agent/agent.go

package agent

import (
	"bytes"
	"context"
	"encoding/binary"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"net"
	"net/http"
	"net/netip"
	"os"
	"os/user"
	"path/filepath"
	"runtime"
	"runtime/debug"
	"sort"
	"strconv"
	"strings"
	"sync"
	"time"

	"github.com/go-chi/chi/v5"
	"github.com/google/uuid"
	"github.com/prometheus/client_golang/prometheus"
	"github.com/prometheus/common/expfmt"
	"github.com/spf13/afero"
	"go.uber.org/atomic"
	"golang.org/x/exp/slices"
	"golang.org/x/sync/errgroup"
	"golang.org/x/xerrors"
	"storj.io/drpc"
	"tailscale.com/net/speedtest"
	"tailscale.com/tailcfg"
	"tailscale.com/types/netlogtype"
	"tailscale.com/util/clientmetric"

	"cdr.dev/slog"
	"github.com/coder/coder/v2/agent/agentproc"
	"github.com/coder/coder/v2/agent/agentscripts"
	"github.com/coder/coder/v2/agent/agentssh"
	"github.com/coder/coder/v2/agent/proto"
	"github.com/coder/coder/v2/agent/reconnectingpty"
	"github.com/coder/coder/v2/buildinfo"
	"github.com/coder/coder/v2/cli/gitauth"
	"github.com/coder/coder/v2/coderd/database/dbtime"
	"github.com/coder/coder/v2/codersdk"
	"github.com/coder/coder/v2/codersdk/agentsdk"
	"github.com/coder/coder/v2/codersdk/workspacesdk"
	"github.com/coder/coder/v2/tailnet"
	tailnetproto "github.com/coder/coder/v2/tailnet/proto"
	"github.com/coder/retry"
)

const (
	ProtocolReconnectingPTY = "reconnecting-pty"
	ProtocolSSH             = "ssh"
	ProtocolDial            = "dial"
)

// EnvProcPrioMgmt determines whether we attempt to manage
// process CPU and OOM Killer priority.
const (
	EnvProcPrioMgmt = "CODER_PROC_PRIO_MGMT"
	EnvProcOOMScore = "CODER_PROC_OOM_SCORE"
)

type Options struct {
	Filesystem                   afero.Fs
	LogDir                       string
	TempDir                      string
	ScriptDataDir                string
	ExchangeToken                func(ctx context.Context) (string, error)
	Client                       Client
	ReconnectingPTYTimeout       time.Duration
	EnvironmentVariables         map[string]string
	Logger                       slog.Logger
	IgnorePorts                  map[int]string
	PortCacheDuration            time.Duration
	SSHMaxTimeout                time.Duration
	TailnetListenPort            uint16
	Subsystems                   []codersdk.AgentSubsystem
	Addresses                    []netip.Prefix
	PrometheusRegistry           *prometheus.Registry
	ReportMetadataInterval       time.Duration
	ServiceBannerRefreshInterval time.Duration
	Syscaller                    agentproc.Syscaller
	// ModifiedProcesses is used for testing process priority management.
	ModifiedProcesses chan []*agentproc.Process
	// ProcessManagementTick is used for testing process priority management.
	ProcessManagementTick <-chan time.Time
}

type Client interface {
	ConnectRPC(ctx context.Context) (drpc.Conn, error)
	RewriteDERPMap(derpMap *tailcfg.DERPMap)
}

type Agent interface {
	HTTPDebug() http.Handler
	// TailnetConn may be nil.
	TailnetConn() *tailnet.Conn
	io.Closer
}

func New(options Options) Agent {
	if options.Filesystem == nil {
		options.Filesystem = afero.NewOsFs()
	}
	if options.TempDir == "" {
		options.TempDir = os.TempDir()
	}
	if options.LogDir == "" {
		if options.TempDir != os.TempDir() {
			options.Logger.Debug(context.Background(), "log dir not set, using temp dir", slog.F("temp_dir", options.TempDir))
		} else {
			options.Logger.Debug(context.Background(), "using log dir", slog.F("log_dir", options.LogDir))
		}
		options.LogDir = options.TempDir
	}
	if options.ScriptDataDir == "" {
		if options.TempDir != os.TempDir() {
			options.Logger.Debug(context.Background(), "script data dir not set, using temp dir", slog.F("temp_dir", options.TempDir))
		} else {
			options.Logger.Debug(context.Background(), "using script data dir", slog.F("script_data_dir", options.ScriptDataDir))
		}
		options.ScriptDataDir = options.TempDir
	}
	if options.ExchangeToken == nil {
		options.ExchangeToken = func(ctx context.Context) (string, error) {
			return "", nil
		}
	}
	if options.ReportMetadataInterval == 0 {
		options.ReportMetadataInterval = time.Second
	}
	if options.ServiceBannerRefreshInterval == 0 {
		options.ServiceBannerRefreshInterval = 2 * time.Minute
	}
	if options.PortCacheDuration == 0 {
		options.PortCacheDuration = 1 * time.Second
	}

	prometheusRegistry := options.PrometheusRegistry
	if prometheusRegistry == nil {
		prometheusRegistry = prometheus.NewRegistry()
	}

	if options.Syscaller == nil {
		options.Syscaller = agentproc.NewSyscaller()
	}

	hardCtx, hardCancel := context.WithCancel(context.Background())
	gracefulCtx, gracefulCancel := context.WithCancel(hardCtx)
	a := &agent{
		tailnetListenPort:                  options.TailnetListenPort,
		reconnectingPTYTimeout:             options.ReconnectingPTYTimeout,
		logger:                             options.Logger,
		gracefulCtx:                        gracefulCtx,
		gracefulCancel:                     gracefulCancel,
		hardCtx:                            hardCtx,
		hardCancel:                         hardCancel,
		coordDisconnected:                  make(chan struct{}),
		environmentVariables:               options.EnvironmentVariables,
		client:                             options.Client,
		exchangeToken:                      options.ExchangeToken,
		filesystem:                         options.Filesystem,
		logDir:                             options.LogDir,
		tempDir:                            options.TempDir,
		scriptDataDir:                      options.ScriptDataDir,
		lifecycleUpdate:                    make(chan struct{}, 1),
		lifecycleReported:                  make(chan codersdk.WorkspaceAgentLifecycle, 1),
		lifecycleStates:                    []agentsdk.PostLifecycleRequest{{State: codersdk.WorkspaceAgentLifecycleCreated}},
		ignorePorts:                        options.IgnorePorts,
		portCacheDuration:                  options.PortCacheDuration,
		reportMetadataInterval:             options.ReportMetadataInterval,
		announcementBannersRefreshInterval: options.ServiceBannerRefreshInterval,
		sshMaxTimeout:                      options.SSHMaxTimeout,
		subsystems:                         options.Subsystems,
		addresses:                          options.Addresses,
		syscaller:                          options.Syscaller,
		modifiedProcs:                      options.ModifiedProcesses,
		processManagementTick:              options.ProcessManagementTick,
		logSender:                          agentsdk.NewLogSender(options.Logger),

		prometheusRegistry: prometheusRegistry,
		metrics:            newAgentMetrics(prometheusRegistry),
	}
	// Initially, we have a closed channel, reflecting the fact that we are not initially connected.
	// Each time we connect we replace the channel (while holding the closeMutex) with a new one
	// that gets closed on disconnection.  This is used to wait for graceful disconnection from the
	// coordinator during shut down.
	close(a.coordDisconnected)
	a.announcementBanners.Store(new([]codersdk.BannerConfig))
	a.sessionToken.Store(new(string))
	a.init()
	return a
}

type agent struct {
	logger            slog.Logger
	client            Client
	exchangeToken     func(ctx context.Context) (string, error)
	tailnetListenPort uint16
	filesystem        afero.Fs
	logDir            string
	tempDir           string
	scriptDataDir     string
	// ignorePorts tells the api handler which ports to ignore when
	// listing all listening ports. This is helpful to hide ports that
	// are used by the agent, that the user does not care about.
	ignorePorts       map[int]string
	portCacheDuration time.Duration
	subsystems        []codersdk.AgentSubsystem

	reconnectingPTYs       sync.Map
	reconnectingPTYTimeout time.Duration

	// we track 2 contexts and associated cancel functions: "graceful" which is Done when it is time
	// to start gracefully shutting down and "hard" which is Done when it is time to close
	// everything down (regardless of whether graceful shutdown completed).
	gracefulCtx       context.Context
	gracefulCancel    context.CancelFunc
	hardCtx           context.Context
	hardCancel        context.CancelFunc
	closeWaitGroup    sync.WaitGroup
	closeMutex        sync.Mutex
	coordDisconnected chan struct{}

	environmentVariables map[string]string

	manifest                           atomic.Pointer[agentsdk.Manifest] // manifest is atomic because values can change after reconnection.
	reportMetadataInterval             time.Duration
	scriptRunner                       *agentscripts.Runner
	announcementBanners                atomic.Pointer[[]codersdk.BannerConfig] // announcementBanners is atomic because it is periodically updated.
	announcementBannersRefreshInterval time.Duration
	sessionToken                       atomic.Pointer[string]
	sshServer                          *agentssh.Server
	sshMaxTimeout                      time.Duration

	lifecycleUpdate            chan struct{}
	lifecycleReported          chan codersdk.WorkspaceAgentLifecycle
	lifecycleMu                sync.RWMutex // Protects following.
	lifecycleStates            []agentsdk.PostLifecycleRequest
	lifecycleLastReportedIndex int // Keeps track of the last lifecycle state we successfully reported.

