<?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 j3ffyang on Medium]]></title>
        <description><![CDATA[Stories by j3ffyang on Medium]]></description>
        <link>https://medium.com/@j3ffyang?source=rss-78472ebe2e19------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*h6P-zBhSPW6t7TKlug3LXw.png</url>
            <title>Stories by j3ffyang on Medium</title>
            <link>https://medium.com/@j3ffyang?source=rss-78472ebe2e19------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 09:00:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@j3ffyang/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[Projecting a Page Onto Another Machine: a Local MCP Over SSH]]></title>
            <link>https://j3ffyang.medium.com/projecting-a-page-onto-another-machine-a-local-mcp-over-ssh-9a2e74bdf9cb?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/9a2e74bdf9cb</guid>
            <category><![CDATA[omarchy]]></category>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[arch-linux]]></category>
            <category><![CDATA[open-code]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 14:51:01 GMT</pubDate>
            <atom:updated>2026-11-17T14:51:01.542Z</atom:updated>
            <content:encoded><![CDATA[<p><em>2026–11–10, updated 2026–11–17</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NjEfFqIl8-0DqFubdp-23w.png" /></figure><p><strong>Thesis:</strong> “Open this page on the other laptop” is a one-line goal that turned into four separate faults, none of them about SSH. A local MCP server that hides them is 390 lines; finding them took a differential measurement at every step, and each answer invalidated the previous guess.</p><p><strong>Status:</strong> verified end to end on both machines, with the proxy cold-start, the browser launch, the fullscreen dispatch and the missing-binary fallback exercised in the failing state rather than merely observed working.</p><h3>The goal</h3><p>Two machines on one LAN. The handheld runs the agent; the laptop, xps, has the screen. The task: put watch.plex.tv/live-tv on that screen, in a browser, fullscreen, on demand.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NDgplP8X_rqOO_eK6Hu1kQ.png" /></figure><p>The page is not the point; the network path is. xps runs a proxy that exits through one chosen VPN, and launching an app through that path is the test. Plex live TV is just the app that makes the requirement visible.</p><p>Carrying the handheld’s session across was rejected first: Chromium-family browsers encrypt cookies with a key in the OS keyring, so a copied profile is undecryptable bytes unless the secret travels too — and then the second machine holds every Google session the first one has. Not worth one login. The laptop launches its own browser instead, and the login happens once, there.</p><h3>The four faults</h3><p>Each one looked like the fault before it was measured.</p><h3>1. The name resolves to the wrong address</h3><p>The first launch produced a window that never loaded. Firewall, DNS, routing — none of them.</p><p>watch.plex.tv does not resolve to Plex. It came back as 157.240.20.18 when measured, 199.59.149.230 a week later; the real answer, from a public resolver, is 104.18.36.51. The wrong answer moves, which is the tell: something other than the site&#39;s DNS is answering, so the page cannot load until traffic leaves through a tunnel.</p><p>That is the whole job, restated: not “open a URL” but “open a URL through a proxy”.</p><p>The confirming measurement was a control group — the same request from both machines, before and after the proxy existed:</p><pre>direct:            curl: (28) Connection timed out after 8001 milliseconds<br>through the proxy: 200 in 15.385030s</pre><h3>2. Firefox has no --proxy-server</h3><p>There is no flag for it; proxy settings live in prefs.js. Two consequences, both of which bite:</p><ul><li>The setting has to be <em>written</em>, not passed. The server writes a dedicated profile’s user.js on every call, which makes the configuration idempotent and keeps the proxy scoped to one profile instead of routing the whole browser through it.</li><li>A running browser cannot be retrofitted. A second invocation only forwards the URL to the existing process, so the server restarts a browser that lacks the setting, reuses the window when exactly one already shows the page, and clears the leftovers when the window count has drifted.</li></ul><p>The proxy also solves the DNS problem for free: with an HTTP proxy, the CONNECT carries a hostname and the name is resolved at the far end, so the poisoned local answer is never consulted.</p><h3>3. Fullscreen is not a browser flag here</h3><p>--start-fullscreen looked correct and did nothing. Under a tiling compositor, and especially when the launch is forwarded to a running instance, the flag is not applied to a new window.</p><p>Hyprland 0.56 also changed the dispatcher language. The old form is rejected outright:</p><pre>$ hyprctl dispatch fullscreen 1<br>error: [string &quot;return hl.dispatch(fullscreen 1)&quot;]:1: &#39;)&#39; expected near &#39;0&#39;</pre><p>Dispatchers are now Lua expressions, hl.dsp.*, and the numbering is not what it looks like: <strong>mode 2 is fullscreen, mode 1 is only maximize.</strong> Dispatching 1 maximises the window and reports fullscreen = 1, which is easy to read as success.</p><p>The working sequence focuses a specific window by address, then fullscreens it:</p><pre>hyprctl dispatch &quot;hl.dsp.focus({ window = \&quot;address:$addr\&quot; })&quot;<br>hyprctl dispatch &#39;hl.dsp.window.fullscreen()&#39;</pre><p>Focusing by address matters: the target page sometimes opens a second window of its own — a sign-in step — and fullscreening “the active window” then lands on the wrong one.</p><h3>4. v2rayN is X11-only</h3><p>Starting the proxy on demand looked trivial and failed immediately:</p><pre>Unhandled exception. System.Exception: XOpenDisplay failed<br>   at Avalonia.X11.AvaloniaX11Platform.Initialize(X11PlatformOptions options)</pre><p>The Avalonia build of v2rayN has an X11 backend and nothing else, so it needs DISPLAY. An SSH session has none. Hyprland does not export DISPLAY into its own environment either, so reading it from the compositor returns nothing. And the fallback :0 is worse than no display at all, because another X server exists there and answers:</p><pre>Authorization required, but no authorization protocol specified</pre><p>XWayland was on :1. The number changes across boots, so it is read from the Xwayland command line rather than assumed.</p><p>One more detail: v2rayN resolves bin/ and guiConfigs/ relative to its working directory, so it has to be started from its own directory, not by absolute path from $HOME.</p><h3>The fifth symptom: a silent page</h3><p>Four faults, and the projection still was not finished. The page came up playing, in fullscreen, and mute: a speaker with a cross sat in the corner of the video. The obvious reading is that the laptop is muted, and it is wrong in a way that costs three rounds of troubleshooting.</p><p>A crossed-out speaker can come from one of four places, and each has its own command:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G_BYhhK3YFKNUYT10qZ9vA.png" /></figure><p>The last row is the one that decides it. An empty Streams: means nothing was ever handed to the sink, so no change at the mixer or the ALSA layer could have produced sound. Raising the volume was the instinct, and it is wrong: a page that is not playing stays silent at any volume. The mute was a control <em>inside</em> the player — grim puts it inside the video frame, bottom-left, around (141, 551) on a 1920×1080 screen, which is a player control and not a desktop widget.</p><p>The fix is that control, and there is no pointer to click it with:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N7LXXpeaZXjdbEF75pfGfg.png" /></figure><p>So the lever is the site’s own keyboard shortcut. Plex web binds m to mute-toggle, and the three steps are the same shape as the fullscreen dispatch — discover the session, then act on a named window:</p><pre># an SSH login carries no session env: without this, hyprctl and wtype fail silently<br>RUNDIR=&quot;/run/user/$(id -u)&quot;<br>WD=$(ls &quot;$RUNDIR&quot; | grep -m1 &#39;^wayland-[0-9]*$&#39;)<br>HIS=$(ls &quot;$RUNDIR/hypr&quot; | head -1)<br>export XDG_RUNTIME_DIR=&quot;$RUNDIR&quot; WAYLAND_DISPLAY=&quot;$WD&quot; HYPRLAND_INSTANCE_SIGNATURE=&quot;$HIS&quot;</pre><pre># focus by address, never by title — titles change with the page<br>hyprctl dispatch &quot;hl.dsp.focus({ window = \&quot;address:0x55b4d64a01d0\&quot; })&quot;</pre><pre>wtype -k m</pre><p>Focusing first is not politeness. Without it the keystroke lands on whatever holds focus — here the v2rayN window parked on its workspace — and the failure is completely silent.</p><p>Which is why the check is by effect, not by exit code:</p><pre>pactl list sink-inputs | grep -E &#39;Sink Input|Mute:|application.name&#39;<br>#   Sink Input #99 · Mute: no · Volume: 100% · application.name = &quot;Firefox&quot;<br>wpctl status | sed -n &#39;/Audio/,/Video/p&#39;<br>#   68. Firefox → output_FL / output_FR  [active]</pre><p>wtype returns 0 whether or not the page listened. <strong>The stream appearing is the proof.</strong> Before the keypress there was no sink-input at all; after it, the hardware path was live.</p><p>The mute is re-asserted on every load, and the unmute dies at the next navigation. Measured: Ctrl+R on the just-unmuted page brought the crossed speaker back and left zero sink-inputs. Two mechanisms would produce that, and they were not separated — Plex&#39;s own default mute, and Firefox&#39;s default autoplay policy, which blocks audio until a user gesture and grants only a <em>temporary</em> permission that resets on navigation. That profile carries no autoplay override, and permissions.sqlite holds no autoplay grant for watch.plex.tv, which is consistent with the second and does not exclude the first. What is established is narrower and sufficient: the mute returns on its own, and one key event clears it for the session.</p><p>The other open item is the sink volume — 15% on the day, 25% when re-checked two days later — quiet enough after the unmute to look like a further fault. That is a separate decision, so it was reported and left alone.</p><p>Measured 2026–11–15, on a page projected minutes earlier.</p><h3>The design</h3><p>The server exposes three tools, and deliberately no generic “run a command on the remote host” tool, which would be remote code execution behind an agent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Um4zgYYkv7T3xbWxeS5_jg.png" /></figure><p>Everything machine-specific is an environment variable with a sane default: the ssh destination, the URL, the proxy address, and the directory holding the proxy binary. The committed file therefore carries no address and no username; the one home-relative default is the proxy directory, and XPS_V2RAY_DIR overrides it.