<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ramu Narasinga on Medium]]></title>
        <description><![CDATA[Stories by Ramu Narasinga on Medium]]></description>
        <link>https://medium.com/@ramunarasinga?source=rss-a0a9d4ff1c65------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Ez0wHw_TJH_NLtUueLaLCQ.jpeg</url>
            <title>Stories by Ramu Narasinga on Medium</title>
            <link>https://medium.com/@ramunarasinga?source=rss-a0a9d4ff1c65------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 16:18:21 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ramunarasinga/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[`.apparmor` type file in OpenMausBot codebase.]]></title>
            <link>https://medium.com/@ramunarasinga/apparmor-type-file-in-openmausbot-codebase-82d1334f1f1d?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/82d1334f1f1d</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[mausbot]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[grok]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 18:11:17 GMT</pubDate>
            <atom:updated>2026-11-10T18:11:17.674Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review .apparmor type file in OpenMasuBot codebase. You will learn:</p><ol><li>What is OpenMausBot?</li><li>What is AppArmor?</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-Y3x5cQhq90bzvIxtxZSdA.png" /></figure><h3>What is OpenMausBot?</h3><p>MausBot turns agent CLIs into a familiar desktop chat app. Each bot can have its own model, instructions, working folder, connected apps, computer, and recurring work.</p><p>You can build a team of AI agents that can talk, use tools, operate computers, and keep working on a schedule.</p><p>MausBot is an open-source Grok Bot alternative with a virtual machine that bots can use.</p><blockquote>Learn more about <a href="https://mausbot.com/">MausBot</a>.</blockquote><h3>What is AppArmor?</h3><p>I found a file named <a href="https://github.com/milind-soni/OpenMausBot/blob/main/connector/profiles/omb-home-narrow.apparmor">omb-home-narrow.apparmor</a> in the MausBot codebase. I never came across such a file type ‘.apparmor’, so I googled what it means and found this <a href="https://apparmor.net/">apparmor.net/</a></p><p>AppArmor is an effective and easy-to-use Linux application security system. AppArmor proactively protects the operating system and applications from external or internal threats, even zero-day attacks, by enforcing good behavior and preventing both known and unknown application flaws from being exploited.</p><p>well, I also wondered if omb-home-narrow file name had any special meaning, but it did not. Below is a comment I picked from that file:</p><pre># omb-home-narrow: omb-home, plus a refusal of every filesystem type a user <br># namespace may mount that bubblewrap<br># never does (it mounts only tmpfs, proc and devpts, and binds). Overlay and <br># FUSE in user namespaces are where most<br># unprivileged-namespace kernel bugs have been; a bot cannot reach them <br># through this container.</pre><p>and this file also deny commands, preventing access to senstive files on your machine.</p><pre>deny mount fstype=overlay,<br>  deny mount fstype=fuse,<br>  deny mount fstype=fuseblk,<br>  deny mount fstype=sysfs,<br>  deny mount fstype=mqueue,<br>  deny mount fstype=cgroup,<br>  deny mount fstype=cgroup2,<br>  deny mount fstype=ramfs,<br>  deny mount fstype=binfmt_misc,<br>  deny mount fstype=bpf,</pre><p>I found this snippet that could help us understand more about this file</p><pre>profile omb-home-narrow flags=(attach_disconnected,mediate_deleted) {</pre><p><a href="https://apparmor.net/profiles/profile-types-and-syntax/">AppArmor profiles</a> are the fundamental building blocks of system security. They act as a set of rules that dictate exactly what a confined application is allowed to do.</p><blockquote>Learn more about <a href="https://apparmor.net/profiles/profile-types-and-syntax/#attached-profiles">AppArmor profiles</a>.</blockquote><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. I spent 4+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade.</p><blockquote>Need help with a project? drop me an email — ramu.narasinga@gmail.com or <a href="https://thinkthroo.com/contact">book a meeting</a>.</blockquote><h3>References:</h3><ol><li><a href="https://apparmor.net/">apparmor.net</a></li><li><a href="https://github.com/milind-soni/OpenMausBot/blob/main/connector/profiles/omb-home-narrow.apparmor">github.com/milind-soni/OpenMausBot/connector/profiles/omb-home-narrow.apparmor</a></li><li><a href="https://mausbot.com/">mausbot.com/</a></li><li><a href="https://apparmor.net/profiles/profile-types-and-syntax/">apparmor.net/profiles/profile-types-and-syntax/</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82d1334f1f1d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[REA — Reverse Engineer Anything.]]></title>
            <link>https://medium.com/@ramunarasinga/rea-reverse-engineer-anything-c25cbbf447d7?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/c25cbbf447d7</guid>
            <category><![CDATA[research]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[reverse-engineering]]></category>
            <category><![CDATA[ai-coding-agent]]></category>
            <category><![CDATA[ears]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 18:00:10 GMT</pubDate>
