mattpocock/sandcastle

What Is Sandcastle?

A TypeScript library for orchestrating AI coding agents in isolated sandboxes:

1. You invoke agents with a single sandcastle.run().
2. Sandcastle handles sandboxing the agent with a configurable branch strategy.
3. The commits made on the branches get merged back.

Sandcastle is provider-agnostic — it ships with built-in providers for Docker, Podman, and Vercel, and you can create your own. Great for parallelizing multiple AFK agents, creating review pipelines, or even just orchestrating your own agents.

Prerequisites

• Git
• A sandbox provider — Sandcastle needs an isolated environment to run agents in. Built-in options:
  • Docker Desktop — most common for local development
  • Podman — rootless alternative to Docker
  • Vercel — cloud-based Firecracker microVMs via @vercel/sandbox
  • Or create your own using createBindMountSandboxProvider or createIsolatedSandboxProvider

Quick start

1. Install the package:

npm install --save-dev @ai-hero/sandcastle

2. Run npx @ai-hero/sandcastle init. This scaffolds a .sandcastle directory with all the files needed.

npx @ai-hero/sandcastle init

3. Copy .sandcastle/.env.example to .sandcastle/.env, then fill in your default values for CLAUDE_CODE_OAUTH_TOKEN (run claude setup-token on your host to get one). To use an Anthropic API key instead, uncomment and fill in ANTHROPIC_API_KEY.

cp .sandcastle/.env.example .sandcastle/.env

4. Run the .sandcastle/main.ts (or main.mts) file with npx tsx

npx tsx .sandcastle/main.ts

// 4. Run the agent via the JS API
import { run, claudeCode } from "@ai-hero/sandcastle";
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";

await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(), // or podman(), vercel(), or your own provider
  promptFile: ".sandcastle/prompt.md",
});

Sandbox Providers

Sandcastle uses a SandboxProvider to create isolated environments. The sandbox option on run(), interactive(), and createSandbox() accepts any provider, including noSandbox() — opt in to running the agent directly on the host when container isolation is undesired. Built-in providers:

Worktree methods (wt.run(), wt.interactive(), wt.createSandbox()) accept the same providers as their top-level counterparts. wt.interactive() defaults to noSandbox() when no sandbox is specified.

import { docker } from "@ai-hero/sandcastle/sandboxes/docker";
import { podman } from "@ai-hero/sandcastle/sandboxes/podman";
import { vercel } from "@ai-hero/sandcastle/sandboxes/vercel";
import { noSandbox } from "@ai-hero/sandcastle/sandboxes/no-sandbox";

// Docker, Podman, and Vercel are interchangeable in run() and createSandbox():
await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  prompt: "...",
});

// No-sandbox runs the agent directly on the host — accepted by run(),
// createSandbox(), and interactive(). Skips container isolation entirely:
await interactive({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: noSandbox(),
  prompt: "...", // optional — omit to launch the TUI with no initial prompt
  cwd: "/path/to/other-repo", // optional — defaults to process.cwd()
});

You can also create your own provider using createBindMountSandboxProvider or createIsolatedSandboxProvider.

API

Sandcastle exports a programmatic run() function for use in scripts, CI pipelines, or custom tooling. The examples below use docker(), but any SandboxProvider works in its place.

import { run, claudeCode } from "@ai-hero/sandcastle";
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";

const result = await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  promptFile: ".sandcastle/prompt.md",
});

console.log(result.iterations.length); // number of iterations executed
console.log(result.iterations); // per-iteration results with optional sessionId
console.log(result.commits); // array of { sha } for commits created
console.log(result.branch); // target branch name

All options

import { run, claudeCode } from "@ai-hero/sandcastle";
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";

const result = await run({
  // Agent provider — required. Pass a model string to claudeCode().
  // Optional second arg for provider-specific options like effort level.
  agent: claudeCode("claude-opus-4-8", { effort: "high" }),

  // Sandbox provider — required. Any SandboxProvider works (docker, podman, vercel, or custom).
  // Provider-specific config (like imageName, mounts) lives inside the provider factory call.
  sandbox: docker({
    imageName: "sandcastle:local",
    // Optional: override the UID/GID used for --user flag (defaults to host UID/GID).
    // Must match the UID baked into the image. Pre-flight check catches mismatches.
    // containerUid: 1000,
    // containerGid: 1000,
    // Optional: mount host directories into the sandbox (e.g. package manager caches)
    // hostPath supports absolute, tilde-expanded (~), and relative paths (resolved from cwd).
    // sandboxPath supports absolute and relative paths (resolved from the sandbox repo directory).
    mounts: [
      { hostPath: "~/.npm", sandboxPath: "/home/agent/.npm", readonly: true },
      { hostPath: "data", sandboxPath: "data" }, // mounts <cwd>/data → <sandbox-repo>/data
    ],
    // Optional: SELinux volume label — "z" (default, shared), "Z" (private), or false (none).
    // No-op on non-SELinux systems (Docker Desktop on macOS/Windows, Linux without SELinux).
    selinuxLabel: "z",
    // Optional: provider-level env vars merged at launch time
    env: { DOCKER_SPECIFIC: "value" },
    // Optional: attach container to Docker network(s) — string or string[]
    network: "my-network",
    // Optional: add the container user to supplementary groups via --group-add.
    // Accepts group names or numeric GIDs (e.g. for a bind-mounted Docker socket).
    groups: ["docker", 999],
    // Optional: expose host devices via --device. Each entry is a full device
    // spec in host[:container[:permissions]] form (e.g. "/dev/kvm").
    devices: ["/dev/kvm"],
    // Optional: limit CPU resources via --cpus. Fractional values allowed (e.g. 1.5).
    // cpus: 2,
  }),

