kandev/apps/backend/internal/agent/runtime/lifecycle/manager_passthrough.go

package lifecycle

import (
	"context"
	"fmt"
	"os"
	"path/filepath"
	"strings"
	"time"

	"go.uber.org/zap"

	"github.com/kandev/kandev/internal/agent/agents"
	"github.com/kandev/kandev/internal/agent/executor"
	"github.com/kandev/kandev/internal/agent/mcpconfig"
	agentctl "github.com/kandev/kandev/internal/agent/runtime/agentctl"
	"github.com/kandev/kandev/internal/agent/settings/cliflags"
	"github.com/kandev/kandev/internal/agentctl/server/process"
	agentctltypes "github.com/kandev/kandev/internal/agentctl/types"
	"github.com/kandev/kandev/internal/events"
	v1 "github.com/kandev/kandev/pkg/api/v1"
)

// MarkPassthroughRunning marks a passthrough execution as running when user submits input.
// This is called when Enter key is detected in the terminal handler.
// It updates the execution status and publishes an AgentRunning event.
func (m *Manager) MarkPassthroughRunning(sessionID string) error {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		return fmt.Errorf("no agent execution found for session: %s", sessionID)
	}

	if execution.PassthroughProcessID == "" {
		return fmt.Errorf("session %s is not in passthrough mode", sessionID)
	}

	// Only publish if not already running (prevents duplicate events)
	if execution.Status != v1.AgentStatusRunning {
		if err := m.UpdateStatus(execution.ID, v1.AgentStatusRunning); err != nil {
			return err
		}
		m.eventPublisher.PublishAgentEvent(context.Background(), events.AgentRunning, execution)
	}

	return nil
}

// WritePassthroughStdin writes data to the agent process stdin in passthrough mode.
// Returns an error if the session is not in passthrough mode or if writing fails.
// Note: For terminal handler input, use MarkPassthroughRunning directly since
// the terminal handler writes to PTY directly for performance.
func (m *Manager) WritePassthroughStdin(ctx context.Context, sessionID string, data string) error {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		return fmt.Errorf("no agent execution found for session: %s", sessionID)
	}

	if execution.PassthroughProcessID == "" {
		return fmt.Errorf("session %s is not in passthrough mode", sessionID)
	}

	// Get the interactive runner from runtime
	interactiveRunner := m.GetInteractiveRunner()
	if interactiveRunner == nil {
		return fmt.Errorf("interactive runner not available")
	}

	// Write to stdin
	if err := interactiveRunner.WriteStdin(execution.PassthroughProcessID, data); err != nil {
		return err
	}

	return nil
}

// IsPassthroughSession checks if the given session is running in passthrough (PTY) mode.
func (m *Manager) IsPassthroughSession(ctx context.Context, sessionID string) bool {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		return false
	}
	return execution.PassthroughProcessID != ""
}

// ResizePassthroughPTY resizes the PTY for a passthrough process.
// Returns an error if the session is not in passthrough mode or if resizing fails.
func (m *Manager) ResizePassthroughPTY(ctx context.Context, sessionID string, cols, rows uint16) error {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		return fmt.Errorf("no agent execution found for session: %s", sessionID)
	}

	if execution.PassthroughProcessID == "" {
		return fmt.Errorf("session %s is not in passthrough mode", sessionID)
	}

	// Get the interactive runner from runtime
	interactiveRunner := m.GetInteractiveRunner()
	if interactiveRunner == nil {
		return fmt.Errorf("interactive runner not available")
	}

	return interactiveRunner.ResizeBySession(sessionID, cols, rows)
}

// GetPassthroughBuffer returns the buffered output from the passthrough process.
// This is used for new subscribers to catch up on output.
func (m *Manager) GetPassthroughBuffer(ctx context.Context, sessionID string) (string, error) {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		return "", fmt.Errorf("no agent execution found for session: %s", sessionID)
	}

	if execution.PassthroughProcessID == "" {
		return "", fmt.Errorf("session %s is not in passthrough mode", sessionID)
	}

	// Get the interactive runner from runtime
	interactiveRunner := m.GetInteractiveRunner()
	if interactiveRunner == nil {
		return "", fmt.Errorf("interactive runner not available")
	}

	chunks, ok := interactiveRunner.GetBuffer(execution.PassthroughProcessID)
	if !ok {
		return "", fmt.Errorf("passthrough process not found")
	}

	// Concatenate all chunks into a single string
	var buffer strings.Builder
	for _, chunk := range chunks {
		buffer.WriteString(chunk.Data)
	}

	return buffer.String(), nil
}

// buildPassthroughEnv builds the environment map for a passthrough session,
// including Kandev metadata and required credentials from the agent runtime config.
func (m *Manager) buildPassthroughEnv(ctx context.Context, execution *AgentExecution, requiredEnv []string) map[string]string {
	env := make(map[string]string)
	env["KANDEV_TASK_ID"] = execution.TaskID
	env["KANDEV_SESSION_ID"] = execution.SessionID
	env["KANDEV_AGENT_PROFILE_ID"] = execution.AgentProfileID
	m.mergeAgentProfileEnv(ctx, execution.AgentProfileID, env)
	if m.credsMgr != nil {
		for _, credKey := range requiredEnv {
			if value, err := m.credsMgr.GetCredentialValue(ctx, credKey); err == nil && value != "" {
				env[credKey] = value
			}
		}
	}
	// Merge env vars contributed by the passthrough MCP strategy (e.g. opencode's
	// OPENCODE_CONFIG). Set during command building in applyPassthroughMCP.
	for key, value := range getPassthroughMCPEnv(execution) {
		env[key] = value
	}
	return env
}

// startPassthroughShell starts the shell session for a passthrough execution.
// Non-fatal errors are logged with the provided warning message.
func (m *Manager) startPassthroughShell(ctx context.Context, execution *AgentExecution, shellWarnMsg string) {
	if execution.agentctl == nil {
		return
	}
	if err := execution.agentctl.StartShell(ctx); err != nil {
		m.logger.Warn(shellWarnMsg,
			zap.String("execution_id", execution.ID),
			zap.Error(err))
	} else {
		m.logger.Info("shell session started for passthrough mode",
			zap.String("execution_id", execution.ID))
	}
}

// resolvedPassthrough holds the agent config, passthrough config, runtime config, and profile
// info resolved from an execution. Used as the basis for building passthrough commands.
type resolvedPassthrough struct {
	agentID     string
	agentConfig agents.Agent
	agent       agents.PassthroughAgent
	pt          agents.PassthroughConfig
	rt          *agents.RuntimeConfig
	profile     *AgentProfileInfo
}

// resolvePassthroughAgent loads the agent config and profile for a passthrough execution.
// Shared by passthroughAgentCommand, freshPassthroughCommand, and ResumePassthroughSession.
func (m *Manager) resolvePassthroughAgent(ctx context.Context, execution *AgentExecution) (*resolvedPassthrough, error) {
	agentConfig, err := m.getAgentConfigForExecution(execution)
	if err != nil {
		return nil, fmt.Errorf("failed to get agent config: %w", err)
	}

	ptAgent, ok := agentConfig.(agents.PassthroughAgent)
	if !ok {
		return nil, fmt.Errorf("agent %s does not support passthrough mode", agentConfig.ID())
	}

	var profileInfo *AgentProfileInfo
	if m.profileResolver != nil && execution.AgentProfileID != "" {
		profileInfo, _ = m.profileResolver.ResolveProfile(ctx, execution.AgentProfileID)
	}

	return &resolvedPassthrough{
		agentID:     agentConfig.ID(),
		agentConfig: agentConfig,
		agent:       ptAgent,
		pt:          ptAgent.PassthroughConfig(),
		rt:          agentConfig.Runtime(),
		profile:     profileInfo,
	}, nil
}