            <atom:updated>2026-10-26T18:00:10.356Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review the tool, REA. You will learn:</p><ol><li>What is REA?</li><li><a href="https://github.com/morluto/rea/blob/main/.agents/skills/reverse-engineer-anything/SKILL.md">reverse-engineer-anything/SKILL.md</a></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RefZhovMoW1ojrudc8WCNA.png" /></figure><h3>What is REA?</h3><p>Ask your coding agent about an app, a game or a website. REA gives it the tools to inspect the code behind the answer.</p><p>You can find out how software works.</p><blockquote>Learn more about <a href="https://rea.tools/">REA</a>.</blockquote><h4>Set up REA.</h4><p>Copy this into your coding agent:</p><pre>Install REA and connect it to this coding agent using <br>npx rea-agents@latest setup. <br>Show me the setup plan for approval, then verify the installation.</pre><p>or you can run the below command in your terminal:</p><pre>npx rea-agents@latest setup</pre><h4>Showcases:</h4><p>I found the below examples at <a href="http://rea.tools/showcase">rea.tools/showcase</a></p><ol><li><a href="https://rea.tools/showcase/aegis/">How an authenticator generates a login code</a></li><li><a href="https://rea.tools/showcase/ctf/">Solve a CTF flag checker</a></li><li><a href="https://rea.tools/showcase/notion/">How Notion copies text and HTML</a></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bbFD9TaccFqroxvSAGf17g.png" /></figure><p>I took this image screenshot from Notion’s showcase.</p><h3>reverse-engineer-anything/SKILL.md</h3><p>I found this <a href="https://github.com/morluto/rea/blob/main/.agents/skills/reverse-engineer-anything/SKILL.md">SKILL in the REA codebase</a>. Below is the metadata of this skill.</p><pre>---<br>name: reverse-engineer-anything<br>description: Reverse engineer native, managed, Electron/JavaScript, packaged, <br>firmware, and browser targets with REA, and build agent environments <br>(AgentEnv bundles) of web applications from a live instance or a capture. <br>Use shipped-artifact or requested runtime evidence to explain features, <br>compare versions, decompile code, or guide a reconstruction. <br>Skip REA for ordinary source-repository architecture analysis.<br>metadata:<br>  version: &quot;49&quot;<br>---</pre><p>Below is the table of contents of this skill.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/694/1*093sy7e52gMfBj-VFIuH7Q.png" /></figure><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 4+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://github.com/morluto/rea">github.com/morluto/rea</a></li><li><a href="https://rea.tools/get-started/">rea.tools/get-started/</a></li><li><a href="https://rea.tools/">rea.tools/</a></li><li><a href="https://github.com/morluto/rea/blob/main/.agents/skills/reverse-engineer-anything/SKILL.md">github.com/morluto/rea/.agents/skills/reverse-engineer-anything/SKILL.md</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c25cbbf447d7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ponytail, the lazy senior dev for your AI agent.]]></title>
            <link>https://medium.com/@ramunarasinga/ponytail-the-lazy-senior-dev-for-your-ai-agent-9f8c465f0696?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f8c465f0696</guid>
            <category><![CDATA[prompt]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[skills]]></category>
            <category><![CDATA[ponytail]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 18:41:58 GMT</pubDate>
            <atom:updated>2026-10-21T18:41:58.116Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review Ponytail. You will learn:</p><ol><li>What is Ponytail?</li><li>The Prompt</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hOQCn0heyFOTwHd9eI7weQ.png" /></figure><h3>What is Ponytail?</h3><p>Ponytail makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.</p><p><a href="https://github.com/DietrichGebert/ponytail#already-built-with-ponytail">Retriever</a> is an app, built using the Ponytail.</p><p>I found this below benchmark in <a href="https://github.com/DietrichGebert/ponytail#numbers">Ponytail’s readme</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*55dIjTW-kMkMTQd7i4axOQ.png" /></figure><p>I mean, at this point, I was wondering how would Ponytail achieve these results? is there any framework/library? any skills?. If you take a closer look at <a href="https://ponytail.dev/#how">How it works section on Ponytail site</a>, it does not reveal much about its internals.</p><p>Github Readme is the right place to dig deeper usually, and that’s how I found the prompt.</p><blockquote>Learn more about <a href="https://ponytail.dev/">Ponytail.dev</a></blockquote><h3>The Prompt</h3><p>Ponytail is one prompt: <a href="https://github.com/DietrichGebert/ponytail/blob/main/skills/ponytail/SKILL.md">skills/ponytail/SKILL.md</a>. The compact version, for agents that read a rules file, is <a href="https://github.com/DietrichGebert/ponytail/blob/main/AGENTS.md">AGENTS.md</a>. Everything else in this repo loads that prompt into different agents.</p><p>Below is the <a href="https://raw.githubusercontent.com/DietrichGebert/ponytail/refs/heads/main/skills/ponytail/SKILL.md">entire ponytail SKILL.md</a>.</p><pre>---<br>name: ponytail<br>description: &gt;<br>  Lazy senior dev mode: the smallest change that fully solves the task, and a<br>  reply a busy human understands in one read. Use on any coding task (writing,<br>  fixing, refactoring, reviewing, choosing dependencies) and when the user says<br>  &quot;ponytail&quot;, &quot;be lazy&quot;, &quot;simplest solution&quot;, &quot;yagni&quot;, or complains about<br>  over-engineering or bloat. Levels: lite, full (default), ultra.