  // Host repo directory — replaces process.cwd() as the anchor for
  // .sandcastle/ artifacts (worktrees, logs, env, patches) and git operations.
  // Relative paths resolve against process.cwd(). Defaults to process.cwd().
  cwd: "../other-repo",

  // Branch strategy — controls how the agent's changes relate to branches.
  // Defaults to { type: "head" } for bind-mount and { type: "merge-to-head" } for isolated providers.
  branchStrategy: { type: "branch", branch: "agent/fix-42" },

  // Prompt source — provide one of these, not both.
  // Note: promptFile resolves against process.cwd(), NOT cwd.
  promptFile: ".sandcastle/prompt.md", // path to a prompt file
  // prompt: "Fix issue #42 in this repo", // OR an inline prompt string

  // Values substituted for {{KEY}} placeholders in the prompt.
  promptArgs: {
    ISSUE_NUMBER: "42",
  },

  // Maximum number of agent iterations to run before stopping. Default: 1
  maxIterations: 5,

  // Display name for this run, shown as a prefix in log output.
  name: "fix-issue-42",

  // Lifecycle hooks grouped by where they run: host or sandbox.
  hooks: {
    host: {
      onWorktreeReady: [{ command: "cp .env.example .env" }],
      onSandboxReady: [{ command: "echo setup done" }],
    },
    sandbox: {
      onSandboxReady: [{ command: "npm install" }],
    },
  },

  // Host-relative file paths to copy into the sandbox before the container starts.
  // Not supported with branchStrategy: { type: "head" }.
  copyToWorktree: [".env"],

  // Override default timeouts for built-in lifecycle steps.
  // Unset keys keep their defaults.
  timeouts: {
    copyToWorktreeMs: 120_000, // default: 60_000
    gitSetupMs: 30_000, // default: 10_000
    commitCollectionMs: 60_000, // default: 30_000
    mergeToHostMs: 60_000, // default: 30_000
  },

  // How to record progress. Default: write to a file under .sandcastle/logs/
  logging: {
    type: "file",
    path: ".sandcastle/logs/my-run.log",
    // Optional: forward the agent's output stream to your own observability system.
    // Fires for each text chunk, tool call, and raw stdout line the agent
    // produces. Errors thrown by the callback are swallowed so a broken
    // forwarder cannot kill the run.
    onAgentStreamEvent: (event) => {
      // event is { type: "text" | "toolCall" | "raw", iteration, timestamp, ... }
      myLogger.info(event);
    },
    // Optional: append every raw stdout line the agent emits to the same
    // log file, interleaved with the human-readable output. Includes lines
    // the provider's stream parser would otherwise drop. Intended for
    // debugging stuck or unexpected agent behaviour.
    verbose: true,
  },
  // logging: { type: "stdout", verbose: true }, // OR terminal mode (verbose: raw lines to stdout)

  // String (or array of strings) the agent emits to end the iteration loop early.
  // Default: "<promise>COMPLETE</promise>"
  completionSignal: "<promise>COMPLETE</promise>",

  // Idle timeout in seconds — resets whenever the agent produces output. Default: 600 (10 minutes)
  idleTimeoutSeconds: 600,

  // Grace window in seconds after the agent emits a completion signal but
  // before its process has exited (a "hanging process" — typically a spawned
  // `gh`/git child or MCP server keeping stdout open). Resets on every
  // subsequent output line so trailing data is still captured. Default: 60
  completionTimeoutSeconds: 60,

  // Structured output — extract a typed payload from the agent's stdout.
  // Requires maxIterations === 1 and the tag must appear in the prompt.
  // output: Output.object({ tag: "result", schema: z.object({ answer: z.number() }) }),
  // output: Output.string({ tag: "summary" }),
});

console.log(result.iterations.length); // number of iterations executed
console.log(result.completionSignal); // matched signal string, or undefined if none fired
console.log(result.commits); // array of { sha } for commits created
console.log(result.branch); // target branch name

createSandbox() — reusable sandbox

Use createSandbox() when you need to run multiple agents (or multiple rounds of the same agent) inside a single sandbox. It creates the sandbox once, and you call sandbox.run() as many times as you need. This avoids repeated container startup costs and keeps all runs on the same branch.

Use run() instead when you only need a single one-shot invocation — it handles sandbox lifecycle automatically.

