Refuse lead tools that would restart an interrupted or errored agent

internal/agents/lead.go

@@ -45,6 +45,10 @@ func leadActor(login string) string {
 	return login + " (lead)"
 }
 
+func isLeadActor(actor string) bool {
+	return strings.HasSuffix(actor, " (lead)")
+}
+
 func humanActor(actor string) bool {
 	return actor != "byre" && actor != agentActor && actor != boardActor && !strings.HasSuffix(actor, leadActor(""))
 }

internal/agents/leadtools.go

@@ -111,7 +111,7 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 		tool("answer_question",
 			"Answer an open question on the human's behalf. Only when the answer is clear or the human told you it. "+
 				"type option: options holds the chosen choices' exact text (one for pick-one), and text may add notes."+
-				"type other: free text.",
+				"type other: free text."+stoppedRefusal,
 			`{"type": "object", "properties": {
 				"item_id": {"type": "string"},
 				"type": {"type": "string", "enum": ["option", "other"]},
@@ -124,6 +124,18 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 				if err != nil {
 					return "", err
 				}
+				item, err := m.store.GetInboxItem(ctx, c.ItemId)
+				if err != nil {
+					return "", err
+				}
+				sess, err := m.store.GetSession(ctx, item.AgentSessionId)
+				if err != nil {
+					return "", err
+				}
+				err = m.checkNotStopped(ctx, sess)
+				if err != nil {
+					return "", err
+				}
 				m.mu.Lock()
 				m.leadAnswered[c.ItemId] = true
 				m.mu.Unlock()
@@ -141,7 +153,8 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 			"Send a message to a task's agent; it starts a turn. Starts the agent when the task has none. "+
 				"If the agent is mid-turn, the message is queued like the human's: the agent reads it with its next tool call's result, "+
 				"or as its next turn if it makes no further tool call, and gets it only once. "+
-				"On a team task, the message is posted to the team's board instead, so every agent of the team gets it the same way; model is refused there.",
+				"On a team task, the message is posted to the team's board instead, so every agent of the team gets it the same way; model is refused there, "+
+				"and so is the post while any agent of the team is stopped."+stoppedRefusal,
 			`{"type": "object", "properties": {
 				"task_id": {"type": "string"},
 				"text": {"type": "string"},
@@ -172,6 +185,10 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 					if args.Model != "" {
 						return "", fmt.Errorf("%w: model can't be given for a team task; its agents keep their own models", ErrInvalid)
 					}
+					err = m.checkTeamNotStopped(ctx, task.Id)
+					if err != nil {
+						return "", err
+					}
 					err = m.Post(ctx, task, leadActor(user.Login), args.Text)
 					if err != nil {
 						return "", err
@@ -359,7 +376,7 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 				return "Interrupted.", nil
 			}),
 
-		tool("compact_task", "Have a task's idle agent write a handoff and continue in a fresh session.",
+		tool("compact_task", "Have a task's idle agent write a handoff and continue in a fresh session."+stoppedRefusal,
 			`{"type": "object", "properties": {
 				"task_id": {"type": "string"},
 				"model": {"type": "string", "description": "Start the fresh session on this model, on either vendor; the current agent still writes the handoff. Defaults to the current model."}
@@ -376,6 +393,10 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 				if err != nil {
 					return "", err
 				}
+				err = m.checkNotStopped(ctx, sess)
+				if err != nil {
+					return "", err
+				}
 				_, err = m.Compact(ctx, sess.Id, args.Model)
 				if err != nil {
 					return "", err
@@ -387,7 +408,7 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 			"Merge a task as the human: its agent merges the default branch in, then byre squashes it into one commit, pushes it to origin and closes the task. "+
 				"While another task in the project is merging, byre holds the merge and starts it when that one is done. "+
 				"For a subtask, target parent instead asks its parent task's agent to merge it into the parent's branch, and the subtask closes once the parent has. "+
-				"Refused for a task with active subtasks. Only when the human asks. You get a signal when it is merged; failures show as errors in the task's transcript, and you get a signal for each.",
+				"Refused for a task with active subtasks. Only when the human asks. You get a signal when it is merged; failures show as errors in the task's transcript, and you get a signal for each."+stoppedRefusal,
 			`{"type": "object", "properties": {
 				"task_id": {"type": "string"},
 				"target": {"type": "string", "enum": ["main", "parent"], "description": "Default main."}
@@ -406,7 +427,17 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 				}
 				switch args.Target {
 				case "", "main":
+					err = m.checkTaskNotStopped(ctx, task.Id)
+					if err != nil {
+						return "", err
+					}
 				case "parent":
+					if task.ParentTaskId != nil {
+						err = m.checkTaskNotStopped(ctx, *task.ParentTaskId)
+						if err != nil {
+							return "", err
+						}
+					}
 					_, err = m.MergeIntoParent(ctx, task, user)
 					if err != nil {
 						return "", err
@@ -440,13 +471,17 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
 
 		tool("discard_task",
 			"Discard a task: its agent gets one turn to write the task summary, then byre stops it and deletes its worktree and branch without merging. This destroys its unmerged work. Refused for a task with active subtasks. Only when the human asks. "+
-				"If the agent is mid-turn, the summary turn starts when that turn ends; no other turn runs in between.",
+				"If the agent is mid-turn, the summary turn starts when that turn ends; no other turn runs in between."+stoppedRefusal,
 			taskOnly,
 			func(ctx context.Context, user api.User, raw json.RawMessage) (string, error) {
 				task, _, err := m.taskArg(ctx, raw)
 				if err != nil {
 					return "", err
 				}
+				err = m.checkTaskNotStopped(ctx, task.Id)
+				if err != nil {
+					return "", err
+				}
 				_, err = m.Discard(ctx, task, user.Login)
 				if err != nil {
 					return "", err
@@ -784,6 +819,38 @@ func (m *Manager) liveSession(ctx context.Context, taskID string) (api.AgentSess
 	return sess, true, nil
 }
 
