kandev/apps/backend/internal/orchestrator/session_launch.go

package orchestrator

import (
	"context"
	"errors"
	"fmt"
	"strings"

	"go.uber.org/zap"

	"github.com/kandev/kandev/internal/orchestrator/executor"
	"github.com/kandev/kandev/internal/task/models"
	wfmodels "github.com/kandev/kandev/internal/workflow/models"
	v1 "github.com/kandev/kandev/pkg/api/v1"
)

// sessionTerminalErrText mirrors lifecycle.ErrSessionTerminal's message. It is
// duplicated as a string rather than imported so this higher-level orchestrator
// file does not take a direct dependency on internal/agent/runtime/lifecycle
// (ARCH-RUNTIME-IMPORT): the launch failures reach here only as wrapped-error
// strings or a stringified persisted session error, so a string match is both
// sufficient and required (see IsBenignLaunchTeardownErr).
const sessionTerminalErrText = "session is terminal"

// SessionIntent represents the type of session operation requested.
type SessionIntent string

const (
	IntentPrepare          SessionIntent = "prepare"           // Create session, optionally launch workspace, NO agent
	IntentStart            SessionIntent = "start"             // Create session + launch agent (new session)
	IntentStartCreated     SessionIntent = "start_created"     // Start agent on existing CREATED session
	IntentResume           SessionIntent = "resume"            // Restart stopped session with resume token
	IntentWorkflowStep     SessionIntent = "workflow_step"     // Start session with workflow step prompt config
	IntentRestoreWorkspace SessionIntent = "restore_workspace" // Restore workspace access for terminal-state session
)

// LaunchSessionRequest is the unified request for session.launch.
type LaunchSessionRequest struct {
	TaskID            string        `json:"task_id"`
	Intent            SessionIntent `json:"intent,omitempty"`
	SessionID         string        `json:"session_id,omitempty"`
	AgentProfileID    string        `json:"agent_profile_id,omitempty"`
	ExecutorID        string        `json:"executor_id,omitempty"`
	ExecutorProfileID string        `json:"executor_profile_id,omitempty"`
	Prompt            string        `json:"prompt,omitempty"`
	PlanMode          bool          `json:"plan_mode,omitempty"`
	WorkflowStepID    string        `json:"workflow_step_id,omitempty"`
	Priority          string        `json:"priority,omitempty"`
	LaunchWorkspace   bool          `json:"launch_workspace,omitempty"`
	SkipMessageRecord bool          `json:"skip_message_record,omitempty"`
	AutoStart         bool          `json:"auto_start,omitempty"`
	// NoAgentLaunch marks a prepare request that must NEVER be upgraded into an
	// agent launch, even for passthrough profiles (whose prepare would normally
	// be eagerly upgraded so the PTY exists). It backs the session.ensure
	// auto_start=false override used by the prevent-auto-start-on-open
	// preference: the session is created workspace-only (CREATED) and the
	// Start agent button launches it later. It is an internal server-side flag
	// set from EnsureSessionOptions, kept off the wire protocol (`json:"-"`).
	NoAgentLaunch bool `json:"-"`
	// DeferredStart marks a prepare whose caller will follow up with an explicit
	// IntentStartCreated that carries the prompt (the two-phase create flow:
	// cheap sync prepare + async start). It suppresses the passthrough
	// launchPrepare→launchStart upgrade so the eager launch doesn't spawn a
	// promptless PTY and pre-empt the prompt-bearing start. It is an internal
	// server-side coordination flag set by the deferred-start handlers, so it is
	// kept off the wire protocol (`json:"-"`) — a client must not be able to
	// suppress the upgrade and strand a passthrough session without a PTY.
	DeferredStart bool                   `json:"-"`
	Attachments   []v1.MessageAttachment `json:"attachments,omitempty"`
	// SpawnOrigin identifies the agent session that requested this launch via
	// spawn_session_kandev, so the new session's first turn can carry spawner
	// attribution and reply instructions. Like DeferredStart it is kept off the
	// wire protocol (`json:"-"`): the launch site turns it into a *trusted*
	// <kandev-system> block that survives first-turn canonicalization, so a WS
	// client must not be able to forge one and fabricate server authority.
	SpawnOrigin *SpawnOrigin `json:"-"`
}

// SpawnOrigin describes the agent session that spawned a new sibling session.
// The identifiers are resolved server-side by the MCP layer from the calling
// agent's own session, never read from the tool arguments.
type SpawnOrigin struct {
	TaskID      string
	SessionID   string
	SessionName string
}

