fix: composite nested uses clone token (#1041)

## Summary

A nested `uses:` action inside a **local composite action** fails to clone with a bare `authentication required: Unauthorized` (401) when the instance resolves host-less action references against itself (`DEFAULT_ACTIONS_URL = self`). The job fails at the composite **main→post boundary**, even though the composite's own steps and all post-steps report success.

## Root cause

`newCompositeRunContext` nils `Config.Secrets` so composite steps do not see job secrets. But the action-clone path in `prepareActionExecutor` sourced its token via `getGitCloneToken` → `Config.GetToken()` → `Config.Secrets`, which is empty inside a composite RunContext. The nested action is therefore cloned anonymously → 401 against the authenticated instance.

Two details explain the exact symptom:

- It surfaces at the composite **main→post boundary** because the swallowed nested-step error is re-emitted by `common.JobError` at the end of the composite main pipeline.
- It carries **no clone URL** because, on a warm action cache, only `r.Fetch` runs (not `PlainClone`), and the go-git fetch error is returned verbatim.

## Fix

Source the clone token from `github.Token` instead of `Config.Secrets`. It is preserved across the composite config copy (`Config.Token` / `PresetGitHubContext`) and is identical to `Config.GetToken()` at the top level, so top-level and `act exec` behaviour is unchanged. The `shouldCloneURLUseToken` host gate is kept so the token is never sent to a foreign host. This also aligns the git-clone path with the ActionCache fetch path, which already uses `github.Token`.

Reusable workflows are unaffected — their RunContext keeps `Config.Secrets`.

## Before / after

Local composite action with a nested `uses:` (+ post step), followed by a marker step. Same workflow, same runner host — only the runner fix differs.

| Job step | Before fix | After fix |
| --- | --- | --- |
| `Run actions/checkout@v6` | ✅ success | ✅ success |
| `Run ./.gitea/actions/probe-composite` (nested `uses:` + post) | ❌ **failure** — bare 401 at the main→post boundary | ✅ success |
| `Step after composite` | ⊘ **skipped** | ✅ **ran** |
| Job result | ❌ **failed** | ✅ **succeeded** |

### Before — boundary log

```text
::endgroup::
##[error]authentication required: Unauthorized   ← bare 401, no clone target
Run Post ./.gitea/actions/probe-composite
Success - Post ./.gitea/actions/probe-composite  ← post steps run and succeed
Success - Post actions/checkout@v6
Job failed
```

The composite's own steps and both post-steps report `Success`; the job fails solely on the bare 401 emitted at the main→post boundary, and the next step is skipped.

### After — boundary log

```text
Run ./.gitea/actions/probe-composite
  ...composite steps...
Success - ./.gitea/actions/probe-composite
Run Marker after composite
PASSED composite boundary — no 401 (runner fix confirmed)
Success - Marker after composite
Job succeeded
```

## Reproduction

`.gitea/actions/probe-composite/action.yml`:

```yaml
name: probe-composite
runs:
  using: composite
  steps:
    - uses: actions/setup-node@v6   # nested uses → has a post step
      with: { node-version: 22 }
    - run: echo "composite inner step OK"
      shell: bash
```

`.gitea/workflows/probe.yaml`:

```yaml
on: [push]
jobs:
  composite-boundary:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
      - uses: ./.gitea/actions/probe-composite
      - run: echo "step after composite"   # skipped before the fix; runs after
```

## Verification

- **Before:** bare 401 at the boundary, reproduced with a `delay=0` tail — rules out a token-TTL / expiry effect; it is a missing credential.
- **After:** the composite step succeeds, the step after the composite runs, the job succeeds, and there is no `authentication required` in the log.
- A reusable workflow (`workflow_call`) with the same nested `uses:` + post step never hit the 401, which isolated the bug to the composite main/post path.

## Tests

Adds `TestStepActionRemoteCloneTokenSurvivesNilSecrets` (`act/runner`): asserts the clone token is forwarded when the RunContext mirrors a composite (`Secrets == nil`, token via `Config.Token`), and that the host gate still withholds the token for a foreign host. Verified to fail without the fix and pass with it.

---------

Reviewed-on: https://gitea.com/gitea/runner/pulls/1041
Reviewed-by: Zettat123 <39446+zettat123@noreply.gitea.com>
Co-authored-by: Christian Heim <christian@heimdaheim.de>
Co-committed-by: Christian Heim <christian@heimdaheim.de>

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

@@ -18,7 +18,7 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
     - uses: actions/setup-node@v6
       with:
         node-version: 24

.gitea/workflows/release-nightly.yml

@@ -17,7 +17,7 @@ jobs:
   goreleaser:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
         with:
           fetch-depth: 0
       - uses: actions/setup-go@v6
@@ -52,7 +52,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0 # all history for all branches and tags
 

.gitea/workflows/release-tag.yml

@@ -9,7 +9,7 @@ jobs:
   goreleaser:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
         with:
           fetch-depth: 0 # all history for all branches and tags
       - uses: actions/setup-go@v6
@@ -55,7 +55,7 @@ jobs:
       DOCKER_LATEST: latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0 # all history for all branches and tags
 

.gitea/workflows/test.yml

@@ -17,7 +17,7 @@ jobs:
       # to ~/.docker with the stale credentials.
       DOCKER_CONFIG: /tmp/docker-noauth
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - uses: actions/setup-go@v6
         with:
           go-version-file: 'go.mod'

Dockerfile

@@ -17,7 +17,7 @@ RUN make clean && make build
 ### DIND VARIANT
 #
 #
-FROM docker:29.5.3-dind AS dind
+FROM docker:29.6.0-dind AS dind
 
 ARG VERSION=dev
 
@@ -37,7 +37,7 @@ ENTRYPOINT ["s6-svscan","/etc/s6"]
 ### DIND-ROOTLESS VARIANT
 #
 #
-FROM docker:29.5.3-dind-rootless AS dind-rootless
+FROM docker:29.6.0-dind-rootless AS dind-rootless
 
 ARG VERSION=dev
 

Makefile

@@ -38,12 +38,15 @@ endif
 ifeq ($(OS), Windows_NT)
 	GOFLAGS := -v -buildmode=exe
 	EXECUTABLE ?= $(EXECUTABLE).exe
+	GO_ENV_WINDOWS := set GOOS=windows&&
 else ifeq ($(OS), Windows)
 	GOFLAGS := -v -buildmode=exe
 	EXECUTABLE ?= $(EXECUTABLE).exe
+	GO_ENV_WINDOWS := set GOOS=windows&&
 else
 	GOFLAGS := -v
 	EXECUTABLE ?= $(EXECUTABLE)
+	GO_ENV_WINDOWS := GOOS=windows
 endif
 
 STORED_VERSION_FILE := VERSION
@@ -108,12 +111,17 @@ deps-tools: ## install tool dependencies
 	wait
 
 .PHONY: lint
-lint: lint-go ## lint everything
+lint: lint-go lint-go-windows ## lint everything
 
 .PHONY: lint-go
 lint-go: ## lint go files
 	$(GO) run $(GOLANGCI_LINT_PACKAGE) run
 
+.PHONY: lint-go-windows
+lint-go-windows: ## lint Windows go files
+	$(GO) install $(GOLANGCI_LINT_PACKAGE)
+	$(GO_ENV_WINDOWS) golangci-lint run
+
 .PHONY: lint-go-fix
 lint-go-fix: ## lint go files and fix issues
 	$(GO) run $(GOLANGCI_LINT_PACKAGE) run --fix

README.md

@@ -160,9 +160,42 @@ Prefer a YAML file for all settings.
 
 If `runner.labels` is set in the YAML file, those labels are used during `register` and the `--labels` CLI flag is ignored.
 
-#### External cache (`actions/cache`)
+#### Caching (`actions/cache`)
 
-If `cache.external_server` is set, you must set `cache.external_secret` to the same value on this runner and on the standalone cache server. Run the server with `gitea-runner cache-server` using a config that defines `cache.external_secret` (and matching `cache.dir` / host / port as needed). Flags `--dir`, `--host`, and `--port` on `cache-server` override the file.
+Each runner starts its own cache server automatically. Cache entries are local to that runner — runners do not share a cache by default.
+
+**Shared cache across multiple runners**
+
+Run one dedicated `gitea-runner cache-server` that all runners point at.
+
+1. Create a config file for the cache server host:
+
+   ```yaml
+   cache:
+     dir: /data/actcache
+     port: 8088
+     external_secret: "replace-with-a-strong-random-secret"
+   ```
+
+2. Start the server:
+
+   ```bash
+   gitea-runner -c cache-server-config.yaml cache-server
+   ```
+
+3. On every runner:
+
+   ```yaml
+   cache:
+     external_server: "http://<cache-server-host>:8088/"
+     external_secret: "replace-with-a-strong-random-secret"  # must match the server
+   ```
+
+Alternatively, mount the same NFS/CIFS share on every runner and point `cache.dir` at it — simpler, but with weaker isolation between repositories.
+
+**S3 / MinIO** — mount object storage as a FUSE filesystem (e.g. [s3fs](https://github.com/s3fs-fuse/s3fs-fuse) or [goofys](https://github.com/kahing/goofys)) and set `cache.dir` to the mount point.
+
+Flags `--dir`, `--host`, and `--port` on `cache-server` override the corresponding `cache.*` YAML keys; all other settings, including `external_secret`, require the config file.
 
 #### Official Docker image
 
@@ -176,6 +209,16 @@ When `container.bind_workdir` is enabled, stale task workspace directories can b
 - only purely numeric subdirectories under `container.workdir_parent` are treated as task workspaces and may be removed
 - cleanup assumes `container.workdir_parent` is not shared across multiple runners
 
+#### Post-task script (`runner.post_task_script`)
+
+Optional host script that runs **after** each task's built-in cleanup (post-steps, container teardown, bind-workdir removal). Use it for extra machine housekeeping — Docker pruning, disk cleanup, and similar.
+
+**While the script runs, the runner stops task heartbeats and stays offline from Gitea's perspective until the script exits (or hits `runner.post_task_script_timeout`, default `5m`).** A script that blocks without exiting keeps the runner from taking new work for up to that timeout. Script output goes to the runner log, not the job log; a non-zero exit is warned but does not change the job result.
+
+On Windows, use `.exe`, `.bat`, or `.cmd` paths; **PowerShell (`.ps1`) is not supported yet** as the configured path — wrap commands in a `.cmd` file instead.
+
+See **[docs/post-task-script.md](docs/post-task-script.md)** for lifecycle details, environment variables, timeout interaction, and platform notes.
+
 ### Example Deployments
 
 Check out the [examples](examples) directory for sample deployment types.

act/common/git/git.go

@@ -257,6 +257,10 @@ type NewGitCloneExecutorInput struct {
 	Token       string
 	OfflineMode bool
 
+	// Depth limits the clone/fetch to the given number of commits from the tip of the requested ref.
+	// 0 for full clone.
+	Depth int
+
 	// For Gitea
 	InsecureSkipTLS bool
 }
@@ -309,7 +313,7 @@ func CloneIfRequired(ctx context.Context, refName plumbing.ReferenceName, input
 		}
 	}
 
-	r, err = git.PlainCloneContext(ctx, input.Dir, false, &cloneOptions)
+	r, err = cloneAtDepth(ctx, input, cloneOptions, logger)
 	if err != nil {
 		logger.Errorf("Unable to clone %v %s: %v", input.URL, refName, err)
 		return nil, false, err
@@ -364,6 +368,16 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 			pullOptions.InsecureSkipTLS = true
 		}
 
+		// Action clones only ever need the tip commit, so keep a shallow cache cheap on update at depth 1 regardless of its original depth
+		// Turning action_shallow_clone off does not convert an existing shallow cache; evict it for a full clone.
+		shallow := isShallow(r)
+		if shallow {
+			fetchOptions.Depth = 1
+			if spec, ok := shallowFetchRefSpec(r, input.Ref); ok {
+				fetchOptions.RefSpecs = []config.RefSpec{spec}
+			}
+		}
+
 		if !isOfflineMode {
 			err = r.Fetch(&fetchOptions)
 			if err != nil && !errors.Is(err, git.NoErrAlreadyUpToDate) {
@@ -431,11 +445,13 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 
 		reusedMsg := ""
 
-		if !isOfflineMode {
+		switch {
+		case !isOfflineMode && !shallow:
+			// In shallow mode the depth-limited fetch above already advanced the ref.
 			if err = w.Pull(&pullOptions); err != nil && err != git.NoErrAlreadyUpToDate {
 				logger.Debugf("Unable to pull %s: %v", refName, err)
 			}
-		} else if reused {
+		case isOfflineMode && reused:
 			reusedMsg = " (reused in offline mode)"
 		}
 
@@ -468,3 +484,53 @@ func NewGitCloneExecutor(input NewGitCloneExecutorInput) common.Executor {
 		return nil
 	}
 }
+
+// cloneAtDepth clones input.URL into input.Dir using opts.
+// With input.Depth > 0 it first tries a shallow, single-branch clone of input.Ref, falling back when error.
+func cloneAtDepth(ctx context.Context, input NewGitCloneExecutorInput, opts git.CloneOptions, logger log.FieldLogger) (*git.Repository, error) {
+	if input.Depth > 0 {
+		for _, refName := range []plumbing.ReferenceName{
+			plumbing.NewBranchReferenceName(input.Ref),
+			plumbing.NewTagReferenceName(input.Ref),
+		} {
+			shallowOpts := opts
+			shallowOpts.Depth = input.Depth
+			shallowOpts.SingleBranch = true
+			shallowOpts.ReferenceName = refName
+			shallowOpts.Tags = git.NoTags
+
+			r, err := git.PlainCloneContext(ctx, input.Dir, false, &shallowOpts)
+			if err == nil {
+				return r, nil
+			}
+			logger.Debugf("Shallow clone of %s as %s failed: %v", input.URL, refName, err)
+			if rmErr := os.RemoveAll(input.Dir); rmErr != nil {
+				return nil, fmt.Errorf("remove partial clone %s: %w", input.Dir, rmErr)
+			}
+		}
+		logger.Debugf("Falling back to a full clone of %s for ref %q", input.URL, input.Ref)
+	}
+
+	return git.PlainCloneContext(ctx, input.Dir, false, &opts)
+}
+
+// isShallow reports whether the local repository was cloned with a limited depth.
+func isShallow(r *git.Repository) bool {
+	shallows, err := r.Storer.Shallow()
+	return err == nil && len(shallows) > 0
+}
+
+// shallowFetchRefSpec returns the single refspec that updates only input.Ref, keeping a shallow clone from re-downloading every branch's history.
+// ok is false when the ref is not present locally as a tag or remote-tracking branch, in which case the broad default refspec is used.
+func shallowFetchRefSpec(r *git.Repository, ref string) (config.RefSpec, bool) {
+	tagRef := plumbing.NewTagReferenceName(ref)
+	if _, err := r.Reference(tagRef, false); err == nil {
+		return config.RefSpec(fmt.Sprintf("+%s:%s", tagRef, tagRef)), true
+	}
+	remoteRef := plumbing.NewRemoteReferenceName("origin", ref)
+	if _, err := r.Reference(remoteRef, false); err == nil {
+		branchRef := plumbing.NewBranchReferenceName(ref)
+		return config.RefSpec(fmt.Sprintf("+%s:%s", branchRef, remoteRef)), true
+	}
+	return "", false
+}

act/common/git/git_test.go

@@ -10,6 +10,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"sync"
 	"syscall"
@@ -380,6 +381,96 @@ func TestGitCloneExecutorOfflineMode(t *testing.T) {
 	})
 }
 
+func TestGitCloneExecutorShallow(t *testing.T) {
+	// Build a local "remote" with several commits on main plus a tag, so a full clone would pull noticeably more history than a shallow one.
+	remoteDir := t.TempDir()
+	require.NoError(t, gitCmd("init", "--bare", "--initial-branch=main", remoteDir))
+	workDir := t.TempDir()
+	require.NoError(t, gitCmd("clone", remoteDir, workDir))
+	require.NoError(t, gitCmd("-C", workDir, "checkout", "-b", "main"))
+	for _, m := range []string{"c1", "c2", "c3"} {
+		require.NoError(t, gitCmd("-C", workDir, "commit", "--allow-empty", "-m", m))
+	}
+	require.NoError(t, gitCmd("-C", workDir, "tag", "v1"))
+	sha := gitRevParse(t, workDir, "HEAD~1") // c2, a SHA that go-git cannot shallow-clone
+	require.NoError(t, gitCmd("-C", workDir, "push", "-u", "origin", "main"))
+	require.NoError(t, gitCmd("-C", workDir, "push", "origin", "v1"))
+
+	shallowMarker := func(dir string) string { return filepath.Join(dir, ".git", "shallow") }
+
+	t.Run("branch is cloned shallowly", func(t *testing.T) {
+		dir := t.TempDir()
+		require.NoError(t, NewGitCloneExecutor(NewGitCloneExecutorInput{
+			URL: remoteDir, Ref: "main", Dir: dir, Depth: 1,
+		})(t.Context()))
+		assert.FileExists(t, shallowMarker(dir), "clone should be shallow")
+		assert.Equal(t, 1, gitRevCount(t, dir), "only the tip commit should be present")
+		assert.Equal(t, "c3", gitHeadSubject(t, dir))
+	})
+
+	t.Run("tag is cloned shallowly", func(t *testing.T) {
+		dir := t.TempDir()
+		require.NoError(t, NewGitCloneExecutor(NewGitCloneExecutorInput{
+			URL: remoteDir, Ref: "v1", Dir: dir, Depth: 1,
+		})(t.Context()))
+		assert.FileExists(t, shallowMarker(dir), "clone should be shallow")
+		assert.Equal(t, 1, gitRevCount(t, dir))
+		assert.Equal(t, "c3", gitHeadSubject(t, dir))
+	})
+
+	t.Run("SHA falls back to a full clone", func(t *testing.T) {
+		dir := t.TempDir()
+		require.NoError(t, NewGitCloneExecutor(NewGitCloneExecutorInput{
+			URL: remoteDir, Ref: sha, Dir: dir, Depth: 1,
+		})(t.Context()))
+		// go-git cannot shallow-clone a raw SHA, so it falls back to a full clone; the absence of a shallow marker proves the fallback happened.
+		assert.NoFileExists(t, shallowMarker(dir), "a SHA ref must not produce a shallow clone")
+		assert.Equal(t, sha, gitRevParse(t, dir, "HEAD"))
+	})
+
+	t.Run("moving branch updates while staying shallow", func(t *testing.T) {
+		dir := t.TempDir()
+		require.NoError(t, NewGitCloneExecutor(NewGitCloneExecutorInput{
+			URL: remoteDir, Ref: "main", Dir: dir, Depth: 1,
+		})(t.Context()))
+		require.Equal(t, "c3", gitHeadSubject(t, dir))
+
+		// Advance main on the remote, then reuse the existing shallow clone.
+		require.NoError(t, gitCmd("-C", workDir, "commit", "--allow-empty", "-m", "c4"))
+		require.NoError(t, gitCmd("-C", workDir, "push", "origin", "main"))
+
+		require.NoError(t, NewGitCloneExecutor(NewGitCloneExecutorInput{
+			URL: remoteDir, Ref: "main", Dir: dir, Depth: 1,
+		})(t.Context()))
+		assert.Equal(t, "c4", gitHeadSubject(t, dir), "reused shallow clone should update to the new tip")
+		assert.FileExists(t, shallowMarker(dir), "repo should remain shallow after update")
+		assert.Equal(t, 1, gitRevCount(t, dir))
+	})
+}
+
+func gitRevParse(t *testing.T, dir, rev string) string {
+	t.Helper()
+	out, err := exec.Command("git", "-C", dir, "rev-parse", rev).Output()
+	require.NoError(t, err)
+	return strings.TrimSpace(string(out))
+}
+
+func gitRevCount(t *testing.T, dir string) int {
+	t.Helper()
+	out, err := exec.Command("git", "-C", dir, "rev-list", "--count", "HEAD").Output()
+	require.NoError(t, err)
+	n, err := strconv.Atoi(strings.TrimSpace(string(out)))
+	require.NoError(t, err)
+	return n
+}
+
+func gitHeadSubject(t *testing.T, dir string) string {
+	t.Helper()
+	out, err := exec.Command("git", "-C", dir, "log", "-1", "--format=%s").Output()
+	require.NoError(t, err)
+	return strings.TrimSpace(string(out))
+}
+
 func gitCmd(args ...string) error {
 	cmd := exec.Command("git", args...)
 	cmd.Stdout = os.Stdout

act/common/job_error.go

@@ -24,7 +24,9 @@ func JobError(ctx context.Context) error {
 }
 
 func SetJobError(ctx context.Context, err error) {
-	ctx.Value(jobErrorContextKeyVal).(map[string]error)["error"] = err
+	if container, ok := ctx.Value(jobErrorContextKeyVal).(map[string]error); ok {
+		container["error"] = err
+	}
 }
 