</p><p>The ssh destination defaults to an alias, not an address. On a LAN the address is usually a DHCP lease, so putting it in the alias keeps the churn in ~/.ssh/config, where changing it is a one-line edit, instead of in Python.</p><p>The proxy start is <strong>best effort</strong>, not a precondition. If the port is down the server starts the binary and waits; if the binary is not there at all, it launches the browser anyway and says so. A projection helper that refuses to open a window because an unrelated daemon is missing is less useful than one that opens it and reports the truth.</p><h3>One call, end to end</h3><p>Every branch below exists because one of the four faults broke the naive order: the proxy comes first and never blocks, the GUI is parked on its own workspace, the browser is reused only when it is already correct, and fullscreen is dispatched rather than passed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yIQ0NF0ouXla-jqAbb7S8g.png" /></figure><h3>Setup</h3><p><strong>The code here is reference, not homework.</strong> Do not retype it: hand the server or the skill to your agent and let it do the work on your machine. The snippets are here so you can check what it did.</p><p>On the remote host: Hyprland with hyprctl, jq, ss from iproute2, and Firefox — installed for this test, so the proxy can live in its own profile and leave the machine&#39;s default Chromium alone. Optionally v2rayN, if the network needs it.</p><p>On the machine running the agent: uv and ssh with a key that already works. In ~/.ssh/config:</p><pre>Host xps<br>    HostName 10.0.0.42<br>    User you<br>    IdentityFile ~/.ssh/your_key</pre><p>The server itself is one directory:</p><pre>.opencode/mcp/xps-plex/<br>├── README.md<br>└── server.py</pre><p>Registered in opencode.jsonc. The SDK is pulled at run time rather than installed, so nothing changes on the system Python:</p><pre>&quot;mcp&quot;: {<br>  &quot;servers&quot;: {<br>    &quot;xps-plex&quot;: {<br>      &quot;type&quot;: &quot;local&quot;,<br>      &quot;command&quot;: [<br>        &quot;uv&quot;, &quot;run&quot;, &quot;--with&quot;, &quot;mcp&quot;,<br>        &quot;/path/to/xps-plex/server.py&quot;<br>      ],<br>      &quot;environment&quot;: {<br>        &quot;XPS_SSH_HOST&quot;: &quot;xps&quot;,<br>        &quot;XPS_PROXY_WORKSPACE&quot;: &quot;2&quot;<br>      }<br>    }<br>  }<br>}</pre><p>V2 reads names only under mcp.servers; older docs show the bare mcp.xps-plex shape, which this schema does not define. The field for keeping a server configured without connecting it is disabled, not enabled — leave it out and the server connects.</p><p>MCP servers load at startup, so the client needs a restart before the tools appear.</p><p>Every machine-specific value is an environment variable in that block, with a default that works if you set none of them:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E1BbfBUFTURtAhr10RK62A.png" /></figure><h3>Using it</h3><p>Three tools, and no fourth. Each is one call with at most two arguments, and none of them accepts a command to run.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YuH-lR389xtyBC0fA3-w8g.png" /></figure><p>Both arguments are optional, so the ordinary call is the bare one:</p><pre>open_page()</pre><p>A cold run reports each part, including the ones that went well:</p><pre>Firefox was not running; started on the proxy profile<br>window: Live TV Guide - Plex — Mozilla Firefox<br>fullscreen: mode 2<br>proxy: was down, v2rayN started<br>proxy window: workspace 2</pre><p>The proxy fallback reports itself too, rather than failing:</p><pre>window: Live TV Guide - Plex — Mozilla Firefox<br>fullscreen: mode 2<br>proxy: v2rayN not found, launched without it</pre><p>Three behaviours are worth knowing before you call it twice.</p><p><strong>It is idempotent, and it cleans up after itself.</strong> If the same profile already shows the page in exactly one window, the window is reused and nothing is relaunched. If it has drifted — two windows from earlier calls, a blank tab, a crash-restore dialog left by a killed browser — the browser is cleared and started once. Calling it in a loop does not pile up windows.</p><p><strong>Fullscreen is mode 2, and it is dispatched, not requested.</strong> The window is focused by address first, so the sign-in popup the site sometimes opens cannot steal the fullscreen from the page you asked for.</p><p><strong>The proxy GUI is parked out of the way.</strong> v2rayN is moved to the workspace named by XPS_PROXY_WORKSPACE (2 by default), silently, so the browser owns the visible workspace instead of sharing it with a proxy window.</p><p>Ask for it in the same words you would use to a person — the tool schema does the routing:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/983/1*oBObA4oE0Vdd4jChGXp3Pw.png" /></figure><p>An agent that skips the status call and goes straight to open_page is not doing anything wrong: the server checks the proxy itself and reports what it found.</p><h3>The debugging trail</h3><p>Every row is a measurement that overturned the previous assumption.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7UrAiJfeGsxkEgh32K8J4A.png" /></figure><p>The DISPLAY discovery hid inside a Python string: a sed backreference written as \1 in a non-raw string became the octal escape U+0001, so the regex silently matched nothing and the :0 fallback returned. The compiler had already warned about it, twice, and the warning was ignored. A SyntaxWarning is a reading instruction.</p><h3>What this does not do</h3><p>It does not move a browser session: the login lives in the dedicated profile on the remote machine, and the server never sees a credential.</p><p>It does not start the proxy if the binary is missing, and it does not verify that the page rendered. It waits for a window whose title is not a placeholder — a blank tab, New Tab, the crash-restore dialog — then reports what it found. Whether the right content sits behind that title is the caller&#39;s judgement.</p><p>It does not check that the page is audible. A launch can be complete — one window, fullscreen mode 2, the right URL — and still be mute inside the player, which no amount of status reporting will reveal. That has to be read off the screen, and fixed with the page’s own controls.</p><p>It assumes one monitor and one compositor. Multiple outputs would need the dispatch to name a monitor.</p><h3>Files</h3><p>The server and its README: <a href="https://github.com/j3ffyang/ai-thoughts/tree/main/.opencode/mcp/xps-plex">ai-thoughts/.opencode/mcp/xps-plex</a>.</p><p>The skill that does the same job, with the same script logic and no process: <a href="https://github.com/j3ffyang/ai-thoughts/tree/main/.opencode/skills/xps-plex">ai-thoughts/.opencode/skills/xps-plex</a>.</p><p>Which of the two to reach for, and why both exist: <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261111-skill-vs-mcp.md">A Skill or an MCP Server?</a>.</p><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9a2e74bdf9cb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A TradingView MCP: Three Tools Over an Endpoint Nobody Documents]]></title>
            <link>https://j3ffyang.medium.com/a-tradingview-mcp-three-tools-over-an-endpoint-nobody-documents-5e69417d6929?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e69417d6929</guid>
            <category><![CDATA[tradingview]]></category>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[open-code]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 14:46:01 GMT</pubDate>
            <atom:updated>2026-11-12T14:46:01.467Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OnJAEmscnb6hDF3XU8ovhw.png" /></figure><p><em>2026–11–11</em></p><p><strong>Thesis:</strong> <strong>TradingView</strong> publishes no API, so every TradingView tool is built on the endpoints its own web app uses. Wrapping one in an MCP is small work; the work is discovering which of its answers are quietly wrong.</p><p><strong>Status:</strong> running on both — <strong>◧ OpenCode</strong> locally and the <strong>⚕ Hermes</strong> bot — three read-only tools, verified against live data on each.</p><h3>Why this exists</h3><p>I wanted the agent to answer “what is NVDA doing right now” without me opening a chart. TradingView is where that answer lives, and TradingView has no official API — no key, no documented endpoint, no developer programme. What exists is a set of endpoints the site itself calls, and a handful of libraries that talk to them.</p><p>The libraries are not equivalent, and the difference is maintenance:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BkO5jRiyE8JhA3HXfBd__Q.png" /></figure><p>LibraryLast releaseVerdicttradingview-screener2026-09maintained; drives the scanner endpointtradingview-ta2022stale, and its ratings are already scanner fieldstradingview-scraper2026-09scrapes chart data, more fragilepyTradingView2026-02websocket client, more moving parts</p><p>One library covers it. The scanner returns quotes, technical ratings <em>and</em> screening, which is everything I actually wanted — no chart drawing, no websocket, no cookies.</p><h3>Three tools</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QPGN-KOdquAWG7uDxzSdBA.png" /></figure><pre>NASDAQ:AAPL  AAPL  348.51  -0.62  29,756,060  5.06T  0.26<br>NASDAQ:NVDA  NVDA  168.89  -5.47  429,273,195  4.07T  -0.51</pre><p>Symbols work bare: AAPL resolves to NASDAQ:AAPL, BTC to BINANCE:BTCUSDT.</p><p>A rating runs −1 to +1. What makes it worth printing side by side is that the three columns disagree: NVDA at one point read overall −0.51, moving averages −0.93, oscillators −0.09 — the trend is down, the momentum has stopped falling. A single number hides that.</p><h3>Three things the first version got wrong</h3><p><strong>Chained filters overwrite each other.</strong> A screen asking for RSI &lt; 25 returned AAPL at RSI 55. The client&#39;s where() replaces the filter rather than adding to it, so only the last condition survived; passing them as where(*conditions) ANDs them. The lesson generalises: a query builder that quietly keeps one condition produces plausible, wrong results — no error, just the wrong list.</p><p><strong>type == &quot;stock&quot; does not mean common stock.</strong> Preferred shares such as BAC/PE and MS/PA report type: stock, and they dominated an oversold screen because preferreds trade thin and fall fast. The discriminating field is typespecs, so stock screens now add typespecs has &#39;common&#39; unless you filter on it yourself, and the table says so underneath.</p><p><strong>Numbers are not presentation.</strong> The scanner returns -0.6216316404801981. Rounding at the edge, grouping thousands, and writing market caps as 5.06T is the difference between a table a person reads and a dump.</p><h3>Why an MCP and not a skill</h3><p>The same question came up for the projection tool, and the answer is <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261111-skill-vs-mcp.md">in the other piece</a>: a skill teaches a procedure, an MCP adds a capability. Here the deciding factor is the filter grammar. tv_screen takes JSON with fields, operators, between, and has — as a tool schema that grammar arrives with the call and a bad field is rejected in words the model can act on. Prose in a skill would drift the first time I added a field.