }()
+	transferServers[configured] = server
+	return server, nil
+}
+
+// offer serves write once to one of clients until done, which reports whether a client took it and what write failed with.
+func (s *transferServer) offer(clients []net.IP, write func(io.Writer) error) (request transferRequest, done func() (taken bool, err error)) {
+	token := rand.Text()
+	result := make(chan error, 1)
+	s.mu.Lock()
+	s.pending[token] = transferOffer{clients: clients, write: write, result: result}
+	s.mu.Unlock()
+	return transferRequest{URL: s.base + token, Pin: s.pin}, func() (bool, error) {
+		s.mu.Lock()
+		_, pending := s.pending[token]
+		delete(s.pending, token)
+		s.mu.Unlock()
+		select {
+		case err := <-result:
+			return true, err
+		default:
+			return !pending, nil
+		}
+	}
+}
+
+func (s *transferServer) ServeHTTP(w http.ResponseWriter, r *http.Request) {
+	host, _, _ := net.SplitHostPort(r.RemoteAddr)
+	token := strings.TrimPrefix(r.URL.Path, "/")
+	s.mu.Lock()
+	offer, ok := s.pending[token]
+	ok = ok && r.Method == http.MethodGet && slices.ContainsFunc(offer.clients, net.ParseIP(host).Equal)
+	if ok {
+		delete(s.pending, token)
+	}
+	s.mu.Unlock()
+	if !ok {
+		http.NotFound(w, r)
+		return
+	}
+	w.Header().Set("Content-Encoding", "gzip")
+	err := writeGzip(w, offer.write)
+	offer.result <- err
+	if err != nil {
+		panic(http.ErrAbortHandler) // a cut-off body fails the fetch, a short tar might not
+	}
+}
+
+// FetchKubernetesTransfer reads a transferRequest from request and extracts the tar it points to into dest.
+func FetchKubernetesTransfer(ctx context.Context, request io.Reader, dest string) error {
+	var transfer transferRequest
+	if err := json.NewDecoder(request).Decode(&transfer); err != nil {
+		return fmt.Errorf("read transfer request: %w", err)
+	}
+	client := &http.Client{Transport: &http.Transport{
+		DialContext:           (&net.Dialer{Timeout: 10 * time.Second}).DialContext,
+		TLSHandshakeTimeout:   10 * time.Second,
+		ResponseHeaderTimeout: time.Minute,
+		TLSClientConfig: &tls.Config{
+			MinVersion:         tls.VersionTLS13,
+			InsecureSkipVerify: true, // VerifyConnection checks the pin instead
+			VerifyConnection: func(state tls.ConnectionState) error {
+				if sum := sha256.Sum256(state.PeerCertificates[0].Raw); hex.EncodeToString(sum[:]) != transfer.Pin {
+					return errors.New("runner certificate does not match the pin")
+				}
+				return nil
+			},
+		},
+	}}
+	httpRequest, err := http.NewRequestWithContext(ctx, http.MethodGet, transfer.URL, nil)
+	if err != nil {
+		return err
+	}
+	response, err := client.Do(httpRequest)
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	if response.StatusCode != http.StatusOK {
+		return fmt.Errorf("fetch: %s", response.Status)
+	}
+	if err := os.MkdirAll(dest, 0o777); err != nil {
+		return err
+	}
+	cmd := exec.CommandContext(ctx, "tar", "-xf", "-", "-C", dest)
+	cmd.Stdin, cmd.Stdout, cmd.Stderr = response.Body, os.Stderr, os.Stderr
+	if err := cmd.Run(); err != nil {
+		return err
+	}
+	_, err = io.Copy(io.Discard, response.Body) // tar stops at the first zero blocks, a cut-off body after them still fails
+	return err
+}
+
+// InstallKubernetesTransferHelper copies the running binary to target.
+func InstallKubernetesTransferHelper(target string) error {
+	executable, err := os.Executable()
+	if err != nil {
+		return err
+	}
+	source, err := os.Open(executable)
+	if err != nil {
+		return err
+	}
+	defer source.Close()
+	if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
+		return err
+	}
+	_ = os.Remove(target) // a file left in a reused act volume would keep its mode
+	file, err := os.OpenFile(target, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o755)
+	if err != nil {
+		return err
+	}
+	_, err = io.Copy(file, source)
+	return errors.Join(err, file.Close())
+}
+
+// setupTransfer turns on transfers for the pod, with the runner pod's image unless one is set.
+func (p *KubernetesPod) setupTransfer(runnerImage string) error {
+	settings := p.options.Transfer
+	if settings == nil {
+		return nil
+	}
+	image := cmp.Or(settings.Image, runnerImage)
+	var err error
+	var server *transferServer
+	if image == "" {
+		err = errors.New("set kubernetes.transfer.image, the runner is no single-container pod in kubernetes.namespace to take it from")
+	} else {
+		server, err = startTransferServer(settings.Host, settings.Port)
+	}
+	if err != nil {
+		if !settings.Fallback {
+			return fmt.Errorf("kubernetes.transfer: %w", err)
+		}
+		logging.Warnf("Job pod copies go through the API server: %v", err)
+		return nil
+	}
+	p.transfer, p.transferImage = server, image
+	p.transferHelper = path.Join(p.job.GetActPath(), "bin", "gitea-runner")
+	return nil
+}