 // WithJobErrorContainer adds a value to the context as a container for an error

act/common/line_writer.go

@@ -12,6 +12,13 @@ import (
 // LineHandler is a callback function for handling a line
 type LineHandler func(line string) bool
 
+// Flusher is implemented by writers that buffer a trailing, not-yet-terminated
+// line. Callers should flush once the underlying stream has reached EOF so the
+// final line (when it is not newline-terminated) is not lost.
+type Flusher interface {
+	Flush()
+}
+
 type lineWriter struct {
 	buffer   bytes.Buffer
 	handlers []LineHandler
@@ -24,6 +31,14 @@ func NewLineWriter(handlers ...LineHandler) io.Writer {
 	return w
 }
 
+// FlushWriter flushes w if it implements Flusher. It is a no-op otherwise, so
+// callers can flush an io.Writer without knowing its concrete type.
+func FlushWriter(w io.Writer) {
+	if f, ok := w.(Flusher); ok {
+		f.Flush()
+	}
+}
+
 func (lw *lineWriter) Write(p []byte) (n int, err error) {
 	pBuf := bytes.NewBuffer(p)
 	written := 0
@@ -44,6 +59,17 @@ func (lw *lineWriter) Write(p []byte) (n int, err error) {
 	return written, nil
 }
 
+// Flush emits any buffered, not-yet-newline-terminated content as a final line.
+// It is safe to call multiple times; subsequent calls with an empty buffer are
+// no-ops.
+func (lw *lineWriter) Flush() {
+	if lw.buffer.Len() == 0 {
+		return
+	}
+	lw.handleLine(lw.buffer.String())
+	lw.buffer.Reset()
+}
+
 func (lw *lineWriter) handleLine(line string) {
 	for _, h := range lw.handlers {
 		ok := h(line)

act/common/line_writer_test.go

@@ -5,6 +5,7 @@
 package common
 
 import (
+	"io"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -39,3 +40,33 @@ func TestLineWriter(t *testing.T) {
 	assert.Equal(" and another\n", lines[2])
 	assert.Equal("last line\n", lines[3])
 }
+
+func TestLineWriterFlush(t *testing.T) {
+	lines := make([]string, 0)
+	lineHandler := func(s string) bool {
+		lines = append(lines, s)
+		return true
+	}
+
+	lineWriter := NewLineWriter(lineHandler)
+
+	assert := assert.New(t)
+	_, err := lineWriter.Write([]byte("complete line\npartial line without newline"))
+	assert.NoError(err) //nolint:testifylint // pre-existing pattern from nektos/act
+
+	// Only the newline-terminated line is emitted before flushing.
+	assert.Equal([]string{"complete line\n"}, lines)
+
+	// Flushing emits the buffered, not-yet-terminated trailing line.
+	FlushWriter(lineWriter)
+	assert.Equal([]string{"complete line\n", "partial line without newline"}, lines)
+
+	// Flushing again is a no-op: nothing is buffered.
+	FlushWriter(lineWriter)
+	assert.Len(lines, 2)
+}
+
+func TestFlushWriterIgnoresNonFlusher(t *testing.T) {
+	// FlushWriter must be a safe no-op for writers that do not buffer lines.
+	assert.NotPanics(t, func() { FlushWriter(io.Discard) })
+}

act/container/container_types.go

@@ -84,6 +84,12 @@ type NewDockerBuildExecutorInput struct {
 	Platform     string
 }
 
+// NewDockerNetworkCreateExecutorInput the input for the NewDockerNetworkCreateExecutor function
+type NewDockerNetworkCreateExecutorInput struct {
+	EnableIPv4 *bool
+	EnableIPv6 *bool
+}
+
 // NewDockerPullExecutorInput the input for the NewDockerPullExecutor function
 type NewDockerPullExecutorInput struct {
 	Image     string

act/container/docker_network.go

@@ -14,7 +14,7 @@ import (
 	"github.com/moby/moby/client"
 )
 
-func NewDockerNetworkCreateExecutor(name string) common.Executor {
+func NewDockerNetworkCreateExecutor(name string, opts NewDockerNetworkCreateExecutorInput) common.Executor {
 	return func(ctx context.Context) error {
 		cli, err := GetDockerClient(ctx)
 		if err != nil {
@@ -39,6 +39,8 @@ func NewDockerNetworkCreateExecutor(name string) common.Executor {
 		_, err = cli.NetworkCreate(ctx, name, client.NetworkCreateOptions{
 			Driver:     "bridge",
 			Scope:      "local",
+			EnableIPv4: opts.EnableIPv4,
+			EnableIPv6: opts.EnableIPv6,
 		})
 		if err != nil {
 			return err

act/container/docker_run.go

@@ -20,6 +20,7 @@ import (
 	"slices"
 	"strconv"
 	"strings"
+	"time"
 
 	"gitea.com/gitea/runner/act/common"
 	"gitea.com/gitea/runner/act/filecollector"
@@ -45,6 +46,13 @@ import (
 	"github.com/spf13/pflag"
 )
 
+// drainGracePeriod bounds how long we wait for an output-copy goroutine to
+// finish draining a container's output before returning, so that neither a
+// cancellation (waitForCommand) nor a normal container exit (wait) truncates
+// the tail of the log. It is a safety bound: in the common case the stream
+// reaches EOF and the goroutine returns well before this elapses.
+const drainGracePeriod = 2 * time.Second
+
 // NewContainer creates a reference to a container
 func NewContainer(input *NewContainerInput) ExecutionsEnvironment {
 	cr := new(containerReference)
@@ -229,6 +237,10 @@ type containerReference struct {
 	input *NewContainerInput
 	UID   int
 	GID   int
+	// attachDone is closed by the attach() streaming goroutine once it has
+	// drained and flushed the container's output. wait() blocks on it so the
+	// tail of the log lands before the step proceeds.
+	attachDone chan struct{}
 	LinuxContainerEnvironmentExtensions
 }
 
@@ -730,7 +742,9 @@ func (cr *containerReference) tryReadGID() common.Executor {
 func (cr *containerReference) waitForCommand(ctx context.Context, isTerminal bool, resp client.HijackedResponse, _ client.ExecCreateResult, _, _ string) error {
 	logger := common.Logger(ctx)
 
-	cmdResponse := make(chan error)
+	// Buffered so the copy goroutine never blocks on send if the grace-period
+	// drain below times out and no one is left to receive.
+	cmdResponse := make(chan error, 1)
 
 	go func() {
 		var outWriter io.Writer
@@ -749,6 +763,11 @@ func (cr *containerReference) waitForCommand(ctx context.Context, isTerminal boo
 		} else {
 			_, err = io.Copy(outWriter, resp.Reader)
 		}
+		// Flush any buffered, not-yet-newline-terminated trailing line so the
+		// final line of a command's output is not lost (e.g. an error message
+		// printed without a trailing newline before the process exits).
+		common.FlushWriter(outWriter)
+		common.FlushWriter(errWriter)
 		cmdResponse <- err
 	}()
 
@@ -760,6 +779,16 @@ func (cr *containerReference) waitForCommand(ctx context.Context, isTerminal boo
 			logger.Warnf("Failed to send CTRL+C: %+s", err)
 		}
 
+		// Give the copy goroutine a brief grace period to drain output already
+		// produced by the command before we return, so cancellation does not
+		// 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 <-time.After(drainGracePeriod):
+			logger.Warn("Timed out draining command output after cancellation")
+		}
+
 		// we return the context canceled error to prevent other steps
 		// from executing
 		return ctx.Err()
@@ -945,14 +974,23 @@ func (cr *containerReference) attach() common.Executor {
 		if errWriter == nil {
 			errWriter = os.Stderr
 		}
+		done := make(chan struct{})
+		cr.attachDone = done
 		go func() {
+			defer close(done)
+			var copyErr error
 			if !isTerminal || os.Getenv("NORAW") != "" {
-				_, err = stdcopy.StdCopy(outWriter, errWriter, out.Reader)
+				_, copyErr = stdcopy.StdCopy(outWriter, errWriter, out.Reader)
 			} else {
-				_, err = io.Copy(outWriter, out.Reader)
+				_, copyErr = io.Copy(outWriter, out.Reader)
 			}
-			if err != nil {
+			// Flush any buffered, not-yet-newline-terminated trailing line once
-				common.Logger(ctx).Error(err)
+			// the stream reaches EOF, so the final line of the container's
+			// output is not lost when it is not newline-terminated.
+			common.FlushWriter(outWriter)
+			common.FlushWriter(errWriter)
+			if copyErr != nil {
+				common.Logger(ctx).Error(copyErr)
 			}
 		}()
 		return nil
@@ -991,6 +1029,18 @@ func (cr *containerReference) wait() common.Executor {
 
 		logger.Debugf("Return status: %v", statusCode)
 
+		// The container has exited; wait for the attach() streaming goroutine to
+		// finish draining and flushing its output before returning, so the tail
+		// of the log is not lost. Bounded so a stuck stream cannot hang the step.
+		if cr.attachDone != nil {
+			select {
+			case <-cr.attachDone:
+			case <-time.After(drainGracePeriod):
+				logger.Warn("Timed out draining container output")
+			}
+			cr.attachDone = nil
+		}
+
 		if statusCode == 0 {
 			return nil
 		}

act/container/docker_run_test.go

@@ -8,6 +8,7 @@ import (
 	"bufio"
 	"bytes"
 	"context"
+	"encoding/binary"
 	"errors"
 	"io"
 	"net"
@@ -20,6 +21,7 @@ import (
 	"gitea.com/gitea/runner/act/common"
 
 	cerrdefs "github.com/containerd/errdefs"
+	"github.com/moby/moby/api/pkg/stdcopy"
 	"github.com/moby/moby/api/types/container"
 	mobyclient "github.com/moby/moby/client"
 	"github.com/sirupsen/logrus/hooks/test"
@@ -89,6 +91,11 @@ func (m *mockDockerClient) ExecInspect(ctx context.Context, execID string, opts
 	return args.Get(0).(mobyclient.ExecInspectResult), args.Error(1)
 }
 
+func (m *mockDockerClient) ContainerAttach(ctx context.Context, containerID string, opts mobyclient.ContainerAttachOptions) (mobyclient.ContainerAttachResult, error) {
+	args := m.Called(ctx, containerID, opts)
+	return args.Get(0).(mobyclient.ContainerAttachResult), args.Error(1)
+}
+
 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)
@@ -206,6 +213,71 @@ func TestDockerExecFailure(t *testing.T) {
 	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.
+func stdcopyFrame(stream stdcopy.StdType, payload string) []byte {
+	b := make([]byte, 8+len(payload))
+	b[0] = byte(stream)
+	binary.BigEndian.PutUint32(b[4:8], uint32(len(payload)))
+	copy(b[8:], payload)
+	return b
+}
+
+// TestDockerAttachFlushesTrailingLine verifies that wait() blocks until the
+// attach() streaming goroutine has drained and flushed the container's output,
+// so a final line without a trailing newline is not lost.
+func TestDockerAttachFlushesTrailingLine(t *testing.T) {
+	ctx := context.Background()
+
+	framed := bytes.NewBuffer(stdcopyFrame(stdcopy.Stdout, "line one\nlast line without newline"))
+
+	var lines []string
+	logWriter := common.NewLineWriter(func(s string) bool {
+		lines = append(lines, s)
+		return true
+	})
+
+	client := &mockDockerClient{}
+	client.On("ContainerAttach", ctx, "123", mock.AnythingOfType("client.ContainerAttachOptions")).
+		Return(mobyclient.ContainerAttachResult{
+			HijackedResponse: mobyclient.HijackedResponse{
+				Conn:   &mockConn{},
+				Reader: bufio.NewReader(framed),
+			},
+		}, nil)
+
+	statusCh := make(chan container.WaitResponse, 1)
+	statusCh <- container.WaitResponse{StatusCode: 0}
+	errCh := make(chan error, 1)
+	client.On("ContainerWait", ctx, "123", mobyclient.ContainerWaitOptions{Condition: container.WaitConditionNotRunning}).
+		Return(mobyclient.ContainerWaitResult{
+			Result: (<-chan container.WaitResponse)(statusCh),
+			Error:  (<-chan error)(errCh),
+		})
+
+	cr := &containerReference{
+		id:  "123",
+		cli: client,
+		input: &NewContainerInput{
+			Image:  "image",
+			Stdout: logWriter,
+			Stderr: logWriter,
+		},
+	}
+
+	require.NoError(t, cr.attach()(ctx))
+	require.NoError(t, cr.wait()(ctx))
+
+	// wait() must have blocked until the goroutine drained AND flushed; the
+	// trailing, non-newline-terminated line must therefore be present. Reading
+	// lines here is race-free because wait() synchronizes on attachDone, which
+	// the goroutine closes after the final append.
+	assert.Equal(t, []string{"line one\n", "last line without newline"}, lines)
+
+	client.AssertExpectations(t)
+}
+
 func TestDockerWaitFailure(t *testing.T) {
 	ctx := context.Background()
 

act/container/docker_stub.go

@@ -61,7 +61,7 @@ func NewDockerVolumeRemoveExecutor(volume string, force bool) common.Executor {
 	}
 }
 
-func NewDockerNetworkCreateExecutor(name string) common.Executor {
+func NewDockerNetworkCreateExecutor(name string, opts NewDockerNetworkCreateExecutorInput) common.Executor {
 	return func(ctx context.Context) error {
 		return nil
 	}

act/container/host_environment.go

@@ -23,6 +23,7 @@ import (
 	"gitea.com/gitea/runner/act/common"
 	"gitea.com/gitea/runner/act/filecollector"
 	"gitea.com/gitea/runner/act/lookpath"
+	"gitea.com/gitea/runner/internal/pkg/process"
 
 	"github.com/go-git/go-billy/v5/helper/polyfill"
 	"github.com/go-git/go-billy/v5/osfs"
@@ -261,7 +262,7 @@ func setupPty(cmd *exec.Cmd, cmdline string) (*os.File, *os.File, error) {
 	cmd.Stdin = tty
 	cmd.Stdout = tty
 	cmd.Stderr = tty
-	cmd.SysProcAttr = getSysProcAttr(cmdline, true)
+	cmd.SysProcAttr = process.SysProcAttr(cmdline, true)
 	return ppty, tty, nil
 }
 
@@ -321,30 +322,14 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
 	cmd.Env = envList
 	cmd.Stderr = e.StdOut
 	cmd.Dir = wd
-	cmd.SysProcAttr = getSysProcAttr(cmdline, false)
+	cmd.SysProcAttr = process.SysProcAttr(cmdline, false)
 
-	// On Windows a step often launches a process tree (a shell that starts a
+	// Kill the step's whole process tree on cancellation (a step often launches a
-	// child which spawns further GUI or background processes). The default
+	// shell that spawns further background or GUI children) and bound the post-exit
-	// context cancellation only kills the direct child, leaving the rest of the
+	// I/O wait, so an orphan inheriting cmd's stdout/stderr pipe can never hang
-	// tree running; and because the orphans inherit cmd's stdout/stderr pipe,
+	// cmd.Wait() and the runner. See process.TreeKill. The PTY path below may
-	// cmd.Wait() would block forever, hanging the runner. Kill the whole tree
+	// override SysProcAttr, but never touches Cancel/WaitDelay.
-	// via a Job Object on cancellation, and bound the wait so a leftover pipe
+	treeKill := process.NewTreeKill(cmd)
-	// writer can never hang Wait indefinitely.
-	var killer atomic.Pointer[processKiller]
-	if runtime.GOOS == "windows" {
-		cmd.Cancel = func() error {
-			if k := killer.Load(); k != nil {
-				return k.Kill()
-			}
-			if cmd.Process != nil {
-				return cmd.Process.Kill()
-			}
-			return nil
-		}
-		// Once the step process has exited, give its I/O pipes at most this long
-		// to drain before Wait force-closes them and returns (Go's WaitDelay).
-		cmd.WaitDelay = 10 * time.Second
-	}
 
 	var ppty *os.File
 	var tty *os.File
@@ -375,18 +360,11 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
 	if err := cmd.Start(); err != nil {
 		return err
 	}
-	if runtime.GOOS == "windows" {
+	if k, kerr := treeKill.Capture(cmd.Process); kerr != nil {
-		// Assign the started process to a Job Object so cmd.Cancel can kill the
-		// whole descendant tree. Children spawned afterwards are auto-included.
-		// On failure (e.g. nested-job restrictions) we fall back to the default
-		// single-process kill; WaitDelay + end-of-job cleanup still apply.
-		if k, kerr := newProcessKiller(cmd.Process); kerr != nil {
 		common.Logger(ctx).Warnf("process tree kill setup failed, falling back to single-process kill: %v", kerr)
 	} else {
-			killer.Store(k)
 		defer k.Close()
 	}
-	}
 	err = cmd.Wait()
 	if err != nil {
 		var exitErr *exec.ExitError

act/container/process_other.go

@@ -1,19 +0,0 @@
-// Copyright 2026 The Gitea Authors. All rights reserved.
-// SPDX-License-Identifier: MIT
-
-//go:build !windows
-
-package container
-
-import "os"
-
-// processKiller is a no-op on non-Windows platforms. The Job Object based
-// tree-kill is only wired in on Windows (see exec()); elsewhere the default
-// exec.CommandContext cancellation and Setpgid handling apply.
-type processKiller struct{}
-
-func newProcessKiller(_ *os.Process) (*processKiller, error) { return &processKiller{}, nil }
-
-func (k *processKiller) Kill() error { return nil }
-
-func (k *processKiller) Close() error { return nil }

act/container/process_windows.go

@@ -1,71 +0,0 @@
-// Copyright 2026 The Gitea Authors. All rights reserved.
-// SPDX-License-Identifier: MIT
-
-package container
-
-import (
-	"os"
-
-	"golang.org/x/sys/windows"
-)
-
-// processKiller terminates a step process together with its entire descendant
-// tree via a Windows Job Object.
-//
-// Background: a step often launches a process tree (a shell that starts a
-// child which in turn spawns further GUI or background processes). The default
-// exec.CommandContext cancellation only kills the direct child, so cancelling a
-// job left the rest of the tree running. Because those orphans inherited the
-// step's stdout/stderr pipe, cmd.Wait() also blocked forever and the runner hung.
-//
-// Assigning the step process to a Job Object lets us kill the whole tree
-// atomically on cancellation (TerminateJobObject), which also closes the
-// inherited pipe handles so cmd.Wait() can return.
-type processKiller struct {
-	job windows.Handle
-}
-
-// newProcessKiller creates a Job Object and assigns p (an already-started
-// process) to it. Children spawned by p afterwards are automatically part of
-// the job. The job does NOT use JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE, so closing
-// the handle on normal completion does not kill legitimate background
-// processes; the tree is only torn down by an explicit Kill (cancellation).
-func newProcessKiller(p *os.Process) (*processKiller, error) {
-	job, err := windows.CreateJobObject(nil, nil)
-	if err != nil {
-		return nil, err
-	}
-
-	h, err := windows.OpenProcess(windows.PROCESS_SET_QUOTA|windows.PROCESS_TERMINATE, false, uint32(p.Pid))
-	if err != nil {
-		windows.CloseHandle(job)
-		return nil, err
-	}
-	defer windows.CloseHandle(h)
-
-	if err := windows.AssignProcessToJobObject(job, h); err != nil {
-		windows.CloseHandle(job)
-		return nil, err
-	}
-
-	return &processKiller{job: job}, nil
-}
-
-// Kill terminates every process currently assigned to the job (the step process
-// and all of its descendants).
-func (k *processKiller) Kill() error {
-	if k == nil || k.job == 0 {
-		return nil
-	}
-	return windows.TerminateJobObject(k.job, 1)
-}
-
-// Close releases the job handle. It does not terminate the processes.
-func (k *processKiller) Close() error {
-	if k == nil || k.job == 0 {
-		return nil
-	}
-	h := k.job
-	k.job = 0
-	return windows.CloseHandle(h)
-}

act/container/util.go

@@ -8,23 +8,10 @@ package container
 
 import (
 	"os"
-	"syscall"
 
 	"github.com/creack/pty"
 )
 
-func getSysProcAttr(_ string, tty bool) *syscall.SysProcAttr {
-	if tty {
-		return &syscall.SysProcAttr{
-			Setsid:  true,
-			Setctty: true,
-		}
-	}
-	return &syscall.SysProcAttr{
-		Setpgid: true,
-	}
-}
-
 func openPty() (*os.File, *os.File, error) {
 	return pty.Open()
 }

act/container/util_openbsd_mips64.go

@@ -7,15 +7,8 @@ package container
 import (
 	"errors"
 	"os"
-	"syscall"
 )
 