	network       *tailnet.Conn
	addresses     []netip.Prefix
	statsReporter *statsReporter
	logSender     *agentsdk.LogSender

	connCountReconnectingPTY atomic.Int64

	prometheusRegistry *prometheus.Registry
	// metrics are prometheus registered metrics that will be collected and
	// labeled in Coder with the agent + workspace.
	metrics   *agentMetrics
	syscaller agentproc.Syscaller

	// modifiedProcs is used for testing process priority management.
	modifiedProcs chan []*agentproc.Process
	// processManagementTick is used for testing process priority management.
	processManagementTick <-chan time.Time
}

func (a *agent) TailnetConn() *tailnet.Conn {
	return a.network
}

func (a *agent) init() {
	// pass the "hard" context because we explicitly close the SSH server as part of graceful shutdown.
	sshSrv, err := agentssh.NewServer(a.hardCtx, a.logger.Named("ssh-server"), a.prometheusRegistry, a.filesystem, &agentssh.Config{
		MaxTimeout:          a.sshMaxTimeout,
		MOTDFile:            func() string { return a.manifest.Load().MOTDFile },
		AnnouncementBanners: func() *[]codersdk.BannerConfig { return a.announcementBanners.Load() },
		UpdateEnv:           a.updateCommandEnv,
		WorkingDirectory:    func() string { return a.manifest.Load().Directory },
	})
	if err != nil {
		panic(err)
	}
	a.sshServer = sshSrv
	a.scriptRunner = agentscripts.New(agentscripts.Options{
		LogDir:      a.logDir,
		DataDirBase: a.scriptDataDir,
		Logger:      a.logger,
		SSHServer:   sshSrv,
		Filesystem:  a.filesystem,
		GetScriptLogger: func(logSourceID uuid.UUID) agentscripts.ScriptLogger {
			return a.logSender.GetScriptLogger(logSourceID)
		},
	})
	// Register runner metrics. If the prom registry is nil, the metrics
	// will not report anywhere.
	a.scriptRunner.RegisterMetrics(a.prometheusRegistry)
	go a.runLoop()
}

// runLoop attempts to start the agent in a retry loop.
// Coder may be offline temporarily, a connection issue
// may be happening, but regardless after the intermittent
// failure, you'll want the agent to reconnect.
func (a *agent) runLoop() {
	go a.manageProcessPriorityUntilGracefulShutdown()

	// need to keep retrying up to the hardCtx so that we can send graceful shutdown-related
	// messages.
	ctx := a.hardCtx
	for retrier := retry.New(100*time.Millisecond, 10*time.Second); retrier.Wait(ctx); {
		a.logger.Info(ctx, "connecting to coderd")
		err := a.run()
		if err == nil {
			continue
		}
		if ctx.Err() != nil {
			// Context canceled errors may come from websocket pings, so we
			// don't want to use `errors.Is(err, context.Canceled)` here.
			return
		}
		if a.isClosed() {
			return
		}
		if errors.Is(err, io.EOF) {
			a.logger.Info(ctx, "disconnected from coderd")
			continue
		}
		a.logger.Warn(ctx, "run exited with error", slog.Error(err))
	}
}

func (a *agent) collectMetadata(ctx context.Context, md codersdk.WorkspaceAgentMetadataDescription, now time.Time) *codersdk.WorkspaceAgentMetadataResult {
	var out bytes.Buffer
	result := &codersdk.WorkspaceAgentMetadataResult{
		// CollectedAt is set here for testing purposes and overrode by
		// coderd to the time of server receipt to solve clock skew.
		//
		// In the future, the server may accept the timestamp from the agent
		// if it can guarantee the clocks are synchronized.
		CollectedAt: now,
	}
	cmdPty, err := a.sshServer.CreateCommand(ctx, md.Script, nil)
	if err != nil {
		result.Error = fmt.Sprintf("create cmd: %+v", err)
		return result
	}
	cmd := cmdPty.AsExec()

	cmd.Stdout = &out
	cmd.Stderr = &out
	cmd.Stdin = io.LimitReader(nil, 0)

	// We split up Start and Wait instead of calling Run so that we can return a more precise error.
	err = cmd.Start()
	if err != nil {
		result.Error = fmt.Sprintf("start cmd: %+v", err)
		return result
	}

	// This error isn't mutually exclusive with useful output.
	err = cmd.Wait()
	const bufLimit = 10 << 10
	if out.Len() > bufLimit {
		err = errors.Join(
			err,
			xerrors.Errorf("output truncated from %v to %v bytes", out.Len(), bufLimit),
		)
		out.Truncate(bufLimit)
	}

	// Important: if the command times out, we may see a misleading error like
	// "exit status 1", so it's important to include the context error.
	err = errors.Join(err, ctx.Err())

	if err != nil {
		result.Error = fmt.Sprintf("run cmd: %+v", err)
	}
	result.Value = out.String()
	return result
}

type metadataResultAndKey struct {
	result *codersdk.WorkspaceAgentMetadataResult
	key    string
}

type trySingleflight struct {
	mu sync.Mutex
	m  map[string]struct{}
}

func (t *trySingleflight) Do(key string, fn func()) {
	t.mu.Lock()
	_, ok := t.m[key]
	if ok {
		t.mu.Unlock()
		return
	}

	t.m[key] = struct{}{}
	t.mu.Unlock()
	defer func() {
		t.mu.Lock()
		delete(t.m, key)
		t.mu.Unlock()
	}()

	fn()
}

func (a *agent) reportMetadata(ctx context.Context, conn drpc.Conn) error {
	tickerDone := make(chan struct{})
	collectDone := make(chan struct{})
	ctx, cancel := context.WithCancel(ctx)
	defer func() {
		cancel()
		<-collectDone
		<-tickerDone
	}()

	var (
		logger          = a.logger.Named("metadata")
		report          = make(chan struct{}, 1)
		collect         = make(chan struct{}, 1)
		metadataResults = make(chan metadataResultAndKey, 1)
	)

	// Set up collect and report as a single ticker with two channels,
	// this is to allow collection and reporting to be triggered
	// independently of each other.
	go func() {
		t := time.NewTicker(a.reportMetadataInterval)
		defer func() {
			t.Stop()
			close(report)
			close(collect)
			close(tickerDone)
		}()
		wake := func(c chan<- struct{}) {
			select {
			case c <- struct{}{}:
			default:
			}
		}
		wake(collect) // Start immediately.

		for {
			select {
			case <-ctx.Done():
				return
			case <-t.C:
				wake(report)
				wake(collect)
			}
		}
	}()

	go func() {
		defer close(collectDone)

		var (
			// We use a custom singleflight that immediately returns if there is already
			// a goroutine running for a given key. This is to prevent a build-up of
			// goroutines waiting on Do when the script takes many multiples of
			// baseInterval to run.
			flight            = trySingleflight{m: map[string]struct{}{}}
			lastCollectedAtMu sync.RWMutex
			lastCollectedAts  = make(map[string]time.Time)
		)
		for {
			select {
			case <-ctx.Done():
				return
			case <-collect:
			}

			manifest := a.manifest.Load()
			if manifest == nil {
				continue
			}

			// If the manifest changes (e.g. on agent reconnect) we need to
			// purge old cache values to prevent lastCollectedAt from growing
			// boundlessly.
			lastCollectedAtMu.Lock()
			for key := range lastCollectedAts {
				if slices.IndexFunc(manifest.Metadata, func(md codersdk.WorkspaceAgentMetadataDescription) bool {
					return md.Key == key
				}) < 0 {
					logger.Debug(ctx, "deleting lastCollected key, missing from manifest",
						slog.F("key", key),
					)
					delete(lastCollectedAts, key)
				}
			}
			lastCollectedAtMu.Unlock()

			// Spawn a goroutine for each metadata collection, and use a
			// channel to synchronize the results and avoid both messy
			// mutex logic and overloading the API.
			for _, md := range manifest.Metadata {
				md := md
				// We send the result to the channel in the goroutine to avoid
				// sending the same result multiple times. So, we don't care about
				// the return values.
				go flight.Do(md.Key, func() {
					ctx := slog.With(ctx, slog.F("key", md.Key))
					lastCollectedAtMu.RLock()
					collectedAt, ok := lastCollectedAts[md.Key]
					lastCollectedAtMu.RUnlock()
					if ok {
						// If the interval is zero, we assume the user just wants
						// a single collection at startup, not a spinning loop.
						if md.Interval == 0 {
							return
						}
						intervalUnit := time.Second
						// reportMetadataInterval is only less than a second in tests,
						// so adjust the interval unit for them.
						if a.reportMetadataInterval < time.Second {
							intervalUnit = 100 * time.Millisecond
						}
						// The last collected value isn't quite stale yet, so we skip it.
						if collectedAt.Add(time.Duration(md.Interval) * intervalUnit).After(time.Now()) {
							return
						}
					}

					timeout := md.Timeout
					if timeout == 0 {
						if md.Interval != 0 {
							timeout = md.Interval
						} else if interval := int64(a.reportMetadataInterval.Seconds()); interval != 0 {
							// Fallback to the report interval
							timeout = interval * 3
						} else {
							// If the interval is still 0 (possible if the interval
							// is less than a second), default to 5. This was
							// randomly picked.
							timeout = 5
						}
					}
					ctxTimeout := time.Duration(timeout) * time.Second
					ctx, cancel := context.WithTimeout(ctx, ctxTimeout)
					defer cancel()

					now := time.Now()
					select {
					case <-ctx.Done():
						logger.Warn(ctx, "metadata collection timed out", slog.F("timeout", ctxTimeout))
					case metadataResults <- metadataResultAndKey{
						key:    md.Key,
						result: a.collectMetadata(ctx, md, now),
					}:
						lastCollectedAtMu.Lock()
						lastCollectedAts[md.Key] = now
						lastCollectedAtMu.Unlock()
					}
				})
			}
		}
	}()

