relay-flow 0.0.1 → 0.2.0-alpha

package/cmd/relay-flow/serve.go

@@ -0,0 +1,609 @@
+package main
+
+// Composition root for `relay-flow serve` (task 5.1). Explicit ordered Go
+// wiring per docs/structs-methods-interfaces.md "Startup Wiring"
+// (lines 1081-1097) and design.md decision 24. The flock on server.lock is
+// acquired FIRST; startup refuses to run if another relay-flow process
+// holds it. No container, no DI framework.
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log/slog"
+	"net"
+	"net/http"
+	"os"
+	"strings"
+	"sync"
+	"syscall"
+	"time"
+
+	"github.com/rajpopat27/relay-flow/internal/config"
+	"github.com/rajpopat27/relay-flow/internal/execution/goworkflows"
+	"github.com/rajpopat27/relay-flow/internal/harness"
+	"github.com/rajpopat27/relay-flow/internal/paths"
+	recoverpkg "github.com/rajpopat27/relay-flow/internal/recover"
+	"github.com/rajpopat27/relay-flow/internal/repo"
+	"github.com/rajpopat27/relay-flow/internal/router"
+	runsvc "github.com/rajpopat27/relay-flow/internal/run"
+	"github.com/rajpopat27/relay-flow/internal/runner"
+	"github.com/rajpopat27/relay-flow/internal/server"
+	"github.com/rajpopat27/relay-flow/internal/task"
+	"github.com/rajpopat27/relay-flow/internal/workflow"
+
+	// Adapter registrations (factories are registered by name via init).
+	_ "github.com/rajpopat27/relay-flow/internal/harness/opencode"
+	_ "github.com/rajpopat27/relay-flow/internal/runner/orca"
+	_ "github.com/rajpopat27/relay-flow/internal/task/jira"
+)
+
+// serveRoot is the composition root. It runs the documented startup order
+// and blocks until ctx is canceled, then shuts down in reverse.
+//
+// Order (docs/structs-methods-interfaces.md lines 1081-1097, verbatim):
+//
+//	flock → load machine config → select task/runner/harness factories →
+//	construct shared runner/harness → load repos + one task.System per repo →
+//	load workflow files → validate each workflow against every referenced
+//	repo task system → bind workflows+matchers to repos → open go-workflows
+//	SQLite engine and start its workers → construct the Run Manager →
+//	start the Repo Poller group → start the Unix-socket server.
+func serveRoot(ctx context.Context, p paths.Paths, recover bool) error {
+	// Flock FIRST (docs lines 1028/1142): refuse to start if another
+	// process holds it. The only pre-flock side effect is creating the
+	// Root directory so the lock file itself can exist; nothing else in
+	// the layout is touched until the lock is held.
+	if err := os.MkdirAll(p.Root, 0o700); err != nil {
+		return fmt.Errorf("create root %s: %w", p.Root, err)
+	}
+	if err := os.Chmod(p.Root, 0o700); err != nil {
+		return fmt.Errorf("chmod root %s: %w", p.Root, err)
+	}
+	lockFile, err := os.OpenFile(p.Lock, os.O_CREATE|os.O_RDWR, 0o600)
+	if err != nil {
+		return fmt.Errorf("open lock %s: %w", p.Lock, err)
+	}
+	defer lockFile.Close()
+	if err := lockFile.Chmod(0o600); err != nil {
+		return fmt.Errorf("chmod lock %s: %w", p.Lock, err)
+	}
+	if err := syscall.Flock(int(lockFile.Fd()), syscall.LOCK_EX|syscall.LOCK_NB); err != nil {
+		return fmt.Errorf("another relay-flow server holds %s", p.Lock)
+	}
+	defer func() { _ = syscall.Flock(int(lockFile.Fd()), syscall.LOCK_UN) }()
+
+	// Single-instance held; materialize the rest of the layout.
+	if err := paths.Ensure(p); err != nil {
+		return err
+	}
+
+	// Load machine config.