-func getSysProcAttr(cmdLine string, tty bool) *syscall.SysProcAttr {
-	return &syscall.SysProcAttr{
-		Setpgid: true,
-	}
-}
-
 func openPty() (*os.File, *os.File, error) {
 	return nil, nil, errors.New("Unsupported")
 }

act/container/util_plan9.go

@@ -7,15 +7,8 @@ package container
 import (
 	"errors"
 	"os"
-	"syscall"
 )
 
-func getSysProcAttr(cmdLine string, tty bool) *syscall.SysProcAttr {
-	return &syscall.SysProcAttr{
-		Rfork: syscall.RFNOTEG,
-	}
-}
-
 func openPty() (*os.File, *os.File, error) {
 	return nil, nil, errors.New("Unsupported")
 }

act/container/util_windows.go

@@ -7,13 +7,8 @@ package container
 import (
 	"errors"
 	"os"
-	"syscall"
 )
 
-func getSysProcAttr(cmdLine string, tty bool) *syscall.SysProcAttr {
-	return &syscall.SysProcAttr{CmdLine: cmdLine, CreationFlags: syscall.CREATE_NEW_PROCESS_GROUP}
-}
-
 func openPty() (*os.File, *os.File, error) {
 	return nil, nil, errors.New("Unsupported")
 }

act/exprparser/functions.go

@@ -266,7 +266,7 @@ func (impl *interperterImpl) jobSuccess() (bool, error) { //nolint:unparam // pr
 	jobNeeds := impl.getNeedsTransitive(impl.config.Run.Job())
 
 	for _, needs := range jobNeeds {
-		if jobs[needs].Result != "success" {
+		if jobs[needs].NeedsResult() != "success" {
 			return false, nil
 		}
 	}
@@ -283,7 +283,7 @@ func (impl *interperterImpl) jobFailure() (bool, error) { //nolint:unparam // pr
 	jobNeeds := impl.getNeedsTransitive(impl.config.Run.Job())
 
 	for _, needs := range jobNeeds {
-		if jobs[needs].Result == "failure" {
+		if jobs[needs].NeedsResult() == "failure" {
 			return true, nil
 		}
 	}

act/lookpath/lp_windows.go

@@ -62,7 +62,7 @@ func LookPath2(file string, lenv Env) (string, error) {
 	var exts []string
 	x := lenv.Getenv(`PATHEXT`)
 	if x != "" {
-		for _, e := range strings.Split(strings.ToLower(x), `;`) {
+		for e := range strings.SplitSeq(strings.ToLower(x), `;`) {
 			if e == "" {
 				continue
 			}

act/model/workflow.go

@@ -197,6 +197,7 @@ type Job struct {
 	If                 yaml.Node                 `yaml:"if"`
 	Steps              []*Step                   `yaml:"steps"`
 	TimeoutMinutes     string                    `yaml:"timeout-minutes"`
+	RawContinueOnError string                    `yaml:"continue-on-error"`
 	Services           map[string]*ContainerSpec `yaml:"services"`
 	Strategy           *Strategy                 `yaml:"strategy"`
 	RawContainer       yaml.Node                 `yaml:"container"`
@@ -207,6 +208,34 @@ type Job struct {
 	RawSecrets         yaml.Node                 `yaml:"secrets"`
 	RawPermissions     yaml.Node                 `yaml:"permissions"`
 	Result             string
+	// Runtime fields set during execution (not from YAML):
+	ContinueOnError bool // true when all failing matrix combinations had continue-on-error=true
+	hasFirmFailure  bool // true once any combination failed without continue-on-error
+}
+
+// SetContinueOnError records whether this combination's failure should not fail the workflow.
+// Must be called under the job lock. Safe across parallel matrix combinations.
+func (j *Job) SetContinueOnError(continueOnErr bool) {
+	if continueOnErr {
+		if !j.hasFirmFailure {
+			j.ContinueOnError = true
+		}
+	} else {
+		j.hasFirmFailure = true
+		j.ContinueOnError = false
+	}
+}
+
+// NeedsResult returns the job result as seen by dependent jobs through the
+// `needs` context. A job that failed but was tolerated via continue-on-error
+// reports "success" to its dependents, matching GitHub: such a failure must not
+// block jobs gated on the default `if: success()`, even though the overall
+// workflow run is still marked as failed.
+func (j *Job) NeedsResult() string {
+	if j.Result == "failure" && j.ContinueOnError {
+		return "success"
+	}
+	return j.Result
 }
 
 // Strategy for the job

act/model/workflow_test.go

@@ -32,6 +32,32 @@ func TestStepCloneIsolatesMutableFields(t *testing.T) {
 	assert.Equal(t, "original", orig.With["arg"], "With map must not be shared with the clone")
 }
 
+// TestJobNeedsResult guards the continue-on-error semantics exposed to dependent
+// jobs through the `needs` context: a failed-but-tolerated job reports "success"
+// so it does not block dependents gated on the default `if: success()`, matching
+// GitHub. A firm failure and any non-failure result are reported verbatim.
+func TestJobNeedsResult(t *testing.T) {
+	cases := []struct {
+		name            string
+		result          string
+		continueOnError bool
+		want            string
+	}{
+		{"tolerated failure reports success", "failure", true, "success"},
+		{"firm failure reports failure", "failure", false, "failure"},
+		{"success is unchanged", "success", false, "success"},
+		{"success with continue-on-error is unchanged", "success", true, "success"},
+		{"empty result is unchanged", "", true, ""},
+		{"skipped is unchanged", "skipped", true, "skipped"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			j := &Job{Result: tc.result, ContinueOnError: tc.continueOnError}
+			assert.Equal(t, tc.want, j.NeedsResult())
+		})
+	}
+}
+
 func TestReadWorkflow_ScheduleEvent(t *testing.T) {
 	yaml := `
 name: local-action-docker-url

act/runner/action.go

@@ -6,7 +6,9 @@ package runner
 
 import (
 	"context"
+	"crypto/sha256"
 	"embed"
+	"encoding/hex"
 	"errors"
 	"fmt"
 	"io"
@@ -272,6 +274,36 @@ func removeGitIgnore(ctx context.Context, directory string) error {
 	return nil
 }
 
+// dockerActionImageTag derives the local docker image tag used when an action
+// is built from a Dockerfile.
+//
+// For Gitea: a local action (`uses: ./` or `uses: ./path`) has an actionName
+// that is the workspace-relative path of the action. That path is identical
+// across repositories (e.g. "./" for a self-referencing action), so without
+// namespacing, every repository's local docker action would build and reuse the
+// same `act-dockeraction:latest` image on a shared docker daemon. A subsequent
+// repository would then silently run the image built for an earlier one.
+// Including the repository keeps the tag stable for caching within a repository
+// while preventing cross-repository collisions.
+// See https://gitea.com/gitea/runner/issues/1039.
+func dockerActionImageTag(repository, actionName string, localAction bool) string {
+	name := actionName
+	if localAction {
+		name = path.Join(repository, actionName)
+	}
+	// The human-readable name is sanitized by collapsing every non-alphanumeric character to "-".
+	sanitized := regexp.MustCompile("[^a-zA-Z0-9]").ReplaceAllString(name, "-")
+	if localAction {
+		// For local actions a short hash of the raw repository and action path is appended so the tag stays unique per repository.
+		sum := sha256.Sum256([]byte(repository + "\x00" + actionName))
+		sanitized += "-" + hex.EncodeToString(sum[:])[:12]
+	}
+	// "-dockeraction" ensures that "./", "./test " won't get converted to "act-:latest", "act-test-:latest" which are invalid docker image names
+	image := fmt.Sprintf("%s-dockeraction:%s", sanitized, "latest")
+	image = "act-" + strings.TrimLeft(image, "-")
+	return strings.ToLower(image)
+}
+
 // TODO: break out parts of function to reduce complexicity
 func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, basedir string, localAction bool) error {
 	logger := common.Logger(ctx)
@@ -286,10 +318,7 @@ func execAsDocker(ctx context.Context, step actionStep, actionName, actionDir, b
 		// Apply forcePull only for prebuild docker images
 		forcePull = rc.Config.ForcePull
 	} else {
-		// "-dockeraction" enshures that "./", "./test " won't get converted to "act-:latest", "act-test-:latest" which are invalid docker image names
+		image = dockerActionImageTag(step.getGithubContext(ctx).Repository, actionName, localAction)
-		image = fmt.Sprintf("%s-dockeraction:%s", regexp.MustCompile("[^a-zA-Z0-9]").ReplaceAllString(actionName, "-"), "latest")
-		image = "act-" + strings.TrimLeft(image, "-")
-		image = strings.ToLower(image)
 		contextDir, fileName := filepath.Split(filepath.Join(basedir, action.Runs.Image))
 
 		anyArchExists, err := ContainerImageExistsLocally(ctx, image, "any")

act/runner/action_test.go

@@ -455,3 +455,50 @@ func TestExecAsDockerHoldsCloneLockForRemoteUncached(t *testing.T) {
 		t.Fatal("execAsDocker did not return after inner was released and ctx was canceled")
 	}
 }
+
+func TestDockerActionImageTag(t *testing.T) {
+	// Remote actions already carry a unique, ref-scoped actionName (the uses
+	// hash), so the tag must be left untouched for backwards compatibility.
+	assert.Equal(t,
+		"act-abc123-dockeraction:latest",
+		dockerActionImageTag("owner/repo", "abc123", false),
+	)
+
+	// Local actions keep a human-readable, repository-namespaced prefix and gain a short hash suffix that makes the tag unique per (repository, actionName).
+	// See https://gitea.com/gitea/runner/issues/1039.
+	assert.Equal(t,
+		"act-owner-repo-baca2daaa2fe-dockeraction:latest",
+		dockerActionImageTag("owner/repo", "./", true),
+	)
+	assert.Equal(t,
+		"act-owner-repo-sub-e847b61255a8-dockeraction:latest",
+		dockerActionImageTag("owner/repo", "./sub", true),
+	)
+
+	// Sanitizing every non-alphanumeric character to "-" is lossy, so distinct inputs can collapse to the same readable prefix.
+	// The hash suffix must keep such cases apart, otherwise an image built for one repository is reused for another.
+	collisions := [][2]struct {
+		repoName   string
+		actionName string
+	}{
+		// Two different repositories, both `uses: ./`: "a/b-c" and "a-b/c" both sanitize to "a-b-c".
+		{{"a/b-c", "./"}, {"a-b/c", "./"}},
+		// A repository's root action vs another repository's sub-path action:
+		// "owner/repo-a" + "./" and "owner/repo" + "./a" both sanitize to "owner-repo-a".
+		{{"owner/repo-a", "./"}, {"owner/repo", "./a"}},
+	}
+	for _, c := range collisions {
+		assert.NotEqual(t,
+			dockerActionImageTag(c[0].repoName, c[0].actionName, true),
+			dockerActionImageTag(c[1].repoName, c[1].actionName, true),
+			"local docker action tags must differ for %q/%q vs %q/%q",
+			c[0].repoName, c[0].actionName, c[1].repoName, c[1].actionName,
+		)
+	}
+
+	// Distinct local actions within the same repository keep distinct tags.
+	assert.NotEqual(t,
+		dockerActionImageTag("owner/repo", "./", true),
+		dockerActionImageTag("owner/repo", "./sub", true),
+	)
+}

act/runner/cancellation_test.go

@@ -0,0 +1,285 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package runner
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/exprparser"
+	"gitea.com/gitea/runner/act/model"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"go.yaml.in/yaml/v4"
+)
+
+// TestCancelledJobStatusEnablesAlwaysAndCancelledSteps verifies that once a job is
+// cancelled, getJobContext reports the "cancelled" status so the step `if` functions
+// evaluate the way GitHub Actions does: cancelled()/always() are true, success()/failure()
+// are false. A step that defaults to success() is therefore skipped while an always() step
+// still runs. Before the fix the status could only ever be success/failure, so cancelled()
+// was structurally impossible and cancel-only cleanup steps never ran.
+func TestCancelledJobStatusEnablesAlwaysAndCancelledSteps(t *testing.T) {
+	rc := createIfTestRunContext(map[string]*model.Job{
+		"job1": createJob(t, `runs-on: ubuntu-latest`, ""),
+	})
+	rc.markCancelled()
+
+	// The core fix: the job status context now reports "cancelled" instead of being
+	// pinned to success/failure.
+	jobCtx := rc.getJobContext()
+	require.Equal(t, "cancelled", jobCtx.Status)
+
+	// Feed that status through the step-context expression functions, which is what a
+	// step `if` evaluates. On a cancelled job only always()/cancelled() are true.
+	interp := exprparser.NewInterpeter(
+		&exprparser.EvaluationEnvironment{Job: jobCtx},
+		exprparser.Config{Context: "step"},
+	)
+	for expr, want := range map[string]bool{
+		"cancelled()":  true,
+		"always()":     true,
+		"success()":    false,
+		"failure()":    false,
+		"!cancelled()": false,
+	} {
+		got, err := interp.Evaluate(expr, exprparser.DefaultStatusCheckNone)
+		require.NoErrorf(t, err, "Evaluate(%q)", expr)
+		assert.Equalf(t, want, got, "Evaluate(%q) on a cancelled job", expr)
+	}
+
+	// A step without an `if` defaults to success() and must be skipped on cancel,
+	// while an `if: always()` step must still run.
+	disabled, err := interp.Evaluate("", exprparser.DefaultStatusCheckSuccess)
+	require.NoError(t, err)
+	assert.Equal(t, false, disabled, "default-success step must be skipped on a cancelled job")
+
+	enabled, err := interp.Evaluate("always()", exprparser.DefaultStatusCheckSuccess)
+	require.NoError(t, err)
+	assert.Equal(t, true, enabled, "`if: always()` step must run on a cancelled job")
+}
+
+// TestMainStepsExecutorRunsAlwaysStepsAfterCancel verifies that newMainStepsExecutor does
+// not abandon the remaining steps when the run is cancelled mid-pipeline. The later step
+// still runs (so a main-stage always() step is reached), it runs under a fresh,
+// non-cancelled context, and the job is marked cancelled. The interrupt error is still
+// propagated so callers up the chain see the cancellation.
+func TestMainStepsExecutorRunsAlwaysStepsAfterCancel(t *testing.T) {
+	rc := createIfTestRunContext(map[string]*model.Job{
+		"job1": createJob(t, `runs-on: ubuntu-latest`, ""),
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	var ran []string
+	var laterStepCtxErr error
+	steps := []common.Executor{
+		func(_ context.Context) error {
+			ran = append(ran, "step1")
+			cancel() // server cancellation lands while step1 runs
+			return nil
+		},
+		func(c context.Context) error {
+			ran = append(ran, "always-step")
+			laterStepCtxErr = c.Err()
+			return nil
+		},
+	}
+
+	err := newMainStepsExecutor(rc, steps)(ctx)
+
+	require.ErrorIs(t, err, context.Canceled, "interrupt error is propagated")
+	assert.Equal(t, []string{"step1", "always-step"}, ran, "the always() step still runs after cancel")
+	require.NoError(t, laterStepCtxErr, "remaining steps run under a fresh, non-cancelled context")
+	assert.True(t, rc.jobCancelled, "the job is marked cancelled")
+}
+
+// TestMainStepsExecutorMarksFailedOnTimeoutBetweenSteps guards the timeout path's symmetry with the cancel path.
+// When the job deadline (timeout-minutes) lands in the gap between two steps, the job must be marked as failed (not cancelled),
+// so always()/failure() cleanup steps run while default success() steps skip, and so the timed-out job is not reported as success.
+func TestMainStepsExecutorMarksFailedOnTimeoutBetweenSteps(t *testing.T) {
+	rc := createIfTestRunContext(map[string]*model.Job{
+		"job1": createJob(t, `runs-on: ubuntu-latest`, ""),
+	})
+
+	// A short deadline that we let elapse between steps, so no step records the error itself.
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Millisecond)
+	defer cancel()
+
+	var ran []string
+	var laterStepCtxErr error
+	steps := []common.Executor{
+		func(c context.Context) error {
+			ran = append(ran, "step1")
+			// Block until the job deadline elapses, then return cleanly: the interrupt lands in the loop's between-steps check, not inside a step.
+			<-c.Done()
+			return nil
+		},
+		func(c context.Context) error {
+			ran = append(ran, "always-step")
+			laterStepCtxErr = c.Err()
+			return nil
+		},
+	}
+
+	err := newMainStepsExecutor(rc, steps)(ctx)
+
+	require.ErrorIs(t, err, context.DeadlineExceeded, "the timeout error is propagated")
+	assert.Equal(t, []string{"step1", "always-step"}, ran, "the always() step still runs after a timeout")
+	require.NoError(t, laterStepCtxErr, "remaining steps run under a fresh, non-expired context")
+	assert.True(t, rc.jobFailed, "a job timeout marks the job failed")
+	assert.False(t, rc.jobCancelled, "a timeout is not a cancellation")
+
+	// The status the real main-step `if` evaluation sees: "failure", so default success() steps skip while always()/failure() steps run.
+	assert.Equal(t, "failure", rc.getJobContext().Status)
+}
+
+// TestStepsExecutorRunsMainStepsAfterPreCancel verifies that a cancellation landing during the
+// pre phase does not abandon the main steps: newStepsExecutor still runs the main-steps executor,
+// so a main-stage always()/cancelled() step is reached (under a fresh, non-cancelled context),
+// the job is marked cancelled, and the cancellation is propagated. Before the fix the `.Then(...)`
+// short-circuit skipped the main steps entirely when a pre step was cancelled.
+func TestStepsExecutorRunsMainStepsAfterPreCancel(t *testing.T) {
+	rc := createIfTestRunContext(map[string]*model.Job{
+		"job1": createJob(t, `runs-on: ubuntu-latest`, ""),
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	var ran []string
+	var mainStepCtxErr error
+	preSteps := []common.Executor{
+		func(_ context.Context) error {
+			ran = append(ran, "pre1")
+			cancel() // server cancellation lands during the pre phase
+			return nil
+		},
+	}
+	steps := []common.Executor{
+		func(c context.Context) error {
+			ran = append(ran, "always-step")
+			mainStepCtxErr = c.Err()
+			return nil
+		},
+	}
+
+	err := newStepsExecutor(rc, preSteps, steps)(ctx)
+
+	require.ErrorIs(t, err, context.Canceled, "the cancellation is propagated")
+	assert.Equal(t, []string{"pre1", "always-step"}, ran, "the main always() step runs after a pre-phase cancel")
+	require.NoError(t, mainStepCtxErr, "the main step runs under a fresh, non-cancelled context")
+	assert.True(t, rc.jobCancelled, "the job is marked cancelled")
+}
+
+// TestStepsExecutorRunsMainStepsAfterPreFailure verifies that a failing pre step does not abandon
+// the main steps: they still run (so a main-stage always()/failure() step is reached), and the
+// pre-step error is propagated so the job is reported as failed. The main steps' own `if`
+// evaluation is what skips success()-default steps, so running them here is safe.
+func TestStepsExecutorRunsMainStepsAfterPreFailure(t *testing.T) {
+	rc := createIfTestRunContext(map[string]*model.Job{
+		"job1": createJob(t, `runs-on: ubuntu-latest`, ""),
+	})
+
+	var ran []string
+	preSteps := []common.Executor{
+		func(_ context.Context) error {
+			ran = append(ran, "pre1")
+			return assert.AnError
+		},
+	}
+	steps := []common.Executor{
+		func(_ context.Context) error {
+			ran = append(ran, "always-step")
+			return nil
+		},
+	}
+
+	err := newStepsExecutor(rc, preSteps, steps)(context.Background())
+
+	require.ErrorIs(t, err, assert.AnError, "the pre-step error is propagated")
+	assert.Equal(t, []string{"pre1", "always-step"}, ran, "the main always() step runs after a pre-step failure")
+	assert.False(t, rc.jobCancelled, "a pre-step failure is not a cancellation")
+}
+
+// TestPreStepFailureAffectsMainStepIfStatus verifies the status path used by real
+// main-step `if` evaluation. A pre-step failure is not present in StepResults, so
+// recording only the context job error is not enough: getJobContext must also report
+// failure so success()-default main steps skip and failure() steps run.
+func TestPreStepFailureAffectsMainStepIfStatus(t *testing.T) {
+	rc := createIfTestRunContext(map[string]*model.Job{
+		"job1": createJob(t, `runs-on: ubuntu-latest`, ""),
+	})
+	ctx := common.WithJobErrorContainer(context.Background())
+
+	reportStepError(ctx, rc, assert.AnError)
+
+	assert.Equal(t, "failure", rc.getJobContext().Status)
+	require.ErrorIs(t, common.JobError(ctx), assert.AnError)
+
+	defaultStep := &stepRun{
+		RunContext: rc,
+		Step:       &model.Step{ID: "default-step"},
+		env:        map[string]string{},
+	}
+	defaultEnabled, err := isStepEnabled(ctx, defaultStep.getIfExpression(ctx, stepStageMain), defaultStep, stepStageMain)
+	require.NoError(t, err)
+	assert.False(t, defaultEnabled, "default success() main step must skip after a pre-step failure")
+
+	failureStep := &stepRun{
+		RunContext: rc,
+		Step: &model.Step{
+			ID: "failure-step",
+			If: yaml.Node{Value: "failure()"},
+		},
+		env: map[string]string{},
+	}
+	failureEnabled, err := isStepEnabled(ctx, failureStep.getIfExpression(ctx, stepStageMain), failureStep, stepStageMain)
+	require.NoError(t, err)
+	assert.True(t, failureEnabled, "failure() main step must run after a pre-step failure")
+}
+
+// TestPostStepsContextCancelledIsUsableForFailingStep guards against a panic: post/cleanup
+// steps run on a context derived from the cancelled job context, and a failing post step
+// records its error via common.SetJobError. If that derived context lacks a job-error container,
+// SetJobError dereferences a nil map and panics. The post context must therefore be detached
+// from cancellation (so the steps run) yet still carry a usable error container.
+func TestPostStepsContextCancelledIsUsableForFailingStep(t *testing.T) {
+	cancelled, cancel := context.WithCancel(common.WithJobErrorContainer(context.Background()))
+	cancel()
+	require.ErrorIs(t, cancelled.Err(), context.Canceled)
+
+	postCtx, done := postStepsContext(cancelled)
+	defer done()
+
+	// Detached from cancellation, so the post steps actually run.
+	require.NoError(t, postCtx.Err(), "post context must not be cancelled")
+
+	// A failing post step records its error instead of panicking.
+	require.NotPanics(t, func() {
+		common.SetJobError(postCtx, assert.AnError)
+	}, "a failing post step must not panic on the cancel path")
+	assert.ErrorIs(t, common.JobError(postCtx), assert.AnError)
+}
+
+// TestPostStepsContextDeadlinePreservesJobError verifies the job-timeout path keeps the original
+// job-error container (via context.WithoutCancel), so the timeout failure and any post-step error
+// survive into the post phase and the job is still reported as failed.
+func TestPostStepsContextDeadlinePreservesJobError(t *testing.T) {
+	base := common.WithJobErrorContainer(context.Background())
+	common.SetJobError(base, assert.AnError)
+	expired, cancel := context.WithDeadline(base, time.Now().Add(-time.Hour))
+	defer cancel()
+	require.ErrorIs(t, expired.Err(), context.DeadlineExceeded)
+
+	postCtx, done := postStepsContext(expired)
+	defer done()
+
+	require.NoError(t, postCtx.Err(), "post context must not carry the expired deadline")
+	assert.ErrorIs(t, common.JobError(postCtx), assert.AnError, "the timeout job error must be preserved")
+}

