https://minhhh.github.io/tags/ai/Recent content in AI on Huy Minh HaHugo -- gohugo.ioen-usMon, 22 Jun 2026 00:00:00 +0700https://minhhh.github.io/posts/initial-thoughts-on-antigravity-cli/Mon, 22 Jun 2026 00:00:00 +0700https://minhhh.github.io/posts/initial-thoughts-on-antigravity-cli/<p>I recently transitioned to using the new <strong>Antigravity CLI</strong> (<code>agy</code>), which has officially replaced the older <strong>Gemini CLI</strong> as the main terminal interface for agentic coding workflows. Here are some of my initial thoughts on my experience using it.</p> <h2 id="the-good-bits">The Good Bits </h2><ul> <li><strong>Performance</strong>: It&rsquo;s generally faster and much more responsive than the old Gemini CLI. Command execution and responses feel near-instant.</li> <li><strong>Context breakdown</strong>: The <code>/context</code> command displays a clean breakdown of loaded files and symbols, making it easy to track what is currently in the agent&rsquo;s context window.</li> <li><strong>Conversation resets</strong>: The <code>/clear</code> (or <code>/new</code>) command resets the conversation history without requiring you to quit and restart the CLI session.</li> <li><strong>Sub-agents</strong>: Spawning sub-agents to handle background research or run A/B tests (like comparing two different implementations) is extremely easy and doesn&rsquo;t interrupt the main thread.</li> </ul> <h2 id="context-bloat--other-issues">Context Bloat &amp; other issues </h2><ul> <li> <p>In sequential coding workflows, context windows bloat rapidly as tool execution outputs and file reads are appended to history, causing latency and token waste. A big multi-file refactoring session can run out of tokens quickly because of the lack of a compression command. The older Gemini CLI resolved this with the <code>/compress</code> command, which replaced the chat context with a high-level summary of the conversation. Similar developer interfaces (like OpenCode with <code>/compact</code>) implement a matching pattern to keep reasoning loops tight.</p> </li> <li> <p>Another problem is that <code>agy</code> can often mess up the application of moderate-to-large diffs; it sometimes removes more lines than expected, or adds extra lines that result in syntax errors.</p> </li> <li> <p>The request-review mode for tool permission is not &ldquo;smart enough&rdquo;. Many times I have commands that I don&rsquo;t want to apply globally, but at least in a session I might want to have it allowed. Antigravity often suggest 2 options, one is to allow the exact command, and another to add the exact command to settings, but it doesn&rsquo;t have the option to allow the command general pattern, such as &ldquo;cp *&rdquo; in this particular session and that can become annoying as I have to approve the command again and again for slightly different parameters</p> </li> <li> <p>Another issue related to the request-review mode is that the list of choice can be messed up if the choice is <code>python3 run -c ....</code>. Here, the LLM is using a python script as a string literal on the command line, so the choice is a huge block of text and therefore I can only see the last 2 choice and I can&rsquo;t even scroll up to see what the first choices are</p> </li> </ul> <h3 id="workaround-for-context-bloat">Workaround for context bloat </h3><p>Currently, <code>agy</code> supports <code>/rewind</code> (to undo the session to a previous checkpoint) but has no native <code>/compact</code> or <code>/compress</code> command. The current workaround requires manually managing state:</p> <ol> <li>Keeping a tracking file like <code>todo.md</code>.</li> <li>Having the agent update it after each task.</li> <li>Clearing the session history via <code>/new</code>.</li> <li>Reading <code>todo.md</code> in the new session to resume.</li> </ol> <p>This workaround works, but having to manually cycle through session restarts feels unnecessarily clunky. Context hygiene should be a native feature of the harness.</p>