// promptForPassthroughCommand returns the prompt that should be passed to
// BuildPassthroughCommand. When the agent uses idle-based auto-inject and has
// no PromptFlag, the prompt would otherwise be appended as a positional arg
// (putting TUIs like Claude into non-interactive `-p` mode and exiting before
// auto-inject fires). In that case we return "" so the prompt is delivered via
// PTY stdin in autoInjectInitialPrompt.
func promptForPassthroughCommand(pt agents.PassthroughConfig, taskDescription string) string {
	if pt.AutoInjectPrompt && pt.PromptFlag.IsEmpty() {
		return ""
	}
	return taskDescription
}

const (
	// metadataKeyPassthroughMCPFiles tracks config files kandev wrote for the
	// passthrough MCP injection so they can be removed when the execution ends.
	metadataKeyPassthroughMCPFiles = "passthrough_mcp_files"
	// metadataKeyPassthroughMCPEnv carries env vars the MCP strategy needs on the
	// agent process (e.g. opencode's OPENCODE_CONFIG), merged in buildPassthroughEnv.
	metadataKeyPassthroughMCPEnv = "passthrough_mcp_env"
	// kandevMCPServerName is the reserved name of kandev's own HTTP MCP server,
	// which exposes the task tools to the agent.
	kandevMCPServerName = "kandev"
)

// redactPassthroughArgs masks secret-bearing MCP override values before the
// command is logged. Codex injects MCP servers via `-c mcp_servers.<name>.<key>=<json>`
// argv (no file-based option), so env vars and HTTP headers — which commonly
// carry tokens — would otherwise be written verbatim into backend logs. The
// real (unredacted) args are still what's executed; only the log copy is masked.
func redactPassthroughArgs(args []string) []string {
	out := make([]string, len(args))
	for i, a := range args {
		out[i] = redactMCPArg(a)
	}
	return out
}

func redactMCPArg(arg string) string {
	if !strings.HasPrefix(arg, "mcp_servers.") {
		return arg
	}
	eq := strings.IndexByte(arg, '=')
	if eq < 0 {
		return arg
	}
	key := arg[:eq]
	if strings.HasSuffix(key, ".env") || strings.HasSuffix(key, ".http_headers") {
		return key + "=<redacted>"
	}
	return arg
}

func passthroughMCPConfigPort(execution *AgentExecution) int {
	if execution == nil {
		return 0
	}
	if execution.standalonePort > 0 {
		return execution.standalonePort
	}
	if execution.Metadata == nil {
		return 0
	}
	switch value := execution.Metadata["standalone_port"].(type) {
	case int:
		return value
	case int64:
		return int(value)
	case float64:
		return int(value)
	}
	return 0
}

func safePassthroughMCPConfigName(value string) string {
	if value == "" {
		return "session"
	}
	var out strings.Builder
	for _, r := range value {
		switch {
		case r >= 'a' && r <= 'z':
			out.WriteRune(r)
		case r >= 'A' && r <= 'Z':
			out.WriteRune(r)
		case r >= '0' && r <= '9':
			out.WriteRune(r)
		case r == '-' || r == '_' || r == '.':
			out.WriteRune(r)
		default:
			out.WriteByte('_')
		}
	}
	return out.String()
}

// applyPassthroughMCP resolves the session's MCP servers (kandev's own server
// plus the profile's configured servers), runs the agent's passthrough MCP
// strategy, materializes any config files, records them for cleanup, stores the
// strategy's env vars on the execution (merged later in buildPassthroughEnv),
// and returns the extra CLI args to append to the passthrough command. It is a
// no-op for agents that declare no strategy.
func (m *Manager) applyPassthroughMCP(ctx context.Context, execution *AgentExecution, pt agents.PassthroughConfig, agentConfig agents.Agent) ([]string, error) {
	if pt.MCPStrategy == nil {
		return nil, nil
	}
	// passthroughMCPServers always returns at least the kandev server (or an
	// error when the port is unavailable), so the strategy receives a non-empty
	// list; each strategy guards its own empty-after-filtering case.
	servers, err := m.passthroughMCPServers(ctx, execution, agentConfig)
	if err != nil {
		return nil, err
	}
	artifacts, err := pt.MCPStrategy.BuildPassthroughMCP(servers, m.passthroughMCPPaths(execution))
	if err != nil {
		return nil, fmt.Errorf("build passthrough MCP config: %w", err)
	}
	if err := m.writePassthroughMCPFiles(execution, artifacts.Files); err != nil {
		return nil, err
	}
	setPassthroughMCPEnv(execution, artifacts.Env)
	return artifacts.Args, nil
}

// passthroughMCPServers returns kandev's own HTTP MCP server followed by the
// profile's resolved MCP servers. The kandev server requires the standalone
// port; a profile server named "kandev" is dropped so it cannot shadow ours.
func (m *Manager) passthroughMCPServers(ctx context.Context, execution *AgentExecution, agentConfig agents.Agent) ([]agentctltypes.McpServer, error) {
	port := passthroughMCPConfigPort(execution)
	if port <= 0 {
		return nil, fmt.Errorf("standalone port unavailable for passthrough MCP config")
	}
	servers := []agentctltypes.McpServer{{
		Name: kandevMCPServerName,
		Type: string(mcpconfig.ServerTypeHTTP),
		URL:  fmt.Sprintf("http://localhost:%d/mcp", port),
	}}
	profileServers, err := m.resolveMcpServersWithParams(ctx, execution.AgentProfileID, execution.Metadata, agentConfig)
	if err != nil {
		return nil, err
	}
	for _, srv := range profileServers {
		if srv.Name == kandevMCPServerName {
			continue
		}
		servers = append(servers, srv)
	}
	return servers, nil
}

// passthroughMCPPaths computes the filesystem locations a strategy may use: a
// kandev-owned temp config path (for file+flag / file+env strategies) and the
// workspace dir (for project-local file strategies like Cursor).
func (m *Manager) passthroughMCPPaths(execution *AgentExecution) mcpconfig.PassthroughPaths {
	root := m.dataDir
	if root == "" {
		root = filepath.Join(os.TempDir(), "kandev")
	}
	name := safePassthroughMCPConfigName(execution.SessionID)
	if execution.SessionID == "" {
		name = safePassthroughMCPConfigName(execution.ID)
	}
	return mcpconfig.PassthroughPaths{
		TempConfigPath: filepath.Join(root, "passthrough-mcp", name+".json"),
		WorkspaceDir:   execution.WorkspacePath,
	}
}

// writePassthroughMCPFiles materializes the strategy's config files and records
// every file kandev OWNS (created or overwrote) for cleanup. Files merged into a
// pre-existing user file (Cursor) are not tracked — kandev must not delete the
// user's file on teardown.
func (m *Manager) writePassthroughMCPFiles(execution *AgentExecution, files []mcpconfig.PassthroughConfigFile) error {
	written := getPassthroughMCPFiles(execution)
	for _, f := range files {
		if f.Path == "" {
			continue
		}
		ok, err := m.materializePassthroughFile(execution, f)
		if err != nil {
			return err
		}
		if ok {
			written = appendUnique(written, f.Path)
		}
	}
	setPassthroughMCPFiles(execution, written)
	return nil
}

