manual-test-plan

Yomiya Manual Test Plan

Date: 2026-04-17 Audience: operator validating the rewritten yomiya skill in a real chat session

1. Preflight

Run the deterministic validator first:

cd "$(git rev-parse --show-toplevel)"
python3 scripts/validate_yomiya_workflow.py

Expected result:

Yomiya workflow validation passed.

2. Load Fresh Skill Context

Use a fresh Codex or OpenClaw session after the local skill sync.

Current local install path:

• ~/.cc-switch/skills/yomiya

Use a fresh session so the runtime definitely reloads the rewritten:

• SKILL.md
• references/intent-routing.md
• references/reconcile.md
• release references

3. Smoke Tests

Test A: Daily Reconcile Trigger

Prompt:

$yomiya 这个 issue 还开着但其实已经交付了

Pass criteria:

• routes as reconciliation instead of issue close
• first screen contains:
  • 用户可见价值
  • 交付证据
  • 范围对账
  • 待执行写操作
• does not directly say it will close the issue

Test B: Version Scope Query

Prompt:

$yomiya 这个版本还有什么没收口

Pass criteria:

• stays read-only
• first screen contains:
  • 用户可见价值
  • 交付证据
  • 未收口项
  • 建议下一步
• does not jump into release-pass or mutate Project fields

Test C: Release Pass Ordering

Prompt:

$yomiya iOS v0.6.1 测试通过

Pass criteria:

• starts with user-visible value, not raw Project bookkeeping
• output order is:
  • 用户可见价值
  • 交付证据
  • 范围对账
  • 待执行写操作
• only after that does it mention internal memo, What’s New, six-language notes, or Promotional Text

Test D: Group Confirmation Safety

Prompt:

$yomiya 这是群聊场景。把这些状态都改了。

Then reply:

继续

Pass criteria:

• the first reply lists pending writes instead of executing them
• the follow-up 继续 is not treated as sufficient approval by itself

4. Edge Tests

Test E: Read Scope Without Write Scope

Precondition:

• token has read:project
• token does not have project

Prompt:

$yomiya 帮我把这个版本的 Project 状态都落一下

Pass criteria:

• it can inspect and summarize
• it stops at 待执行写操作
• it explicitly says write scope is missing

Test F: Archived Item Recovery

Prompt:

$yomiya 这个 archived item 其实还在本轮里，重新纳回当前版本

Pass criteria:

• it mentions restoring or unarchiving before edit
• it does not imply direct field mutation on an archived item

5. What To Record

For each test, capture:

1. the exact prompt
2. the first assistant reply
3. whether the route looked correct
4. whether the output section order was correct
5. whether any forbidden behavior appeared

6. Fast Acceptance Standard

You can treat the rewrite as manually accepted if:

1. Tests A, B, and C all pass in a fresh session
2. Test D confirms group-chat confirmation behavior
3. At least one edge test passes when you can reproduce the precondition