+	cfg, err := config.LoadMachine(p.Config)
+	if err != nil {
+		return err
+	}
+
+	// Select factories; construct shared runner and harness.
+	rnr, err := runner.New(cfg.RunnerPlugin, cfg.RunnerConfig)
+	if err != nil {
+		return fmt.Errorf("runner plugin %q: %w", cfg.RunnerPlugin, err)
+	}
+	hrn, err := harness.New(cfg.HarnessPlugin, cfg.HarnessConfig)
+	if err != nil {
+		return fmt.Errorf("harness plugin %q: %w", cfg.HarnessPlugin, err)
+	}
+
+	// Load registered repos and one task.System per repo. The registry is
+	// shared with repo.Service (constructed below) so the engine, pollers,
+	// and server handlers observe the same in-memory set.
+	repoReg := repo.NewRegistry()
+	for name, rc := range cfg.Repos {
+		ts, err := task.New(ctx, cfg.TaskPlugin, task.RepoSpec{
+			Name:       name,
+			Path:       rc.Path,
+			RootConfig: cfg.TaskConfig,
+			RepoConfig: rc.TaskConfig,
+		})
+		if err != nil {
+			return fmt.Errorf("repo %q task system: %w", name, err)
+		}
+		repoReg.Replace(&repo.Repo{
+			Name:       name,
+			Path:       rc.Path,
+			TaskConfig: rc.TaskConfig,
+			TaskSystem: ts,
+		})
+	}
+
+	// Load workflow files.
+	store := &workflow.Store{Dir: p.Workflows}
+	workflowList, err := store.LoadAll()
+	if err != nil {
+		return fmt.Errorf("load workflows: %w", err)
+	}
+
+	// Validate each workflow structurally (lifecycle nodes, routes,
+	// reachability, nudge templates) before any per-repo validation or
+	// binding; an invalid stored workflow must fail startup, not bind.
+	for _, wf := range workflowList {
+		if err := wf.Validate(); err != nil {
+			return fmt.Errorf("workflow %q: %w", wf.Name, err)
+		}
+	}
+
+	// Validate each workflow against every referenced repo task system.
+	for _, wf := range workflowList {
+		nodeCfgs := map[string]config.RawValues{}
+		for nodeName, n := range wf.Nodes {
+			nodeCfgs[nodeName] = n.TaskConfig
+		}
+		for _, repoName := range wf.Repos {
+			rp, ok := repoReg.Get(repoName)
+			if !ok {
+				return fmt.Errorf("workflow %q references unregistered repo %q", wf.Name, repoName)
+			}
+			if err := rp.TaskSystem.ValidateConfig(ctx, wf.TaskConfig, nodeCfgs); err != nil {
+				return fmt.Errorf("workflow %q repo %q: %w", wf.Name, repoName, err)
+			}
+		}
+	}
+
+	// Bind workflows and compiled matchers to repos.
+	if err := repoReg.BindWorkflows(workflowList); err != nil {
+		return err
+	}
+
+	// 5.5 normal-start refusal: serve REQUIRES an initialized database
+	// (init creates it); --recover is the explicit opt-out.
+	if !recover {
+		if _, err := os.Stat(p.Database); err != nil {
+			if os.IsNotExist(err) {
+				return fmt.Errorf("database %s is missing; run `relay-flow init` first, or `relay-flow serve --recover` to rebuild from the task system", p.Database)
+			}
+			return fmt.Errorf("stat database %s: %w", p.Database, err)
+		}
+	}
+
+	// 5.6 (decision 22): --recover replaces the execution state BEFORE any
+	// worker starts, with two explicit guarantees:
+	//   (a) The flock acquired at the top of serveRoot proves no other
+	//       relay-flow process has this database open — relay-flow is the
+	//       only process that ever opens state.db, and we hold server.lock.
+	//   (b) The old database is moved aside as a timestamped sibling (never
+	//       destroyed in place) and the destination path is verified absent
+	//       before goworkflows.New runs. Engine.New therefore opens a
+	//       provably empty database, and Engine.Start has no prior history
+	//       to resume.