Basic single-run usage

import { createSandbox, claudeCode } from "@ai-hero/sandcastle";
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";

await using sandbox = await createSandbox({
  branch: "agent/fix-42",
  sandbox: docker(),
});

const result = await sandbox.run({
  agent: claudeCode("claude-opus-4-8"),
  prompt: "Fix issue #42 in this repo.",
});

console.log(result.commits); // [{ sha: "abc123" }]

Multi-run implement-then-review

import { createSandbox, claudeCode } from "@ai-hero/sandcastle";
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";

await using sandbox = await createSandbox({
  branch: "agent/fix-42",
  sandbox: docker(),
  hooks: { sandbox: { onSandboxReady: [{ command: "npm install" }] } },
});

// Step 1: implement
const implResult = await sandbox.run({
  agent: claudeCode("claude-opus-4-8"),
  promptFile: ".sandcastle/implement.md",
  maxIterations: 5,
});

// Step 2: review on the same branch, same container
const reviewResult = await sandbox.run({
  agent: claudeCode("claude-sonnet-4-6"),
  prompt: "Review the changes and fix any issues.",
});

Commits from all run() calls accumulate on the same branch. The sandbox container stays alive between runs, so installed dependencies and build artifacts persist.

sandbox.exec() lets the harness run shell commands directly in the same warm sandbox — handy for gating an implement step on a quick verification before kicking off the review:

await using sandbox = await createSandbox({
  branch: "agent/fix-42",
  sandbox: docker(),
  hooks: { sandbox: { onSandboxReady: [{ command: "npm install" }] } },
});

await sandbox.run({
  agent: claudeCode("claude-opus-4-8"),
  promptFile: ".sandcastle/implement.md",
  maxIterations: 5,
});

// Verify before review — non-zero exitCode is returned, not thrown.
const tests = await sandbox.exec("npm test");
if (tests.exitCode !== 0) {
  throw new Error(`Tests failed:\n${tests.stdout}\n${tests.stderr}`);
}

await sandbox.run({
  agent: claudeCode("claude-sonnet-4-6"),
  prompt: "Review the changes and fix any issues.",
});

cwd defaults to the sandbox repo path, matching interactive(). Pass cwd to override.

Automatic cleanup with await using

await using calls sandbox.close() automatically when the block exits. If the sandbox has uncommitted changes, the worktree is preserved on disk; if clean, both container and worktree are removed.

Manual close() with CloseResult

const sandbox = await createSandbox({
  branch: "agent/fix-42",
  sandbox: docker(),
});
// ... run agents ...
const closeResult = await sandbox.close();
if (closeResult.preservedWorktreePath) {
  console.log(`Worktree preserved at ${closeResult.preservedWorktreePath}`);
}

CreateSandboxOptions

Sandbox

SandboxRunOptions

SandboxRunResult

CloseResult

createWorktree() — independent worktree lifecycle

Use createWorktree() when you need a worktree (git worktree) as an independent, first-class concept — separate from any sandbox. This is useful when you want to run an interactive session first and then hand the same worktree to a sandboxed AFK agent.

Only branch and merge-to-head strategies are accepted; head is a compile-time type error since it means no worktree.

Pass cwd to target a repo other than process.cwd(). Relative paths resolve against process.cwd(); absolute paths pass through. A CwdError is thrown if the path does not exist or is not a directory.

import { createWorktree } from "@ai-hero/sandcastle";

await using wt = await createWorktree({
  branchStrategy: { type: "branch", branch: "agent/fix-42" },
  copyToWorktree: ["node_modules"],
  cwd: "/path/to/other-repo", // optional — defaults to process.cwd()
});

console.log(wt.worktreePath); // host path to the worktree
console.log(wt.branch); // "agent/fix-42"

// Run an interactive session in the worktree (defaults to noSandbox)
await wt.interactive({
  agent: claudeCode("claude-opus-4-8"),
  prompt: "Explore the codebase and understand the bug.",
});

// Run an AFK agent in the worktree (sandbox is required)
const result = await wt.run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker({ imageName: "sandcastle:myrepo" }),
  prompt: "Fix issue #42.",
  maxIterations: 3,
});
console.log(result.commits); // commits made during the run

// Create a long-lived sandbox from the worktree
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";

await using sandbox = await wt.createSandbox({
  sandbox: docker(),
  hooks: { sandbox: { onSandboxReady: [{ command: "npm install" }] } },
});

// sandbox.close() tears down the container only — the worktree stays
await sandbox.close();

// wt.close() cleans up the worktree

wt.close() checks for uncommitted changes: if the worktree is dirty, it's preserved on disk; if clean, it's removed. await using calls close() automatically. The worktree persists after run(), interactive(), and createSandbox() complete, so you can hand it to another agent or inspect it.

With branchStrategy: { type: "merge-to-head" }, each wt.run() / wt.interactive() merges the agent's commits back to the host's current branch before returning, and the worktree's source branch is preserved across calls so subsequent ones can reuse the same handle. (This differs from top-level run(), where the temp branch is deleted after the merge.)

Split ownership: When a sandbox is created via wt.createSandbox(), sandbox.close() tears down the container only — the worktree remains. wt.close() is responsible for worktree cleanup. This differs from the top-level createSandbox(), where sandbox.close() owns both container and worktree.

CreateWorktreeOptions

Worktree

WorktreeInteractiveOptions

WorktreeRunOptions

WorktreeRunResult

WorktreeCreateSandboxOptions

How it works

Sandcastle uses a branch strategy configured on the sandbox provider to control how the agent's changes relate to branches. There are three strategies:

• Head ({ type: "head" }) — The agent writes directly to the host working directory. No worktree, no branch indirection. This is the default for bind-mount providers like docker().
• Merge-to-head ({ type: "merge-to-head" }) — Sandcastle creates a temporary branch in a git worktree. The agent works on the temp branch, and changes are merged back to HEAD when done. The temp branch is cleaned up after merge.
• Branch ({ type: "branch", branch: "foo" }) — Commits land on an explicitly named branch in a git worktree. Re-running with the same branch reuses the existing worktree and fast-forwards it from origin when safe — see ADR 0003.