<br>argument-hint: &quot;[lite|full|ultra|off]&quot;<br>license: MIT<br>---<br><br># Ponytail<br><br>You are a lazy senior developer. The best code is the code never written. You solve the whole problem with the least new code. End your reply with one or two lines: what you skipped or did not check, and any risk the user must know.<br><br>Active for the whole session until the user says &quot;stop ponytail&quot; or &quot;normal mode&quot;. Switch level: `/ponytail lite|full|ultra`.<br><br>## Before you write<br><br>Read the task and the code it touches. List every place your change must reach: callers, tests, fixtures, config, exports. Check what your change could break for users: data it would destroy or expose, callers that stop working. That is scope. Extra features are not.<br><br>## The smallest complete change<br><br>Take the first option that fully works:<br><br>1. Does it need to exist? Skip features, options and flexibility nobody asked for, and name them in one line. A vague request (&quot;build me X&quot;) gets the smallest version that does the core job.<br>2. Already in this codebase (a helper, component, service, pattern)? Use it the way the surrounding code does.<br>3. Standard library or a platform feature? Use it, unless the project has its own. A house component beats a native widget.<br>4. An installed dependency? Use it. Never add a dependency for a few lines.<br>5. Can it be one line a reader gets at a glance? One line.<br>6. Otherwise: the minimum code that works.<br><br>- Be lazy about the solution, never about the change itself: finish every part the task needs, including the callers, tests and fixtures your change breaks.<br>- No abstraction, wrapper, type conversion, option, config, boilerplate or &quot;for later&quot; code nobody asked for. Keep values in the form the platform already gives you. Deletion beats addition. Keep the structure the codebase already has: its layers, interfaces and conventions.<br>- The shortest working diff wins, once you know everything it must touch. A one-liner that needs decoding is not short.<br>- Comment only the why the code cannot show, in one line.<br>- Bug fix: before you edit, grep every caller of the function you touch, then fix the root cause once in the shared code.<br>- Code you move or merge keeps its error handling and validation.<br>- Between options of equal size, take the one that is correct on edge cases.<br>- Lazy code without its check is unfinished: new non-trivial logic (a branch, a loop, a parser, money or security, or a whole new script or app) leaves one small test or an assert-based self-check. Trivial changes need none.<br>- A shortcut with a known limit gets a code comment in this form: `shortcut: &lt;the limit&gt;, &lt;when to upgrade&gt;`.<br><br>Never cut: validation at trust boundaries, error handling that prevents data loss, security, accessibility, the calibration real hardware needs, anything the user asked for.<br><br>## Levels<br><br>| Level | Behavior |<br>|-------|----------|<br>| **lite** | Build what was asked. Name the smaller option in one line and let the user pick. |<br>| **full** | The rules above. Default. |<br>| **ultra** | Also question the request: before building, push back on any part the need does not justify. |</pre><p>There’s also examples in the Ponytail repo, explaining how the Ponytail works/helps in reducing the lines of code:</p><ol><li><a href="https://github.com/DietrichGebert/ponytail/blob/main/examples/email-validation.md">examples/email-validation.md</a></li><li><a href="https://github.com/DietrichGebert/ponytail/blob/main/examples/react-countdown.md">examples/react-countdown.md</a></li></ol><p>and <a href="https://github.com/DietrichGebert/ponytail/tree/main/examples">much more</a>.</p><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 4+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://github.com/DietrichGebert/ponytail">github.com/DietrichGebert/ponytail</a></li><li><a href="https://ponytail.dev/#install">ponytail.dev</a></li><li><a href="https://github.com/DietrichGebert/ponytail/blob/main/skills/ponytail/SKILL.md">github.com/DietrichGebert/ponytail/skills/ponytail/SKILL.md</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f8c465f0696" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fallow, a codebase intelligence for TypeScript and JavaScript.]]></title>
            <link>https://medium.com/@ramunarasinga/fallow-a-codebase-intelligence-for-typescript-and-javascript-fde9711d0261?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/fde9711d0261</guid>
            <category><![CDATA[hyperframes]]></category>
            <category><![CDATA[fallow]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[codebase]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 18:41:01 GMT</pubDate>
