Files
kandev/apps/backend/internal/agent/runtime/lifecycle/manager.go
T
João Salavisa 103878047c fix(auth): scope in-session agent MCP calls to the task owner (#1937)
* fix(auth): scope in-session agent MCP calls to the task owner

The MCP tools an agent gets automatically inside its own session are
relayed over the agent's WebSocket stream, which carries no credential.
The backend dispatched them on the raw stream context, so the task
service saw no identity — its signal for an internal caller (event bus,
pollers, office schedulers) — and served the request unscoped. An agent
that supplied another user's task_id or workflow_id was given their
data, violating the rule that knowing an ID must not grant access.

Resolve the owning user from the stream's own AgentExecution (task ->
workspace -> owner) and attach that real stored identity before
dispatch, so the existing authorize* checks apply exactly as they do on
the PAT-authenticated /mcp endpoint. The owning task never comes from
the agent-supplied payload: honoring a payload session_id would let an
agent name another user's session and inherit their identity.

Tool handlers are unchanged. Auth disabled stays unscoped, unclaimed
pre-auth workspaces stay visible to everyone, and a lookup failure
denies the dispatch rather than falling back to full access.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(auth): never leave in-session MCP dispatch unscoped

Review follow-up. Three paths still returned an identity-free context
under enforced auth, and "no identity" is exactly what the task service
reads as an internal caller and serves unscoped:

- Unowned workspace. CreateWorkspace only stamps an owner for scoped
  callers, so internal callers keep producing owner_id='' rows after
  setup. Those streams got full cross-user access. Scope them to a
  sentinel user ID no account can hold: the service's "empty owner_id or
  my own ID" rule then limits them to the unowned rows that are public
  by the compatibility contract, and nothing else.
- Task with no workspace. Same treatment, same reasoning.
- Missing task or workspace row. A lookup that reports not-found as
  (nil, nil) fell through to unscoped; treat it as a resolution failure.

Also deny, rather than fabricate a member identity, when the owner's
account is gone or disabled. Disabling a user revokes their sessions and
PATs, so a still-running agent session was the one remaining way into
their workspace.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Kandev Agent <agent@kandev.dev>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-07-25 09:56:47 +01:00

462 lines
19 KiB
Go

// Package lifecycle manages agent instance lifecycles including tracking,
// state transitions, and cleanup.
package lifecycle
import (
"context"
"sync"
"sync/atomic"
"time"
"go.uber.org/zap"
"golang.org/x/sync/singleflight"
"github.com/kandev/kandev/internal/agent/docker"
"github.com/kandev/kandev/internal/agent/executor"
"github.com/kandev/kandev/internal/agent/registry"
"github.com/kandev/kandev/internal/agent/runtime/activity"
agentctl "github.com/kandev/kandev/internal/agent/runtime/agentctl"
"github.com/kandev/kandev/internal/agent/runtime/routingerr"
"github.com/kandev/kandev/internal/common/logger"
"github.com/kandev/kandev/internal/events/bus"
"github.com/kandev/kandev/internal/secrets"
"github.com/kandev/kandev/internal/task/models"
"github.com/kandev/kandev/internal/worktree"
)
// ExecutorFallbackPolicy controls behavior when a requested runtime is unavailable.
type ExecutorFallbackPolicy string
const (
// ExecutorFallbackAllow silently falls back to the default runtime (current behavior).
ExecutorFallbackAllow ExecutorFallbackPolicy = "allow"
// ExecutorFallbackWarn falls back but logs a warning (current behavior, explicit).
ExecutorFallbackWarn ExecutorFallbackPolicy = "warn"
// ExecutorFallbackDeny returns an error if the requested runtime is unavailable.
ExecutorFallbackDeny ExecutorFallbackPolicy = "deny"
)
// Manager manages agent instance lifecycles
type Manager struct {
registry *registry.Registry
eventBus bus.EventBus
credsMgr CredentialsManager
profileResolver ProfileResolver
worktreeMgr *worktree.Manager
mcpProvider McpConfigProvider
logger *logger.Logger
// dataDir is the kandev root directory. Misnamed for historical reasons:
// cmd/kandev/agents.go passes cfg.ResolvedHomeDir() (the kandev root —
// typically ~/.kandev) here, not ResolvedDataDir(). Used for:
// - Session history storage (SessionHistoryManager)
// - Ephemeral workspace creation (quick chat) at <root>/quick-chat/<sessionID>
// - Scratch workspaces for repo-less tasks at <root>/tasks/<workspaceID>/<taskID>
dataDir string
// ExecutorRegistry manages multiple runtimes (Docker, Standalone, etc.)