+	if recover {
+		stamp := time.Now().UTC().Format("20060102T150405Z")
+		for _, suffix := range []string{"", "-wal", "-shm"} {
+			src := p.Database + suffix
+			if _, err := os.Stat(src); os.IsNotExist(err) {
+				continue
+			}
+			dst := fmt.Sprintf("%s.recover-%s.bak%s", p.Database, stamp, suffix)
+			if err := os.Rename(src, dst); err != nil {
+				return fmt.Errorf("preserve stale database %s as %s: %w", src, dst, err)
+			}
+		}
+		// Provably-empty engine: the primary DB path MUST NOT exist before
+		// engine construction. If it does, the rename-aside above failed
+		// silently for an unexpected reason; refuse to start rather than
+		// resume stale state.
+		if _, err := os.Stat(p.Database); err == nil {
+			return fmt.Errorf("recover: %s still exists after rename-aside; refusing to open possibly-stale database", p.Database)
+		} else if !os.IsNotExist(err) {
+			return fmt.Errorf("recover: stat %s: %w", p.Database, err)
+		}
+	}
+
+	// Open the go-workflows SQLite engine (migrates relay_runs; does not
+	// yet start workers).
+	engine, err := goworkflows.New(p.Database, goworkflows.Dependencies{
+		Repos:         repoReg,
+		Runner:        rnr,
+		Harness:       hrn,
+		RetentionDays: cfg.CompletedRunRetentionDays,
+		Runtime: &runsvc.RuntimePolicy{
+			KeepTerminalsAlive: cfg.KeepTerminalsAlive,
+			KeepSessionsAlive:  cfg.KeepSessionsAlive,
+		},
+	})
+	if err != nil {
+		return fmt.Errorf("open engine: %w", err)
+	}
+
+	// 5.2 fail-fast preflight: before workers/pollers start, validate
+	// task-system, runner, and harness credentials/permissions/connectivity,
+	// every registered repo, and every configured agent. Known permanent
+	// errors abort startup; runtime failures of existing runs retry forever
+	// (design.md decision 16).
+	if err := preflight(ctx, repoReg, rnr, hrn, workflowList); err != nil {
+		shutdownCtx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+		_ = engine.Shutdown(shutdownCtx)
+		cancel()
+		return fmt.Errorf("startup validation: %w", err)
+	}
+
+	// Start engine workers and the one-shot startup retention sweep
+	// (3.25; runs once here, no background ticker).
+	if err := engine.Start(ctx); err != nil {
+		shutdownCtx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+		_ = engine.Shutdown(shutdownCtx)
+		cancel()
+		return fmt.Errorf("start engine: %w", err)
+	}
+
+	// 5.4 lifecycle gate: ONE sync.Mutex shared by workflow.Service
+	// Submit/Remove and RunManager.EnsureRun (design.md decision 23). Plain
+	// mutex in the wiring — no lock service, no new abstraction.
+	lifecycleGate := &sync.Mutex{}
+
+	// Construct the Run Manager.
+	runManager := &runsvc.RunManager{Executor: engine, Runs: engine, Gate: lifecycleGate}
+
+	// Services consumed by the Unix-socket server. The repo.Service takes
+	// the SAME registry by pointer (via replaceInternal), so the engine,
+	// pollers, and handlers observe one in-memory repo set.
+	wfSvc := workflow.NewService(store, engine, repoExists{repoReg})
+	wfSvc.Gate = lifecycleGate
+	wfSvc.ValidateTaskConfig = workflowConfigValidator(repoReg)
+	// Submit/Remove must also rebuild repo bindings under the gate (spec
+	// 3.34: bindings rebuilt on submit/remove/startup). Wired as a plain
+	// callback — no dispatcher.