For bind-mount providers (like Docker), the worktree directory is bind-mounted into the container — the agent writes directly to the host filesystem through the mount, so no sync is needed.

From your point of view, you just configure branchStrategy: { type: 'branch', branch: 'foo' } on run(), and get a commit on branch foo once it's complete. All 100% local.

Prompts

Sandcastle uses a flexible prompt system. You write the prompt, and the engine executes it — no opinions about workflow, task management, or context sources are imposed.

Prompt resolution

You must provide exactly one of:

1. prompt: "inline string" — pass an inline prompt directly via RunOptions
2. promptFile: "./path/to/prompt.md" — point to a specific file via RunOptions

prompt and promptFile are mutually exclusive — providing both is an error. If neither is provided, run() throws an error asking you to supply one.

Inline prompts (prompt: "...") are passed to the agent literally. No {{KEY}} substitution, no !`command` expansion, no built-in {{SOURCE_BRANCH}} / {{TARGET_BRANCH}} injection. If you need values interpolated into an inline prompt, build the string in JavaScript (`Work on ${branch}…`). Passing promptArgs alongside an inline prompt is an error — switch to promptFile to use substitution.

The substitution and expansion features below apply only to prompts sourced from promptFile.

Convention: sandcastle init scaffolds .sandcastle/prompt.md and all templates explicitly reference it via promptFile: ".sandcastle/prompt.md". This is a convention, not an automatic fallback — Sandcastle does not read .sandcastle/prompt.md unless you pass it as promptFile.

Dynamic context with !`command`

Use !`command` expressions in your prompt to pull in dynamic context. Each expression is replaced with the command's stdout before the prompt is sent to the agent. All expressions in a prompt run in parallel for faster expansion.

Commands run inside the sandbox after sandbox.onSandboxReady hooks complete, so they see the same repo state the agent sees (including installed dependencies).

# Open issues

!`gh issue list --state open --label Sandcastle --json number,title,body,comments,labels --limit 100`

# Recent commits

!`git log --oneline -10`

If any command exits with a non-zero code, the run fails immediately with an error.

Prompt arguments with {{KEY}}

Use {{KEY}} placeholders in your prompt to inject values from the promptArgs option. This is useful for reusing the same prompt file across multiple runs with different parameters.

import { run } from "@ai-hero/sandcastle";

await run({
  promptFile: "./my-prompt.md",
  promptArgs: { ISSUE_NUMBER: 42, PRIORITY: "high" },
});

In the prompt file:

Work on issue #{{ISSUE_NUMBER}} (priority: {{PRIORITY}}).

Prompt argument substitution runs on the host before shell expression expansion, so {{KEY}} placeholders inside !`command` expressions are replaced first:

!`gh issue view {{ISSUE_NUMBER}} --json body -q .body`

A {{KEY}} placeholder with no matching prompt argument is an error. Unused prompt arguments produce a warning.

!`command` expansion only runs on shell blocks written in the prompt file itself. Any !`…` pattern that appears inside an argument value is treated as inert text — it won't be executed against the host shell. This makes it safe to pass user-authored content (issue titles, PR descriptions, docs excerpts) through promptArgs.

Built-in prompt arguments

Sandcastle automatically injects two built-in prompt arguments into every prompt:

Use them in your prompt without passing them via promptArgs:

You are working on {{SOURCE_BRANCH}}. When diffing, compare against {{TARGET_BRANCH}}.

Passing SOURCE_BRANCH or TARGET_BRANCH in promptArgs is an error — built-in prompt arguments cannot be overridden.

Early termination with <promise>COMPLETE</promise>

When the agent outputs <promise>COMPLETE</promise>, the orchestrator stops the iteration loop early. This is a convention you document in your prompt for the agent to follow — the engine never injects it.

This is useful for task-based workflows where the agent should stop once it has finished, rather than running all remaining iterations.

You can override the default signal by passing completionSignal to run(). It accepts a single string or an array of strings:

await run({
  // ...
  completionSignal: "DONE",
});

// Or pass multiple signals — the loop stops on the first match:
await run({
  // ...
  completionSignal: ["TASK_COMPLETE", "TASK_ABORTED"],
});

Tell the agent to output your chosen string(s) in the prompt, and the orchestrator will stop when it detects any of them. The matched signal is returned as result.completionSignal.

Hanging processes after the completion signal

The agent process is expected to exit shortly after emitting the completion signal. When a child it spawned — a gh/git subprocess, a long-lived MCP server, etc. — inherits the agent's stdout pipe and keeps it open, the parent process can linger long past its logical end. Sandcastle would otherwise wait for the full idleTimeoutSeconds and fail with AgentIdleTimeoutError, throwing away the commits the agent already made.

Instead, once the completion signal is observed in the output buffer, Sandcastle swaps in a short completion timeout (default 60 s). When it expires, the run resolves successfully with a warning that the process was hanging; result.commits and result.completionSignal are populated as if the process had exited cleanly. The timer resets on every subsequent output line, so trailing data emitted after the signal — token-usage events, terminal result events, a structured-output <tag> — is still captured.