+const stoppedRefusal = " Refused while the agent is stopped: the human interrupted its last turn, or that turn ended in an error, and no turn has started since. Then put it to the human in chat instead."
+
+func (m *Manager) checkTaskNotStopped(ctx context.Context, taskID string) error {
+	sess, found, err := m.liveSession(ctx, taskID)
+	if err != nil || !found {
+		return err
+	}
+	return m.checkNotStopped(ctx, sess)
+}
+
+func (m *Manager) checkTeamNotStopped(ctx context.Context, taskID string) error {
+	sessions, err := m.store.ListSessions(ctx, taskID)
+	if err != nil {
+		return err
+	}
+	for _, s := range teamSessions(sessions) {
+		err = m.checkNotStopped(ctx, s)
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (m *Manager) checkNotStopped(ctx context.Context, sess api.AgentSession) error {
+	stopped, err := m.store.SessionStopped(ctx, sess.TaskId, sess.Id)
+	if err != nil || !stopped {
+		return err
+	}
+	return fmt.Errorf("%w: the human interrupted task %s's agent (or its last turn ended in an error) and hasn't messaged it since, so byre won't start a turn on it for you; put what you wanted to send or do to the human in chat instead", ErrConflict, sess.TaskId)
+}
+
 func (m *Manager) taskArg(ctx context.Context, raw json.RawMessage) (api.Task, api.Project, error) {
 	args := struct {
 		TaskID string `json:"task_id"`

internal/agents/manager.go

@@ -881,7 +881,7 @@ func (m *Manager) pump(l *live) {
 		}
 	}
 	m.endMergeTurn(l.id, true)
-	next, _, err := m.store.EndTurn(ctx, l.id, agentActor, map[string]any{"error": "agent process exited mid-turn"}, nil, false, false)
+	next, _, err := m.store.EndTurn(ctx, l.id, agentActor, map[string]any{"error": "agent process exited mid-turn", "stopped": true}, nil, false, false)
 	if err != nil {
 		slog.Error("record session exit", "session", l.id, "err", err)
 	}
@@ -1186,8 +1186,12 @@ func (m *Manager) turnEnded(ctx context.Context, l *live, res endResult) error {
 	unlock()
 
 	held := len(l.subagents) > 0 || interrupted
+	var state map[string]any
+	if interrupted || res.isError {
+		state = map[string]any{"stopped": true}
+	}
 	turn := m.event(l, agentActor, "turn.ended", payload)
-	next, finalizeBy, err := m.store.EndTurn(ctx, l.id, agentActor, nil, &turn, review && !held, finished && !handoff)
+	next, finalizeBy, err := m.store.EndTurn(ctx, l.id, agentActor, state, &turn, review && !held, finished && !handoff)
 	if err != nil {
 		return err
 	}

internal/agents/queue.go

@@ -47,6 +47,12 @@ func (m *Manager) Message(ctx context.Context, sessionID, actor, text, model str
 	if err != nil {
 		return api.AgentSession{}, false, err
 	}
+	if sess.State == api.AgentSessionStateIdle && isLeadActor(actor) {
+		err = m.checkNotStopped(ctx, sess)
+		if err != nil {
+			return api.AgentSession{}, false, err
+		}
+	}
 	held := sess.State == api.AgentSessionStateIdle && sess.Role == api.Implementer && automatedActor(actor) && m.interrupted(sess.Id)
 	if !held && (sess.State != api.AgentSessionStateRunning || sess.Role != api.Implementer) {
 		sess, err = m.Send(ctx, sessionID, actor, text, model, images...)

internal/store/store.go

@@ -1061,6 +1061,24 @@ func (s *Store) TaskHandoffs(ctx context.Context, taskID string) ([]Handoff, err
 	return handoffs, rows.Err()
 }
 
+func (s *Store) SessionStopped(ctx context.Context, taskID, sessionID string) (bool, error) {
+	stopped := false
+	err := s.db.QueryRowContext(ctx, `
+		SELECT coalesce(json_extract(e.payload, '$.stopped'), 0) = 1 AND NOT EXISTS (
+			SELECT 1 FROM events a
+			WHERE a.task_id = e.task_id AND a.agent_session_id = e.agent_session_id AND a.seq > e.seq
+				AND a.kind = 'inbox.updated' AND json_extract(a.payload, '$.state') = 'answered'
+		)
+		FROM events e
+		WHERE e.task_id = ? AND e.agent_session_id = ? AND e.kind = 'session.state'
+		ORDER BY e.seq DESC LIMIT 1
+	`, taskID, sessionID).Scan(&stopped)
+	if errors.Is(err, sql.ErrNoRows) {
+		return false, nil
+	}
+	return stopped, err
+}
+
 func (s *Store) LeadEvents(ctx context.Context, owner string, after int64, limit int) ([]api.Event, error) {
 	rows, err := s.db.QueryContext(ctx, `
 		SELECT `+eventColumns+`