iii-sandbox
Ephemeral microVM sandboxes for running untrusted or agent-generated code in isolation — a one-call run path, a create/exec/stop lifecycle, and a set of filesystem operations.
What this skill does
# iii-sandbox
The `iii-sandbox` worker boots ephemeral libkrun microVMs and runs code inside them, isolated from the host. Each sandbox boots in a few hundred milliseconds, runs commands or file operations scoped to the engine it lives on, and is reaped when idle; the overlay filesystem is discarded on stop. Its `sandbox::*` functions are called like any other iii function.
There are two ways in. `sandbox::run` is the fast path: it boots a VM, runs a code snippet, captures stdout/stderr, and tears the VM down in a single call. For multi-step work — several commands, multiple files, or inspecting a VM between steps — use the `sandbox::create` → `sandbox::exec` / `sandbox::fs::*` → `sandbox::stop` lifecycle and carry the returned sandbox id across calls.
Prerequisites: the worker is enabled by adding `iii-sandbox` to `config.yaml`, and the host needs hardware virtualization (Apple Silicon, or `/dev/kvm` on Linux) — Intel Macs and Windows cannot boot sandboxes.
## Get the contract from the engine
This page tells you WHEN and HOW to use the sandbox; the engine is the source of truth for every function's exact arguments and responses. Before you call any `sandbox::*` function, fetch its contract and build your payload from that — do not guess field names, types, or formats from memory:
```
engine::functions::info { function_id: "sandbox::<fn>" }
```
(e.g. `{ function_id: "sandbox::fs::write" }`). The returned schema is authoritative; this skill never restates it.
## Running an iii worker inside a sandbox
To author an iii worker (read the `iii` skill first), you write the worker code into the sandbox, install its deps, and run it as a process there — but the worker must still join the HOST engine bus to be useful.
- Reaching the host engine: enable networking when you create the sandbox (the field is in the `sandbox::create` contract — fetch it), AND set the worker's engine-URL env (`III_ENGINE_URL`, e.g. `ws://localhost:49134`) IN THE `sandbox::create` `env` — NOT later at exec/run time. The localhost→host rewrite happens ONLY for the create-time `env`: a `localhost` / `127.0.0.1` engine URL set there is rewritten to the per-sandbox gateway so the worker reaches the host bus. The same `localhost` URL passed at `sandbox::exec` time is NOT rewritten — it points at the sandbox's own loopback, the worker silently fails to connect, and you will waste turns debugging TCP. So: put the engine URL in the create `env`; the worker process inherits it already rewritten.
- Run it backgrounded, never as a foreground `exec`: starting the worker with `sandbox::exec "node index.mjs"` blocks and times out (`S200 exec timed out`), because the process never returns. Launch it detached (`… &`, `nohup`, or `sandbox::run` shell mode) so the exec call returns while the worker keeps running — see the serialized-exec boundary below.
- Find the result: once the worker connects, its functions appear in `engine::functions::list` and any `http` trigger it binds is served by the `iii-http` worker; verify endpoints with `web::fetch`, not `curl`.
- EPHEMERAL by nature: a worker hosted in a sandbox dies when the sandbox is stopped or reaped (the overlay is discarded). That is fine for a demo, test, or one-off. For a worker that must stay up, install it as a managed worker (`worker::add`, see the `iii` skill) instead of hosting it in a sandbox.
## When to Use
- Execute untrusted or LLM-generated code without exposing the host.
- One-shot "code in, stdout out" — reach for `sandbox::run`.
- Spin up an ephemeral build worker, integration-test fixture, or throwaway iii worker.
- Read, write, search, or edit files inside an isolated filesystem.
## Boundaries
- A sandbox is for running code and build steps, NOT for hosting a long-lived server. To expose an HTTP API or any always-on endpoint, author an iii worker and register a trigger for it (an `http` trigger is served by the `iii-http` worker) — do NOT start a server process (`express`, `http.createServer`, a framework `listen()`) inside the sandbox. A server you start here is not routed by iii, is unreachable as an iii endpoint, and as a foreground process hangs the exec slot until it times out. If the task is "build an iii worker," the deliverable is registered functions plus a trigger, not a running server.
- Not for long-lived services or durable state — the overlay is wiped on stop. Use a regular worker for daemons and `iii-state` for persistence.
- Bootable images are catalog NAMES (`node`, `python`, or an operator-registered custom image), never arbitrary OCI references. Discover the live set with `sandbox::catalog::list`; an unknown name fails fast.
- `sandbox::exec` runs one command at a time per sandbox (serialized). A concurrent call is rejected, and waiting does NOT free the slot when the in-flight command is long-running or foreground (a server, `npm install`, a build) — it holds the slot until it exits. Run those in the background, or use `sandbox::run` in shell mode, or recover by replacing the sandbox (`sandbox::stop` then `sandbox::create`).
- `sandbox::exec` is not a shell — for pipes, `&&`, redirects, or variable expansion use `sandbox::run` in shell mode, or wrap the command in `sh -c`.
- Requires host virtualization; boot failures on unsupported hosts are reported as errors, not silent hangs.
- Sandboxes are network-isolated by default; reaching host services is opt-in at create time — see `engine::functions::info` on `sandbox::create` for how to enable it.
