feat(tasks): background TaskRunner executor + expanded task states

cmd/chat_tools.go

@@ -96,6 +96,7 @@ func optionalTools() []tool.Tool {
 		tool.TaskGetTool{},
 		tool.TaskListTool{},
 		tool.TaskUpdateTool{},
+		tool.TaskRunTool{},
 		tool.SleepTool{},
 		tool.CronCreateTool{},
 		tool.CronDeleteTool{},

internal/engine/session.go

@@ -2,6 +2,7 @@ package engine
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"log/slog"
 	"os"
@@ -184,6 +185,7 @@ func NewSessionWithClient(chat ChatClient, provider, model, systemPrompt string,
 		recordVerification: s.recordVerificationObservation,
 		lifecycle:          s.life,
 		appendSystem:       s.AppendSystemContext,
+		taskExec:           s.taskExecFromAgentSpawn(),
 	})
 	s.refreshContextWindowCache()
 	s.life.SetAgentsAccumulator(agentsAccum)
@@ -209,6 +211,49 @@ func (s *Session) ReattachTransport(chat ChatClient, provider string, deployment
 	_ = deploymentRouting
 }
 
+// taskExecFromAgentSpawn returns the TaskRun executor: it spawns a general
+// sub-agent to perform a stored task, feeding the task's description, active
+// form, and checkpoint as the prompt. Nil agent-spawn capability disables the
+// executor (the TaskRun tool then reports that no executor is configured).
+func (s *Session) taskExecFromAgentSpawn() tool.TaskExecutorFunc {
+	return func(ctx context.Context, t *tool.Task) (string, error) {
+		if s == nil || s.tools == nil || s.tools.AgentSpawnFn() == nil {
+			return "", fmt.Errorf("task execution unavailable: no agent spawn capability")
+		}
+		prompt := "Execute the following task and report results.\n\nSubject: " + t.Subject +
+			"\n\nDescription: " + t.Description
+		if t.ActiveForm != "" {
+			prompt += "\n\n(You are working on: " + t.ActiveForm + ")"
+		}
+		if len(t.Checkpoint) > 0 {
+			b, _ := json.Marshal(t.Checkpoint)
+			prompt += "\n\nPrior progress (checkpoint): " + string(b)
+		}
+		res, err := s.tools.AgentSpawnFn()(ctx, agentcontracts.SpawnRequest{
+			Prompt:       prompt,
+			Description:  "Execute task " + t.ID,
+			SubagentType: "general",
+		})
+		if err != nil {
+			return "", err
+		}
+		if res.Status == agentcontracts.StatusFailed {
+			if res.Output != "" {
+				return "", fmt.Errorf("%s", res.Output)
+			}
+			return "", fmt.Errorf("task agent reported failure")
+		}
+		out := res.Output
+		if res.Summary != "" {
+			if out != "" {
+				out += "\n"
+			}
+			out += res.Summary
+		}
+		return out, nil
+	}
+}
+
 // SubSession clones transport and routing mode for explore/general sub-agents.
 func (s *Session) SubSession(model, systemPrompt string, registry *tool.Registry) *Session {
 	if registry == nil {

internal/engine/tool_service.go

@@ -70,6 +70,7 @@ type toolExecutionDeps struct {
 	recordVerification func(types.ToolCall, string, bool)
 	lifecycle          *LifecycleService
 	appendSystem       func(string)
+	taskExec           tool.TaskExecutorFunc
 }
 
 // NewToolService constructs a ToolService with the given registry.
@@ -395,6 +396,7 @@ func (s *ToolService) ExecuteOne(ctx context.Context, tc types.ToolCall, overrid
 		AvailableTools:      available,
 		Registry:            s.registry,
 		AutoCommit:          s.AutoCommit(),
+		TaskExecutor:        s.deps.taskExec,
 	})
 	// Bridge session sandbox policy onto the context so Bash/PowerShell
 	// WrapCommand actually applies. Path guards already read ToolContext.SandboxMode;

internal/tool/task_create.go

@@ -15,8 +15,11 @@ type TaskStatus string
 const (
 	TaskStatusPending    TaskStatus = "pending"
 	TaskStatusInProgress TaskStatus = "in_progress"
+	TaskStatusReviewing  TaskStatus = "reviewing"
 	TaskStatusCompleted  TaskStatus = "completed"
 	TaskStatusFailed     TaskStatus = "failed"
+	TaskStatusSkipped    TaskStatus = "skipped"
+	TaskStatusCancelled  TaskStatus = "cancelled"
 )
 
 // DefaultMaxAttempts is the retry budget used when a task does not declare
@@ -399,7 +402,7 @@ func (TaskUpdateTool) Parameters() map[string]interface{} {
 		"type": "object",
 		"properties": map[string]interface{}{
 			"taskId":       map[string]interface{}{"type": "string", "description": "The ID of the task to update"},
-			"status":       map[string]interface{}{"type": "string", "enum": []string{"pending", "in_progress", "completed", "failed"}, "description": "New task status"},
+			"status":       map[string]interface{}{"type": "string", "enum": []string{"pending", "in_progress", "reviewing", "completed", "failed", "skipped", "cancelled"}, "description": "New task status"},
 			"owner":        map[string]interface{}{"type": "string", "description": "Agent name to assign"},
 			"dependencies": map[string]interface{}{"type": "array", "items": map[string]interface{}{"type": "object", "properties": map[string]interface{}{"targetId": map[string]interface{}{"type": "string"}, "type": map[string]interface{}{"type": "string", "enum": []string{"blocks", "related", "parent-child"}}}}, "description": "Replace dependencies"},
 		},
@@ -422,9 +425,9 @@ func (TaskUpdateTool) Execute(_ context.Context, input json.RawMessage) (string,
 	}
 	ok := globalTaskStore.Update(p.TaskID, func(t *Task) {
 		if p.Status != "" {
-			// Moving a failed task back to pending is an explicit replan
+			// Any explicit status change out of failed is a replan or terminal
 			// signal: reset the retry budget and error so it starts fresh.
-			if TaskStatus(p.Status) == TaskStatusPending && t.Status == TaskStatusFailed {
+			if t.Status == TaskStatusFailed && TaskStatus(p.Status) != TaskStatusFailed {
 				t.Attempts = 0
 				t.LastError = ""
 			}