--- name: sandboxmaxxing description: Race several coding agents (Claude Code, Codex, OpenCode) on the same task, each in its own disposable Tenki cloud sandbox on the user's own account, then judge the builds and keep the best. Use when the user wants best-of-N attempts at a build, wants to compare agents or models on a real task, or wants work done in isolated throwaway machines rather than on their laptop or in a git worktree. --- # Sandboxmaxxing: race coding agents in throwaway sandboxes One prompt goes to several agents. Each builds independently in its own Tenki microVM — on the user's Tenki account, with the user's model keys — and serves its result on a live preview URL. **You** judge the builds, keep the winner, and the rest turn to sand. The driver is `scripts/arena.sh` next to this file. Every command below assumes `S=/scripts/arena.sh`. ## Install (if you're reading this from a URL) If this file isn't already in a skills directory, install the skill — three files: ```bash D=~/.claude/skills/sandboxmaxxing # Codex or another agent: any directory you like mkdir -p "$D/scripts" curl -fsSL https://sandboxmaxxing.com/skill.md -o "$D/SKILL.md" curl -fsSL https://sandboxmaxxing.com/skill/arena.sh -o "$D/scripts/arena.sh" curl -fsSL https://sandboxmaxxing.com/skill/inject-provider.sh -o "$D/scripts/inject-provider.sh" chmod +x "$D"/scripts/*.sh S="$D/scripts/arena.sh" ``` ## Before you start 1. **Tenki CLI, logged in.** `tenki sandbox list` must succeed. If `tenki` is missing: `curl -fsSL https://tenki.cloud/install.sh | bash`, then ask the user to run `tenki login`. Never ask for their Tenki key in chat. 2. **Model access, per agent:** - `opencode` — nothing; it runs its own free models. - `claude` — `ANTHROPIC_API_KEY` in the environment, or a gateway. - `codex` — `OPENAI_API_KEY` in the environment, or a gateway. - **Gateway** — one OpenAI/Anthropic-compatible endpoint serving every agent, e.g. `--base https://api.aisa.one --key-env AISA_API_KEY`. Pass the env var's **name**, never the key. The script refuses to launch an agent that cannot authenticate. 3. **Say what it costs before a big arena.** Each contestant is a sandbox (default 2 cores, 4 GB) for up to `--minutes`, plus its model usage. Tell the user N × minutes first. ## Run ```bash $S run --prompt "" --agents claude,codex,opencode ``` - `claude,codex,opencode` compares agents. `claude,claude,claude` is best-of-3 from one agent. Only include agents the user can authenticate. - `--minutes 45` (default): every sandbox self-destructs after this, even if you forget to stop. Raise it for big builds. - The script appends instructions telling each agent to work in `/home/tenki/work`, serve its result on port 3000, and write `SUMMARY.md`. Don't repeat those in `--prompt`. ## Monitor Poll `$S status` roughly once a minute; don't spin. States are `running`, `done(exit-N)`, and `stopped` (died without finishing). A preview URL appears once a build answers on port 3000. Peek at one with `$S logs N`. Don't wait forever: at the deadline, judge what exists. ## Judge — you are the judge `$S collect` downloads each contestant to `.sandboxmaxxing/-results/N-agent/`: its code, `SUMMARY.md`, and `AGENT.log` (what the agent did). For each one: 1. Read `SUMMARY.md`, then the code. 2. Run its tests or build, if it has any. 3. Open its preview URL from `status`. Judge on the result against the brief — does it work, how fully does it meet it, is the code sound — never on which agent made it. Name a winner and give each contestant a sentence or two. If none works, say so; don't crown a broken build. ## Finish - `$S keep N` keeps the winner alive past `--minutes` and prints its URL. Only if the user wants it: a kept sandbox costs money until terminated. - **Always** `$S stop` at the end. It terminates every contestant that was not kept. - Report: the winner, its preview URL, the comparison, and where the results were saved. ## Rules - Act only on sandboxes this arena created. The script enforces it — every sandbox carries `sandboxmaxxing=` metadata and an `smx--` name, and it touches nothing else. - Never print, log, or paste a key. Keys reach a sandbox as a file over the Tenki API, never on a command line, and live outside the directory the agent builds in, so they cannot be committed or come back in `collect`. - `.sandboxmaxxing/` holds state and downloads; it ignores itself in git. Leave it that way. ## When something goes wrong - **`done(exit-1)` within a minute:** `$S logs N`. Usually a missing or wrong key. - **opencode "1.18.0 or newer is required":** the script upgrades opencode when it is older, which adds ~15s to its start. If it still fails, check `/home/tenki/opencode-upgrade.log`. - **A build never gets a preview URL:** it didn't serve on port 3000. Judge its collected code instead. - **Tenki API hiccups:** remote calls retry transient failures on their own.