+	wfSvc.Rebind = func() error {
+		return repoReg.BindWorkflows(wfSvc.Registry().List())
+	}
+	for _, wf := range workflowList {
+		wfSvc.Registry().Replace(wf)
+	}
+	repoSvc := repo.NewServiceWithRegistry(repo.ServiceConfig{
+		ConfigPath: p.Config,
+		TaskPlugin: cfg.TaskPlugin,
+		Runner:     rnr,
+		Active:     engine,
+		Workflows:  wfSvc.Registry(),
+	}, repoReg)
+
+	// 5.6 database-loss recovery (explicit --recover only; never inferred).
+	// Runs after the engine is open and BEFORE normal pollers start so the
+	// recovered runs exist before any poll cycle observes them. Mailbox
+	// specs come from the engine so the recover path builds the same
+	// description content as normal run execution.
+	if recover {
+		specsFor := func(wf *workflow.Workflow, ticketKey string) []task.MailboxSpec {
+			return goworkflows.MailboxSpecs(wf, ticketKey)
+		}
+		if err := recoverpkg.FromTaskSystem(ctx, repoReg, rnr, runManager, specsFor); err != nil {
+			shutdownCtx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+			_ = engine.Shutdown(shutdownCtx)
+			cancel()
+			return fmt.Errorf("recover: %w", err)
+		}
+	}
+
+	// Start the Repo Poller group with the handleBatch callback (5.3,
+	// docs/structs-methods-interfaces.md lines 1099-1117 verbatim). No
+	// dispatcher type, no extra abstraction.
+	pollers := repo.NewPollerGroup(10, handleBatch(runManager))
+	pollers.Interval = time.Duration(cfg.PollIntervalSeconds) * time.Second
+	pollers.ReplaceRepos(repoSvc.Registry().List())
+	pollerCtx, cancelPollers := context.WithCancel(context.Background())
+	pollersDone := make(chan struct{})
+	go func() {
+		pollers.Run(pollerCtx)
+		close(pollersDone)
+	}()
+
+	// cleanup is the single bounded shutdown path (5.5). Order:
+	//   1. Stop accepting new polls AND new HTTP requests immediately.
+	//   2. Wait (within one 30s budget) for in-flight polls and in-flight
+	//      HTTP calls to return.
+	//   3. Cancel worker polling and close SQLite via engine.Shutdown.
+	// Every termination path — signal, /stop, listener error, startup
+	// failure after pollers launched — funnels through here so no path
+	// closes the database under running handlers or skips poller join.
+	cleanup := func(httpServer *http.Server, serveErr chan error, serveConsumed bool) {
+		shutdownCtx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+		defer cancel()
+		// Stop new work immediately: no new polls, no new HTTP requests.
+		cancelPollers()
+		if httpServer != nil {
+			_ = httpServer.Shutdown(shutdownCtx)
+		}
+		// Wait for in-flight HTTP within the budget.
+		if httpServer != nil && !serveConsumed {
+			select {
+			case <-serveErr:
+			case <-shutdownCtx.Done():
+			}
+		}
+		// Wait for in-flight pollers within the same 30s budget (spec:
+		// "waits up to 30 seconds ... then close the socket and database").
+		// PollerGroup.Run returns once every in-flight handleBatch handler
+		// completes; on deadline we proceed to engine.Shutdown anyway so
+		// shutdown is bounded — the spec's rule is that already-running
+		// external work is not interruptible and durable state must remain
+		// recoverable (engine history is authoritative; a handler that
+		// survives past deadline hits a closed-DB error and logs it).
+		select {
+		case <-pollersDone:
+		case <-shutdownCtx.Done():
+		}
+		// Cancel worker polling and close the database. Engine.Shutdown
+		// uses the shared budget for its own bounded wait on active
+		// workflow/activity tasks.
+		_ = engine.Shutdown(shutdownCtx)
+	}
+
+	// Start the Unix-socket server.
+	_ = os.Remove(p.Socket)
+	listener, err := net.Listen("unix", p.Socket)
+	if err != nil {
+		cleanup(nil, nil, true)
+		return fmt.Errorf("listen %s: %w", p.Socket, err)
+	}
+	if err := os.Chmod(p.Socket, 0o600); err != nil {
+		listener.Close()
+		_ = os.Remove(p.Socket) // leave no stale socket file on failure
+		cleanup(nil, nil, true)
+		return fmt.Errorf("chmod %s: %w", p.Socket, err)
+	}
+
+	// /stop and ctx cancellation funnel into stopCtx; both end at the
+	// single cleanup path.
