github-issue-agent

English | 日本語

An HTTP server agent that automatically decomposes GitHub issues and ships implementation PRs. The WebUI lets you trigger the equivalent of --issue 21925 --repo CyberAgentSRG/server from your browser.

Overview

github-issue-agent accepts a GitHub issue as a parent task, automatically breaks it down into multiple child issues (sub-tasks), implements each of them, and finally consolidates the work into a single pull request.

It is built on top of the Anthropic Managed Agents API (@anthropic-ai/sdk); the agents operate on your GitHub repository directly to drive each task to completion.

Quick Start

bun install
export ANTHROPIC_API_KEY=...
export GITHUB_TOKEN=...
bun run start
# → Listening on http://127.0.0.1:3000

Open http://127.0.0.1:3000 in your browser.

Screenshots

Environment Variables

Variable	Default	Description
PORT	3000	listen port
HOST	127.0.0.1	bind host (set to 0.0.0.0 to expose)
DB_PATH	.github-issue-agent/dashboard.db	SQLite db
CONFIG_PATH	(none)	optional config TS path
ANTHROPIC_API_KEY	(required)	Anthropic API key
GITHUB_TOKEN	(required)	GitHub PAT (repo for classic; contents:read, issues:write, pull_requests:write for fine-grained)
LOG_LEVEL	info	log level
LOG_FILE	stderr	log file path

HTTP API

POST /api/runs

{ "issue": 42, "repo": "owner/name", "dryRun": false }

returns { "runId": "uuid" }

GET /api/runs

returns { "runs": [...], "total": N }

GET /api/runs/:runId

returns full run detail

POST /api/runs/:runId/stop

returns { "outcome": "stopped" | "..." }

GET /api/runs/:runId/events

Server-Sent Events stream. Supports Last-Event-ID for resume. Event kinds: phase, session, subIssue, log, complete, error.

WebUI flow

1. Open http://127.0.0.1:3000/runs/new in your browser
2. Enter Issue Number (e.g. 21925) and Repo (e.g. CyberAgentSRG/server)
3. (Optional) Check Dry-run to compute the decomposition plan only
4. Submit → you are redirected to /runs/:runId/live, which shows real-time progress over SSE

Prompt Management

You can view and edit the agents' system prompts from the WebUI. Edited values are persisted to SQLite (.github-issue-agent/dashboard.db) and automatically loaded into the Anthropic-side agent definition on the next run.

Open the Prompts tab in the header to navigate to the prompt list (/prompts) and edit them.

Configuration

Customize behavior by creating a github-issue-agent.config.ts file.

import type { Config } from "./src/shared/config";

const config: Config = {
  models: { parent: "claude-opus-4-7", child: "claude-sonnet-4-6" },
  maxSubIssues: 10,
  maxRunMinutes: 120,
  maxChildMinutes: 30,
  pr: { draft: true, base: "main" },
  commitStyle: "conventional",
  git: {
    authorName: "claude-agent[bot]",
    authorEmail: "claude-agent@users.noreply.github.com",
  },
};

export default config;

Cost

As a rough guideline, expect roughly 0.08 USD per session-hour (based on claude-opus-4-7 session pricing as of 2026-04).

The total cost from issue decomposition to sub-task completion depends on the size of the issue and the number of child tasks generated. Anthropic's pricing changes over time, so check the official docs for the latest rates.

Deployment

See docs/deploy-fly.md for Fly.io deployment steps. The repository ships with Dockerfile, fly.toml, and scripts/start.sh.

E2E Tests

E2E=1 TEST_REPO=<owner>/<repo> TEST_ISSUE=<n> bun run scripts/e2e-real.ts

See docs/e2e-setup.md for details.

Troubleshooting

• Stale lockfile: rm .github-issue-agent/run.lock.lock
• No history in the WebUI: you need to run an issue at least once to populate the DB
• Port conflict: PORT=3097 bun run start