tiny-forge/internal/workload/plugin/source/static/deploy.go

package static

import (
	"context"
	"encoding/json"
	"fmt"
	"io"
	"log/slog"
	"os"
	"path/filepath"
	"strconv"
	"strings"
	"time"

	"github.com/moby/moby/api/types/mount"

	"github.com/alexei/tinyforge/internal/crypto"
	"github.com/alexei/tinyforge/internal/docker"
	"github.com/alexei/tinyforge/internal/events"
	"github.com/alexei/tinyforge/internal/notify"
	"github.com/alexei/tinyforge/internal/proxy"
	"github.com/alexei/tinyforge/internal/staticsite"
	"github.com/alexei/tinyforge/internal/store"
	"github.com/alexei/tinyforge/internal/workload/plugin"
)

// healthCheckDelay is the brief grace window after StartContainer
// before we probe IsContainerRunning. Short enough not to bog down a
// healthy deploy; long enough to catch crash-on-boot failures
// (missing env var, bad Dockerfile, port conflict).
const healthCheckDelay = 3 * time.Second

// deploy runs one full sync of a static workload: fetch the latest
// commit, optionally rebuild the image, recreate the container, and
// reconfigure the proxy. Ported from internal/staticsite/manager.go to
// operate directly on plugin.Workload + container row state, without
// the legacy static_sites table.
//
// Behavior parity with the legacy path is the explicit goal — the
// log-line format ("Static site \"%s\": %s") and event payload shapes
// are preserved so log scrapers and SSE clients keep working through
// the cutover.
func deploy(ctx context.Context, deps plugin.Deps, w plugin.Workload, intent plugin.DeploymentIntent) (retErr error) {
	cfg, err := plugin.SourceConfigOf[Config](w)
	if err != nil {
		return fmt.Errorf("static source: decode config: %w", err)
	}

	prev, prevContainer, err := loadState(deps, w)
	if err != nil {
		return err
	}

	// Manual / first-time deploys force a full rebuild even when the
	// commit SHA is unchanged. The legacy Manager.Deploy was called
	// with force=true from the adapter; preserve that semantic by
	// treating any non-cron / non-git intent as forcing.
	force := intent.Reason == "" || intent.Reason == "manual" || intent.Reason == "promote"

	// Decrypt the access token if present. Kept in a local so the
	// sanitizer can scrub it from any error string before persisting.
	token := ""
	if cfg.AccessToken != "" {
		decrypted, derr := crypto.Decrypt(deps.EncKey, cfg.AccessToken)
		if derr != nil {
			slog.Warn("static source: failed to decrypt access token", "site", w.Name, "error", derr)
		} else {
			token = decrypted
		}
	}

	provider, err := staticsite.NewGitProvider(staticsite.ProviderType(cfg.Provider), cfg.BaseURL, token)
	if err != nil {
		updateStatus(deps, w, "failed", prev.LastCommitSHA,
			sanitizeError(fmt.Sprintf("create provider: %v", err), token))
		return fmt.Errorf("create provider: %w", err)
	}

	latestSHA, err := provider.GetLatestCommitSHA(ctx, cfg.RepoOwner, cfg.RepoName, cfg.Branch)
	if err != nil {
		updateStatus(deps, w, "failed", prev.LastCommitSHA,
			sanitizeError(fmt.Sprintf("fetch commit SHA: %v", err), token))
		return fmt.Errorf("get latest commit: %w", err)
	}

	// Resolve the public-facing domain from the workload's first enabled
	// public face. Mirrors the synthetic-row adapter's logic so the
	// proxy registration sees the same FQDN it did before.
	domain := primaryDomain(deps, w)

	// Commit-status reporter (best-effort; gated on cfg.ReportCommitStatus).
	// Built here once latestSHA + domain are known. A deferred terminal
	// report fires Success/Failure based on the deploy's outcome, but ONLY
	// once an actual deploy began (deployStarted) — the unchanged-SHA
	// short-circuit below returns before that flips, so no status is
	// reported when nothing was deployed. retErr is the named return the
	// defer inspects.
	reporter := newCommitStatusReporter(provider, cfg, latestSHA, statusTargetURL(domain))
	deployStarted := false
	defer func() {
		if !deployStarted {
			return
		}
		if retErr != nil {
			reporter.report(ctx, w, staticsite.CommitStatusFailure, "Tinyforge: deploy failed")
		} else {
			reporter.report(ctx, w, staticsite.CommitStatusSuccess, "Tinyforge: deployed")
		}
	}()