act/container/kubernetes_transfer_test.go

@@ -0,0 +1,185 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package container
+
+import (
+	"bytes"
+	"context"
+	"crypto/rand"
+	"crypto/tls"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+const transferHelperEnv = "GITEA_RUNNER_TEST_TRANSFER_HELPER"
+
+func TestMain(m *testing.M) {
+	if os.Getenv(transferHelperEnv) == "1" && len(os.Args) == 4 && os.Args[1] == KubernetesTransferCommand && os.Args[2] == "fetch" {
+		if err := FetchKubernetesTransfer(context.Background(), os.Stdin, os.Args[3]); err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			os.Exit(1)
+		}
+		os.Exit(0)
+	}
+	os.Exit(m.Run())
+}
+
+func fetch(t *testing.T, request transferRequest) error {
+	t.Helper()
+	body, err := json.Marshal(request)
+	require.NoError(t, err)
+	return FetchKubernetesTransfer(t.Context(), bytes.NewReader(body), t.TempDir())
+}
+
+func TestTransferServer(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("extracts with tar")
+	}
+	server, err := startTransferServer("127.0.0.1", 0)
+	require.NoError(t, err)
+	again, err := startTransferServer("127.0.0.1", 0)
+	require.NoError(t, err)
+	assert.Same(t, server, again)
+	loopback := []net.IP{net.ParseIP("127.0.0.1")}
+	nothing := func(io.Writer) error { return nil }
+
+	t.Run("serves an offer gzipped once to an IP of its job pod", func(t *testing.T) {
+		request, done := server.offer([]net.IP{net.ParseIP("fd00::1"), loopback[0]}, nothing)
+		client := &http.Client{Transport: &http.Transport{TLSClientConfig: &tls.Config{InsecureSkipVerify: true}, DisableCompression: true}}
+		response, err := client.Get(request.URL)
+		require.NoError(t, err)
+		require.NoError(t, response.Body.Close())
+		assert.Equal(t, "gzip", response.Header.Get("Content-Encoding"))
+		require.ErrorContains(t, fetch(t, request), "404")
+		taken, _ := done()
+		assert.True(t, taken)
+	})
+
+	t.Run("another pod can neither take nor block an offer", func(t *testing.T) {
+		request, done := server.offer([]net.IP{net.ParseIP("10.1.2.3")}, nothing)
+		require.ErrorContains(t, fetch(t, request), "404")
+		taken, _ := done()
+		assert.False(t, taken)
+	})
+
+	t.Run("a failed write fails the fetch even past the end blocks tar stops at, and done reports it", func(t *testing.T) {
+		request, done := server.offer(loopback, func(w io.Writer) error {
+			_, _ = w.Write(make([]byte, 1024))
+			_, _ = io.CopyN(w, rand.Reader, 1<<20)
+			return errors.New("source gone")
+		})
+		require.Error(t, fetch(t, request))
+		taken, err := done()
+		assert.True(t, taken)
+		require.EqualError(t, err, "source gone")
+	})
+
+	t.Run("the helper refuses a runner whose certificate does not match its pin", func(t *testing.T) {
+		request, done := server.offer(loopback, nothing)
+		request.Pin = strings.Repeat("0", 64)
+		require.ErrorContains(t, fetch(t, request), "does not match the pin")
+		taken, _ := done()
+		assert.False(t, taken)
+	})
+}
+
+func TestKubernetesPodTransferImage(t *testing.T) {
+	newPod := func(image string, fallback bool) *KubernetesPod {
+		pod := &KubernetesPod{options: KubernetesOptions{Transfer: &KubernetesTransfer{Host: "127.0.0.1", Image: image, Fallback: fallback}}}
+		pod.JobContainer(&NewContainerInput{})
+		return pod
+	}
+
+	pod := newPod("", false)
+	require.NoError(t, pod.setupTransfer("gitea/runner:latest"))
+	assert.Equal(t, "gitea/runner:latest", pod.transferImage)
+
+	pod = newPod("registry.example/runner:1", false)
+	require.NoError(t, pod.setupTransfer("gitea/runner:latest"))
+	assert.Equal(t, "registry.example/runner:1", pod.transferImage)
+
+	pod = newPod("", true)
