Merge pull request #343 from scgopi/fix/provider-not-on-path

Stop a loop whose backend CLI is not on PATH, and say so

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KMVJyUtec3hpH5hF44xAgh

Author: scgopi

GraphcodeKit/Sources/Domain/LoopNode.swift

@@ -192,6 +192,9 @@ public struct LoopNode: Identifiable, Codable, Equatable, Sendable {
   /// `GraphStore` at the moment of the stall; `nil` for loops stalled before the field
   /// existed, and for stalls whose cause the graph had nothing to say about.
   public var stallReason: String?
+  /// Set when the daemon stopped this loop because its backend's CLI is not on the
+  /// launch shell's PATH; cleared by the restart that follows the fix.
+  public var launchFailure: LaunchFailure?
   public var state: LoopState
   public var createdAt: Date
 
@@ -492,7 +495,7 @@ public struct LoopNode: Identifiable, Codable, Equatable, Sendable {
     case lastMailroomRead, mailroomWatch
     case state, createdAt, activity, presence, firstInstruction, pausesBeforeWritesOnly
     case summary, board, heartbeatIntervalSeconds, stallReason
-    case createdFromTemplateID, templateFollow, sessionRestarts
+    case createdFromTemplateID, templateFollow, sessionRestarts, launchFailure
   }
 
   /// Hand-written for the same reason `LoopEdge`'s is: `ProjectPersistence.loadGraph`
@@ -548,6 +551,7 @@ public struct LoopNode: Identifiable, Codable, Equatable, Sendable {
       try container.decodeIfPresent(MailroomWatch.self, forKey: .mailroomWatch)
       ?? decoder.legacyMailroomValue(MailroomWatch.self, "artifactoryWatch")
     stallReason = try container.decodeIfPresent(String.self, forKey: .stallReason)
+    launchFailure = try container.decodeIfPresent(LaunchFailure.self, forKey: .launchFailure)
     state = try container.decodeIfPresent(LoopState.self, forKey: .state) ?? .idle
     createdAt = try container.decodeIfPresent(Date.self, forKey: .createdAt) ?? Date()
   }