	// Gather metadata updates and report them once every interval. If a
	// previous report is in flight, wait for it to complete before
	// sending a new one. If the network conditions are bad, we won't
	// benefit from canceling the previous send and starting a new one.
	var (
		updatedMetadata = make(map[string]*codersdk.WorkspaceAgentMetadataResult)
		reportTimeout   = 30 * time.Second
		reportError     = make(chan error, 1)
		reportInFlight  = false
		aAPI            = proto.NewDRPCAgentClient(conn)
	)

	for {
		select {
		case <-ctx.Done():
			return ctx.Err()
		case mr := <-metadataResults:
			// This can overwrite unsent values, but that's fine because
			// we're only interested about up-to-date values.
			updatedMetadata[mr.key] = mr.result
			continue
		case err := <-reportError:
			a.logger.Debug(ctx, "batch update metadata complete", slog.Error(err))
			if err != nil {
				return xerrors.Errorf("failed to report metadata: %w", err)
			}
			reportInFlight = false
		case <-report:
			if len(updatedMetadata) == 0 {
				continue
			}
			if reportInFlight {
				// If there's already a report in flight, don't send
				// another one, wait for next tick instead.
				a.logger.Debug(ctx, "skipped metadata report tick because report is in flight")
				continue
			}
			metadata := make([]*proto.Metadata, 0, len(updatedMetadata))
			for key, result := range updatedMetadata {
				pr := agentsdk.ProtoFromMetadataResult(*result)
				metadata = append(metadata, &proto.Metadata{
					Key:    key,
					Result: pr,
				})
				delete(updatedMetadata, key)
			}

			reportInFlight = true
			go func() {
				a.logger.Debug(ctx, "batch updating metadata")
				ctx, cancel := context.WithTimeout(ctx, reportTimeout)
				defer cancel()

				_, err := aAPI.BatchUpdateMetadata(ctx, &proto.BatchUpdateMetadataRequest{Metadata: metadata})
				reportError <- err
			}()
		}
	}
}

// reportLifecycle reports the current lifecycle state once. All state
// changes are reported in order.
func (a *agent) reportLifecycle(ctx context.Context, conn drpc.Conn) error {
	aAPI := proto.NewDRPCAgentClient(conn)
	for {
		select {
		case <-a.lifecycleUpdate:
		case <-ctx.Done():
			return ctx.Err()
		}

		for {
			a.lifecycleMu.RLock()
			lastIndex := len(a.lifecycleStates) - 1
			report := a.lifecycleStates[a.lifecycleLastReportedIndex]
			if len(a.lifecycleStates) > a.lifecycleLastReportedIndex+1 {
				report = a.lifecycleStates[a.lifecycleLastReportedIndex+1]
			}
			a.lifecycleMu.RUnlock()

			if lastIndex == a.lifecycleLastReportedIndex {
				break
			}
			l, err := agentsdk.ProtoFromLifecycle(report)
			if err != nil {
				a.logger.Critical(ctx, "failed to convert lifecycle state", slog.F("report", report))
				// Skip this report; there is no point retrying.  Maybe we can successfully convert the next one?
				a.lifecycleLastReportedIndex++
				continue
			}
			payload := &proto.UpdateLifecycleRequest{Lifecycle: l}
			logger := a.logger.With(slog.F("payload", payload))
			logger.Debug(ctx, "reporting lifecycle state")

			_, err = aAPI.UpdateLifecycle(ctx, payload)
			if err != nil {
				return xerrors.Errorf("failed to update lifecycle: %w", err)
			}

			logger.Debug(ctx, "successfully reported lifecycle state")
			a.lifecycleLastReportedIndex++
			select {
			case a.lifecycleReported <- report.State:
			case <-a.lifecycleReported:
				a.lifecycleReported <- report.State
			}
			if a.lifecycleLastReportedIndex < lastIndex {
				// Keep reporting until we've sent all messages, we can't
				// rely on the channel triggering us before the backlog is
				// consumed.
				continue
			}
			break
		}
	}
}

// setLifecycle sets the lifecycle state and notifies the lifecycle loop.
// The state is only updated if it's a valid state transition.
func (a *agent) setLifecycle(state codersdk.WorkspaceAgentLifecycle) {
	report := agentsdk.PostLifecycleRequest{
		State:     state,
		ChangedAt: dbtime.Now(),
	}

	a.lifecycleMu.Lock()
	lastReport := a.lifecycleStates[len(a.lifecycleStates)-1]
	if slices.Index(codersdk.WorkspaceAgentLifecycleOrder, lastReport.State) >= slices.Index(codersdk.WorkspaceAgentLifecycleOrder, report.State) {
		a.logger.Warn(context.Background(), "attempted to set lifecycle state to a previous state", slog.F("last", lastReport), slog.F("current", report))
		a.lifecycleMu.Unlock()
		return
	}
	a.lifecycleStates = append(a.lifecycleStates, report)
	a.logger.Debug(context.Background(), "set lifecycle state", slog.F("current", report), slog.F("last", lastReport))
	a.lifecycleMu.Unlock()

	select {
	case a.lifecycleUpdate <- struct{}{}:
	default:
	}
}

// fetchServiceBannerLoop fetches the service banner on an interval.  It will
// not be fetched immediately; the expectation is that it is primed elsewhere
// (and must be done before the session actually starts).
func (a *agent) fetchServiceBannerLoop(ctx context.Context, conn drpc.Conn) error {
	aAPI := proto.NewDRPCAgentClient(conn)
	ticker := time.NewTicker(a.announcementBannersRefreshInterval)
	defer ticker.Stop()
	for {
		select {
		case <-ctx.Done():
			return ctx.Err()
		case <-ticker.C:
			bannersProto, err := aAPI.GetAnnouncementBanners(ctx, &proto.GetAnnouncementBannersRequest{})
			if err != nil {
				if ctx.Err() != nil {
					return ctx.Err()
				}
				a.logger.Error(ctx, "failed to update notification banners", slog.Error(err))
				return err
			}
			banners := make([]codersdk.BannerConfig, 0, len(bannersProto.AnnouncementBanners))
			for _, bannerProto := range bannersProto.AnnouncementBanners {
				banners = append(banners, agentsdk.BannerConfigFromProto(bannerProto))
			}
			a.announcementBanners.Store(&banners)
		}
	}
}

func (a *agent) run() (retErr error) {
	// This allows the agent to refresh it's token if necessary.
	// For instance identity this is required, since the instance
	// may not have re-provisioned, but a new agent ID was created.
	sessionToken, err := a.exchangeToken(a.hardCtx)
	if err != nil {
		return xerrors.Errorf("exchange token: %w", err)
	}
	a.sessionToken.Store(&sessionToken)

	// ConnectRPC returns the dRPC connection we use for the Agent and Tailnet v2+ APIs
	conn, err := a.client.ConnectRPC(a.hardCtx)
	if err != nil {
		return err
	}
	defer func() {
		cErr := conn.Close()
		if cErr != nil {
			a.logger.Debug(a.hardCtx, "error closing drpc connection", slog.Error(err))
		}
	}()

	// A lot of routines need the agent API / tailnet API connection.  We run them in their own
	// goroutines in parallel, but errors in any routine will cause them all to exit so we can
	// redial the coder server and retry.
	connMan := newAPIConnRoutineManager(a.gracefulCtx, a.hardCtx, a.logger, conn)

	connMan.start("init notification banners", gracefulShutdownBehaviorStop,
		func(ctx context.Context, conn drpc.Conn) error {
			aAPI := proto.NewDRPCAgentClient(conn)
			bannersProto, err := aAPI.GetAnnouncementBanners(ctx, &proto.GetAnnouncementBannersRequest{})
			if err != nil {
				return xerrors.Errorf("fetch service banner: %w", err)
			}
			banners := make([]codersdk.BannerConfig, 0, len(bannersProto.AnnouncementBanners))
			for _, bannerProto := range bannersProto.AnnouncementBanners {
				banners = append(banners, agentsdk.BannerConfigFromProto(bannerProto))
			}
			a.announcementBanners.Store(&banners)
			return nil
		},
	)

	// sending logs gets gracefulShutdownBehaviorRemain because we want to send logs generated by
	// shutdown scripts.
	connMan.start("send logs", gracefulShutdownBehaviorRemain,
		func(ctx context.Context, conn drpc.Conn) error {
			err := a.logSender.SendLoop(ctx, proto.NewDRPCAgentClient(conn))
			if xerrors.Is(err, agentsdk.LogLimitExceededError) {
				// we don't want this error to tear down the API connection and propagate to the
				// other routines that use the API.  The LogSender has already dropped a warning
				// log, so just return nil here.
				return nil
			}
			return err
		})

	// part of graceful shut down is reporting the final lifecycle states, e.g "ShuttingDown" so the
	// lifecycle reporting has to be via gracefulShutdownBehaviorRemain
	connMan.start("report lifecycle", gracefulShutdownBehaviorRemain, a.reportLifecycle)