// Each task can select its runtime based on executor type.
executorRegistry *ExecutorRegistry
// executorFallbackPolicy controls behavior when a requested runtime is unavailable.
executorFallbackPolicy ExecutorFallbackPolicy
// Refactored components for separation of concerns
executionStore *ExecutionStore // Thread-safe execution tracking
commandBuilder *CommandBuilder // Builds agent commands from registry config
sessionManager *SessionManager // Handles ACP session initialization
streamManager *StreamManager // Manages WebSocket streams
eventPublisher *EventPublisher // Publishes lifecycle events
historyManager *SessionHistoryManager // Stores session history for context injection (fork_session pattern)
// Workspace info provider for on-demand instance creation
workspaceInfoProvider WorkspaceInfoProvider
// bootMessageService creates boot messages displayed in chat during agent startup.
bootMessageService BootMessageService
// preparerRegistry maps executor types to environment preparers.
preparerRegistry *PreparerRegistry
// mcpHandler is the MCP request dispatcher for handling MCP requests
// from agentctl instances through the agent stream.
mcpHandler agentctl.MCPHandler
// sessionAccessCheck enforces per-user workspace scoping on session-scoped
// surfaces (opt-in auth). Nil = no scoping. See SetSessionAccessChecker.
sessionAccessCheck func(ctx context.Context, sessionID string) error
// environmentAccessCheck is the environment-keyed sibling of
// sessionAccessCheck, used by the terminal environment-shell route which
// resolves executions by environment ID. Nil = no scoping.
environmentAccessCheck func(ctx context.Context, environmentID string) error
// singleflight deduplicates concurrent GetOrEnsureExecution calls for the same session
ensureExecutionGroup singleflight.Group
// Background remote status polling
remoteStatusPollInterval time.Duration
remoteStatusMu sync.RWMutex
remoteStatusBySession map[string]*RemoteStatus
stopCh chan struct{}
stopOnce sync.Once
wg sync.WaitGroup
// shuttingDown is flipped true when graceful shutdown begins (see
// StopAllAgents) so handlers running in detached goroutines can
// short-circuit work that would otherwise race the teardown and log
// confusing errors against children already being stopped.
shuttingDown atomic.Bool
// pollAggregator routes hub session-mode events to agentctl. See
// manager_subscription.go.
pollAggregator *workspacePollAggregator
// secretStore encrypts/decrypts runtime auth tokens (e.g., agentctl handshake tokens).
// Used to persist tokens across backend restarts for remote executor recovery.
secretStore secrets.SecretStore
// runningWriter persists the executors_running row in lockstep with executionStore.
// See SetExecutorRunningWriter and persistence.go. The lifecycle manager is the
// only component allowed to write the lifecycle-owned columns of this table.
runningWriter ExecutorRunningWriter
// agentProfileReader resolves the full agent_profiles row (including the
// office-enrichment fields added in ADR 0005 Wave A) for the launch-prep
// SkillDeployer hook. Nil → skill deploy is skipped.
agentProfileReader AgentProfileReader
// skillDeployer materialises per-profile skills + custom prompt before
// the agent process starts. Defaults to a no-op deployer; office wires
// its concrete implementation via SetSkillDeployer.
skillDeployer SkillDeployer
// remediateNpxCache is the hook fired when the routing classifier
// returns CodeNpxCacheCorrupted. NewManager wires routingerr.RemediateNpxCache;
// tests override it to avoid touching the real filesystem.
remediateNpxCache func(path string, log *zap.Logger) error
// standaloneHostPID is the OS process id of the standalone agentctl
// control-server this backend spawned on the local host. It is the
// host-local liveness handle recorded in executors_running.local_pid for
// local/standalone rows (see persistence.go / #1597 truthful executor rows).