// LaunchSessionResponse is the unified response for session.launch.
type LaunchSessionResponse struct {
	Success          bool    `json:"success"`
	TaskID           string  `json:"task_id"`
	SessionID        string  `json:"session_id,omitempty"`
	AgentExecutionID string  `json:"agent_execution_id,omitempty"`
	AgentProfileID   string  `json:"agent_profile_id,omitempty"`
	State            string  `json:"state"`
	WorktreePath     *string `json:"worktree_path,omitempty"`
	WorktreeBranch   *string `json:"worktree_branch,omitempty"`
}

// ResolveIntent infers the session intent from request fields when Intent is empty.
func ResolveIntent(req *LaunchSessionRequest) SessionIntent {
	if req.Intent != "" {
		return req.Intent
	}
	if req.SessionID != "" && req.WorkflowStepID != "" {
		return IntentWorkflowStep
	}
	if req.SessionID != "" && req.Prompt == "" && req.AgentProfileID == "" {
		return IntentResume
	}
	if req.SessionID != "" {
		return IntentStartCreated
	}
	if req.LaunchWorkspace && req.Prompt == "" {
		return IntentPrepare
	}
	return IntentStart
}

// IsBenignLaunchTeardownErr reports whether a session-launch failure is an
// expected graceful-shutdown teardown race rather than a genuine fault. During
// shutdown the root context is cancelled and terminal sessions reject launches,
// so an in-flight session.launch fails predictably; those should log WARN
// without a stack trace, not ERROR.
//
// Two shapes reach the launch handler on shutdown:
//   - restore_workspace wraps lifecycle.ErrSessionTerminal with %w and the
//     cancelled root context surfaces context.Canceled, so errors.Is matches
//     the context sentinel.
//   - resume stringifies a persisted session error (task_operations.go uses %s
//     on sess.ErrorMessage), destroying any sentinel, so a bounded string
//     fallback for "context canceled"/"session is terminal" is required. The
//     terminal-session text is matched as a string (not errors.Is) to avoid a
//     higher-level import of the runtime/lifecycle seam.
func IsBenignLaunchTeardownErr(err error) bool {
	if err == nil {
		return false
	}
	if errors.Is(err, context.Canceled) {
		return true
	}
	msg := err.Error()
	return strings.Contains(msg, context.Canceled.Error()) ||
		strings.Contains(msg, sessionTerminalErrText)
}

// LaunchSession is the unified entry point for all session operations.
func (s *Service) LaunchSession(ctx context.Context, req *LaunchSessionRequest) (*LaunchSessionResponse, error) {
	// Every intent funnels through here. SessionID is empty when creating, so
	// that case is carried by the task check alone.
	if err := s.authorizeTask(ctx, req.TaskID); err != nil {
		return nil, err
	}
	if err := s.authorizeSession(ctx, req.SessionID); err != nil {
		return nil, err
	}
	intent := ResolveIntent(req)
	req.Prompt = strings.TrimSpace(req.Prompt)

	switch intent {
	case IntentPrepare:
		return s.launchPrepare(ctx, req)
	case IntentStart:
		return s.launchStart(ctx, req)
	case IntentStartCreated:
		return s.launchStartCreated(ctx, req)
	case IntentResume:
		return s.launchResume(ctx, req)
	case IntentWorkflowStep:
		return s.launchWorkflowStep(ctx, req)
	case IntentRestoreWorkspace:
		return s.launchRestoreWorkspace(ctx, req)
	default:
		return nil, fmt.Errorf("unknown intent: %s", intent)
	}
}

// launchPrepare creates a session entry without launching the agent.
// Passthrough profiles can't be "prepared" without a running PTY — the terminal
// has nothing to attach to until the agent process exists. Upgrade those calls
// to a full start so the PTY is ready by the time the user sees the terminal.
//
// AutoStart=true means we arrived here from launchStart's blocked-auto-start
// downgrade path; skipping the upgrade in that case avoids a launchStart ↔
// launchPrepare bounce.
//
// DeferredStart=true means a prompt-bearing IntentStartCreated will follow this
// prepare (the two-phase create flow); skipping the upgrade there leaves the
// session CREATED so that follow-up start launches the passthrough agent WITH
// the prompt — eagerly launching here would spawn a promptless PTY and the
// later start would be rejected against the now-running session.
func (s *Service) launchPrepare(ctx context.Context, req *LaunchSessionRequest) (*LaunchSessionResponse, error) {
	if s.shouldUpgradePassthroughPrepare(ctx, req) {
		return s.launchStart(ctx, req)
	}
	sessionID, err := s.PrepareTaskSession(
		ctx, req.TaskID, req.AgentProfileID, req.ExecutorID,
		req.ExecutorProfileID, req.WorkflowStepID, req.LaunchWorkspace,
	)
	if err != nil {
		return nil, err
	}
	return &LaunchSessionResponse{
		Success:   true,
		TaskID:    req.TaskID,
		SessionID: sessionID,
		State:     string(models.TaskSessionStateCreated),
	}, nil
}