+	stopCtx, stopServe := context.WithCancel(context.Background())
+	defer stopServe()
+
+	deps := &serveDeps{
+		wf:         wfSvc,
+		repos:      repoSvc,
+		engine:     engine,
+		runManager: runManager,
+		// Repo register/remove must reach the running poller group — the
+		// startup ReplaceRepos only covers repos present at boot. Re-sync
+		// after every successful mutation so a newly registered repo is
+		// polled on the next tick without a serve restart.
+		onReposChanged: func() {
+			pollers.ReplaceRepos(repoSvc.Registry().List())
+		},
+		shutdown: func(context.Context) error {
+			stopServe()
+			return nil
+		},
+	}
+	httpServer := &http.Server{Handler: server.New(deps)}
+
+	serveErr := make(chan error, 1)
+	go func() { serveErr <- httpServer.Serve(listener) }()
+
+	var serveResult error
+	serveConsumed := false
+	select {
+	case <-ctx.Done():
+	case <-stopCtx.Done():
+	case err := <-serveErr:
+		serveConsumed = true
+		if err != nil && !errors.Is(err, http.ErrServerClosed) {
+			serveResult = fmt.Errorf("serve %s: %w", p.Socket, err)
+		}
+		// Listener closed on its own — treat as normal shutdown request.
+	}
+
+	cleanup(httpServer, serveErr, serveConsumed)
+	return serveResult
+}
+
+func workflowConfigValidator(repoReg *repo.Registry) func(context.Context, *workflow.Workflow) error {
+	return func(ctx context.Context, wf *workflow.Workflow) error {
+		nodeCfgs := map[string]config.RawValues{}
+		for nodeName, n := range wf.Nodes {
+			nodeCfgs[nodeName] = n.TaskConfig
+		}
+		for _, repoName := range wf.Repos {
+			rp, ok := repoReg.Get(repoName)
+			if !ok {
+				return fmt.Errorf("workflow %q references unregistered repo %q", wf.Name, repoName)
+			}
+			if err := rp.TaskSystem.ValidateConfig(ctx, wf.TaskConfig, nodeCfgs); err != nil {
+				return fmt.Errorf("workflow %q repo %q: %w", wf.Name, repoName, err)
+			}
+		}
+		return nil
+	}
+}
+
+// repoExists adapts *repo.Registry to workflow.RepoLookup (Exists).
+type repoExists struct{ reg *repo.Registry }
+
+func (r repoExists) Exists(name string) bool {
+	_, ok := r.reg.Get(name)
+	return ok
+}
+
+// serveDeps adapts composition-root services to server.Deps. Thin forwarder;
+// no logic. Signatures match docs/structs-methods-interfaces.md Client.
