Sandcastle

Sandcastle (@ai-hero/sandcastle) orchestrates AI coding agents inside isolated sandbox environments. It manages git worktrees, boots containers, runs agents with structured prompts, collects commits, and merges results back to branches — all from a single run() call.

Installation

npm i @ai-hero/sandcastle
npx sandcastle init          # scaffolds .sandcastle/ with Dockerfile and prompt templates
npx sandcastle docker build-image   # builds the Docker image from .sandcastle/Dockerfile

Quick Start

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

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

console.log(result.commits);   // [{ sha: "abc123" }, ...]
console.log(result.branch);    // branch the agent worked on

Core Concepts

The Four Entry Points

Agent Providers

import { claudeCode, codex, opencode, pi } from "@ai-hero/sandcastle";

claudeCode()                           // default: claude-opus-4-7
claudeCode("claude-sonnet-4-6")        // specify model
claudeCode("claude-sonnet-4-6", { effort: "low" })  // with options
codex("o4-mini")

Sandbox Providers

Sandbox providers are subpath imports — not from the main entry point:

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

Two categories:

• Bind-mount (docker, podman): mount the worktree into a container. Default branch strategy: head.
• Isolated (vercel, daytona): sync files into a remote environment. Default branch strategy: merge-to-head.
• noSandbox: host-only, no isolation. Only valid with interactive() — run() and createSandbox() reject it.

Branch Strategies

Pipeline Patterns

Pattern 1: Simple One-Shot

await run({
  agent: claudeCode(),
  sandbox: docker(),
  promptFile: "./.sandcastle/prompt.md",
});

Pattern 2: Iteration Loop

Agent runs multiple iterations, checking for a completion signal between each:

await run({
  name: "worker",
  agent: claudeCode("claude-sonnet-4-6"),
  sandbox: docker(),
  promptFile: "./.sandcastle/prompt.md",
  maxIterations: 5,
  branchStrategy: { type: "merge-to-head" },
  copyToWorktree: ["node_modules"],
  hooks: {
    sandbox: {
      onSandboxReady: [{ command: "npm install" }],
    },
  },
});

The agent sees a completion signal instruction in the prompt. When it outputs <promise>COMPLETE</promise> (default), iteration stops. Otherwise it runs up to maxIterations.

Pattern 3: Sequential Reviewer

Reuse one sandbox for implement → review → fix cycles:

const sandbox = await createSandbox({
  branch: "feature/my-branch",
  sandbox: docker(),
});

await sandbox.run({
  agent: claudeCode("claude-sonnet-4-6"),
  promptFile: "./.sandcastle/implement.md",
});

await sandbox.run({
  agent: claudeCode("claude-opus-4-7"),
  promptFile: "./.sandcastle/review.md",
});

await sandbox.close();

Pattern 4: Parallel Agents

Multiple agents work in parallel on separate worktrees, then a merge agent combines results:

const tasks = issues.map(async (issue) => {
  const wt = await createWorktree({
    branchStrategy: { type: "branch", branch: `fix/${issue.id}` },
  });

  const result = await wt.run({
    agent: claudeCode("claude-sonnet-4-6"),
    sandbox: docker(),
    prompt: `Fix issue: ${issue.title}\n\n${issue.description}`,
  });

  await wt.close();
  return result;
});

const results = await Promise.allSettled(tasks);

Pattern 5: Structured Output

Extract typed data from agent runs:

import { Output } from "@ai-hero/sandcastle";
import { z } from "zod";

const result = await run({
  agent: claudeCode("claude-opus-4-7"),
  sandbox: docker(),
  promptFile: "./.sandcastle/plan.md",
  maxIterations: 1,   // required for structured output
  output: Output.object({
    tag: "plan",
    schema: z.object({
      issues: z.array(z.object({
        title: z.string(),
        priority: z.enum(["high", "medium", "low"]),
      })),
    }),
  }),
});

console.log(result.output.issues);  // fully typed

The prompt must contain the XML tag (e.g., <plan>) — sandcastle validates this at startup. The agent wraps its structured response in that tag. Also available: Output.string({ tag }) for plain string extraction.