// shouldUpgradePassthroughPrepare reports whether a prepare request for a
// passthrough profile should be eagerly upgraded to a full launch so a PTY
// exists for the terminal to attach to. It is the single decision point for the
// upgrade documented on launchPrepare: only genuine prepare-only callers (no
// imminent prompt-bearing start) get the eager launch. See launchPrepare for
// why AutoStart and DeferredStart each suppress it.
func (s *Service) shouldUpgradePassthroughPrepare(ctx context.Context, req *LaunchSessionRequest) bool {
	if req.NoAgentLaunch {
		return false
	}
	return !req.AutoStart && !req.DeferredStart && s.isPassthroughProfile(ctx, req.AgentProfileID)
}

// isPassthroughProfile reports whether the agent profile is a CLI
// passthrough provider.
func (s *Service) isPassthroughProfile(ctx context.Context, profileID string) bool {
	if profileID == "" || s.agentManager == nil {
		return false
	}
	info, err := s.agentManager.ResolveAgentProfile(ctx, profileID)
	if err != nil || info == nil {
		return false
	}
	return info.CLIPassthrough
}

// launchStart creates a new session and launches the agent.
// If the request is an auto-start and the task's current workflow step does not
// have auto_start_agent, the request is downgraded to a prepare (workspace-only,
// no agent) to prevent unwanted auto-starts from the frontend's useAutoStartSession hook.
func (s *Service) launchStart(ctx context.Context, req *LaunchSessionRequest) (*LaunchSessionResponse, error) {
	if req.AutoStart && s.shouldBlockAutoStart(ctx, req) {
		req.LaunchWorkspace = true
		return s.launchPrepare(ctx, req)
	}

	execution, err := s.startTask(
		ctx, req.TaskID, req.AgentProfileID, req.ExecutorID,
		req.ExecutorProfileID, req.Priority, req.Prompt,
		req.WorkflowStepID, req.PlanMode, req.AutoStart, req.Attachments,
		startTaskOptions{SpawnOrigin: req.SpawnOrigin},
	)
	if err != nil {
		return nil, err
	}
	return executionToLaunchResponse(req.TaskID, execution), nil
}

// shouldBlockAutoStart checks whether the task's workflow step allows auto-starting
// the agent. Returns true when the step exists but does not have auto_start_agent
// in its on_enter events. Tasks without a workflow step are never blocked.
func (s *Service) shouldBlockAutoStart(ctx context.Context, req *LaunchSessionRequest) bool {
	if s.workflowStepGetter == nil {
		return false
	}

	task, err := s.repo.GetTask(ctx, req.TaskID)
	if err != nil || task.WorkflowStepID == "" {
		return false
	}

	step, err := s.workflowStepGetter.GetStep(ctx, task.WorkflowStepID)
	if err != nil || step == nil {
		return false
	}

	if step.HasOnEnterAction(wfmodels.OnEnterAutoStartAgent) {
		return false
	}

	s.logger.Info("auto-start downgraded to prepare: step lacks auto_start_agent",
		zap.String("task_id", req.TaskID),
		zap.String("workflow_step_id", task.WorkflowStepID),
		zap.String("step_name", step.Name))

	return true
}

// launchStartCreated starts agent execution on an existing CREATED session.
func (s *Service) launchStartCreated(ctx context.Context, req *LaunchSessionRequest) (*LaunchSessionResponse, error) {
	execution, err := s.StartCreatedSession(
		ctx, req.TaskID, req.SessionID, req.AgentProfileID,
		req.Prompt, req.SkipMessageRecord, req.PlanMode, req.AutoStart, req.Attachments, nil,
	)
	if err != nil {
		return nil, err
	}
	return executionToLaunchResponse(req.TaskID, execution), nil
}

// launchResume resumes a stopped session.
func (s *Service) launchResume(ctx context.Context, req *LaunchSessionRequest) (*LaunchSessionResponse, error) {
	execution, err := s.ResumeTaskSession(ctx, req.TaskID, req.SessionID)
	if err != nil {
		return nil, err
	}
	return executionToLaunchResponse(req.TaskID, execution), nil
}