	// Skip redeploy when nothing changed AND we have a live container +
	// (if applicable) live proxy route. Manual deploys always force.
	prevContainerID := ""
	prevProxyRouteID := ""
	if prevContainer != nil {
		prevContainerID = prevContainer.ContainerID
		prevProxyRouteID = prevContainer.ProxyRouteID
	}
	if !force && latestSHA == prev.LastCommitSHA && prev.Status == "deployed" && prevContainerID != "" {
		running, _ := deps.Docker.IsContainerRunning(ctx, prevContainerID)
		if !running {
			slog.Info("static site: container not running, forcing redeploy", "site", w.Name)
		} else if domain != "" {
			proxyOK, perr := deps.Proxy.RouteExists(ctx, domain)
			if perr != nil {
				slog.Warn("static site: proxy check failed, forcing redeploy", "site", w.Name, "error", perr)
			} else if !proxyOK {
				slog.Info("static site: proxy route missing, forcing redeploy", "site", w.Name)
			} else {
				slog.Info("static site: no changes", "site", w.Name, "sha", latestSHA)
				return nil
			}
		} else {
			slog.Info("static site: no changes", "site", w.Name, "sha", latestSHA)
			return nil
		}
	}

	// Mark syncing. From here on a deploy is genuinely underway, so the
	// deferred terminal status report should fire. Push a "pending" commit
	// status (best-effort) and arm the deferred Success/Failure report.
	updateStatus(deps, w, "syncing", prev.LastCommitSHA, "")
	publishEvent(deps, w, "syncing")
	deployStarted = true
	reporter.report(ctx, w, staticsite.CommitStatusPending, "Tinyforge: deploying")

	// Build context — temp dir cleaned up on every exit path.
	buildDir, err := os.MkdirTemp("", "dw-site-"+idShort(w)+"-*")
	if err != nil {
		updateStatus(deps, w, "failed", prev.LastCommitSHA,
			sanitizeError(fmt.Sprintf("create temp dir: %v", err), token))
		return fmt.Errorf("create temp dir: %w", err)
	}
	defer os.RemoveAll(buildDir)

	if err := provider.DownloadFolder(ctx, cfg.RepoOwner, cfg.RepoName, cfg.Branch, cfg.FolderPath, buildDir); err != nil {
		updateStatus(deps, w, "failed", prev.LastCommitSHA,
			sanitizeError(fmt.Sprintf("download folder: %v", err), token))
		return fmt.Errorf("download folder: %w", err)
	}

	// Defense in depth: providers should never write outside buildDir,
	// but a hostile self-hosted Gitea/GitLab the operator pointed at
	// could in principle return a tree entry that escapes. Verify
	// before the copy step materializes the build context.
	if err := verifyDownloadInsideRoot(buildDir); err != nil {
		updateStatus(deps, w, "failed", prev.LastCommitSHA,
			sanitizeError(fmt.Sprintf("downloaded tree rejected: %v", err), token))
		return fmt.Errorf("downloaded tree rejected: %w", err)
	}

	if cfg.RenderMarkdown {
		if err := staticsite.RenderMarkdownFiles(buildDir); err != nil {
			slog.Warn("static site: markdown rendering failed", "site", w.Name, "error", err)
		}
	}

	// Detect mode: deno requires an api/ folder. Fall back to static if
	// the operator declared deno but the repo doesn't carry routes.
	mode := cfg.Mode
	apiDir := filepath.Join(buildDir, "api")
	hasAPI := false
	if info, err := os.Stat(apiDir); err == nil && info.IsDir() {
		hasAPI = true
	}
	if mode == "deno" && !hasAPI {
		mode = "static"
		slog.Info("static site: no api/ folder found, falling back to static mode", "site", w.Name)
	}

	imageTag := imageTagFor(w)
	contextDir, err := os.MkdirTemp("", "dw-site-build-*")
	if err != nil {
		updateStatus(deps, w, "failed", latestSHA,
			sanitizeError(fmt.Sprintf("create build context: %v", err), token))
		return fmt.Errorf("create build context dir: %w", err)
	}
	defer os.RemoveAll(contextDir)

