The latest articles on DEV Community by Hasegawa Takuya (@rluisr). https://dev.to/rluisr https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4043932%2Feec9a41e-279d-4724-8c7d-64a5bd35b027.png https://dev.to/rluisr en Hasegawa Takuya Thu, 23 Jul 2026 13:31:54 +0000 https://dev.to/rluisr/issue-in-pr-out-maestro-orchestrates-claude-agents-for-github-linear-2l1n https://dev.to/rluisr/issue-in-pr-out-maestro-orchestrates-claude-agents-for-github-linear-2l1n <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2pi6zt4o3bo8f1wmi6yk.jpeg" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2pi6zt4o3bo8f1wmi6yk.jpeg" alt="maestro — Decompose GitHub issues. Ship pull requests. Autonomously." width="800" height="447"></a></p> <p>Coding agents are great at <strong>writing code</strong>.</p> <p>They're much weaker at <strong>owning a ticket end-to-end</strong>: scoping the work, carving it into the right units, implementing in parallel, opening one clean PR, and leaving a trail you can audit later.</p> <p>That gap costs real time. Teams paste issues into a chat box, babysit with "please continue," and still get sprawling diffs that are painful to review.</p> <p><strong><a href="https://github.com/ca-srg/maestro" rel="noopener noreferrer">maestro</a></strong> fills that gap.</p> <blockquote> <p><strong>Drop in a GitHub or Linear issue. Get a reviewable pull request.</strong></p> </blockquote> <p>It's an open-source multi-agent system on <a href="https://docs.anthropic.com/" rel="noopener noreferrer">Claude Managed Agents</a> that turns issues into shipped code — with orchestration, not another chatbot loop.</p> <h2> Why bother? </h2> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>You</th> <th>maestro</th> </tr> </thead> <tbody> <tr> <td>Open a <strong>GitHub</strong> or <strong>Linear</strong> issue</td> <td>Treats it as the parent task</td> </tr> <tr> <td>Glance at the WebUI</td> <td> <strong>Optimally decomposes</strong> into focused sub-issues</td> </tr> <tr> <td>Review <strong>one</strong> PR</td> <td>Implementers code; work lands as a single PR</td> </tr> <tr> <td>Ask "what did the agent actually do?"</td> <td> <strong>Audit logs + run events</strong> keep every phase</td> </tr> </tbody> </table></div> <p>No chat babysitting. No "please continue."<br> <strong>Issue in → PR out</strong> — plus a receipt for how you got there.</p> <h3> What breaks with "just throw it at an agent" </h3> <ol> <li> <strong>Chat doesn't scale for tickets</strong> — threads drift, context dies, ownership vanishes.</li> <li> <strong>One giant loop on a fat issue</strong> yields messy diffs and fuzzy review boundaries.</li> <li> <strong>Trust needs receipts</strong> — which tools ran, what changed, and why.</li> </ol> <p>maestro is built around those three failure modes.</p> <h2> Screenshots </h2> <h3> Kick off a run </h3> <p>Pick a repo + issue. Hit run. No thread to babysit.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2vx94gm6k67huqzo9xca.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2vx94gm6k67huqzo9xca.png" alt="Start a new run" width="800" height="663"></a></p> <h3> Follow the multi-agent pipeline </h3> <p>Orchestrator splits work; implementers ship in parallel; everything consolidates into one PR.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ff4iz94186tstmtvxe5ho.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ff4iz94186tstmtvxe5ho.png" alt="Runs list" width="800" height="893"></a></p> <h3> Connected repositories </h3> <p>GitHub App auth per installation — skip the shared-PAT mess.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwokwz1vi8vnbse6flnea.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwokwz1vi8vnbse6flnea.png" alt="Repositories" width="800" height="500"></a></p> <h3> MCP client settings </h3> <p>GitHub MCP is built in. Add Figma, Linear, or your own stdio servers through the MCP Gateway.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fa3051f0lddakv5hxheuv.