// launchWorkflowStep starts a session with workflow step prompt configuration.
func (s *Service) launchWorkflowStep(ctx context.Context, req *LaunchSessionRequest) (*LaunchSessionResponse, error) {
	err := s.StartSessionForWorkflowStep(ctx, req.TaskID, req.SessionID, req.WorkflowStepID)
	if err != nil {
		return nil, err
	}
	return &LaunchSessionResponse{
		Success:   true,
		TaskID:    req.TaskID,
		SessionID: req.SessionID,
		State:     string(v1.TaskSessionStateRunning),
	}, nil
}

// launchRestoreWorkspace restores workspace access for a terminal-state session (COMPLETED, FAILED, CANCELLED).
// It creates a lightweight agentctl execution so the frontend can browse files, open terminals, and view git status.
func (s *Service) launchRestoreWorkspace(ctx context.Context, req *LaunchSessionRequest) (*LaunchSessionResponse, error) {
	if req.SessionID == "" {
		return nil, fmt.Errorf("session_id is required for workspace restore")
	}

	session, err := s.repo.GetTaskSession(ctx, req.SessionID)
	if err != nil {
		return nil, fmt.Errorf("session not found: %w", err)
	}
	if session.TaskID != req.TaskID {
		return nil, fmt.Errorf("session does not belong to task")
	}

	if err := s.agentManager.EnsureWorkspaceExecutionForSession(ctx, req.TaskID, req.SessionID); err != nil {
		return nil, fmt.Errorf("failed to restore workspace: %w", err)
	}

	resp := &LaunchSessionResponse{
		Success:   true,
		TaskID:    req.TaskID,
		SessionID: req.SessionID,
		State:     string(session.State),
	}
	if len(session.Worktrees) > 0 {
		wt := session.Worktrees[0]
		if wt.WorktreePath != "" {
			resp.WorktreePath = &wt.WorktreePath
		}
		if wt.WorktreeBranch != "" {
			resp.WorktreeBranch = &wt.WorktreeBranch
		}
	}
	return resp, nil
}

// RecoverSession handles user-initiated recovery after an agent CLI failure.
// action is "resume" (retry with existing ACP session) or "fresh_start" (clear token, start fresh).
func (s *Service) RecoverSession(ctx context.Context, taskID, sessionID, action string) (*LaunchSessionResponse, error) {
	// Guard before the switch: "fresh_start" clears the resume token, so an
	// unauthorized call would mutate the session even if the launch failed.
	if err := s.authorizeSession(ctx, sessionID); err != nil {
		return nil, err
	}
	if err := s.authorizeTask(ctx, taskID); err != nil {
		return nil, err
	}
	switch action {
	case "fresh_start":
		s.clearResumeToken(ctx, sessionID)
	case "resume":
		// no-op — relaunch with existing resume token
	default:
		return nil, fmt.Errorf("invalid recovery action: %s", action)
	}

	resp, err := s.LaunchSession(ctx, &LaunchSessionRequest{
		TaskID:    taskID,
		SessionID: sessionID,
		Intent:    IntentResume,
	})
	if err != nil {
		return nil, normalizeRecoverSessionError(err)
	}
	return resp, nil
}

// normalizeRecoverSessionError maps a missing-profile resume failure to a
// user-actionable message.
func normalizeRecoverSessionError(err error) error {
	if err == nil {
		return nil
	}
	if isMissingProfileResumeError(err) {
		return fmt.Errorf("the agent profile used by this session was deleted; start a new session and choose an available agent profile: %w", err)
	}
	return err
}

// isMissingProfileResumeError reports whether the error indicates the
// session's agent profile no longer exists.
func isMissingProfileResumeError(err error) bool {
	msg := strings.ToLower(err.Error())
	return strings.Contains(msg, "failed to resolve agent profile") ||
		strings.Contains(msg, "agent profile not found")
}

// executionToLaunchResponse converts a TaskExecution to a LaunchSessionResponse.
func executionToLaunchResponse(taskID string, exec *executor.TaskExecution) *LaunchSessionResponse {
	resp := &LaunchSessionResponse{
		Success:          true,
		TaskID:           taskID,
		SessionID:        exec.SessionID,
		AgentExecutionID: exec.AgentExecutionID,
		AgentProfileID:   exec.AgentProfileID,
		State:            string(exec.SessionState),
	}
	if exec.WorktreePath != "" {
		resp.WorktreePath = &exec.WorktreePath
	}
	if exec.WorktreeBranch != "" {
		resp.WorktreeBranch = &exec.WorktreeBranch
	}
	return resp
}