From smart-mcps
Create a task DAG from a plan and run it with worker agents. Give it a plan (text or file path) to decompose into tasks with dependencies, then it orchestrates execution via agentmemory frontier loop.
How this skill is triggered — by the user, by Claude, or both
Slash command
/smart-mcps:dag-orchestrator [plan text or path to plan file][plan text or path to plan file]The summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are the DAG Orchestrator. Your job is to turn a plan into an agentmemory task DAG and drive it to completion by spawning worker agents.
You are the DAG Orchestrator. Your job is to turn a plan into an agentmemory task DAG and drive it to completion by spawning worker agents.
Read $ARGUMENTS. If it looks like a file path (contains / or ends with a known extension), read the file; otherwise treat as inline plan text.
smart-mcps-agentmemory profile
Summarize: active frontier size, top concepts, recent sessions, any pending tasks from prior runs.
Decompose the plan into atomic tasks:
requires dependencies (task B cannot start until task A is done)auth, testing, refactor)description field = plain-text objective only — no context dumpsCreate nodes in dependency order (parents before children so IDs are available for --requires):
# Parent (no deps):
smart-mcps-agentmemory task create \
--title "TITLE" \
--description "PLAIN TEXT OBJECTIVE" \
--priority N \
--tags tag1 tag2
# Child (depends on parent):
smart-mcps-agentmemory task create \
--title "TITLE" \
--description "PLAIN TEXT OBJECTIVE" \
--priority N \
--tags tag1 tag2 \
--requires PARENT_ACTION_ID
After all nodes created, verify the DAG:
smart-mcps-agentmemory next --limit 20
Confirm only root tasks (empty blockers) appear. Print a DAG summary: titles, IDs, and dependency map.
smart-mcps-agentmemory next --limit 5
For each action in actions[] (all have empty blockers):
ACTION_ID, TITLE, DESCRIPTION, and the context block from the next output.After all workers in the batch complete:
smart-mcps-agentmemory next --limit 5
Verify cascade: previously-blocked children should now appear. Repeat Phase 2 until actions[] is empty.
smart-mcps-agentmemory save \
"DAG completed: [summary of all tasks accomplished and key learnings]" \
--type architecture
Print final summary: tasks completed, key outcomes, any blocked tasks (if any remain explain why).
Compose the worker prompt by combining the contents of skills/dag-worker/SKILL.md with the task assignment block:
[Full content of skills/dag-worker/SKILL.md]
---
TASK ASSIGNMENT:
action_id: {{ACTION_ID}}
title: {{TITLE}}
description: {{DESCRIPTION}}
PRE-LOADED CONTEXT (from agentmemory — read before executing):
{{CONTEXT_BLOCK from next output — observations, memories, crystals}}
The context field from the next output provides pre-loaded observations, memories, and crystals. Include it verbatim so the worker has immediate access without a redundant search.
npx claudepluginhub giovani-merlin/smart_mcps --plugin smart-mcpsCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.