// materializePassthroughFile writes one config file and reports whether kandev
// OWNS the result (true = track for cleanup). It refuses to write through an
// existing symlink (a malicious repo could point it outside the worktree),
// guards against a symlinked parent escaping the worktree, and creates new files
// with O_EXCL. For MergeKey files (Cursor) that already exist, kandev's servers
// are merged into the user's file (preserving their entries) and the file is NOT
// tracked for cleanup since it is the user's.
func (m *Manager) materializePassthroughFile(execution *AgentExecution, f mcpconfig.PassthroughConfigFile) (bool, error) {
	if escapes, err := workspacePathEscapes(execution.WorkspacePath, f.Path); err != nil {
		return false, fmt.Errorf("validate passthrough MCP config path: %w", err)
	} else if escapes {
		m.logger.Warn("passthrough MCP config path escapes workspace via symlink; skipping",
			zap.String("path", f.Path))
		return false, nil
	}

	info, statErr := os.Lstat(f.Path)
	switch {
	case statErr == nil && info.Mode()&os.ModeSymlink != 0:
		// Never write through an existing symlink — it could redirect the write
		// outside the worktree. Applies to both merge and create.
		m.logger.Warn("passthrough MCP config is a symlink; leaving it untouched",
			zap.String("path", f.Path))
		return false, nil
	case statErr == nil && f.MergeKey != "":
		// Merge kandev's servers into the user's existing regular file. Not
		// tracked — it's the user's file; we only appended our entries.
		return false, m.mergePassthroughConfig(f)
	case statErr == nil:
		// Existing kandev-owned temp file (Claude/OpenCode) — overwrite it.
		if err := os.WriteFile(f.Path, f.Content, 0o600); err != nil {
			return false, fmt.Errorf("write passthrough MCP config: %w", err)
		}
		return true, nil
	case !os.IsNotExist(statErr):
		return false, fmt.Errorf("lstat passthrough MCP config: %w", statErr)
	default:
		if err := os.MkdirAll(filepath.Dir(f.Path), 0o700); err != nil {
			return false, fmt.Errorf("create passthrough MCP config dir: %w", err)
		}
		return m.writeFileNoFollow(f.Path, f.Content)
	}
}

// mergePassthroughConfig merges kandev's servers (f.Content's f.MergeKey object)
// into the existing regular file at f.Path, preserving the user's other entries.
// A malformed or unreadable existing file is left untouched (logged), never
// clobbered. The file is confirmed to be a regular file (not a symlink) by the
// caller before this runs.
func (m *Manager) mergePassthroughConfig(f mcpconfig.PassthroughConfigFile) error {
	existing, err := os.ReadFile(f.Path)
	if err != nil {
		m.logger.Warn("cannot read existing MCP config to merge; leaving it untouched",
			zap.String("path", f.Path), zap.Error(err))
		return nil
	}
	merged, err := mcpconfig.MergeJSONUnderKey(existing, f.Content, f.MergeKey)
	if err != nil {
		m.logger.Warn("cannot merge into existing MCP config; leaving it untouched",
			zap.String("path", f.Path), zap.Error(err))
		return nil
	}
	if err := os.WriteFile(f.Path, merged, 0o600); err != nil {
		return fmt.Errorf("write merged passthrough MCP config: %w", err)
	}
	return nil
}

// writeFileNoFollow creates path with O_EXCL so it never follows or overwrites
// an existing leaf (including a symlink). A concurrently-created file is treated
// as "leave it alone" rather than an error.
func (m *Manager) writeFileNoFollow(path string, content []byte) (bool, error) {
	file, err := os.OpenFile(path, os.O_WRONLY|os.O_CREATE|os.O_EXCL, 0o600)
	if err != nil {
		if os.IsExist(err) {
			m.logger.Info("passthrough MCP config appeared concurrently; leaving it untouched",
				zap.String("path", path))
			return false, nil
		}
		return false, fmt.Errorf("create passthrough MCP config: %w", err)
	}
	if _, werr := file.Write(content); werr != nil {
		_ = file.Close()
		// Remove the empty file so a later SkipIfExists probe doesn't see it and
		// silently skip writing the real config.
		_ = os.Remove(path)
		return false, fmt.Errorf("write passthrough MCP config: %w", werr)
	}
	if cerr := file.Close(); cerr != nil {
		_ = os.Remove(path)
		return false, fmt.Errorf("close passthrough MCP config: %w", cerr)
	}
	return true, nil
}

// workspacePathEscapes reports whether path — assumed to live under
// workspaceDir — would, after resolving symlinks on its deepest existing
// ancestor, land outside workspaceDir. Files not lexically under workspaceDir
// (kandev's own temp configs) are exempt and return false. An empty
// workspaceDir disables the check.
func workspacePathEscapes(workspaceDir, path string) (bool, error) {
	if workspaceDir == "" {
		return false, nil
	}
	cleanWS := filepath.Clean(workspaceDir)
	cleanPath := filepath.Clean(path)
	if rel, err := filepath.Rel(cleanWS, cleanPath); err != nil || rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) {
		return false, nil // not a workspace-relative file; not our concern
	}
	canonWS, err := filepath.EvalSymlinks(cleanWS)
	if err != nil {
		return false, fmt.Errorf("resolve workspace dir: %w", err)
	}
	// Resolve the deepest existing ancestor of the target (the leaf and some
	// parents may not exist yet — those get created as real dirs).
	ancestor := filepath.Dir(cleanPath)
	for {
		resolved, err := filepath.EvalSymlinks(ancestor)
		if err == nil {
			rel, relErr := filepath.Rel(canonWS, resolved)
			escaped := relErr != nil || rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator))
			return escaped, nil
		}
		if !os.IsNotExist(err) {
			return false, fmt.Errorf("resolve ancestor %q: %w", ancestor, err)
		}
		parent := filepath.Dir(ancestor)
		if parent == ancestor {
			return false, nil // reached filesystem root with nothing existing
		}
		ancestor = parent
	}
}

func (m *Manager) cleanupPassthroughMCPConfig(execution *AgentExecution) {
	for _, path := range getPassthroughMCPFiles(execution) {
		if err := os.Remove(path); err != nil && !os.IsNotExist(err) {
			m.logger.Warn("failed to remove passthrough MCP config",
				zap.String("path", path),
				zap.Error(err))
		}
	}
	if execution.Metadata != nil {
		delete(execution.Metadata, metadataKeyPassthroughMCPFiles)
		delete(execution.Metadata, metadataKeyPassthroughMCPEnv)
	}
}

func appendUnique(list []string, value string) []string {
	for _, v := range list {
		if v == value {
			return list
		}
	}
	return append(list, value)
}

// getPassthroughMCPFiles reads the recorded config-file list, tolerating both
// []string (in-memory) and []interface{} (JSON-decoded after a restart).
func getPassthroughMCPFiles(execution *AgentExecution) []string {
	if execution == nil || execution.Metadata == nil {
		return nil
	}
	switch v := execution.Metadata[metadataKeyPassthroughMCPFiles].(type) {
	case []string:
		return append([]string(nil), v...)
	case []interface{}:
		out := make([]string, 0, len(v))
		for _, item := range v {
			if s, ok := item.(string); ok {
				out = append(out, s)
			}
		}
		return out
	default:
		return nil
	}
}

func setPassthroughMCPFiles(execution *AgentExecution, files []string) {
	if execution.Metadata == nil {
		execution.Metadata = map[string]interface{}{}
	}
	execution.Metadata[metadataKeyPassthroughMCPFiles] = files
}

// getPassthroughMCPEnv reads the recorded MCP env map, tolerating both
// map[string]string and map[string]interface{} (JSON-decoded after a restart).
func getPassthroughMCPEnv(execution *AgentExecution) map[string]string {
	if execution == nil || execution.Metadata == nil {
		return nil
	}
	switch v := execution.Metadata[metadataKeyPassthroughMCPEnv].(type) {
	case map[string]string:
		return v
	case map[string]interface{}:
		out := make(map[string]string, len(v))
		for key, item := range v {
			if s, ok := item.(string); ok {
				out[key] = s
			}
		}
		return out
	default:
		return nil
	}
}

func setPassthroughMCPEnv(execution *AgentExecution, env map[string]string) {
	if len(env) == 0 {
		return
	}
	if execution.Metadata == nil {
		execution.Metadata = map[string]interface{}{}
	}
	execution.Metadata[metadataKeyPassthroughMCPEnv] = env
}