	// metadata reporting can cease as soon as we start gracefully shutting down
	connMan.start("report metadata", gracefulShutdownBehaviorStop, a.reportMetadata)

	// channels to sync goroutines below
	//  handle manifest
	//       |
	//    manifestOK
	//      |      |
	//      |      +----------------------+
	//      V                             |
	//      app health reporter           |
	//                                    V
	//                               create or update network
	//                                             |
	//                                          networkOK
	//                                             |
	//     coordination <--------------------------+
	//        derp map subscriber <----------------+
	//           stats report loop <---------------+
	networkOK := newCheckpoint(a.logger)
	manifestOK := newCheckpoint(a.logger)

	connMan.start("handle manifest", gracefulShutdownBehaviorStop, a.handleManifest(manifestOK))

	connMan.start("app health reporter", gracefulShutdownBehaviorStop,
		func(ctx context.Context, conn drpc.Conn) error {
			if err := manifestOK.wait(ctx); err != nil {
				return xerrors.Errorf("no manifest: %w", err)
			}
			manifest := a.manifest.Load()
			NewWorkspaceAppHealthReporter(
				a.logger, manifest.Apps, agentsdk.AppHealthPoster(proto.NewDRPCAgentClient(conn)),
			)(ctx)
			return nil
		})

	connMan.start("create or update network", gracefulShutdownBehaviorStop,
		a.createOrUpdateNetwork(manifestOK, networkOK))

	connMan.start("coordination", gracefulShutdownBehaviorStop,
		func(ctx context.Context, conn drpc.Conn) error {
			if err := networkOK.wait(ctx); err != nil {
				return xerrors.Errorf("no network: %w", err)
			}
			return a.runCoordinator(ctx, conn, a.network)
		},
	)

	connMan.start("derp map subscriber", gracefulShutdownBehaviorStop,
		func(ctx context.Context, conn drpc.Conn) error {
			if err := networkOK.wait(ctx); err != nil {
				return xerrors.Errorf("no network: %w", err)
			}
			return a.runDERPMapSubscriber(ctx, conn, a.network)
		})

	connMan.start("fetch service banner loop", gracefulShutdownBehaviorStop, a.fetchServiceBannerLoop)

	connMan.start("stats report loop", gracefulShutdownBehaviorStop, func(ctx context.Context, conn drpc.Conn) error {
		if err := networkOK.wait(ctx); err != nil {
			return xerrors.Errorf("no network: %w", err)
		}
		return a.statsReporter.reportLoop(ctx, proto.NewDRPCAgentClient(conn))
	})

	return connMan.wait()
}

// handleManifest returns a function that fetches and processes the manifest
func (a *agent) handleManifest(manifestOK *checkpoint) func(ctx context.Context, conn drpc.Conn) error {
	return func(ctx context.Context, conn drpc.Conn) error {
		var (
			sentResult = false
			err        error
		)
		defer func() {
			if !sentResult {
				manifestOK.complete(err)
			}
		}()
		aAPI := proto.NewDRPCAgentClient(conn)
		mp, err := aAPI.GetManifest(ctx, &proto.GetManifestRequest{})
		if err != nil {
			return xerrors.Errorf("fetch metadata: %w", err)
		}
		a.logger.Info(ctx, "fetched manifest", slog.F("manifest", mp))
		manifest, err := agentsdk.ManifestFromProto(mp)
		if err != nil {
			a.logger.Critical(ctx, "failed to convert manifest", slog.F("manifest", mp), slog.Error(err))
			return xerrors.Errorf("convert manifest: %w", err)
		}
		if manifest.AgentID == uuid.Nil {
			return xerrors.New("nil agentID returned by manifest")
		}
		a.client.RewriteDERPMap(manifest.DERPMap)

		// Expand the directory and send it back to coderd so external
		// applications that rely on the directory can use it.
		//
		// An example is VS Code Remote, which must know the directory
		// before initializing a connection.
		manifest.Directory, err = expandDirectory(manifest.Directory)
		if err != nil {
			return xerrors.Errorf("expand directory: %w", err)
		}
		subsys, err := agentsdk.ProtoFromSubsystems(a.subsystems)
		if err != nil {
			a.logger.Critical(ctx, "failed to convert subsystems", slog.Error(err))
			return xerrors.Errorf("failed to convert subsystems: %w", err)
		}
		_, err = aAPI.UpdateStartup(ctx, &proto.UpdateStartupRequest{Startup: &proto.Startup{
			Version:           buildinfo.Version(),
			ExpandedDirectory: manifest.Directory,
			Subsystems:        subsys,
		}})
		if err != nil {
			return xerrors.Errorf("update workspace agent startup: %w", err)
		}

		oldManifest := a.manifest.Swap(&manifest)
		manifestOK.complete(nil)
		sentResult = true

		// The startup script should only execute on the first run!
		if oldManifest == nil {
			a.setLifecycle(codersdk.WorkspaceAgentLifecycleStarting)

			// Perform overrides early so that Git auth can work even if users
			// connect to a workspace that is not yet ready. We don't run this
			// concurrently with the startup script to avoid conflicts between
			// them.
			if manifest.GitAuthConfigs > 0 {
				// If this fails, we should consider surfacing the error in the
				// startup log and setting the lifecycle state to be "start_error"
				// (after startup script completion), but for now we'll just log it.
				err := gitauth.OverrideVSCodeConfigs(a.filesystem)
				if err != nil {
					a.logger.Warn(ctx, "failed to override vscode git auth configs", slog.Error(err))
				}
			}

			err = a.scriptRunner.Init(manifest.Scripts)
			if err != nil {
				return xerrors.Errorf("init script runner: %w", err)
			}
			err = a.trackGoroutine(func() {
				start := time.Now()
				// here we use the graceful context because the script runner is not directly tied
				// to the agent API.
				err := a.scriptRunner.Execute(a.gracefulCtx, func(script codersdk.WorkspaceAgentScript) bool {
					return script.RunOnStart
				})
				// Measure the time immediately after the script has finished
				dur := time.Since(start).Seconds()
				if err != nil {
					a.logger.Warn(ctx, "startup script(s) failed", slog.Error(err))
					if errors.Is(err, agentscripts.ErrTimeout) {
						a.setLifecycle(codersdk.WorkspaceAgentLifecycleStartTimeout)
					} else {
						a.setLifecycle(codersdk.WorkspaceAgentLifecycleStartError)
					}
				} else {
					a.setLifecycle(codersdk.WorkspaceAgentLifecycleReady)
				}

				label := "false"
				if err == nil {
					label = "true"
				}
				a.metrics.startupScriptSeconds.WithLabelValues(label).Set(dur)
				a.scriptRunner.StartCron()
			})
			if err != nil {
				return xerrors.Errorf("track conn goroutine: %w", err)
			}
		}
		return nil
	}
}

// createOrUpdateNetwork waits for the manifest to be set using manifestOK, then creates or updates
// the tailnet using the information in the manifest
func (a *agent) createOrUpdateNetwork(manifestOK, networkOK *checkpoint) func(context.Context, drpc.Conn) error {
	return func(ctx context.Context, _ drpc.Conn) (retErr error) {
		if err := manifestOK.wait(ctx); err != nil {
			return xerrors.Errorf("no manifest: %w", err)
		}
		var err error
		defer func() {
			networkOK.complete(retErr)
		}()
		manifest := a.manifest.Load()
		a.closeMutex.Lock()
		network := a.network
		a.closeMutex.Unlock()
		if network == nil {
			// use the graceful context here, because creating the tailnet is not itself tied to the
			// agent API.
			network, err = a.createTailnet(a.gracefulCtx, manifest.AgentID, manifest.DERPMap, manifest.DERPForceWebSockets, manifest.DisableDirectConnections)
			if err != nil {
				return xerrors.Errorf("create tailnet: %w", err)
			}
			a.closeMutex.Lock()
			// Re-check if agent was closed while initializing the network.
			closed := a.isClosed()
			if !closed {
				a.network = network
				a.statsReporter = newStatsReporter(a.logger, network, a)
			}
			a.closeMutex.Unlock()
			if closed {
				_ = network.Close()
				return xerrors.New("agent is closed")
			}
		} else {
			// Update the wireguard IPs if the agent ID changed.
			err := network.SetAddresses(a.wireguardAddresses(manifest.AgentID))
			if err != nil {
				a.logger.Error(a.gracefulCtx, "update tailnet addresses", slog.Error(err))
			}
			// Update the DERP map, force WebSocket setting and allow/disallow
			// direct connections.
			network.SetDERPMap(manifest.DERPMap)
			network.SetDERPForceWebSockets(manifest.DERPForceWebSockets)
			network.SetBlockEndpoints(manifest.DisableDirectConnections)
		}
		return nil
	}
}

// updateCommandEnv updates the provided command environment with the
// following set of environment variables:
// - Predefined workspace environment variables
// - Environment variables currently set (overriding predefined)
// - Environment variables passed via the agent manifest (overriding predefined and current)
// - Agent-level environment variables (overriding all)
func (a *agent) updateCommandEnv(current []string) (updated []string, err error) {
	manifest := a.manifest.Load()
	if manifest == nil {
		return nil, xerrors.Errorf("no manifest")
	}

	executablePath, err := os.Executable()
	if err != nil {
		return nil, xerrors.Errorf("getting os executable: %w", err)
	}
	unixExecutablePath := strings.ReplaceAll(executablePath, "\\", "/")

