القائمة
使用 RapidOCR 对本地图片做 OCR。当用户要从 PNG/JPG/JPEG/WEBP/BMP/TIFF 图片中提取文字、返回结构化 JSON、标记低置信度或疑似金额数字，或把表格截图粗略还原成 Markdown 表格时使用；只处理本地文件，不下载远程 URL。
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | rapidocr |
| description | 使用 RapidOCR 对本地图片做 OCR。当用户要从 PNG/JPG/JPEG/WEBP/BMP/TIFF 图片中提取文字、返回结构化 JSON、标记低置信度或疑似金额数字，或把表格截图粗略还原成 Markdown 表格时使用；只处理本地文件，不下载远程 URL。 |
用本 skill 从本地图片中提取文字。运行入口是 Python 脚本，依赖通过 uv tool 安装和隔离。
png、jpg、jpeg、webp、bmp、tif、tiff。--table 做启发式表格还原。--table 只是基于 OCR 框的启发式重排。Python 工具依赖统一使用 uv tool：
uv tool install --force rapidocr --with onnxruntime
不要改成 python -m pip install ...。
python3 "$SKILL_DIR/scripts/run_rapidocr.py" "/absolute/path/to/image.png"
结构化输出：
python3 "$SKILL_DIR/scripts/run_rapidocr.py" "/absolute/path/to/image.png" --json
表格截图：
python3 "$SKILL_DIR/scripts/run_rapidocr.py" "/absolute/path/to/table.png" --json --table
脚本会先尝试当前 Python。如果当前 Python 没有 rapidocr，会自动 re-exec 到 uv tool 环境：
$(uv tool dir)/rapidocr/bin/python$(uv tool dir)/rapidocr/Scripts/python.exe普通模式输出纯文本，一行一个识别文本块。
JSON 模式包含：
text: 合并后的纯文本lines: 识别行数组boxes: OCR 框scores: 置信度items: 每行文本、框、置信度、中心点review.low_confidence: 低置信度行review.amount_like: 疑似金额/数量/日期等数字行review.needs_review: 是否建议人工复核table.markdown: 使用 --table 时的启发式 Markdown 表格review.needs_review 为 true，回复里要提醒用户复核低置信度或金额数字。--table，明确说明这是启发式表格，还需要人工核对结构。uv tool install --force rapidocr --with onnxruntime
python3 "$SKILL_DIR/scripts/run_rapidocr.py" "/absolute/path/to/image.png" --json
能输出可解析 JSON，且包含 review 字段，就算 wiring 正常。
Autonomously audit native local Codex session JSONL for workflow quality, latency, cost shape, tool use, recurring failures, and eval candidates. Use when the user wants scheduled harness observability or evidence-driven harness improvement; do not use for ordinary task execution, application telemetry, or full transcript export.
Verify an existing build artifact itself: metadata, complete download, checksum or size, archive integrity, unpacked contents, and source or lock provenance. Use when artifact contents or completeness must be proven; do not use to trigger builds, releases, deploys, or health checks.
Diagnose the live remote-access layer: SSH resolution, reachability, Tailscale, login-shell PATH, service state, package-manager locks, and remote CLI auth. Use when a named remote host or service is unreachable or unhealthy; do not use for deploy planning or ordinary local repo work.
Establish the current repo, instruction, branch, dirty-worktree, and linked live-object facts before non-trivial work. Use once at the start of repo work when current state or scope is unclear; do not stack it onto every later workflow phase.
Carry already-scoped code changes through verification, commit, push, PR/MR, authorized merge, and authorized cleanup. Use when the requested outcome includes delivery actions; do not use for review-only work, initial repo discovery, CI diagnosis, or deploy execution.
GitLab CLI workflows with `glab` for issues, merge requests, CI/CD, and GitLab API operations. Use when the user needs GitLab command-line automation, issue or MR triage, pipeline checks, API queries, threaded discussion replies, release checks, or work item conversion and hierarchy updates. Prefer `glab api` for reliable scripted flows, while keeping direct CLI commands and local helper scripts available for common MR and pipeline tasks.