GraphcodeKit/Sources/GraphStore.swift

@@ -290,6 +290,7 @@ public actor GraphStore {
     deliveryDeadline: Duration = .seconds(45),
     onGraphChanged: (@Sendable (LoopGraph) -> Void)? = nil,
     onEnsureSession: (@Sendable (LoopNode, String?) -> Void)? = nil,
+    onFindMissingProvider: (@Sendable (LoopNode, String?) async -> LaunchFailure?)? = nil,
     onTerminateSession: (@Sendable (LoopNode, String?) -> Void)? = nil,
     onRestartSession: (@Sendable (LoopNode, String?) async -> Bool)? = nil,
     onEvaluatePredicate: (@Sendable (ShellPredicate) async -> Bool)? = nil,
@@ -325,6 +326,7 @@ public actor GraphStore {
     self.subGraphDepth = subGraphDepth
     self.onGraphChanged = onGraphChanged
     self.onEnsureSession = onEnsureSession
+    self.onFindMissingProvider = onFindMissingProvider
     self.onTerminateSession = onTerminateSession
     self.onRestartSession = onRestartSession
     self.onEvaluatePredicate = onEvaluatePredicate
@@ -423,6 +425,54 @@ public actor GraphStore {
   /// it; see `resolvedForLaunch`.
   private func ensureSession(_ node: LoopNode) {
     onEnsureSession?(resolvedForLaunch(node), graph.project.path)
+    guard onFindMissingProvider != nil else { return }
+    Task { await self.stopIfProviderMissing(node) }
+  }
+
+  /// Whether the node's backend CLI is missing from the launch shell's PATH
+  /// (`ProviderPath`). Asked beside the launch rather than before it: `ensureSession` is
+  /// synchronous and a login shell takes a moment, and a launch whose CLI is missing
+  /// only makes a session that exits at once — which the stop kills anyway.
+  private let onFindMissingProvider: (@Sendable (LoopNode, String?) async -> LaunchFailure?)?
+
+  private func stopIfProviderMissing(_ node: LoopNode) async {
+    guard let onFindMissingProvider,
+      let failure = await onFindMissingProvider(node, graph.project.path),
+      stopForMissingProvider(node.id, failure)
+    else { return }
+    broadcast()
+  }
+
+  /// A stop rather than a failure: nothing the loop did went wrong, and the restart that
+  /// follows the fix must be allowed (`restartNode`). Killed rather than asked, because
+  /// there is no agent in the session to ask.
+  @discardableResult
+  private func stopForMissingProvider(_ nodeID: UUID, _ failure: LaunchFailure) -> Bool {
+    guard let node = graph.nodes[id: nodeID], !node.isResolved else { return false }
+    setNodeState(nodeID, .stopped)
+    graph.nodes[id: nodeID]?.launchFailure = failure
+    cancelGoalPoller(nodeID)
+    cancelHeartbeat(nodeID)
+    recordMemory(
+      nodeID,
+      "stopped: \(failure.title) — install \(failure.backend.displayName) or add the folder "
+        + "containing \(failure.executable) to the login shell's PATH, then restart the loop")
+    terminateSession(node)
+    fireOutgoingEdges(from: nodeID, sourceSucceeded: false)
+    return true
+  }
+
+  /// The restart after the fix. `sessionRestarts` moves because it is the app's cue to
+  /// remount the workspace it closed for the restart (`SessionRestart.pendingReopen`).
+  private func relaunchAfterMissingProvider(_ node: LoopNode, _ failure: LaunchFailure) {
+    graph.nodes[id: node.id]?.launchFailure = nil
+    graph.nodes[id: node.id]?.sessionRestarts += 1
+    setNodeState(node.id, node.runsUnattended ? .running : .idle)
+    recordMemory(node.id, "restarted after \(failure.title) — launching again")
+    guard node.runsUnattended, let relaunched = graph.nodes[id: node.id] else { return }
+    if relaunched.loopType == .goalBased { armGoalPoller(for: relaunched) }
+    armHeartbeat(for: relaunched)
+    ensureSession(relaunched)
   }
 
   // MARK: - Template follows
@@ -1273,6 +1323,11 @@ public actor GraphStore {
       guard graph.nodes[id: node.id]?.presence != reading else { continue }
       graph.nodes[id: node.id]?.presence = reading
       changed = true
+      // The backstop for a session no launch of ours checked: a zsh that exits 127 could
+      // not find its command, and the probe says whether that command was the agent.
+      if reading.exitCode == ProviderPath.commandNotFoundStatus, onFindMissingProvider != nil {
+        Task { await self.stopIfProviderMissing(node) }
+      }
     }
     if refreshActiveDependents() { changed = true }
     return changed
@@ -2116,12 +2171,18 @@ public actor GraphStore {
 