	if mode == "deno" {
		if err := prepareDenoBuild(buildDir, contextDir); err != nil {
			updateStatus(deps, w, "failed", latestSHA,
				sanitizeError(fmt.Sprintf("prepare deno build: %v", err), token))
			return fmt.Errorf("prepare deno build: %w", err)
		}
	} else {
		if err := prepareStaticBuild(buildDir, contextDir); err != nil {
			updateStatus(deps, w, "failed", latestSHA,
				sanitizeError(fmt.Sprintf("prepare static build: %v", err), token))
			return fmt.Errorf("prepare static build: %w", err)
		}
	}

	if err := deps.Docker.BuildImage(ctx, contextDir, imageTag); err != nil {
		updateStatus(deps, w, "failed", latestSHA,
			sanitizeError(fmt.Sprintf("build image: %v", err), token))
		return fmt.Errorf("build image: %w", err)
	}

	env := buildEnv(deps, w.ID)

	containerPort := "80"
	if mode == "deno" {
		containerPort = "8000"
	}

	settings, err := deps.Store.GetSettings()
	if err != nil {
		updateStatus(deps, w, "failed", latestSHA,
			sanitizeError(fmt.Sprintf("get settings: %v", err), token))
		return fmt.Errorf("get settings: %w", err)
	}

	networkName := settings.Network
	networkID, err := deps.Docker.EnsureNetwork(ctx, networkName)
	if err != nil {
		updateStatus(deps, w, "failed", latestSHA,
			sanitizeError(fmt.Sprintf("ensure network: %v", err), token))
		return fmt.Errorf("ensure network: %w", err)
	}

	containerName := containerNameFor(w)

	var mounts []mount.Mount
	if cfg.StorageEnabled && mode == "deno" {
		volName, volErr := deps.Docker.EnsureSiteVolume(ctx, siteVolumeKey(w))
		if volErr != nil {
			slog.Warn("static site: failed to ensure storage volume", "site", w.Name, "error", volErr)
		} else {
			mounts = append(mounts, mount.Mount{
				Type:   mount.TypeVolume,
				Source: volName,
				Target: "/app/data",
			})
			slog.Info("static site: storage volume attached", "site", w.Name, "volume", volName)
		}
	}

	// Per-face proxy labels (Traefik picks these up; NPM ignores them).
	// Static workloads have at most one face today, but iterate for
	// future multi-face parity with the image source.
	labels := map[string]string{}
	if domain != "" {
		port, _ := strconv.Atoi(containerPort)
		if l := deps.Proxy.ContainerLabels(domain, port); l != nil {
			for k, v := range l {
				labels[k] = v
			}
		}
	}

	cc := docker.ContainerConfig{
		Name:         containerName,
		Image:        imageTag,
		Env:          env,
		ExposedPorts: []string{containerPort + "/tcp"},
		NetworkName:  networkName,
		NetworkID:    networkID,
		Mounts:       mounts,
		Labels:       labels,
		WorkloadID:   w.ID,
		WorkloadKind: string(store.WorkloadKindSite),
		Role:         "",
	}

	containerID, err := deps.Docker.CreateContainer(ctx, cc)
	if err != nil {
		// Container with this name might already exist — best-effort
		// cleanup of any prior container by ID and by name, then retry.
		if prevContainerID != "" {
			deps.Docker.StopContainer(ctx, prevContainerID, 10)
			deps.Docker.RemoveContainer(ctx, prevContainerID, true)
		}
		removeContainerByName(ctx, deps, containerName)

		containerID, err = deps.Docker.CreateContainer(ctx, cc)
		if err != nil {
			updateStatus(deps, w, "failed", latestSHA,
				sanitizeError(fmt.Sprintf("create container: %v", err), token))
			return fmt.Errorf("create container: %w", err)
		}
	}

	if err := deps.Docker.StartContainer(ctx, containerID); err != nil {
		deps.Docker.RemoveContainer(ctx, containerID, true)
		updateStatus(deps, w, "failed", latestSHA,
			sanitizeError(fmt.Sprintf("start container: %v", err), token))
		return fmt.Errorf("start container: %w", err)
	}