// passthroughAgentCommand validates passthrough support and builds the command for a passthrough session.
// Returns the PassthroughAgent, PassthroughConfig, RuntimeConfig pointer, command, and any error.
func (m *Manager) passthroughAgentCommand(ctx context.Context, execution *AgentExecution, profileInfo *AgentProfileInfo) (agents.PassthroughAgent, agents.PassthroughConfig, *agents.RuntimeConfig, agents.Command, error) {
	agentConfig, err := m.getAgentConfigForExecution(execution)
	if err != nil {
		return nil, agents.PassthroughConfig{}, nil, agents.Command{}, fmt.Errorf("failed to get agent config: %w", err)
	}

	ptAgent, ok := agentConfig.(agents.PassthroughAgent)
	if !ok {
		return nil, agents.PassthroughConfig{}, nil, agents.Command{}, fmt.Errorf("agent %s does not support passthrough mode", agentConfig.ID())
	}

	pt := ptAgent.PassthroughConfig()
	rt := agentConfig.Runtime()
	taskDescription := getTaskDescriptionFromMetadata(execution)
	promptForCmd := promptForPassthroughCommand(pt, taskDescription)
	mcpArgs, err := m.applyPassthroughMCP(ctx, execution, pt, agentConfig)
	if err != nil {
		return nil, agents.PassthroughConfig{}, nil, agents.Command{}, err
	}

	cmd := ptAgent.BuildPassthroughCommand(agents.PassthroughOptions{
		Model:            effectivePassthroughModel(execution, profileInfo),
		SessionID:        execution.ACPSessionID,
		Prompt:           promptForCmd,
		PermissionValues: profilePermissionValues(profileInfo),
		MCPArgs:          mcpArgs,
		CLIFlagTokens:    m.profileCLIFlagTokens(profileInfo),
	})
	if cmd.IsEmpty() {
		return nil, agents.PassthroughConfig{}, nil, agents.Command{}, fmt.Errorf("passthrough command is empty for agent %s", agentConfig.ID())
	}
	return ptAgent, pt, rt, cmd, nil
}

// profileCLIFlagTokens resolves the user-configured CLI flag argv tokens from
// a profile, mirroring the ACP launch path (manager_launch.go). Returns nil
// on resolve error and logs a warning so a malformed flag does not block the
// session — matches the warn-and-continue behaviour of the ACP path.
func (m *Manager) profileCLIFlagTokens(p *AgentProfileInfo) []string {
	if p == nil {
		return nil
	}
	tokens, err := cliflags.Resolve(p.CLIFlags)
	if err != nil {
		m.logger.Warn("failed to resolve cli_flags for passthrough profile, launching without user-configured flags",
			zap.String("profile_id", p.ProfileID),
			zap.Error(err))
		return nil
	}
	return tokens
}

// buildInteractiveStartRequest builds the InteractiveStartRequest for a passthrough session.
// immediateStart overrides pt.WaitForTerminal when true (used for restart/resume where the
// terminal WebSocket is already connected).
func buildInteractiveStartRequest(sessionID string, execution *AgentExecution, pt agents.PassthroughConfig, env map[string]string, cmd agents.Command, stripEnv []string, immediateStart bool) process.InteractiveStartRequest {
	return process.InteractiveStartRequest{
		SessionID: sessionID,
		Command:   cmd.Args(),
		// Redacted copy logged in place of Command by the interactive runner so
		// Codex MCP `-c` overrides (env/headers tokens) never reach process logs.
		LogCommand:      redactPassthroughArgs(cmd.Args()),
		WorkingDir:      execution.WorkspacePath,
		Env:             env,
		StripEnv:        stripEnv,
		PromptPattern:   pt.PromptPattern,
		IdleTimeout:     pt.IdleTimeout,
		BufferMaxBytes:  pt.BufferMaxBytes,
		StatusDetector:  pt.StatusDetector,
		CheckInterval:   pt.CheckInterval,
		StabilityWindow: pt.StabilityWindow,
		ImmediateStart:  immediateStart,
		DefaultCols:     120,
		DefaultRows:     40,
	}
}

// startInteractiveProcess launches the interactive PTY process for a passthrough session.
// Returns the process info on success.
func (m *Manager) startInteractiveProcess(ctx context.Context, execution *AgentExecution, pt agents.PassthroughConfig, env map[string]string, cmd agents.Command, stripEnv []string) (*process.InteractiveProcessInfo, error) {
	interactiveRunner := m.GetInteractiveRunner()
	if interactiveRunner == nil {
		return nil, fmt.Errorf("interactive runner not available for passthrough mode")
	}

	// Always start immediately with default dimensions (120×40). The first resize
	// from the terminal WebSocket will correct the size. Without immediate start,
	// WaitForTerminal agents deadlock: the frontend won't connect the terminal
	// until the session leaves STARTING, but the process never starts without a resize.
	// This matches ResumePassthroughSession and restartPassthroughProcess.
	startReq := buildInteractiveStartRequest(execution.SessionID, execution, pt, env, cmd, stripEnv, true)

	processInfo, err := interactiveRunner.Start(ctx, startReq)
	if err != nil {
		m.updateExecutionError(execution.ID, "failed to start passthrough session: "+err.Error())
		return nil, fmt.Errorf("failed to start passthrough session: %w", err)
	}
	return processInfo, nil
}

// startPassthroughSession starts an agent in passthrough mode (direct terminal interaction).
// Instead of using ACP protocol, the agent's stdin/stdout is passed through directly.
func (m *Manager) startPassthroughSession(ctx context.Context, execution *AgentExecution, profileInfo *AgentProfileInfo) error {
	_, pt, rt, cmd, err := m.passthroughAgentCommand(ctx, execution, profileInfo)
	if err != nil {
		return err
	}

	m.logger.Info("passthrough command built",
		zap.String("session_id", execution.SessionID),
		zap.Strings("full_command", redactPassthroughArgs(cmd.Args())))

	env := m.buildPassthroughEnv(ctx, execution, rt.RequiredEnv)

	processInfo, err := m.startInteractiveProcess(ctx, execution, pt, env, cmd, rt.StripEnv)
	if err != nil {
		return err
	}

	execution.PassthroughProcessID = processInfo.ID
	execution.PassthroughStartedAt = time.Now()
	execution.passthroughLaunchUsedResume = false

	m.logger.Info("passthrough session started",
		zap.String("execution_id", execution.ID),
		zap.String("task_id", execution.TaskID),
		zap.String("session_id", execution.SessionID),
		zap.String("process_id", processInfo.ID),
		zap.Strings("command", redactPassthroughArgs(cmd.Args())))

	m.eventPublisher.PublishAgentctlEvent(ctx, events.AgentctlReady, execution, "")
	m.startPassthroughShell(ctx, execution, "failed to start shell for passthrough session")

	if m.streamManager != nil && execution.agentctl != nil {
		m.streamManager.ConnectWorkspaceStream(execution, nil)
		// Also open the agent updates stream so the agentctl instance can proxy
		// kandev MCP tool calls to the backend (passthrough has no ACP stream
		// otherwise, so MCP tool calls would hang).
		if !execution.agentctl.HasAgentStream() {
			m.streamManager.ConnectMCPStream(execution)
		}
	}

	go m.autoInjectInitialPrompt(execution, pt)

	return nil
}

// profileModel extracts the model from profile info, returning empty string if nil.
func profileModel(p *AgentProfileInfo) string {
	if p == nil {
		return ""
	}
	return p.Model
}

// effectivePassthroughModel returns the model that should be passed to the next
// passthrough launch: a model_override on the execution wins over the profile's
// model. SetSessionModel sets the override for passthrough sessions because the
// model is baked into the CLI command at launch time — there is no live channel
// to swap it, so the PTY must be relaunched with a new --model.
func effectivePassthroughModel(execution *AgentExecution, profile *AgentProfileInfo) string {
	if execution != nil && execution.Metadata != nil {
		if override, ok := execution.Metadata[MetadataKeyModelOverride].(string); ok && override != "" {
			return override
		}
	}
	return profileModel(profile)
}