	// Define environment variables that should be set for all commands,
	// and then merge them with the current environment.
	envs := map[string]string{
		// Set env vars indicating we're inside a Coder workspace.
		"CODER":                      "true",
		"CODER_WORKSPACE_NAME":       manifest.WorkspaceName,
		"CODER_WORKSPACE_AGENT_NAME": manifest.AgentName,

		// Specific Coder subcommands require the agent token exposed!
		"CODER_AGENT_TOKEN": *a.sessionToken.Load(),

		// Git on Windows resolves with UNIX-style paths.
		// If using backslashes, it's unable to find the executable.
		"GIT_SSH_COMMAND": fmt.Sprintf("%s gitssh --", unixExecutablePath),
		// Hide Coder message on code-server's "Getting Started" page
		"CS_DISABLE_GETTING_STARTED_OVERRIDE": "true",
	}

	// This adds the ports dialog to code-server that enables
	// proxying a port dynamically.
	// If this is empty string, do not set anything. Code-server auto defaults
	// using its basepath to construct a path based port proxy.
	if manifest.VSCodePortProxyURI != "" {
		envs["VSCODE_PROXY_URI"] = manifest.VSCodePortProxyURI
	}

	// Allow any of the current env to override what we defined above.
	for _, env := range current {
		parts := strings.SplitN(env, "=", 2)
		if len(parts) != 2 {
			continue
		}
		if _, ok := envs[parts[0]]; !ok {
			envs[parts[0]] = parts[1]
		}
	}

	// Load environment variables passed via the agent manifest.
	// These override all variables we manually specify.
	for k, v := range manifest.EnvironmentVariables {
		// Expanding environment variables allows for customization
		// of the $PATH, among other variables. Customers can prepend
		// or append to the $PATH, so allowing expand is required!
		envs[k] = os.ExpandEnv(v)
	}

	// Agent-level environment variables should take over all. This is
	// used for setting agent-specific variables like CODER_AGENT_TOKEN
	// and GIT_ASKPASS.
	for k, v := range a.environmentVariables {
		envs[k] = v
	}

	// Prepend the agent script bin directory to the PATH
	// (this is where Coder modules place their binaries).
	if _, ok := envs["PATH"]; !ok {
		envs["PATH"] = os.Getenv("PATH")
	}
	envs["PATH"] = fmt.Sprintf("%s%c%s", a.scriptRunner.ScriptBinDir(), filepath.ListSeparator, envs["PATH"])

	for k, v := range envs {
		updated = append(updated, fmt.Sprintf("%s=%s", k, v))
	}
	return updated, nil
}

func (a *agent) wireguardAddresses(agentID uuid.UUID) []netip.Prefix {
	if len(a.addresses) == 0 {
		return []netip.Prefix{
			// This is the IP that should be used primarily.
			netip.PrefixFrom(tailnet.IPFromUUID(agentID), 128),
			// We also listen on the legacy codersdk.WorkspaceAgentIP. This
			// allows for a transition away from wsconncache.
			netip.PrefixFrom(workspacesdk.AgentIP, 128),
		}
	}

	return a.addresses
}

func (a *agent) trackGoroutine(fn func()) error {
	a.closeMutex.Lock()
	defer a.closeMutex.Unlock()
	if a.isClosed() {
		return xerrors.New("track conn goroutine: agent is closed")
	}
	a.closeWaitGroup.Add(1)
	go func() {
		defer a.closeWaitGroup.Done()
		fn()
	}()
	return nil
}

func (a *agent) createTailnet(ctx context.Context, agentID uuid.UUID, derpMap *tailcfg.DERPMap, derpForceWebSockets, disableDirectConnections bool) (_ *tailnet.Conn, err error) {
	network, err := tailnet.NewConn(&tailnet.Options{
		ID:                  agentID,
		Addresses:           a.wireguardAddresses(agentID),
		DERPMap:             derpMap,
		DERPForceWebSockets: derpForceWebSockets,
		Logger:              a.logger.Named("net.tailnet"),
		ListenPort:          a.tailnetListenPort,
		BlockEndpoints:      disableDirectConnections,
	})
	if err != nil {
		return nil, xerrors.Errorf("create tailnet: %w", err)
	}
	defer func() {
		if err != nil {
			network.Close()
		}
	}()

	sshListener, err := network.Listen("tcp", ":"+strconv.Itoa(workspacesdk.AgentSSHPort))
	if err != nil {
		return nil, xerrors.Errorf("listen on the ssh port: %w", err)
	}
	defer func() {
		if err != nil {
			_ = sshListener.Close()
		}
	}()
	if err = a.trackGoroutine(func() {
		_ = a.sshServer.Serve(sshListener)
	}); err != nil {
		return nil, err
	}

	reconnectingPTYListener, err := network.Listen("tcp", ":"+strconv.Itoa(workspacesdk.AgentReconnectingPTYPort))
	if err != nil {
		return nil, xerrors.Errorf("listen for reconnecting pty: %w", err)
	}
	defer func() {
		if err != nil {
			_ = reconnectingPTYListener.Close()
		}
	}()
	if err = a.trackGoroutine(func() {
		logger := a.logger.Named("reconnecting-pty")
		var wg sync.WaitGroup
		for {
			conn, err := reconnectingPTYListener.Accept()
			if err != nil {
				if !a.isClosed() {
					logger.Debug(ctx, "accept pty failed", slog.Error(err))
				}
				break
			}
			clog := logger.With(
				slog.F("remote", conn.RemoteAddr().String()),
				slog.F("local", conn.LocalAddr().String()))
			clog.Info(ctx, "accepted conn")
			wg.Add(1)
			closed := make(chan struct{})
			go func() {
				select {
				case <-closed:
				case <-a.hardCtx.Done():
					_ = conn.Close()
				}
				wg.Done()
			}()
			go func() {
				defer close(closed)
				// This cannot use a JSON decoder, since that can
				// buffer additional data that is required for the PTY.
				rawLen := make([]byte, 2)
				_, err = conn.Read(rawLen)
				if err != nil {
					return
				}
				length := binary.LittleEndian.Uint16(rawLen)
				data := make([]byte, length)
				_, err = conn.Read(data)
				if err != nil {
					return
				}
				var msg workspacesdk.AgentReconnectingPTYInit
				err = json.Unmarshal(data, &msg)
				if err != nil {
					logger.Warn(ctx, "failed to unmarshal init", slog.F("raw", data))
					return
				}
				_ = a.handleReconnectingPTY(ctx, clog, msg, conn)
			}()
		}
		wg.Wait()
	}); err != nil {
		return nil, err
	}

	speedtestListener, err := network.Listen("tcp", ":"+strconv.Itoa(workspacesdk.AgentSpeedtestPort))
	if err != nil {
		return nil, xerrors.Errorf("listen for speedtest: %w", err)
	}
	defer func() {
		if err != nil {
			_ = speedtestListener.Close()
		}
	}()
	if err = a.trackGoroutine(func() {
		var wg sync.WaitGroup
		for {
			conn, err := speedtestListener.Accept()
			if err != nil {
				if !a.isClosed() {
					a.logger.Debug(ctx, "speedtest listener failed", slog.Error(err))
				}
				break
			}
			clog := a.logger.Named("speedtest").With(
				slog.F("remote", conn.RemoteAddr().String()),
				slog.F("local", conn.LocalAddr().String()))
			clog.Info(ctx, "accepted conn")
			wg.Add(1)
			closed := make(chan struct{})
			go func() {
				select {
				case <-closed:
				case <-a.hardCtx.Done():
					_ = conn.Close()
				}
				wg.Done()
			}()
			go func() {
				defer close(closed)
				sErr := speedtest.ServeConn(conn)
				if sErr != nil {
					clog.Error(ctx, "test ended with error", slog.Error(sErr))
					return
				}
				clog.Info(ctx, "test ended")
			}()
		}
		wg.Wait()
	}); err != nil {
		return nil, err
	}

	apiListener, err := network.Listen("tcp", ":"+strconv.Itoa(workspacesdk.AgentHTTPAPIServerPort))
	if err != nil {
		return nil, xerrors.Errorf("api listener: %w", err)
	}
	defer func() {
		if err != nil {
			_ = apiListener.Close()
		}
	}()
	if err = a.trackGoroutine(func() {
		defer apiListener.Close()
		server := &http.Server{
			Handler:           a.apiHandler(),
			ReadTimeout:       20 * time.Second,
			ReadHeaderTimeout: 20 * time.Second,
			WriteTimeout:      20 * time.Second,
			ErrorLog:          slog.Stdlib(ctx, a.logger.Named("http_api_server"), slog.LevelInfo),
		}
		go func() {
			select {
			case <-ctx.Done():
			case <-a.hardCtx.Done():
			}
			_ = server.Close()
		}()

		err := server.Serve(apiListener)
		if err != nil && !xerrors.Is(err, http.ErrServerClosed) && !strings.Contains(err.Error(), "use of closed network connection") {
			a.logger.Critical(ctx, "serve HTTP API server", slog.Error(err))
		}
	}); err != nil {
		return nil, err
	}

	return network, nil
}

// runCoordinator runs a coordinator and returns whether a reconnect
// should occur.
func (a *agent) runCoordinator(ctx context.Context, conn drpc.Conn, network *tailnet.Conn) error {
	defer a.logger.Debug(ctx, "disconnected from coordination RPC")
	tClient := tailnetproto.NewDRPCTailnetClient(conn)
	// we run the RPC on the hardCtx so that we have a chance to send the disconnect message if we
	// gracefully shut down.
	coordinate, err := tClient.Coordinate(a.hardCtx)
	if err != nil {
		return xerrors.Errorf("failed to connect to the coordinate endpoint: %w", err)
	}
	defer func() {
		cErr := coordinate.Close()
		if cErr != nil {
			a.logger.Debug(ctx, "error closing Coordinate client", slog.Error(err))
		}
	}()
	a.logger.Info(ctx, "connected to coordination RPC")

