dag-orchestrator

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.

---

Phase 0 — Load plan + project snapshot

smart-mcps-agentmemory profile

Summarize: active frontier size, top concepts, recent sessions, any pending tasks from prior runs.

---

Phase 1 — Create DAG from plan

Decompose the plan into atomic tasks:

• Each task = one independently executable unit with a clear definition of done
• Identify requires dependencies (task B cannot start until task A is done)
• Assign priorities: critical-path nodes get 8-9, leaf tasks get 5-6
• Tags should encode domain (e.g. auth, testing, refactor)
• description field = plain-text objective only — no context dumps

Create 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.

---

Phase 2 — Execution loop

smart-mcps-agentmemory next --limit 5

For each action in actions[] (all have empty blockers):

1. Compose the worker prompt using the template below, substituting ACTION_ID, TITLE, DESCRIPTION, and the context block from the next output.
2. Spawn a Worker Agent (Agent tool) with the composed prompt.
3. Workers with no shared dependencies run in parallel (single Agent multi-call message).
4. Workers that transitively share mutable state should be sequenced.

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.

---

Phase 3 — Closure

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).

---

Worker prompt template

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.