+	require.NoError(t, pod.setupTransfer(""))
+	assert.Empty(t, pod.transferHelper)
+
+	pod = newPod("", false)
+	require.ErrorContains(t, pod.setupTransfer(""), "kubernetes.transfer.image")
+}
+
+func TestKubernetesPodCopiesLargeFilesOverPodNetwork(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("uses POSIX shell")
+	}
+	ctx := t.Context()
+	workdir, err := filepath.EvalSymlinks(t.TempDir())
+	require.NoError(t, err)
+	kubeconfig, cluster := fakeKubernetes(t, `{"spec":{"containers":[{}]},"status":{"phase":"Running","podIPs":[{"ip":"127.0.0.1"}],"containerStatuses":[
+		{"name":"job","ready":true,"state":{"running":{}}}]}}`, transferHelperEnv+"=1")
+	pod, err := NewKubernetesPod(KubernetesOptions{Kubeconfig: kubeconfig, ForcePull: true, Transfer: &KubernetesTransfer{Host: "127.0.0.1", Image: "gitea/runner:latest", Fallback: true}})
+	require.NoError(t, err)
+	job := pod.JobContainer(&NewContainerInput{Name: "job", Image: "node", WorkingDir: workdir, Stdout: io.Discard, Stderr: io.Discard})
+	require.NoError(t, job.Create(nil, nil)(ctx))
+
+	var manifest struct {
+		Spec struct {
+			InitContainers []struct {
+				Image, ImagePullPolicy string
+				Command                []string
+				VolumeMounts           []struct{ Name, MountPath string }
+			}
+		}
+	}
+	require.NoError(t, json.Unmarshal([]byte(cluster.requests["POST /api/v1/namespaces/ci/pods"][0]), &manifest))
+	require.Len(t, manifest.Spec.InitContainers, 1)
+	initContainer := manifest.Spec.InitContainers[0]
+	assert.Equal(t, "gitea/runner:latest", initContainer.Image)
+	assert.Equal(t, "Always", initContainer.ImagePullPolicy)
+	assert.Equal(t, []string{"gitea-runner", KubernetesTransferCommand, "install", "/var/run/act/bin/gitea-runner"}, initContainer.Command)
+	assert.Equal(t, []struct{ Name, MountPath string }{{"act", "/var/run/act"}}, initContainer.VolumeMounts)
+	require.NoError(t, job.Start(false)(ctx))
+
+	helper := filepath.Join(t.TempDir(), "bin", "gitea-runner")
+	require.NoError(t, InstallKubernetesTransferHelper(helper))
+	pod.transferHelper = helper
+	lastExec := func() string { return cluster.requests["exec"][len(cluster.requests["exec"])-1] }
+	large := &FileEntry{Name: "large", Mode: 0o644, Body: strings.Repeat("a", compressCopyMin)}
+	copyLarge := func(dir string) error { return job.Copy(filepath.Join(workdir, dir)+"/", large)(ctx) }
+	require.NoError(t, copyLarge("transferred"))
+	content, err := os.ReadFile(filepath.Join(workdir, "transferred", "large"))
+	require.NoError(t, err)
+	assert.Equal(t, large.Body, string(content))
+	assert.Contains(t, lastExec(), "command="+KubernetesTransferCommand)
+	assert.NotContains(t, lastExec(), "127.0.0.1")
+
+	require.NoError(t, job.Copy(workdir+"/act/", &FileEntry{Name: "event.json", Mode: 0o644, Body: `{}`})(ctx))
+	assert.NotContains(t, lastExec(), "command="+KubernetesTransferCommand)
+
+	require.NoError(t, os.WriteFile(filepath.Join(workdir, "file"), nil, 0o644))
+	require.ErrorContains(t, copyLarge("file"), "over the pod network")
+	assert.Equal(t, helper, pod.transferHelper)
+
+	unreachable := filepath.Join(workdir, "unreachable")
+	require.NoError(t, os.WriteFile(unreachable, []byte("#!/bin/sh\necho 'connection refused' >&2\nexit 1\n"), 0o755))
+	pod.transferHelper = unreachable
+	pod.options.Transfer.Fallback = false
+	require.ErrorContains(t, copyLarge("direct"), "connection refused")
+	assert.NoFileExists(t, filepath.Join(workdir, "direct", "large"))
+
+	pod.options.Transfer.Fallback = true
+	require.NoError(t, copyLarge("auto"))
+	assert.FileExists(t, filepath.Join(workdir, "auto", "large"))
+	assert.Empty(t, pod.transferHelper)
+}