	// This allows the Close() routine to wait for the coordinator to gracefully disconnect.
	a.closeMutex.Lock()
	if a.isClosed() {
		return nil
	}
	disconnected := make(chan struct{})
	a.coordDisconnected = disconnected
	defer close(disconnected)
	a.closeMutex.Unlock()

	coordination := tailnet.NewRemoteCoordination(a.logger, coordinate, network, uuid.Nil)

	errCh := make(chan error, 1)
	go func() {
		defer close(errCh)
		select {
		case <-ctx.Done():
			err := coordination.Close()
			if err != nil {
				a.logger.Warn(ctx, "failed to close remote coordination", slog.Error(err))
			}
			return
		case err := <-coordination.Error():
			errCh <- err
		}
	}()
	return <-errCh
}

// runDERPMapSubscriber runs a coordinator and returns if a reconnect should occur.
func (a *agent) runDERPMapSubscriber(ctx context.Context, conn drpc.Conn, network *tailnet.Conn) error {
	defer a.logger.Debug(ctx, "disconnected from derp map RPC")
	ctx, cancel := context.WithCancel(ctx)
	defer cancel()
	tClient := tailnetproto.NewDRPCTailnetClient(conn)
	stream, err := tClient.StreamDERPMaps(ctx, &tailnetproto.StreamDERPMapsRequest{})
	if err != nil {
		return xerrors.Errorf("stream DERP Maps: %w", err)
	}
	defer func() {
		cErr := stream.Close()
		if cErr != nil {
			a.logger.Debug(ctx, "error closing DERPMap stream", slog.Error(err))
		}
	}()
	a.logger.Info(ctx, "connected to derp map RPC")
	for {
		dmp, err := stream.Recv()
		if err != nil {
			return xerrors.Errorf("recv DERPMap error: %w", err)
		}
		dm := tailnet.DERPMapFromProto(dmp)
		a.client.RewriteDERPMap(dm)
		network.SetDERPMap(dm)
	}
}

func (a *agent) handleReconnectingPTY(ctx context.Context, logger slog.Logger, msg workspacesdk.AgentReconnectingPTYInit, conn net.Conn) (retErr error) {
	defer conn.Close()
	a.metrics.connectionsTotal.Add(1)

	a.connCountReconnectingPTY.Add(1)
	defer a.connCountReconnectingPTY.Add(-1)

	connectionID := uuid.NewString()
	connLogger := logger.With(slog.F("message_id", msg.ID), slog.F("connection_id", connectionID))
	connLogger.Debug(ctx, "starting handler")

	defer func() {
		if err := retErr; err != nil {
			a.closeMutex.Lock()
			closed := a.isClosed()
			a.closeMutex.Unlock()

			// If the agent is closed, we don't want to
			// log this as an error since it's expected.
			if closed {
				connLogger.Info(ctx, "reconnecting pty failed with attach error (agent closed)", slog.Error(err))
			} else {
				connLogger.Error(ctx, "reconnecting pty failed with attach error", slog.Error(err))
			}
		}
		connLogger.Info(ctx, "reconnecting pty connection closed")
	}()

	var rpty reconnectingpty.ReconnectingPTY
	sendConnected := make(chan reconnectingpty.ReconnectingPTY, 1)
	// On store, reserve this ID to prevent multiple concurrent new connections.
	waitReady, ok := a.reconnectingPTYs.LoadOrStore(msg.ID, sendConnected)
	if ok {
		close(sendConnected) // Unused.
		connLogger.Debug(ctx, "connecting to existing reconnecting pty")
		c, ok := waitReady.(chan reconnectingpty.ReconnectingPTY)
		if !ok {
			return xerrors.Errorf("found invalid type in reconnecting pty map: %T", waitReady)
		}
		rpty, ok = <-c
		if !ok || rpty == nil {
			return xerrors.Errorf("reconnecting pty closed before connection")
		}
		c <- rpty // Put it back for the next reconnect.
	} else {
		connLogger.Debug(ctx, "creating new reconnecting pty")

		connected := false
		defer func() {
			if !connected && retErr != nil {
				a.reconnectingPTYs.Delete(msg.ID)
				close(sendConnected)
			}
		}()

		// Empty command will default to the users shell!
		cmd, err := a.sshServer.CreateCommand(ctx, msg.Command, nil)
		if err != nil {
			a.metrics.reconnectingPTYErrors.WithLabelValues("create_command").Add(1)
			return xerrors.Errorf("create command: %w", err)
		}

		rpty = reconnectingpty.New(ctx, cmd, &reconnectingpty.Options{
			Timeout: a.reconnectingPTYTimeout,
			Metrics: a.metrics.reconnectingPTYErrors,
		}, logger.With(slog.F("message_id", msg.ID)))

		if err = a.trackGoroutine(func() {
			rpty.Wait()
			a.reconnectingPTYs.Delete(msg.ID)
		}); err != nil {
			rpty.Close(err)
			return xerrors.Errorf("start routine: %w", err)
		}

		connected = true
		sendConnected <- rpty
	}
	return rpty.Attach(ctx, connectionID, conn, msg.Height, msg.Width, connLogger)
}

// Collect collects additional stats from the agent
func (a *agent) Collect(ctx context.Context, networkStats map[netlogtype.Connection]netlogtype.Counts) *proto.Stats {
	a.logger.Debug(context.Background(), "computing stats report")
	stats := &proto.Stats{
		ConnectionCount:    int64(len(networkStats)),
		ConnectionsByProto: map[string]int64{},
	}
	for conn, counts := range networkStats {
		stats.ConnectionsByProto[conn.Proto.String()]++
		stats.RxBytes += int64(counts.RxBytes)
		stats.RxPackets += int64(counts.RxPackets)
		stats.TxBytes += int64(counts.TxBytes)
		stats.TxPackets += int64(counts.TxPackets)
	}

	// The count of active sessions.
	sshStats := a.sshServer.ConnStats()
	stats.SessionCountSsh = sshStats.Sessions
	stats.SessionCountVscode = sshStats.VSCode
	stats.SessionCountJetbrains = sshStats.JetBrains

	stats.SessionCountReconnectingPty = a.connCountReconnectingPTY.Load()

	// Compute the median connection latency!
	a.logger.Debug(ctx, "starting peer latency measurement for stats")
	var wg sync.WaitGroup
	var mu sync.Mutex
	status := a.network.Status()
	durations := []float64{}
	pingCtx, cancelFunc := context.WithTimeout(ctx, 5*time.Second)
	defer cancelFunc()
	for nodeID, peer := range status.Peer {
		if !peer.Active {
			continue
		}
		addresses, found := a.network.NodeAddresses(nodeID)
		if !found {
			continue
		}
		if len(addresses) == 0 {
			continue
		}
		wg.Add(1)
		go func() {
			defer wg.Done()
			duration, _, _, err := a.network.Ping(pingCtx, addresses[0].Addr())
			if err != nil {
				return
			}
			mu.Lock()
			defer mu.Unlock()
			durations = append(durations, float64(duration.Microseconds()))
		}()
	}
	wg.Wait()
	sort.Float64s(durations)
	durationsLength := len(durations)
	if durationsLength == 0 {
		stats.ConnectionMedianLatencyMs = -1
	} else if durationsLength%2 == 0 {
		stats.ConnectionMedianLatencyMs = (durations[durationsLength/2-1] + durations[durationsLength/2]) / 2
	} else {
		stats.ConnectionMedianLatencyMs = durations[durationsLength/2]
	}
	// Convert from microseconds to milliseconds.
	stats.ConnectionMedianLatencyMs /= 1000

	// Collect agent metrics.
	// Agent metrics are changing all the time, so there is no need to perform
	// reflect.DeepEqual to see if stats should be transferred.