act/runner/command.go

@@ -48,8 +48,11 @@ func (rc *RunContext) commandHandler(ctx context.Context) common.LineHandler {
 		if resumeCommand != "" && command != resumeCommand {
 			// There should not be any emojis in the log output for Gitea.
 			// The code in the switch statement is the same.
+			// Return true (not false) so the line still reaches the raw_output
+			// log handler; otherwise everything between ::stop-commands:: and
+			// its end token is silently dropped from the step log.
 			logger.Infof("%s", line)
-			return false
+			return true
 		}
 		arg = UnescapeCommandData(arg)
 		kvPairs = unescapeKvPairs(kvPairs)

act/runner/command_test.go

@@ -28,6 +28,29 @@ func TestSetEnv(t *testing.T) {
 	a.Equal("valz", rc.Env["x"])
 }
 
+func TestStopCommandsKeepsSuppressedLinesInLog(t *testing.T) {
+	a := assert.New(t)
+	ctx := context.Background()
+	rc := new(RunContext)
+	handler := rc.commandHandler(ctx)
+
+	// Stop command processing until the matching end token is seen.
+	a.True(handler("::stop-commands::my-end-token\n"))
+
+	// A command-shaped line while stopped must not be executed (env unchanged),
+	// but must still return true so it reaches the raw_output log handler and is
+	// not dropped from the step log.
+	a.True(handler("::set-env name=x::valz\n"))
+	a.NotContains(rc.Env, "x")
+
+	// The matching end token resumes command processing.
+	a.True(handler("::my-end-token::\n"))
+
+	// Commands are processed again after resuming.
+	a.True(handler("::set-env name=y::valy\n"))
+	a.Equal("valy", rc.Env["y"])
+}
+
 func TestSetOutput(t *testing.T) {
 	a := assert.New(t)
 	ctx := context.Background()

act/runner/expression.go

@@ -56,7 +56,7 @@ func (rc *RunContext) NewExpressionEvaluatorWithEnv(ctx context.Context, env map
 		for _, needs := range jobNeeds {
 			using[needs] = exprparser.Needs{
 				Outputs: jobs[needs].Outputs,
-				Result:  jobs[needs].Result,
+				Result:  jobs[needs].NeedsResult(),
 			}
 		}
 
@@ -127,7 +127,7 @@ func (rc *RunContext) NewStepExpressionEvaluator(ctx context.Context, step step)
 	for _, needs := range jobNeeds {
 		using[needs] = exprparser.Needs{
 			Outputs: jobs[needs].Outputs,
-			Result:  jobs[needs].Result,
+			Result:  jobs[needs].NeedsResult(),
 		}
 	}
 

act/runner/job_executor.go

@@ -22,6 +22,7 @@ import (
 
 	"gitea.com/gitea/runner/act/common"
 	"gitea.com/gitea/runner/act/container"
+	"gitea.com/gitea/runner/act/exprparser"
 	"gitea.com/gitea/runner/act/model"
 )
 
@@ -57,9 +58,10 @@ type jobInfo interface {
 
 // reportStepError emits the GitHub Actions ##[error] annotation and records
 // the error against the job so the job is reported as failed.
-func reportStepError(ctx context.Context, err error) {
+func reportStepError(ctx context.Context, rc *RunContext, err error) {
 	common.Logger(ctx).Errorf("##[error]%v", err)
 	common.SetJobError(ctx, err)
+	rc.markFailed()
 }
 
 func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executor {
@@ -117,9 +119,9 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 			rc.CurrentStepIndex = stepIdx
 			preErr := preExec(ctx)
 			if preErr != nil {
-				reportStepError(ctx, preErr)
+				reportStepError(ctx, rc, preErr)
 			} else if ctx.Err() != nil {
-				reportStepError(ctx, ctx.Err())
+				reportStepError(ctx, rc, ctx.Err())
 			}
 			return preErr
 		}))
@@ -129,9 +131,9 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 			rc.CurrentStepIndex = stepIdx
 			err := stepExec(ctx)
 			if err != nil {
-				reportStepError(ctx, err)
+				reportStepError(ctx, rc, err)
 			} else if ctx.Err() != nil {
-				reportStepError(ctx, ctx.Err())
+				reportStepError(ctx, rc, ctx.Err())
 			}
 			return nil
 		}))
@@ -141,9 +143,9 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 			rc.CurrentStepIndex = stepIdx
 			err := postFn(ctx)
 			if err != nil {
-				reportStepError(ctx, err)
+				reportStepError(ctx, rc, err)
 			} else if ctx.Err() != nil {
-				reportStepError(ctx, ctx.Err())
+				reportStepError(ctx, rc, ctx.Err())
 			}
 			return err
 		})
@@ -158,7 +160,12 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 	postExecutor = postExecutor.Finally(func(ctx context.Context) error {
 		jobError := common.JobError(ctx)
 		var err error
-		if rc.Config.AutoRemove || jobError == nil {
+		// jobError == nil keeps a failed job's container alive for post-mortem debugging when
+		// AutoRemove is off (the act-CLI --rm behavior; the shipped runner always sets
+		// AutoRemove). A cancelled run is not a failure to inspect, and the cancel-path post
+		// context now carries its own error container so a failing post step makes jobError
+		// non-nil — OR in rc.jobCancelled so cancellation still always tears the container down.
+		if rc.Config.AutoRemove || jobError == nil || rc.jobCancelled {
 			// always allow 1 min for stopping and removing the runner, even if we were cancelled
 			ctx, cancel := context.WithTimeout(common.WithLogger(context.Background(), common.Logger(ctx)), time.Minute)
 			defer cancel()
@@ -197,25 +204,107 @@ func newJobExecutor(info jobInfo, sf stepFactory, rc *RunContext) common.Executo
 		return err
 	})
 
-	pipeline := make([]common.Executor, 0)
+	stepsExecutor := newStepsExecutor(rc, preSteps, steps)
-	pipeline = append(pipeline, preSteps...)
-	pipeline = append(pipeline, steps...)
 
-	return common.NewPipelineExecutor(info.startContainer(), common.NewPipelineExecutor(pipeline...).
+	return common.NewPipelineExecutor(info.startContainer(), stepsExecutor.
 		Finally(func(ctx context.Context) error {
-			var cancel context.CancelFunc
+			// Record an interrupt (backstop for interrupts that land outside the main
-			if ctx.Err() == context.Canceled {
+			// step loop) so the post steps observe the cancelled/failed job status.
-				// in case of an aborted run, we still should execute the
+			rc.markInterrupted(ctx.Err())
-				// post steps to allow cleanup.
+			postCtx, cancel := postStepsContext(ctx)
-				ctx, cancel = context.WithTimeout(common.WithLogger(context.Background(), common.Logger(ctx)), 5*time.Minute)
 			defer cancel()
-			}
+			return postExecutor(postCtx)
-			return postExecutor(ctx)
 		}).
 		Finally(info.interpolateOutputs()).
 		Finally(info.closeContainer()))
 }
 
+// postStepsContext derives the context used to run the job's post/cleanup steps from the
+// finished main-pipeline context. Cleanup has to run even when the run was interrupted, so the
+// returned context always carries a fresh bounded deadline and is never itself cancelled.
+//
+//   - context.Canceled (server cancel): detach from the cancelled context via a fresh root so
+//     the post steps can run.
+//   - context.DeadlineExceeded (job timeout): detach the deadline with WithoutCancel, which
+//     keeps the original values — including the job-error container — so the timeout failure and
+//     any post-step error are preserved and the job is still reported as failed.
+//   - otherwise: run on the live context unchanged.
+func postStepsContext(ctx context.Context) (context.Context, context.CancelFunc) {
+	switch ctx.Err() {
+	case context.Canceled:
+		// The cancelled context is abandoned for a fresh root, which drops the job-error
+		// container installed at the job root. Re-attach a fresh one so a failing post step
+		// records its error via SetJobError instead of panicking on a nil container.
+		return context.WithTimeout(common.WithJobErrorContainer(common.WithLogger(context.Background(), common.Logger(ctx))), 5*time.Minute)
+	case context.DeadlineExceeded:
+		return context.WithTimeout(context.WithoutCancel(ctx), 5*time.Minute)
+	default:
+		return ctx, func() {}
+	}
+}
+
+// newStepsExecutor sequences the job's pre steps and main steps.
+//
+// The pre steps run as a normal pipeline that short-circuits on the first failure or
+// cancellation. The main-steps executor then runs unconditionally — even if a pre step failed
+// or the job was interrupted — so always()/cancelled()/failure() main steps still run, mirroring
+// GitHub Actions. This is safe because each main step re-evaluates its own `if` (a pre-step
+// failure flips the expression job status to failure, so success()-default steps skip) and
+// newMainStepsExecutor detaches from an interrupted context before running the remaining steps.
+//
+// A pre-step failure or interrupt is still propagated so the job is reported with the correct
+// conclusion; the pre error takes precedence since it happened first.
+func newStepsExecutor(rc *RunContext, preSteps, steps []common.Executor) common.Executor {
+	preExecutor := common.NewPipelineExecutor(preSteps...)
+	mainExecutor := newMainStepsExecutor(rc, steps)
+	return func(ctx context.Context) error {
+		preErr := preExecutor(ctx)
+		mainErr := mainExecutor(ctx)
+		if preErr != nil {
+			return preErr
+		}
+		return mainErr
+	}
+}
+
+// newMainStepsExecutor runs the job's main-stage step executors in order. Unlike a plain
+// pipeline, an interruption (context.Canceled from a server cancel, or context.DeadlineExceeded
+// from the job timeout) does not abandon the remaining steps: it marks the job cancelled when
+// appropriate and keeps iterating under a fresh, bounded context so steps whose `if` still
+// evaluates true — always() and cancelled() — run for cleanup, mirroring GitHub Actions. Steps
+// that default to success() skip themselves because success() is false once the job is no longer
+// successful. The main-step wrappers report their own errors and return nil, so the loop drives
+// step ordering off the context, not return values.
+func newMainStepsExecutor(rc *RunContext, steps []common.Executor) common.Executor {
+	return func(ctx context.Context) error {
+		for i, step := range steps {
+			if ctx.Err() != nil {
+				return runMainStepsAfterInterrupt(ctx, rc, steps[i:])
+			}
+			_ = step(ctx)
+		}
+		// An interrupt can land during the final step, after the loop's last context
+		// check; record it so the post steps still observe the cancelled/failed status.
+		rc.markInterrupted(ctx.Err())
+		return nil
+	}
+}
+
+// runMainStepsAfterInterrupt runs the remaining main steps after the job context was cancelled or
+// timed out. It detaches from the interrupted context (keeping its values: logger and job error)
+// and applies a fresh deadline so always()/cancelled() steps run to completion. The original
+// interrupt error is returned so callers up the chain still see the job as cancelled/timed out.
+func runMainStepsAfterInterrupt(ctx context.Context, rc *RunContext, steps []common.Executor) error {
+	interruptErr := ctx.Err()
+	rc.markInterrupted(interruptErr)
+	freshCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 5*time.Minute)
+	defer cancel()
+	for _, step := range steps {
+		_ = step(freshCtx)
+	}
+	return interruptErr
+}
+
 func setJobResult(ctx context.Context, info jobInfo, rc *RunContext, success bool) {
 	logger := common.Logger(ctx)
 
@@ -223,6 +312,12 @@ func setJobResult(ctx context.Context, info jobInfo, rc *RunContext, success boo
 	// read-modify-write of the job result so a failing combination is not lost-updated by a
 	// concurrent succeeding one.
 	job := rc.Run.Job()
+	var continueOnError bool
+	if !success {
+		// Use a fresh context so an expired job timeout cannot block expression evaluation.
+		evalCtx := common.WithLogger(context.Background(), common.Logger(ctx))
+		continueOnError = evaluateJobContinueOnError(evalCtx, rc, job)
+	}
 	jobResult := func() string {
 		defer lockJob(job)()
 		result := "success"
@@ -233,6 +328,7 @@ func setJobResult(ctx context.Context, info jobInfo, rc *RunContext, success boo
 		}
 		if !success {
 			result = "failure"
+			job.SetContinueOnError(continueOnError)
 		}
 		info.result(result)
 		return result
@@ -271,6 +367,32 @@ func setJobOutputs(ctx context.Context, rc *RunContext) {
 	}
 }
 
+// applyJobTimeout applies the job-level timeout-minutes to ctx, mirroring the
+// step-level evaluateStepTimeout in step.go.
+func applyJobTimeout(ctx context.Context, rc *RunContext, job *model.Job) (context.Context, context.CancelFunc) {
+	timeout := rc.ExprEval.Interpolate(ctx, job.TimeoutMinutes)
+	if timeout != "" {
+		if timeoutMinutes, err := strconv.ParseInt(timeout, 10, 64); err == nil {
+			return context.WithTimeout(ctx, time.Duration(timeoutMinutes)*time.Minute)
+		}
+	}
+	return ctx, func() {}
+}
+
+// evaluateJobContinueOnError evaluates the job-level continue-on-error expression.
+func evaluateJobContinueOnError(ctx context.Context, rc *RunContext, job *model.Job) bool {
+	expr := strings.TrimSpace(job.RawContinueOnError)
+	if expr == "" {
+		return false
+	}
+	continueOnError, err := EvalBool(ctx, rc.NewExpressionEvaluator(ctx), expr, exprparser.DefaultStatusCheckNone)
+	if err != nil {
+		common.Logger(ctx).Warnf("continue-on-error expression %q evaluation failed: %v", expr, err)
+		return false
+	}
+	return continueOnError
+}
+
 func tryUploadJobSummary(ctx context.Context, rc *RunContext) {
 	if rc == nil || rc.JobContainer == nil || rc.Config == nil {
 		return
@@ -462,6 +584,11 @@ func useStepLogger(rc *RunContext, stepModel *model.Step, stage stepStage, execu
 		oldout, olderr := rc.JobContainer.ReplaceLogWriter(logWriter, logWriter)
 		defer rc.JobContainer.ReplaceLogWriter(oldout, olderr)
 
+		// Flush any buffered, not-yet-newline-terminated trailing line once the
+		// step has finished, so the final line of the step's output is not lost
+		// when it is not newline-terminated.
+		defer common.FlushWriter(logWriter)
+
 		return executor(ctx)
 	}
 }

act/runner/job_executor_test.go

@@ -28,6 +28,7 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/mock"
 	"github.com/stretchr/testify/require"
+	yaml "go.yaml.in/yaml/v4"
 )
 
 func TestJobExecutor(t *testing.T) {
@@ -347,6 +348,133 @@ func TestNewJobExecutor(t *testing.T) {
 	}
 }
 