act/container/parse_env_file.go

@@ -7,6 +7,7 @@ package container
 import (
 	"archive/tar"
 	"bufio"
+	"bytes"
 	"context"
 	"fmt"
 	"io"
@@ -18,7 +19,8 @@ import (
 	"golang.org/x/text/transform"
 )
 
-func parseEnvFile(e Container, srcPath string, env *map[string]string) common.Executor {
+// parseEnvFile reads a GITHUB_ENV style file, splitting lines like GitHub's runner on Windows or elsewhere.
+func parseEnvFile(e Container, srcPath string, env *map[string]string, windows bool) common.Executor {
 	localEnv := *env
 	return func(ctx context.Context) error {
 		envTar, err := e.GetContainerArchive(ctx, srcPath)
@@ -38,8 +40,22 @@ func parseEnvFile(e Container, srcPath string, env *map[string]string) common.Ex
 		s := bufio.NewScanner(decoded)
 		// Default 64 KiB max token size is too small for realistic env-file lines; allow up to 16 MiB.
 		s.Buffer(make([]byte, 0, 64*1024), 16*1024*1024)
+		scanned := 0
+		s.Split(func(data []byte, atEOF bool) (int, []byte, error) {
+			if index := bytes.IndexByte(data[scanned:], '\n'); index >= 0 {
+				advance := scanned + index + 1
+				scanned = 0
+				return advance, data[:advance], nil
+			}
+			scanned = len(data)
+			if atEOF && len(data) > 0 {
+				scanned = 0
+				return len(data), data, nil
+			}
+			return 0, nil, nil
+		})
 		for s.Scan() {
-			line := s.Text()
+			line := trimEnvLineEnding(s.Text(), windows)
 			// GitHub's runner ignores blank lines
 			if strings.TrimSpace(line) == "" {
 				continue
@@ -52,10 +68,13 @@ func parseEnvFile(e Container, srcPath string, env *map[string]string) common.Ex
 			case multiLineEnv != -1:
 				var multiLineEnvContent []string
 				multiLineEnvDelimiter := line[multiLineEnv+2:]
+				if multiLineEnv == 0 || multiLineEnvDelimiter == "" {
+					return fmt.Errorf("invalid format '%v', name and delimiter must not be empty", line)
+				}
 				delimiterFound := false
 				for s.Scan() {
 					content := s.Text()
-					if content == multiLineEnvDelimiter {
+					if trimEnvLineEnding(content, windows) == multiLineEnvDelimiter {
 						delimiterFound = true
 						break
 					}
@@ -67,7 +86,7 @@ func parseEnvFile(e Container, srcPath string, env *map[string]string) common.Ex
 				if !delimiterFound {
 					return fmt.Errorf("invalid format delimiter '%v' not found before end of file", multiLineEnvDelimiter)
 				}
-				localEnv[line[:multiLineEnv]] = strings.Join(multiLineEnvContent, "\n")
+				localEnv[line[:multiLineEnv]] = trimEnvLineEnding(strings.Join(multiLineEnvContent, ""), windows)
 			default:
 				return fmt.Errorf("invalid format '%v', expected a line with '=' or '<<'", line)
 			}
@@ -78,3 +97,10 @@ func parseEnvFile(e Container, srcPath string, env *map[string]string) common.Ex
 		return nil
 	}
 }
+
+func trimEnvLineEnding(line string, windows bool) string {
+	if windows && strings.HasSuffix(line, "\r\n") {
+		return line[:len(line)-2]
+	}
+	return strings.TrimSuffix(line, "\n")
+}

act/container/parse_env_file_test.go

@@ -22,6 +22,7 @@ func newTestHostEnv(t *testing.T) (*HostEnvironment, string) {
 }
 