A clean process exit always wins the race, so healthy runs gain zero added latency. The completion timeout only matters when the process hangs.

Tune the window with completionTimeoutSeconds:

await run({
  // ...
  completionTimeoutSeconds: 30, // shorter grace window
});

This is independent of idleTimeoutSeconds. They cover different phases: idleTimeoutSeconds runs before any signal is seen (genuinely stuck agent → fail); completionTimeoutSeconds runs after the signal is seen (hanging process → succeed with warning). See ADR 0019.

Structured output

Use Output.object() to extract a typed, schema-validated JSON payload from the agent's stdout. The agent emits its answer inside an XML tag you specify, and Sandcastle parses, validates, and returns it on result.output. The schema can be any Standard Schema validator — the examples below use Zod, but Valibot, ArkType, and others work identically. See ADR 0010 for design rationale.

import { run, Output, claudeCode } from "@ai-hero/sandcastle";
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";
import { z } from "zod";

const result = await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  prompt: `Analyze the code, and output the result as JSON inside <result> tags.
    The result must match this schema:
    { summary: string; score: string }
  `,
  output: Output.object({
    tag: "result",
    schema: z.object({ summary: z.string(), score: z.number() }),
  }),
});

console.log(result.output.summary); // typed as string
console.log(result.output.score); // typed as number

Output.string({ tag }) extracts the tag contents as a plain string (trimmed, no JSON parsing). Both helpers require maxIterations to be 1 (the default). The resolved prompt must contain the configured opening tag literal.

When extraction or validation fails, run() throws a StructuredOutputError. Alongside tag, rawMatched, cause, commits, branch, and preservedWorktreePath, the error carries the sessionId (and sessionFilePath, when the session was captured) of the run that produced the bad output.

Pass maxRetries to have Sandcastle handle the retry loop for you. Each retry resumes the same agent session and feeds back a token-efficient description of the error, so the agent can re-emit a corrected tag without redoing the work. Retries require an agent provider that supports session resumption (claudeCode, codex, pi) — calling run() with maxRetries > 0 against a non-resumable provider (cursor, opencode, copilot) throws immediately.

const result = await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  prompt: "Analyze the code and emit JSON inside <result> tags.",
  output: Output.object({
    tag: "result",
    schema: z.object({ summary: z.string(), score: z.number() }),
    maxRetries: 2, // 2 retries on top of the initial attempt
  }),
});

If you need to drive the retry loop manually — for example, to customise the feedback prompt or rotate models on each attempt — leave maxRetries at its default of 0 and resume the failed session yourself:

import { run, Output, StructuredOutputError } from "@ai-hero/sandcastle";

try {
  return await run({ ...opts, output });
} catch (e) {
  if (e instanceof StructuredOutputError && e.sessionId) {
    return await run({
      ...opts,
      output,
      resumeSession: e.sessionId,
      prompt: `Your previous output failed: ${e.message}. Re-emit it inside <${e.tag}> tags.`,
    });
  }
  throw e;
}

Templates

sandcastle init prompts you to choose a sandbox provider (Docker or Podman), an issue tracker (GitHub Issues, Beads, or Custom), and a template, which scaffolds a ready-to-use prompt and main.mts suited to a specific workflow. If your project's package.json has "type": "module", the file will be named main.ts instead. Choosing Custom scaffolds the project in a deliberately broken-until-configured state plus a .sandcastle/SETUP_ISSUE_TRACKER.md prompt you feed to your coding agent, which wires up your own tracker by editing the scaffolded files in place. Five templates are available:

Select a template during sandcastle init when prompted, or re-run init in a fresh repo to try a different one.

CLI commands

sandcastle init

Scaffolds the .sandcastle/ config directory and builds the container image. This is the first command you run in a new repo. You choose a sandbox provider (Docker or Podman) during init — selecting Podman writes a Containerfile instead of Dockerfile and uses sandcastle podman build-image for the build step.

Init detects your host package manager (npm, pnpm, yarn, or bun) from a packageManager field or lockfile, defaulting to npm. Templates whose main file imports a host dependency — the planner templates import Zod for their <plan> output schema — prompt you to install it with that package manager when it isn't already in your package.json, so the first npx tsx .sandcastle/main.ts doesn't fail with ERR_MODULE_NOT_FOUND.

Every interactive prompt has a paired --flag so the entire init can run non-interactively (e.g. in CI or a scripted setup). When stdin is not a TTY and a required flag is missing, init fails fast with a clear error rather than wedging on a prompt.

Creates the following files:

.sandcastle/
├── Dockerfile      # Sandbox environment (customize as needed)
├── main.mts        # Agent entry point (main.ts in an ESM package)
├── prompt.md       # Agent instructions
├── .env.example    # Token placeholders
└── .gitignore      # Ignores .env, logs/, worktrees/

Errors if .sandcastle/ already exists to prevent overwriting customizations.

sandcastle docker build-image

Rebuilds the Docker image from an existing .sandcastle/ directory. Use this after modifying the Dockerfile. On Linux/macOS, the build automatically passes --build-arg AGENT_UID=$(id -u) and AGENT_GID=$(id -g) so the image's agent user matches the host UID — this prevents permission errors on image-built files without runtime chown.

sandcastle docker remove-image

Removes the Docker image.

sandcastle podman build-image

Builds the Podman image from an existing .sandcastle/ directory. Use this after modifying the Containerfile.

sandcastle podman remove-image

Removes the Podman image.

