Brain

sandbox

2 min read

🧪 Sandbox — Company Playground

Repo: /root/repos/sandbox → patbhakta/sandbox (private) Paperclip: Project “Sandbox” Agent: Sandbox Dev (BHAA issues) Model: openrouter/deepseek/deepseek-v4-flash-0731

Purpose

The sandbox is the company’s experimental engine. Not a product. Not a project. It’s where we test everything before it touches production.

What It’s For

  • Agent Testing — New agents prove they can follow the pipeline here first
  • New Hires — Onboarding tasks that verify a worker understands the flow
  • R&D — Experimental approaches, proofs of concept, dead ends worth documenting
  • Integrations — New tools (Kestra, MCP servers, APIs) tested before wiring to real projects
  • Benchmarks — Model comparisons, cost analysis, capability evaluations

Pipeline Verification Results

Test 1: BHAA-1339 — Hello World ✅

  • Agent created src/hello.py + tests/test_hello.py
  • Tests passed, committed to git, posted Paperclip comment, marked done
  • Cost: $0.01, Duration: ~2 min
  • Minor gap: didn’t push to remote (committed locally)

Test 2: BHAA-1340 — Calculator Module ✅

  • Agent created src/calculator.py (4 functions) + tests/test_calculator.py (5 tests)
  • All 6 tests pass (including divide-by-zero error case)
  • Committed to git, posted Paperclip comment
  • Cost: $0.003, Duration: ~1 min

Issues Found & Fixed

  1. Stale OpenCode binary — Paperclip was running v1.15.10 via old path. Fixed: symlinked to v1.18.11.
  2. Workspace resolution — Paperclip used internal agent_home instead of git repo. Fixed: symlinked agent_home → /root/repos/sandbox.
  3. Model cache — After binary update, needed Paperclip restart to clear cached model list.
  4. Push gap — Agents commit but don’t always push. Needs reinforcement in AGENTS.md.

How To Use

Testing a New Agent

  1. Create agent in Paperclip with adapterType: opencode_local
  2. Set cwd: /root/repos/sandbox and model: openrouter/deepseek/deepseek-v4-flash-0731
  3. Symlink agent_home: ln -s /root/repos/sandbox /root/.paperclip/instances/default/workspaces/{agent_id}
  4. Create an issue with clear instructions
  5. Trigger wakeup: POST /api/agents/{id}/wakeup
  6. Monitor: check heartbeat_runs table
  7. Verify: git log, test results, Paperclip comments

Running an Experiment

  1. Create folder: experiments/YYYY-MM-DD-name/
  2. Add README.md explaining what’s being tested
  3. Run the experiment
  4. Commit results (even failures)
  5. Document findings in vault

Graph View