planner

Files (1)

• SKILL.md

SKILL.md (first lines)

# Planner

**Language:** English | [Français](../../docs/fr/skills/planner/SKILL.md)

Activated automatically when `task.assess` classifies the goal as **complex**
(also selectable manually). For hierarchical Deep Thinking mode
(`cognitive_mode: deep_thinking`), prefer the **deep-thinking** skill and
`plan.create` / `plan.delegate_step` instead of flat `plan.update`.

1. Call `plan.update` first with atomic nodes — required before any side effect.
2. Mark independence: if two nodes do not need each other's output, they are **parallel**.
3. For each **heavy or parallel** node: `agent.spawn` with a **short self-contained brief** (≤3 sentences; limited skills/tools/docs — never dump parent tool results). Prefer parallel spawns over serial solo work when nodes are independent.
4. For each node / spawned brief: `memory.recall` on that narrow query (not the whole goal).
5. When a child finishes, the runtime injects `[child-done]` into the parent. `agent.await` is optional; integrate results as soon as they arrive. Keep sequential only when a node truly depends on another.
6. Never pass more caps/tools/docs to a sub-agent than necessary; briefs stay lean so the parent context does not bloat.
7. Finish with `goal.complete` when all nodes are Done.

From azerothl/akasha-os, MIT.

More from azerothl/akasha-os