AgentSkillsCN

human-approval-gate

>

SKILL.md
--- frontmatter
name: human-approval-gate
description: >
  Before applying changes (or before merging), check the Risk Level.
  ONLY produce a Human Approval Sheet for "Level 2: Approve" changes.
  If anything is unclear, STOP and ask Human instead of guessing.
inputs:
  - Task Intake (recommended)
  - Diff summary or planned change summary
  - Tests/bench results (if already run)
outputs:
  - docs/approvals/APPROVAL-<YYYYMMDD>-<slug>.md (Only for Level 2)
rules:
  - **Approval Policy (3 Levels)**:
    - **Level 0: Auto** (No gate needed): Cleanup, comments, docs fix, internal refactor.
    - **Level 1: Notify** (Report only): Minor UI tweaks, non-breaking additions.
    - **Level 2: Approve** (MUST use this skill):
      - Changes to Coordinates/Units/Grid.
      - Crossing SceneContext boundary / Direct ECS access.
      - Affects State Transitions / Dirty Resolution.
      - Affects Undo/Redo / Persistence.
      - SRBA Layers (A-F) or new modules (axiom/composition/disposition).
      - Concurrency/Locking changes.
      - User Experience flow changes.
  - Must be understandable without reading code.
  - If design/spec is unclear, do not proceed; ask <=3 questions with A/B/C options.

human-approval-gate

Output format (required)

Approval Sheet: <title>

  • •Date:
  • •Approver: Human
  • •Related Task Intake:
  • •Related START_HERE / design docs:

What we are trying to achieve (1-3 lines)

  • •...

What we will NOT do (Non-goals)

  • •...

Scope / Impact (bullets)

  • •Areas touched (high-level): ...
  • •User-visible changes: ...
  • •Data/compat risks (if any): ...

Design alignment (checklist)

  • • Does not violate core invariants (coords/units, SceneContext boundary, state rules)
  • • If it changes design, docs will be updated first OR an ADR will be added
  • • No “silent” scope expansion

Validation (what was checked)

  • •Tests run:
  • •Manual checks / screenshots:
  • •Performance notes (if relevant):

Risks (bullets)

  • •...

STOP-if-unclear questions (<=3, A/B/C choices)

  1. •...
  2. •...
  3. •...

Decision (Human)

  • • APPROVE (proceed)
  • • REJECT (revise)
  • • HOLD (need more info)