	// Brief health-check window — verify the container survives a few
	// seconds after start, surfacing the tail of its logs as the
	// failure reason if it crashes. Honor ctx so a cancelled deploy
	// returns promptly instead of waiting out the full delay.
	select {
	case <-ctx.Done():
		deps.Docker.RemoveContainer(ctx, containerID, true)
		updateStatus(deps, w, "failed", latestSHA, "deploy cancelled before health check")
		return ctx.Err()
	case <-time.After(healthCheckDelay):
	}
	running, runErr := deps.Docker.IsContainerRunning(ctx, containerID)
	if runErr != nil || !running {
		logMsg := "container exited immediately after start"
		if logs, logErr := deps.Docker.ContainerLogs(ctx, containerID, false, "20"); logErr == nil {
			buf, _ := io.ReadAll(logs)
			logs.Close()
			if len(buf) > 0 {
				logMsg = sanitizeError(string(buf), token)
			}
		}
		deps.Docker.RemoveContainer(ctx, containerID, true)
		updateStatus(deps, w, "failed", latestSHA, logMsg)
		return fmt.Errorf("container not running: %s", logMsg)
	}

	// Resolve proxy target. Default to in-network DNS (containerName);
	// switch to (settings.ServerIP, hostPort) under NPM remote mode.
	internalPort, _ := strconv.Atoi(containerPort)
	forwardHost := containerName
	forwardPort := internalPort
	if settings.NpmRemote && settings.ProxyProvider == "npm" {
		if settings.ServerIP != "" {
			hostPort, hpErr := deps.Docker.InspectContainerPort(ctx, containerID, containerPort+"/tcp")
			if hpErr != nil {
				slog.Warn("static site: could not get host port for remote NPM", "site", w.Name, "error", hpErr)
			} else {
				forwardHost = settings.ServerIP
				forwardPort = int(hostPort)
			}
		}
	}

	// Configure proxy if a domain is set. Replace any prior route in-
	// place so traffic shifts atomically.
	proxyRouteID := prevProxyRouteID
	if domain != "" {
		if prevProxyRouteID != "" {
			deps.Proxy.DeleteRoute(ctx, prevProxyRouteID)
		}
		routeID, rerr := deps.Proxy.ConfigureRoute(ctx, domain, forwardHost, forwardPort, proxy.RouteOptions{
			SSLCertificateID: settings.SSLCertificateID,
		})
		if rerr != nil {
			slog.Warn("static site: failed to configure proxy",
				"site", w.Name, "domain", domain,
				"target", fmt.Sprintf("%s:%d", forwardHost, forwardPort), "error", rerr)
		} else {
			proxyRouteID = routeID
			slog.Info("static site: proxy configured",
				"site", w.Name, "domain", domain,
				"target", fmt.Sprintf("%s:%d", forwardHost, forwardPort), "routeID", routeID)
		}
	}

	// Drop the old container if a fresh one was created (different ID).
	if prevContainerID != "" && prevContainerID != containerID {
		deps.Docker.StopContainer(ctx, prevContainerID, 10)
		deps.Docker.RemoveContainer(ctx, prevContainerID, true)
	}

	// Single transactional write of the new state + container metadata.
	// On failure: tear down the just-created container and proxy route
	// so we don't leave orphans behind. The next deploy would otherwise
	// see no row and try to create a third container.
	if err := saveState(deps, w, func(rs *runtimeState, c *store.Container) {
		rs.LastCommitSHA = latestSHA
		rs.LastSyncAt = store.Now()
		rs.LastError = ""
		rs.Status = "deployed"

		c.ContainerID = containerID
		c.ProxyRouteID = proxyRouteID
		c.Subdomain = domain
		c.State = "running"
		c.Port = internalPort
		c.ImageRef = imageTag
	}); err != nil {
		slog.Error("static site: failed to persist deploy state — rolling back",
			"site", w.Name, "error", err)
		if proxyRouteID != "" {
			deps.Proxy.DeleteRoute(ctx, proxyRouteID)
		}
		deps.Docker.StopContainer(ctx, containerID, 10)
		deps.Docker.RemoveContainer(ctx, containerID, true)
		// Best-effort failure-state write so the operator sees the
		// deploy failed instead of a silent gap. If even this fails
		// we have nothing left to log.
		updateStatus(deps, w, "failed", latestSHA,
			sanitizeError(fmt.Sprintf("persist deploy state: %v", err), token))
		return fmt.Errorf("persist deploy state: %w", err)
	}

	publishEvent(deps, w, "deployed")