</p><p>The cost is honest: the three tool definitions sit in context permanently, about 670 tokens. For a tool used in most trading conversations that is cheap.</p><h3>Registering it</h3><p>Locally, OpenCode reads mcp.servers.&lt;name&gt; from its config. Two traps, both hit:</p><ul><li>The flat mcp.&lt;name&gt; form still works and is marked deprecated in the schema; mixing both shapes in one file is how you end up debugging the copy that does nothing.</li><li>After editing the config, opencode mcp list read the <strong>background service&#39;s</strong> cached state and showed the old server list. The file was right the whole time. Restart the service, or check with opencode debug config, before believing a missing server.</li></ul><p>On the Hermes VPS the shape is different: hermes mcp add &lt;name&gt; --command … --args …, which connects, lists the tools, and asks whether to enable them — non-interactively it cancels at that prompt, so pipe y or add the tools afterwards.</p><p>The host had no uv and no ensurepip, so the server lives in its own venv under ~/.hermes/mcp/tradingview/ with pip bootstrapped into that one directory. Nothing was installed system-wide.</p><h3>What enabling it on a bot actually means</h3><p>Hermes exposes each MCP server as a toolset — mcp-tradingview — and a platform gets a list of toolsets. Adding it to platform_toolsets.discord gives it to <strong>everyone in the channel</strong>, because Hermes configures tools per platform, never per user. There is no way to hand it to one member.</p><p>That is fine here: the tools read a scanner and write nothing. It does mean each member’s question spends model tokens on my account. Worth knowing before adding the next toolset.</p><h3>What it does not do</h3><p>No chart data — candles belong to another endpoint and a different library, and this answers “what is it doing now”, not “draw me a chart”. No watchlists, alerts, or ideas: those need a logged-in session, and cookies in a config file are another risk class. Nothing writes.</p><p>And the honest constraint: this is unofficial tooling against endpoints TradingView can change any morning. The data is theirs, the queries are for personal use, and nothing here should be redistributed.</p><h3>The files</h3><ul><li>GitHub: <a href="https://github.com/j3ffyang/ai-thoughts/tree/main/.opencode/mcp/tradingview">tradingview MCP</a> — server.py and its README.</li></ul><p>The lesson is not about TradingView. Any service without a documented API still has a shape, and the work is finding where that shape lies to you quietly.</p><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e69417d6929" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Bot Runs as You: Hardening a ⚕ Hermes Agent for Guests]]></title>
            <link>https://j3ffyang.medium.com/the-bot-runs-as-you-hardening-a-hermes-agent-for-guests-dd25c87be46e?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd25c87be46e</guid>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[hermes]]></category>
            <category><![CDATA[security-hardening]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[open-source]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 14:26:01 GMT</pubDate>
            <atom:updated>2026-11-11T14:26:01.592Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i023owVdtZqZ60LaxNMX_g.png" /></figure><p><em>2026–11–08</em></p><p><strong>Thesis:</strong> a messaging bot is not a separate user. It is your account with a Discord adapter attached, and adding people to it adds people to your shell — so the boundary has to be drawn on the platform, not on the person.</p><p><strong>Goal:</strong> members keep the bot in Discord and can add nothing that runs; you keep the full toolset over SSH.</p><p><strong>Status:</strong> config-only. The Discord list below is five toolsets — four read-only, plus skills behind the write gate — and the CLI keeps the full set.</p><p>Related: <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261107-hermes-discord-integration.md">the private Discord setup</a> is the invite side, and <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261105-hermes-skill-index.md">the skill index</a> is what fills the library this article gates.</p><h3>A skill appeared in ~/.hermes</h3><p>Yesterday a member asked the bot for a custom skill. The bot wrote one, and it landed in /home/jeff/.hermes/skills/. Nothing broke. That was the problem: it worked exactly as designed, and the design is &quot;the agent may modify any skill it can see&quot;.</p><p>Skills are files. The agent creates, patches, and deletes them through skill_manage. If that tool is enabled on the platform the guest is talking on, a guest puts a file on your VPS by asking politely.</p><h3>Why it happened</h3><p>The gateway runs as jeff. Every tool call — terminal, write_file, skill_manage, cronjob — inherits that uid. There is no privilege boundary inside the process, because there is no second identity to switch to.</p><p>So the invite was the mistake, not the prompt. Adding a member to the bot added them to the account that holds your SSH keys, your git config, and your home directory. <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261107-hermes-discord-integration.md">The private Discord setup</a> is how to issue an invite that does not hand over the account.</p><h3>The limitation worth knowing first</h3><p>Hermes configures tools <strong>per platform</strong>, not per user. Gateway authorization is binary: a user ID is allowed or denied. There is no “this member may run a shell, that one may not” for plain chat — only slash commands split into admin and non-admin.</p><p>So everyone on one platform shares one tool list: grant yourself terminal on Discord and every guest has it; deny it and you lose it there too. The shell is the only place you can hold more power.</p><p>That decides the architecture. The trust boundary has to be the <em>platform</em>, or a second gateway. It cannot be the person.</p><h3>Three layers</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*suA5bQvClHPLJ-r1hKF7Jg.png" /></figure><p>Identity is the only layer that survives a tool you have not audited yet. Capability is the one that stops today’s incident. The control plane keeps a guest from re-enabling what you disabled.</p><p>This article applies the second and third. The first needs a home outside your own: a tree inside /home/jeff cannot coexist with chmod 700 /home/jeff, which would lock the service user out of its own home. Everything below is the config-only path, and the tool list is the barrier.</p><h3>1. Audit what is already there</h3><pre># pick the window yourself — the day you noticed, not a date from this article<br>find ~/.hermes -newermt &quot;-7 days&quot; -type f \<br>  | grep -vE &#39;state\.db|/logs/|gateway_state\.json&#39; | head -40<br>hermes skills list<br>hermes cron list</pre><p>Read every new file before deleting it. A planted skill runs on every future use, so a blind rm is not an audit. Then check ~/.hermes/config.yaml for keys you did not write, and .env for values you did not set.</p><h3>2. Take the writes away</h3><p>Every toolset that can write, and what it writes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mAWpHcwc6jlePe4tMSnk0g.png" /></figure><p>session_search is not in that table because it writes nothing — but it reads past conversations, so it belongs off on privacy grounds alone.</p><p>The selection is the whole list, not a set of removals: whatever you put in platform_toolsets.discord is exactly what that platform gets. Put the gate first and the writer last, in one file:</p><pre># ~/.hermes/config.yaml<br>skills:<br>  write_approval: true           # nothing installs without your approval</pre><pre>platform_toolsets:<br>  discord:<br>    - web        # web_search, web_extract<br>    - vision     # vision_analyze<br>    - clarify    # ask a question<br>    - todo       # in-session task list<br>    - skills     # skill_view, skills_list — writes staged, step 3</pre><p><strong>Do not paste that block in as a new section.</strong> skills: is already there, and platform_toolsets: is there once hermes tools has written it. A second copy of either makes the file unparseable, and then none of its settings load. Set the keys instead:</p><pre>hermes config set skills.write_approval true<br>hermes config set platform_toolsets.discord &#39;[&quot;web&quot;,&quot;vision&quot;,&quot;clarify&quot;,&quot;todo&quot;,&quot;skills&quot;]&#39;</pre><p>hermes config set writes into the block that already exists, and creates the parent path if the key is not there at all. If it had to recreate platform_toolsets, the cli: and telegram: lists that lived in that block are gone, and both fall back to defaults.</p><p>That is “as much as possible” under the rule that nothing installs without you. browser, image_gen, video_gen, and tts look harmless and are not: they write browser profiles, downloads, and media files.</p><p>For the maximum instead, discord: [] removes everything, tools and conversation aids alike.</p><p>The same selection, if you would rather click than type: tick those five boxes in hermes tools. Set the gate <strong>before the restart that applies the toolset</strong> — ticking the boxes writes the file, the restart activates it.</p><pre>hermes gateway restart<br>hermes tools --summary</pre><p><strong>A broken config does not fail loudly.</strong> One duplicate top-level key makes the file unparseable. Older builds fell back to defaults, and a duplicate skills: that way reverted everything — every tool back, no gate, no lock. Current builds keep serving the last config that parsed instead, reading ~/.hermes/backups/config/config.yaml.good.* in a fresh process. Either way your new block never loads: the hardening you just wrote is silently not in effect, and the permissive config from before it may be the one actually running. After any hand edit, hermes config check must report no error and hermes tools --summary must still show five. Both are one line, and they are the only check that catches a silent no-op.</p><p>The CLI keeps its full set either way, so SSH sessions still have everything.</p><h3>3. The skill gate</h3><p>Why skills is there at all: members need skills_list and skill_view to see and use the library, and those travel with skill_manage in the same toolset — toolsets are not split by tool. The gate is what keeps authorship yours. Every skill_manage write — create, patch, delete, file write — is staged under ~/.hermes/pending/skills/ instead of installed, and the skill is inert until you act. That directory appears on the first staged write, so ls on a fresh install answering <em>No such file or directory</em> means nobody has proposed anything yet, not that the gate is off.