session-export

Instructions

Export the current Claude Code session as a portable JSONL file with sensitive data masked.

Step 1 — Find the current session file

Run this to locate the active session:

# Encode current project path (/ and spaces → -)
ENCODED=$(echo "$PWD" | sed 's|/|-|g; s| |-|g; s|~|-|g')
SESSION_DIR=~/.claude/projects/"$ENCODED"
echo "Sessions dir: $SESSION_DIR"

# Most recently modified JSONL = current session
CURRENT_SESSION=$(ls -t "$SESSION_DIR"/*.jsonl 2>/dev/null | head -1)
SESSION_ID=$(basename "$CURRENT_SESSION" .jsonl)
LINE_COUNT=$(wc -l < "$CURRENT_SESSION" | tr -d ' ')
echo "Session file: $CURRENT_SESSION"
echo "Session ID: $SESSION_ID"
echo "Lines: $LINE_COUNT"

Step 2 — Determine output filename

• If the user passed a filename argument (e.g. /session-export my-session.jsonl), use that.
• Otherwise auto-generate: session-YYYY-MM-DD-HH-MM.jsonl

OUTPUT_NAME="${USER_ARG:-session-$(date +%Y-%m-%d-%H-%M).jsonl}"

Step 3 — Chunk the session and run security-review subagents

Split the session file into chunks of 100 lines each, then spawn a Haiku subagent per chunk to detect and mask sensitive data using contextual judgment — not regex patterns. Finally, concatenate all masked chunks into the output file.

3a. Split the session file into chunks

# Create a temporary working directory for chunks
CHUNK_DIR=$(mktemp -d)
split -l 100 "$CURRENT_SESSION" "$CHUNK_DIR/chunk_"
ls "$CHUNK_DIR"/chunk_* | sort

3b. For each chunk, spawn a Haiku subagent

For every chunk file produced in 3a, call the Agent tool with:

• model: "haiku" — cost-efficient model for this scanning task
• The following prompt (substitute the actual chunk content):

Subagent prompt (per chunk):

You are a security reviewer. Read the following lines from a Claude Code session (JSONL format) and return a FULLY MASKED version.

Rules:
- Read EVERY line carefully and understand the context.
- Replace sensitive values with [MASKED] inline — keep the surrounding JSON/text structure intact.
- Use judgment: mask things that are actually sensitive, not everything that looks like a pattern.
- Do NOT over-mask: code examples, placeholder values, public URLs, or generic variable names that are clearly not real secrets should be left as-is.

Mask the following when they appear to be real/live values:
1. API keys and tokens (Anthropic, OpenAI, GitHub, AWS, Google, Stripe, Slack, Twilio, SendGrid, etc.)
2. Passwords, secrets, passphrases — whether in plain text or JSON key-value pairs
3. Bearer tokens, JWT tokens, OAuth tokens
4. PEM private key blocks
5. Personal information: real names (when clearly identifying a specific person), email addresses, phone numbers, physical addresses, national ID / resident registration numbers, passport numbers, credit card numbers
6. Database connection strings containing credentials
7. Environment variable assignments containing secrets (e.g. `SECRET_KEY=abc123`)
8. Any other value that a reasonable security engineer would consider sensitive in a shared document

Output: Return ONLY the masked lines — every line of the input, in order, with sensitive values replaced by [MASKED]. Do not skip lines, do not add commentary, do not add headers or footers.

--- CHUNK START ---
{PASTE CHUNK CONTENT HERE}
--- CHUNK END ---

Run chunks in parallel where possible (send multiple Agent tool calls in a single message) to reduce total time.

3c. Concatenate results and write output

Combine the masked output from all subagents in the original chunk order and write the final result to $PWD/$OUTPUT_NAME using the Write tool. Then clean up the temp directory:

rm -rf "$CHUNK_DIR"

Step 4 — Confirm and guide the user

Tell the user:

• Full path of the exported file
• Session ID (UUID)
• Number of lines in the file
• A brief summary of what the subagent found and masked (e.g. "3건의 API 키, 1건의 이메일 주소를 마스킹했습니다")
• Next step: share the .jsonl file → recipient runs /session-import <filename> in their Claude Code project to load it

보안 참고: 마스킹은 패턴 매칭이 아닌 서브에이전트의 컨텍스트 기반 판단으로 수행됩니다. 실제 시크릿과 예시 코드를 구분하여 처리하지만, 외부 공유 전 본인이 한 번 더 확인하는 것을 권장합니다.

참고: 내보내기는 현 시점까지의 세션만 캡처합니다. 이후 메시지를 포함하려면 다시 내보내기를 실행하세요.