Build Agents for GitHub Actions

Last updated Jul 2, 2026View as Markdown

Build and run Flue agents in GitHub Actions. This guide walks you through creating your first agent, running it locally with the CLI, and wiring it into a CI workflow.

By the end, you will have a Flue agent running inside GitHub Actions, and you will know how to use local sandbox context, external CLIs, subagents, skills, and typed results to build CI automations.

CI is flue run’s home turf: one agent module, one message, no server, no port. The command executes the agent transport-free and prints the reply to stdout, so a workflow step can pipe it anywhere.

Hello World

A minimal agent that runs in CI whenever an issue is opened.

1. Set up your project

mkdir my-flue-project && cd my-flue-project
npm init -y
npm install @flue/runtime valibot
npm install -D @flue/cli

2. Create your first agent

import { useModel, useSandbox } from '@flue/runtime';
import { local } from '@flue/runtime/node';

export function Hello() {
  useModel('anthropic/claude-sonnet-4-6');
  useSandbox(local());
  return 'Greet the person the user names and share an interesting fact.';
}

A few things to note:

This agent is never mounted over HTTP — it exists to be run from the CLI, which is perfect for CI. (The 'use agent' directive matters only for building a deployable app with Vite; flue run takes the module path directly.)
local() — The local() sandbox runs the agent directly against the host filesystem and shell. In CI, that’s the checked-out repo plus whatever binaries are on $PATH (gh, git, npm, etc.). Skills and AGENTS.md are discovered automatically from the project root. By default only shell-essential env vars (PATH, HOME, locale, etc.) are inherited from process.env — pass local({ env: { GH_TOKEN: process.env.GH_TOKEN } }) to expose more. Use local() only when the runner itself provides the isolation boundary.

3. Test it locally

npx flue run src/agents/hello.ts --message "Say hello to World"

flue run executes the agent module in-process — no HTTP listener, no build — streams progress to stderr, and prints the final reply to stdout. Pass --json for a machine-readable envelope instead.

4. Wire it into GitHub Actions

.github/workflows/hello.yml:

name: Hello Flue

on:
  issues:
    types: [opened]

jobs:
  hello:
    runs-on: ubuntu-latest
    permissions:
      issues: read
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 22
      - run: npm ci
      - name: Run agent
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          npx flue run src/agents/hello.ts \
            --message 'Say hello to ${{ github.event.issue.user.login }}'

Add ANTHROPIC_API_KEY as a repository secret (Settings > Secrets and variables > Actions). Open an issue and you’ll see the agent’s greeting in the job logs.

Building a real agent

Now let’s build something useful — an issue triage agent that analyzes an issue and reports back. This is where Flue’s agent features start to shine.

Structured work with skills and actions

An agent’s deterministic orchestration lives in harness tools — finite, schema-validated jobs the model calls — and skills, reusable instruction files. Inside either, the harness gives you two core methods:

harness.sandbox.exec(cmd) — Run a shell command in the sandbox. Returns { stdout, stderr, exitCode }.
harness.prompt(text, opts) — Send a prompt to the agent and get back a result. harness.prompt(...) runs in the same conversation context as the agent’s own turns, so naming a mounted skill in the instruction is enough to direct the model to apply it.

prompt() accepts a result option — a Valibot schema that defines the expected output shape. Flue parses the agent’s response and returns it on response.data, fully typed:

import * as v from 'valibot';

// summary: string
const { data: summary } = await harness.prompt(`Summarize this diff:\n${diff}`, {
  result: v.string(),
});

// diagnosis: { reproducible: boolean, skipped: boolean }
const { data: diagnosis } = await harness.prompt(
  `Apply the triage skill to issue #${issueNumber}:\n\n${issue}`,
  {
    result: v.object({
      reproducible: v.boolean(),
      skipped: v.boolean(),
    }),
  },
);

Connecting external CLIs

Your agent often needs to interact with tools like gh, npm, or git. With local(), the agent’s bash tool runs against the host shell directly — anything on $PATH is reachable. Host env vars are opt-in: only shell essentials (PATH, HOME, locale, etc.) are inherited by default, so you pass the specific vars your CLIs need via local({ env: { ... } }).