            <atom:updated>2026-10-12T18:41:01.674Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review Fallow. You will learn:</p><ol><li>What is Fallow?</li><li>.fallowrc.jsonc in hyperframes.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1yAA_HW0d8reFDe-VNL59A.png" /></figure><h3>What is Fallow?</h3><p><a href="https://fallow.tools/docs/">Fallow</a> is codebase intelligence for TypeScript and JavaScript. One command reads your whole repository as one graph and shows:</p><ul><li>where the code is hard to change (complexity hotspots and a health score)</li><li>where the architecture drifts (boundary rules and circular dependencies)</li><li>what is copied (duplication)</li><li>where the UI drifts from your design system (styling drift)</li><li>what nothing uses (unused files, exports, and dependencies)</li><li>where the feature flags are</li></ul><h4>Quickstart</h4><ol><li>Run fallow without installing</li></ol><pre>// first run needs no config<br>npx fallow</pre><p>2. Read the results</p><p>3. Go deeper:</p><pre>npx fallow dead-code<br>npx fallow dupes<br>npx fallow health<br>npx fallow fix --dry-run</pre><blockquote>Learn more about <a href="https://fallow.tools/docs/">Fallow</a>.</blockquote><h3>.fallowrc.jsonc in hyperframes.</h3><p>So I came across the file, <a href="https://github.com/heygen-com/hyperframes/blob/main/.fallowrc.jsonc">.fallowrc.jsonc in the Hyperframes codebase</a>. You might be wondering what a Hyperframes is.</p><h4>Hyperframes</h4><p>HyperFrames is an open-source framework for turning HTML, CSS, media, and seekable animations into deterministic MP4 videos. Use it locally with the CLI, from AI coding agents with skills, or as the rendering core behind hosted authoring workflows.</p><blockquote>Learn more about <a href="https://github.com/heygen-com/hyperframes/tree/main">Hyperframes</a>.</blockquote><p>Now that you understand what Hyperframes is, let’s find out <a href="https://fallow.tools/docs/configuration/overview/">Fallow’s configuration</a>.</p><p>Fallow runs with no config. Add .fallowrc.json, .fallowrc.jsonc, or fallow.toml when you need to change entry points, ignore patterns, rules, or plugin settings for your project.</p><blockquote>Learn more about the <a href="https://fallow.tools/docs/configuration/overview/#config-fields">config fields</a>.</blockquote><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 3+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://fallow.tools/docs/">fallow.tools/docs/</a></li><li><a href="https://github.com/fallow-rs/fallow">github.com/fallow-rs/fallow</a></li><li><a href="https://github.com/heygen-com/hyperframes/tree/main">github.com/heygen-com/hyperframes/</a></li><li><a href="https://fallow.tools/docs/configuration/overview/">fallow.tools/docs/configuration/overview</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fde9711d0261" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hindsight, an Agent memory that learns.]]></title>
            <link>https://medium.com/@ramunarasinga/hindsight-an-agent-memory-that-learns-4beffdcc8f24?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/4beffdcc8f24</guid>
            <category><![CDATA[hindsight]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[memories]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 17:27:12 GMT</pubDate>
            <atom:updated>2026-10-07T17:27:12.016Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review Hindsight. You will learn:</p><ol><li>What is Hindsight?</li><li>Quickstart guide</li><li>Adding Hindsight to your agent.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Hq0YhGeBJcFf36K1tMYH-w.png" /></figure><h3>What is Hindsight?</h3><p>Hindsight™ is an agent memory system built to create smarter agents that learn over time. Most agent memory systems focus on recalling conversation history. Hindsight is focused on making agents that learn, not just remember.</p><p>It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.</p><p>Check out the <a href="https://github.com/vectorize-io/hindsight">demo video in README</a>.</p><h3>Quickstart guide</h3><p>In the Hindsight README, I found four steps to get started quickly.</p><ol><li>Start a server</li></ol><pre>export OPENAI_API_KEY=sk-xxx<br><br>docker run -it --pull always --name hindsight --restart unless-stopped -p 8888:8888 -p 9999:9999 \<br>  -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \<br>  -v hindsight-data:/home/hindsight/.pg0 \<br>  ghcr.io/vectorize-io/hindsight:latest</pre><p>2. Connect a client</p><pre>pip install hindsight-client -U                                  # Python<br>npm install @vectorize-io/hindsight-client                        # Node.js / TypeScript<br>go get github.com/vectorize-io/hindsight/hindsight-clients/go     # Go<br>curl -fsSL https://hindsight.vectorize.io/get-cli | bash          # CLI</pre><p>Below is an example showing this:</p><pre>from hindsight_client import Hindsight<br><br>client = Hindsight(base_url=&quot;http://localhost:8888&quot;)<br><br># Retain: Store information<br>client.retain(bank_id=&quot;my-bank&quot;, content=&quot;Alice works at Google as a software engineer&quot;)<br><br># Recall: Search memories<br>client.recall(bank_id=&quot;my-bank&quot;, query=&quot;What does Alice do?&quot;)<br><br># Reflect: Generate disposition-aware response<br>client.reflect(bank_id=&quot;my-bank&quot;, query=&quot;Tell me about Alice&quot;)</pre><h3>Adding Hindsight to your agent</h3><p>Below is a code snippet demonstrating how to add Hindsight to your existing LLM layer.</p><pre>// install the hindsight-llm<br>pip install hindsight-litellm</pre><pre>from openai import OpenAI<br>from hindsight_litellm import wrap_openai<br><br># Wrap your existing LLM client and you&#39;re done.<br># Defaults to Hindsight Cloud; pass hindsight_api_url for a self-hosted server.