UI Test Explore
Purpose
Use Midscene as the first visual exploration layer and preserve what it learns for later verification.
Operating Rules
- Midscene-first means first real UI exploration uses Midscene, not that Midscene alone proves success.
- Send only approved screenshots and visible page content to the model gateway.
- Never send credentials, cookies, tokens, hidden DOM, or private data to a model.
- Every Midscene action that claims success needs a planned deterministic verification.
- If Midscene returns
passed but URL, selected state, input value, route, visible content, or API evidence disagrees, mark the step degraded.
- Consume the generated Midscene View and current Source/build fingerprints; do not maintain a separate manually edited case body.
- A ready current-session ModuleReadyContext may start exploration inside the module and prohibits repeating stable module-entry clicks. It never authorizes formal Playwright to skip login/menu Flow.
- Persist exploration output only in a governed
D:\UI-Test\_tmp or run staging path selected by Path Planner.
Workflow
- Read the incoming
ui-test-packet.
- Convert test intent into concise Midscene prompts with operation, target, constraints, expected visible state, and forbidden actions.
- Prefer stage-based exploration: login state, project switch, module entry, page or modal open, form recognition, form fill, submit/result.
- Record successful strategy, failed strategy, reason, retry, and fallback candidate.
- Output updated packet with
next_action: verify or next_action: review when blocked.
Validation
- Prompt includes target, visible anchor, allowed operation, expected state, and forbidden actions.
- Repeated failures preserve root cause and changed variable.
- Exploration does not skip deterministic verification requirements.
Read the Midscene prompt rules before preparing model-facing operations.
Safety
Ask before sending private-page visible content to a model or performing an R2 action. Never send credentials or hidden state, and never treat exploration output as write approval or final business proof.
1---2name: ui-test-explore3description: Prepare and review Midscene-first UI exploration from a ui-test-packet. Use for semantic browser exploration, prompt packet shaping, weak locator discovery, visible-state observation, and preserving successful or failed exploration strategies before deterministic Playwright verification.4---56# UI Test Explore78## Purpose910Use Midscene as the first visual exploration layer and preserve what it learns for later verification.1112## Operating Rules1314- Midscene-first means first real UI exploration uses Midscene, not that Midscene alone proves success.15- Send only approved screenshots and visible page content to the model gateway.16- Never send credentials, cookies, tokens, hidden DOM, or private data to a model.17- Every Midscene action that claims success needs a planned deterministic verification.18- If Midscene returns `passed` but URL, selected state, input value, route, visible content, or API evidence disagrees, mark the step `degraded`.19- Consume the generated Midscene View and current Source/build fingerprints; do not maintain a separate manually edited case body.20- A ready current-session ModuleReadyContext may start exploration inside the module and prohibits repeating stable module-entry clicks. It never authorizes formal Playwright to skip login/menu Flow.21- Persist exploration output only in a governed `D:\UI-Test\_tmp` or run staging path selected by Path Planner.2223## Workflow24251. Read the incoming `ui-test-packet`.262. Convert test intent into concise Midscene prompts with operation, target, constraints, expected visible state, and forbidden actions.273. Prefer stage-based exploration: login state, project switch, module entry, page or modal open, form recognition, form fill, submit/result.284. Record successful strategy, failed strategy, reason, retry, and fallback candidate.295. Output updated packet with `next_action: verify` or `next_action: review` when blocked.3031## Validation3233- Prompt includes target, visible anchor, allowed operation, expected state, and forbidden actions.34- Repeated failures preserve root cause and changed variable.35- Exploration does not skip deterministic verification requirements.3637Read [the Midscene prompt rules](references/midscene-prompt-rules.md) before preparing model-facing operations.3839## Safety4041Ask before sending private-page visible content to a model or performing an R2 action. Never send credentials or hidden state, and never treat exploration output as write approval or final business proof.
Zengxiansheng007