Pattern 6: Interactive Then Automated

Explore interactively, then automate implementation:

const wt = await createWorktree({
  branchStrategy: { type: "branch", branch: "feature/new" },
});

await wt.interactive({
  agent: claudeCode("claude-opus-4-7"),
  sandbox: noSandbox(),  // noSandbox only valid with interactive()
});

await wt.run({
  agent: claudeCode("claude-sonnet-4-6"),
  sandbox: docker(),
  promptFile: "./.sandcastle/implement.md",
});

await wt.close();

Prompt System

Inline Prompts

await run({
  agent: claudeCode(),
  sandbox: docker(),
  prompt: "Refactor the auth module to use dependency injection",
});

Inline prompts skip all processing — no {{KEY}} substitution, no shell expressions. Passing promptArgs with an inline prompt is an error.

Prompt Files

await run({
  agent: claudeCode(),
  sandbox: docker(),
  promptFile: "./.sandcastle/prompt.md",
  promptArgs: { ISSUE_TITLE: "Fix login bug", ISSUE_BODY: "Users can't log in" },
});

Prompt files support two features:

1. Template arguments — {{KEY}} placeholders replaced from promptArgs:

Fix the following issue:
Title: {{ISSUE_TITLE}}
Body: {{ISSUE_BODY}}

2. Shell expressions — !`command` evaluated inside the sandbox before each iteration:

The current test output is:
!`npm test 2>&1 | tail -50`

The current git diff is:
!`git diff`

Shell expressions run per-iteration, so each iteration sees fresh state.

Built-in args (auto-injected, cannot be overridden): {{SOURCE_BRANCH}}, {{TARGET_BRANCH}}.

Processing order: Prompt file resolution → Arg substitution (host, once) → Shell expansion (sandbox, per iteration).

Configuration

Docker Options

docker({
  imageName: "my-custom-image",
  containerUid: 1000,
  containerGid: 1000,
  mounts: [
    { hostPath: "~/.npmrc", sandboxPath: "~/.npmrc", readonly: true },
    { hostPath: "./secrets", sandboxPath: "/app/secrets" },
  ],
  env: { NODE_ENV: "development", CI: "true" },
  network: "host",
  selinuxLabel: "z",
})

Hooks

Lifecycle hooks run commands at specific points:

await run({
  agent: claudeCode(),
  sandbox: docker(),
  promptFile: "./.sandcastle/prompt.md",
  hooks: {
    host: {
      onWorktreeReady: [{ command: "cp .env.local .sandcastle/.env" }],
      onSandboxReady: [{ command: "echo 'sandbox is up'" }],
    },
    sandbox: {
      onSandboxReady: [
        { command: "npm install", timeoutMs: 120_000 },
        { command: "npm run build", sudo: false },
      ],
    },
  },
});

Execution order: copyToWorktree → host.onWorktreeReady (sequential) → sandbox created → host.onSandboxReady + sandbox.onSandboxReady (parallel).

Timeouts

Completion Signal

await run({
  completionSignal: "<done>FINISHED</done>",  // custom signal
  // or: completionSignal: ["<done>FINISHED</done>", "<done>SKIPPED</done>"],
  // default: "<promise>COMPLETE</promise>"
});

Cancellation

const controller = new AbortController();
setTimeout(() => controller.abort("timeout"), 300_000);

await run({
  agent: claudeCode(),
  sandbox: docker(),
  prompt: "Fix all lint errors",
  signal: controller.signal,
});

Automatic Cleanup

Both Sandbox and Worktree support Symbol.asyncDispose:

await using sandbox = await createSandbox({
  branch: "feature/x",
  sandbox: docker(),
});
// automatically cleaned up when scope exits

RunResult

Every run() returns:

{
  iterations: IterationResult[];     // per-iteration details
  completionSignal?: string;         // matched signal, or undefined if maxIterations hit
  stdout: string;                    // combined agent output
  commits: { sha: string }[];       // commits made by agent
  branch: string;                    // branch agent worked on
  logFilePath?: string;              // log file path (file logging mode)
  preservedWorktreePath?: string;    // set when worktree had uncommitted changes
  output?: T;                        // present when Output config was provided
}

Session Management

Claude Code sessions can be captured and resumed across runs:

import { hostSessionStore, transferSession } from "@ai-hero/sandcastle";

const result = await run({
  agent: claudeCode("claude-opus-4-7"),
  sandbox: docker(),
  promptFile: "./.sandcastle/prompt.md",
});

// Resume a prior session in a new run
await run({
  agent: claudeCode("claude-opus-4-7"),
  sandbox: docker(),
  promptFile: "./.sandcastle/continue.md",
  resumeSession: "session-id-from-prior-run",
  maxIterations: 1,  // required when resuming
});

CLI Commands

npx sandcastle init                    # scaffold .sandcastle/ directory
npx sandcastle docker build-image      # build Docker image from .sandcastle/Dockerfile
npx sandcastle docker remove-image     # remove the Docker image
npx sandcastle podman build-image      # build with Podman
npx sandcastle podman remove-image     # remove Podman image

sandcastle init offers templates:

• blank — minimal single run() call
• simple-loop — iteration loop picking work items
• sequential-reviewer — implement-then-review via createSandbox()
• parallel-planner — plan/execute/merge with parallel agents
• parallel-planner-with-review — parallel planner plus review phase

Critical Rules

1. Sandbox providers are subpath imports — import { docker } from "@ai-hero/sandcastle/sandboxes/docker", never from the main entry point.
2. noSandbox() only works with interactive() — run() and createSandbox() reject it. AFK work requires real isolation.
3. head strategy is incompatible with isolated providers — throws at runtime. Use merge-to-head or branch instead.
4. copyToWorktree is incompatible with head strategy — head mode bind-mounts the host directory directly, no worktree exists.
5. resumeSession requires maxIterations: 1 — resume applies to iteration 1 only.
6. Structured output requires maxIterations: 1 — and the XML tag must appear in the prompt text. Sandcastle validates both at startup.
7. Inline prompts skip all processing — no {{KEY}} substitution, no !`command` expansion. Passing promptArgs with an inline prompt is an error.
8. Built-in prompt args (SOURCE_BRANCH, TARGET_BRANCH) cannot be overridden — attempting to do so is a runtime error.
9. Docker UID must match image UID — sandcastle runs a pre-flight check and errors on mismatch. Default is host UID or 1000.
10. Worktrees are preserved on error — if the agent leaves uncommitted changes or an error occurs, the worktree is not cleaned up. The path is returned in preservedWorktreePath.

See references/api-types.md for full type signatures and references/prompt-system.md for prompt file details.

Deliverables

When asked to build a sandcastle pipeline:

1. Pipeline script — TypeScript file using sandcastle's API (e.g., pipeline.ts or .sandcastle/run.ts)
2. Prompt files — .sandcastle/*.md prompt templates with {{ARG}} placeholders and !`command` expressions as needed
3. Docker setup — .sandcastle/Dockerfile if not already present (or instruct user to run npx sandcastle init)
4. package.json script — e.g., "pipeline": "npx tsx .sandcastle/run.ts"

Verification

After generating a pipeline:

1. Confirm @ai-hero/sandcastle is in package.json dependencies
2. Confirm sandbox provider imports use subpath syntax
3. Confirm prompt files exist at the referenced paths
4. Confirm any {{ARG}} placeholders in prompts have matching promptArgs
5. If using Docker: confirm .sandcastle/Dockerfile exists or sandcastle init was run
6. If using structured output: confirm maxIterations: 1 and the XML tag appears in the prompt