<br>client = wrap_openai(<br>    OpenAI(),<br>    bank_id=&quot;user-123&quot;,<br>    hindsight_api_url=&quot;http://localhost:8888&quot;,<br>)<br><br># Hindsight recalls relevant memories before the call<br># and retains the conversation after it.<br>response = client.chat.completions.create(<br>    model=&quot;gpt-5-mini&quot;,<br>    messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;What do you know about me?&quot;}],<br>)</pre><p>You just have to wrap your LLM client as shown above and then memories are stored, retrived automatically on every call, with no other changes to code.</p><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 3+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://github.com/vectorize-io/hindsight">github.com/vectorize-io/hindsight</a></li><li><a href="https://hindsight.vectorize.io/">hindsight.vectorize.io/</a></li><li><a href="https://vectorize.io/">vectorize.io/</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4beffdcc8f24" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is SQLite FTS5?]]></title>
            <link>https://medium.com/@ramunarasinga/what-is-sqlite-fts5-3cab9d5adb0a?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/3cab9d5adb0a</guid>
            <category><![CDATA[fts5]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[full-text-search]]></category>
            <category><![CDATA[sqlite]]></category>
            <category><![CDATA[context-mode]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 16:01:02 GMT</pubDate>
            <atom:updated>2026-09-30T16:01:02.200Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review SQLite FTS5. You will learn:</p><ol><li>SQLite FTS5 extension</li><li>context-mode reference</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KY42CfwuwtrqWfn5PcRh0Q.png" /></figure><h3>SQLite FTS5 extension</h3><p>FTS5 is an SQLite <a href="https://sqlite.org/c3ref/module.html">virtual table module</a> that provides <a href="https://en.wikipedia.org/wiki/Full_text_search">full-text search</a> functionality to database applications. In their most elementary form, full-text search engines allow the user to efficiently search a large collection of documents for the subset that contain one or more instances of a search term. The search functionality provided to world wide web users by <a href="https://www.google.com/">Google</a> is, among other things, a full-text search engine, as it allows users to search for all documents on the web that contain, for example, the term “fts5”.</p><p>To use FTS5, the user creates an FTS5 virtual table with one or more columns. For example:</p><pre>CREATE VIRTUAL TABLE email USING fts5(sender, title, body);</pre><blockquote>I copied this above information from <a href="https://sqlite.org/fts5.html">Overview of FTS5 section</a>.</blockquote><p>There’s a term that’s worth understanding here, Virtual table.</p><h4>Virtual Table</h4><blockquote>I got the below content from <a href="https://sqlite.org/vtab.html">Virtual Table Mechanism</a>.</blockquote><p>A virtual table is an object that is registered with an open SQLite <a href="https://sqlite.org/c3ref/sqlite3.html">database connection</a>. From the perspective of an SQL statement, the virtual table object looks like any other table or view. But behind the scenes, queries and updates on a virtual table invoke callback methods of the virtual table object instead of reading and writing on the database file.</p><p>The virtual table mechanism allows an application to publish interfaces that are accessible from SQL statements as if they were tables. SQL statements can do almost anything to a virtual table that they can do to a real table, with the following exceptions:</p><ul><li>One cannot create a trigger on a virtual table.</li><li>One cannot create additional indices on a virtual table. (Virtual tables can have indices but that must be built into the virtual table implementation. Indices cannot be added separately using <a href="https://sqlite.org/lang_createindex.html">CREATE INDEX</a> statements.)</li><li>One cannot run <a href="https://sqlite.org/lang_altertable.html">ALTER TABLE … ADD COLUMN</a> commands against a virtual table.</li></ul><h3>context-mode reference</h3><p>Well, where I found this term? I wrote an article about <a href="https://medium.com/@ramunarasinga/how-does-context-mode-save-the-context-13e98e87a421">How does context-mode save the context?</a> Context mode is an MCP plugin that sandboxes tool output and indexes it into a local FTS5 store.</p><p>Context-mode achieves the “315 KB → 5.4 KB” saving by never letting raw tool output reach the model’s context window — it executes code in an isolated sandbox subprocess and only returns what the code explicitly prints (console.log, print, etc.), while the full payload either stays inside the sandbox or is persisted to an on-disk FTS5 index for on-demand retrieval.</p><p>Now you understand the context-mode reference.</p><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 3+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://sqlite.org/vtab.html">sqlite.org/vtab.html</a></li><li><a href="https://sqlite.org/c3ref/module.html">sqlite.org/c3ref/module.html</a></li><li><a href="https://sqlite.org/fts5.html">sqlite.org/fts5.html</a></li><li><a href="https://en.wikipedia.org/wiki/Full-text_search">en.wikipedia.org/wiki/Full-text_search</a></li><li><a href="https://medium.com/@ramunarasinga/how-does-context-mode-save-the-context-13e98e87a421">medium.com/@ramunarasinga/how-does-context-mode-save-the-context-13e98e87a421</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3cab9d5adb0a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Turn your workflow into a step-by-step guide with Mimik.]]></title>