</p><pre>/skills pending      # staged writes, one line each<br>/skills diff &lt;id&gt;    # the full diff — read it in the CLI, chat truncates<br>/skills approve &lt;id&gt;<br>/skills reject &lt;id&gt;<br>/skills approval on  # the same switch as write_approval, from chat</pre><p>Members keep skills_list and skill_view; only your account is admin, so only you approve. The cost is one staged JSON per proposal — a write, but not an installation — plus one review per skill. Approved skills land in ~/.hermes/skills/, so git-init that directory: every approval then has a diff and a revert. Lock what you depend on with hermes curator lock &lt;skill&gt;, and an approved skill appears to a new session; /reload-skills forces a rescan in the current one. What the library costs before anyone proposes anything is the other half of this: <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261105-hermes-skill-index.md">the skill index</a> measures it.</p><h3>4. Close the control plane</h3><p>allow_admin_from decides who may run <strong>slash commands</strong>. It has no effect on tools: a slash command is a command, not a capability. Scope matters — allow_admin_from covers DMs, group_allow_admin_from covers server channels, and both are needed.</p><p>Both platforms.&lt;name&gt; and gateway.platforms.&lt;name&gt; are read, but the top-level block wins on shared keys, extra included. So put it in one place, next to enabled:</p><pre>platforms:<br>  discord:<br>    enabled: true<br>    extra:<br>      allow_admin_from:<br>        - &quot;&lt;your-user-id&gt;&quot;<br>      group_allow_admin_from:<br>        - &quot;&lt;your-user-id&gt;&quot;<br>      user_allowed_commands:<br>        - whoami<br>        - status<br>      group_user_allowed_commands:<br>        - whoami<br>        - status</pre><p>Write it once. If both blocks carry an extra, the top-level one wins and you will debug the copy that does nothing.</p><p>Non-admins get /help, /whoami, and that list. They do not get /yolo, which would switch approvals off in their session.</p><p><strong>Slash commands are a second write surface.</strong> They are not tools, so platform_toolsets never sees them, and several of them write on their own:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sfTdmWDZjF5zgJGgj0dnRA.png" /></figure><p>The admin split is the only thing that gates them. That makes it more important than the tool list, not less.</p><p>/tools is not on that list and never was: it is a <strong>CLI-only</strong> command, so it does not exist in Discord and the tool selection cannot be re-opened from chat at all.</p><p>Check the split from the channel, not the shell:</p><pre>/whoami</pre><p><strong>Tier: admin</strong> means the block was read. Anyone else should get <strong>Tier: user</strong> and a short list of commands after it; that line is the split working. <strong>Tier: unrestricted</strong> means no admin list is in effect for that scope — the YAML landed in the shadowed block, or you set the DM pair and are testing in a channel. In that state every command in the table above is open to your guests.</p><h3>5. Verify</h3><p>From a guest account, ask for a shell and a file write. Both should fail because the tool is <strong>absent</strong>, not because the model refused — a refusal means the tool is still there and the model is being polite. Then ask for a skill: it should stage under ~/.hermes/pending/skills/ and install nothing.</p><pre>hermes tools --summary                          # Discord: five toolsets, nothing installed<br>find ~/.hermes/skills -newermt &#39;-10 minutes&#39;    # empty<br>ls ~/.hermes/pending/skills                     # the staged proposal, if any</pre><p>Run hermes tools --summary in a real terminal — piped, or as a non-interactive SSH command, it refuses with <em>&quot;requires an interactive terminal&quot;</em>.</p><p>Confirm your own DM still answers, since the change applies to you on that platform too.</p><h3>What the finished shape looks like</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*USEDgHwA7mX7nY0a4I8J_g.png" /></figure><p>The fifth toolset is there so guests can see and use the library; the gate is what stops them installing into it. Nothing a guest writes reaches ~/.hermes/skills/ without your /skills approve.</p><p>You and the other members are equal on Discord. The only difference is slash commands, and the only place you hold more power is the shell — which is the point.</p><h3>What this does not buy</h3><ul><li><strong>The agent still writes.</strong> Sessions, logs, memory, and state.db are written by the gateway user as normal operation. The goal is &quot;no guest-directed writes&quot;, not &quot;read-only host&quot;.</li><li><strong>Prompt injection is unchanged.</strong> If web stays enabled, a page the agent fetches can still steer it inside the tools that remain. web_extract can also reach internal addresses from your VPS; drop web as well if guests only need conversation.</li><li><strong>Approvals are not a boundary.</strong> Upstream says it plainly: deny rules are a shell policy, not a sandbox. approvals.mode shapes what runs; it does not contain what runs.</li><li><strong>Your own Discord access loses the same tools.</strong> The platform is the boundary. Admin work belongs in SSH and the CLI, where your own install keeps the full toolset.</li><li><strong>Splitting by person needs a second gateway.</strong> Separate bot, separate home, separate toolset. Twice the surface, so do it only if admin-via-Discord is a real requirement.</li><li><strong>There is no second identity here.</strong> The bot still runs as you, so the tool list is the only barrier between a guest and your files. The identity layer needs a home outside your own, which a tree under /home/jeff is not.</li><li><strong>state.db has a second door.</strong> Everything above governs guests in Discord; it does not govern a client you run yourself. hermes mcp-serve gives an MCP client ten tools — read every message in ~/.hermes/state.db, post as the bot, answer pending approvals — and --tools adds model spend on top. That is a decision made at a shell, not a hole in this config, and <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261109-opencode-remote-hermes-mcp.md">the MCP piece</a> measures what it hands over.</li></ul><h3>Rollback</h3><p>Re-enable a toolset in hermes tools and restart; /skills approval off ungates skill writes the same way.</p><p>The lesson is smaller than the setup: an invite is a grant. The bot does not have a users table. It has a uid.</p><p>I will leave this running for a while before calling it settled. The question that matters is not whether the config is correct on paper, but whether the bot stays healthy in a group where other people talk to it: five toolsets, writes staged, nothing installable. Anything practice teaches me that this config does not cover goes here.</p><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd25c87be46e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Disable the Skills You Will Never Use in Hermes ⚕]]></title>
            <link>https://j3ffyang.medium.com/disable-the-skills-you-will-never-use-in-hermes-d4ac344dd9c4?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4ac344dd9c4</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[bloatware]]></category>
            <category><![CDATA[open-source]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 14:11:01 GMT</pubDate>
            <atom:updated>2026-11-10T14:11:01.763Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eh_D9Uzc1XGksVzNEqzKrQ.png" /></figure><p><em>2026–11–05</em></p><p><strong>Thesis:</strong> A fresh Hermes install seeds a catalog of skills nobody asked for, and their descriptions load before you say a word. Switch off the categories you will never touch, and load the rest on demand.</p><h3>A 4 GB VPS, a 2-core CPU, and a catalog</h3><p>I installed Hermes again on a small VPS: 4 GB of RAM, 2 cores, no GPU, running Debian. The install itself was unremarkable. What I noticed came later, in the skill list.</p><p>A small box is the right place to notice this, because it rules out the usual excuse. If anything were going to be tight, 4 GB would show it first — and nothing did. The problem was not the machine.</p><p>For the record, top on that box after a couple of hours — the unrelated sshd row trimmed:</p><pre>top - 09:15:22 up  2:33,  1 user,  load average: 0.08, 0.02, 0.01<br>Tasks: 116 total,   2 running, 114 sleeping,   0 stopped,   0 zombie<br>%Cpu(s):  0.5 us,  0.0 sy,  0.0 ni, 99.2 id,  0.0 wa,  0.0 hi,  0.3 si,  0.0 st<br>MiB Mem :   3817.6 total,   1085.3 free,    601.0 used,   2320.4 buff/cache<br>MiB Swap:      0.0 total,      0.0 free,      0.0 used.   3216.5 avail Mem</pre><pre>    PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+<br>   3197 jeff      20   0  756044 302472  41080 S   0.7   7.7   1:20.99 hermes</pre><p>Read it plainly: this is hermes <strong>idle</strong> — 0.7% CPU, no task running — and it already holds 295 MiB resident (302,472 KiB), 7.7% of 3.8 GB, with 3.2 GB available and no swap in use. Load average 0.08 across two cores, so nothing here is straining. The numbers to remember are the idle baseline: 295 MiB of RAM held, not freed — 255 MiB of that private once the 40 MiB of shared pages are discounted — and a 3k-token catalog index before I touched it. Both are paid before you ask the first question.</p><h3>What the index costs</h3><p>Hermes loads skills by progressive disclosure: nothing arrives whole. The first level is an index — name, description and category for every skill — and the full text loads only when the agent asks for that skill. Hermes’ own docs put the full-catalog index at roughly 3k tokens.</p><p>Three thousand tokens is not a crisis. It is also not free, and it is spent before I have asked anything, on descriptions of skills I will never call. An index is context: it competes with the instructions I actually wrote, and it nudges the agent toward whatever looks adjacent to the task at hand.</p><p>On a 4 GB box this never shows up as RAM pressure. It shows up as an agent that has already been told about most of a catalog it does not need.</p><p><strong>Disabling a skill saves tokens, never megabytes.</strong></p><h3>What I disabled</h3><p>The install seeds a full catalog. I opened the config, chose <strong>Toggle by category</strong>, and switched four off:</p><pre>hermes skills config<br># then: Toggle by category → emails, productivity, social-media, software-development</pre><p>That list is mine, not a recommendation. I do not run an email workflow from an agent session, I do not build social-media pipelines, and the software development I do belongs to OpenCode, where the project’s own AGENTS.md and skills already describe it. Productivity is the vaguest of the four, and vague is the last thing I want an agent guessing at.</p><p>Leave whatever you use weekly. The point is not a minimal install for its own sake; it is that a skill you have never opened is noise wearing a tool’s clothing.