// profilePermissionValues builds a permission values map from profile info.
func profilePermissionValues(p *AgentProfileInfo) map[string]bool {
	if p == nil {
		return nil
	}
	values := map[string]bool{
		"dangerously_skip_permissions": p.DangerouslySkipPermissions,
		"allow_indexing":               p.AllowIndexing,
	}
	values[agents.PermissionKeyAutoApprove] = p.AutoApprove
	return values
}

// freshPassthroughCommand resolves the agent config and profile, and builds a
// bare passthrough command with no session, resume, or prompt flags.
func (m *Manager) freshPassthroughCommand(ctx context.Context, execution *AgentExecution) (agents.PassthroughConfig, *agents.RuntimeConfig, agents.Command, error) {
	resolved, err := m.resolvePassthroughAgent(ctx, execution)
	if err != nil {
		return agents.PassthroughConfig{}, nil, agents.Command{}, err
	}
	mcpArgs, err := m.applyPassthroughMCP(ctx, execution, resolved.pt, resolved.agentConfig)
	if err != nil {
		return agents.PassthroughConfig{}, nil, agents.Command{}, err
	}

	cmd := resolved.agent.BuildPassthroughCommand(agents.PassthroughOptions{
		Model:            effectivePassthroughModel(execution, resolved.profile),
		PermissionValues: profilePermissionValues(resolved.profile),
		MCPArgs:          mcpArgs,
		CLIFlagTokens:    m.profileCLIFlagTokens(resolved.profile),
	})
	if cmd.IsEmpty() {
		return agents.PassthroughConfig{}, nil, agents.Command{}, fmt.Errorf("passthrough command is empty for agent %s", resolved.agentID)
	}

	return resolved.pt, resolved.rt, cmd, nil
}

func (m *Manager) resumePassthroughCommand(ctx context.Context, execution *AgentExecution, resolved *resolvedPassthrough, useResume bool) (agents.Command, error) {
	mcpArgs, err := m.applyPassthroughMCP(ctx, execution, resolved.pt, resolved.agentConfig)
	if err != nil {
		return agents.Command{}, err
	}
	cmd := resolved.agent.BuildPassthroughCommand(agents.PassthroughOptions{
		Model:            effectivePassthroughModel(execution, resolved.profile),
		Resume:           useResume,
		PermissionValues: profilePermissionValues(resolved.profile),
		MCPArgs:          mcpArgs,
		CLIFlagTokens:    m.profileCLIFlagTokens(resolved.profile),
	})
	if cmd.IsEmpty() {
		return agents.Command{}, fmt.Errorf("passthrough resume command is empty for agent %s", resolved.agentID)
	}
	return cmd, nil
}

// restartPassthroughProcess kills the current PTY process and relaunches a fresh one
// without --resume, effectively clearing the agent's conversation context.
// The workflow step prompt is delivered afterwards via stdin (autoStartPassthroughPrompt).
func (m *Manager) restartPassthroughProcess(ctx context.Context, execution *AgentExecution) error {
	m.logger.Info("restarting passthrough process for context reset",
		zap.String("execution_id", execution.ID),
		zap.String("session_id", execution.SessionID),
		zap.String("old_process_id", execution.PassthroughProcessID))

	// 1. Stop the current PTY process.
	// Clear PassthroughProcessID before stopping so that handlePassthroughStatus
	// doesn't trigger auto-restart for the deliberately-killed process.
	interactiveRunner := m.GetInteractiveRunner()
	if interactiveRunner == nil {
		return fmt.Errorf("interactive runner not available")
	}

	oldProcessID := execution.PassthroughProcessID
	execution.PassthroughProcessID = ""
	execution.PassthroughStartedAt = time.Time{}

	if err := interactiveRunner.Stop(ctx, oldProcessID); err != nil {
		m.logger.Warn("failed to stop passthrough process during context reset",
			zap.String("execution_id", execution.ID),
			zap.String("process_id", oldProcessID),
			zap.Error(err))
	}

	// 2. Build fresh command (no SessionID, no Resume, no Prompt)
	pt, rt, cmd, err := m.freshPassthroughCommand(ctx, execution)
	if err != nil {
		return err
	}

	// 3. Start new PTY process with ImmediateStart (terminal is already connected)
	env := m.buildPassthroughEnv(ctx, execution, rt.RequiredEnv)
	startReq := buildInteractiveStartRequest(execution.SessionID, execution, pt, env, cmd, rt.StripEnv, true)

	processInfo, err := interactiveRunner.Start(ctx, startReq)
	if err != nil {
		m.updateExecutionError(execution.ID, "failed to restart passthrough session: "+err.Error())
		return fmt.Errorf("failed to restart passthrough session: %w", err)
	}

	// 4. Update execution with new process ID
	execution.PassthroughProcessID = processInfo.ID
	execution.PassthroughStartedAt = time.Now()
	execution.passthroughLaunchUsedResume = false

	m.logger.Info("passthrough process restarted with fresh context",
		zap.String("execution_id", execution.ID),
		zap.String("session_id", execution.SessionID),
		zap.String("new_process_id", processInfo.ID))

	// 5. Reconnect existing WebSocket to the new process
	if interactiveRunner.ConnectSessionWebSocket(processInfo.ID) {
		m.logger.Debug("reconnected WebSocket to restarted passthrough process",
			zap.String("session_id", execution.SessionID),
			zap.String("process_id", processInfo.ID))
	}

	// 6. Publish context reset event
	m.eventPublisher.PublishAgentEvent(ctx, events.AgentContextReset, execution)

	return nil
}

// ResumePassthroughSession restarts a passthrough session after backend restart.
// This is called when user reconnects to a terminal but the PTY process is no longer running.
// If the agent supports resume, it uses the resume flag to continue the last conversation.
// Otherwise, it starts a fresh CLI session with the same profile settings.
func (m *Manager) ResumePassthroughSession(ctx context.Context, sessionID string) error {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		return fmt.Errorf("%w: %s", ErrNoExecutionForSession, sessionID)
	}

	resolved, err := m.resolvePassthroughAgent(ctx, execution)
	if err != nil {
		return err
	}

	interactiveRunner := m.GetInteractiveRunner()
	if interactiveRunner == nil {
		return fmt.Errorf("interactive runner not available")
	}

	// Skip the resume flag if a previous resume already fast-failed for this
	// execution: re-attaching `-c` / `--resume` would just reproduce the same
	// "No conversation found to continue" exit on every WS reconnect after
	// backend restart. Once the sticky flag is set, every subsequent launch
	// for this execution starts fresh.
	useResume := !execution.passthroughResumeFailed
	cmd, err := m.resumePassthroughCommand(ctx, execution, resolved, useResume)
	if err != nil {
		return err
	}

	m.logger.Info("resuming passthrough session",
		zap.String("session_id", sessionID),
		zap.String("execution_id", execution.ID),
		zap.Bool("use_resume", useResume),
		zap.Strings("command", redactPassthroughArgs(cmd.Args())))

	env := m.buildPassthroughEnv(ctx, execution, resolved.rt.RequiredEnv)

	// Always use immediate start on resume — the terminal WebSocket is already connected,
	// so we don't need to wait for a resize to get exact dimensions. The first resize
	// from the terminal will correct the dimensions. Without this, TUI apps that use
	// WaitForTerminal would never start because the frontend may not send resizes
	// to a process it doesn't know about yet.
	startReq := buildInteractiveStartRequest(sessionID, execution, resolved.pt, env, cmd, resolved.rt.StripEnv, true)

	processInfo, err := interactiveRunner.Start(ctx, startReq)
	if err != nil {
		return fmt.Errorf("failed to start passthrough session: %w", err)
	}

	execution.PassthroughStartedAt = time.Now()
	execution.passthroughLaunchUsedResume = useResume
	execution.PassthroughProcessID = processInfo.ID

	m.logger.Info("passthrough session resumed",
		zap.String("session_id", sessionID),
		zap.String("execution_id", execution.ID),
		zap.String("process_id", processInfo.ID))