In GitHub Actions, this means you set the secrets you want the agent’s CLIs to see in the workflow env: block, then forward them explicitly into the sandbox. The runner is your isolation boundary; flue makes the inner boundary (host → spawned shell) explicit.

import { useModel, useSandbox } from '@flue/runtime';
import { local } from '@flue/runtime/node';

export function Triage() {
  useModel('anthropic/claude-opus-4-7');
  useSandbox(
    local({
      env: {
        GH_TOKEN: process.env.GH_TOKEN,
        NPM_TOKEN: process.env.NPM_TOKEN,
      },
    }),
  );
  return 'When given an issue number, run the `triage` skill on it and report severity, reproducibility, and a summary.';
}

If you want a tighter boundary — the agent can call a specific operation but never see the underlying token — call useTool(...) inside the agent function. The tool implementation reads the secret from process.env; the agent only sees the tool’s parameters and result.

Subagents

Named subagents can run focused detached tasks. useSubagent(...) declares one, backed by its own agent function:

'use agent';
import { useModel, useSubagent } from '@flue/runtime';

function Reviewer() {
  return 'Focus on correctness, security, and project standards.';
}

export function Triage() {
  useSubagent({
    name: 'reviewer',
    description: 'Reviews a pull request for correctness, security, and project standards.',
    agent: Reviewer,
  });
  return 'Delegate PR reviews to the `reviewer` subagent via a task.';
}

Sandbox context

The agent reads AGENTS.md and skills from its sandbox at runtime. CI agents typically use local(), which gives direct access to the runner’s checkout — so any files in your repo are visible automatically.

Skills are reusable agent tasks defined as markdown files in .agents/skills/. They give the agent a focused instruction set for a specific job:

.agents/skills/triage/SKILL.md:

---
name: triage
description: Triage a GitHub issue — reproduce, assess severity, and optionally fix.
---

Given the issue number in the arguments:

1. Use `gh issue view` to fetch the issue details
2. Read the codebase to understand the relevant area
3. Attempt to reproduce the issue
4. Assess severity and write a summary
5. If the fix is straightforward, apply it and open a PR

AGENTS.md at your project root is the agent’s system prompt — it provides global context about the project:

You are a helpful assistant working on the my-project codebase.

## Project structure

- `src/` — Application source code
- `tests/` — Test suite

## Guidelines

- Always run tests before suggesting a fix is complete
- Use the project's existing patterns and conventions

Wiring it into GitHub Actions

.github/workflows/issue-triage.yml:

name: Issue Triage

on:
  issues:
    types: [opened]

jobs:
  triage:
    runs-on: ubuntu-latest
    timeout-minutes: 30
    permissions:
      contents: read
      issues: write
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 22
      - run: npm ci
      - name: Run triage agent
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          npx flue run src/agents/triage.ts \
            --message 'Triage issue #${{ github.event.issue.number }}'

GITHUB_TOKEN is provided automatically by GitHub Actions.

Typed results and orchestration

Result schemas aren’t just for type safety — they’re how you orchestrate multi-step work. Wrap the orchestration in a harness-connected tool (useTool({ harness: true })): you get typed data back from prompt() and can branch on it in plain code, all inside one durable conversation:

'use agent';
import { useModel, useSandbox, useTool } from '@flue/runtime';
import { local } from '@flue/runtime/node';
import * as v from 'valibot';

export function AutoTriage() {
  useModel('anthropic/claude-sonnet-4-6');
  useSandbox(local());
  useTool({
    name: 'triage-issue',
    description: 'Triage one GitHub issue and auto-fix critical reproducible ones.',
    input: v.object({ issueNumber: v.number() }),
    harness: true,
    async run({ harness, data }) {
      const { data: triage } = await harness.prompt(
        `Apply the triage skill to issue #${data.issueNumber}.`,
        {
          result: v.object({
            severity: v.picklist(['low', 'medium', 'high', 'critical']),
            reproducible: v.boolean(),
            summary: v.string(),
          }),
        },
      );

      if (triage.severity === 'critical' && triage.reproducible) {
        await harness.prompt(`Apply the auto-fix skill to issue #${data.issueNumber}.`, {
          result: v.object({ fix_applied: v.boolean(), pr_url: v.optional(v.string()) }),
        });
      }
      return triage;
    },
  });
  return 'When given an issue number, call the `triage-issue` tool and report its result.';
}

This pattern — prompt, check the result, decide what to do next — is how you build sophisticated agents that go beyond single-shot prompts.

Running agents locally

During development, flue run executes the same module the CI step will run:

# One-shot run
npx flue run src/agents/auto-triage.ts --message "Triage issue #42"

# Machine-readable output for scripting
npx flue run src/agents/auto-triage.ts --message "Triage issue #42" --json | jq -r '.message'

# Continue the same conversation across invocations
npx flue run src/agents/auto-triage.ts --message "What did you conclude?" --id issue-42

Progress streams to stderr; only the final reply (or the --json envelope) lands on stdout. See flue run for the full contract.