 func TestParseEnvFile(t *testing.T) {
+	t.Parallel()
 	largeValue := strings.Repeat("x", 2*1024*1024)
 	encodeUTF16 := func(endianness unicode.Endianness) string {
 		content, err := unicode.UTF16(endianness, unicode.UseBOM).NewEncoder().Bytes([]byte("FOO=bar\r\nMULTI<<EOF\r\nline1\r\nEOF\r\n"))
@@ -33,41 +34,65 @@ func TestParseEnvFile(t *testing.T) {
 		want          map[string]string
 	}{
 		{"single line", "FOO=bar\nBAZ=qux\n", map[string]string{"FOO": "bar", "BAZ": "qux"}},
+		{"single line CRLF and final bare CR", "FOO=bar\r\nBAZ=qux\r", map[string]string{"FOO": "bar\r", "BAZ": "qux\r"}},
+		{"single line without final newline", "FOO=bar", map[string]string{"FOO": "bar"}},
 		{"blank lines", "\nFOO=bar\n\n   \nBAZ=qux\n\n", map[string]string{"FOO": "bar", "BAZ": "qux"}},
 		{"multiline", "FOO<<EOF\nline1\nline2\nEOF\n", map[string]string{"FOO": "line1\nline2"}},
+		{"multiline without final newline", "FOO<<EOF\nline1\r\nline2\r\nEOF", map[string]string{"FOO": "line1\r\nline2\r"}},
+		{"empty multiline without final newline", "FOO<<EOF\nEOF", map[string]string{"FOO": ""}},
 		{"multiline blank lines", "FOO<<EOF\nline1\n\nline2\nEOF\n", map[string]string{"FOO": "line1\n\nline2"}},
 		{"multiline leading blank line", "FOO<<EOF\n\nline2\nEOF\n", map[string]string{"FOO": "\nline2"}},
+		{"multiline CRLF", "FOO<<EOF\r\nline1\r\nline2\r\nEOF\r\n", map[string]string{"FOO": "line1\r\nline2\r"}},
+		{"multiline mixed line endings", "FOO<<EOF\nline1\r\nline2\n\r\nEOF\n", map[string]string{"FOO": "line1\r\nline2\n\r"}},
 		{"large value within limit", "FOO<<EOF\n" + largeValue + "\nEOF\n", map[string]string{"FOO": largeValue}},
 		{"UTF8 BOM", "\xef\xbb\xbfFOO=bar\n", map[string]string{"FOO": "bar"}},
-		{"UTF16 little endian", encodeUTF16(unicode.LittleEndian), map[string]string{"FOO": "bar", "MULTI": "line1"}},
-		{"UTF16 big endian", encodeUTF16(unicode.BigEndian), map[string]string{"FOO": "bar", "MULTI": "line1"}},
+		{"UTF16 little endian", encodeUTF16(unicode.LittleEndian), map[string]string{"FOO": "bar\r", "MULTI": "line1\r"}},
+		{"UTF16 big endian", encodeUTF16(unicode.BigEndian), map[string]string{"FOO": "bar\r", "MULTI": "line1\r"}},