</p><p>Disabled is the right verb, by the way. The files stay where they are, and one toggle brings the category back — nothing is deleted, so nothing has to be reinstalled later.</p><h3>Why disable instead of ignore</h3><p>Because ignoring does not work. The index is read every session, so a category you never touch still votes on what the agent considers available on that platform, and it is still paid for at the start of every conversation there.</p><p>My preference is the other default: load them when I need them. If I want the email skill next month, one toggle brings the category back — cheaper than carrying it through a hundred sessions in case, and cheaper than reinstalling it after a purge.</p><h3>If your build differs</h3><p>The category toggles are what my install offers, and they only disable. The documented commands cover the other two shapes of the same idea:</p><pre>hermes skills opt-out            # stop future seeding, touch nothing on disk<br>hermes skills opt-in --sync      # undo, and re-seed now<br>hermes skills opt-out --remove   # the destructive one: delete unmodified bundled skills<br>hermes profile create research --no-skills   # start a profile with no catalog</pre><p>The <a href="https://hermes-agent.nousresearch.com/docs/user-guide/features/skills">skills documentation</a> is where I would start if a command has moved.</p><p>None of this touches the toolkit itself. Disabling a catalog entry hides one skill; removing the skills toolset from a platform hides skills_list, skill_view and skill_manage together; and the write gate is a third thing again — tools kept, installs staged. If you want the bot to stop offering skills at all, that is a toolset decision, not a catalog one. <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261108-hermes-guest-hardening.md">The hardening guide</a> covers the other two.</p><h3>The index is the smallest of three costs</h3><p>Weight in an install comes in three kinds, and only two of them are paid in tokens:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/784/1*YRX10pqiu3cNMWdXd4APCQ.png" /></figure><p>Tool schemas are the quiet one. Every enabled toolset ships a JSON schema before you speak, and a full hermes-cli platform carries dozens of them; turn off the six you never call and the prompt shrinks by more than the skill index ever cost. Plugins are a different currency — a platform plugin for Telegram costs process memory and startup time, not prompt space, and on this 4 GB box the idle figure in the first section is already comfortable.</p><pre>hermes tools                     # per platform; interactive, needs a real terminal<br>hermes plugins list              # what is installed, and what is switched on<br>hermes plugins disable &lt;name&gt;    # switch one off without removing it</pre><p>/context all inside a session prints the per-skill and per-toolset breakdown, which is the only way to see whether a removal actually helped. Guesswork here runs in the wrong direction: the plugin list looks like the fattest thing and is usually the cheapest.</p><h3>Closing</h3><p>A small VPS is a good place to notice this, because it removes the usual excuse: it is not the hardware, and it never was. The agent reads the skill index before you speak. Everything in it should earn its place there.</p><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4ac344dd9c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Running a Private ⚕ Hermes Agent on a VPS, Wired to Discord]]></title>
            <link>https://j3ffyang.medium.com/running-a-private-hermes-agent-on-a-vps-wired-to-discord-f1057962cc59?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1057962cc59</guid>
            <category><![CDATA[hermes]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 14:31:01 GMT</pubDate>
            <atom:updated>2026-11-09T14:31:01.460Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I4rn5bWrgD9BwOLP2F0YFw.png" /></figure><p><em>2026–11–07</em></p><p><strong>Goal:</strong> private everything. A private server, a private channel, a private app that only you can install. Nobody reaches the bot without an invite you issued and an allowlist entry you made.</p><p><strong>Status:</strong> working — a friend joined the server and held a working conversation with the bot. The bot answers @mentions inline in its channel, and DMs for anyone on the allowlist. This article documents that private variant: owner-only app install, private channel, allowlist.</p><p>Upstream reference: <a href="https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/messaging/discord.md">https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/messaging/discord.md</a></p><h3>Why not Google Chat</h3><p>Discord was the second attempt. The first was Google Chat, and it swallowed two or three days.</p><p>Hermes ships a Google Chat adapter, so the work looked ordinary: a Google Cloud project, the Chat API enabled, a service account, an endpoint. The wall sits underneath. A Chat app needs a Business or Enterprise Google Workspace account — Google’s own quickstart lists that as a prerequisite. A free Gmail account cannot enable the Google Chat API at all, and nothing in the flow says so. The API just stays disabled, in a project you own, with no switch that turns it on.</p><p>The Google Cloud console then added insult. The Chat API dashboard looks like a Microsoft Windows product: panels from different design generations, boxed-in controls everywhere, dead space where a control should be, and no sign that anyone ever looked at the whole screen at once. It is one of the ugliest interfaces I have used, and if a Microsoft team built it, that would explain everything.</p><p>I gave up on Google Chat. Discord took a fraction of the time, and what follows is the setup that worked.</p><p><strong>The leftover.</strong> Giving up does not remove the adapter. It stays configured and reports a dead Pub/Sub subscription on every health check, so turn it off:</p><pre>hermes config set platforms.google_chat.enabled false<br>hermes gateway restart</pre><p>Delete the GOOGLE_CHAT_* lines from ~/.hermes/.env while you are there. If the warning survives the restart, it is not a live failure: gateway status prints the last saved snapshot from ~/.hermes/gateway_state.json. Stop the gateway, move that file aside, start it again.</p><h3>What the pieces are</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*r9F3kHj1XoiPP8bWkiCczQ.png" /></figure><p>The bot token is the only Discord secret. It lives on the VPS and never in the server, the repo, or a chat message. The OpenRouter key is the other one, and it lives in the same place.</p><h3>Part 1 — Hermes on the VPS</h3><p>Run everything as a <strong>normal user</strong>, not root. The installer writes to ~/.hermes/ and ~/.local/bin/.</p><pre># prerequisites + install<br>curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash<br>source ~/.bashrc<br>hermes doctor</pre><h3>The model: Nemotron 3 Ultra, free</h3><p>Nothing on Discord works until the agent itself can answer. This install runs an OpenRouter free variant, so the only cost of a conversation is the daily request cap.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/869/1*qgSGXRLZaSxhMWonQ4XaOw.png" /></figure><pre>hermes config set OPENROUTER_API_KEY sk-or-v1-...<br>hermes config set model.provider openrouter<br>hermes config set model &quot;nvidia/nemotron-3-ultra-550b-a55b:free&quot;<br>hermes config show</pre><p>Check what is left in the day’s budget. hermes config set writes the key to ~/.hermes/.env, and your shell does not read that file, so load it first:</p><pre>set -a; . ~/.hermes/.env; set +a<br>curl -s https://openrouter.ai/api/v1/key \<br>  -H &quot;Authorization: Bearer $OPENROUTER_API_KEY&quot; \<br>  | jq &#39;.data.free_model_daily_requests&#39;<br># {&quot;used&quot;: 12, &quot;limit&quot;: 1000, &quot;remaining&quot;: 988}</pre><p>The cap is one counter across every :free model on the account, so switching to another free model does not restore service. When it is gone, replies stop until the reset.</p><p>Two things worth separating when the bot looks broken: a 429 is a rate limit (the daily cap, or the 20-requests-per-minute free-tier burst), and a slow or refused reply is free-tier priority at peak hours. Neither is a Discord problem.</p><p>Put the gateway under systemd so it survives logout and reboot:</p><pre>hermes gateway install     # user unit: hermes-gateway<br>hermes gateway start<br>hermes gateway status<br><br>sudo loginctl enable-linger $USER   # keep it running after SSH disconnects</pre><p>Follow the log with journalctl --user -u hermes-gateway -f.</p><p>On a headless VPS you can use a boot-time system unit instead: sudo hermes gateway install --system, then --system on start/status and journalctl -u hermes-gateway -f without --user. It also makes every restart need root, including the one at the end of hermes update — the user unit plus linger avoids that.</p><p><strong>Do not</strong> add a ExecStopPost=/bin/kill -9 drop-in. The installed unit already stops cleanly and sets Restart=always; a hard-kill drop-in turns every restart into a respawn loop.</p><p>From here on, every config or .env change takes effect with hermes gateway restart.</p><h3>Part 2 — The Discord application</h3><p>At <a href="https://discord.com/developers/applications">https://discord.com/developers/applications</a> → <strong>New Application</strong>.</p><p><strong>Bot page:</strong></p><ul><li><strong>Public Bot → OFF.</strong> This install is private: with Public Bot off, <strong>only you can add the app to a guild</strong>. Nobody else can authorize it, whatever URL they hold. Set the install link to <strong>None</strong> below first — Discord rejects a private app that keeps a default install link.</li><li><strong>Require OAuth2 Code Grant → OFF.</strong></li><li><strong>Privileged Gateway Intents:</strong> enable <strong>Server Members Intent</strong> and <strong>Message Content Intent</strong>. Message Content is the hard one: with it off Discord refuses the connection and the bot never comes online — the number-one failure. Server Members is what lets Hermes resolve usernames and roles. Save changes.</li><li><strong>Token → Reset Token</strong>, copy it once. It is never shown again; a lost token means resetting it.</li></ul><p>If the app was created earlier with Public Bot on, it is still installable by anyone holding the URL — see <strong>Hardening after the fact</strong> at the end of this part.</p><p><strong>Installation page</strong> (invite link):</p><ul><li><strong>Installation Contexts:</strong> Guild Install.</li><li><strong>Install Link → None.</strong> No “Add App” button, no App Directory listing, no URL to leak. A manual authorize URL still works for you, the owner.</li><li><strong>Scopes:</strong> bot + applications.commands.</li><li><strong>Permissions:</strong> 309237763136 (recommended) — View Channels, Send Messages, Embed Links, Attach Files, Read Message History, Create Public Threads, Send Messages in Threads, Add Reactions. 