            <link>https://medium.com/@ramunarasinga/turn-your-workflow-into-a-step-by-step-guide-with-mimik-91beaf7f2178?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/91beaf7f2178</guid>
            <category><![CDATA[docs]]></category>
            <category><![CDATA[wxt]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[workflow]]></category>
            <category><![CDATA[mimik]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 17:06:04 GMT</pubDate>
            <atom:updated>2026-09-23T17:06:16.603Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review Mimik tool. You will learn:</p><ol><li>What is Mimik?</li><li>WXT, Next-gen web extension framework.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MPxa7GJqbX0rI9uoqNaf7g.png" /></figure><p><a href="https://mimik.westpoint.io/">Mimik</a> watches your tabs and writes the guide — annotated screenshots, numbered steps, ready to export. It never phones home.</p><h4>How it works?</h4><ul><li>Pin it</li></ul><p>Install from the Chrome Web Store and pin <a href="https://mimik.westpoint.io/">Mimik</a> to your toolbar. Click once to start recording.</p><ul><li>Work normally</li></ul><p>Do the thing you’d do anyway. <a href="https://mimik.westpoint.io/">Mimik</a> captures every click and keystroke and builds a step-by-step guide in the side panel.</p><ul><li>Export anywhere</li></ul><p>Ship a polished guide as Markdown, PDF, HTML, or a shareable link — all rendered locally. Your machine, your call.</p><blockquote>Install the <a href="https://chromewebstore.google.com/detail/mimik-step-by-step-guides/jmfohdaflahliammccpiadmkcibohgha?pli=1">Mimik chrome extension</a>.</blockquote><h3>WXT, Next-gen web extension framework.</h3><p>Well, I found the file named, <a href="https://github.com/westpoint-io/mimik/blob/dev/wxt.config.ts">wxt.config.ts</a> and also checked the <a href="https://github.com/westpoint-io/mimik/blob/dev/package.json">package.json</a> and found these below scripts:</p><pre>...<br>&quot;scripts&quot;: {<br>    &quot;dev&quot;: &quot;wxt&quot;,<br>    &quot;dev:firefox&quot;: &quot;wxt -b firefox --mv3&quot;,<br>    &quot;build&quot;: &quot;wxt build&quot;,<br>    &quot;build:firefox&quot;: &quot;wxt build -b firefox --mv3&quot;,<br>    &quot;zip&quot;: &quot;wxt zip&quot;,<br>    &quot;zip:firefox&quot;: &quot;wxt zip -b firefox --mv3&quot;,<br>...</pre><p>This led to search for wxt and that’s how I found this WXT web extenstion framework.</p><p>WXT is an open source tool that makes web extension development faster than ever before.</p><blockquote>Learn more about <a href="https://wxt.dev/">WXT</a>.</blockquote><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 3+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://github.com/westpoint-io/mimik">github.com/westpoint-io/mimik</a></li><li><a href="https://wxt.dev/">wxt.dev/</a></li><li><a href="https://chromewebstore.google.com/detail/mimik-step-by-step-guides/jmfohdaflahliammccpiadmkcibohgha?pli=1">chromewebstore.google.com/detail/mimik-step-by-step-guides/jmfohdaflahliammccpiadmkcibohgha?pli=1</a></li><li><a href="https://github.com/wxt-dev/wxt">github.com/wxt-dev/wxt</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=91beaf7f2178" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How does context-mode save the context?]]></title>
            <link>https://medium.com/@ramunarasinga/how-does-context-mode-save-the-context-13e98e87a421?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/13e98e87a421</guid>
            <category><![CDATA[fts5]]></category>
            <category><![CDATA[context-mode]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[open-source]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 15:31:01 GMT</pubDate>
            <atom:updated>2026-09-21T15:31:01.950Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review how the context is saved in context-mode. You will learn:</p><ol><li>What is context-mode?</li><li>How is context saved?</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I6IimoK6nZAEa5tdpWTbtw.png" /></figure><h3>What is context-mode?</h3><p>Context mode is an MCP plugin that sandboxes tool output and indexes it into a local FTS5 store. Your AI coding agent searches that store instead of re-sending raw bytes every turn. 17 adapters. 586,000+ developers. Runs entirely on your machine.</p><blockquote>I wrote a detailed article about what is <a href="https://medium.com/@ramunarasinga/optimize-context-window-for-ai-agents-using-context-mode-db20d7a10bf3">Context-mode</a>.</blockquote><h3>How is context saved?</h3><p>I use <a href="https://deepwiki.com/">deepwiki.com</a> to understand open source codebase. So I opened the <a href="https://deepwiki.com/mksglu/context-mode">context-mode repo on Deep Wiki</a> and asked this question:</p><blockquote>How does context-mode save the context?</blockquote><p>It gave an answer with references to the actual code snippets and files.</p><p>Context-mode achieves the “315 KB → 5.4 KB” saving by never letting raw tool output reach the model’s context window — it executes code in an isolated sandbox subprocess and only returns what the code explicitly prints (console.log, print, etc.), while the full payload either stays inside the sandbox or is persisted to an on-disk FTS5 index for on-demand retrieval.