 	} {
 		t.Run(test.name, func(t *testing.T) {
 			e, envPath := newTestHostEnv(t)
 			require.NoError(t, os.WriteFile(envPath, []byte(test.content), 0o600))
 			env := map[string]string{}
-			require.NoError(t, parseEnvFile(e, envPath, &env)(t.Context()))
+			require.NoError(t, parseEnvFile(e, envPath, &env, false)(t.Context()))
 			assert.Equal(t, test.want, env)
 		})
 	}
+	t.Run("Windows line endings", func(t *testing.T) {
+		hostEnv, envPath := newTestHostEnv(t)
+		require.NoError(t, os.WriteFile(envPath, []byte("FOO=bar\r\nMULTI<<EOF\r\nline1\r\nline2\nEOF\r\n"), 0o600))
+		env := map[string]string{}
+		require.NoError(t, parseEnvFile(hostEnv, envPath, &env, true)(t.Context()))
+		assert.Equal(t, map[string]string{"FOO": "bar", "MULTI": "line1\r\nline2"}, env)
+	})
 }
 
 func TestParseEnvFileLineExceedsBufferReportsScannerError(t *testing.T) {
+	t.Parallel()
 	e, envPath := newTestHostEnv(t)
 	require.NoError(t, os.WriteFile(envPath, []byte("FOO<<EOF\n"+strings.Repeat("x", 17*1024*1024)+"\nEOF\n"), 0o600))
 
 	env := map[string]string{}
-	err := parseEnvFile(e, envPath, &env)(t.Context())
+	err := parseEnvFile(e, envPath, &env, false)(t.Context())
 	require.ErrorIs(t, err, bufio.ErrTooLong)
 	assert.Contains(t, err.Error(), "reading env file")
 }
 
-func TestParseEnvFileMissingDelimiter(t *testing.T) {
+func TestParseEnvFileInvalidHeredoc(t *testing.T) {
+	for name, content := range map[string]string{
+		"missing delimiter":              "FOO<<EOF\nline1\nline2\n",
+		"empty name":                     "<<EOF\nline1\nEOF\n",
+		"empty delimiter":                "FOO<<\nline1\n\n",
+		"delimiter with carriage return": "FOO<<EOF\nline1\nEOF\r\n",
+	} {
+		t.Run(name, func(t *testing.T) {
 			e, envPath := newTestHostEnv(t)
-	require.NoError(t, os.WriteFile(envPath, []byte("FOO<<EOF\nline1\nline2\n"), 0o600))
+			require.NoError(t, os.WriteFile(envPath, []byte(content), 0o600))
 
 			env := map[string]string{}
-	err := parseEnvFile(e, envPath, &env)(t.Context())
+			err := parseEnvFile(e, envPath, &env, false)(t.Context())
 			require.Error(t, err)
 			assert.Contains(t, err.Error(), "delimiter")
+			assert.Empty(t, env)
+		})
+	}
 }