	metricsCtx, cancelFunc := context.WithTimeout(ctx, 5*time.Second)
	defer cancelFunc()
	a.logger.Debug(ctx, "collecting agent metrics for stats")
	stats.Metrics = a.collectMetrics(metricsCtx)

	return stats
}

var prioritizedProcs = []string{"coder agent"}

func (a *agent) manageProcessPriorityUntilGracefulShutdown() {
	// process priority can stop as soon as we are gracefully shutting down
	ctx := a.gracefulCtx
	defer func() {
		if r := recover(); r != nil {
			a.logger.Critical(ctx, "recovered from panic",
				slog.F("panic", r),
				slog.F("stack", string(debug.Stack())),
			)
		}
	}()

	if val := a.environmentVariables[EnvProcPrioMgmt]; val == "" || runtime.GOOS != "linux" {
		a.logger.Debug(ctx, "process priority not enabled, agent will not manage process niceness/oom_score_adj ",
			slog.F("env_var", EnvProcPrioMgmt),
			slog.F("value", val),
			slog.F("goos", runtime.GOOS),
		)
		return
	}

	if a.processManagementTick == nil {
		ticker := time.NewTicker(time.Second)
		defer ticker.Stop()
		a.processManagementTick = ticker.C
	}

	oomScore := unsetOOMScore
	if scoreStr, ok := a.environmentVariables[EnvProcOOMScore]; ok {
		score, err := strconv.Atoi(strings.TrimSpace(scoreStr))
		if err == nil && score >= -1000 && score <= 1000 {
			oomScore = score
		} else {
			a.logger.Error(ctx, "invalid oom score",
				slog.F("min_value", -1000),
				slog.F("max_value", 1000),
				slog.F("value", scoreStr),
			)
		}
	}

	debouncer := &logDebouncer{
		logger:   a.logger,
		messages: map[string]time.Time{},
		interval: time.Minute,
	}

	for {
		procs, err := a.manageProcessPriority(ctx, debouncer, oomScore)
		// Avoid spamming the logs too often.
		if err != nil {
			debouncer.Error(ctx, "manage process priority",
				slog.Error(err),
			)
		}
		if a.modifiedProcs != nil {
			a.modifiedProcs <- procs
		}

		select {
		case <-a.processManagementTick:
		case <-ctx.Done():
			return
		}
	}
}

// unsetOOMScore is set to an invalid OOM score to imply an unset value.
const unsetOOMScore = 1001

func (a *agent) manageProcessPriority(ctx context.Context, debouncer *logDebouncer, oomScore int) ([]*agentproc.Process, error) {
	const (
		niceness = 10
	)

	// We fetch the agent score each time because it's possible someone updates the
	// value after it is started.
	agentScore, err := a.getAgentOOMScore()
	if err != nil {
		agentScore = unsetOOMScore
	}
	if oomScore == unsetOOMScore && agentScore != unsetOOMScore {
		// If the child score has not been explicitly specified we should
		// set it to a score relative to the agent score.
		oomScore = childOOMScore(agentScore)
	}

	procs, err := agentproc.List(a.filesystem, a.syscaller)
	if err != nil {
		return nil, xerrors.Errorf("list: %w", err)
	}

	modProcs := []*agentproc.Process{}

	for _, proc := range procs {
		containsFn := func(e string) bool {
			contains := strings.Contains(proc.Cmd(), e)
			return contains
		}

		// If the process is prioritized we should adjust
		// it's oom_score_adj and avoid lowering its niceness.
		if slices.ContainsFunc(prioritizedProcs, containsFn) {
			continue
		}

		score, niceErr := proc.Niceness(a.syscaller)
		if niceErr != nil && !xerrors.Is(niceErr, os.ErrPermission) {
			debouncer.Warn(ctx, "unable to get proc niceness",
				slog.F("cmd", proc.Cmd()),
				slog.F("pid", proc.PID),
				slog.Error(niceErr),
			)
			continue
		}

		// We only want processes that don't have a nice value set
		// so we don't override user nice values.
		// Getpriority actually returns priority for the nice value
		// which is niceness + 20, so here 20 = a niceness of 0 (aka unset).
		if score != 20 {
			// We don't log here since it can get spammy
			continue
		}

		if niceErr == nil {
			err := proc.SetNiceness(a.syscaller, niceness)
			if err != nil && !xerrors.Is(err, os.ErrPermission) {
				debouncer.Warn(ctx, "unable to set proc niceness",
					slog.F("cmd", proc.Cmd()),
					slog.F("pid", proc.PID),
					slog.F("niceness", niceness),
					slog.Error(err),
				)
			}
		}

		// If the oom score is valid and it's not already set and isn't a custom value set by another process then it's ok to update it.
		if oomScore != unsetOOMScore && oomScore != proc.OOMScoreAdj && !isCustomOOMScore(agentScore, proc) {
			oomScoreStr := strconv.Itoa(oomScore)
			err := afero.WriteFile(a.filesystem, fmt.Sprintf("/proc/%d/oom_score_adj", proc.PID), []byte(oomScoreStr), 0o644)
			if err != nil && !xerrors.Is(err, os.ErrPermission) {
				debouncer.Warn(ctx, "unable to set oom_score_adj",
					slog.F("cmd", proc.Cmd()),
					slog.F("pid", proc.PID),
					slog.F("score", oomScoreStr),
					slog.Error(err),
				)
			}
		}
		modProcs = append(modProcs, proc)
	}
	return modProcs, nil
}

// isClosed returns whether the API is closed or not.
func (a *agent) isClosed() bool {
	return a.hardCtx.Err() != nil
}

func (a *agent) requireNetwork() (*tailnet.Conn, bool) {
	a.closeMutex.Lock()
	defer a.closeMutex.Unlock()
	return a.network, a.network != nil
}

func (a *agent) HandleHTTPDebugMagicsock(w http.ResponseWriter, r *http.Request) {
	network, ok := a.requireNetwork()
	if !ok {
		w.WriteHeader(http.StatusInternalServerError)
		_, _ = w.Write([]byte("network is not ready yet"))
		return
	}
	network.MagicsockServeHTTPDebug(w, r)
}

func (a *agent) HandleHTTPMagicsockDebugLoggingState(w http.ResponseWriter, r *http.Request) {
	state := chi.URLParam(r, "state")
	stateBool, err := strconv.ParseBool(state)
	if err != nil {
		w.WriteHeader(http.StatusBadRequest)
		_, _ = fmt.Fprintf(w, "invalid state %q, must be a boolean", state)
		return
	}

	network, ok := a.requireNetwork()
	if !ok {
		w.WriteHeader(http.StatusInternalServerError)
		_, _ = w.Write([]byte("network is not ready yet"))
		return
	}

	network.MagicsockSetDebugLoggingEnabled(stateBool)
	a.logger.Info(r.Context(), "updated magicsock debug logging due to debug request", slog.F("new_state", stateBool))

	w.WriteHeader(http.StatusOK)
	_, _ = fmt.Fprintf(w, "updated magicsock debug logging to %v", stateBool)
}

func (a *agent) HandleHTTPDebugManifest(w http.ResponseWriter, r *http.Request) {
	sdkManifest := a.manifest.Load()
	if sdkManifest == nil {
		a.logger.Error(r.Context(), "no manifest in-memory")
		w.WriteHeader(http.StatusInternalServerError)
		_, _ = fmt.Fprintf(w, "no manifest in-memory")
		return
	}

	w.WriteHeader(http.StatusOK)
	if err := json.NewEncoder(w).Encode(sdkManifest); err != nil {
		a.logger.Error(a.hardCtx, "write debug manifest", slog.Error(err))
	}
}

func (a *agent) HandleHTTPDebugLogs(w http.ResponseWriter, r *http.Request) {
	logPath := filepath.Join(a.logDir, "coder-agent.log")
	f, err := os.Open(logPath)
	if err != nil {
		a.logger.Error(r.Context(), "open agent log file", slog.Error(err), slog.F("path", logPath))
		w.WriteHeader(http.StatusInternalServerError)
		_, _ = fmt.Fprintf(w, "could not open log file: %s", err)
		return
	}
	defer f.Close()

	// Limit to 10MB.
	w.WriteHeader(http.StatusOK)
	_, err = io.Copy(w, io.LimitReader(f, 10*1024*1024))
	if err != nil && !errors.Is(err, io.EOF) {
		a.logger.Error(r.Context(), "read agent log file", slog.Error(err))
		return
	}
}

func (a *agent) HTTPDebug() http.Handler {
	r := chi.NewRouter()

	r.Get("/debug/logs", a.HandleHTTPDebugLogs)
	r.Get("/debug/magicsock", a.HandleHTTPDebugMagicsock)
	r.Get("/debug/magicsock/debug-logging/{state}", a.HandleHTTPMagicsockDebugLoggingState)
	r.Get("/debug/manifest", a.HandleHTTPDebugManifest)
	r.NotFound(func(w http.ResponseWriter, r *http.Request) {
		w.WriteHeader(http.StatusNotFound)
		_, _ = w.Write([]byte("404 not found"))
	})

	return r
}

func (a *agent) Close() error {
	a.closeMutex.Lock()
	defer a.closeMutex.Unlock()
	if a.isClosed() {
		return nil
	}

	a.logger.Info(a.hardCtx, "shutting down agent")
	a.setLifecycle(codersdk.WorkspaceAgentLifecycleShuttingDown)

	// Attempt to gracefully shut down all active SSH connections and
	// stop accepting new ones.
	err := a.sshServer.Shutdown(a.hardCtx)
	if err != nil {
		a.logger.Error(a.hardCtx, "ssh server shutdown", slog.Error(err))
	}
	err = a.sshServer.Close()
	if err != nil {
		a.logger.Error(a.hardCtx, "ssh server close", slog.Error(err))
	}
	// wait for SSH to shut down before the general graceful cancel, because
	// this triggers a disconnect in the tailnet layer, telling all clients to
	// shut down their wireguard tunnels to us. If SSH sessions are still up,
	// they might hang instead of being closed.
	a.gracefulCancel()

	lifecycleState := codersdk.WorkspaceAgentLifecycleOff
	err = a.scriptRunner.Execute(a.hardCtx, func(script codersdk.WorkspaceAgentScript) bool {
		return script.RunOnStop
	})
	if err != nil {
		a.logger.Warn(a.hardCtx, "shutdown script(s) failed", slog.Error(err))
		if errors.Is(err, agentscripts.ErrTimeout) {
			lifecycleState = codersdk.WorkspaceAgentLifecycleShutdownTimeout
		} else {
			lifecycleState = codersdk.WorkspaceAgentLifecycleShutdownError
		}
	}
	a.setLifecycle(lifecycleState)

	err = a.scriptRunner.Close()
	if err != nil {
		a.logger.Error(a.hardCtx, "script runner close", slog.Error(err))
	}