   /// Kills a loop's session and brings it back on the same transcript — see
   /// `GraphCommand.restartNode`. A resolved loop has no session worth bringing back and
-  /// a stopped one was told to stay down, so both are refused rather than revived.
+  /// a stopped one was told to stay down, so both are refused rather than revived — except
+  /// a loop stopped for a missing CLI, which nobody told to stay down and whose restart
+  /// is exactly what its dialog asks the human for once the CLI is installed.
   private func restartNode(_ nodeID: UUID) async {
     guard let node = graph.nodes[id: nodeID] else {
       announceError("no loop \(nodeID) in this graph")
       return
     }
+    if node.state == .stopped, let failure = node.launchFailure {
+      relaunchAfterMissingProvider(node, failure)
+      return
+    }
     guard !node.isResolved else {
       announceError("\(node.title) has finished — there is no session to restart")
       return
@@ -2241,6 +2302,15 @@ public actor GraphStore {
   /// memory, so a state nobody expected can be traced to the report that caused it.
   private func sessionPermitsResolution(_ nodeID: UUID, succeeded: Bool) async -> Bool {
     guard let node = graph.nodes[id: nodeID], !node.isResolved else { return true }
+    // An agent the launch shell could not find exits at once, which a pane reports exactly
+    // like an agent that finished — the loop resolved SUCCEEDED having never run. Asked
+    // before the restart grace, because a restart whose CLI is still missing exits in it.
+    if succeeded, let onFindMissingProvider,
+      let failure = await onFindMissingProvider(node, graph.project.path)
+    {
+      stopForMissingProvider(nodeID, failure)
+      return false
+    }
     let report =
       "surface reported its pane "
       + (succeeded ? "finished" : "closed with its process still running")
@@ -3525,7 +3595,9 @@ public actor GraphStore {
   /// existing session first — `zmx run` itself is *not* idempotent, and re-running it
   /// against a live session types the prompt in a second time.
   public func ensureUnattendedSessions() {
-    for node in graph.nodes where node.runsUnattended {
+    // A loop stopped for a missing CLI waits for the human's restart: relaunching it at
+    // boot would only reach the same missing CLI and raise the same dialog.
+    for node in graph.nodes where node.runsUnattended && node.launchFailure == nil {
       if node.loopType == .goalBased {
         guard !node.isResolved else { continue }
         armGoalPoller(for: node)

GraphcodeKit/Sources/ProjectRegistry.swift

@@ -642,6 +642,9 @@ public actor ProjectRegistry {
         Task { await self?.refreshAwakeAssertion() }
       },
       onEnsureSession: ensureSession,
+      onFindMissingProvider: { node, path in
+        await ProviderPath.missingProvider(for: node, projectPath: path)
+      },
       onTerminateSession: terminateSession,
       onRestartSession: restartSession,
       onEvaluatePredicate: evaluatePredicate,

GraphcodeKit/Sources/Sessions/ProviderPath.swift

@@ -0,0 +1,94 @@
+import Foundation
+
+/// Why the daemon stopped a loop whose backend CLI the launch shell could not find —
+/// carried on the node (`LoopNode.launchFailure`) so every client can say so, and so
+/// `GraphStore.restartNode` knows this stop is one the human is allowed to undo.
+public struct LaunchFailure: Codable, Equatable, Sendable {
+  public var executable: String
+  public var backend: CLISessionBackendKind
+  public var occurredAt: Date
+
+  public init(executable: String, backend: CLISessionBackendKind, occurredAt: Date = Date()) {
+    self.executable = executable
+    self.backend = backend
+    self.occurredAt = occurredAt
+  }
+
+  public var title: String { "\(executable) is not on your PATH" }
+
+  public var message: String {
+    "GraphCode couldn't find \(executable), the \(backend.displayName) command-line tool, so "
+      + "the loop was stopped instead of left running without an agent. Loops start their "
+      + "agent from a login shell (/bin/zsh -i -l): install \(backend.displayName), or add "
+      + "the folder containing \(executable) to PATH in ~/.zshrc or ~/.zprofile, then "
+      + "restart the loop."
+  }
+}
+
+/// Whether a backend's CLI resolves the way a session launch resolves it. Without this a
+/// missing CLI produced a session whose shell exited 127 at once while the graph went on
+/// reporting the loop as running — or, reported by its pane, as SUCCEEDED.
+public enum ProviderPath {
+  /// What zsh exits with for a command it cannot find.
+  public static let commandNotFoundStatus = 127
+
+  /// The same `-i -l` shell the launches use (`ZmxSessionLauncher.loginShellInvocation`,
+  /// `GhosttyTerminalView.interactiveLoginShell`), since a developer's PATH usually comes
+  /// from `~/.zshrc`. `whence -p` rather than `command -v`: the launch `exec`s the agent,
+  /// which only a file on PATH satisfies, so an alias of the same name must not count.
+  public static func probeInvocation(for executable: String) -> [String] {
+    [
+      "/bin/zsh", "-i", "-l", "-c",
+      "whence -p -- \(RemoteProjectLocation.shellQuoted(executable)) >/dev/null 2>&1",
+    ]
+  }
+
+  /// `nil` when the shell did not answer in time: a slow `~/.zshrc` says nothing about
+  /// PATH, and must never be what stops a loop.
+  public static func isOnPath(_ executable: String, deadline: Duration = .seconds(15)) async
+    -> Bool?
+  {
+    if await FoundCache.shared.isFresh(executable) { return true }
+    let invocation = probeInvocation(for: executable)
+    guard
+      let shell = invocation.first,
+      let session = try? PTYProcessSession(
+        executable: shell, arguments: Array(invocation.dropFirst()))
+    else { return nil }
+    guard let found = await withDeadline(deadline, { await session.waitUntilFinished() }) else {
+      session.terminate()
+      return nil
+    }
+    if found { await FoundCache.shared.record(executable) }
+    return found
+  }
+
+  /// The failure launching `node` would hit, or `nil`. Always `nil` for a remote project:
+  /// its PATH belongs to another machine, which this shell cannot see.
+  public static func missingProvider(for node: LoopNode, projectPath: String?) async
+    -> LaunchFailure?
+  {
+    if let projectPath, RemoteProjectLocation.parse(projectPath: projectPath) != nil {
+      return nil
+    }
+    guard let executable = node.backend.executableName else { return nil }
+    guard await isOnPath(executable) == false else { return nil }
+    return LaunchFailure(executable: executable, backend: node.backend)
+  }
+
+  /// Found answers only, and briefly: a daemon loading a graph ensures every unattended
+  /// loop at once, and one login shell per loop for the same answer is waste. A missing
+  /// CLI is never cached, so the check after a fix sees the fix.
+  private actor FoundCache {
+    static let shared = FoundCache()
+    private static let lifetime: TimeInterval = 300
+    private var foundAt: [String: Date] = [:]
+
+    func isFresh(_ executable: String) -> Bool {
+      guard let found = foundAt[executable] else { return false }
+      return Date().timeIntervalSince(found) < Self.lifetime
+    }
+
+    func record(_ executable: String) { foundAt[executable] = Date() }
+  }
+}