117760 is the minimum; 309240908864 adds voice.</li></ul><p>Manual URL — keep it to yourself:</p><pre>https://discord.com/oauth2/authorize?client_id=YOUR_APP_ID&amp;scope=bot+applications.commands&amp;permissions=309237763136</pre><p><strong>Install into your server:</strong> open that URL <strong>yourself</strong>, pick the private server, Continue → Authorize. The app is in the server from that moment — <strong>Server Settings → Integrations</strong> lists it. A channel’s member list is a different thing: it only shows members who can view that channel, so the bot appears there after the grant in Part 4, and online after the gateway starts. This is the only time the app is installed; friends never do it. Permissions granted at install time do not grow later — re-open the URL to grant more.</p><p><strong>Hardening after the fact.</strong> An app created with the defaults is public: Public Bot on, a Discord-provided install link, and anyone holding the URL can add it to their own server. Making it private has an order, because Discord refuses the second step while the first is undone:</p><ol><li><strong>Installation</strong> → <strong>Install Link</strong> → <strong>None</strong>, Save Changes.</li><li><strong>Bot</strong> → <strong>Public Bot</strong> → OFF, Save Changes.</li></ol><p>Leave it half-done and the app stays installable — the error Discord gives is <em>Private application cannot have a default authorization link</em>. Neither switch removes the app from a server where it is already installed; they only block new authorisations. After that, install with the manual URL above. If Discord answers <em>Verified apps must be public</em>, the app carries a Verified badge, and verification exists only for bots in 100 or more servers — it should not apply here.</p><h3>Part 3 — Wire Hermes to the bot</h3><p>Fast path, which verifies the token, checks the intents, prints an authorize URL, and allowlists you:</p><pre>hermes gateway setup     # choose Discord, paste the token</pre><p>Manual path — ~/.hermes/.env:</p><pre>DISCORD_BOT_TOKEN=your-bot-token<br>DISCORD_ALLOWED_USERS=284102345871466496</pre><p>DISCORD_ALLOWED_USERS is a comma-separated list of <strong>user IDs</strong> (Discord Settings → Advanced → Developer Mode → right-click a user → Copy User ID). Without it — or DISCORD_ALLOWED_ROLES, or DISCORD_ALLOWED_CHANNELS scoping the bot to specific channels — the gateway denies everyone.</p><p>Then:</p><pre>hermes gateway restart</pre><h3>Part 4 — Giving the bot its own channel</h3><p>Discord offers no CLI for channel creation, so this part is one GUI errand, one permission grant, and one config key.</p><p><strong>Create it.</strong> Right-click the category — or the server name, for a top-level channel — then <strong>Create Channel</strong> → <strong>Text</strong>, named for the bot, e.g. hermes.</p><p><strong>Private or not.</strong> This build makes the channel private: @everyone — the default role every member of the server holds — gets nothing, and visibility is granted per role. Server Settings → Roles → @everyone: turn View Channels off, and turn off the guild-level permissions too. <strong>Create Invite</strong> matters most: a member who can invite can bring in someone you never approved, whatever your allowlist says. Keep Manage Server, Manage Roles, and Manage Channels for yourself. A public channel would work, but then every member sees the bot&#39;s conversations, which defeats the point of a private server.</p><p><strong>Let the bot into a private channel.</strong> <strong>Edit Channel → Permissions → Add members or roles</strong>, pick the app’s role (named after the application, e.g. Hermes, and listed among roles once the app is installed), and grant:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/869/1*CLMBMZMUwNS7pzbaj56H5g.png" /></figure><p>An install-time permission only sets the ceiling; on a private channel this override is what lets the bot in. Friends need View Channel on the same channel from a member role, independently of the bot.</p><p><strong>Pick how the channel behaves.</strong> Copy the channel ID first (Developer Mode → right-click the channel → <strong>Copy Channel ID</strong>), then choose one of three modes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/1*x5bdZRxpqHuvpU1_Oz6Y-g.png" /></figure><pre># ~/.hermes/config.yaml — the mode this install runs<br>discord:<br>  no_thread_channels:<br>    - 123456789012345678   # &lt;channel id&gt; — reply inline instead of opening a thread</pre><p>A mention-free channel already replies inline, so it needs no no_thread_channels. To force one thread per top-level message in one, add free_response_auto_thread: true (it needs discord.auto_thread, which is on by default).</p><pre>hermes gateway restart<br>hermes gateway status</pre><p>The same knobs exist as DISCORD_NO_THREAD_CHANNELS and DISCORD_FREE_RESPONSE_CHANNELS in .env; env wins when both are set.</p><p><strong>From the API</strong> — the REST call needs <strong>Manage Channels</strong>, which permissions=309237763136 does not include, so the bot gets a 403. Create the channel from your own client, or re-invite the app with Manage Channels (309237763136 + 16) first.</p><h3>Part 5 — Letting friends in</h3><p>The app stays yours. A friend never opens the Developer Portal, never installs anything, and never receives a bot URL. Four independent gates decide whether they can talk to it, and every one has to be open. The channel scope is the only one that starts open.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/728/1*87Imf4YWJBxC2YCHoV4iSw.png" /></figure><p>An allowlisted friend reaches the agent’s tool surface on this VPS, not just a chat window — treat the allowlist like a shell account, and see <em>Keeping it private</em> below for what that means.</p><p><strong>Commands are a second surface.</strong> With no admin split configured, every allowlisted user also gets the full slash-command set — including /learn, which writes a skill into ~/.hermes/skills, plus /save, /rollback, and /update. Tools and commands are gated separately: platform_toolsets never sees a slash command. Set allow_admin_from and group_allow_admin_from to your ID, and give guests a short user_allowed_commands. The exact block — and why it belongs in platforms.discord rather than gateway.platforms.discord — is in <a href="https://file+.vscode-resource.vscode-cdn.net/home/jeff/pool/git/negtivSpace/ai-thoughts/docs/261108-hermes-guest-hardening.md">the hardening guide</a>.</p><p>Before the first invite, leave <strong>Community</strong> off — no discovery, no vanity link. Access comes from the roles you grant.</p><ol><li>Create an invite: right-click the server name → <strong>Invite People</strong> → <strong>Edit invite link</strong> → set an expiry and a max number of uses. Send that, and revoke it once they have joined. Server Settings → Invites lists and kills outstanding links.</li><li>They accept and land in the server. They see only the channels their role allows — #hermes if you granted it.</li><li>Collect their <strong>user ID</strong> (Developer Mode → right-click their name → Copy User ID) and append it:</li></ol><pre># ~/.hermes/.env<br>DISCORD_ALLOWED_USERS=284102345871466496,198765432109876543</pre><pre>hermes gateway restart<br>hermes gateway status</pre><ol><li>4. They post in #hermes. The first reply means all four gates are open; a silence means one is still shut, and the troubleshooting table names which.</li></ol><p><strong>Roles instead of a list.</strong> Create a role such as hermes-user, grant it View Channel on #hermes, and set DISCORD_ALLOWED_ROLES=&lt;role-id&gt;. Granting or removing that one role then adds or revokes access with no config push and no restart.</p><p><strong>Keep the bot in one channel.</strong> The channel-scope gate, set separately from the user allowlist:</p><pre># ~/.hermes/.env<br>DISCORD_ALLOWED_CHANNELS=123456789012345678<br>DISCORD_IGNORED_CHANNELS=987654321098765432   # optional hard deny</pre><p>With DISCORD_ALLOWED_CHANNELS set, the bot answers in those channels and in DMs, and nowhere else. DMs do not honour that list — the user allowlist is their only gate.</p><h3>Keeping it private</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/871/1*5LkKu5jmNgaQzLmX4tiV0Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I4rn5bWrgD9BwOLP2F0YFw.png" /></figure><p>What this posture does <strong>not</strong> buy, stated plainly:</p><ul><li>Discord is not end-to-end encrypted. Channel messages, attachments, and DMs sit on Discord’s servers, and copies sit in ~/.hermes/state.db on the VPS.</li><li>Every prompt leaves the VPS for the model provider (here OpenRouter’s free tier) and is subject to that provider’s retention.</li><li>Allowlisting a person hands them the agent’s tool surface on that VPS — terminal, files, browser — <strong>by default</strong>. The Discord privacy settings control <em>who can reach the bot</em>; platform_toolsets and approvals control <em>what it may do</em> once reached. <a href="https://file+.vscode-resource.vscode-cdn.net/home/jeff/pool/git/negtivSpace/ai-thoughts/docs/261108-hermes-guest-hardening.md">The hardening guide</a> is the second half of this one: it takes the write tools off Discord and leaves them in the CLI.</li></ul><h3>How the bot behaves</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/869/1*FAdp1KjQ6P51I9rLYbsrCw.png" /></figure><p>Useful knobs in ~/.hermes/config.yaml: discord.free_response_channels, discord.no_thread_channels, discord.auto_thread, discord.reactions, display.tool_progress.</p><p>Three blocks share the name. discord: holds behaviour — the knobs above. platforms.discord: holds the connection and extra, which is where the admin split goes. gateway.platforms.discord: is another nesting the gateway still reads, shadowed by the top-level block on shared keys. Behaviour keys never belong in extra.</p><h3>Troubleshooting</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/1*fZovQzq-4J8QyjeRgJLpxg.png" /></figure><p>Health checks: hermes gateway status (online state, heartbeat, degraded reason) and journalctl --user -u hermes-gateway -f.</p><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1057962cc59" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Agent, Not the Model]]></title>
            <link>https://j3ffyang.medium.com/the-agent-not-the-model-22dc10e2a354?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/22dc10e2a354</guid>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 15:06:01 GMT</pubDate>