+// TestNewJobExecutorRunsPostStepsAfterTimeout guards the timeout-minutes cleanup
+// path: when a job exceeds its timeout the job context is DeadlineExceeded, but
+// the post steps (cleanup hooks like actions/checkout post and cache save) must
+// still run against a fresh, non-expired context, and the job must still be
+// reported as failed.
+func TestNewJobExecutorRunsPostStepsAfterTimeout(t *testing.T) {
+	ctx := common.WithJobErrorContainer(context.Background())
+	// The timeout is generous so the main step (which blocks on ctx.Done below) is
+	// always reached before the deadline fires; otherwise the pipeline would
+	// short-circuit before the step runs and the job error would never be set.
+	ctx, cancel := context.WithTimeout(ctx, 200*time.Millisecond)
+	defer cancel()
+
+	jim := &jobInfoMock{}
+	sfm := &stepFactoryMock{}
+	rc := &RunContext{
+		JobContainer: &jobContainerMock{},
+		Run: &model.Run{
+			JobID: "test",
+			Workflow: &model.Workflow{
+				Jobs: map[string]*model.Job{
+					"test": {},
+				},
+			},
+		},
+		Config: &Config{},
+	}
+	rc.ExprEval = rc.NewExpressionEvaluator(ctx)
+
+	stepModel := &model.Step{ID: "1"}
+	jim.On("steps").Return([]*model.Step{stepModel})
+	jim.On("matrix").Return(map[string]any{})
+	jim.On("startContainer").Return(func(ctx context.Context) error { return nil })
+	jim.On("interpolateOutputs").Return(func(ctx context.Context) error { return nil })
+	jim.On("closeContainer").Return(func(ctx context.Context) error { return nil })
+	// The job timed out, so it must be reported as failed. stopContainer is left
+	// unexpected on purpose: a timed-out (failed) job preserves its error state, so
+	// the graceful stop is skipped exactly like any other failure without AutoRemove.
+	jim.On("result", "failure")
+
+	sm := &stepMock{}
+	sfm.On("newStep", stepModel, rc).Return(sm, nil)
+	sm.On("pre").Return(func(ctx context.Context) error { return nil })
+	// The main step runs past the job timeout: it blocks until the job context is
+	// done, mirroring a step that overruns timeout-minutes.
+	sm.On("main").Return(func(ctx context.Context) error {
+		<-ctx.Done()
+		return ctx.Err()
+	})
+
+	var postRan bool
+	var postCtxErr error
+	sm.On("post").Return(func(ctx context.Context) error {
+		postRan = true
+		postCtxErr = ctx.Err()
+		return nil
+	})
+
+	executor := newJobExecutor(jim, sfm, rc)
+	// The executor itself returns nil on timeout: the failure is surfaced through
+	// the job result ("failure", asserted via the result mock below), not the
+	// return value.
+	require.NoError(t, executor(ctx))
+
+	assert.True(t, postRan, "post step must run after a job timeout")
+	require.NoError(t, postCtxErr, "post step must run against a fresh, non-expired context")
+
+	jim.AssertExpectations(t)
+	sfm.AssertExpectations(t)
+	sm.AssertExpectations(t)
+}
+
+// TestSetJobResultMatrixContinueOnError exercises the parallel-matrix path
+// end-to-end: two combinations share one *model.Job and continue-on-error is
+// keyed on matrix.experimental, so one combination tolerates its failure and the
+// other does not. The job is reported as continue-on-error only when EVERY failing
+// combination was tolerated; a single firm failure makes the whole job firm, and
+// handleFailure then fails the run.
+func TestSetJobResultMatrixContinueOnError(t *testing.T) {
+	const jobYAML = "continue-on-error: ${{ matrix.experimental }}\nruns-on: ubuntu-latest"
+
+	newSharedJob := func(t *testing.T) (*model.Job, *model.Workflow) {
+		t.Helper()
+		var job *model.Job
+		require.NoError(t, yaml.Unmarshal([]byte(jobYAML), &job))
+		return job, &model.Workflow{
+			Name: "workflow1",
+			Jobs: map[string]*model.Job{"job1": job},
+		}
+	}
+
+	planFor := func(wf *model.Workflow) *model.Plan {
+		return &model.Plan{Stages: []*model.Stage{{Runs: []*model.Run{{Workflow: wf, JobID: "job1"}}}}}
+	}
+
+	ctx := context.Background()
+
+	// fail drives a single matrix combination through the failure path; each
+	// RunContext is its own jobInfo (rc implements jobInfo) and shares the job.
+	fail := func(wf *model.Workflow, experimental bool) {
+		rc := newTestRC(wf, map[string]any{"experimental": experimental})
+		setJobResult(ctx, rc, rc, false)
+	}
+
+	t.Run("one tolerated and one firm failure fails the run", func(t *testing.T) {
+		job, wf := newSharedJob(t)
+		// Order is intentional: the tolerated combination finishes first, then the
+		// firm one. The firm-failure latch must still win regardless of order.
+		fail(wf, true)
+		fail(wf, false)
+
+		assert.Equal(t, "failure", job.Result)
+		assert.False(t, job.ContinueOnError, "a single firm failure must make the whole job firm")
+		assert.Error(t, handleFailure(planFor(wf))(ctx))
+	})
+
+	t.Run("all tolerated failures do not fail the run", func(t *testing.T) {
+		job, wf := newSharedJob(t)
+		fail(wf, true)
+		fail(wf, true)
+
+		assert.Equal(t, "failure", job.Result)
+		assert.True(t, job.ContinueOnError, "every failing combination was tolerated")
+		assert.NoError(t, handleFailure(planFor(wf))(ctx))
+	})
+}
+
 func TestHasJobSummaryCapability(t *testing.T) {
 	assert.True(t, hasJobSummaryCapability("cache,job-summary artifacts"))
 	assert.True(t, hasJobSummaryCapability("cache,\njob-summary\tartifacts"))
@@ -674,3 +802,104 @@ func tarArchive(t *testing.T, entries ...tarEntry) []byte {
 	require.NoError(t, tw.Close())
 	return buf.Bytes()
 }
+
+func newTestRC(wf *model.Workflow, matrix map[string]any) *RunContext {
+	return &RunContext{
+		Config: &Config{
+			Workdir: ".",
+			Platforms: map[string]string{
+				"ubuntu-latest": "ubuntu-latest",
+			},
+		},
+		StepResults: map[string]*model.StepResult{},
+		Env:         map[string]string{},
+		Matrix:      matrix,
+		Run:         &model.Run{JobID: "job1", Workflow: wf},
+	}
+}
+
+func makeTestRC(t *testing.T, jobYAML string) *RunContext {
+	t.Helper()
+	var job *model.Job
+	require.NoError(t, yaml.Unmarshal([]byte(jobYAML), &job))
+	rc := newTestRC(&model.Workflow{
+		Name: "workflow1",
+		Jobs: map[string]*model.Job{"job1": job},
+	}, nil)
+	rc.ExprEval = rc.NewExpressionEvaluator(context.Background())
+	return rc
+}
+
+func TestApplyJobTimeout(t *testing.T) {
+	cases := []struct {
+		name        string
+		yaml        string
+		wantTimeout bool
+	}{
+		{"empty", "runs-on: ubuntu-latest", false},
+		{"integer", "timeout-minutes: 5\nruns-on: ubuntu-latest", true},
+		{"non-numeric ignored", "timeout-minutes: abc\nruns-on: ubuntu-latest", false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			rc := makeTestRC(t, tc.yaml)
+			ctx := context.Background()
+			newCtx, cancel := applyJobTimeout(ctx, rc, rc.Run.Job())
+			defer cancel()
+			_, hasDeadline := newCtx.Deadline()
+			assert.Equal(t, tc.wantTimeout, hasDeadline)
+		})
+	}
+}
+
+func TestEvaluateJobContinueOnError(t *testing.T) {
+	cases := []struct {
+		name string
+		yaml string
+		want bool
+	}{
+		{"absent", "runs-on: ubuntu-latest", false},
+		{"true", "continue-on-error: true\nruns-on: ubuntu-latest", true},
+		{"false", "continue-on-error: false\nruns-on: ubuntu-latest", false},
+		{"expression true", "continue-on-error: ${{ 'x' == 'x' }}\nruns-on: ubuntu-latest", true},
+		{"expression false", "continue-on-error: ${{ 'x' != 'x' }}\nruns-on: ubuntu-latest", false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			rc := makeTestRC(t, tc.yaml)
+			got := evaluateJobContinueOnError(context.Background(), rc, rc.Run.Job())
+			assert.Equal(t, tc.want, got)
+		})
+	}
+}
+
+func TestJobSetContinueOnError(t *testing.T) {
+	t.Run("first call true", func(t *testing.T) {
+		j := &model.Job{}
+		j.SetContinueOnError(true)
+		assert.True(t, j.ContinueOnError)
+	})
+	t.Run("first call false", func(t *testing.T) {
+		j := &model.Job{}
+		j.SetContinueOnError(false)
+		assert.False(t, j.ContinueOnError)
+	})
+	t.Run("true then false locks to false", func(t *testing.T) {
+		j := &model.Job{}
+		j.SetContinueOnError(true)
+		j.SetContinueOnError(false)
+		assert.False(t, j.ContinueOnError)
+	})
+	t.Run("false then true stays false", func(t *testing.T) {
+		j := &model.Job{}
+		j.SetContinueOnError(false)
+		j.SetContinueOnError(true)
+		assert.False(t, j.ContinueOnError)
+	})
+	t.Run("true then true stays true", func(t *testing.T) {
+		j := &model.Job{}
+		j.SetContinueOnError(true)
+		j.SetContinueOnError(true)
+		assert.True(t, j.ContinueOnError)
+	})
+}

act/runner/reusable_workflow.go

@@ -141,6 +141,7 @@ func cloneRemoteReusableWorkflow(rc *RunContext, cloneURL, ref, targetDirectory,
 			Dir:         targetDirectory,
 			Token:       token,
 			OfflineMode: rc.Config.ActionOfflineMode,
+			Depth:       rc.Config.ActionCloneDepth,
 		})(ctx)
 	}
 }

act/runner/run_context.go

@@ -73,6 +73,39 @@ type RunContext struct {
 	// captured before execution so each matrix combo interpolates from the originals rather
 	// than from a sibling's already-resolved values written into the shared Job.Outputs.
 	outputTemplate map[string]string
+	// jobCancelled records that this job's run was cancelled (context.Canceled). It makes
+	// getJobContext report the "cancelled" status so cancelled()/always() evaluate the way
+	// GitHub Actions does, letting cleanup and always() steps run while normal steps skip.
+	jobCancelled bool
+	// jobFailed records failures outside normal main-step results, such as action pre-step
+	// failures. Those failures must still make success() false and failure() true for later
+	// main-step if evaluation.
+	jobFailed bool
+}
+
+// markCancelled flags the job as cancelled so subsequent step `if` evaluations and the
+// job status context observe the "cancelled" state.
+func (rc *RunContext) markCancelled() {
+	rc.jobCancelled = true
+}
+
+// markFailed flags the job as failed so subsequent step `if` evaluations observe
+// failure even when the error happened outside a main step result.
+func (rc *RunContext) markFailed() {
+	rc.jobFailed = true
+}
+
+// markInterrupted records the job's interruption status from a context error so later step `if` evaluations and the job result observe it,
+// keeping the timeout path symmetric with the cancel path:
+//   - context.Canceled (server cancel) marks the job cancelled, matching GitHub's "only always()/cancelled() run on cancel".
+//   - context.DeadlineExceeded (job timeout-minutes) marks the job failed, matching the "Timeout -> FAILURE" reporting semantics.
+func (rc *RunContext) markInterrupted(err error) {
+	switch {
+	case errors.Is(err, context.Canceled):
+		rc.markCancelled()
+	case errors.Is(err, context.DeadlineExceeded):
+		rc.markFailed()
+	}
 }
 
 func (rc *RunContext) AddMask(mask string) {
@@ -189,6 +222,9 @@ func (rc *RunContext) GetBindsAndMounts() ([]string, map[string]string) {
 	if job := rc.Run.Job(); job != nil {
 		if container := job.Container(); container != nil {
 			for _, v := range container.Volumes {
+				if rc.ExprEval != nil {
+					v = rc.ExprEval.Interpolate(context.Background(), v)
+				}
 				if !strings.Contains(v, ":") || filepath.IsAbs(v) {
 					// Bind anonymous volume or host file.
 					binds = append(binds, v)
@@ -471,7 +507,8 @@ func (rc *RunContext) startJobContainer() common.Executor {
 			rc.pullServicesImages(rc.Config.ForcePull),
 			rc.JobContainer.Pull(rc.Config.ForcePull),
 			rc.stopJobContainer(),
-			container.NewDockerNetworkCreateExecutor(networkName).IfBool(createAndDeleteNetwork),
+			container.NewDockerNetworkCreateExecutor(networkName, rc.Config.ContainerNetworkCreateOptions).
+				IfBool(createAndDeleteNetwork),
 			rc.startServiceContainers(networkName),
 			rc.JobContainer.Create(rc.Config.ContainerCapAdd, rc.Config.ContainerCapDrop),
 			rc.JobContainer.Start(false),
@@ -900,12 +937,21 @@ func trimToLen(s string, l int) string {
 
 func (rc *RunContext) getJobContext() *model.JobContext {
 	jobStatus := "success"
+	if rc.jobFailed {
+		jobStatus = "failure"
+	}
 	for _, stepStatus := range rc.StepResults {
 		if stepStatus.Conclusion == model.StepStatusFailure {
 			jobStatus = "failure"
 			break
 		}
 	}
+	// A cancelled run takes precedence over success/failure so cancelled() is true and
+	// success()/failure() are false, matching GitHub Actions: on cancellation only
+	// always() and cancelled() steps run.
+	if rc.jobCancelled {
+		jobStatus = "cancelled"
+	}
 	return &model.JobContext{
 		Status: jobStatus,
 	}

act/runner/run_context_test.go

@@ -276,6 +276,37 @@ func TestRunContext_GetBindsAndMounts(t *testing.T) {
 			{"MountExistingVolume", []string{"volume-id:/volume"}, "", map[string]string{"volume-id": "/volume"}},
 		}
 
+		t.Run("InterpolatedContainerVolumes", func(t *testing.T) {
+			job := &model.Job{}
+			err := job.RawContainer.Encode(map[string][]string{
+				"volumes": {"${{ secrets.MAME }}:/root/.mame/roms:ro"},
+			})
+			require.NoError(t, err)
+
+			rc := &RunContext{
+				Name: "TestRCName",
+				Run: &model.Run{
+					Workflow: &model.Workflow{
+						Name: "TestWorkflowName",
+					},
+				},
+				Config: &Config{
+					BindWorkdir: false,
+					Secrets: map[string]string{
+						"MAME": "/host/mame/roms",
+					},
+				},
+			}
+			rc.Run.JobID = "job1"
+			rc.Run.Workflow.Jobs = map[string]*model.Job{"job1": job}
+			rc.ExprEval = rc.NewExpressionEvaluator(context.Background())
+
+			gotbind, gotmount := rc.GetBindsAndMounts()
+			assert.Contains(t, gotbind, "/host/mame/roms:/root/.mame/roms:ro")
+			assert.NotContains(t, gotbind, "${{ secrets.MAME }}")
+			assert.NotContains(t, gotmount, "${{ secrets.MAME }}")
+		})
+
 		for _, testcase := range tests {
 			t.Run(testcase.name, func(t *testing.T) {
 				job := &model.Job{}

act/runner/runner.go

@@ -15,6 +15,7 @@ import (
 	"time"
 
 	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/container"
 	"gitea.com/gitea/runner/act/model"
 
 	docker_container "github.com/moby/moby/api/types/container"
@@ -32,6 +33,7 @@ type Config struct {
 	Workdir                            string                                        // path to working directory
 	ActionCacheDir                     string                                        // path used for caching action contents
 	ActionOfflineMode                  bool                                          // when offline, use cached action contents
+	ActionCloneDepth                   int                                           // limit history when cloning an action repo; 0 clones every branch in full
 	BindWorkdir                        bool                                          // bind the workdir to the job container
 	EventName                          string                                        // name of event to run
 	EventPath                          string                                        // path to JSON file to use for event.json in containers
@@ -68,6 +70,7 @@ type Config struct {
 	ReplaceGheActionTokenWithGithubCom string                                        // Token of private action repo on GitHub.
 	Matrix                             map[string]map[string]bool                    // Matrix config to run
 	ContainerNetworkMode               docker_container.NetworkMode                  // the network mode of job containers (the value of --network)
+	ContainerNetworkCreateOptions      container.NewDockerNetworkCreateExecutorInput // the default network create options
 	ActionCache                        ActionCache                                   // Use a custom ActionCache Implementation
 
 	PresetGitHubContext   *model.GithubContext         // the preset github context, overrides some fields like DefaultBranch, Env, Secrets etc.
@@ -248,7 +251,10 @@ func (runner *runnerImpl) NewPlanExecutor(plan *model.Plan) common.Executor {
 							return err
 						}
 
-						return executor(common.WithJobErrorContainer(WithJobLogger(ctx, rc.Run.JobID, jobName, rc.Config, &rc.Masks, matrix)))
+						jobCtx := common.WithJobErrorContainer(WithJobLogger(ctx, rc.Run.JobID, jobName, rc.Config, &rc.Masks, matrix))
+						jobCtx, cancelTimeout := applyJobTimeout(jobCtx, rc, job)
+						defer cancelTimeout()
+						return executor(jobCtx)
 					})
 				}
 				// Run all matrix combinations of this job, then drop its aggregation mutex: the
@@ -303,7 +309,7 @@ func handleFailure(plan *model.Plan) common.Executor {
 	return func(ctx context.Context) error {
 		for _, stage := range plan.Stages {
 			for _, run := range stage.Runs {
-				if run.Job().Result == "failure" {
+				if run.Job().Result == "failure" && !run.Job().ContinueOnError {
 					return fmt.Errorf("Job '%s' failed", run.String())
 				}
 			}

act/runner/step_action_remote.go

@@ -114,13 +114,23 @@ func (sar *stepActionRemote) prepareActionExecutor() common.Executor {
 
 		actionDir := fmt.Sprintf("%s/%s", sar.RunContext.ActionCacheDir(), sar.Step.UsesHash())
 		defaultActionURL := sar.RunContext.Config.DefaultActionURL()
-		token := getGitCloneToken(sar.getRunContext().Config, sar.remoteAction.CloneURL(defaultActionURL))
+		// For Gitea
+		// A composite RunContext nils Config.Secrets, so getGitCloneToken would yield an
+		// empty token and clone the action anonymously (401 against the authenticated
+		// instance). github.Token survives the composite config copy and matches the
+		// top-level token; keep the shouldCloneURLUseToken host gate to avoid leaking it.
+		cloneURL := sar.remoteAction.CloneURL(defaultActionURL)
+		token := ""
+		if shouldCloneURLUseToken(sar.RunContext.Config.GitHubInstance, cloneURL) {
+			token = github.Token
+		}
 		gitClone := stepActionRemoteNewCloneExecutor(git.NewGitCloneExecutorInput{
-			URL:         sar.remoteAction.CloneURL(defaultActionURL),
+			URL:         cloneURL,
 			Ref:         sar.remoteAction.Ref,
 			Dir:         actionDir,
 			Token:       token,
 			OfflineMode: sar.RunContext.Config.ActionOfflineMode,
+			Depth:       sar.RunContext.Config.ActionCloneDepth,
 
 			InsecureSkipTLS: sar.cloneSkipTLS(), // For Gitea
 		})
@@ -312,7 +322,7 @@ func (ra *remoteAction) IsCheckout() bool {
 
 func newRemoteAction(action string) *remoteAction {
 	// support http(s)://host/owner/repo@v3
-	for _, schema := range []string{"https://", "http://"} {
+	for _, schema := range []string{"https://", "http://", "ssh://"} {
 		if after, ok := strings.CutPrefix(action, schema); ok {
 			splits := strings.SplitN(after, "/", 2)
 			if len(splits) != 2 {

act/runner/step_action_remote_test.go

@@ -778,6 +778,32 @@ func Test_newRemoteAction(t *testing.T) {
 			},
 			wantCloneURL: "http://gitea.com/actions/aws",
 		},
+		{
+			action: "ssh://git@gitea.com/actions/heroku@main", // it's invalid for GitHub, but gitea supports it
+			want: &remoteAction{
+				URL:  "ssh://git@gitea.com",
+				Org:  "actions",
+				Repo: "heroku",
+				Path: "",
+				Ref:  "main",
+			},
+			wantCloneURL: "ssh://git@gitea.com/actions/heroku",
+		},
+		{
+			action: "ssh://git@gitea.com/actions/aws/ec2@main", // the ssh user is kept as part of the host segment
+			want: &remoteAction{
+				URL:  "ssh://git@gitea.com",
+				Org:  "actions",
+				Repo: "aws",
+				Path: "ec2",
+				Ref:  "main",
+			},
+			wantCloneURL: "ssh://git@gitea.com/actions/aws",
+		},
+		{
+			action: "ssh://gitea.com/onlyonesegment@main", // missing org/repo after the host
+			want:   nil,
+		},
 	}
 	for _, tt := range tests {
 		t.Run(tt.action, func(t *testing.T) {
@@ -812,3 +838,83 @@ func Test_safeFilename(t *testing.T) {
 		})
 	}
 }
+
+// Regression: a nested action in a composite cloned anonymously (401) because the
+// composite RunContext nils Config.Secrets. The token must come from github.Token,
+// which survives the config copy; the host gate must still withhold it cross-host.
+func TestStepActionRemoteCloneTokenSurvivesNilSecrets(t *testing.T) {
+	const wantToken = "job-token"
+
+	table := []struct {
+		name                  string
+		gitHubInstance        string
+		defaultActionInstance string
+		wantCloneToken        string
+	}{
+		{
+			name:           "same host forwards token despite nil secrets",
+			gitHubInstance: "gitea.example.com",
+			wantCloneToken: wantToken,
+		},
+		{
+			name:                  "foreign host is not given the token",
+			gitHubInstance:        "gitea.example.com",
+			defaultActionInstance: "github.com",
+			wantCloneToken:        "",
+		},
+	}
+
+	for _, tt := range table {
+		t.Run(tt.name, func(t *testing.T) {
+			ctx := context.Background()
+
+			var capturedToken string
+			origStepAtionRemoteNewCloneExecutor := stepActionRemoteNewCloneExecutor
+			stepActionRemoteNewCloneExecutor = func(input git.NewGitCloneExecutorInput) common.Executor {
+				capturedToken = input.Token
+				return func(ctx context.Context) error { return nil }
+			}
+			defer (func() {
+				stepActionRemoteNewCloneExecutor = origStepAtionRemoteNewCloneExecutor
+			})()
+
+			sarm := &stepActionRemoteMocks{}
+			sar := &stepActionRemote{
+				Step: &model.Step{Uses: "org/repo@v1"},
+				RunContext: &RunContext{
+					Config: &Config{
+						GitHubInstance:        tt.gitHubInstance,
+						DefaultActionInstance: tt.defaultActionInstance,
+						ActionCacheDir:        "/tmp/test-cache",
+						// Mirrors the state of a composite RunContext: job secrets are
+						// stripped, but the job token is still reachable via Config.Token.
+						Secrets: nil,
+						Token:   wantToken,
+					},
+					Run: &model.Run{
+						JobID: "1",
+						Workflow: &model.Workflow{
+							Jobs: map[string]*model.Job{"1": {}},
+						},
+					},
+					StepResults: map[string]*model.StepResult{},
+				},
+				readAction: sarm.readAction,
+			}
+			sar.RunContext.ExprEval = sar.RunContext.NewExpressionEvaluator(ctx)
+
+			suffixMatcher := func(suffix string) any {
+				return mock.MatchedBy(func(actionDir string) bool {
+					return strings.HasSuffix(actionDir, suffix)
+				})
+			}
+			sarm.On("readAction", sar.Step, suffixMatcher(sar.Step.UsesHash()), "", mock.Anything, mock.Anything).Return(&model.Action{}, nil)
+
+			err := sar.prepareActionExecutor()(ctx)
+			require.NoError(t, err)
+			assert.Equal(t, tt.wantCloneToken, capturedToken)
+
+			sarm.AssertExpectations(t)
+		})
+	}
+}