	// Start shell session for workspace shell access (right panel terminal).
	// This needs to be done after resume since the shell process was killed on backend restart.
	m.startPassthroughShell(ctx, execution, "failed to start shell for resumed passthrough session")

	// Connect to workspace stream for shell/git/file features.
	// Only connect if not already connected (process restart reuses the same agentctl).
	if m.streamManager != nil && execution.agentctl != nil && execution.GetWorkspaceStream() == nil {
		m.streamManager.ConnectWorkspaceStream(execution, nil)
	}
	// Re-open the MCP proxy stream too (drains kandev MCP tool calls).
	if m.streamManager != nil && execution.agentctl != nil && !execution.agentctl.HasAgentStream() {
		m.streamManager.ConnectMCPStream(execution)
	}

	return nil
}

// handlePassthroughTurnComplete is called when turn detection fires for a passthrough session.
// This marks the execution as ready for follow-up prompts when the agent finishes processing.
func (m *Manager) handlePassthroughTurnComplete(sessionID string) {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		m.logger.Debug("turn complete for unknown session (may have ended)",
			zap.String("session_id", sessionID))
		return
	}

	m.logger.Info("passthrough turn complete",
		zap.String("session_id", sessionID),
		zap.String("execution_id", execution.ID))

	// Mark execution as ready for follow-up prompts
	// This publishes AgentReady event to notify subscribers
	if err := m.MarkReady(execution.ID); err != nil {
		m.logger.Error("failed to mark execution as ready after passthrough turn complete",
			zap.String("execution_id", execution.ID),
			zap.Error(err))
	}
}

// handlePassthroughOutput handles output from a passthrough process and publishes it to the event bus.
// This is called when running in standalone mode without a WorkspaceTracker.
func (m *Manager) handlePassthroughOutput(output *agentctltypes.ProcessOutput) {
	if output == nil {
		return
	}

	execution, exists := m.executionStore.GetBySessionID(output.SessionID)
	if !exists {
		m.logger.Debug("passthrough output for unknown session",
			zap.String("session_id", output.SessionID))
		return
	}

	// Convert to agentctl client type for event publisher
	clientOutput := &agentctl.ProcessOutput{
		SessionID: output.SessionID,
		ProcessID: output.ProcessID,
		Kind:      output.Kind,
		Stream:    output.Stream,
		Data:      output.Data,
		Timestamp: output.Timestamp,
	}

	m.eventPublisher.PublishProcessOutput(execution, clientOutput)
}

// handlePassthroughStatus handles status updates from a passthrough process and publishes to the event bus.
// This is called when running in standalone mode without a WorkspaceTracker.
// When the process exits while a WebSocket is connected, it attempts auto-restart with rate limiting.
func (m *Manager) handlePassthroughStatus(status *agentctltypes.ProcessStatusUpdate) {
	if status == nil {
		return
	}

	execution, exists := m.executionStore.GetBySessionID(status.SessionID)
	if !exists {
		m.logger.Debug("passthrough status for unknown session",
			zap.String("session_id", status.SessionID))
		return
	}

	// Convert to agentctl client type for event publisher
	clientStatus := &agentctl.ProcessStatusUpdate{
		SessionID:  status.SessionID,
		ProcessID:  status.ProcessID,
		Kind:       status.Kind,
		Command:    status.Command,
		ScriptName: status.ScriptName,
		WorkingDir: status.WorkingDir,
		Status:     status.Status,
		ExitCode:   status.ExitCode,
		Timestamp:  status.Timestamp,
	}

	m.eventPublisher.PublishProcessStatus(execution, clientStatus)

	// Check if process exited and should be auto-restarted
	// Only restart if this is the ACTUAL passthrough process, not user shell terminals
	// Run asynchronously to allow the old process to be cleaned up first
	if status.Status == agentctltypes.ProcessStatusExited || status.Status == agentctltypes.ProcessStatusFailed {
		// Only trigger auto-restart for the passthrough process, not for user shell terminals
		if execution.PassthroughProcessID != "" && status.ProcessID == execution.PassthroughProcessID {
			// Snapshot the start time and resume-launch flag synchronously so the
			// goroutine doesn't race the next launch's writes to those fields.
			startedAt := execution.PassthroughStartedAt
			usedResume := execution.passthroughLaunchUsedResume
			// Detect fast-fail synchronously so we can flip the resume-failed
			// flag before the next WS reconnect arrives. The goroutine below
			// would otherwise miss this race: when the PTY exits the WS bridge
			// closes too, and by the time the goroutine runs (after a 100ms
			// cleanup delay) HasActiveWebSocketBySession returns false and it
			// bails without setting any flags. Scoped to fast-fail+usedResume
			// so a healthy resumed session that exits cleanly or crashes long
			// after launch keeps its resume intent for auto-restart.
			if usedResume && passthroughExitIsFastFail(startedAt, status) {
				execution.passthroughResumeFailed = true
				execution.isResumedSession = false
				execution.passthroughLaunchUsedResume = false
			}
			go m.handlePassthroughExit(execution, status, startedAt, usedResume)
		} else {
			m.logger.Debug("process exited but not the passthrough process, skipping auto-restart",
				zap.String("session_id", status.SessionID),
				zap.String("exited_process_id", status.ProcessID),
				zap.String("passthrough_process_id", execution.PassthroughProcessID))
		}
	}
}

// handlePassthroughExit handles auto-restart logic when a passthrough process exits.
// This function is called asynchronously to allow the old process to be cleaned up first.
// startedAt and usedResume are snapshots of the matching execution fields taken
// synchronously at the call site — passed in rather than re-read here to avoid
// racing with the next launch's writes to those fields.
func (m *Manager) handlePassthroughExit(execution *AgentExecution, status *agentctltypes.ProcessStatusUpdate, startedAt time.Time, usedResume bool) {
	const restartDelay = 500 * time.Millisecond
	const cleanupDelay = 100 * time.Millisecond // Wait for old process cleanup
	// fastFailWindow is short enough to catch launch-time failures (bad CLI
	// flag, missing binary, immediate auth rejection) but long enough that a
	// healthy agent that does any startup work won't be mistaken for one.
	const fastFailWindow = 2 * time.Second

	sessionID := execution.SessionID

	if m.IsShuttingDown() {
		m.logger.Debug("skipping passthrough auto-restart during shutdown",
			zap.String("session_id", sessionID))
		return
	}

	// Wait a bit for the old process to be cleaned up from the process map
	time.Sleep(cleanupDelay)

	// Shutdown may have started during cleanupDelay; re-check before emitting
	// the "attempting auto-restart" log and the terminal banner, which would
	// otherwise mislead the user during a clean shutdown.
	if m.IsShuttingDown() {
		m.logger.Debug("skipping passthrough auto-restart during shutdown",
			zap.String("session_id", sessionID))
		return
	}

	interactiveRunner := m.GetInteractiveRunner()
	if interactiveRunner == nil {
		m.logger.Debug("no interactive runner available for auto-restart",
			zap.String("session_id", sessionID))
		return
	}

	// Check if WebSocket is still connected (use session-level tracking which survives process deletion)
	if !interactiveRunner.HasActiveWebSocketBySession(sessionID) {
		m.logger.Debug("no active WebSocket, skipping auto-restart",
			zap.String("session_id", sessionID))
		return
	}

	exitCode := 0
	if status.ExitCode != nil {
		exitCode = *status.ExitCode
	}

	// Use the exit timestamp from the status event (set when the child
	// actually exited), not time.Now() — the cleanupDelay sleep and goroutine
	// hops above would otherwise inflate the measured uptime by ~100 ms.
	exitedAt := status.Timestamp
	if exitedAt.IsZero() {
		exitedAt = time.Now()
	}