graphcode/Sources/Features/App/AppFeature+LoopSessions.swift

@@ -19,6 +19,11 @@ struct SessionRestart: Equatable {
   var isConfirmingAll = false
   /// The workspace closed for a restart, to be remounted once the daemon confirms.
   var pendingReopen: PendingReopen?
+  /// A loop the daemon stopped because its CLI is not on PATH — see `LaunchFailure`.
+  var launchFailureNotice: LaunchFailureNotice?
+  /// Each node's failure already raised, so a snapshot that still carries it does not
+  /// raise it again on every broadcast.
+  var announcedLaunchFailures: [UUID: Date] = [:]
 
   struct PendingReopen: Equatable {
     var projectPath: String
@@ -28,12 +33,23 @@ struct SessionRestart: Equatable {
     var seenRestarts: Int
   }
 
+  struct LaunchFailureNotice: Equatable {
+    var title: String
+    var message: String
+
+    init(node: LoopNode, failure: LaunchFailure) {
+      title = failure.title
+      message = "“\(node.title)” was stopped. \(failure.message)"
+    }
+  }
+
   @CasePathable
   enum Action: Equatable {
     case openLoopTapped
     case allTapped
     case allConfirmed
     case allCancelled
+    case launchFailureNoticeDismissed
   }
 }
 
@@ -116,7 +132,12 @@ extension AppFeature {
           }
         }
 
+      case .sessionRestart(.launchFailureNoticeDismissed):
+        state.sessionRestart.launchFailureNotice = nil
+        return .none
+
       case .daemonEvent(.graphChanged(let graph)):
+        announceLaunchFailures(in: graph, &state)
         guard let pending = state.sessionRestart.pendingReopen,
           pending.projectPath == graph.project.path
         else { return .none }
@@ -144,6 +165,23 @@ extension AppFeature {
     }
   }
 