docs/post-task-script.md

@@ -0,0 +1,155 @@
+# Post-task script
+
+The post-task script is an optional host hook that runs **once after every task**, after the runner has already finished its normal per-task cleanup. Typical uses include pruning Docker images, vacuuming ephemeral disks, or resetting VM state between jobs.
+
+It is configured under `runner.post_task_script` in the runner YAML config (see [config.example.yaml](../internal/pkg/config/config.example.yaml)).
+
+## When it runs
+
+For each task, execution order is:
+
+1. Workflow runs (steps, actions, containers).
+2. In-job cleanup (action `post:` steps, container stop/remove).
+3. Job outputs are reported to Gitea.
+4. Bind-workdir workspace removal, when `container.bind_workdir` is enabled.
+5. **Post-task script** (this hook).
+6. Final task acknowledgement to Gitea (`reporter.Close()`).
+
+The script is **additive**: it does not replace any built-in cleanup. When `container.bind_workdir` is enabled, the task workspace directory has usually already been deleted before the script starts. `GITEA_WORKSPACE` is still set to the path the job used, for reference.
+
+## Runner stays offline until the script finishes
+
+This is the most important operational detail.
+
+When the post-task script starts, the runner **stops sending task heartbeats** to Gitea (the same mechanism used during cancel/cleanup). From Gitea's perspective, the runner is **not available for new work** until:
+
+1. The script exits (success or failure), **and**
+2. The runner sends the final task flush to Gitea.
+
+While the script runs:
+
+- **Gitea will not assign another task** to this runner for the current job slot (heartbeats are stopped).
+- **The runner capacity slot stays occupied** locally — with `capacity: 1`, the poller will not start another task until the script completes.
+- **Runner shutdown** (`shutdown_timeout`) counts this phase as part of the in-flight task; a long or stuck script delays graceful shutdown.
+
+If the script **never exits**, the runner remains in this state until `runner.post_task_script_timeout` elapses (default **5 minutes** when a script is configured). The runner then kills the script process and proceeds to the final acknowledgement. Until that timeout fires, **the runner effectively stays offline**.
+
+Set `post_task_script_timeout` to a value that matches how long your housekeeping is allowed to take — not how long you wish it could take. Prefer short, bounded scripts.
+
+### Recommendations
+
+- Keep scripts **fast and bounded** (seconds, not minutes).
+- Avoid interactive prompts, blocking network calls without timeouts, or waiting on user input.
+- Use **idempotent** operations (the script may run after success, failure, or cancellation).
+- Test failure modes: hung script, non-zero exit, missing executable.
+- Watch the **runner process log** for script output (it is not written to the Gitea job log).
+- On shutdown, ensure scripts respond to process termination within `post_task_script_timeout`.
+
+## Configuration
+
+```yaml
+runner:
+  # Path to an executable on the host. Empty or omitted disables the hook.
+  post_task_script: /usr/local/bin/gitea-post-task.sh
+
+  # Hard limit on script runtime. Default when post_task_script is set: 5m.
+  # If the script exceeds this, it is killed and the runner continues.
+  post_task_script_timeout: 2m
+```
+
+| Option | Default | Description |
+| --- | --- | --- |
+| `runner.post_task_script` | *(disabled)* | Host path to the script or binary. Relative paths are resolved from the runner process working directory. |
+| `runner.post_task_script_timeout` | `5m` (only when script is set) | Maximum time the script may run before the runner kills it and moves on. |
+
+The script must be **executable** on the host (shebang on Linux/macOS, or a native `.exe` / `.bat` / `.cmd` on Windows). **PowerShell (`.ps1`) is not supported yet** as the value of `post_task_script`; the runner executes the configured path directly and does not invoke `powershell.exe` for you.
+
+`gitea-runner exec` does **not** load runner YAML and will not run this hook.
+
+## Environment variables
+
+The script receives `runner.envs` / `runner.env_file` values plus:
+
+| Variable | Description |
+| --- | --- |
+| `GITEA_TASK_ID` | Numeric task ID. |
+| `GITEA_RUN_ID` | Workflow run ID, when provided by the server. |
+| `GITEA_REPOSITORY` | Repository slug (`owner/name`). |
+| `GITEA_WORKSPACE` | Workspace path used for the job (may already be deleted). |
+| `GITEA_JOB_RESULT` | `success`, `failure`, `cancelled`, `skipped`, or `unknown`. |
+
+The script environment is **not** a full copy of the job container environment. System variables such as `PATH` are only present if you define them in `runner.envs` or `runner.env_file`.
+
+## Output and errors
+
+- **Stdout/stderr** are written to the **runner process log** (logrus), prefixed with `post-task script stdout:` / `post-task script stderr:`.
+- **Non-zero exit codes** are logged as warnings only. They do **not** change the job result already reported to Gitea.
+- **Timeouts and start failures** are logged as warnings; the runner still completes the task acknowledgement.
+
+## Interaction with other timeouts
+
+| Timeout | Effect on post-task script |
+| --- | --- |
+| `runner.post_task_script_timeout` | Kills the script if it runs too long. This is the **only** timeout that bounds the script. |
+| `runner.timeout` | Caps the task **up to** the script. The script detaches from the task deadline, so a job near the runner timeout limit does **not** cut the script short — it still gets its full `post_task_script_timeout`. |
+| `runner.shutdown_timeout` | On SIGINT/SIGTERM, bounds how long the runner waits for the **task** to finish. The post-task script detaches from cancellation, so it is **not** interrupted by this window and may extend shutdown until its own `post_task_script_timeout` elapses. |
+
+## Examples
+
+### Linux — prune dangling Docker resources
+
+`/usr/local/bin/gitea-post-task.sh`:
+
+```sh
+#!/bin/sh
+set -eu
+docker image prune -f
+docker builder prune -f --filter 'until=24h'
+```
+
+`config.yaml`:
+
+```yaml
+runner:
+  post_task_script: /usr/local/bin/gitea-post-task.sh
+  post_task_script_timeout: 3m
+```
+
+### Windows — batch file (`.cmd`)
+
+Use a `.cmd` or `.bat` file. PowerShell scripts are **not supported yet** as `post_task_script`; call PowerShell from a batch wrapper if needed:
+
+`C:\gitea-runner\scripts\post-task.cmd`:
+
+```bat
+@echo off
+docker image prune -f
+```
+
+```yaml
+runner:
+  post_task_script: C:\gitea-runner\scripts\post-task.cmd
+  post_task_script_timeout: 3m
+```
+
+PowerShell workaround until native `.ps1` support exists:
+
+`C:\gitea-runner\scripts\post-task.cmd`:
+
+```bat
+@echo off
+powershell.exe -NoProfile -NonInteractive -ExecutionPolicy Bypass -File "%~dp0post-task.ps1"
+```
+
+## Windows notes
+
+- Supported as `post_task_script`: `.exe`, `.bat`, `.cmd`.
+- **Not supported yet:** `.ps1` as the configured path (use a `.cmd` wrapper; see above).
+- `.sh` files require a Unix shell on the PATH unless you point `post_task_script` at the interpreter.
+- Use backslashes or forward slashes in YAML paths; both work in Go on Windows.
+
+## See also
+
+- [Configuration](../README.md#configuration) — generating and loading `config.yaml`
+- [config.example.yaml](../internal/pkg/config/config.example.yaml) — all runner options
+- Bind-workdir idle cleanup (`runner.workdir_cleanup_age`) — separate from this hook; runs only when the runner is idle

go.mod

@@ -11,7 +11,7 @@ require (
 	github.com/containerd/errdefs v1.0.0
 	github.com/creack/pty v1.1.24
 	github.com/distribution/reference v0.6.0
-	github.com/docker/cli v29.5.3+incompatible
+	github.com/docker/cli v29.6.0+incompatible
 	github.com/docker/go-connections v0.7.0
 	github.com/go-git/go-billy/v5 v5.9.0
 	github.com/go-git/go-git/v5 v5.19.1
@@ -22,8 +22,8 @@ require (
 	github.com/kballard/go-shellquote v0.0.0-20180428030007-95032a82bc51
 	github.com/mattn/go-isatty v0.0.22
 	github.com/moby/go-archive v0.2.0
-	github.com/moby/moby/api v1.54.2
+	github.com/moby/moby/api v1.55.0
-	github.com/moby/moby/client v0.4.1
+	github.com/moby/moby/client v0.5.0
 	github.com/moby/patternmatcher v0.6.1
 	github.com/opencontainers/image-spec v1.1.1
 	github.com/opencontainers/selinux v1.15.1
@@ -35,7 +35,7 @@ require (
 	github.com/spf13/pflag v1.0.10
 	github.com/stretchr/testify v1.11.1
 	github.com/timshannon/bolthold v0.0.0-20240314194003-30aac6950928
-	go.etcd.io/bbolt v1.4.3
+	go.etcd.io/bbolt v1.5.0
 	go.yaml.in/yaml/v4 v4.0.0-rc.3
 	golang.org/x/sys v0.46.0
 	golang.org/x/term v0.44.0

go.sum

@@ -49,6 +49,8 @@ github.com/distribution/reference v0.6.0 h1:0IXCQ5g4/QMHHkarYzh5l+u8T3t73zM5Qvfr
 github.com/distribution/reference v0.6.0/go.mod h1:BbU0aIcezP1/5jX/8MP0YiH4SdvB5Y4f/wlDRiLyi3E=
 github.com/docker/cli v29.5.3+incompatible h1:nbEFfz774vBwQ5KRYv7c/AghjReqnGISvrRhzjV0evs=
 github.com/docker/cli v29.5.3+incompatible/go.mod h1:JLrzqnKDaYBop7H2jaqPtU4hHvMKP+vjCwu2uszcLI8=
+github.com/docker/cli v29.6.0+incompatible h1:nw9himxMMZ7eIeherJNlKQq+acnlzGgHd+4uf10QRSc=
+github.com/docker/cli v29.6.0+incompatible/go.mod h1:JLrzqnKDaYBop7H2jaqPtU4hHvMKP+vjCwu2uszcLI8=
 github.com/docker/docker-credential-helpers v0.9.6 h1:cT2PbRPSlnMmNTfT2TDMXRyQ1KMWHG7xoTLBcn1ZNv0=
 github.com/docker/docker-credential-helpers v0.9.6/go.mod h1:v1S+hepowrQXITkEfw6o4+BMbGot02wiKpzWhGUZK6c=
 github.com/docker/go-connections v0.7.0 h1:6SsRfJddP22WMrCkj19x9WKjEDTB+ahsdiGYf0mN39c=
@@ -129,8 +131,12 @@ github.com/moby/go-archive v0.2.0 h1:zg5QDUM2mi0JIM9fdQZWC7U8+2ZfixfTYoHL7rWUcP8
 github.com/moby/go-archive v0.2.0/go.mod h1:mNeivT14o8xU+5q1YnNrkQVpK+dnNe/K6fHqnTg4qPU=
 github.com/moby/moby/api v1.54.2 h1:wiat9QAhnDQjA7wk1kh/TqHz2I1uUA7M7t9SAl/JNXg=
 github.com/moby/moby/api v1.54.2/go.mod h1:+RQ6wluLwtYaTd1WnPLykIDPekkuyD/ROWQClE83pzs=
+github.com/moby/moby/api v1.55.0 h1:2/sexvQyqIWS8pRSCFddBfpW2qE7vR7FCL+vN8pxwMc=
+github.com/moby/moby/api v1.55.0/go.mod h1:+RQ6wluLwtYaTd1WnPLykIDPekkuyD/ROWQClE83pzs=
 github.com/moby/moby/client v0.4.1 h1:DMQgisVoMkmMs7fp3ROSdiBnoAu8+vo3GggFl06M/wY=
 github.com/moby/moby/client v0.4.1/go.mod h1:z52C9O2POPOsnxZAy//WtKcQ32P+jT/NGeXu/7nfjGQ=
+github.com/moby/moby/client v0.5.0 h1:5XhyPk2fuOWf6RlSFa3MkIIgDZkF25xToXW8Q/BH7cc=
+github.com/moby/moby/client v0.5.0/go.mod h1:rcVpF8ncl9vo5gaIBdol6CnbEtSj1uxMvEV/UrykF/s=
 github.com/moby/patternmatcher v0.6.1 h1:qlhtafmr6kgMIJjKJMDmMWq7WLkKIo23hsrpR3x084U=
 github.com/moby/patternmatcher v0.6.1/go.mod h1:hDPoyOpDY7OrrMDLaYoY3hf52gNCR/YOUYxkhApJIxc=
 github.com/moby/sys/sequential v0.6.0 h1:qrx7XFUd/5DxtqcoH1h438hF5TmOvzC/lspjy7zgvCU=
@@ -209,6 +215,8 @@ github.com/xeipuuv/gojsonschema v1.2.0/go.mod h1:anYRn/JVcOK2ZgGU+IjEV4nwlhoK5sQ
 go.etcd.io/bbolt v1.3.8/go.mod h1:N9Mkw9X8x5fupy0IKsmuqVtoGDyxsaDlbk4Rd05IAQw=
 go.etcd.io/bbolt v1.4.3 h1:dEadXpI6G79deX5prL3QRNP6JB8UxVkqo4UPnHaNXJo=
 go.etcd.io/bbolt v1.4.3/go.mod h1:tKQlpPaYCVFctUIgFKFnAlvbmB3tpy1vkTnDWohtc0E=
+go.etcd.io/bbolt v1.5.0 h1:S7GAl7Fxv12yohbwFfIbQCGDWbQbtDGPET4P/bD4lxU=
+go.etcd.io/bbolt v1.5.0/go.mod h1:mkltfYE5aUHQxUct9N9V+Kp7aSjFqjgrhcXIS70Lrdk=
 go.etcd.io/gofail v0.1.0/go.mod h1:VZBCXYGZhHAinaBiiqYvuDynvahNsAyLFwB3kEHKz1M=
 go.opentelemetry.io/auto/sdk v1.2.1 h1:jXsnJ4Lmnqd11kwkBV2LgLoFMZKizbCi5fNZ/ipaZ64=
 go.opentelemetry.io/auto/sdk v1.2.1/go.mod h1:KRTj+aOaElaLi+wW1kO/DZRXwkF4C5xPbEe3ZiIhN7Y=
@@ -253,8 +261,6 @@ golang.org/x/sys v0.15.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
 golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.43.0 h1:S4RLU2sB31O/NCl+zFN9Aru9A/Cq2aqKpTZJ6B+DwT4=
-golang.org/x/term v0.43.0/go.mod h1:lrhlHNdQJHO+1qVYiHfFKVuVioJIheAc3fBSMFYEIsk=
 golang.org/x/term v0.44.0 h1:0rLvDRCtNj0gZkyIXhCyOb2OAzEhLVqc4B+hrsBhrmc=
 golang.org/x/term v0.44.0/go.mod h1:7ze4MdzUzLXpSAoFP1H0bOI9aXDqveSvatT5vKcFh2Y=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=

internal/app/run/post_task_script.go

@@ -0,0 +1,132 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package run
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"os/exec"
+	"strconv"
+	"strings"
+	"time"
+
+	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/internal/pkg/config"
+	"gitea.com/gitea/runner/internal/pkg/metrics"
+	"gitea.com/gitea/runner/internal/pkg/process"
+	"gitea.com/gitea/runner/internal/pkg/report"
+
+	runnerv1 "gitea.dev/actions-proto-go/runner/v1"
+	log "github.com/sirupsen/logrus"
+)
+
+func (r *Runner) runPostTaskScript(ctx context.Context, reporter *report.Reporter, task *runnerv1.Task, workdir string) {
+	script := r.cfg.Runner.PostTaskScript
+	if script == "" {
+		return
+	}
+
+	timeout := r.cfg.Runner.PostTaskScriptTimeout
+	if timeout <= 0 {
+		timeout = config.DefaultPostTaskScriptTimeout
+	}
+
+	scriptCtx, cancel := postTaskScriptContext(ctx, timeout)
+	defer cancel()
+
+	env := r.postTaskScriptEnv(reporter, task, workdir)
+	log.Infof("running post-task script %q for task %d", script, task.Id)
+
+	cmd := exec.CommandContext(scriptCtx, script)
+	cmd.Env = envListFromMap(env)
+	cmd.SysProcAttr = process.SysProcAttr(script, false)
+
+	stdout := postTaskScriptLogWriter("stdout")
+	stderr := postTaskScriptLogWriter("stderr")
+	cmd.Stdout = stdout
+	cmd.Stderr = stderr
+
+	// Kill the script's whole process tree on cancellation and bound the post-exit
+	// I/O wait, so a backgrounded child inheriting cmd's stdout/stderr pipe can
+	// never hang cmd.Wait() and the runner. See process.TreeKill.
+	treeKill := process.NewTreeKill(cmd)
+
+	if err := cmd.Start(); err != nil {
+		log.Warnf("post-task script %q for task %d: %v", script, task.Id, err)
+		return
+	}
+	if k, kerr := treeKill.Capture(cmd.Process); kerr != nil {
+		log.Warnf("post-task script %q for task %d: process tree kill setup failed, falling back to single-process kill: %v", script, task.Id, kerr)
+	} else {
+		defer k.Close()
+	}
+
+	err := cmd.Wait()
+	// Flush any trailing, not-yet-newline-terminated output now that the I/O
+	// copiers have finished (cmd.Wait, bounded by WaitDelay above, guarantees it).
+	common.FlushWriter(stdout)
+	common.FlushWriter(stderr)
+	if err != nil {
+		if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+			log.Warnf("post-task script %q for task %d: %v", script, task.Id, err)
+			return
+		}
+		var exitErr *exec.ExitError
+		if errors.As(err, &exitErr) {
+			log.Warnf("post-task script %q for task %d exited with code %d", script, task.Id, exitErr.ExitCode())
+			return
+		}
+		log.Warnf("post-task script %q for task %d: %v", script, task.Id, err)
+	}
+}
+
+func postTaskScriptContext(ctx context.Context, timeout time.Duration) (context.Context, context.CancelFunc) {
+	// Detach from the task context's deadline and cancellation: the task has
+	// already finished by the time the post-task script runs, so the script must
+	// get its full configured timeout. Inheriting the task deadline would silently
+	// truncate that budget when the job completed close to its own timeout (and an
+	// already-cancelled task context would skip the script entirely).
+	// context.WithoutCancel keeps the context values while dropping the deadline.
+	return context.WithTimeout(context.WithoutCancel(ctx), timeout)
+}
+
+func (r *Runner) postTaskScriptEnv(reporter *report.Reporter, task *runnerv1.Task, workdir string) map[string]string {
+	env := r.cloneEnvs()
+	env["GITEA_TASK_ID"] = strconv.FormatInt(task.Id, 10)
+	env["GITEA_WORKSPACE"] = workdir
+	// GITEA_JOB_RESULT shares the runner's canonical result vocabulary
+	// (success/failure/cancelled/skipped/unknown), the same strings the reporter
+	// parses and the metrics labels use.
+	env["GITEA_JOB_RESULT"] = metrics.ResultToStatusLabel(reporter.Result())
+	if v := task.Context.Fields["run_id"].GetStringValue(); v != "" {
+		env["GITEA_RUN_ID"] = v
+	}
+	if v := task.Context.Fields["repository"].GetStringValue(); v != "" {
+		env["GITEA_REPOSITORY"] = v
+	}
+	return env
+}
+
+func envListFromMap(env map[string]string) []string {
+	envList := make([]string, 0, len(env))
+	for k, v := range env {
+		envList = append(envList, fmt.Sprintf("%s=%s", k, v))
+	}
+	return envList
+}
+
+// postTaskScriptLogWriter returns an io.Writer that logs the script's output one
+// line at a time, tagged with the stream name. It is passed as cmd.Stdout/Stderr
+// (rather than a StdoutPipe) so that cmd.WaitDelay governs the copying goroutine:
+// a backgrounded process holding the pipe open can never block cmd.Wait()
+// indefinitely. Flush any trailing partial line with common.FlushWriter after
+// cmd.Wait() returns.
+func postTaskScriptLogWriter(stream string) io.Writer {
+	return common.NewLineWriter(func(line string) bool {
+		log.Infof("post-task script %s: %s", stream, strings.TrimRight(line, "\r\n"))
+		return true
+	})
+}