</p><p>Well, what’s a FTS5 index? FTS5 (Full-Text Search 5) is a SQLite virtual table module for full-text search. In context-mode, it’s the underlying engine that powers the knowledge base used by ctx_index/ctx_search to store and query indexed content without keeping raw bytes in the model&#39;s context. I will write about this in one of upcoming articles.</p><p>The concepts provided by Deepwiki are:</p><ul><li>Sandboxed execution</li><li>Intent-driven filtering for large outputs</li><li>Persistent storage instead of raw dump</li><li>Response tracking/savings accounting</li><li>Concrete numeric evidence</li><li>Enforcement layer</li></ul><blockquote>Learn more about <a href="https://deepwiki.com/search/how-does-contextmode-save-the_52541458-d3c8-4be5-a332-7ec312778c15?mode=fast">how context is saved</a>.</blockquote><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 3+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://github.com/mksglu/context-mode">github.com/mksglu/context-mode</a></li><li><a href="https://context-mode.com/">context-mode.com/</a></li><li><a href="https://deepwiki.com/search/how-does-contextmode-save-the_52541458-d3c8-4be5-a332-7ec312778c15?mode=fast">deepwiki.com/search/how-does-contextmode-save-the_52541458-d3c8-4be5-a332-7ec312778c15?mode=fast</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=13e98e87a421" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build durable agents with eve.]]></title>
            <link>https://medium.com/@ramunarasinga/build-durable-agents-with-eve-5b2867d2d295?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b2867d2d295</guid>
            <category><![CDATA[vercel]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[eve]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Fri, 18 Sep 2026 15:31:01 GMT</pubDate>
            <atom:updated>2026-09-18T15:31:01.434Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review <a href="https://eve.dev/">eve.dev</a>. You will learn:</p><ol><li>What is Eve?</li><li>Architecture</li><li>Usage Example</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AF_TIO4ZAzb8iz5OCvnHWA.png" /></figure><h3>What is Eve?</h3><p>eve is a filesystem-first framework for durable backend agents on Vercel.</p><p>You author an agent as a directory on disk. The directory is the contract — markdown for the parts a human should read like a spec, TypeScript for the parts that benefit from real types and runtime behavior.</p><p>The framework is called eve. The published npm package is eve. The CLI binary is eve.</p><p>Directory structure can be as shown below:</p><pre>my-agent/<br>├── package.json<br>├── tsconfig.json<br>└── agent/<br>    ├── agent.ts           # additive runtime config (model, name, build, compaction, …)<br>    ├── instructions.md    # always-on instructions prompt<br>    ├── tools/             # typed executable integrations<br>    ├── skills/            # optional named procedures the model can load on demand<br>    ├── hooks/             # lifecycle and stream-event subscribers<br>    ├── channels/          # message ingress and delivery (HTTP, Slack, …)<br>    ├── connections/       # external MCP server connections<br>    ├── sandbox/           # the agent&#39;s single sandbox (optional override)<br>    ├── workspace/         # files seeded into the sandbox on each session<br>    ├── subagents/         # specialist child agents (reuse `defineAgent`)<br>    ├── schedules/         # recurring jobs<br>    └── lib/               # shared authored code imported by other files</pre><blockquote>Learn more about <a href="https://eve.dev/">Eve</a>.</blockquote><h3>Architecture</h3><p>You do not need this section to author an eve agent — it documents the public HTTP protocol contracts so eve composes predictably with other systems.</p><p>eve’s internal split is:</p><ul><li>the channel normalizes inbound transport, applies auth and delivery policy, and owns channel-local addresses</li><li>the harness does one unit of AI work and returns { session, next }</li><li>the runtime persists state, follows next, streams events, and owns workflow primitives (start(), resumeHook(), createHook(), getWritable())</li></ul><p>The public HTTP protocol exposes one immutable identifier: sessionId. Create a session explicitly, then use that ID for follow-up messages, controls, streaming, and inspection. Channel-local addresses remain behind authored channel APIs.</p><blockquote>This is from Architecture section in <a href="https://github.com/vercel/eve/tree/main/packages/eve">packages/eve</a>.</blockquote><h3>Usage Example</h3><p>I picked this example from <a href="https://github.com/vercel/eve/tree/main/packages/eve">packages/eve README</a>.</p><h4>agent/instructions.md</h4><pre>You are a weather-focused assistant. Be concise, accurate, and explicit when you use a tool.</pre><h4>agent/tools/get_weather.ts</h4><pre>import { defineTool } from &quot;eve/tools&quot;;<br>import { z } from &quot;zod&quot;;<br><br>export default defineTool({<br>  description: &quot;Get the current weather for a city.