Sdd Agents Edit

---
name: sdd-agents-edit
description: Quản lý và cập nhật AGENTS.md: Agent Constitution, vai trò, phạm vi và ma trận quyền công cụ
user-invocable: true
---

# SDD Agent Constitution Editor (`/sdd-agents-edit`)

**Output language:** Mirror the invoking prompt. Canonical tokens, paths and commands remain language-invariant.

Dùng khi cập nhật `AGENTS.md`: vai trò, scope, tool permission, security hoặc escalation của AI Agent. Skill này chỉ quản lý governance sau review; không thay thế feature pipeline hoặc `/add-execute`, không cấp execution grant và không được dùng để nới quyền qua shortcut.

## Khi nào dùng / không dùng

- **Dùng khi:** thay đổi có chủ đích đối với Agent Constitution, permission, security hoặc escalation.
- **Không dùng khi:** sửa source/feature behavior, thay đổi Constitution Layer 1/2 (dùng `/sdd-rfc`), hoặc cần bỏ qua review/checkpoint.

## Parameters

- `--section=<name>`: Optional canonical section: `identity`, `scope`, `tool-permissions`, `security`, `communication`, `error-handling`, `escalation`, `changelog`.
- `--reason=<reason>`: Required reason for permission or rule update.

## Canonical structure

`AGENTS.md` has exactly eight ordered sections: Identity & Persona, Scope & Boundaries, Tool Permissions, Security Rules, Communication Style, Error Handling, Escalation Protocol and Changelog. Modify only the applicable section. Customize generated content to observed stack, paths and exact commands; never copy generic stack commands as evidence.

## Ownership and execution model

Read `.sdd/shared_context.md` before changing roles:

- `Project Ownership: solo` means one Human project owner. It may simplify Human accountability, but never limits agent roles or prevents worker orchestration.
- `Project Ownership: team` means multiple Human collaborators.
- `Agent Execution: direct` executes in the current agent session.
- `Agent Execution: orchestrated` allows `/add-execute` to coordinate workers with exclusive boundaries after observed runtime capability.

Agent roles, worker count, Lead→worker escalation and ownership boundaries depend on approved execution/tasks, not Human ownership count. Direct execution retains Human gates, profile evidence, checkpoint and no-push restrictions.

## Procedure

1. Read `AGENTS.md`, `CONSTITUTION.md`, `CLAUDE.md`, Architecture Profile, shared context and relevant constraints.
2. Identify the canonical section; do not expand permissions beyond safety boundaries.
3. Update only that section and record execution/ownership effects where relevant.
4. Bump SemVer and add a Changelog entry with reason and updater.
5. For execution-orchestration changes, distinguish policy from observed host enforcement; record affected role, no-self-approval/no-commit-push, bounded retry and escalation behavior.
6. Report permission impact, security risk, affected skill and escalation step.

## AI Recommendation and Human Final Review

Before editing `AGENTS.md`, create the canonical recommendation at `.sdd/reviews/agents-edit.md` with permission impact, security risk, execution/ownership effect, alternative and affected skills. A Human review must be `APPROVED` before editing. A solo owner may persist their own Human Final Review; in team ownership an authorized Human collaborator persists it. Agent never self-approves permissions.

## Hướng dẫn sử dụng
- **Khi dùng:** Đề xuất bounded governance change trong một section của `AGENTS.md`.
- **Không dùng:** Không dùng cho source/feature behavior, Constitution Layer 1/2, hoặc để nới quyền execution shortcut.
- **Input:** `--reason` bắt buộc; tùy chọn `--section`; reject requests that expand permissions without evidence or review.
- **Điều kiện trước:** `AGENTS.md`, governance, profile and constraints readable; recommendation must be `PENDING HUMAN REVIEW` before edit.
- **Evidence:** `.sdd/reviews/agents-edit.md`, changed section, permission/security impact, and refreshed recommendation.
- **Dừng khi:** Missing approval, unsafe permission expansion, conflicting profile, or request crosses Constitution/RFC boundary.
- **Human quyết định:** Authorized Human reviews and approves permission impact via `/sdd-review`; Agent only applies approved bounded edit.
- **Lệnh tiếp theo:** `/sdd-review --target=.sdd/reviews/agents-edit.md --status=APPROVED ...`, then the persisted Follow-up.
- **Ví dụ:** `/sdd-agents-edit --section=tool-permissions --reason="Document observed read-only audit boundary"`.

## Completion output

Use the [Completion output contract](../_shared/ai-review-protocol.md#completion-output-contract). Report affected section, permission/security impact, affected skills and persisted review path. Before edit, choose `Human decision required`; an authorized Human reviewer records the selected non-placeholder decision, reviewer identity and persisted Follow-up with `/sdd-review --target=.sdd/reviews/agents-edit.md --status=<APPROVED|REVISE|REJECTED> --decision="<human decision>" --reviewer="<authorized human reviewer>" --follow-up="<exact next command or required action>"`. After bounded edit, report the refreshed recommendation and required review state. Missing approval or an unsafe permission expansion is `BLOCKED`; do not modify unrelated sections or self-approve.