verify-replay

Replay Checklist Solidification

Solidify a successfully completed verification execution into a YAML step checklist. Default output to:

docs/verification/replay.yml

When there is only a URL and no repository, write to the current working directory replay.yml。

Input threshold (fail closed)

• The input is the verification report produced by verify-run (docs/verification/verify-run-report.md or a user-specified path).
• When the report contains any fail itemrefuse to solidify, and clearly inform: the replay checklist is an execution artifact, and it must be fixed, rerun, and all green before solidifying. This is a hard threshold; do not do it even if the user insists.
• When the report does not exist, prompt to run verify-run first.

YAML Format

One entry per step, with fixed fields:

# 回放清单：由 verify-replay 于 2026-10-10 从 verify-run 报告固化
# 来源报告：docs/verification/verify-run-report.md（全绿）
meta:
  target: http://localhost:5173        # 被验证目标（URL 或本地服务说明）
  setup:                               # 重放前的环境准备（启动命令等，可空）
    - bash scripts/dev.sh
  teardown:                            # 重放后的清理（可空）
    - 停止 dev 服务

steps:
  - id: 1
    action: open                       # open | click | fill | upload | curl | run | wait | screenshot
    target: http://localhost:5173/login
    expect: 页面出现"登录"标题           # 断言描述，agent 重放时据此判定
    evidence: screenshots/01-login.png # 期望留存的证据（可选）
  - id: 2
    action: fill
    target: 用户名
    value: demo@example.com
  - id: 3
    action: curl
    target: GET http://localhost:8790/api/health
    expect: "HTTP 200，响应含 {\"status\":\"ok\"}"

Solidification rules:

• Only include the verification points that passed in the report and their operation paths; skip items are appended at the end in comment form and do not enter steps.
• target / URL / selector descriptions must all use the values from this execution that wereactually verified, and do not retroactively change them into idealized wording.
• Assertion descriptions retain human-eye-level judgments ("order list appears", "no console errors") and are not forcibly converted into hardcoded assertions—during replay, the agent makes the judgment.
• Add as few waits as possible; for steps that truly need waiting, write the condition concretely ("X text appears" instead of "wait 3 seconds").

Replay Method

Replay is not executed by this skill. When the user wants to replay:

• Directly drive opencli / curl execution step by step according to replay.yml, with each step judged by the agent using expect, and evidence retained according to the evidence field.
• By default, replay results are written to docs/verification/replay-run-report.md(overwrite update).
• When a step fails, record it and continue with subsequent steps, then summarize at the end; if the same step fails across multiple consecutive rounds, it indicates the target has drifted, so prompt the user to rerun verify-run and resolidify, rather than repeatedly patching the YAML.

Output Requirements

• Use Chinese.
• After generation is complete, list in the final reply: the replay.yml path, the number of included steps, skipped items, and the invocation method for replay.