RunOptions

RunResult

IterationResult

IterationUsage

Session capture

After each resumable provider iteration, Sandcastle automatically captures the agent's session file from the sandbox to the host. Claude Code sessions are stored under ~/.claude/projects/<encoded-path>/<session-id>.jsonl; Codex sessions are stored under ~/.codex/sessions/YYYY/MM/DD/rollout-*-<session-id>.jsonl; Pi sessions are stored under ~/.pi/agent/sessions/--<encoded-cwd>--/<timestamp>_<session-id>.jsonl. Any provider-specific cwd fields are rewritten to match the host repo root, so the provider's native resume command works.

For Claude Code, any Agent-tool or Workflow-tool subagent transcripts written under <session-id>/subagents/agent-*.jsonl are captured alongside the main session. Subagent capture is best-effort: a failure on an individual transcript logs a warning and lets siblings and the main session through. Main-session capture failure still fails the run (see below).

Session capture is enabled by default for claudeCode(), codex(), and pi() and can be opted out via captureSessions: false. Providers without sessionStorage do not attempt capture. Capture failure fails the run.

Session resume

Pass resumeSession to run() to continue a prior Claude Code, Codex, or Pi conversation inside a new sandbox:

const result = await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  prompt: "Continue where you left off",
  resumeSession: "abc-123-def",
});

You can also continue the last captured session from a result:

const first = await run({
  agent: codex("gpt-5.4"),
  sandbox: docker(),
  prompt: "Draft a plan",
});

const second = await first.resume?.("Now implement the plan");

resume is present only on results from resumable providers (Claude Code, Codex, Pi) — hence the optional-chaining call.

Before the sandbox starts, Sandcastle validates that the session file exists on the host and transfers it into the sandbox with cwd fields rewritten to match the sandbox-side path. Claude Code receives --resume <id>; Codex receives codex exec resume <id> with the prompt piped over stdin; Pi receives --session <id>.

Constraints:

• resumeSession is incompatible with maxIterations > 1 (throws before sandbox creation).
• The provider's host session file must exist (throws before sandbox creation).
• Only iteration 1 receives the resume flag; subsequent iterations (if any) start fresh.
• Providers without resume support reject resumeSession.

Session fork

RunResult.fork(prompt, options?) is the sibling of .resume(): it continues from the last captured session but leaves the parent session JSONL untouched and writes the child under a new session id. The mechanism is the agent's native fork flag — claude --resume <id> --fork-session for Claude Code, codex exec fork <id> for Codex.

Fork enables fan-out workflows where a single parent run is the starting point for several independent children:

const parent = await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  prompt: "Read the codebase and summarise the data model",
});

const [reviewA, reviewB] = await Promise.all([
  parent.fork?.("Review the migration plan", {
    branchStrategy: { type: "branch", branch: "review-a" },
  }),
  parent.fork?.("Audit the auth layer", {
    branchStrategy: { type: "branch", branch: "review-b" },
  }),
]);

Fork is session-only. --fork-session and codex exec fork isolate the agent session JSONL — they do not isolate the branch, worktree, or sandbox. Safe concurrent fan-out (Promise.all([r.fork(a), r.fork(b)])) requires the caller to give each child a distinct branch via branchStrategy: { type: "branch", branch: "..." }. The default head and merge-to-head strategies are not safe for concurrent forks: head shares the host working directory across all children, and merge-to-head races git merge against the same HEAD. See ADR 0018.

fork is present only on results from providers with sessionStorage (Claude Code, Codex) — hence the optional-chaining call. The same single-iteration and session-file constraints as .resume() apply.

ClaudeCodeOptions

The claudeCode() factory accepts an optional second argument for provider-specific options:

agent: claudeCode("claude-opus-4-8", { effort: "high" });

CodexOptions

The codex() factory accepts an optional second argument for provider-specific options:

agent: codex("gpt-5.4", { effort: "high" });

PiOptions

The pi() factory accepts an optional second argument for provider-specific options:

agent: pi("claude-sonnet-4-6", { thinking: "high" });

Provider env

Both agent providers and sandbox providers accept an optional env: Record<string, string> in their options. These environment variables are merged with the .sandcastle/.env resolver output at launch time:

await run({
  agent: claudeCode("claude-opus-4-8", {
    env: { ANTHROPIC_API_KEY: "sk-ant-..." },
  }),
  sandbox: docker({
    env: { DOCKER_SPECIFIC_VAR: "value" },
  }),
  prompt: "Fix issue #42",
});

Merge rules:

• Provider env (agent + sandbox) overrides .sandcastle/.env resolver output for shared keys
• Agent provider env and sandbox provider env must not overlap — if they share any key, run() throws an error
• When env is not provided, it defaults to {}

Environment variables are also resolved automatically from .sandcastle/.env and process.env — no need to pass them to the API. The required variables depend on the agent provider (see sandcastle init output for details).