internal/app/run/post_task_script_test.go

@@ -0,0 +1,157 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package run
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"gitea.com/gitea/runner/internal/pkg/config"
+	"gitea.com/gitea/runner/internal/pkg/metrics"
+	"gitea.com/gitea/runner/internal/pkg/report"
+
+	runnerv1 "gitea.dev/actions-proto-go/runner/v1"
+	log "github.com/sirupsen/logrus"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"google.golang.org/protobuf/types/known/structpb"
+)
+
+func TestRunPostTaskScriptSkippedWhenEmpty(t *testing.T) {
+	r := &Runner{
+		cfg: &config.Config{},
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	taskCtx, err := structpb.NewStruct(map[string]any{})
+	require.NoError(t, err)
+	task := &runnerv1.Task{Id: 1, Context: taskCtx}
+	reporter := report.NewReporter(ctx, cancel, nil, task, r.cfg)
+
+	require.NotPanics(t, func() {
+		r.runPostTaskScript(ctx, reporter, task, "/workspace/owner/repo")
+	})
+}
+
+func TestRunPostTaskScriptNonZeroExitDoesNotPanic(t *testing.T) {
+	dir := t.TempDir()
+	scriptPath := filepath.Join(dir, "fail.sh")
+	require.NoError(t, os.WriteFile(scriptPath, []byte("#!/bin/sh\nexit 2\n"), 0o700))
+
+	cfg, err := config.LoadDefault("")
+	require.NoError(t, err)
+	cfg.Runner.PostTaskScript = scriptPath
+
+	r := &Runner{cfg: cfg}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	taskCtx, err := structpb.NewStruct(map[string]any{})
+	require.NoError(t, err)
+	task := &runnerv1.Task{Id: 1, Context: taskCtx}
+	reporter := report.NewReporter(ctx, cancel, nil, task, cfg)
+
+	require.NotPanics(t, func() {
+		r.runPostTaskScript(ctx, reporter, task, "/workspace/owner/repo")
+	})
+}
+
+func TestPostTaskScriptContextUsesFullTimeout(t *testing.T) {
+	const timeout = 5 * time.Minute
+
+	// A task context that finished close to its own deadline must not truncate the
+	// script's budget: the script should still get its full configured timeout.
+	near, cancelNear := context.WithTimeout(context.Background(), time.Second)
+	defer cancelNear()
+	scriptCtx, cancel := postTaskScriptContext(near, timeout)
+	defer cancel()
+	deadline, ok := scriptCtx.Deadline()
+	require.True(t, ok)
+	assert.Greater(t, time.Until(deadline), time.Minute, "script timeout truncated to task deadline")
+
+	// An already-cancelled task context must not cancel the script either.
+	cancelledCtx, cancelIt := context.WithCancel(context.Background())
+	cancelIt()
+	scriptCtx2, cancel2 := postTaskScriptContext(cancelledCtx, timeout)
+	defer cancel2()
+	assert.NoError(t, scriptCtx2.Err(), "script context inherited the cancelled task context")
+}
+
+func TestPostTaskScriptEnv(t *testing.T) {
+	cfg, err := config.LoadDefault("")
+	require.NoError(t, err)
+
+	r := &Runner{
+		cfg:  cfg,
+		envs: map[string]string{"BASE": "1"},
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	taskCtx, err := structpb.NewStruct(map[string]any{
+		"run_id":     "99",
+		"repository": "acme/widget",
+	})
+	require.NoError(t, err)
+	task := &runnerv1.Task{Id: 3, Context: taskCtx}
+	reporter := report.NewReporter(ctx, cancel, nil, task, cfg)
+	setReporterJobResult(t, reporter, runnerv1.Result_RESULT_FAILURE)
+
+	env := r.postTaskScriptEnv(reporter, task, "/tmp/workspace")
+	assert.Equal(t, "1", env["BASE"])
+	assert.Equal(t, "3", env["GITEA_TASK_ID"])
+	assert.Equal(t, "99", env["GITEA_RUN_ID"])
+	assert.Equal(t, "acme/widget", env["GITEA_REPOSITORY"])
+	assert.Equal(t, "/tmp/workspace", env["GITEA_WORKSPACE"])
+	assert.Equal(t, "failure", env["GITEA_JOB_RESULT"])
+}
+
+func TestRunPostTaskScriptIntegration(t *testing.T) {
+	dir := t.TempDir()
+	outFile := filepath.Join(dir, "out.txt")
+	scriptPath := filepath.Join(dir, "post-task.sh")
+	script := "#!/bin/sh\nprintf '%s %s %s' \"$GITEA_TASK_ID\" \"$GITEA_JOB_RESULT\" \"$CUSTOM\" > \"" + outFile + "\"\n"
+	require.NoError(t, os.WriteFile(scriptPath, []byte(script), 0o700))
+
+	cfg, err := config.LoadDefault("")
+	require.NoError(t, err)
+	cfg.Runner.PostTaskScript = scriptPath
+
+	r := &Runner{
+		cfg:  cfg,
+		envs: map[string]string{"CUSTOM": "runner-env"},
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	taskCtx, err := structpb.NewStruct(map[string]any{})
+	require.NoError(t, err)
+	task := &runnerv1.Task{Id: 11, Context: taskCtx}
+	reporter := report.NewReporter(ctx, cancel, nil, task, cfg)
+	setReporterJobResult(t, reporter, runnerv1.Result_RESULT_SUCCESS)
+
+	r.runPostTaskScript(ctx, reporter, task, "/workspace/acme/repo")
+
+	content, err := os.ReadFile(outFile)
+	require.NoError(t, err)
+	assert.Equal(t, "11 success runner-env", string(content))
+}
+
+func setReporterJobResult(t *testing.T, reporter *report.Reporter, result runnerv1.Result) {
+	t.Helper()
+	require.NoError(t, reporter.Fire(&log.Entry{
+		Time:    time.Now(),
+		Message: "job finished",
+		Data: log.Fields{
+			"stage":     "Post",
+			"jobResult": metrics.ResultToStatusLabel(result),
+		},
+	}))
+}

internal/app/run/runner.go

@@ -22,6 +22,7 @@ import (
 
 	"gitea.com/gitea/runner/act/artifactcache"
 	"gitea.com/gitea/runner/act/common"
+	"gitea.com/gitea/runner/act/container"
 	"gitea.com/gitea/runner/act/model"
 	"gitea.com/gitea/runner/act/runner"
 	"gitea.com/gitea/runner/internal/pkg/client"
@@ -33,7 +34,7 @@ import (
 
 	"connectrpc.com/connect"
 	runnerv1 "gitea.dev/actions-proto-go/runner/v1"
-	"github.com/moby/moby/api/types/container"
+	docker_container "github.com/moby/moby/api/types/container"
 	log "github.com/sirupsen/logrus"
 )
 
@@ -395,6 +396,12 @@ func (r *Runner) run(ctx context.Context, task *runnerv1.Task, reporter *report.
 		maxLifetime = time.Until(deadline)
 	}
 
+	// shallow clones the requested ref at depth 1, otherwise 0 means a full clone
+	actionCloneDepth := 1
+	if r.cfg.Runner.ActionShallowClone != nil && !*r.cfg.Runner.ActionShallowClone {
+		actionCloneDepth = 0
+	}
+
 	workdirParent := strings.TrimLeft(r.cfg.Container.WorkdirParent, "/")
 	if r.cfg.Container.BindWorkdir {
 		// Append the task ID to isolate concurrent jobs from the same repo.
@@ -417,6 +424,7 @@ func (r *Runner) run(ctx context.Context, task *runnerv1.Task, reporter *report.
 		ActionCacheDir:    filepath.FromSlash(r.cfg.Host.WorkdirParent),
 		AllocatePTY:       r.cfg.Runner.AllocatePTY,
 		ActionOfflineMode: r.cfg.Cache.OfflineMode,
+		ActionCloneDepth:  actionCloneDepth,
 
 		ReuseContainers:      false,
 		ForcePull:            r.cfg.Container.ForcePull,
@@ -433,7 +441,11 @@ func (r *Runner) run(ctx context.Context, task *runnerv1.Task, reporter *report.
 		ContainerNamePrefix:  fmt.Sprintf("GITEA-ACTIONS-TASK-%d", task.Id),
 		ContainerMaxLifetime: maxLifetime,
 		CleanWorkdir:         true,
-		ContainerNetworkMode:  container.NetworkMode(r.cfg.Container.Network),
+		ContainerNetworkMode: docker_container.NetworkMode(r.cfg.Container.Network),
+		ContainerNetworkCreateOptions: container.NewDockerNetworkCreateExecutorInput{
+			EnableIPv4: r.cfg.Container.NetworkCreateOptions.EnableIPv4,
+			EnableIPv6: r.cfg.Container.NetworkCreateOptions.EnableIPv6,
+		},
 		ContainerOptions:      r.cfg.Container.Options,
 		ContainerDaemonSocket: r.cfg.Container.DockerHost,
 		Privileged:            r.cfg.Container.Privileged,
@@ -470,6 +482,9 @@ func (r *Runner) run(ctx context.Context, task *runnerv1.Task, reporter *report.
 		}
 	}
 
+	reporter.StopHeartbeats()
+	r.runPostTaskScript(ctx, reporter, task, workdir)
+
 	return execErr
 }
 

internal/pkg/config/config.example.yaml

@@ -69,6 +69,9 @@ runner:
   # and github_mirror is not empty. In this case,
   # it replaces https://github.com with the value here, which is useful for some special network environments.
   github_mirror: ''
+  # When true (the default), fetch only the requested ref of an action repository (e.g. actions/checkout@v4) at depth 1 instead of cloning every branch's full history. 
+  # Set to false to clone the full history.
+  action_shallow_clone: true
   # The labels of a runner are used to determine which jobs the runner can run, and how to run them.
   # Like: "macos-arm64:host" or "ubuntu-latest:docker://docker.gitea.com/runner-images:ubuntu-latest"
   # Find more images provided by Gitea at https://gitea.com/gitea/runner-images .
@@ -83,31 +86,47 @@ runner:
   # terminal; tools like `docker build` emit redrawing progress frames into the captured log
   # when a TTY is present.
   allocate_pty: false
+  # Optional executable on the host, run once after each task's built-in cleanup
+  # (post-steps, container teardown, bind-workdir removal). Additive only.
+  #
+  # IMPORTANT: While this script runs the runner stops task heartbeats and stays
+  # offline from Gitea's perspective until the script exits. A script that never
+  # returns blocks new work until post_task_script_timeout kills it (default 5m).
+  # Keep scripts short; set post_task_script_timeout to a safe upper bound.
+  #
+  # Output -> runner process log (not the job log). Non-zero exit -> warning only.
+  # Windows: use .exe, .bat, or .cmd. PowerShell (.ps1) is not supported yet as
+  # the configured path; wrap PowerShell commands in a .cmd file instead.
+  # Full guide: docs/post-task-script.md
+  post_task_script: ''
+  # Hard limit on post_task_script runtime. Default if omitted: 5m.
+  post_task_script_timeout: 5m
 
 cache:
-  # Enable cache server to use actions/cache.
+  # Enable the built-in cache server (used by actions/cache and similar actions).
   enabled: true
-  # The directory to store the cache data.
+  # Directory where cache blobs are stored on disk. Default: $HOME/.cache/actcache
-  # If it's empty, the cache data will be stored in $HOME/.cache/actcache.
+  # Ignored when external_server is set.
   dir: ""
-  # The host of the cache server.
+  # Outbound IP or hostname that job containers use to reach this runner's cache server.
-  # It's not for the address to listen, but the address to connect from job containers.
+  # Leave empty to detect automatically. 0.0.0.0 is not valid here.
-  # So 0.0.0.0 is a bad choice, leave it empty to detect automatically.
+  # Ignored when external_server is set.
   host: ""
-  # The port of the cache server.
+  # Port for the built-in cache server. 0 picks a random free port.
-  # 0 means to use a random available port.
+  # Ignored when external_server is set.
   port: 0
-  # The external cache server URL. Valid only when enable is true.
+  # URL of a shared `gitea-runner cache-server` to use instead of starting a local one.
-  # If it's specified, runner will use this URL as the ACTIONS_CACHE_URL rather than start a server by itself.
+  # Set on every runner that should share a cache pool. Must end with "/".
-  # The URL should generally end with "/".
+  # Example: "http://cache-host:8088/"
-  # Requires external_secret below to be set to the same value on both this runner and the cache-server.
+  # Requires external_secret (below) to match the value on the cache-server.
   external_server: ""
-  # Shared secret between this runner and the external `gitea-runner cache-server`. Required when external_server
+  # Shared secret between this runner and the external cache-server.
-  # (or `gitea-runner cache-server`) is in use: the runner pre-registers each job's ACTIONS_RUNTIME_TOKEN with the
+  # Required when external_server is set. Must be identical on every runner and the cache-server.
-  # cache-server, and the cache-server enforces bearer auth + per-repo cache isolation.
+  # Generate with: openssl rand -hex 32
   external_secret: ""
-  # When true, reuse a cached action instead of fetching from the remote on every job. Note: a moved tag
+  # When true, reuse a cached action instead of fetching from the remote on every job.
-  # (e.g. a re-tagged "v6") or an updated branch stays at the cached commit until its cache entry is removed.
+  # A moved tag (e.g. a re-tagged "v6") or an updated branch stays at the cached commit
+  # until its cache entry expires or is manually removed.
   offline_mode: false
 
 container:
@@ -116,6 +135,13 @@ container:
   # If it's empty, runner will create a network automatically.
   # Deprecated: `network_mode` is still accepted for old configs; use `network` instead.
   network: ""
+  # network_create_options only apply when `network` is left empty and the runner
+  # auto-creates a per-job network that does not already exist. They have no effect
+  # when a custom `network` name is set, because that network is used as-is and never
+  # created by the runner. Omit the entire block to use Docker's defaults.
+  network_create_options:
+    enable_ipv4: true  # Omit to use Docker's default (IPv4 enabled). Set false to disable IPv4.
+    enable_ipv6: false # Omit to use Docker's default (IPv6 disabled). Enabling it requires dockerd started with --ipv6.
   # Whether to use privileged mode or not when launching task containers (privileged mode is required for Docker-in-Docker).
   privileged: false
   # Any other options to be used when the container is started (e.g., --add-host=my.gitea.url:host-gateway).

internal/pkg/config/config.go

@@ -16,6 +16,12 @@ import (
 	"go.yaml.in/yaml/v4"
 )
 
+// DefaultPostTaskScriptTimeout is the fallback cap on how long the post-task
+// script may run when post_task_script is set without an explicit timeout. It is
+// applied both at config load (for a configured script) and at the point of use
+// (so a programmatically built config still gets a sane bound).
+const DefaultPostTaskScriptTimeout = 5 * time.Minute
+
 // Log represents the configuration for logging.
 type Log struct {
 	Level string `yaml:"level"` // Level indicates the logging level.
@@ -42,7 +48,10 @@ type Runner struct {
 	ReportCloseTimeout    time.Duration     `yaml:"report_close_timeout"`     // ReportCloseTimeout caps each RPC attempt when flushing the final logs and task state at job completion, on a detached context so a server cancel can't block the acknowledgement.
 	Labels                []string          `yaml:"labels"`                   // Labels specify the labels of the runner. Labels are declared on each startup
 	GithubMirror          string            `yaml:"github_mirror"`            // GithubMirror defines what mirrors should be used when using github
+	ActionShallowClone    *bool             `yaml:"action_shallow_clone"`     // ActionShallowClone fetches only the requested ref of an action repository at depth 1 instead of cloning every branch's full history. It is a pointer to distinguish between false and not set; if not set, it defaults to true.
 	AllocatePTY           bool              `yaml:"allocate_pty"`             // AllocatePTY allocates a pseudo-TTY for each step's process. Default is false, matching GitHub's actions/runner. Enable only for jobs that need an interactive terminal; tools like docker build emit redrawing progress frames into the captured log when a TTY is present. Applies to both host and docker backends.
+	PostTaskScript        string            `yaml:"post_task_script"`         // PostTaskScript is the path to an executable script run on the host after each task's cleanup completes. Empty disables the hook. On Windows use .exe/.bat/.cmd; PowerShell (.ps1) is not supported yet as the configured path.
+	PostTaskScriptTimeout time.Duration     `yaml:"post_task_script_timeout"` // PostTaskScriptTimeout caps how long the post-task script may run. Default is 5m when post_task_script is set.
 }
 
 // Cache represents the configuration for caching.
@@ -59,6 +68,7 @@ type Cache struct {
 // Container represents the configuration for the container.
 type Container struct {
 	Network              string                        `yaml:"network"`                // Network specifies the network for the container.
+	NetworkCreateOptions ContainerNetworkCreateOptions `yaml:"network_create_options"` // Add options when the network need to be created by the runner
 	NetworkMode          string                        `yaml:"network_mode"`           // Deprecated: use Network instead. Could be removed after Gitea 1.20
 	Privileged           bool                          `yaml:"privileged"`             // Privileged indicates whether the container runs in privileged mode.
 	Options              string                        `yaml:"options"`                // Options specifies additional options for the container.
@@ -72,6 +82,11 @@ type Container struct {
 	BindWorkdir          bool                          `yaml:"bind_workdir"`           // BindWorkdir binds the workspace to the host filesystem instead of using Docker volumes. Required for DinD when jobs use docker compose with bind mounts.
 }
 
+type ContainerNetworkCreateOptions struct {
+	EnableIPv4 *bool `yaml:"enable_ipv4"` // Enable or disable IPv4 for the network (true for docker by default)
+	EnableIPv6 *bool `yaml:"enable_ipv6"` // Enable or disable IPv6 for the network (false for docker by default)
+}
+
 // Host represents the configuration for the host.
 type Host struct {
 	WorkdirParent string `yaml:"workdir_parent"` // WorkdirParent specifies the parent directory for the host's working directory.
@@ -137,6 +152,10 @@ func LoadDefault(file string) (*Config, error) {
 	if cfg.Runner.Timeout <= 0 {
 		cfg.Runner.Timeout = 3 * time.Hour
 	}
+	if cfg.Runner.ActionShallowClone == nil {
+		b := true
+		cfg.Runner.ActionShallowClone = &b
+	}
 	if cfg.Cache.Enabled == nil {
 		b := true
 		cfg.Cache.Enabled = &b
@@ -187,6 +206,9 @@ func LoadDefault(file string) (*Config, error) {
 	if cfg.Runner.ReportCloseTimeout <= 0 {
 		cfg.Runner.ReportCloseTimeout = 10 * time.Second
 	}
+	if cfg.Runner.PostTaskScript != "" && cfg.Runner.PostTaskScriptTimeout <= 0 {
+		cfg.Runner.PostTaskScriptTimeout = DefaultPostTaskScriptTimeout
+	}
 	if cfg.Metrics.Addr == "" {
 		cfg.Metrics.Addr = "127.0.0.1:9101"
 	}

internal/pkg/config/config_test.go

@@ -107,6 +107,34 @@ runner:
 // TestLoadDefault_MalformedYAMLReturnsParseError pins the error surfaced for
 // invalid YAML to the canonical "parse config file" message rather than the
 // "for defaults metadata" variant — i.e. the main yaml.Unmarshal runs first.
+func TestLoadDefault_LoadsPostTaskScript(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	require.NoError(t, os.WriteFile(path, []byte(`
+runner:
+  post_task_script: /usr/local/bin/post-task.sh
+  post_task_script_timeout: 2m
+`), 0o600))
+
+	cfg, err := LoadDefault(path)
+	require.NoError(t, err)
+	assert.Equal(t, "/usr/local/bin/post-task.sh", cfg.Runner.PostTaskScript)
+	assert.Equal(t, 2*time.Minute, cfg.Runner.PostTaskScriptTimeout)
+}
+
+func TestLoadDefault_DefaultsPostTaskScriptTimeout(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	require.NoError(t, os.WriteFile(path, []byte(`
+runner:
+  post_task_script: /usr/local/bin/post-task.sh
+`), 0o600))
+
+	cfg, err := LoadDefault(path)
+	require.NoError(t, err)
+	assert.Equal(t, 5*time.Minute, cfg.Runner.PostTaskScriptTimeout)
+}
+
 func TestLoadDefault_MalformedYAMLReturnsParseError(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "config.yaml")
@@ -117,3 +145,50 @@ func TestLoadDefault_MalformedYAMLReturnsParseError(t *testing.T) {
 	assert.Contains(t, err.Error(), "parse config file")
 	assert.NotContains(t, err.Error(), "defaults metadata")
 }
+
+func TestContainerNetworkCreateOptions(t *testing.T) {
+	// Verify that the enable_ipv4/enable_ipv6 YAML keys unmarshal into the *bool fields,
+	// distinguishing an explicit true/false from an omitted key (nil). A nil here is
+	// forwarded as-is to Docker, which applies its own default.
+	loadOptions := func(t *testing.T, yaml string) ContainerNetworkCreateOptions {
+		t.Helper()
+		dir := t.TempDir()
+		path := filepath.Join(dir, "config.yaml")
+		require.NoError(t, os.WriteFile(path, []byte(yaml), 0o600))
+
+		cfg, err := LoadDefault(path)
+		require.NoError(t, err)
+		return cfg.Container.NetworkCreateOptions
+	}
+
+	t.Run("enable_ipv6 true unmarshals to non-nil true", func(t *testing.T) {
+		opts := loadOptions(t, "container:\n  network_create_options:\n    enable_ipv6: true\n")
+		require.NotNil(t, opts.EnableIPv6)
+		assert.True(t, *opts.EnableIPv6)
+	})
+
+	t.Run("enable_ipv6 false unmarshals to non-nil false", func(t *testing.T) {
+		opts := loadOptions(t, "container:\n  network_create_options:\n    enable_ipv6: false\n")
+		require.NotNil(t, opts.EnableIPv6)
+		assert.False(t, *opts.EnableIPv6)
+	})
+
+	t.Run("enable_ipv4 false unmarshals to non-nil false", func(t *testing.T) {
+		opts := loadOptions(t, "container:\n  network_create_options:\n    enable_ipv4: false\n")
+		require.NotNil(t, opts.EnableIPv4)
+		assert.False(t, *opts.EnableIPv4)
+	})
+
+	t.Run("omitted keys stay nil", func(t *testing.T) {
+		opts := loadOptions(t, "container:\n  network_create_options:\n    enable_ipv4: true\n")
+		require.NotNil(t, opts.EnableIPv4)
+		assert.True(t, *opts.EnableIPv4)
+		assert.Nil(t, opts.EnableIPv6, "an omitted enable_ipv6 must remain nil so Docker's default applies")
+	})
+
+	t.Run("omitted block leaves both nil", func(t *testing.T) {
+		opts := loadOptions(t, "container:\n  network: \"\"\n")
+		assert.Nil(t, opts.EnableIPv4)
+		assert.Nil(t, opts.EnableIPv6)
+	})
+}

