SB-Box

curl -fsSL https://www.genspark.ai/sb-box/opencode/install.sh | sh

wrote  .opencode/skills/sb-box/SKILL.md

One file, nothing else. Needs gsk on PATH (npm i -g @genspark/cli && gsk login); sandboxes come from the public base image.

How to use me

1. One sandbox per task, named by --session. Pick a short name for the task (the project folder name is a good one) and pass the same --session value to every command. The first run-code, exec or write creates the sandbox from the public base image (Python 3 with pandas, numpy, matplotlib; Node; git); later commands reuse it, also after the platform froze it. You never handle a sandbox id.

   gsk sb-box run-code --session "$TASK" --code 'print("hello")'
   

2. Files first. The sandbox is a separate machine: nothing from the working directory is visible there. Copy in what the code needs, then refer to it under /home/user/.

   gsk sb-box write --session "$TASK" --path /home/user/data.csv --content "$(cat data.csv)"
   

3. Run, then read the answer from data. For step-by-step analysis prefer the stateful interpreter: variables and DataFrames persist between calls like a notebook, and plots come back as chart data. The first call starts it; later cells reuse the running kernel.

   gsk sb-box run-code --session "$TASK" --code 'import pandas as pd; df = pd.read_csv("/home/user/data.csv"); df.describe()'
   

   gsk sb-box run-code --session "$TASK" --code 'df.groupby("region").margin.mean()'      # df is still there
   

4. Bring results back with gsk sb-box read --session "$TASK" --path /home/user/out.json

5. Kill it when the task is done, unless the user wants to keep it:

   gsk sb-box kill --session "$TASK"
   

gsk sb-box exec --session dash --timeout 300 -c 'git clone https://github.com/acme/dashboard && cd dashboard && npm ci'

{
  "stdout": "added 412 packages in 18s",
  "exitCode": 0,
  "session": { "key": "dash", "sandbox_id": "sbx-a41d02…", "created": true }
}

The session has no deadline; only the idle timer applies. The repo lives on the sandbox from here on — nothing is cloned to your laptop, and no key is copied anywhere: the agent runs locally with the sb-box skill from ACT 01 and reaches this machine by name.

opencode run "In sandbox session dash, pytest fails in ~/dashboard. Find the bug, fix it, run pytest again and tell me what it was. Use gsk sb-box exec --session dash for every command — do not run anything on this machine."

⚙ bash    gsk sb-box exec --session dash -c 'cd dashboard && pytest -q'
          → 1 failed, 23 passed   # FAILED tests/test_todo.py::test_complete — assert done is True
⚙ bash    gsk sb-box read --session dash -p /home/user/dashboard/todo.py
⚙ bash    gsk sb-box write --session dash -p /home/user/dashboard/todo.py -d "$(…fixed file…)"
⚙ bash    gsk sb-box exec --session dash -c 'cd dashboard && pytest -q'
          → 24 passed
The bug was in complete() at todo.py:6 — it set done to the string "yes" instead of True. All tests pass.

Fixed. Leave the dev server running on its machine and go home:

gsk sb-box exec --session dash -c 'cd dashboard && (npm run dev -- --port 3000 >/tmp/dev.log 2>&1 &) && sleep 2'

Preview, alive while the sandbox is, at its port URL (the record's portUrl with port 3000): https://3000-sbx-a41d02.sandbox.westus2-prod.sspark.ai. Then you close the laptop.

Her machine is the same kind of sandbox with two extras: an egress allow-list and a port the ticket webhook calls. Network rules are REST-only today.

API=https://www.genspark.ai/api/sb-box/v1; H='X-Api-Key: gsk-…'
curl -s -X POST "$API/sandboxes" -H "$H" -H 'Content-Type: application/json' -d '{
  "templateId": "tpl-4755…",
  "timeout": 0,
  "network": {
    "allowInternetAccess": false,
    "allowOut": ["10.20.0.0/16"]
  }
}'

{
  "id": "sbx-91be47",
  "state": "running"
}

gsk sb-box write -s sbx-91be47 -p /home/user/anna.py -d "$(cat anna.py)"

{
  "written": 2140,
  "path": "/home/user/anna.py"
}

gsk sb-box exec -s sbx-91be47 -c 'cd /home/user && (python3 anna.py --port 8080 &) && sleep 2'

Every ticket webhook goes to the sandbox's port URL — the record's portUrl with port 8080, https://8080-sbx-91be47.sandbox.westus2-prod.sspark.ai/tickets; a frozen Anna is resumed by the request itself.

gsk sb-box exec --session task-0417 --timeout 300 -c 'git clone https://github.com/acme/calc && cd calc && pip install -q -r requirements.txt && pytest -q; true'
gsk sb-box exec --session task-0417 -c 'cd calc && (./services/start.sh >/tmp/services.log 2>&1 &) && sleep 2'
gsk sb-box freeze --session task-0417

{
  "id": "sbx-0417a1…",
  "state": "frozen",
  "versions": [
    { "version": 1 }
  ]
}

The frozen state is the environment: dependencies installed, services up, tests already run once so every cache is warm (the freeze archived the bare template as v1; the environment itself is the live frozen snapshot). It stays frozen for as long as the task is in the training set.

ATTEMPTS=$(for i in $(seq 8); do gsk sb-box clone -s sbx-0417a1… | jq -r .data.id & done; wait)

Eight identical machines. The policy on your GPUs now acts in each one; every action is an exec, its output the next observation.

gsk sb-box exec -s $A -c 'cat calc/calc.py'                                   # the policy reads
gsk sb-box exec -s $A -c 'sed -i "s/return a - b/return a + b/" calc/calc.py'   # the policy edits

When the attempt ends, the grader runs in the same machine. Tests pass → 1.0, anything else → 0.0.

while read -r A; do gsk sb-box exec -s $A -c 'cd calc && pytest -q' | jq -c '{reward: (if .data.exitCode == 0 then 1.0 else 0.0 end)}' & done <<<"$ATTEMPTS"; wait

{"reward": 1.0}
{"reward": 0.0}
{"reward": 1.0}
{"reward": 1.0}
{"reward": 0.0}
{"reward": 1.0}
{"reward": 0.0}
{"reward": 1.0}

Rewards and trajectories go to the trainer; the eight machines go away. The environment is still frozen, untouched, for the next step.

while read -r A; do gsk sb-box kill -s $A & done <<<"$ATTEMPTS"; wait