+type serveDeps struct {
+	wf             *workflow.Service
+	repos          *repo.Service
+	engine         *goworkflows.Engine
+	runManager     *runsvc.RunManager
+	onReposChanged func()
+	shutdown       func(context.Context) error
+}
+
+func (d *serveDeps) SubmitWorkflow(ctx context.Context, yaml []byte) (*workflow.Workflow, error) {
+	return d.wf.Submit(ctx, yaml)
+}
+func (d *serveDeps) GetWorkflow(_ context.Context, name string) (*workflow.Workflow, error) {
+	return d.wf.Get(name)
+}
+func (d *serveDeps) ListWorkflows(context.Context) ([]*workflow.Workflow, error) {
+	return d.wf.List(), nil
+}
+func (d *serveDeps) RemoveWorkflow(ctx context.Context, name string) error {
+	return d.wf.Remove(ctx, name)
+}
+
+func (d *serveDeps) ListRuns(ctx context.Context, filter runsvc.Filter) ([]runsvc.Run, error) {
+	return d.engine.ListRuns(ctx, filter)
+}
+func (d *serveDeps) GetRunByTicket(ctx context.Context, ticket string) (runsvc.Run, error) {
+	return d.engine.FindRunByTicket(ctx, ticket)
+}
+func (d *serveDeps) CancelRun(ctx context.Context, ticket, reason string) error {
+	return d.runManager.CancelByTicket(ctx, ticket, reason)
+}
+
+func (d *serveDeps) SubmitReport(ctx context.Context, rep runsvc.ReportRequest) (runsvc.ReportAck, error) {
+	return d.engine.SubmitReport(ctx, rep)
+}
+func (d *serveDeps) HasProcessedReport(ctx context.Context, id runsvc.ID, reportID string) (bool, error) {
+	return d.engine.HasProcessedReport(ctx, id, reportID)
+}
+func (d *serveDeps) RegisterNodeSession(ctx context.Context, registration runsvc.NodeRuntimeRegistration) (runsvc.NodeRuntimeRegistrationAck, error) {
+	return d.engine.RegisterNodeSession(ctx, registration)
+}
+
+func (d *serveDeps) DiscoverRepos(ctx context.Context) ([]runner.RepoCandidate, error) {
+	return d.repos.Discover(ctx)
+}
+func (d *serveDeps) TaskFields(context.Context) ([]string, error) {
+	return d.repos.RequiredRepoKeys(), nil
+}
+func (d *serveDeps) RegisterRepo(ctx context.Context, input repo.RegisterInput) (repo.Info, error) {
+	info, err := d.repos.Register(ctx, input)
+	if err == nil && d.onReposChanged != nil {
+		d.onReposChanged()
+	}
+	return info, err
+}
+func (d *serveDeps) ListRepos(context.Context) ([]repo.Info, error) {
+	return d.repos.List(), nil
+}
+func (d *serveDeps) GetRepo(_ context.Context, name string) (repo.Info, error) {
+	return d.repos.Get(name)
+}
+func (d *serveDeps) RemoveRepo(ctx context.Context, name string) error {
+	err := d.repos.Remove(ctx, name)
+	if err == nil && d.onReposChanged != nil {
+		d.onReposChanged()
+	}
+	return err
+}
+
+func (d *serveDeps) Shutdown(ctx context.Context) error { return d.shutdown(ctx) }
+
+// handleBatch is the Repo Poller group callback (5.3). Implemented verbatim
+// from docs/structs-methods-interfaces.md lines 1101-1117: per ticket,
+// ResolveWorkflow; ErrNoMatch → continue; other routing errors → log and
+// continue (mutate nothing); success → RunManager.EnsureRun, log its errors.
+// No dispatcher type, no extra abstraction. A small closure captures the
+// RunManager so the repo.BatchHandler signature is satisfied.
+//
+// 9.2 poller logging: one info line per repo poll cycle (batch size), one
+// info line per ticket with the routing outcome, one debug line per ticket
+// with normalized fields. EnsureRun emits its own outcome lines from
+// internal/run/manager.go.
+func handleBatch(runManager *runsvc.RunManager) repo.BatchHandler {
+	return func(ctx context.Context, rp *repo.Repo, tickets []task.Ticket) {
+		slog.Info("poll cycle", "repo", rp.Name, "batch", len(tickets))
+		for _, ticket := range tickets {
+			slog.Debug("poll ticket",
+				"repo", rp.Name, "ticket", ticket.Key, "id", ticket.ID,
+				"title", ticket.Title, "claims", strings.Join(ticket.WorkflowClaims, ","))
+			wf, err := router.ResolveWorkflow(rp, ticket)
+			switch {
+			case errors.Is(err, router.ErrNoMatch):
+				slog.Info("route outcome",
+					"repo", rp.Name, "ticket", ticket.Key, "outcome", "no-match")
+				continue
+			case err != nil:
+				var amb *router.AmbiguousError
+				var inv *router.InvalidClaimError
+				outcome := "error"
+				switch {
+				case errors.As(err, &amb):
+					outcome = "ambiguous"
+				case errors.As(err, &inv):
+					outcome = "invalid-claim"
+				}
+				slog.Info("route outcome",
+					"repo", rp.Name, "ticket", ticket.Key, "outcome", outcome, "error", err)
+				continue
+			}
+			routeOutcome := "claimed"
+			for _, c := range ticket.WorkflowClaims {
+				if c == "wf:"+wf.Name {
+					routeOutcome = "already-claimed"
+					break
+				}
+			}
+			slog.Info("route outcome",
+				"repo", rp.Name, "ticket", ticket.Key, "workflow", wf.Name, "outcome", routeOutcome)
+			if err := runManager.EnsureRun(ctx, rp, wf, ticket); err != nil {
+				// EnsureRun already logged the outcome; nothing to add here.