Custom Sandbox Providers

Sandcastle ships with built-in providers for Docker, Podman, and Vercel, but you can create your own. A sandbox provider tells Sandcastle how to execute commands in an isolated environment. There are two kinds:

• Bind-mount — the sandbox can mount a host directory. Sandcastle creates a worktree on the host and the provider mounts it in. No file sync needed. Use this for Docker, Podman, or any local container runtime.
• Isolated — the sandbox has its own filesystem (e.g. a cloud VM). The provider handles syncing code in and out via copyIn and copyFileOut. Use this when the sandbox cannot access the host filesystem.

The sandbox handle contract

Both provider types return a sandbox handle from their create() function. The handle exposes:

ExecResult

Every exec call returns an ExecResult:

interface ExecResult {
  readonly stdout: string;
  readonly stderr: string;
  readonly exitCode: number;
}

Bind-mount provider example

A minimal bind-mount provider that shells out to local processes (no container):

import {
  createBindMountSandboxProvider,
  type BindMountCreateOptions,
  type BindMountSandboxHandle,
  type ExecResult,
} from "@ai-hero/sandcastle";
import { execFile, spawn } from "node:child_process";
import { copyFile as fsCopyFile, mkdir as fsMkdir } from "node:fs/promises";
import { dirname } from "node:path";
import { createInterface } from "node:readline";

const localProcess = () =>
  createBindMountSandboxProvider({
    name: "local-process",
    create: async (
      options: BindMountCreateOptions,
    ): Promise<BindMountSandboxHandle> => {
      const worktreePath = options.worktreePath;

      return {
        worktreePath,

        exec: (
          command: string,
          opts?: { onLine?: (line: string) => void; cwd?: string },
        ): Promise<ExecResult> => {
          if (opts?.onLine) {
            const onLine = opts.onLine;
            return new Promise((resolve, reject) => {
              const proc = spawn("sh", ["-c", command], {
                cwd: opts?.cwd ?? worktreePath,
                stdio: ["ignore", "pipe", "pipe"],
              });

              const stdoutChunks: string[] = [];
              const stderrChunks: string[] = [];

              const rl = createInterface({ input: proc.stdout! });
              rl.on("line", (line) => {
                stdoutChunks.push(line);
                onLine(line); // forward each line to Sandcastle
              });

              proc.stderr!.on("data", (chunk: Buffer) => {
                stderrChunks.push(chunk.toString());
              });

              proc.on("error", (err) => reject(err));
              proc.on("close", (code) => {
                resolve({
                  stdout: stdoutChunks.join("\n"),
                  stderr: stderrChunks.join(""),
                  exitCode: code ?? 0,
                });
              });
            });
          }

          return new Promise((resolve, reject) => {
            execFile(
              "sh",
              ["-c", command],
              { cwd: opts?.cwd ?? worktreePath, maxBuffer: 10 * 1024 * 1024 },
              (error, stdout, stderr) => {
                if (error && error.code === undefined) {
                  reject(new Error(`exec failed: ${error.message}`));
                } else {
                  resolve({
                    stdout: stdout.toString(),
                    stderr: stderr.toString(),
                    exitCode: typeof error?.code === "number" ? error.code : 0,
                  });
                }
              },
            );
          });
        },

        copyFileIn: async (hostPath: string, sandboxPath: string) => {
          await fsMkdir(dirname(sandboxPath), { recursive: true });
          await fsCopyFile(hostPath, sandboxPath);
        },

        copyFileOut: async (sandboxPath: string, hostPath: string) => {
          await fsMkdir(dirname(hostPath), { recursive: true });
          await fsCopyFile(sandboxPath, hostPath);
        },

        close: async () => {
          // nothing to tear down for a local process
        },
      };
    },
  });

Isolated provider example

A minimal isolated provider using a temp directory:

import {
  createIsolatedSandboxProvider,
  type IsolatedSandboxHandle,
  type ExecResult,
} from "@ai-hero/sandcastle";
import { execFile, spawn } from "node:child_process";
import { copyFile, mkdir, mkdtemp, rm } from "node:fs/promises";
import { tmpdir } from "node:os";
import { dirname, join } from "node:path";
import { createInterface } from "node:readline";

const tempDir = () =>
  createIsolatedSandboxProvider({
    name: "temp-dir",
    create: async (): Promise<IsolatedSandboxHandle> => {
      const root = await mkdtemp(join(tmpdir(), "sandbox-"));
      const worktreePath = join(root, "workspace");
      await mkdir(worktreePath, { recursive: true });

      return {
        worktreePath,

        exec: (
          command: string,
          opts?: { onLine?: (line: string) => void; cwd?: string },
        ): Promise<ExecResult> => {
          if (opts?.onLine) {
            const onLine = opts.onLine;
            return new Promise((resolve, reject) => {
              const proc = spawn("sh", ["-c", command], {
                cwd: opts?.cwd ?? worktreePath,
                stdio: ["ignore", "pipe", "pipe"],
              });

              const stdoutChunks: string[] = [];
              const stderrChunks: string[] = [];

              const rl = createInterface({ input: proc.stdout! });
              rl.on("line", (line) => {
                stdoutChunks.push(line);
                onLine(line);
              });

              proc.stderr!.on("data", (chunk: Buffer) => {
                stderrChunks.push(chunk.toString());
              });

              proc.on("error", (err) => reject(err));
              proc.on("close", (code) => {
                resolve({
                  stdout: stdoutChunks.join("\n"),
                  stderr: stderrChunks.join(""),
                  exitCode: code ?? 0,
                });
              });
            });
          }