// 0 when unset (tests, or before the launcher wires it). Never used for
// SSH/remote rows — their process lives on another host.
standaloneHostPID atomic.Int64
// managedGoCache provides the opt-in GOCACHE for host-local executions.
// System storage wiring installs it after settings persistence is ready.
managedGoCache ManagedGoCacheEnvironmentProvider
activityCoordinator *activity.Coordinator
activityMu sync.Mutex
activityLeases map[string]*activity.TaskLease
activityLeaseOwners map[string]uint64
activityPending map[string]map[uint64]*executionActivityClaim
activityGeneration uint64
}
// ManagedGoCacheEnvironmentProvider supplies the environment for one new
// local execution. Implementations must return an absolute GOCACHE path.
type ManagedGoCacheEnvironmentProvider interface {
ExecutionEnvironment(ctx context.Context) (map[string]string, error)
}
// SetManagedGoCacheEnvironmentProvider wires install-wide managed cache settings.
func (m *Manager) SetManagedGoCacheEnvironmentProvider(provider ManagedGoCacheEnvironmentProvider) {
m.managedGoCache = provider
}
// SetActivityCoordinator wires the install-wide host-resource activity gate.
// It is optional so embedded and test configurations retain legacy behavior.
func (m *Manager) SetActivityCoordinator(coordinator *activity.Coordinator) {
m.activityMu.Lock()
m.activityCoordinator = coordinator
if m.activityLeases == nil {
m.activityLeases = make(map[string]*activity.TaskLease)
}
if m.activityLeaseOwners == nil {
m.activityLeaseOwners = make(map[string]uint64)
}
if m.activityPending == nil {
m.activityPending = make(map[string]map[uint64]*executionActivityClaim)
}
m.activityMu.Unlock()
if m.executorRegistry == nil {
return
}
backend, err := m.executorRegistry.GetBackend(executor.NameDocker)
if err != nil {
return
}
if dockerExecutor, ok := backend.(*DockerExecutor); ok {
dockerExecutor.SetActivityCoordinator(coordinator)
}
}
// SetStandaloneHostPID records the local agentctl control-server PID so
// local/standalone executor rows can carry a real host-local liveness handle.
// Wired during DI from the agentctl launcher (see backendapp). Safe to leave
// unset in tests that don't exercise the persistence path.
func (m *Manager) SetStandaloneHostPID(pid int) {
m.standaloneHostPID.Store(int64(pid))
}
// NewManager creates a new lifecycle manager.
// The executorRegistry manages multiple runtimes (Docker, Standalone, etc.) for task-specific execution.
// The fallbackPolicy controls behavior when a requested runtime is unavailable.
func NewManager(
reg *registry.Registry,
eventBus bus.EventBus,
executorRegistry *ExecutorRegistry,
credsMgr CredentialsManager,
profileResolver ProfileResolver,
mcpProvider McpConfigProvider,
fallbackPolicy ExecutorFallbackPolicy,
dataDir string,
log *logger.Logger,
) *Manager {
componentLogger := log.WithFields(zap.String("component", "lifecycle-manager"))
// Initialize command builder
commandBuilder := NewCommandBuilder()
// Create stop channel for graceful shutdown
stopCh := make(chan struct{})
// Initialize session manager
sessionManager := NewSessionManager(log, stopCh)
// Initialize event publisher
eventPublisher := NewEventPublisher(eventBus, log)
// Initialize execution store
executionStore := NewExecutionStore()
// Initialize session history manager for fork_session pattern (context injection)
historyManager, err := NewSessionHistoryManager("", dataDir, log)
if err != nil {
log.Warn("failed to create session history manager, context injection disabled", zap.Error(err))
}
mgr := &Manager{
registry: reg,
eventBus: eventBus,
executorRegistry: executorRegistry,
executorFallbackPolicy: fallbackPolicy,
credsMgr: credsMgr,
profileResolver: profileResolver,
mcpProvider: mcpProvider,
logger: componentLogger,
dataDir: dataDir,
executionStore: executionStore,
commandBuilder: commandBuilder,
sessionManager: sessionManager,
eventPublisher: eventPublisher,
historyManager: historyManager,
remoteStatusPollInterval: 60 * time.Second,
remoteStatusBySession: make(map[string]*RemoteStatus),
stopCh: stopCh,
skillDeployer: NoopSkillDeployer(),
remediateNpxCache: routingerr.RemediateNpxCache,
}
// Initialize stream manager with callbacks that delegate to manager methods
// mcpHandler will be set later via SetMCPHandler.
