medium

medium — 自分の Cookie を使って Medium で読んだり公開したりできます

Medium が 2023 年に公開書き込み API を廃止したため、サイトと同じ内部 Web API を使って、ユーザーの実際の Medium アカウントを操作します。ACE 拡張機能で取得したログイン Cookie を使って認証します。ブラウザーやサードパーティ製の依存関係はなく、urllib のみを使用します。

コネクターは Cookie jar を環境変数として渡します。

• MEDIUM_COOKIES — Cookie（sid、uid、xsrf）の JSON 配列です。秘密情報です。絶対にエコーしたり、出力したりしないでください。 CLI は書き込み時に xsrf Cookie を x-xsrf-token ヘッダーとして自動的にエコーします。取得した jar に xsrf Cookie がない場合（よくあります）、CLI は書き込み前に自動で作成するため、「Missing xsrf token」による公開エラーは発生しなくなります。

CLI

このスキルには、標準ライブラリのみで動作する単独の scripts/medium.py が含まれています。

# $SKILL_DIR can point at another skill loaded this turn — anchor on our own
# script, and re-run this at the top of every Bash block (fresh shell each time).
MED="$SKILL_DIR/scripts/medium.py"; [ -f "$MED" ] || MED=$(find /tmp -maxdepth 8 -path '*/skills/*/scripts/medium.py' 2>/dev/null | head -1)
[ -f "$MED" ] || { echo "medium script not found (SKILL_DIR=$SKILL_DIR)" >&2; exit 1; }
python3 "$MED" whoami                     # who is logged in
python3 "$MED" articles --limit 20        # my posts + clap/response stats
python3 "$MED" article <post-id>          # one post's details

まず接続を確認してください

MED="$SKILL_DIR/scripts/medium.py"; [ -f "$MED" ] || MED=$(find /tmp -maxdepth 8 -path '*/skills/*/scripts/medium.py' 2>/dev/null | head -1)
python3 "$MED" whoami
# → {"user_id": "...", "name": "...", "username": "..."}

認証エラーが発生した場合、Cookie の有効期限が切れています。ユーザーに https://studio.acedata.cloud/console/connectors で再接続してもらってください。再試行を繰り返してはいけません。

公開 — 要確認（末尾に --confirm がない場合はドライラン）

publish はユーザーの実際のアカウントに書き込みます。コンテンツは Markdown 形式で、Medium の段落ブロックに変換されます（#／##→見出し、###→小見出し、>→引用、 ```→code, -/*→bullet list, 1.→numbered list, tables→aligned code block). Inline bold, italic, code, and [links](url) render as real Medium formatting (clickable links included). Bare https://… URLs in text are auto-linked too (URLs inside code spans are left untouched). Without a trailing --confirm it dry-runs. --confirm is honored only as the last argument. Always show the dry-run, get an explicit "yes", then re-run.

MED="$SKILL_DIR/scripts/medium.py"; [ -f "$MED" ] || MED=$(find /tmp -maxdepth 8 -path '*/skills/*/scripts/medium.py' 2>/dev/null | head -1)
python3 "$MED" publish --title "Title" --content-file a.md                       # dry-run
python3 "$MED" publish --title "Title" --content-file a.md --draft-only --confirm   # private draft
python3 "$MED" publish --title "Title" --content-file a.md --confirm                # PUBLIC story

Publishing is Medium's multi-step editor flow (new-story → write deltas → publish). --draft-only stops at the draft (visible only at the user's /me/stories/drafts). Default to --draft-only unless the user asked to go live.

Images

publish automatically uploads each external markdown image (![](url)) to Medium and inserts it as a real image block — Medium has no markdown-image syntax, so this is the only way images render. Pass --no-rehost-images to degrade images to link-only paragraphs. An image that fails to upload falls back to a link paragraph (never blocks the post).

Gotchas

• This is the user's real Medium account. Confirm before any publish.
• Markdown→Medium conversion covers headings, quotes, fenced code, bullet/ numbered lists, images, and inline markups (bold/italic/code/links). Medium's editor has no table element: a narrow markdown table renders as an aligned monospace code block, while a wide one (long cells / URLs, which would wrap into an unreadable grid) is rendered as per-row records — the first column bolded as a lead-in, the remaining columns as header: value bullets.
• Medium sits behind Cloudflare; an occasional 403/429 is transient — the CLI auto-retries once after a short pause. A persistent 403 means the cookie is genuinely expired (reconnect).
• Never print MEDIUM_COOKIES — it is full account access.
• ToS: acts only on the user's own account with their own captured cookie.

Record the output

After you successfully publish and obtain the live result URL, call the built-in publish_artifact tool ONCE so the user can track this deliverable in My Outputs:

publish_artifact(kind="article", channel="medium", title="<title>", url="<the REAL returned URL>", status="delivered")

Use the real returned URL — never fabricate one. Call it once per published item, only after delivery is confirmed; skip it (or use status="failed") if publishing failed. See _shared/artifacts.md.