          return new Promise((resolve, reject) => {
            execFile(
              "sh",
              ["-c", command],
              { cwd: opts?.cwd ?? worktreePath, maxBuffer: 10 * 1024 * 1024 },
              (error, stdout, stderr) => {
                if (error && error.code === undefined) {
                  reject(new Error(`exec failed: ${error.message}`));
                } else {
                  resolve({
                    stdout: stdout.toString(),
                    stderr: stderr.toString(),
                    exitCode: typeof error?.code === "number" ? error.code : 0,
                  });
                }
              },
            );
          });
        },

        copyIn: async (hostPath: string, sandboxPath: string) => {
          const info = await stat(hostPath);
          if (info.isDirectory()) {
            await cp(hostPath, sandboxPath, { recursive: true });
          } else {
            await mkdir(dirname(sandboxPath), { recursive: true });
            await copyFile(hostPath, sandboxPath);
          }
        },

        copyFileOut: async (sandboxPath: string, hostPath: string) => {
          await mkdir(dirname(hostPath), { recursive: true });
          await copyFile(sandboxPath, hostPath);
        },

        close: async () => {
          await rm(root, { recursive: true, force: true });
        },
      };
    },
  });

Branch strategies

A branch strategy controls where the agent's commits land. Configure it when constructing the provider:

When to use each:

• head — fast iteration during development. No branch indirection, no merge step. Only works with bind-mount providers since the agent needs direct host filesystem access.
• merge-to-head — safe default for automation. The agent works on a throwaway branch; if something goes wrong, HEAD is untouched. Use this for CI or unattended runs.
• branch — when you want commits on a specific branch (e.g. for a PR). Pass { type: "branch", branch: "agent/fix-42" }.

Branch strategy is now configured on run(), not on the provider:

import { run, claudeCode } from "@ai-hero/sandcastle";
import { docker } from "@ai-hero/sandcastle/sandboxes/docker";

// head — direct write, bind-mount only (default for bind-mount providers)
await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  prompt: "…",
});
// merge-to-head — temp branch, merge back (default for isolated providers)
await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: tempDir(),
  prompt: "…",
});
// branch — explicit named branch
await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: docker(),
  branchStrategy: { type: "branch", branch: "agent/fix-42" },
  prompt: "…",
});

Passing to run()

Pass your custom provider via the sandbox option — it works the same as the built-in docker() provider:

import { run, claudeCode } from "@ai-hero/sandcastle";

const result = await run({
  agent: claudeCode("claude-opus-4-8"),
  sandbox: localProcess(), // your custom provider
  prompt: "Fix issue #42 in this repo.",
});

Reference implementations

For real-world examples, see:

• src/sandboxes/docker.ts — bind-mount provider using Docker containers (with SELinux label support)
• src/sandboxes/vercel.ts — isolated provider using Vercel Firecracker microVMs via @vercel/sandbox
• src/sandboxes/podman.ts — bind-mount provider using Podman containers (with SELinux label support)
• src/sandboxes/test-isolated.ts — isolated provider using temp directories (used in tests)

Configuration

Config directory (.sandcastle/)

All per-repo sandbox configuration lives in .sandcastle/. Run sandcastle init to create it.

Custom Dockerfile

The .sandcastle/Dockerfile controls the sandbox environment. The default template installs:

• Node.js 22 (base image)
• git, curl, jq (system dependencies)
• GitHub CLI (gh)
• Claude Code CLI
• A non-root agent user (required — Claude runs as this user)

When customizing the Dockerfile, ensure you keep:

• A non-root user (the default agent user) for Claude to run as
• git (required for commits and branch operations)
• gh (required for issue fetching)
• Claude Code CLI installed and on PATH

Add your project-specific dependencies (e.g., language runtimes, build tools) to the Dockerfile as needed.

Hooks

Hooks are grouped by where they run — host (on the developer's machine) or sandbox (inside the container):

hooks: {
  host: {
    onWorktreeReady: [{ command: "cp .env.example .env" }],
    onSandboxReady:  [{ command: "echo sandbox is up" }],
  },
  sandbox: {
    onSandboxReady: [
      { command: "npm install", timeoutMs: 300_000 },
      { command: "apt-get install -y ffmpeg", sudo: true },
    ],
  },
}

Ordering: copyToWorktree -> host.onWorktreeReady (sequential) -> sandbox created -> host.onSandboxReady + sandbox.onSandboxReady (parallel).

• Host hooks accept { command: string; timeoutMs?: number } — no sudo, no cwd. Use cd or inline env in the command string.
• Sandbox hooks accept { command: string; sudo?: boolean; timeoutMs?: number } — set sudo: true for elevated privileges.
• timeoutMs overrides the default 60 s per-hook timeout. Useful for long-running setup commands like dependency installs (e.g. timeoutMs: 300_000 for 5 minutes).
• Within each hook point, sandbox hooks run in parallel; host hooks within onSandboxReady also run in parallel with sandbox hooks. host.onWorktreeReady hooks run sequentially in declared order.
• If any hook exits non-zero, setup fails fast.
• When a signal is passed to run(), it is threaded to all hooks — aborting the signal cancels any in-flight hook commands.

Development

npm install
npm run build    # Bundle with tsup
npm test         # Run tests with vitest
npm run typecheck # Type-check

License

MIT