🇺🇸English

Runtime Context

Agents run in different environments with different capabilities. This skill detects the current runtime so agents can use the right tools and fail clearly when a required capability is missing.

The Three Runtimes

Detection Script

Run scripts/detect.sh to get a JSON summary of the current environment:

bash skills/runtime-context/scripts/detect.sh

Returns:

{
  "runtime": "sandbox",
  "has_bash": true,
  "has_skill_tool": true,
  "sandbox_id": "sbx_abc123",
  "working_dir": "/vercel/sandbox/bot",
  "node_version": "v22.x.x",
  "bun_version": "1.x.x"
}

For SOUL.md / System Prompt Authors

When writing a system prompt that works across runtimes, structure it like this:

## Environment Awareness

Before taking action, check your environment:
1. If you have `Skill()` and `Agent()` tools, you are in Claude Code
   - Use skills via `Skill(name)`, delegate via `Agent()`
   - Full file access via Read/Write/Edit
2. If you have `bash`, `readFile`, `writeFile` tools, you are in a Sandbox
   - Run scripts via `bash`, read context via `readFile`
   - Skills available via the `skill` tool if configured
3. If you have no tools, you are text-only
   - Answer from your training and system prompt context
   - Tell the user what you would do if you had tools

Do not guess which tools exist. If a tool call fails, state what tool
was missing and what capability it would have provided.

For AI SDK Developers

When building a hosted agent that should leverage skills, use bash-tool:

import {
  experimental_createSkillTool as createSkillTool,
  createBashTool,
} from "bash-tool";

// Skills directory lives alongside your bot code
const { skill, files, instructions } = await createSkillTool({
  skillsDirectory: "./skills",
});

const { tools } = await createBashTool({
  files,
  extraInstructions: instructions,
});

// Now your agent has bash + readFile + writeFile + skill tools
const result = streamText({
  model: gateway("anthropic/claude-sonnet-4.6"),
  tools: { skill, ...tools },
  system: soulPrompt,
  messages,
});

Adapting Behavior

The goal is NOT fallbacks. The goal is clarity.

Good: "I'm in a Sandbox. I have bash and file tools but no Skill() tool. I'll read the skill instructions directly from ./skills/clawnet/SKILL.md."

Bad: Silently trying Skill() first, catching the error, then trying readFile, then trying fetch, then giving up without telling the user.

Pattern: Capability Declaration

At the start of a conversation or task, state what you can do:

I'm running in a Vercel Sandbox with bash, file, and skill tools.
I can: run scripts, read/write files, look up skill instructions.
I cannot: spawn subagents, use Claude Code tools, access the host filesystem.

This sets expectations immediately. Users and orchestrators know what to ask for.

bash-tool: Bridging Claude Code and Sandbox

The bash-tool npm package gives AI SDK agents the same capabilities Claude Code provides natively. It creates bash, readFile, writeFile, and skill tools that work inside Vercel Sandboxes.

Tool Mapping

Skill Declaration Across Runtimes

Skills are declared differently depending on the runtime:

The bot.json skills array is the deployed bot's skill manifest — the sandbox equivalent of agent frontmatter's Skill() declarations.

Integration Pattern

import {
  experimental_createSkillTool as createSkillTool,
  createBashTool,
} from "bash-tool";
import { existsSync } from "node:fs";

let agentTools = {};
let skillInstructions = "";

if (existsSync("./skills")) {
  const { skill, files, instructions } = await createSkillTool({
    skillsDirectory: "./skills",
  });
  const { tools } = await createBashTool({
    files,
    extraInstructions: instructions,
  });
  agentTools = { skill, ...tools };
  skillInstructions = instructions;
}

// Pass to streamText
streamText({
  model: gateway(MODEL_ID),
  system: soulPrompt + skillInstructions,
  tools: agentTools,
  messages,
});

When to Use Each Runtime's Strengths

Weekly Installs

1

Repository

b-open-io/prompts

GitHub Stars

8

First Seen

1 day ago

Security Audits

Gen Agent Trust HubPassSocketFailSnykPass

Installed on

windsurf1

amp1

cline1

opencode1

cursor1

kimi-cli1