	// updateStatus normally fires the terminal-state notification; the
	// success path above wrote state via saveState directly, so dispatch
	// the deployed notification explicitly here.
	dispatchSiteNotification(deps, w, domain, "deployed", "")

	shaDisplay := latestSHA
	if len(shaDisplay) > 8 {
		shaDisplay = shaDisplay[:8]
	}
	slog.Info("static site deployed", "site", w.Name, "sha", shaDisplay, "mode", mode)
	return nil
}

// commitStatusReporter pushes deploy outcomes back to the git provider as
// a commit status, gated on the per-workload report_commit_status flag.
// It is strictly best-effort: every call is wrapped so a reporting failure
// is logged at Warn and NEVER propagates to fail or block the deploy.
//
// The provider + identifiers are captured once at deploy start so the hot
// transition points (pending/success/failure) read as one-liners. A nil
// receiver (reporting disabled) makes report a no-op, so callers don't have
// to guard each site.
type commitStatusReporter struct {
	provider  staticsite.GitProvider
	owner     string
	repo      string
	sha       string
	targetURL string
	enabled   bool
}

// newCommitStatusReporter builds a reporter from the decoded config. When
// report_commit_status is off (or the SHA is empty) the returned reporter's
// report method is inert.
func newCommitStatusReporter(provider staticsite.GitProvider, cfg Config, sha, targetURL string) *commitStatusReporter {
	return &commitStatusReporter{
		provider:  provider,
		owner:     cfg.RepoOwner,
		repo:      cfg.RepoName,
		sha:       sha,
		targetURL: targetURL,
		enabled:   cfg.ReportCommitStatus,
	}
}

// report sends one commit status, swallowing (and logging) any error. Safe
// to call on a disabled reporter or with a nil provider/empty SHA.
func (r *commitStatusReporter) report(ctx context.Context, w plugin.Workload, status staticsite.CommitStatus, description string) {
	if r == nil || !r.enabled || r.provider == nil || r.sha == "" {
		return
	}
	if err := r.provider.SetCommitStatus(ctx, r.owner, r.repo, r.sha, status, r.targetURL, description); err != nil {
		slog.Warn("static site: commit-status report failed (ignored)",
			"site", w.Name, "status", string(status), "error", err)
	}
}

// statusTargetURL derives the https URL the commit status links back to —
// the workload's primary public face, or "" when it has none.
func statusTargetURL(domain string) string {
	if domain == "" {
		return ""
	}
	return "https://" + domain
}

// updateStatus writes the runtime state's status/error/commit fields
// and fires the side effects the legacy Manager.updateStatus did:
// failures land in the event log, and terminal transitions trigger an
// outbound notification.
//
// On the deploy success path saveState is called directly (with the
// full container metadata in the same write); this helper covers the
// failure / intermediate transitions where only state moves.
func updateStatus(deps plugin.Deps, w plugin.Workload, status, commitSHA, errMsg string) {
	if err := saveState(deps, w, func(rs *runtimeState, c *store.Container) {
		rs.Status = status
		rs.LastError = errMsg
		if commitSHA != "" {
			rs.LastCommitSHA = commitSHA
		}
		// Reflect status into the container row state column so the
		// global containers index stays useful for filtered queries.
		switch status {
		case "deployed":
			c.State = "running"
		case "stopped":
			c.State = "stopped"
		case "failed":
			c.State = "failed"
		case "syncing":
			// Don't churn the container row's state during in-progress
			// syncs — leave it on whatever value the previous deploy left.
		}
	}); err != nil {
		slog.Error("static site: failed to update status", "id", w.ID, "status", status, "error", err)
	}

	if status == "failed" {
		publishEvent(deps, w, "failed: "+errMsg)
	}

	if status == "deployed" || status == "failed" {
		dispatchSiteNotification(deps, w, primaryDomain(deps, w), status, errMsg)
	}
}

