Flue

Help users install, configure, and work with Flue workflows. Flue connects a full OpenCode session to AI agents and CI workflows with secure, autonomous execution.

Routing

Determine what the user needs and load the appropriate reference:

• Install Flue / get started — Read ${CLAUDE_SKILL_ROOT}/references/install.md
• Add sandbox isolation — Read ${CLAUDE_SKILL_ROOT}/references/add-sandbox.md
• Create agent skills for workflows — See https://agentskills.io/specification
• Write or modify a workflow — Use the API reference below
• See example patterns — Read ${CLAUDE_SKILL_ROOT}/references/patterns.md

FlueClient API

Workflows are TypeScript files in .flue/workflows/ that export a default async function receiving a FlueClient. The client provides three core methods:

flue.shell(command, options?)

Run shell commands in the target environment. Returns { stdout, stderr, exitCode }.

const result = await flue.shell('gh issue view 123 --json title,body');
const issue = JSON.parse(result.stdout);

// With stdin
await flue.shell('gh issue comment 123 --body-file -', { stdin: commentBody });

Options: env, stdin, cwd, timeout

flue.prompt(text, options?)

Send a one-off prompt to an LLM. Use for quick tasks like summarization, classification, or text generation.

// Without result schema (fire-and-forget)
await flue.prompt('Refactor the tests in src/utils/ to use vitest');

// With typed result via Valibot schema
const summary = await flue.prompt('Categorize this issue: ' + issue.title, {
  result: v.picklist(['bug', 'feature', 'question']),
});

Options: result (Valibot schema), model, timeout

flue.skill(path, options?)

Delegate a complex task to an autonomous agent using a skill file from .agents/skills/. The agent reads the skill instructions and works autonomously with full tool access.

// Fire-and-forget
await flue.skill('triage/reproduce.md', {
  args: { issueNumber: 123 },
});

// With typed result validation
const diagnosis = await flue.skill('triage/diagnose.md', {
  args: { issueNumber: 123 },
  result: v.object({
    confidence: v.picklist(['high', 'medium', 'low']),
    reproducible: v.boolean(),
  }),
});

Options: args, result (Valibot schema), model, timeout

Workflow File Structure

import type { FlueClient } from '@flue/client';
import * as v from 'valibot';

// Optional: typed input args (validated by CLI via --args)
export const args = v.object({
  issueNumber: v.number(),
});

// Optional: proxy declarations for sandbox mode
export const proxies = {
  anthropic: anthropic(),
  github: github({ policy: 'allow-read' }),
};

// Required: default export function
export default async function myWorkflow(
  flue: FlueClient,
  { issueNumber }: v.InferOutput<typeof args>,
) {
  // ...
}

CLI

# Run a workflow
npx flue run .flue/workflows/hello.ts
npx flue run .flue/workflows/triage.ts --args '{"issueNumber": 123}'
npx flue run .flue/workflows/triage.ts --model anthropic/claude-sonnet-4-20250514
npx flue run .flue/workflows/triage.ts --sandbox ghcr.io/org/repo/flue-sandbox:latest

# Install OpenCode CLI (for CI environments)
npx flue install

Reference Files

• Installation & getting started: ${CLAUDE_SKILL_ROOT}/references/install.md
• Adding sandbox isolation: ${CLAUDE_SKILL_ROOT}/references/add-sandbox.md
• Common workflow patterns: ${CLAUDE_SKILL_ROOT}/references/patterns.md
• Example workflow file: ${CLAUDE_SKILL_ROOT}/assets/hello-flue.ts
• Example GitHub Actions workflow: ${CLAUDE_SKILL_ROOT}/assets/hello-flue.yml