lark1115/SKILL.md

name

session-from-r2

description

過去のセッションログ (R2) を読み込んで、セッションコンテキストを復元する。 Cloudflare R2 に保存された JSONL 形式のセッションログを取得・要約する。トリガー: /session-from-r2、「前回の引き継ぎ読んで」「ハンドオーバー確認して」「前回どこまでやった？」「前のセッションの続きから」引数でセッション ID を直接指定可能。省略時は R2 から自動検索。

allowed-tools

Read

Bash

Session from R2

過去のセッションログ (R2) を読み込んで、セッションコンテキストを復元する。

R2 アクセスには sops-secrets 経由の S3 互換クレデンシャルが必要。 session-to-r2 と同じ認証方式 (curl + S3v4 署名) を使用する。 wrangler は不要。

環境変数

変数	必須	説明
`SESSION_R2_BUCKET`	Yes	R2 バケット名 (default: claude-sessions)
`R2_S3_ACCESS_KEY_ID`	Yes	R2 S3 Access Key ID
`R2_S3_SECRET_ACCESS_KEY`	Yes	R2 S3 Secret Access Key
`R2_S3_ENDPOINT`	Yes	R2 S3 endpoint URL

sops-secrets でロード済みであれば自動で利用可能。

Procedure

Minimize commands. Each R2 API call has latency — never fetch the same object twice.

Step 1: Load credentials

# sops-secrets から R2 クレデンシャルをロード (未設定の場合)
if [ -z "${R2_S3_ACCESS_KEY_ID:-}" ]; then
  SOPS_FILE=""
  for _candidate in \
    "$(git rev-parse --show-toplevel 2>/dev/null)/secrets/agents.enc.env" \
    "$HOME/.agents/skills/secrets/agents.enc.env" \
  ; do
    [ -f "$_candidate" ] && SOPS_FILE="$_candidate" && break
  done
  unset _candidate
  if [ -n "$SOPS_FILE" ] && command -v sops &>/dev/null; then
    export SOPS_AGE_KEY_FILE="${SOPS_AGE_KEY_FILE:-$HOME/.config/sops/age/keys.txt}"
    while IFS='=' read -r k v; do [ -n "$k" ] && export "$k=$v"; done \
      < <(sops decrypt --input-type dotenv --output-type dotenv "$SOPS_FILE" 2>/dev/null)
  fi
fi
export SESSION_R2_BUCKET="${SESSION_R2_BUCKET:-claude-sessions}"

Step 2: Find the target session

If a session ID is given as argument, skip to Step 3.

Otherwise, discover available sessions:

REPO_PREFIX=$(git remote get-url origin 2>/dev/null | sed -E 's#.*[:/]([^/]+)/([^/.]+)(\.git)?$#\1/\2#')

# R2 S3 API で session 一覧を取得
~/.agents/scripts/r2-s3v4.sh list "$SESSION_R2_BUCKET" "$REPO_PREFIX/" | \
  python3 -c "
import sys, json
from collections import defaultdict
data = json.loads(sys.stdin.read())
sessions = defaultdict(list)
for obj in data.get('Contents', []):
    key = obj['Key']
    parts = key.split('/')
    if len(parts) >= 3:
        session_id = parts[2]
        sessions[session_id].append(obj)
for sid, objs in sorted(sessions.items(),
    key=lambda x: max(o['Key'] for o in x[1]), reverse=True)[:10]:
    latest = max(o['Key'] for o in objs)
    total_size = sum(int(o.get('Size', 0)) for o in objs)
    print(f'{sid}  chunks={len(objs)}  size={total_size//1024}KB  latest={latest.split(\"/\")[-1]}')
"

Step 3: Read the latest chunk

SESSION_ID="<selected-session-id>"

# 最新チャンクのキーを取得
LATEST_KEY=$(~/.agents/scripts/r2-s3v4.sh list "$SESSION_R2_BUCKET" "$REPO_PREFIX/$SESSION_ID/" | \
  python3 -c "
import sys, json
data = json.loads(sys.stdin.read())
keys = sorted([o['Key'] for o in data.get('Contents', [])])
if keys: print(keys[-1])
")

# ダウンロード
~/.agents/scripts/r2-s3v4.sh get "$SESSION_R2_BUCKET" "$LATEST_KEY" /tmp/session-chunk.jsonl

# パース・要約
python3 -c '
import sys, json
first_ts = last_ts = None
user_msgs, asst_msgs = [], []
for line in open("/tmp/session-chunk.jsonl"):
    try: obj = json.loads(line)
    except: continue
    ts = obj.get("timestamp", "")
    if not first_ts and ts: first_ts = ts
    if ts: last_ts = ts
    role, msg = obj.get("type", ""), obj.get("message", {})
    if not isinstance(msg, dict): continue
    content = msg.get("content", "")
    if isinstance(content, list):
        content = " ".join(c.get("text", "") for c in content if isinstance(c, dict) and c.get("type") == "text")
    if not content: continue
    if role == "user" and len(content) < 500:
        user_msgs.append((ts[11:19] if len(ts) > 19 else ts, content[:200]))
    elif role == "assistant" and len(content) > 20:
        asst_msgs.append((ts[11:19] if len(ts) > 19 else ts, content[:300]))
print(f"Range: {first_ts} -> {last_ts}")
print(f"Events: {len(user_msgs)} user, {len(asst_msgs)} assistant")
print("--- User messages ---")
for ts, m in user_msgs: print(f"[{ts}] {m}")
print("--- Key assistant messages (last 15) ---")
for ts, m in asst_msgs[-15:]: print(f"[{ts}] {m}")
'

Step 4: Summarize + offer to go deeper

From the latest chunk output, report:

Chunk info and total chunk count
What was done, key decisions, issues, outcomes
Any incomplete tasks or next steps

Then ask: if the user needs more context from earlier in the session, fetch additional chunks (work backwards from the second-to-last chunk). Do not pre-fetch all chunks.

Key Rules

Never fetch the same R2 object more than once.
Always read the latest chunk by default.
Keep the summary concise — the user can ask for details.
Focus on assistant text messages (user messages often contain verbose tool results).
Report the session date and time range.
If the session references files, verify they still exist before suggesting actions.
Clean up temp files: rm -f /tmp/session-chunk.jsonl after reading.