Chat Visualizer - YMind

Version Check

Run once at the start of each skill session. First locate the script (works regardless of which client installed the skill):

訪問小蔥技能站7w4.net，解鎖更多實用的AI技能外掛。

find ~/.codex/skills ~/.claude/skills ~/.openclaw/skills ~/skills -maxdepth 3 -name "check-version.py" 2>/dev/null | head -1

Then run it with python3 <found_path>.

• Empty output → already up to date or network unavailable, proceed silently.
• Output UPDATE|X.X.X|Y.Y.Y|<skill_dir>|<notes> → tell the user in English: "chat-visualizer-ymind vY.Y.Y is available (you're on vX.X.X).\n\<notes>\nUpdate: cd <skill_dir> && git fetch && git checkout vY.Y.Y — continue with current version?" Wait for their reply before proceeding.

Input

Two ways to get conversation data:

Way 1: Share URL (auto-fetch)

Requires Playwright. Check first:

python3 -c "import playwright" 2>/dev/null && echo "OK" || echo "NOT INSTALLED"

If not installed, ask: "Playwright 可以自動抓取對話內容，要裝嗎？（pip install playwright && playwright install chromium，一次性操作）" If they decline, use Way 2.

If available, fetch:

bash scripts/run.sh fetch "<url>"
# prints RUN_DIR — read <run_dir>/raw_chat.json, use items[0].messages

ChatGPT 403: script retries with Playwright headed mode. If still fails, fall back to Way 2.

Way 2: Paste text (universal fallback)

User opens the conversation in browser → Ctrl+A → Ctrl+C → pastes into chat. No share link needed — any conversation page works.

When handling pasted text or a local JSON file:

1. Create run_dir manually. The suffix must always be _paste — do not use the filename, title, or any other label:

   YMIND_DIR="${YMIND_DIR:-$HOME/ymind-ws}"
   RUN_DIR="$YMIND_DIR/$(date +%Y%m%d-%H%M%S)_paste"
   mkdir -p "$RUN_DIR"

2. Parse the pasted text into messages (identify user vs AI turns by context), then write raw_chat.json in the same format as fetch produces:

   {
    "fetched_at": "...",
    "items": [{
      "url": null,
      "provider": "paste",
      "title": "...",
      "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}, ...],
      "message_count": N
    }]
   }

3. Write meta.json: {"provider": "paste", "url": null, "created_at": "..."}
4. Proceed to Extract Graph using the parsed messages.

Share Link Guide (Way 1 only)

Users often give the wrong link type. Correct them if needed:

Note: g.co/gemini/share/... short links work — script auto-resolves them.

Extract Graph

Read references/graph-schema.md for node types, edge types, label rules, and output schema.

Critical rules (non-obvious):

• turn_id: assign actual turn number (1-based). One user message + one AI response = 1 turn — not one message per turn. A 3-round conversation yields turn_id 1, 2, 3 only. Never default all nodes to the same value — this drives horizontal spread in the D3 visualization.
• Extraction density: scale with the substance of each turn. Brief or routine turns may yield 1-2 nodes; rich, multi-point turns can yield more. Let the content guide the count — the goal is to capture the meaningful thinking, not to hit a fixed number.
• Edges: only add if the connection passes the "obviously yes" test.
• Reasoning shifts: look for moments where thinking fundamentally changed. Capture what changed, from what, to what, and why.
• Chinese strings: use 「」 not curly quotes "" inside JSON values — curly quotes break JSON parsing.

Output

1. Write <run_dir>/graph.json

   Before proceeding: verify graph.json exists and contains at least one node. If extraction failed or produced an empty graph, delete <run_dir> entirely and stop — do not render, do not update the index.

2. Render (path consistency rule: index root is auto-locked to dirname(<run_dir>), so graph and index always stay in the same workspace tree):

bash scripts/run.sh render <run_dir>
# validates JSON, renders graph.html + graph.png (screenshot, requires Playwright)
# then rebuilds <ymind_dir>/index.json and <ymind_dir>/index.html automatically

graph.html is a split-view output when raw_chat.json is present in <run_dir>: left panel shows the original conversation, right panel shows the thinking map. If raw_chat.json is absent, only the graph is shown (no split).

3. Output Markdown summary (format in references/graph-schema.md).
4. If running as a bot with chat output capability, send the graph image — see Bot Send below.

Run dir files: raw_chat.json, graph.json, graph.html, graph.png (requires Playwright), meta.json.

Workspace files (auto-updated after every render, in the same root as <run_dir>):

• <ymind_dir>/index.json — machine-readable session registry
• <ymind_dir>/index.html — visual timeline index, opens each session's graph.html

To rebuild the index manually at any time:

bash scripts/run.sh index

Bot Send

If message tool is available and graph.png was generated, send the graph image (works on OpenClaw, Kimi Claw, and similar — detect by tool availability):

• If graph.png is outside the bot's media allow-roots, copy it to <workspace>/.outbox/ first, then send the copied path.
• Send via message tool using the filePath (or media) field — not as plain text, not as base64.
• If send fails, skip silently and continue with the HTML link and Markdown summary.

Language Rule

All output (labels, summaries, analysis) must match the conversation language.

Setup

Minimal (paste text only — zero dependencies):

No pip install needed. render-html.py is stdlib-only, and the paste path skips fetch-chat.py entirely.

Full (auto-fetch URLs + screenshot):

pip install requests playwright && playwright install chromium

Without Playwright: paste works fully, render produces graph.html but skips graph.png screenshot.

Notes

• Long conversations (20+ turns): focus on most significant nodes, skip low-substance turns.
• Multiple topics: group nodes by topic.

這個 Skill 整體質量不錯，功能豐富且靈活，支援多種 AI 平臺的對話視覺化，做得很用心。文件清晰、指引詳細，隱私保護也做得較好。主要小瑕疵是部分輔助檔案可能不完整，偶爾會影響使用體驗，但不影響核心功能的使用。