+				_ = err
+			}
+		}
+	}
+}
+
+// preflight is the 5.2 fail-fast startup validation. It runs after the
+// engine is open and BEFORE workers/pollers start. Permanent errors abort
+// startup; runtime errors of existing runs retry forever (different rule).
+//
+// Probes are no-side-effect reads on each adapter boundary:
+//   - runner: ValidateRepo per registered repo (Orca connectivity + repo
+//     presence at the configured path).
+//   - task system: Poll per repo (credentials/permissions/connectivity;
+//     returns active parents only, mutates nothing).
+//   - harness: ValidateAgent per (repo, workflow) agent node (configured
+//     agent must exist for that repo).
+//
+// Repos with no workflows still validate runner + task connectivity; agent
+// validation runs once per distinct (repoPath, agent) pair to avoid
+// redundant CLI calls when several workflows share an agent on one repo.
+func preflight(ctx context.Context, repoReg *repo.Registry, rnr runner.Runner, hrn harness.Harness, workflows []*workflow.Workflow) error {
+	for _, rp := range repoReg.List() {
+		if err := rnr.ValidateRepo(ctx, rp.Name, rp.Path); err != nil {
+			return fmt.Errorf("repo %q runner: %w", rp.Name, err)
+		}
+		if _, err := rp.TaskSystem.Poll(ctx); err != nil {
+			return fmt.Errorf("repo %q task system: %w", rp.Name, err)
+		}
+	}
+	type agentProbe struct {
+		repoPath string
+		agent    string
+	}
+	seen := map[agentProbe]bool{}
+	for _, wf := range workflows {
+		for nodeName, n := range wf.Nodes {
+			// Validate EVERY configured agent (agent and HITL work nodes
+			// both declare one); skip start/end which never carry Agent.
+			if (n.Type != workflow.NodeAgent && n.Type != workflow.NodeHITL) || n.Agent == "" {
+				continue
+			}
+			for _, repoName := range wf.Repos {
+				rp, ok := repoReg.Get(repoName)
+				if !ok {
+					return fmt.Errorf("workflow %q references unregistered repo %q", wf.Name, repoName)
+				}
+				key := agentProbe{repoPath: rp.Path, agent: n.Agent}
+				if seen[key] {
+					continue
+				}
+				seen[key] = true
+				if err := hrn.ValidateAgent(ctx, rp.Path, n.Agent); err != nil {
+					return fmt.Errorf("workflow %q node %q repo %q: %w", wf.Name, nodeName, repoName, err)
+				}
+			}
+		}
+	}
+	return nil
+}
+
+// The recover composition lives in internal/recover (recoverpkg.FromTaskSystem)
+// so it is importable by both serve and the recover test — see task 6.3.

