agent-export

Agent Export — Migration Bundle Creator

Create a structured migration bundle from your current agent data and upload it to the Starchild migration relay. The receiving Starchild agent uses the agent-import skill to load it.

Migration Bundle Format

The bundle is a tar.gz archive with this structure:

migration/
  manifest.json           # required — metadata
  memory/
    agent.json            # agent's own notes & knowledge
    user.json             # what the agent knows about the user
  identity/
    profile.json          # agent name, personality
    soul.md               # behavioral guidelines (free-form markdown)
  user/
    settings.json         # user preferences (name, timezone, language)
  tasks/
    tasks.json            # scheduled/recurring tasks
  env/
    keys.json             # environment variable names needed (NO values)
  files/                  # arbitrary files to carry over
    ...

All files are optional except manifest.json.

---

File Specifications

manifest.json (required)

{
  "version": "1.0",
  "source": "openclaw",
  "created_at": "2025-07-13T10:00:00Z",
  "description": "Migration from OpenClaw agent"
}

• source: identifier of the originating agent/platform (free text)
• version: always "1.0" for now

memory/agent.json

Agent's accumulated knowledge — things the agent learned about the environment, tool quirks, API notes, workflows, conventions.

{
  "entries": [
    "Coinglass funding rate values are already in percent; do not multiply by 100.",
    "User's Hyperliquid account uses cross-margin by default.",
    "For Fly.io deploys, extract FLY_TOKEN via sed from .env in project dir."
  ]
}

Each entry: 1-3 sentences, one concern per entry. Think "what would I need to remember next session?"

memory/user.json

What the agent knows about the user — their role, preferences, communication style, interests.

{
  "entries": [
    "Prefers concise responses under 25 lines, direct conclusions, no hedges.",
    "Technical background in full-stack dev and crypto trading.",
    "Located in Argentina, primary language is Chinese."
  ]
}

identity/profile.json

{
  "name": "MyAgent",
  "vibe": "professional, concise, opinionated",
  "emoji": "🤖",
  "creature": "robot"
}

All fields optional. vibe is a short personality description.

identity/soul.md

Free-form markdown describing how the agent should behave. Keep it under 50 lines. Example:

# Behavior
- Be concise, skip filler phrases
- Have opinions, back them with data
- For trading: present analysis, not financial advice

user/settings.json

{
  "name": "Alice",
  "what_to_call": "Boss",
  "timezone": "Asia/Shanghai",
  "language": "zh-CN"
}

• timezone: IANA format (e.g., America/New_York, Asia/Tokyo)
• language: BCP-47 code (en, zh-CN, ja, etc.)

tasks/tasks.json

{
  "tasks": [
    {
      "title": "BTC Price Alert",
      "schedule": "every 30 minutes",
      "description": "Check BTC price, alert if > $100k or < $80k",
      "channels": ["web", "tg"]
    },
    {
      "title": "Daily Market Summary",
      "schedule": "0 1 * * *",
      "description": "Summarize crypto market at 9am Shanghai time"
    }
  ]
}

• schedule: cron expression (UTC), interval (every 5 minutes), or delay (in 2 hours)
• channels: optional, subset of ["web", "tg", "wechat"]

env/keys.json

List environment variable names the agent needs (values are entered separately for security).

{
  "keys": [
    {"key": "OPENAI_API_KEY", "label": "OpenAI API Key", "required": true},
    {"key": "TELEGRAM_BOT_TOKEN", "label": "Telegram Bot Token", "required": false}
  ]
}

files/

Put any arbitrary files here. They'll be copied to the Starchild workspace as-is, preserving subdirectory structure. Use for custom scripts, configs, dashboards, etc.

---

Export Workflow

Step 1 — Gather Data

Collect information from your current agent. Map your data to the formats above:

• Chat history / memory → extract key facts into memory/agent.json and memory/user.json
• System prompt / persona → transform into identity/profile.json + identity/soul.md
• User preferences → user/settings.json
• Recurring automations → tasks/tasks.json
• API keys needed → env/keys.json (names only!)
• Custom files → copy into files/

Step 2 — Build the Bundle

# Create directory structure
mkdir -p migration/{memory,identity,user,tasks,env,files}

# Write your JSON files (examples — replace with real data)
cat > migration/manifest.json << 'EOF'
{"version": "1.0", "source": "my-agent", "created_at": "2025-07-13T10:00:00Z"}
EOF

# ... write other files ...

# Pack into tar.gz
cd migration && tar czf ../migration-bundle.tar.gz . && cd ..

Step 3 — Upload to Relay

The relay only accepts bundles coming from an external agent platform. Run the guard first and abort the upload if it exits non-zero. Resolve the script against the directory of this SKILL.md (the absolute path shown for this skill when it was loaded) — never against the current working directory, which is usually the workspace/repo root:

SKILL_DIR="<absolute path of the directory containing this SKILL.md>"
python3 "$SKILL_DIR/scripts/export.py" --check-relay || exit 1

If it refuses, the bundle you built in Step 2 is still valid as a local backup — only the relay handoff is unavailable.

RESPONSE=$(curl -s -X POST https://sc-agent-migration.fly.dev/paste \
  -H "Content-Type: application/octet-stream" \
  --data-binary @migration-bundle.tar.gz)

CODE=$(echo "$RESPONSE" | grep -o '"code":"[^"]*"' | cut -d'"' -f4)
TOKEN=$(echo "$RESPONSE" | grep -o '"download_token":"[^"]*"' | cut -d'"' -f4)

echo "Migration code:   $CODE"
echo "Download token:   $TOKEN"

Give both values to the user — they need the code AND the token to import in Starchild.

⚠️ The code expires in 1 hour and is single-use (deleted after first successful download). ⚠️ The token is the only authorization for download — do not lose it.

---

Tips for Other Agents

• Don't overthink it. Even just manifest.json + memory/user.json with 5 entries is useful.
• Summarize, don't dump. Convert raw chat history into concise memory entries.
• Timezone matters. Get the user's timezone right — it affects all scheduled tasks.
• No secrets in the bundle. API keys go in env/keys.json as names only.
• Test your tar.gz — tar tzf migration-bundle.tar.gz should show paths starting with ./.