WeChat Mini List Capture

Use this skill as a human-Codex operating loop, not as a blind crawler. Prefer Android ADB mode when the user does not want Windows mouse takeover. The user keeps the mini program visible in the chosen environment and performs login/tab switching when needed; Codex directly annotates screenshots, applies those annotations to config, runs small samples, and exports review artifacts.

Boundaries

• Use only user-visible, already-open UI state.
• Do not read, copy, or store cookies, tokens, headers, localStorage, private APIs, or credentials.
• Keep raw screenshots, OCR text, checkpoints, and logs under the ignored runtime directory.
• Put only status and sanitized review/export summaries in D:\todo.
• Treat OCR and AI outputs as candidates until human-reviewed.

Local Harness

Default local code path:

D:\agent-workflow-lab\harnesses\wechat-mini-list-capture

Default private runtime:

D:\agent-workflow-lab\runtime\wechat-mini-list-capture

Before running commands, read the harness README.md for current command usage and screenshot-review criteria. If a referenced references/harness-workflow.md exists in a future version, read that as the more specific runbook.

OfferShow List B Fast Profile

For the OfferShow list_b high-salary sample/comment workflow, do not rely on hand-edited runtime\wechat-mini-list-capture\config.local.json values. Before any full or resumed list_b run, apply the saved profile from the harness repo:

cd D:\agent-workflow-lab\harnesses\wechat-mini-list-capture
.\scripts\apply-list-b-fast-profile.ps1

This profile is the durable source for the faster desktop strategy: PageDown-style list/comment movement, a 5s detail-page load wait, salary-table-first detail capture, no routine blank/content-area clicks inside detail pages, OCR-recognized reply-expander clicks, re-screenshot/re-OCR after each reply expansion before clicking another expander, dynamic mini-program back (app-back), and a strict guard requiring the full app OCR to show 顶尖人才计划 plus salary-card structure before scanning a list_b round or opening any next detail. If scrolling appears slow again, first reapply this profile and inspect scroll-events.jsonl / per-item comment_events before changing ad hoc runtime config.

Driver Choice

• Prefer android_adb when the user wants no mouse takeover. It controls a real Android device or emulator through ADB screenshots, taps, swipes, and back key events.
• Use desktop only when the user accepts Windows mouse/focus takeover or Android is unavailable.
• If doctor-android.ps1 reports adb is not installed or not on PATH, stop and ask the user to install Android SDK Platform Tools or provide -AdbPath.

Operating Loop

1. Run the relevant doctor command. Prefer doctor-android.ps1 -Screenshot for no-mouse mode.
2. Inspect the raw current screenshot, usually runtime\wechat-mini-list-capture\calibration-<list-id>-screen.png.
3. Create a Codex annotation JSON with app_region, list_region, and full-card items[] boxes.
4. Apply the annotation with apply-annotation.ps1; this renders a review image and writes config.local.json.
5. Inspect the rendered review image. If boxes are good, run a small sample first, usually -MaxItems 3.
6. If boxes are bad, revise the annotation JSON and reapply it. Do not ask the user to manually mark row slots.
7. After sample capture, export a review CSV and ask the user to confirm field quality before full capture.
8. Write back durable status in D:\todo when the workflow meaningfully progresses.

Codex Annotation Format

Generate JSON like this under the private runtime:

{
  "list_id": "list_a",
  "source_image": "D:\\agent-workflow-lab\\runtime\\wechat-mini-list-capture\\calibration-list_a-screen.png",
  "annotator": "codex",
  "app_region": [765, 96, 1392, 1274],
  "list_region": [765, 690, 1392, 1274],
  "items": [
    {
      "id": "list_a_visible_1",
      "label": "visible card title",
      "box": [790, 690, 1368, 962],
      "confidence": 0.86,
      "note": "full visible card; excludes announcement"
    }
  ]
}

Apply it:

.\scripts\apply-annotation.ps1 -Annotation D:\agent-workflow-lab\runtime\wechat-mini-list-capture\codex-annotation-list_a.json

The harness stores Codex item boxes as candidate_mode=row_slots, but the user does not manually mark them. The harness still OCRs each slot during runtime and uses OCR text for deduplication.

Annotation Rules

• Make app_region cover the mini-program window only.
• Make list_region start below tabs and announcement boxes unless the announcement is an intended item.
• Make each items[].box cover one whole clickable card.
• Exclude announcement text, close buttons, floating share/add buttons, terminal text, and VS Code UI.
• Prefer 1-3 visible full cards over many uncertain partial boxes.
• If a card is partly covered by a floating button, either avoid the covered edge or mark confidence lower.

User Coordination

Ask the user only for actions Codex cannot do reliably:

• open the mini program;
• keep the mini-program window visible and stable;
• switch from list A to list B;
• confirm a rendered Codex annotation review image, sample review CSV, or ambiguous screenshot;
• resolve login, QR, CAPTCHA, or platform prompts.

Do not ask the user to manually mark every row unless Codex annotation repeatedly fails and the user explicitly accepts manual fallback.