Skip to main content

fx

还没翻译译自英文（ai-sdk.dev）本页 37 段都还没译

The fx harness adapter connects HarnessAgent to fx using libfx. The runtime runs on the host, while built-in filesystem and shell tools operate exclusively on the sandbox. No bridge process runs inside the sandbox.

Note

Harness packages are experimental. Expect breaking changes between releases as this early API gets further refined.

Setup

pnpmnpmyarnbun

pnpm add @ai-sdk/harness @ai-sdk/harness-fx @ai-sdk/sandbox-vercel

libfx is installed with the adapter. No fx executable is downloaded or installed inside the sandbox when a session starts.

Import

import { createFx, fx } from '@ai-sdk/harness-fx';

fx is equivalent to createFx() with its default configuration.

Basic Usage

import { HarnessAgent } from '@ai-sdk/harness/agent';
import { fx } from '@ai-sdk/harness-fx';
import { createVercelNetworkSandboxSession } from '@ai-sdk/sandbox-vercel';

const agent = new HarnessAgent({
  harness: fx,
  model: 'openai/gpt-5.6-luna',
});

const sandboxSession = await createVercelNetworkSandboxSession({
  runtime: 'node24',
  template: await agent.getSandboxTemplate(),
});
const session = await agent.createSession({ sandboxSession });

let exitCode = 0;
try {
  const result = await agent.stream({
    session,
    prompt: 'Check the test failures and fix the production code.',
  });

  for await (const part of result.stream) {
    if (part.type === 'text-delta') {
      process.stdout.write(part.text);
    }
  }
} catch (err) {
  exitCode = 1;
  console.error(err);
} finally {
  await session.destroy();
  await sandboxSession.destroy();
  process.exit(exitCode);
}

Adapter Settings

Use createFx() to configure the runtime:

const harness = createFx({
  auth: 'ai-gateway',
  effort: 'low',
  serviceTier: 'default',
});

Settings:

  • auth: selects auto, direct, or ai-gateway authentication, or accepts an authentication environment for programmatically resolved credentials. auto prefers Gateway credentials, direct prefers subscription credentials with Gateway fallback, and ai-gateway uses Gateway only.
  • mcpServers: MCP server definitions keyed by server name. Supports HTTP and SSE servers, with url, optional type, headers, and disabled fields. Connections run on the host; command/stdio servers are unsupported.
  • headers: extra headers sent with AI Gateway requests from the host.
  • effort: reasoning effort for models that support it.
  • serviceTier: selects default, fast, or ultrafast. default is equivalent to omitting the setting and leaves libfx's defaults unchanged. Fast and Ultra require a model that supports the selected tier.
  • backend: libfx runtime backend (auto, native, or wasm). The default is auto, which prefers the native backend. The WASM backend requires JavaScript Promise Integration (JSPI) support.

Authentication

fx supports Vercel AI Gateway authentication and Codex or Grok subscriptions. For Gateway authentication, set one of these environment variables:

  • VERCEL_OIDC_TOKEN
  • AI_GATEWAY_API_KEY

The adapter prefers AI_GATEWAY_API_KEY when both are available. AI_GATEWAY_BASE_URL optionally overrides the Gateway endpoint. Credentials are used on the host, not forwarded into the sandbox.

When Gateway credentials are absent, auto discovers subscription credentials in this order:

  1. fx Codex: ~/.fx/chatgpt-auth.json.
  2. fx Grok: ~/.fx/grok-auth.json.
  3. Native Codex: the credential store under CODEX_HOME or ~/.codex, respecting cli_auth_credentials_store and the native keyring settings.
  4. Native Grok: the xAI OAuth record in GROK_HOME/auth.json or ~/.grok/auth.json.

libfx refreshes subscription tokens near expiry and after an authentication rejection. The adapter saves rotated credentials back to the selected store. Failures after selection do not silently switch subscriptions or fall back to Gateway. Credentials are never included in persisted lifecycle state.

Pass an authentication environment when the host resolves credentials at runtime instead of exposing them through process.env:

const gatewayHarness = createFx({
  auth: {
    AI_GATEWAY_API_KEY: await resolveGatewayToken(),
    AI_GATEWAY_BASE_URL: 'https://ai-gateway.vercel.sh',
  },
});

The supplied record replaces the host environment for authentication discovery. The adapter does not add its credentials to process.env or include their values in persisted lifecycle state.

Use direct to try subscriptions before Gateway, or ai-gateway to disable subscription discovery:

const directHarness = createFx({ auth: 'direct' });
const gatewayHarness = createFx({ auth: 'ai-gateway' });

An explicit authentication environment does not inspect ambient subscription stores. Resumed sessions keep their original provider and require compatible credentials for that provider.

Sandbox

fx runs on the host and accepts either a restricted sandbox session or a network sandbox session. No exposed sandbox ports are required:

const sandboxSession = await createVercelNetworkSandboxSession({
  runtime: 'node24',
  template: await agent.getSandboxTemplate(),
});

The host needs network access for AI Gateway and configured MCP servers. The sandbox does not need network access for model requests. Built-in tools use the sandbox's file and command-execution APIs; the sandbox must provide Bash and GNU command-line utilities such as find, grep, realpath, and base64.

Built-in Tools

The adapter exposes read, write, edit, bash, grep, glob, ls, and askUserQuestions. Filesystem and shell operations run in the session sandbox. The adapter does not expose the fx CLI's other built-in tools.

Built-in tool filtering is supported through activeTools and inactiveTools. With permissionMode: 'allow-reads', file edits and shell commands require approval. With allow-edits, only shell commands require approval. allow-all runs these tools without approval requests.

Session Lifecycle

The adapter stores its libfx checkpoint and unfinished-turn replay records in $HOME/.ai-sdk-harness/.agent-runs/<encoded-sessionId>/fx/state.json. Here, $HOME belongs to the sandbox, and the file is outside sessionWorkDir. A future process restores this state after the caller reattaches the sandbox and supplies the saved lifecycle state.

Known Limitations

  • Manual compaction is unsupported by stable libfx and throws HarnessCapabilityUnsupportedError.

Related

Edit on GitHub
AI 翻译查看英文原文