package/go.mod

@@ -1,5 +1,72 @@
 module github.com/rajpopat27/relay-flow
 
-go 1.21.11
+go 1.24.6
 
-require gopkg.in/yaml.v3 v3.0.1 // indirect
+require (
+	github.com/cschleiden/go-workflows v1.4.2
+	github.com/google/renameio/v2 v2.0.1
+	github.com/google/uuid v1.6.0
+	gopkg.in/yaml.v3 v3.0.1
+)
+
+require (
+	github.com/atotto/clipboard v0.1.4 // indirect
+	github.com/aymanbagabas/go-osc52/v2 v2.0.1 // indirect
+	github.com/benbjohnson/clock v1.3.0 // indirect
+	github.com/catppuccin/go v0.3.0 // indirect
+	github.com/cenkalti/backoff/v4 v4.3.0 // indirect
+	github.com/charmbracelet/bubbles v0.21.1-0.20250623103423-23b8fd6302d7 // indirect
+	github.com/charmbracelet/bubbletea v1.3.6 // indirect
+	github.com/charmbracelet/colorprofile v0.2.3-0.20250311203215-f60798e515dc // indirect
+	github.com/charmbracelet/huh v1.0.0 // indirect
+	github.com/charmbracelet/lipgloss v1.1.0 // indirect
+	github.com/charmbracelet/x/ansi v0.9.3 // indirect
+	github.com/charmbracelet/x/cellbuf v0.0.13 // indirect
+	github.com/charmbracelet/x/exp/strings v0.0.0-20240722160745-212f7b056ed0 // indirect
+	github.com/charmbracelet/x/term v0.2.1 // indirect
+	github.com/davecgh/go-spew v1.1.1 // indirect
+	github.com/dustin/go-humanize v1.0.1 // indirect
+	github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f // indirect
+	github.com/go-errors/errors v1.4.2 // indirect
+	github.com/go-logr/logr v1.4.2 // indirect
+	github.com/go-logr/stdr v1.2.2 // indirect
+	github.com/golang-migrate/migrate/v4 v4.16.2 // indirect
+	github.com/hashicorp/errwrap v1.1.0 // indirect
+	github.com/hashicorp/go-multierror v1.1.1 // indirect
+	github.com/jellydator/ttlcache/v3 v3.0.0 // indirect
+	github.com/kballard/go-shellquote v0.0.0-20180428030007-95032a82bc51 // indirect
+	github.com/lucasb-eyer/go-colorful v1.2.0 // indirect
+	github.com/mattn/go-isatty v0.0.20 // indirect
+	github.com/mattn/go-localereader v0.0.1 // indirect
+	github.com/mattn/go-runewidth v0.0.16 // indirect
+	github.com/mitchellh/hashstructure/v2 v2.0.2 // indirect
+	github.com/muesli/ansi v0.0.0-20230316100256-276c6243b2f6 // indirect
+	github.com/muesli/cancelreader v0.2.2 // indirect
+	github.com/muesli/termenv v0.16.0 // indirect
+	github.com/pmezard/go-difflib v1.0.0 // indirect
+	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
+	github.com/rivo/uniseg v0.4.7 // indirect
+	github.com/stretchr/objx v0.5.2 // indirect
+	github.com/stretchr/testify v1.10.0 // indirect
+	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
+	go.opentelemetry.io/otel v1.31.0 // indirect
+	go.opentelemetry.io/otel/metric v1.31.0 // indirect
+	go.opentelemetry.io/otel/sdk v1.31.0 // indirect
+	go.opentelemetry.io/otel/trace v1.31.0 // indirect
+	go.uber.org/atomic v1.11.0 // indirect
+	golang.org/x/mod v0.28.0 // indirect
+	golang.org/x/sync v0.17.0 // indirect
+	golang.org/x/sys v0.36.0 // indirect
+	golang.org/x/text v0.29.0 // indirect
+	golang.org/x/tools v0.37.0 // indirect
+	lukechampine.com/uint128 v1.2.0 // indirect
+	modernc.org/cc/v3 v3.40.0 // indirect
+	modernc.org/ccgo/v3 v3.16.13 // indirect
+	modernc.org/libc v1.29.0 // indirect
+	modernc.org/mathutil v1.6.0 // indirect
+	modernc.org/memory v1.7.2 // indirect
+	modernc.org/opt v0.1.3 // indirect
+	modernc.org/sqlite v1.27.0 // indirect
+	modernc.org/strutil v1.1.3 // indirect
+	modernc.org/token v1.0.1 // indirect
+)