internal/pkg/process/killer_plan9.go

@@ -0,0 +1,29 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build plan9
+
+package process
+
+import "os"
+
+// Killer falls back to single-process termination on platforms without a
+// process-group / Job Object tree-kill. The Job Object (Windows) and process
+// group (Unix) based tree-kills live in killer_windows.go / killer_unix.go;
+// here we just kill the direct child, matching the previous default behaviour.
+type Killer struct {
+	p *os.Process
+}
+
+func NewKiller(p *os.Process) (*Killer, error) {
+	return &Killer{p: p}, nil
+}
+
+func (k *Killer) Kill() error {
+	if k == nil || k.p == nil {
+		return nil
+	}
+	return k.p.Kill()
+}
+
+func (k *Killer) Close() error { return nil }

internal/pkg/process/killer_unix.go

@@ -0,0 +1,56 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build !windows && !plan9
+
+package process
+
+import (
+	"errors"
+	"os"
+	"syscall"
+)
+
+// Killer terminates a started process together with its whole process group,
+// which is the Unix counterpart of the Windows Job Object tree-kill.
+//
+// Background: a process (a step or a post-task script) often launches a process
+// tree (a shell that starts a child which in turn spawns further background
+// processes). The default exec.CommandContext cancellation only kills the
+// direct child, so cancelling left the rest of the tree running. Because those
+// orphans inherited the parent's stdout/stderr pipe, cmd.Wait() also blocked
+// forever and the runner hung.
+//
+// Processes are started with Setpgid (or Setsid for the PTY path, see
+// SysProcAttr), which makes the process the leader of a new process group whose
+// ID equals its PID. Signalling the negative PID delivers to every process
+// still in that group, so we can tear down the whole tree atomically on
+// cancellation, which also closes the inherited pipe handles so cmd.Wait() can
+// return.
+type Killer struct {
+	pgid int
+}
+
+// NewKiller captures the process group of p (an already-started process).
+// Because the process is launched with Setpgid/Setsid, p is a group leader and
+// its PGID equals its PID; children spawned afterwards stay in the same group
+// unless they explicitly create their own.
+func NewKiller(p *os.Process) (*Killer, error) {
+	return &Killer{pgid: p.Pid}, nil
+}
+
+// Kill sends SIGKILL to the entire process group (the process and every
+// descendant that stayed in the group). A missing group (ESRCH) means the
+// processes already exited and is not treated as an error.
+func (k *Killer) Kill() error {
+	if k == nil || k.pgid <= 0 {
+		return nil
+	}
+	if err := syscall.Kill(-k.pgid, syscall.SIGKILL); err != nil && !errors.Is(err, syscall.ESRCH) {
+		return err
+	}
+	return nil
+}
+
+// Close is a no-op on Unix; there is no job handle to release.
+func (k *Killer) Close() error { return nil }

internal/pkg/process/killer_unix_test.go

@@ -0,0 +1,101 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build !windows && !plan9
+
+package process
+
+import (
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"syscall"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+// processAlive reports whether pid refers to a still-running process. Signal 0
+// performs error checking without delivering a signal: a nil error (or EPERM)
+// means the process exists, ESRCH means it is gone.
+//
+// On Linux, zombie processes (state Z in /proc/<pid>/stat) appear alive to
+// kill(0) but have already terminated — their corpse lingers until the parent
+// calls wait(). In a Docker container the child may be reparented to a PID 1
+// that does not reap promptly, so we treat zombies as not alive.
+func processAlive(pid int) bool {
+	err := syscall.Kill(pid, 0)
+	if err != nil {
+		return false
+	}
+	// On Linux /proc is available; check whether the process is a zombie.
+	if b, readErr := os.ReadFile(fmt.Sprintf("/proc/%d/stat", pid)); readErr == nil {
+		// Format: "pid (comm) state ..." — state follows the closing ')' of the
+		// command name (which may itself contain spaces and parens).
+		rest := string(b)
+		if idx := strings.LastIndex(rest, ") "); idx >= 0 {
+			fields := strings.Fields(rest[idx+2:])
+			if len(fields) > 0 && fields[0] == "Z" {
+				return false // zombie: terminated but not yet reaped
+			}
+		}
+	}
+	return true
+}
+
+// TestKillerKillsTree verifies that a process group captured by the killer is
+// terminated together with a child the process spawns afterwards. This mirrors
+// a step or post-task script that launches a child which spawns further
+// processes, where cancelling must take down the whole tree, not just the
+// direct child.
+func TestKillerKillsTree(t *testing.T) {
+	dir := t.TempDir()
+	pidFile := filepath.Join(dir, "child.pid")
+
+	// Parent shell backgrounds a long-lived child (writing its PID to a file)
+	// and then sleeps. With job control off (non-interactive sh) the backgrounded
+	// child stays in the parent's process group, so the group kill must reach it.
+	script := fmt.Sprintf(`sleep 600 & echo $! > %q; sleep 600`, pidFile)
+	cmd := exec.Command("/bin/sh", "-c", script)
+	// Launch as its own process-group leader, exactly like a real process does
+	// (see SysProcAttr), so the killer's PGID == the process PID.
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	require.NoError(t, cmd.Start())
+	t.Cleanup(func() {
+		_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
+		_ = cmd.Wait()
+	})
+
+	killer, err := NewKiller(cmd.Process)
+	require.NoError(t, err)
+	defer killer.Close()
+
+	// Wait for the backgrounded child PID to be reported.
+	var childPID int
+	require.Eventually(t, func() bool {
+		b, e := os.ReadFile(pidFile)
+		if e != nil {
+			return false
+		}
+		s := strings.TrimSpace(string(b))
+		if s == "" {
+			return false
+		}
+		childPID, _ = strconv.Atoi(s)
+		return childPID > 0 && processAlive(childPID)
+	}, 20*time.Second, 100*time.Millisecond, "child process should start")
+
+	// Killing the group must terminate both the parent and the backgrounded child.
+	require.NoError(t, killer.Kill())
+	// Reap the parent so it does not linger as a zombie (which would still report
+	// as alive); SIGKILL makes Wait return promptly.
+	_ = cmd.Wait()
+
+	require.Eventually(t, func() bool {
+		return !processAlive(childPID)
+	}, 20*time.Second, 100*time.Millisecond, "backgrounded child should be terminated")
+}

internal/pkg/process/killer_windows.go

@@ -0,0 +1,72 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package process
+
+import (
+	"os"
+
+	"golang.org/x/sys/windows"
+)
+
+// Killer terminates a started process together with its entire descendant tree
+// via a Windows Job Object.
+//
+// Background: a process (a step or a post-task script) often launches a process
+// tree (a shell that starts a child which in turn spawns further GUI or
+// background processes). The default exec.CommandContext cancellation only kills
+// the direct child, so cancelling left the rest of the tree running. Because
+// those orphans inherited the parent's stdout/stderr pipe, cmd.Wait() also
+// blocked forever and the runner hung.
+//
+// Assigning the process to a Job Object lets us kill the whole tree atomically
+// on cancellation (TerminateJobObject), which also closes the inherited pipe
+// handles so cmd.Wait() can return.
+type Killer struct {
+	job windows.Handle
+}
+
+// NewKiller creates a Job Object and assigns p (an already-started process) to
+// it. Children spawned by p afterwards are automatically part of the job. The
+// job does NOT use JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE, so closing the handle on
+// normal completion does not kill legitimate background processes; the tree is
+// only torn down by an explicit Kill (cancellation).
+func NewKiller(p *os.Process) (*Killer, error) {
+	job, err := windows.CreateJobObject(nil, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	h, err := windows.OpenProcess(windows.PROCESS_SET_QUOTA|windows.PROCESS_TERMINATE, false, uint32(p.Pid))
+	if err != nil {
+		_ = windows.CloseHandle(job)
+		return nil, err
+	}
+	defer func() { _ = windows.CloseHandle(h) }()
+
+	if err := windows.AssignProcessToJobObject(job, h); err != nil {
+		_ = windows.CloseHandle(job)
+		return nil, err
+	}
+
+	return &Killer{job: job}, nil
+}
+
+// Kill terminates every process currently assigned to the job (the process and
+// all of its descendants).
+func (k *Killer) Kill() error {
+	if k == nil || k.job == 0 {
+		return nil
+	}
+	return windows.TerminateJobObject(k.job, 1)
+}
+
+// Close releases the job handle. It does not terminate the processes.
+func (k *Killer) Close() error {
+	if k == nil || k.job == 0 {
+		return nil
+	}
+	h := k.job
+	k.job = 0
+	return windows.CloseHandle(h)
+}

internal/pkg/process/killer_windows_test.go

@@ -1,7 +1,7 @@
 // Copyright 2026 The Gitea Authors. All rights reserved.
 // SPDX-License-Identifier: MIT
 
-package container
+package process
 
 import (
 	"fmt"
@@ -23,7 +23,7 @@ func processAlive(pid int) bool {
 	if err != nil {
 		return false
 	}
-	defer windows.CloseHandle(h)
+	defer func() { _ = windows.CloseHandle(h) }()
 	var code uint32
 	if err := windows.GetExitCodeProcess(h, &code); err != nil {
 		return false
@@ -32,11 +32,11 @@ func processAlive(pid int) bool {
 	return code == stillActive
 }
 
-// TestProcessKillerKillsTree verifies that a process assigned to the Job Object
+// TestKillerKillsTree verifies that a process assigned to the Job Object is
-// is terminated together with a child it spawns afterwards. This mirrors a step
+// terminated together with a child it spawns afterwards. This mirrors a step or
-// that launches a child which spawns further processes, where cancelling the
+// post-task script that launches a child which spawns further processes, where
-// job must take down the whole tree, not just the direct child.
+// cancelling must take down the whole tree, not just the direct child.
-func TestProcessKillerKillsTree(t *testing.T) {
+func TestKillerKillsTree(t *testing.T) {
 	dir := t.TempDir()
 	pidFile := filepath.Join(dir, "child.pid")
 
@@ -50,7 +50,7 @@ func TestProcessKillerKillsTree(t *testing.T) {
 	require.NoError(t, cmd.Start())
 	t.Cleanup(func() { _ = cmd.Process.Kill() })
 
-	killer, err := newProcessKiller(cmd.Process)
+	killer, err := NewKiller(cmd.Process)
 	require.NoError(t, err)
 	defer killer.Close()
 

internal/pkg/process/sysprocattr_plan9.go

@@ -0,0 +1,17 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build plan9
+
+package process
+
+import "syscall"
+
+// SysProcAttr returns the platform attributes used to start a process. Plan 9
+// has no process-group tree-kill (see Killer), so we only request a new rfork
+// note group here.
+func SysProcAttr(cmdLine string, tty bool) *syscall.SysProcAttr {
+	return &syscall.SysProcAttr{
+		Rfork: syscall.RFNOTEG,
+	}
+}

internal/pkg/process/sysprocattr_unix.go

@@ -0,0 +1,24 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+//go:build !windows && !plan9
+
+package process
+
+import "syscall"
+
+// SysProcAttr returns the platform attributes used to start a process so that a
+// Killer can later tear down its whole process tree. On Unix the process becomes
+// the leader of a new process group (or session, for the PTY path), so a
+// signal to the negative PID reaches every descendant that stayed in the group.
+func SysProcAttr(_ string, tty bool) *syscall.SysProcAttr {
+	if tty {
+		return &syscall.SysProcAttr{
+			Setsid:  true,
+			Setctty: true,
+		}
+	}
+	return &syscall.SysProcAttr{
+		Setpgid: true,
+	}
+}

internal/pkg/process/sysprocattr_windows.go

@@ -0,0 +1,14 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package process
+
+import "syscall"
+
+// SysProcAttr returns the platform attributes used to start a process so that a
+// Killer can later tear down its whole process tree. On Windows the process is
+// placed in a new process group; the descendant tree is reclaimed via the Job
+// Object set up by NewKiller.
+func SysProcAttr(cmdLine string, tty bool) *syscall.SysProcAttr {
+	return &syscall.SysProcAttr{CmdLine: cmdLine, CreationFlags: syscall.CREATE_NEW_PROCESS_GROUP}
+}

internal/pkg/process/treekill.go

@@ -0,0 +1,66 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package process
+
+import (
+	"os"
+	"os/exec"
+	"sync/atomic"
+	"time"
+)
+
+// treeKillWaitDelay bounds how long Wait lingers for the command's I/O pipes to
+// drain after the process exits before force-closing them and returning. It also
+// covers a command that backgrounds a process holding a pipe open after a clean
+// exit.
+const treeKillWaitDelay = 10 * time.Second
+
+// TreeKill wires an exec.Cmd so that cancelling it tears down the command's
+// whole process tree (see Killer) rather than only the direct child, and bounds
+// the post-exit I/O wait so a leftover pipe writer can never hang cmd.Wait.
+//
+// Background: a command often launches a process tree (a shell that starts a
+// child which spawns further background processes). The default
+// exec.CommandContext cancellation only kills the direct child, leaving the rest
+// of the tree running; and because the orphans inherit cmd's stdout/stderr pipe,
+// cmd.Wait() would block forever, hanging the caller.
+//
+// Callers still set cmd.SysProcAttr (via SysProcAttr) themselves, because the
+// value differs between the plain and PTY execution paths.
+type TreeKill struct {
+	killer atomic.Pointer[Killer]
+}
+
+// NewTreeKill sets cmd.Cancel and cmd.WaitDelay. Call it before cmd.Start, then
+// call Capture once after a successful Start.
+func NewTreeKill(cmd *exec.Cmd) *TreeKill {
+	t := &TreeKill{}
+	cmd.Cancel = func() error {
+		if k := t.killer.Load(); k != nil {
+			return k.Kill()
+		}
+		if cmd.Process != nil {
+			return cmd.Process.Kill()
+		}
+		return nil
+	}
+	cmd.WaitDelay = treeKillWaitDelay
+	return t
+}
+
+// Capture assigns the started process (and the descendants it spawns) to a
+// Killer so cancellation can reach the whole tree — a Job Object on Windows
+// (children spawned afterwards are auto-included) and the process group on Unix.
+// Call it once after cmd.Start. On failure the command falls back to the default
+// single-process kill and the returned error is for logging only; WaitDelay
+// still bounds the wait. The returned Killer should be closed when the command
+// finishes (Close is nil-safe).
+func (t *TreeKill) Capture(p *os.Process) (*Killer, error) {
+	k, err := NewKiller(p)
+	if err != nil {
+		return nil, err
+	}
+	t.killer.Store(k)
+	return k, nil
+}

internal/pkg/report/reporter.go

@@ -49,6 +49,8 @@ type Reporter struct {
 	stateMu           sync.RWMutex
 	outputs           sync.Map
 	daemon            chan struct{}
+	heartbeatStop     chan struct{}
+	heartbeatStopOnce sync.Once
 
 	// Unix-nanos of the last successful UpdateTask. Atomic so the heartbeat
 	// guard in ReportState reads it without contending stateMu.
@@ -100,6 +102,7 @@ func NewReporter(ctx context.Context, cancel context.CancelFunc, client client.C
 			Id: task.Id,
 		},
 		daemon:        make(chan struct{}),
+		heartbeatStop: make(chan struct{}),
 	}
 
 	if task.Secrets["ACTIONS_STEP_DEBUG"] == "true" {
@@ -273,6 +276,15 @@ func (r *Reporter) RunDaemon() {
 	go r.runDaemonLoop()
 }
 
+// StopHeartbeats stops periodic UpdateTask heartbeats without cancelling the
+// task context. Close() still delivers the final flush. Safe to call multiple
+// times and when the context is already cancelled.
+func (r *Reporter) StopHeartbeats() {
+	r.heartbeatStopOnce.Do(func() {
+		close(r.heartbeatStop)
+	})
+}
+
 func (r *Reporter) stopLatencyTimer(active *bool, timer *time.Timer) {
 	if *active {
 		if !timer.Stop() {
@@ -339,6 +351,12 @@ func (r *Reporter) runDaemonLoop() {
 			// delivers the final flush on a detached context (flushFinal).
 			close(r.daemon)
 			return
+
+		case <-r.heartbeatStop:
+			// Stop heartbeating during post-task script execution. Close() still
+			// delivers the final flush on a detached context (flushFinal).
+			close(r.daemon)
+			return
 		}
 
 		r.stateMu.RLock()

internal/pkg/report/reporter_test.go

@@ -921,3 +921,65 @@ func TestReporter_CloseReportsCancelledOnCanceledCtx(t *testing.T) {
 	assert.True(t, foundCancelled, "final log must contain a 'Cancelled' row")
 	assert.False(t, foundEarlyTermination, "final log must not contain 'Early termination' on the cancel path")
 }
+
+// TestReporter_StopHeartbeats verifies that StopHeartbeats ends periodic
+// UpdateTask heartbeats while Close() still flushes the final state.
+func TestReporter_StopHeartbeats(t *testing.T) {
+	var updateTaskCalls atomic.Int64
+
+	client := mocks.NewClient(t)
+	client.On("UpdateLog", mock.Anything, mock.Anything).Maybe().Return(
+		func(_ context.Context, req *connect_go.Request[runnerv1.UpdateLogRequest]) (*connect_go.Response[runnerv1.UpdateLogResponse], error) {
+			return connect_go.NewResponse(&runnerv1.UpdateLogResponse{
+				AckIndex: req.Msg.Index + int64(len(req.Msg.Rows)),
+			}), nil
+		},
+	)
+	client.On("UpdateTask", mock.Anything, mock.Anything).Return(
+		func(_ context.Context, _ *connect_go.Request[runnerv1.UpdateTaskRequest]) (*connect_go.Response[runnerv1.UpdateTaskResponse], error) {
+			updateTaskCalls.Add(1)
+			return connect_go.NewResponse(&runnerv1.UpdateTaskResponse{}), nil
+		},
+	)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	taskCtx, err := structpb.NewStruct(map[string]any{})
+	require.NoError(t, err)
+
+	cfg, err := config.LoadDefault("")
+	require.NoError(t, err)
+	cfg.Runner.StateReportInterval = 20 * time.Millisecond
+	cfg.Runner.LogReportInterval = time.Hour
+
+	reporter := NewReporter(ctx, cancel, client, &runnerv1.Task{Context: taskCtx}, cfg)
+	reporter.ResetSteps(1)
+	reporter.RunDaemon()
+
+	reporter.stateMu.Lock()
+	reporter.stateChanged = true
+	reporter.state.Result = runnerv1.Result_RESULT_SUCCESS
+	reporter.state.StoppedAt = timestamppb.Now()
+	reporter.stateMu.Unlock()
+
+	require.Eventually(t, func() bool {
+		return updateTaskCalls.Load() >= 1
+	}, time.Second, 5*time.Millisecond, "daemon must send at least one UpdateTask before StopHeartbeats")
+
+	beforeStop := updateTaskCalls.Load()
+	reporter.StopHeartbeats()
+
+	select {
+	case <-reporter.daemon:
+	case <-time.After(time.Second):
+		t.Fatal("StopHeartbeats must stop the daemon loop")
+	}
+
+	time.Sleep(3 * cfg.Runner.StateReportInterval)
+	assert.Equal(t, beforeStop, updateTaskCalls.Load(),
+		"UpdateTask must not be called after StopHeartbeats")
+
+	require.NoError(t, reporter.Close(""))
+	assert.Greater(t, updateTaskCalls.Load(), beforeStop,
+		"Close() must still send a final UpdateTask after StopHeartbeats")
+}