The latest articles on DEV Community by shuo xiao (@shuo_xiao_c7d1a80c798f8ec). https://dev.to/shuo_xiao_c7d1a80c798f8ec https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3959531%2F8b4d15cb-3d39-43ab-8168-047921c56874.jpg https://dev.to/shuo_xiao_c7d1a80c798f8ec en shuo xiao Sat, 30 May 2026 09:29:43 +0000 https://dev.to/shuo_xiao_c7d1a80c798f8ec/coding-agents-keep-losing-context-between-tools-so-i-built-a-local-first-handoff-cli-4h01 https://dev.to/shuo_xiao_c7d1a80c798f8ec/coding-agents-keep-losing-context-between-tools-so-i-built-a-local-first-handoff-cli-4h01 <h2> The problem </h2> <p>I often switch between Codex, OpenCode, Cline, Claude Desktop, scripts, and terminals.</p> <p>The annoying part is not starting a new tool. The annoying part is explaining the same workspace state again:</p> <ul> <li>what changed</li> <li>what is still pending</li> <li>what should not be touched</li> <li>what tests passed</li> <li>what the next agent should read before editing</li> </ul> <h2> What I built </h2> <p>AgentContextBus (acb) is a local-first CLI for handing off workspace context between coding agents.</p> <p>It saves a local handoff packet, then lets the next agent read it through:</p> <ul> <li>paste-ready prompts</li> <li>brief prompts</li> <li>a local dashboard</li> <li>JSON output</li> <li>explicit MCP tools</li> </ul> <h2> First run </h2> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>npx @xiaoshuo1988/acb verify first-run </code></pre> </div> <p>For Chinese output:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>npx @xiaoshuo1988/acb verify first-run <span class="nt">--lang</span> zh-CN </code></pre> </div> <h2> A normal handoff </h2> <p>From the agent that has context:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>acb handoff <span class="nt">--from</span> codex <span class="nt">--summary</span> <span class="s2">"Ready for the next agent"</span> <span class="nt">--git</span> </code></pre> </div> <p>From the receiving side:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>acb receive <span class="nt">--latest</span> </code></pre> </div> <p>After the receiving agent summarizes the packet:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>acb ack <span class="nt">--latest</span> <span class="nt">--by</span> opencode </code></pre> </div> <h2> What ACB intentionally does not do </h2> <ul> <li>no hidden prompt injection</li> <li>no traffic interception</li> <li>no third-party client config mutation</li> <li>no cloud sync</li> <li>no background daemon</li> </ul> <h2> Why local-first </h2> <p>I want the user to be able to inspect the packet store, copy text manually, and decide exactly when context crosses from one agent to another.</p> <h2> What I want feedback on </h2> <ol> <li>Is the handoff packet concept clear?</li> <li>Is <code>verify first-run</code> enough to understand the tool?</li> <li>Is <code>receive --latest</code> the right receiving-side command?</li> <li>Which client path needs the most work?</li> <li>Would you trust this workflow in a real project?</li> </ol> <p>Repo:<br> <a href="https://github.com/xiaoshuo1988130/acb" rel="noopener noreferrer">https://github.com/xiaoshuo1988130/acb</a></p> <p>Feedback discussion:<br> <a href="https://github.com/xiaoshuo1988130/acb/discussions/1" rel="noopener noreferrer">https://github.com/xiaoshuo1988130/acb/discussions/1</a></p> ai opensource cli devtools