	// Fast-fail short-circuit: a non-zero exit shortly after start almost
	// always means the launch itself was wrong (bad CLI flag, missing binary,
	// auth failure). Restarting just thrashes — the next run hits the same
	// failure at the same speed. Surface the failure to the user instead.
	if isFastFailExit(startedAt, exitedAt, exitCode, fastFailWindow) {
		uptime := exitedAt.Sub(startedAt)
		// If the failed launch was a resume (e.g. `--resume <id>` or `-c`),
		// the most likely cause is a stale conversation ID after a backend
		// restart — "No conversation found to continue". Retry once with a
		// fresh command (no resume flag) before giving up.
		if usedResume {
			// Resume-failed flags have already been flipped synchronously in
			// handlePassthroughStatus (see passthroughExitIsFastFail) so any
			// concurrent WS reconnect that races this goroutine sees the new
			// values immediately.
			m.attemptResumeFallback(execution, interactiveRunner, sessionID, exitCode, uptime)
			return
		}
		m.notifyFastFailExit(interactiveRunner, sessionID, uptime, exitCode, fastFailWindow)
		return
	}

	m.attemptPassthroughRestart(execution, interactiveRunner, sessionID, exitCode, restartDelay)
}

// attemptResumeFallback recovers from a fast-failed resume launch by relaunching
// once with a fresh command (no resume flag). This handles the common case where
// the local CLI's conversation history is gone after a backend restart and
// `claude -c` / `claude --resume <id>` exits with "No conversation found to
// continue". On a successful fallback the user keeps a working session; on
// continued failure we surface the existing red banner so they can fix their
// profile.
func (m *Manager) attemptResumeFallback(execution *AgentExecution, runner *process.InteractiveRunner, sessionID string, exitCode int, uptime time.Duration) {
	m.logger.Info("passthrough resume launch fast-failed, retrying without resume flag",
		zap.String("session_id", sessionID),
		zap.String("execution_id", execution.ID),
		zap.Int("exit_code", exitCode),
		zap.Duration("uptime", uptime))

	if m.IsShuttingDown() {
		m.logger.Debug("skipping passthrough resume fallback during shutdown",
			zap.String("session_id", sessionID))
		return
	}
	if !runner.HasActiveWebSocketBySession(sessionID) {
		m.logger.Debug("no active WebSocket, skipping passthrough resume fallback",
			zap.String("session_id", sessionID))
		return
	}

	banner := "\r\n\x1b[33m[No prior conversation to resume — starting a fresh session...]\x1b[0m\r\n"
	if err := runner.WriteToDirectOutputBySession(sessionID, []byte(banner)); err != nil {
		m.logger.Debug("failed to write resume-fallback banner to terminal",
			zap.String("session_id", sessionID),
			zap.Error(err))
	}

	ctx := context.Background()
	pt, rt, cmd, err := m.freshPassthroughCommand(ctx, execution)
	if err != nil {
		m.notifyFallbackInfrastructureFailure(runner, sessionID, "build fresh command", err)
		return
	}

	env := m.buildPassthroughEnv(ctx, execution, rt.RequiredEnv)
	startReq := buildInteractiveStartRequest(sessionID, execution, pt, env, cmd, rt.StripEnv, true)

	processInfo, err := runner.Start(ctx, startReq)
	if err != nil {
		m.notifyFallbackInfrastructureFailure(runner, sessionID, "start fresh process", err)
		return
	}

	execution.PassthroughStartedAt = time.Now()
	execution.passthroughLaunchUsedResume = false
	execution.PassthroughProcessID = processInfo.ID

	if runner.ConnectSessionWebSocket(processInfo.ID) {
		m.logger.Info("passthrough resume fallback succeeded",
			zap.String("session_id", sessionID),
			zap.String("execution_id", execution.ID),
			zap.String("new_process_id", processInfo.ID))
	} else {
		m.logger.Warn("passthrough resume fallback started but failed to reconnect WebSocket",
			zap.String("session_id", sessionID),
			zap.String("new_process_id", processInfo.ID))
	}

	// Mirror ResumePassthroughSession's post-launch bootstrap so right-panel
	// shell/git/file features come back too — without this the main terminal
	// works but the user's shell session and workspace stream stay torn down.
	m.startPassthroughShell(ctx, execution, "failed to start shell after passthrough resume fallback")
	if m.streamManager != nil && execution.agentctl != nil && execution.GetWorkspaceStream() == nil {
		m.streamManager.ConnectWorkspaceStream(execution, nil)
	}
	if m.streamManager != nil && execution.agentctl != nil && !execution.agentctl.HasAgentStream() {
		m.streamManager.ConnectMCPStream(execution)
	}

	// Fallback path is a fresh session (no --resume) — re-inject the prompt.
	go m.autoInjectInitialPrompt(execution, pt)
}

// attemptPassthroughRestart announces the restart on the terminal, waits the
// restart delay, re-checks shutdown/WebSocket, and resumes the session.
// Reconnects the existing WebSocket to the new process on success.
func (m *Manager) attemptPassthroughRestart(execution *AgentExecution, runner *process.InteractiveRunner, sessionID string, exitCode int, restartDelay time.Duration) {
	m.logger.Info("passthrough process exited with active WebSocket, attempting auto-restart",
		zap.String("session_id", sessionID),
		zap.Int("exit_code", exitCode))

	restartMsg := "\r\n\x1b[33m[Agent exited. Restarting...]\x1b[0m\r\n"
	if err := runner.WriteToDirectOutputBySession(sessionID, []byte(restartMsg)); err != nil {
		m.logger.Debug("failed to write restart message to terminal",
			zap.String("session_id", sessionID),
			zap.Error(err))
	}

	time.Sleep(restartDelay)

	// Shutdown may have started during the sleep; re-check before touching
	// state that the teardown is racing to remove.
	if m.IsShuttingDown() {
		m.logger.Debug("skipping passthrough auto-restart during shutdown",
			zap.String("session_id", sessionID))
		return
	}

	if !runner.HasActiveWebSocketBySession(sessionID) {
		m.logger.Debug("WebSocket disconnected during restart delay, aborting",
			zap.String("session_id", sessionID))
		return
	}

	if err := m.ResumePassthroughSession(context.Background(), sessionID); err != nil {
		m.logger.Error("failed to auto-restart passthrough session",
			zap.String("session_id", sessionID),
			zap.Error(err))

		errorMsg := fmt.Sprintf("\r\n\x1b[31m[Restart failed: %s]\x1b[0m\r\n", err.Error())
		if writeErr := runner.WriteToDirectOutputBySession(sessionID, []byte(errorMsg)); writeErr != nil {
			m.logger.Debug("failed to write restart error message to terminal",
				zap.String("session_id", sessionID),
				zap.Error(writeErr))
		}
		return
	}

	if runner.ConnectSessionWebSocket(execution.PassthroughProcessID) {
		m.logger.Info("passthrough session auto-restarted and reconnected WebSocket",
			zap.String("session_id", sessionID),
			zap.String("new_process_id", execution.PassthroughProcessID))
	} else {
		m.logger.Warn("passthrough session restarted but failed to reconnect WebSocket",
			zap.String("session_id", sessionID),
			zap.String("new_process_id", execution.PassthroughProcessID))
	}
}