            <atom:updated>2026-11-05T15:06:01.571Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iyGW4AaKdg_NghjXmOyl5Q.png" /></figure><p><em>2026–10–22</em></p><p>Thesis: The model is the least interesting part of the stack. What comes out of a session is decided mostly by the loop an agent runs around the model — and that loop is only as good as the instruction it was given. The real skill is not technical. It is saying precisely what you want.</p><p>This is not a survey. It is one person’s experience after about a year of heavy use, and the only measurements in it are my own.</p><h3>The bottleneck isn’t the model</h3><p>Hundreds, maybe thousands of models now sit in the catalogues: free ones, paid ones, confusing ones. Claude, OpenAI and Google are probably still the top three… maybe. New entrants keep arriving with a chart that says they won something.</p><p>The chart has little to do with my work. I write articles, code, review and analyze code, and keep repositories organized. None of it is very complex; all of it needs to be logical and precise. So: benchmark performance is mostly marketing — a way to attract attention and investment. That is not a claim that benchmarks measure nothing. They measure something. It just isn’t the thing I’m paid for.</p><p>The practical version: a cheaper model does not much degrade the output, for work like mine. DeepSeek V4 Pro costs $0.348 / $0.696 per million input / output tokens; GPT-5.2 costs $1.75 / $14. A twentieth of the output price, for a model that is still frontier-capable. The numbers and the method are in <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/260927-opencode-openrouter-model-selection.md">my model-selection notes</a>.</p><p>That is not a claim that every model is equally good. It is that the gap between models is smaller than the gap in their price, and that my work rarely lands in the narrow zone where the top model decides the outcome. If yours does, this article is not for you, and that’s fine.</p><h3>The other half: the human</h3><p>The human half is easier to see in a librarian than in a benchmark table.</p><p>Ask a good librarian for a reading list and you get one. Say what you already know, why you want it, how much time you have — and you get five books you will actually finish. Say “something about history” and you get a cart of thirty. Every one of them is good. You will read none.</p><p>The librarian is not the problem, and the shelf is not the problem — nobody said how much you can carry, and no librarian can read a book on your behalf.</p><p>A bigger context window is a bigger cart. Whether it grows to 10M tokens does not help me: I still cannot hold it, check it, or decide from it. I become the bottleneck exactly where the machine’s output meets my attention. There is one honest exception: when the output goes to another machine — a script, a summarizer, a next model — the cart is fine, because nobody has to read it. That is a different story from this one.</p><p>So the work does not fail at the model’s edge. It fails at the reader.</p><h3>What the agent actually does</h3><p>The agent is the part that got smart. Not the model — the harness that runs the loop, plus the instructions you hand it.</p><p>A turn looks roughly like this:</p><pre>┌──────────────────────────────────────────────┐<br>│ 1  human asks                                │<br>│ 2  harness builds the prompt                 │<br>│    system + AGENTS.md + history + tools      │<br>│ 3  model answers: text, or a tool call       │<br>│ 4  harness runs the tool, appends the result │<br>│ 5  back to 3, until the model stops asking   │<br>│ 6  human reviews, and decides the next step  │<br>└──────────────────────────────────────────────┘</pre><p>Every turn rebuilds that prompt from scratch: the harness keeps no state the model can see, only what it re-sends.</p><p>The tool call is worth being precise about, because it is the heart of the argument: an agent does not search the web because the model “doesn’t know”. The model does not introspect its own knowledge and dispatch a sub-agent when it finds a gap. It emits a tool call; the harness honours it; the result becomes new context. Searching, editing files, running the tests, spawning a sub-process — all of that is the harness side of the loop, not the model’s cleverness.</p><p>So the quality of what comes out is decided by:</p><ul><li>how the agent digested and, sometimes, rewrote your question</li><li>which tools it had, and whether it reached for them at the right moment</li><li>how it read what came back from a search or a file</li><li>how it organized what it collected, and whether it addressed what you actually asked</li><li>whether it followed up the unfinished parts, or quietly declared victory</li></ul><p>None of those are settings on the model. They are behaviour, and behaviour is written down somewhere — in instructions.</p><h3>The only interface you have is the instruction</h3><p>Working well with an agent comes down to one skill: making the thing in your head legible to the thing on the other side. Precisely. Accurately. Clearly. As few gaps as possible for it to fill in on its own.</p><p>It is a conversation with no partner present: you say what you want once, in writing, and a later session re-sends only what the files hold.</p><p>In practice you do it in three files:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/859/1*b-aXcF-5XZxwq6B2LWSYVw.png" /></figure><p>Three different jobs: what must always be true, where things live, and how one job gets done. When they overlap or contradict each other, the agent picks a reading and acts on it — and you find out later which reading it picked. (Whether they should be written like a maintenance manual is <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261019-ste100-agent-instructions.md">a separate question</a>.)</p><p>I don’t have to look far for the evidence. The repository this article lives in is the argument: AGENTS.md at three levels, .opencode/skills/ for the procedures, a README generated from a YAML manifest, and scripts the agent runs to keep the output honest. When the instruction was vague, the output was vague. When I wrote the constraint down — never commit unasked, no hard-wrap, the box must be square — the behaviour changed, and it stayed changed.</p><h3>A loop, not an arrow</h3><p>So the workflow is not a line. It is a loop:</p><pre>instruction ──▶ agent loop ──▶ artifact<br>      ▲                            │<br>      └──────── you review ────────┘</pre><p>The instruction feeds the agent, the agent produces an artifact, and then you look at it and decide. The review step is the whole reason the top model doesn’t decide your outcome: the artifact comes back to a human either way, and the judgment at that point is yours.</p><p>The review is also where the next instruction comes from. Every rule in my AGENTS.md started as me noticing something in an artifact I did not want to see again.</p><h3>Closing</h3><p>Every item here is about the quality of communication between a human and a machine — a strange thing to be the skill, and yet here we are. The prompt is the program, and the agent is the interpreter. For the practices that go with this argument — what to understand, what to never hand over alone, how to keep the agent in the loop — see <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/260922-what-to-know-work-with-ai.md">the earlier article</a>. That one is the how. This one is the why.</p><p>The rules I actually require of my own team are <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/docs/261104-what-i-tell-my-team-about-ai.md">one page of rules</a>.</p><p>The models will keep changing. The loop won’t.</p><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22dc10e2a354" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MEMORY.md Is a Budget, Not a Document]]></title>
            <link>https://j3ffyang.medium.com/memory-md-is-a-budget-not-a-document-ca8290800291?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca8290800291</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[hermes-agent]]></category>
            <category><![CDATA[open-code]]></category>
            <category><![CDATA[ai-memory]]></category>
            <category><![CDATA[openai-codex]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 14:56:01 GMT</pubDate>
            <atom:updated>2026-11-02T14:56:01.714Z</atom:updated>
            <content:encoded><![CDATA[<p><em>OpenCode v2.2.5, November 2026.</em></p><p>Everyone who wants an agent with memory builds the same thing first: one long file where the agent writes down everything it learns, loaded at the start of every session. It feels obvious, and no harness shipped it that way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_5PnJyiJz4J5xd1VMTglpA.png" /></figure><h3>Background</h3><p>I did not come to this by reading documentation. I came to it by copying a file.</p><p>In both Hermes and OpenClaw, I created MEMORY.md by hand. I wrote the entries myself, edited them when they went stale, and pruned them when the file got long, because that is what the docs described: a file in the agent&#39;s workspace that it reads at the start of a session.</p><p>Then I went looking for the equivalent in OpenCode and found nothing — no memory file, and nothing about memory in the docs. I put a MEMORY.md where I expected it to go and the agent never saw it. That was the first hint that &quot;memory file&quot; is not a feature of the category, but a convention inside some products.</p><p>Claude Code confused me in the other direction. It already had a memory file I never wrote, one it maintains itself, and it was not the document I expected. It was a list of one-line pointers into a directory of topic files. I had been treating memory as a place to write things down. This was a table of contents.</p><p>So I read what the four of them actually do, and wrote it down, because I could not find it stated in one place.</p><p>All four end up with a file literally named MEMORY.md — Claude Code, OpenClaw, Codex, and Hermes. None is a document. Each is a small layer loaded at session start, with the bulk kept elsewhere and fetched only when needed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/878/1*4-gThfQMHd2qylAkkO419Q.png" /></figure><p>Claude Code’s file is a real index — the index of its auto-memory directory, not its instruction file, which is CLAUDE.md. One line per memory, usually ending in a pointer to a topic file. OpenClaw&#39;s is a summary you read — durable facts and decisions stated outright. Codex takes the same idea furthest: it injects only memory_summary.md, a compact overview, and the agent searches MEMORY.md and the supporting files when a task looks related. Different flavours, same constraint. Keep it small or it stops being read.</p><p>The literal reading of that table is a ceiling. Three of them cap the memory file itself; Codex caps the summary instead. Either way it is prompt space being measured — lines, kilobytes, characters, tokens — not the file. So MEMORY.md is not a place where memory lives — it is a budget of prompt space that happens to be backed by a file on disk.</p><p>Which leaves the definition worth carrying: MEMORY.md is not a log, not a journal, and not a database. It is a table of contents with a character budget attached — the only two things a session-start injection can be.