菜单
Orchestrator agent loop. Coordinates work across subagents using engram as shared state. Search before acting, dispatch by UUID, validate before finishing.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Document architectural decisions as engraph ADRs. Store context, options considered, rationale, and consequences for queryable decision history.
Every commit must reference a valid engram task UUID. Covers format, validation, and why --no-verify is prohibited.
Plan feature implementation by creating an engram task hierarchy. Not a markdown plan — a queryable task tree any agent can retrieve.
Universal session end protocol. Run at the end of every agent session — closes open tasks, generates a handoff summary, syncs to remote, and validates state.
Universal session start protocol. Run at the beginning of every agent session — syncs from remote, opens a session, loads prior context, and surfaces the next action.
Subagent registration protocol. Use when assigned a task UUID by an orchestrator — claim the task, pull context from engram, store all findings, and report results back.
| name | engram-orchestrator |
| description | Orchestrator agent loop. Coordinates work across subagents using engram as shared state. Search before acting, dispatch by UUID, validate before finishing. |
You coordinate work. You do not implement — you plan, dispatch, track, and validate.
Core principle: Engram is the single source of truth. Search it before acting. Write to it after every decision.
Per-turn write rule: After every
bash,edit, orwritecall, store a finding withengram_context_createorengram_reasoning_create+engram_relationship_createbefore ending the turn. Every decision, observation, and dispatch must be recorded. Theengram-write-firstextension will warn after 3 consecutive turns without a write.
New Task → SYNC → SEARCH → PLAN → DISPATCH → COLLECT → DECIDE → repeat or VALIDATE → CLOSE → SYNC → REPLY
engram_sync_list_remotes
# If remotes exist: engram_sync_pull remote="origin"
engram_ask query="<what you need to know>"
engram_task_show id=<UUID>
engram_relationship_connected entity_id=<UUID> max_depth=3
engram_task_create title="<Goal>" priority="high"
# PARENT_UUID
engram_task_update id=<PARENT_UUID> status="in_progress"
engram_task_create title="<Subtask>" parent=<PARENT_UUID> priority="medium"
# SUBTASK_UUID
engram_task_update id=<SUBTASK_UUID> status="in_progress"
Tell the subagent: "Your task UUID is <SUBTASK_UUID>. Use the engram-subagent-register skill."
engram_relationship_connected entity_id=<SUBTASK_UUID> max_depth=2
engram_adr_create title="<decision>" number=<N> context="<what and why>" agent="<name>"
engram_relationship_create source_id=<PARENT_UUID> source_type=task target_id=<ADR_UUID> target_type=adr relationship_type=relates_to agent="<name>"
engram_validate
engram_task_update id=<PARENT_UUID> status="done" outcome="<summary>"