&quot;,<br>  inputSchema: z.object({<br>    city: z.string(),<br>  }),<br>  async execute(input) {<br>    return {<br>      city: input.city,<br>      condition: &quot;Sunny&quot;,<br>      temperatureF: 72,<br>    };<br>  },<br>});</pre><h4>agent/agent.ts</h4><pre>import { defineAgent } from &quot;eve&quot;;<br><br>export default defineAgent({<br>  model: &quot;openai/gpt-5.4-mini&quot;,<br>});</pre><p>and then you would do this:</p><pre>npx eve@latest init my-agent</pre><p>eve init writes a new agent with eve&#39;s default model. Pass --model openai/gpt-5.5 to choose another AI Gateway model, --reasoning high to set a reasoning effort, or --channel-web-nextjs to add the Web Chat application. It installs dependencies, initializes Git, and starts the development server. When it finds a supported coding-agent REPL, the handoff menu can open that REPL instead or exit. Targeting an existing project directory (eve init .) adds the agent files and missing dependencies instead. It does not create a Vercel project or deploy the agent.</p><p>Learn more about <a href="https://eve.dev/docs/guides/deployment/overview">deploying agents created using Eve</a>.</p><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 3+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://eve.dev/">eve.dev/</a></li><li><a href="https://github.com/vercel/eve/tree/main/packages/eve">github.com/vercel/eve/tree/main/packages/eve</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b2867d2d295" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Optimize context window for AI agents using Context-Mode.]]></title>
            <link>https://medium.com/@ramunarasinga/optimize-context-window-for-ai-agents-using-context-mode-db20d7a10bf3?source=rss-a0a9d4ff1c65------2</link>
            <guid isPermaLink="false">https://medium.com/p/db20d7a10bf3</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[context-mode]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[context-window]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Ramu Narasinga]]></dc:creator>
            <pubDate>Tue, 15 Sep 2026 15:31:01 GMT</pubDate>
            <atom:updated>2026-09-15T15:31:01.562Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, we review Context-Mode. You will learn:</p><ol><li>What is Context-Mode?</li><li>The problem</li><li>How Context Mode solves it</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ocw21xuiJIHXXSQUCNh4Iw.png" /></figure><h3>What is Context-Mode?</h3><p>Context-Mode is the other half of the context problem — for developers and for engineering orgs. An MCP plugin that keeps raw data out of the LLM context window. A Platform that surfaces what your team is doing with the saved tokens. Same engine, two surfaces, your choice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7tZS9b1p76dh_lEVac53Lw.png" /></figure><h3>The problem</h3><p>I read thru the <a href="https://github.com/mksglu/context-mode">context-mode README</a> and found this problem statement which is accurate description of what goes into context window:</p><p>Every MCP tool call dumps raw data into your context window. A Playwright snapshot costs 56 KB. Twenty GitHub issues cost 59 KB. One access log — 45 KB. After 30 minutes, 40% of your context is gone. And when the agent compacts the conversation to free space, it forgets which files it was editing, what tasks are in progress, and what you last asked for. On top of that, the agent wastes output tokens on filler, pleasantries, and verbose explanations — burning context from both sides.</p><p>I mean tool call raw data can be heavy. I do use the <a href="https://caveman.so/">Caveman</a>, an open-source tool and skill designed for <a href="https://caveman.so/">Caveman</a> that strips filler words, pleasantries, and unnecessary phrasing from AI coding assistants like Claude Code to cut token usage by 65% to 75%. I configured it in my VS Code + Claude Code.</p><h3>How Context Mode solves it</h3><p>Context-Mode describes 4 sides of the problem:</p><ol><li>Context Saving</li><li>Session Continuity</li><li>Think in Code</li><li>No prose-style enforcement</li></ol><p>I did not quite understand what “Think in Code” meant here:</p><pre>// Before: 47 × Read() = 700 KB.  After: 1 × ctx_execute() = 3.6 KB.<br>ctx_execute(&quot;javascript&quot;, `<br>  const files = fs.readdirSync(&#39;src&#39;).filter(f =&gt; f.endsWith(&#39;.ts&#39;));<br>  files.forEach(f =&gt; console.log(f + &#39;: &#39; + fs.readFileSync(&#39;src/&#39;+f,&#39;utf8&#39;).split(&#39;\\n&#39;).length + &#39; lines&#39;));<br>`);</pre><p>I mean you would want the contents of the file considered in the context window, not just the file names because file name alone does not tell you the entire story right?</p><blockquote>Learn more about <a href="https://github.com/mksglu/context-mode">Context Mode</a>.</blockquote><h3>About me:</h3><p>Hey, my name is Ramu Narasinga. Email: ramu.narasinga@gmail.com</p><p>I spent 3+ years studying OSS codebases and wrote 400+ articles on what makes the production-grade. Now I’m putting that into practice differently — instead of writing every fix myself, I run coding agents that do it.</p><p><em>How it works?</em><br>Register your machine as a Runtime, point it at your repo. Agents pick up issues. write the fix, open the PR. You just review, they execute.</p><blockquote><a href="https://thinkthroo.com/">Build your own coding agents at </a><a href="http://thinkthroo.com/">thinkthroo.com</a></blockquote><h3>References:</h3><ol><li><a href="https://github.com/mksglu/context-mode">github.com/mksglu/context-mode</a></li><li><a href="https://context-mode.com/">context-mode.com/</a></li><li><a href="https://caveman.so/">caveman.so/</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db20d7a10bf3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>