// stopCh is shared with the manager so workspace-stream backoff drains on Stop.
mgr.streamManager = NewStreamManager(log, StreamCallbacks{
OnAgentEvent: mgr.handleAgentEvent,
OnStreamDisconnect: mgr.handleStreamDisconnect,
OnGitStatus: mgr.handleGitStatusUpdate,
OnGitCommit: mgr.handleGitCommitCreated,
OnGitReset: mgr.handleGitResetDetected,
OnBranchSwitch: mgr.handleBranchSwitch,
OnFileChange: mgr.handleFileChangeNotification,
OnShellOutput: mgr.handleShellOutput,
OnShellExit: mgr.handleShellExit,
OnProcessOutput: mgr.handleProcessOutput,
OnProcessStatus: mgr.handleProcessStatus,
}, nil, stopCh)
// Set session manager dependencies for full orchestration
sessionManager.SetDependencies(eventPublisher, mgr.streamManager, executionStore, historyManager)
sessionManager.SetPromptStarter(mgr.BeginPrompt)
mgr.pollAggregator = newWorkspacePollAggregator(mgr)
if executorRegistry != nil {
mgr.logger.Info("initialized with runtimes", zap.Int("count", len(executorRegistry.List())))
}
return mgr
}
// HandleSessionMode routes a session-level mode transition (from the gateway
// hub) into the per-workspace aggregator, which pushes the resulting
// workspace-effective mode to agentctl. See manager_subscription.go.
func (m *Manager) HandleSessionMode(sessionID string, mode WorkspacePollMode) {
if m.pollAggregator == nil {
return
}
m.pollAggregator.HandleSessionMode(sessionID, mode)
}
// flushCachedPollMode pushes any session mode the gateway cached before this
// execution was ready. Closes the pre-execution-focus race where the frontend
// sent session.focus during execution startup and the cached mode never
// reached agentctl. No-op when nothing was cached.
func (m *Manager) flushCachedPollMode(sessionID string) {
if m.pollAggregator == nil {
return
}
m.pollAggregator.FlushSessionMode(sessionID)
}
// SetWorktreeManager sets the worktree manager for Git worktree isolation.
//
// This must be called before launching agents if Git worktree support is enabled in the runtime.
// The worktree manager creates isolated Git working directories for each agent execution,
// allowing multiple agents to work on the same repository without conflicts.
//
// Call this during initialization, typically when setting up the orchestrator service.
// If not set, agents will work directly in the repository's main working directory.
func (m *Manager) SetWorktreeManager(worktreeMgr *worktree.Manager) {
m.worktreeMgr = worktreeMgr
// Register the worktree preparer so that executor type "worktree" gets
// worktree-specific preparation (create git worktree, checkout PR branch)
// instead of the generic local preparer.
if m.preparerRegistry != nil {
m.preparerRegistry.Register(models.ExecutorTypeWorktree, NewWorktreePreparer(worktreeMgr, m.logger))
}
}
// WorktreeManager returns the configured worktree manager. Returns nil
// when worktree support has not been wired (legacy / tests). Used by
// the office task-handoffs cleaner to translate worktree IDs into
// disk operations.
func (m *Manager) WorktreeManager() *worktree.Manager {
return m.worktreeMgr
}
// SetMCPHandler sets the MCP request handler for dispatching MCP tool calls.
//
// MCP requests from agents flow through the agent stream (WebSocket) to the backend,
// where they are dispatched to this handler. This enables agents to use MCP tools
// like listing workspaces, boards, tasks, and asking user questions.
//
// This must be called before agents start making MCP calls. Typically set during
// initialization after the MCP handlers are created.
func (m *Manager) SetMCPHandler(handler agentctl.MCPHandler) {
m.mcpHandler = handler
// Update the stream manager with the handler
m.streamManager.mcpHandler = handler
}
// SetMCPIdentityScoper installs the per-user scoping hook for in-session MCP
// tool calls.