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fa3051f0lddakv5hxheuv.png" alt="MCP client settings" width="800" height="550"></a></p> <h3> Prompt versioning </h3> <p>Edit system prompts from the UI without redeploying — and keep a diff of what changed.</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpprifwbq6rc2gqd1xg4j.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpprifwbq6rc2gqd1xg4j.png" alt="Prompt version diff" width="800" height="550"></a></p> <h2> Pipeline </h2> <ol> <li> <strong>Ingest</strong> from <strong>GitHub Issues</strong> or <strong>Linear</strong> (identifier or URL).</li> <li> <strong>Orchestrator</strong> plans and <strong>decomposes</strong> into the smallest set of independent, implementable sub-issues — boundaries for parallel work and clean review, not arbitrary chunks.</li> <li> <strong>Implementers</strong> (Claude Managed Agents) take each sub-issue and write real code via <strong>GitHub MCP</strong>.</li> <li> <strong>Finalize</strong> folds the work into <strong>one pull request</strong> for human review.</li> <li> <strong>Audit</strong> — structured logs and per-run event history (phases, tools, outcomes). OpenTelemetry optional.</li> </ol> <p>Start from the <strong>WebUI</strong>, <strong>HTTP API</strong>, or a <strong>labeled-issue poller</strong>. Linear origins use the same path.</p> <h2> What you get </h2> <ul> <li>GitHub Issues <strong>and</strong> Linear on one orchestrate → implement → PR path</li> <li>Optimal task decomposition instead of a mega-agent loop</li> <li>Durable audit / run history in the WebUI</li> <li>Multi-agent by design (coordinator + implementers)</li> <li>Real PRs — branches, commits, reviewable diffs on your GitHub</li> <li>GitHub App auth (installation tokens, not PAT sprawl)</li> <li>MCP-native (GitHub + gateway for Figma / Linear / custom stdio)</li> <li>Self-hosted sandboxes via <code>ANTHROPIC_ENVIRONMENT_ID</code> </li> <li>OpenTelemetry when you want it</li> </ul> <h2> Stack </h2> <ul> <li> <strong>Runtime:</strong> Bun + TypeScript</li> <li> <strong>Agents:</strong> Anthropic Claude Managed Agents</li> <li> <strong>HTTP / UI:</strong> Hono SSR dashboard</li> <li> <strong>GitHub:</strong> GitHub App + official GitHub MCP</li> <li> <strong>Extensibility:</strong> MCP Gateway (stdio → remote MCP)</li> </ul> <h2> Quick start </h2> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>mise <span class="nb">install </span>bun <span class="nb">install export </span><span class="nv">ANTHROPIC_API_KEY</span><span class="o">=</span>... <span class="nb">export </span><span class="nv">GITHUB_APP_ID</span><span class="o">=</span>... <span class="nb">export </span><span class="nv">GITHUB_APP_PRIVATE_KEY_PATH</span><span class="o">=</span>/path/to/github-app.pem bun run start <span class="c"># → http://127.0.0.1:3000</span> </code></pre> </div> <p>Open the dashboard, pick a repo + issue, hit run.</p> <p>Docs: <a href="https://github.com/ca-srg/maestro/wiki/Installation" rel="noopener noreferrer">Installation</a> · <a href="https://github.com/ca-srg/maestro/wiki/Architecture" rel="noopener noreferrer">Architecture</a> · <a href="https://github.com/ca-srg/maestro" rel="noopener noreferrer">Repo</a></p> <h2> Who this is for </h2> <ul> <li>Teams with backlog tickets that are agentable but not chat-friendly</li> <li>Orgs that want <strong>Issue → PR</strong> with humans still in the review loop</li> <li>Anyone who needs <strong>auditability</strong> for autonomous coding agents</li> <li>Builders on Claude / MCP who want orchestration, not another chat UI</li> </ul> <h2> TL;DR </h2> <p>Stop pasting tickets into a chat box.</p> <p>Point <strong>maestro</strong> at a GitHub or Linear issue. Get an optimally decomposed plan, a real PR, and an audit trail.</p> <p>🔗 <strong><a href="https://github.com/ca-srg/maestro" rel="noopener noreferrer">https://github.com/ca-srg/maestro</a></strong></p> opensource ai github devops