## Functions
A map of what exists. For the arguments and response of any one, call `engine::functions::info { function_id: "<id>" }`.
- `sandbox::run` — boot a VM, run a code snippet, capture output, and auto-stop, in one call.
- `sandbox::create` — boot a microVM you then drive with the lifecycle ops; returns the id every other op uses.
- `sandbox::exec` — run a single command inside a running sandbox.
- `sandbox::stop` — destroy a sandbox and reclaim its resources.
- `sandbox::list` — enumerate sandboxes currently running on this engine.
- `sandbox::catalog::list` — list bootable image names (bundled presets plus operator custom images).
- `sandbox::fs::write` — write a file into a sandbox.
- `sandbox::fs::read` — read a file out of a sandbox.
- `sandbox::fs::ls` — list the entries of a directory.
- `sandbox::fs::stat` — read metadata for a single path.
- `sandbox::fs::mkdir` — create a directory.
- `sandbox::fs::rm` — remove a file or directory.
- `sandbox::fs::mv` — move or rename a path.
- `sandbox::fs::chmod` — change a path's permissions (and optionally owner).
- `sandbox::fs::grep` — recursive regex search across files.
- `sandbox::fs::sed` — regex find-and-replace across files.
Every error comes back structured with a machine-readable hint describing how to fix it — read it and apply the fix before retrying, rather than re-sending the same call.
Related in AI Agents
skill-development
IncludedComprehensive meta-skill for creating, managing, validating, auditing, and distributing Claude Code skills and slash commands (unified in v2.1.3+). Provides skill templates, creation workflows, validation patterns, audit checklists, naming conventions, YAML frontmatter guidance, progressive disclosure examples, and best practices lookup. Use when creating new skills, validating existing skills, auditing skill quality, understanding skill architecture, needing skill templates, learning about YAML frontmatter requirements, progressive disclosure patterns, tool restrictions (allowed-tools), skill composition, skill naming conventions, troubleshooting skill activation issues, creating custom slash commands, configuring command frontmatter, using command arguments ($ARGUMENTS, $1, $2), bash execution in commands, file references in commands, command namespacing, plugin commands, MCP slash commands, Skill tool configuration, or deciding between skills vs slash commands. Delegates to docs-management skill for official documentation.
reprompter
IncludedTransform messy prompts into well-structured, effective prompts — single or multi-agent. Use when: "reprompt", "reprompt this", "clean up this prompt", "structure my prompt", rough text needing XML tags and best practices, "reprompter teams", "repromptception", "run with quality", "smart run", "smart agents", multi-agent tasks, audits, parallel work, anything going to agent teams. Don't use when: simple Q&A, pure chat, immediate execution-only tasks. See "Don't Use When" section for details. Outputs: Structured XML/Markdown prompt, quality score (before/after), optional team brief + per-agent sub-prompts, agent team output files. Success criteria: Single mode quality score ≥ 7/10; Repromptception per-agent prompt quality score 8+/10; all required sections present, actionable and specific.
adaptive-compaction
IncludedAdaptive add-on policy and recovery layer that decides WHEN to compact, prune, snapshot, or fork -- replacing fixed-percent auto-compaction across Claude Code, Codex, and MCP-capable hosts. Trigger on auto-compact timing or damage: "when should I compact", "is it safe to compact now or start a fresh session", "auto-compact fires too early/mid-task", "switching to an unrelated task but the window still has space", "context rot", "answers get worse the longer the session runs", "the agent forgot the plan or my decisions after it summarized", "add a layer on top that manages context without changing the agent", raising autoCompactWindow to give the policy room, or installing/tuning a cross-tool compaction policy or PreCompact hook -- even when "compaction" is never said but the problem is context-window pressure or post-summarization memory loss. Do NOT use to summarize a conversation, build RAG, write a summarization prompt (decides WHEN not HOW), or answer max-context-length trivia.
agent-skill-creator
IncludedCreate cross-platform agent skills from workflow descriptions. Activates when users ask to create an agent, automate a repetitive workflow, create a custom skill, or need advanced agent creation. Triggers on phrases like create agent for, automate workflow, create skill for, every day I have to, daily I need to, turn process into agent, need to automate, create a cross-platform skill, validate this skill, export this skill, migrate this skill. Supports single skills, multi-agent suites, transcript processing, template-based creation, interactive configuration, cross-platform export, and spec validation.
llm-wiki
IncludedUse when building or maintaining a persistent personal knowledge base (second brain) in Obsidian where an LLM incrementally ingests sources, updates entity/concept pages, maintains cross-references, and keeps a synthesis current. Triggers include "second brain", "Obsidian wiki", "personal knowledge management", "ingest this paper/article/book", "build a research wiki", "compound knowledge", "Memex", or whenever the user wants knowledge to accumulate across sessions instead of being re-derived by RAG on every query.
skill-master
IncludedAgent Skills authoring, evaluation, and optimization. Create, edit, validate, benchmark, and improve skills following the agentskills.io specification. Use when designing SKILL.md files, structuring skill folders (references, scripts, assets), ingesting external documentation into skills, running trigger evals, benchmarking skill quality, optimizing descriptions, or performing blind A/B comparisons. Keywords: agentskills.io, SKILL.md, skill authoring, eval, benchmark, trigger optimization.