// notifyFallbackInfrastructureFailure surfaces an attemptResumeFallback
// failure that originated from kandev's own machinery (could not build the
// fresh command, could not start the new process) rather than from the
// agent CLI itself. The existing fast-fail banner blames a "bad CLI flag,
// missing binary, or auth failure" — wrong copy for these paths, which
// the user can't fix by editing their profile.
func (m *Manager) notifyFallbackInfrastructureFailure(runner *process.InteractiveRunner, sessionID, stage string, err error) {
	m.logger.Error("passthrough resume fallback failed",
		zap.String("session_id", sessionID),
		zap.String("stage", stage),
		zap.Error(err))
	failMsg := fmt.Sprintf("\r\n\x1b[31m[Resume fallback failed: could not %s — %s. Please reconnect to retry.]\x1b[0m\r\n",
		stage, err.Error())
	if writeErr := runner.WriteToDirectOutputBySession(sessionID, []byte(failMsg)); writeErr != nil {
		m.logger.Debug("failed to write resume-fallback infra-failure banner to terminal",
			zap.String("session_id", sessionID),
			zap.Error(writeErr))
	}
}

// notifyFastFailExit logs the fast-fail decision and writes a one-shot
// banner to the terminal explaining why the auto-restart was skipped.
// uptime is the measured process lifetime (status timestamp minus start
// time), pre-computed by the caller so the log reflects true child uptime.
func (m *Manager) notifyFastFailExit(runner *process.InteractiveRunner, sessionID string, uptime time.Duration, exitCode int, window time.Duration) {
	m.logger.Warn("passthrough process exited fast with non-zero code, skipping auto-restart",
		zap.String("session_id", sessionID),
		zap.Int("exit_code", exitCode),
		zap.Duration("uptime", uptime))
	failMsg := fmt.Sprintf("\r\n\x1b[31m[Agent exited (code %d) within %s. Likely cause: bad CLI flag, missing binary, or auth failure. Edit your profile and reconnect to retry.]\x1b[0m\r\n",
		exitCode, window)
	if err := runner.WriteToDirectOutputBySession(sessionID, []byte(failMsg)); err != nil {
		m.logger.Debug("failed to write fast-fail message to terminal",
			zap.String("session_id", sessionID),
			zap.Error(err))
	}
}

// passthroughExitIsFastFail wraps isFastFailExit for the synchronous
// status-callback path that doesn't yet have the unpacked exit code or
// timestamp. Same window/semantics as the goroutine path.
func passthroughExitIsFastFail(startedAt time.Time, status *agentctltypes.ProcessStatusUpdate) bool {
	const fastFailWindow = 2 * time.Second
	exitCode := 0
	if status.ExitCode != nil {
		exitCode = *status.ExitCode
	}
	exitedAt := status.Timestamp
	if exitedAt.IsZero() {
		exitedAt = time.Now()
	}
	return isFastFailExit(startedAt, exitedAt, exitCode, fastFailWindow)
}

// isFastFailExit reports whether a passthrough process exit looks like a
// launch failure rather than a runtime exit worth restarting. A zero start
// time disables the check (e.g. recovered executions where the start time
// is unknown), so the legacy restart path remains the default. exitedAt is
// the wall-clock time the process actually exited (status.Timestamp), kept
// distinct from time.Now() so the cleanupDelay sleep above the call site
// doesn't shrink the effective window.
func isFastFailExit(startedAt, exitedAt time.Time, exitCode int, window time.Duration) bool {
	if exitCode == 0 || startedAt.IsZero() {
		return false
	}
	return exitedAt.Sub(startedAt) < window
}

// passthroughRunner is the minimal seam autoInjectInitialPrompt needs from
// *process.InteractiveRunner. Defined as an interface so tests can supply a
// fake runner without spinning up a real PTY subprocess.
type passthroughRunner interface {
	WaitForFirstIdle(ctx context.Context, processID string) error
	WriteStdin(processID string, data string) error
}

// autoInjectInitialPrompt writes the task description to the PTY stdin once
// the agent is idle (ready for input). Opt-in per agent via PassthroughConfig.
// Called from startPassthroughSession and attemptResumeFallback only — never
// from ResumePassthroughSession (would duplicate the prompt in agent history).
func (m *Manager) autoInjectInitialPrompt(execution *AgentExecution, pt agents.PassthroughConfig) {
	runner := m.GetInteractiveRunner()
	if runner == nil {
		return
	}
	m.autoInjectInitialPromptWith(runner, execution, pt)
}

// autoInjectInitialPromptWith is the testable inner of autoInjectInitialPrompt,
// taking a runner seam so unit tests can avoid spawning a real PTY.
func (m *Manager) autoInjectInitialPromptWith(runner passthroughRunner, execution *AgentExecution, pt agents.PassthroughConfig) {
	if !pt.AutoInjectPrompt {
		return
	}
	if !pt.PromptFlag.IsEmpty() {
		// The agent already received the prompt as a CLI flag.
		return
	}
	description := getTaskDescriptionFromMetadata(execution)
	if description == "" {
		return
	}
	processID := execution.PassthroughProcessID
	if processID == "" {
		m.logger.Warn("autoInjectInitialPrompt called without passthrough process",
			zap.String("execution_id", execution.ID))
		return
	}
	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
	defer cancel()
	if err := runner.WaitForFirstIdle(ctx, processID); err != nil {
		m.logger.Warn("autoInjectInitialPrompt timed out waiting for idle",
			zap.String("execution_id", execution.ID),
			zap.String("process_id", processID),
			zap.Error(err))
		return
	}
	// WaitForFirstIdle also unblocks when the process exits — skip the write
	// during shutdown so we don't race the lifecycle manager's teardown. If the
	// process is already gone, WriteStdin returns "process not found" and the
	// existing error branch logs it.
	if m.IsShuttingDown() {
		return
	}
	// Mark RUNNING before the chunk loop so a composer/message.add fired during
	// the inter-chunk SubmitDelay window (150ms for Claude) is blocked by
	// checkSessionPromptable instead of racing into the same PTY mid-submit.
	if err := m.MarkPassthroughRunning(execution.SessionID); err != nil {
		m.logger.Warn("failed to mark passthrough as running before auto-inject",
			zap.String("execution_id", execution.ID),
			zap.String("session_id", execution.SessionID),
			zap.Error(err))
	}
	for _, chunk := range agents.PlanPassthroughStdinChunks(description, pt) {
		if chunk.DelayBefore > 0 {
			time.Sleep(chunk.DelayBefore)
		}
		if err := runner.WriteStdin(processID, chunk.Data); err != nil {
			m.logger.Warn("autoInjectInitialPrompt write failed",
				zap.String("execution_id", execution.ID),
				zap.String("process_id", processID),
				zap.Error(err))
			return
		}
	}
	m.logger.Info("autoInjectInitialPrompt wrote task description to PTY",
		zap.String("execution_id", execution.ID),
		zap.String("process_id", processID),
		zap.Int("description_len", len(description)))
}

// ResolvePassthroughConfig returns the PassthroughConfig for a session's agent.
// Used by callers outside this package (e.g. orchestrator) that need the submit
// sequence to write to PTY stdin.
func (m *Manager) ResolvePassthroughConfig(ctx context.Context, sessionID string) (agents.PassthroughConfig, error) {
	execution, exists := m.executionStore.GetBySessionID(sessionID)
	if !exists {
		return agents.PassthroughConfig{}, fmt.Errorf("no execution for session %q", sessionID)
	}
	resolved, err := m.resolvePassthroughAgent(ctx, execution)
	if err != nil {
		return agents.PassthroughConfig{}, err
	}
	return resolved.pt, nil
}

// GetInteractiveRunner returns the interactive runner for passthrough mode.
// Returns nil if the runtime is not available or doesn't support passthrough.
func (m *Manager) GetInteractiveRunner() *process.InteractiveRunner {
	if m.executorRegistry == nil {
		return nil
	}
	standaloneRT, err := m.executorRegistry.GetBackend(executor.NameStandalone)
	if err != nil {
		return nil
	}
	return standaloneRT.GetInteractiveRunner()
}