//
// Unlike the external /mcp endpoint — where the agent presents a personal
// access token and the auth middleware resolves the identity — MCP requests
// relayed over an agent's own stream carry no credential. Without this hook
// they reach the task service with no identity, which that service reads as an
// internal caller and serves unscoped, so an agent supplying another user's
// task_id or workflow_id would be given their data.
//
// Set once during startup wiring, before agents start making MCP calls. Leave
// unset to keep dispatch unscoped (single-user instances).
func (m *Manager) SetMCPIdentityScoper(scoper MCPIdentityScoper) {
m.streamManager.mcpIdentityScoper = scoper
}
// SetSessionAccessChecker installs the per-user session visibility check used
// by GetOrEnsureExecution and EnsurePassthroughExecution. The checker must
// return nil for contexts without a request identity (internal callers). Set
// once during startup wiring, before the HTTP server accepts connections.
func (m *Manager) SetSessionAccessChecker(check func(ctx context.Context, sessionID string) error) {
m.sessionAccessCheck = check
}
// SetEnvironmentAccessChecker installs the per-user environment visibility
// check used by GetOrEnsureExecutionForEnvironment (terminal env-shell route).
func (m *Manager) SetEnvironmentAccessChecker(check func(ctx context.Context, environmentID string) error) {
m.environmentAccessCheck = check
}
// CheckSessionAccess authorizes a session-scoped operation for the ctx
// identity. Handlers that resolve an execution by a bare in-memory lookup
// (vscode/port reverse proxies) must call this before serving, since only the
// GetOrEnsure* paths run the check internally. No-op when no checker is set.
func (m *Manager) CheckSessionAccess(ctx context.Context, sessionID string) error {
if m.sessionAccessCheck == nil {
return nil
}
return m.sessionAccessCheck(ctx, sessionID)
}
// SetWorkspaceInfoProvider sets the provider for workspace information.
//
// The WorkspaceInfoProvider interface allows the lifecycle manager to dynamically create
// agent executions on-demand when the frontend connects to a session that doesn't have
// an active execution yet. This enables session resume after server restart or when
// accessing a session via URL (/task/[id]/[sessionId]).
//
// The provider must implement:
// - GetWorkspaceInfoBySessionID(ctx, sessionID) - Returns workspace path, worktree info,
// and MCP servers configured for the session
//
// This is typically called during initialization, with the task service as the provider.
// Without this, EnsureWorkspaceExecutionForSession will fail.
func (m *Manager) SetWorkspaceInfoProvider(provider WorkspaceInfoProvider) {
m.workspaceInfoProvider = provider
}
// SetBootMessageService sets the service used to create boot messages in chat
// during agent startup. If not set, no boot messages will be created.
func (m *Manager) SetBootMessageService(svc BootMessageService) {
m.bootMessageService = svc
}
// SetPreparerRegistry sets the registry of environment preparers.
func (m *Manager) SetPreparerRegistry(registry *PreparerRegistry) {
m.preparerRegistry = registry
}
// SetSecretStore sets the secret store for encrypting runtime auth tokens.
func (m *Manager) SetSecretStore(store secrets.SecretStore) {
m.secretStore = store
}
// SetAgentProfileReader wires the reader the launch-prep SkillDeployer uses
// to resolve full agent_profiles rows by id. Without it, the skill deploy
// step is silently skipped.
func (m *Manager) SetAgentProfileReader(reader AgentProfileReader) {
m.agentProfileReader = reader
}
// SetSkillDeployer plugs in a concrete deployer that materialises per-profile
// skills + custom prompt into the workspace before launch. Default is a
// no-op deployer; office wires its real implementation here.
func (m *Manager) SetSkillDeployer(deployer SkillDeployer) {
if deployer == nil {
m.skillDeployer = NoopSkillDeployer()
return
}
m.skillDeployer = deployer
}
// DockerClientProvider returns a function that lazily resolves the Docker client
// from the Docker executor in the registry. Returns nil if Docker is unavailable.
func (m *Manager) DockerClientProvider() func() *docker.Client {
return func() *docker.Client {
if m.executorRegistry == nil {
return nil
}
backend, err := m.executorRegistry.GetBackend(executor.NameDocker)
if err != nil {
return nil
}
dockerExec, ok := backend.(*DockerExecutor)
if !ok {
return nil
}
return dockerExec.Client()
}
}