// dispatchSiteNotification fires a site_sync_success or
// site_sync_failure event for the workload via the shared multi-route
// dispatcher in plugin.DispatchNotificationForWorkload. Resolution
// order (workload_notifications → legacy single URL → settings global)
// is identical to the dockerfile plugin's path so receivers see
// consistent fan-out behaviour across source kinds.
func dispatchSiteNotification(deps plugin.Deps, w plugin.Workload, domain, status, errMsg string) {
	eventType := "site_sync_success"
	if status == "failed" {
		eventType = "site_sync_failure"
	}
	siteURL := ""
	if domain != "" {
		siteURL = "https://" + domain
	}
	plugin.DispatchNotificationForWorkload(deps, w, notify.Event{
		Type:    eventType,
		Project: w.Name,
		URL:     siteURL,
		Error:   errMsg,
	})
}

// publishEvent emits a static_site_status event on the bus AND
// persists an event_log row so the dashboard's audit trail picks it
// up. Message format ("Static site \"%s\": %s") is preserved verbatim
// from the legacy Manager.publishEvent so log scrapers and operator-
// configured event triggers keep matching.
func publishEvent(deps plugin.Deps, w plugin.Workload, status string) {
	deps.Events.Publish(events.Event{
		Type: events.EventStaticSiteStatus,
		Payload: events.StaticSiteStatusPayload{
			SiteID: w.ID,
			Name:   w.Name,
			Status: status,
		},
	})

	severity := "info"
	if strings.HasPrefix(status, "failed") {
		severity = "error"
	}
	message := fmt.Sprintf("Static site %q: %s", w.Name, status)

	// Build metadata via json.Marshal so workload names containing
	// quotes or backslashes don't produce invalid JSON for downstream
	// log-scan consumers.
	metaBytes, err := json.Marshal(map[string]string{
		"site_id":   w.ID,
		"site_name": w.Name,
		"status":    status,
	})
	if err != nil {
		slog.Error("static site: marshal event metadata", "error", err)
		metaBytes = []byte("{}")
	}
	metadata := string(metaBytes)

	evt, err := deps.Store.InsertEvent(store.EventLog{
		Source:   "static_site",
		Severity: severity,
		Message:  message,
		Metadata: metadata,
	})
	if err != nil {
		slog.Error("static site: failed to persist event log", "error", err)
		return
	}
	deps.Events.Publish(events.Event{
		Type: events.EventLog,
		Payload: events.EventLogPayload{
			ID:        evt.ID,
			Source:    "static_site",
			Severity:  severity,
			Message:   message,
			Metadata:  metadata,
			CreatedAt: evt.CreatedAt,
		},
	})
}

// removeContainerByName mirrors the legacy helper: enumerate Docker's
// view and best-effort drop the matching container so a name conflict
// in CreateContainer is recoverable. Best-effort.
func removeContainerByName(ctx context.Context, deps plugin.Deps, name string) {
	containers, err := deps.Docker.ListContainers(ctx, nil)
	if err != nil {
		return
	}
	for _, c := range containers {
		if c.Name == name {
			deps.Docker.StopContainer(ctx, c.ID, 10)
			deps.Docker.RemoveContainer(ctx, c.ID, true)
			return
		}
	}
}

// primaryDomain derives the public-facing FQDN from the workload's
// first enabled public face. Static workloads support at most one
// face today, but iterate defensively in case the API contract
// loosens later. An empty return means "no proxy registration"; the
// container still runs and is reachable inside the docker network.
//
// For the bare-subdomain case (Domain == "" but Subdomain != "") the
// helper appends settings.Domain to form a complete FQDN — matching
// the legacy Manager which let settings.Domain fall through silently.
// On a settings lookup failure the bare subdomain is returned as-is
// so the proxy still gets *something* to register.
func primaryDomain(deps plugin.Deps, w plugin.Workload) string {
	for _, f := range w.PublicFaces {
		if f.Subdomain == "" && f.Domain == "" {
			continue
		}
		switch {
		case f.Subdomain != "" && f.Domain != "":
			return f.Subdomain + "." + f.Domain
		case f.Subdomain == "" && f.Domain != "":
			return f.Domain
		case f.Subdomain != "" && f.Domain == "":
			settings, err := deps.Store.GetSettings()
			if err != nil || settings.Domain == "" {
				return f.Subdomain
			}
			return f.Subdomain + "." + settings.Domain
		}
	}
	return ""
}