	// Wait for the graceful shutdown to complete, but don't wait forever so
	// that we don't break user expectations.
	go func() {
		defer a.hardCancel()
		select {
		case <-a.hardCtx.Done():
		case <-time.After(5 * time.Second):
		}
	}()

	// Wait for lifecycle to be reported
lifecycleWaitLoop:
	for {
		select {
		case <-a.hardCtx.Done():
			a.logger.Warn(context.Background(), "failed to report final lifecycle state")
			break lifecycleWaitLoop
		case s := <-a.lifecycleReported:
			if s == lifecycleState {
				a.logger.Debug(context.Background(), "reported final lifecycle state")
				break lifecycleWaitLoop
			}
		}
	}

	// Wait for graceful disconnect from the Coordinator RPC
	select {
	case <-a.hardCtx.Done():
		a.logger.Warn(context.Background(), "timed out waiting for Coordinator RPC disconnect")
	case <-a.coordDisconnected:
		a.logger.Debug(context.Background(), "coordinator RPC disconnected")
	}

	// Wait for logs to be sent
	err = a.logSender.WaitUntilEmpty(a.hardCtx)
	if err != nil {
		a.logger.Warn(context.Background(), "timed out waiting for all logs to be sent", slog.Error(err))
	}

	a.hardCancel()
	if a.network != nil {
		_ = a.network.Close()
	}
	a.closeWaitGroup.Wait()

	return nil
}

// userHomeDir returns the home directory of the current user, giving
// priority to the $HOME environment variable.
func userHomeDir() (string, error) {
	// First we check the environment.
	homedir, err := os.UserHomeDir()
	if err == nil {
		return homedir, nil
	}

	// As a fallback, we try the user information.
	u, err := user.Current()
	if err != nil {
		return "", xerrors.Errorf("current user: %w", err)
	}
	return u.HomeDir, nil
}

// expandDirectory converts a directory path to an absolute path.
// It primarily resolves the home directory and any environment
// variables that may be set
func expandDirectory(dir string) (string, error) {
	if dir == "" {
		return "", nil
	}
	if dir[0] == '~' {
		home, err := userHomeDir()
		if err != nil {
			return "", err
		}
		dir = filepath.Join(home, dir[1:])
	}
	dir = os.ExpandEnv(dir)

	if !filepath.IsAbs(dir) {
		home, err := userHomeDir()
		if err != nil {
			return "", err
		}
		dir = filepath.Join(home, dir)
	}
	return dir, nil
}

// EnvAgentSubsystem is the environment variable used to denote the
// specialized environment in which the agent is running
// (e.g. envbox, envbuilder).
const EnvAgentSubsystem = "CODER_AGENT_SUBSYSTEM"

// eitherContext returns a context that is canceled when either context ends.
func eitherContext(a, b context.Context) context.Context {
	ctx, cancel := context.WithCancel(a)
	go func() {
		defer cancel()
		select {
		case <-a.Done():
		case <-b.Done():
		}
	}()
	return ctx
}

type gracefulShutdownBehavior int

const (
	gracefulShutdownBehaviorStop gracefulShutdownBehavior = iota
	gracefulShutdownBehaviorRemain
)

type apiConnRoutineManager struct {
	logger    slog.Logger
	conn      drpc.Conn
	eg        *errgroup.Group
	stopCtx   context.Context
	remainCtx context.Context
}

func newAPIConnRoutineManager(gracefulCtx, hardCtx context.Context, logger slog.Logger, conn drpc.Conn) *apiConnRoutineManager {
	// routines that remain in operation during graceful shutdown use the remainCtx.  They'll still
	// exit if the errgroup hits an error, which usually means a problem with the conn.
	eg, remainCtx := errgroup.WithContext(hardCtx)

	// routines that stop operation during graceful shutdown use the stopCtx, which ends when the
	// first of remainCtx or gracefulContext ends (an error or start of graceful shutdown).
	//
	// +------------------------------------------+
	// | hardCtx                                  |
	// |  +------------------------------------+  |
	// |  |  stopCtx                           |  |
	// |  | +--------------+  +--------------+ |  |
	// |  | | remainCtx    |  | gracefulCtx  | |  |
	// |  | +--------------+  +--------------+ |  |
	// |  +------------------------------------+  |
	// +------------------------------------------+
	stopCtx := eitherContext(remainCtx, gracefulCtx)
	return &apiConnRoutineManager{
		logger:    logger,
		conn:      conn,
		eg:        eg,
		stopCtx:   stopCtx,
		remainCtx: remainCtx,
	}
}

func (a *apiConnRoutineManager) start(name string, b gracefulShutdownBehavior, f func(context.Context, drpc.Conn) error) {
	logger := a.logger.With(slog.F("name", name))
	var ctx context.Context
	switch b {
	case gracefulShutdownBehaviorStop:
		ctx = a.stopCtx
	case gracefulShutdownBehaviorRemain:
		ctx = a.remainCtx
	default:
		panic("unknown behavior")
	}
	a.eg.Go(func() error {
		logger.Debug(ctx, "starting routine")
		err := f(ctx, a.conn)
		if xerrors.Is(err, context.Canceled) && ctx.Err() != nil {
			logger.Debug(ctx, "swallowing context canceled")
			// Don't propagate context canceled errors to the error group, because we don't want the
			// graceful context being canceled to halt the work of routines with
			// gracefulShutdownBehaviorRemain.  Note that we check both that the error is
			// context.Canceled and that *our* context is currently canceled, because when Coderd
			// unilaterally closes the API connection (for example if the build is outdated), it can
			// sometimes show up as context.Canceled in our RPC calls.
			return nil
		}
		logger.Debug(ctx, "routine exited", slog.Error(err))
		if err != nil {
			return xerrors.Errorf("error in routine %s: %w", name, err)
		}
		return nil
	})
}

func (a *apiConnRoutineManager) wait() error {
	return a.eg.Wait()
}

func PrometheusMetricsHandler(prometheusRegistry *prometheus.Registry, logger slog.Logger) http.Handler {
	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		w.Header().Set("Content-Type", "text/plain")

		// Based on: https://github.com/tailscale/tailscale/blob/280255acae604796a1113861f5a84e6fa2dc6121/ipn/localapi/localapi.go#L489
		clientmetric.WritePrometheusExpositionFormat(w)

		metricFamilies, err := prometheusRegistry.Gather()
		if err != nil {
			logger.Error(context.Background(), "prometheus handler failed to gather metric families", slog.Error(err))
			return
		}

		for _, metricFamily := range metricFamilies {
			_, err = expfmt.MetricFamilyToText(w, metricFamily)
			if err != nil {
				logger.Error(context.Background(), "expfmt.MetricFamilyToText failed", slog.Error(err))
				return
			}
		}
	})
}

// childOOMScore returns the oom_score_adj for a child process. It is based
// on the oom_score_adj of the agent process.
func childOOMScore(agentScore int) int {
	// If the agent has a negative oom_score_adj, we set the child to 0
	// so it's treated like every other process.
	if agentScore < 0 {
		return 0
	}

	// If the agent is already almost at the maximum then set it to the max.
	if agentScore >= 998 {
		return 1000
	}

	// If the agent oom_score_adj is >=0, we set the child to slightly
	// less than the maximum. If users want a different score they set it
	// directly.
	return 998
}

func (a *agent) getAgentOOMScore() (int, error) {
	scoreStr, err := afero.ReadFile(a.filesystem, "/proc/self/oom_score_adj")
	if err != nil {
		return 0, xerrors.Errorf("read file: %w", err)
	}

	score, err := strconv.Atoi(strings.TrimSpace(string(scoreStr)))
	if err != nil {
		return 0, xerrors.Errorf("parse int: %w", err)
	}

	return score, nil
}

// isCustomOOMScore checks to see if the oom_score_adj is not a value that would
// originate from an agent-spawned process.
func isCustomOOMScore(agentScore int, process *agentproc.Process) bool {
	score := process.OOMScoreAdj
	return agentScore != score && score != 1000 && score != 0 && score != 998
}

// logDebouncer skips writing a log for a particular message if
// it's been emitted within the given interval duration.
// It's a shoddy implementation used in one spot that should be replaced at
// some point.
type logDebouncer struct {
	logger   slog.Logger
	messages map[string]time.Time
	interval time.Duration
}

func (l *logDebouncer) Warn(ctx context.Context, msg string, fields ...any) {
	l.log(ctx, slog.LevelWarn, msg, fields...)
}

func (l *logDebouncer) Error(ctx context.Context, msg string, fields ...any) {
	l.log(ctx, slog.LevelError, msg, fields...)
}

func (l *logDebouncer) log(ctx context.Context, level slog.Level, msg string, fields ...any) {
	// This (bad) implementation assumes you wouldn't reuse the same msg
	// for different levels.
	if last, ok := l.messages[msg]; ok && time.Since(last) < l.interval {
		return
	}
	switch level {
	case slog.LevelWarn:
		l.logger.Warn(ctx, msg, fields...)
	case slog.LevelError:
		l.logger.Error(ctx, msg, fields...)
	}
	l.messages[msg] = time.Now()
}