+  /// A snapshot is how the app learns the daemon stopped a loop for a missing CLI,
+  /// whichever process launched its session. A project's first snapshot only records
+  /// what it already carries: those loops say why on their cards, and a stack of alerts
+  /// at launch would be about the past.
+  private func announceLaunchFailures(in graph: LoopGraph, _ state: inout State) {
+    let firstSight = state.projects[id: graph.project.path] == nil
+    for node in graph.nodes {
+      guard let failure = node.launchFailure,
+        state.sessionRestart.announcedLaunchFailures[node.id] != failure.occurredAt
+      else { continue }
+      state.sessionRestart.announcedLaunchFailures[node.id] = failure.occurredAt
+      guard !firstSight, state.sessionRestart.launchFailureNotice == nil else { continue }
+      state.sessionRestart.launchFailureNotice = SessionRestart.LaunchFailureNotice(
+        node: node, failure: failure)
+    }
+  }
+
   private func closeForRestart(_ state: inout State, reopening open: LoopWorkspaceFeature.State) {
     let current =
       state.projects[id: open.projectPath]?.graph.nodes[id: open.node.id]?.sessionRestarts

graphcode/Sources/Features/App/AppFeature.swift

@@ -579,10 +579,7 @@ struct AppFeature {
         }
         closeOpenWorkspace(&state)
         state.selectedProjectPath = projectPath
-        return .run { _ in
-          try? await orchestratorClient.send(
-            .graphCommand(projectPath: projectPath, command: .deleteNode(id)))
-        }
+        return deleteAcknowledgedLoop(id, in: projectPath, state)
 
       // Closing the workspace's last tab — by its x, by ⌘W, or by a plain shell simply
       // exiting. There is nothing left to show, which for a loop means ending the loop,

graphcode/Sources/Features/App/AppView.swift

@@ -255,6 +255,9 @@ struct AppView: View {
     // Why a tap on a gated loop did nothing — same hosting rule again: the tap can
     // come from the sidebar or ⇧⌘] while any detail pane is up.
     .modifier(BlockedLoopDialog(store: store))
+    // A loop stopped because its CLI is not on PATH — the daemon finds it whichever
+    // project the loop is in, so it is hosted with the rest.
+    .modifier(LaunchFailureDialog(store: store))
   }
 
   /// Folders past their worktree notice threshold, for the titlebar chip. Policies

graphcode/Sources/Features/App/LaunchFailureDialog.swift

@@ -0,0 +1,41 @@
+import ComposableArchitecture
+import Foundation
+import GraphcodeKit
+import SwiftUI
+
+/// The alert for `SessionRestart.launchFailureNotice`.
+struct LaunchFailureDialog: ViewModifier {
+  let store: StoreOf<AppFeature>
+
+  func body(content: Content) -> some View {
+    content
+      .alert(
+        store.sessionRestart.launchFailureNotice?.title ?? "",
+        isPresented: Binding(
+          get: { store.sessionRestart.launchFailureNotice != nil },
+          set: { if !$0 { store.send(.sessionRestart(.launchFailureNoticeDismissed)) } }
+        )
+      ) {
+        Button("OK") { store.send(.sessionRestart(.launchFailureNoticeDismissed)) }
+      } message: {
+        Text(store.sessionRestart.launchFailureNotice?.message ?? "")
+      }
+  }
+}
+
+extension AppFeature {
+  /// The deletion a key on a dead agent pane stands for (`.primaryExitAcknowledged`) —
+  /// except for a loop stopped for a missing CLI, which is waiting for the restart its
+  /// dialog asked for. The key only puts that pane away.
+  func deleteAcknowledgedLoop(_ id: UUID, in projectPath: String, _ state: State)
+    -> Effect<Action>
+  {
+    guard state.projects[id: projectPath]?.graph.nodes[id: id]?.launchFailure == nil else {
+      return .none
+    }
+    return .run { _ in
+      try? await orchestratorClient.send(
+        .graphCommand(projectPath: projectPath, command: .deleteNode(id)))
+    }
+  }
+}