</p><h3>The limit is the design</h3><p>Claude Code’s cap has a sharp edge: the write succeeds. The file stays on disk intact, and the error it returns tells the agent to rewrite the index, because everything past the cap is dropped on the next load.</p><p>That is not a limitation to work around. It forces the only split that matters: the small file says what you know about, and the detail files hold what you know. A single document cannot make that split, so it grows until it is 400 lines, gets injected in full every turn, and gets skimmed the way you skim a 400-line file.</p><p>OpenClaw states the same rule as advice rather than an error: if MEMORY.md outgrows its budget, the injected copy is truncated, and that is your signal to move detail into memory/*.md.</p><p>Hermes budgets the same idea in characters: 2,200 for MEMORY.md and 1,375 for USER.md, roughly 800 and 500 tokens, with room for 8–15 entries in the memory file and 5–10 in the profile. Both go in as a <em>frozen snapshot</em> taken once at session start, so a mid-session save reaches the disk immediately but not the prompt until the next session.</p><p>When a write would overflow, its memory tool returns an error and tells the agent to free the stated number of characters — consolidate or delete first, then retry — so nothing is dropped silently. What doesn&#39;t fit needn&#39;t be lost either: every session is kept in SQLite and can be full-text-searched on demand, which is the working layer to MEMORY.md&#39;s curated one.</p><p>That is the difference between a guardrail and a wish. Claude Code and Hermes answer an overrun with an error; OpenClaw truncates and shows you the damage in /context; Codex never loads MEMORY.md at all, only the summary. Where the harness will not stop you, the budget is yours to police, and nothing will tell you the file quietly stopped being read.</p><h3>Why it exists at all</h3><p>Because the context window will be taken from you. The Anthropic memory tool appends this to the system prompt automatically:</p><blockquote><em>ASSUME INTERRUPTION: Your context window might be reset at any moment, so you risk losing any progress that is not recorded in your memory directory.</em></blockquote><p>That is the whole justification. Not “it would be nice to remember” — the window is going to close, and nothing you did not write down survives it.</p><h3>Doing it in OpenCode</h3><p>OpenCode has no memory file. It auto-discovers AGENTS.md, falling back to Claude Code&#39;s CLAUDE.md, and nothing memory-shaped. So you assemble the slot yourself.</p><p>First the rules, in AGENTS.md. Second the small file, injected with one config line in the project&#39;s opencode.json:</p><pre>{<br>  &quot;$schema&quot;: &quot;https://opencode.ai/config.json&quot;,<br>  &quot;instructions&quot;: [&quot;MEMORY.md&quot;]<br>}</pre><p>In the global ~/.config/opencode/opencode.jsonc, use an absolute path instead: the docs do not say what a relative one resolves against.</p><p>Third the detail, in ordinary files read only when something points at them — which is exactly how OpenCode already loads skills. A one-line description of every skill sits in context; the body loads only when the description matches. Same machine, applied to procedures instead of facts. Hermes’ docs reach that conclusion from the other direction, recommending a skill over a memory entry for anything that would compete for the 2,200-character budget.</p><p>The filename is arbitrary. MEMORY.md, memory.md, notes.md — OpenCode does not care, and it will load none of them without that config line. That is where a Claude Code migration quietly breaks: MEMORY.md comes along and does nothing.</p><p>Keep the file an index:</p><pre># Project memory — index only<br><br>One line per entry: `- (YYYY-MM-DD) &lt;fact&gt;. See detail/&lt;file&gt;.md`<br>Delete entries older than six months, or contradicted by the code. Do not archive.<br><br>## project<br>- (2026-11-01) Deploys go to eu-west-1; forced by a 2025-11 data-residency contract. See detail/deploy.md<br>- (2026-10-14) `make test` needs Docker running. Without it the suite exits 0 and tests nothing.<br><br>## feedback<br>- (2026-10-20) Do not reformat YAML I did not ask you to touch. It happened twice.</pre><p>The number is yours to pick. 100 lines is plenty, and well under Claude Code’s 200 — count it with wc -l. Passing it is the signal to move detail into a file.</p><p>If you build only one half, build the read. The usual failure is a rule saying “append your findings to MEMORY.md&quot; with nothing saying &quot;read it first&quot;, which produces a journal nobody opens.</p><p>And the rule that makes it work, also in AGENTS.md, because nothing will enforce it:</p><pre>Before answering a question about project state, deployment, or a past decision,<br>read `MEMORY.md` and open the detail file it points to before stating a fact.<br>Never answer from an index line alone.</pre><p>I built a throwaway one to test this on v2.2.5. The file loads. Asked why a project used eu-west-1 — a question the index cannot answer — the model read the detail file and quoted it. It also re-read MEMORY.md itself, already in context, which is a small tax worth knowing about.</p><h3>Whether to build it</h3><p>The case for it is already made: sessions end. The case against is smaller and more practical.</p><p>An instructions file is injected in full, every session. A 100-line index is roughly a thousand tokens before you have asked anything — cheap, but it loads every time and buys nothing on a day when the answer is already in the code.</p><p>The bigger cost is maintenance. A memory file is a second surface to prune, and it goes stale at a different rate from your rules. Put a fact that changes weekly into AGENTS.md and you are editing the constitution every week.</p><p>So the test is not whether it is good practice. It is whether you are losing facts. Build it when a new session keeps re-deriving something you have already settled. Until then it is a capability you are not using, not a defect in your setup — which is where I left mine: AGENTS.md for the rules, skills for the on-demand half, no facts index, and nothing lost.</p><h3>Where it lives is part of the decision</h3><p>If you build one, choose its home before the first entry. A memory file collects what the code cannot tell you: project names, deploy regions, file paths, decisions, sometimes a customer’s constraint. That is the content you would not paste into a public issue.</p><p>My config is symlinked into a repo that publishes, so a MEMORY.md beside it would have gone out with the next push. Tracking and publishing cannot be separated in a public repo. If you want git history and privacy both, the file needs a private repository, or no repository at all — and that repository is also the only thing that carries it to a new machine.</p><p>And the failure is silent, which is the whole problem: it sits on disk looking authoritative while every session ignores it — the same no-op this article opened with.</p><h3>What it does not do</h3><p>Date every entry. An agent-recorded fact has no auditor; six weeks later it is wrong with the same confidence it was right, and it outranks the code because someone wrote it down. Claude Code at least stamps the write time itself, as a modified field in the memory file&#39;s frontmatter, so the age of a fact is visible to whoever reads it back.</p><p>A lot of what people file under memory is a rule they have not decided yet, or an environment fact that belongs in a config file. A deploy region is memory. “Use tabs, not spaces” is a rule for AGENTS.md. Whether Docker is installed is a question for docker --version — if the machine can answer it, it does not belong in memory.</p><p>Memory is context, not enforcement. Claude Code says so outright, and points at a hook, not a memory entry, when you need to block an action. A memory entry can tell the agent what you decided. It cannot stop the agent doing otherwise.</p><p>Which leaves the honest limit. Memory does not buy you volume: a session that fits everything still ends, and the next one starts empty. Nor does it travel on its own: auto memory is machine-local, so a new laptop starts empty until you carry the file there. What memory buys is a table of contents that survives the reset.</p><h3>Sources</h3><ul><li>Claude Code, <a href="https://code.claude.com/docs/en/memory">How Claude remembers your project</a></li><li>OpenClaw, <a href="https://docs.openclaw.ai/concepts/memory">Memory overview</a></li><li>Codex, <a href="https://developers.openai.com/codex/memories">Memories</a></li><li>Anthropic, <a href="https://platform.claude.com/docs/en/agents-and-tools/tool-use/memory-tool">Memory tool</a></li><li>Hermes, <a href="https://hermes-agent.nousresearch.com/docs/user-guide/features/memory">Persistent Memory</a></li></ul><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca8290800291" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Health Records an Agent Can Audit]]></title>
            <link>https://j3ffyang.medium.com/health-records-an-agent-can-audit-579179464ad6?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/579179464ad6</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[medical]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[audit]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 14:46:01 GMT</pubDate>
            <atom:updated>2026-10-27T14:46:01.486Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A two-layer skill for AI-assisted review of personal and family medical records: a generic procedure, project-local facts, and no clinical authority.</em></p><p>Source: the <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/.opencode/skills/medical-record-review/SKILL.md">medical-record-review</a> skill, first extracted on 21 October 2026 and updated to version 1.5.3 on 27 October 2026, from a private record set that had accumulated several years of lab rounds, prescriptions, imaging, and conflicts. The skill is the how; this article is the why.</p><p>Thesis: the value of an agent in this domain is not cleverer analysis. It is a procedure that refuses to guess — one that stops on conflict, cites what it used, separates record from inference, and cannot silently rewrite a data row.</p><p>Verdict: the evidence discipline transfers, the clinical judgement does not. The clinician’s judgement wins; the skill only structures the evidence. Ship the procedure, keep the facts local, and label the output loudly enough that nobody mistakes it for a doctor.</p><blockquote><em>⚠️ Disclaimer. This is a practice of AI × medical combination, not a medical device and not official medical advice, diagnosis, prescription, or treatment decision. The author assumes no liability for consequences of use or misuse. Read the full statement at the top of the skill before adopting it.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kphqznEZ1VjsQd_vJ7LsBQ.png" /></figure><h3>Why a procedure, not a prompt</h3><p>Medical records do not arrive as a clean dataset. They arrive as a pile of PDFs from several institutions over many years, many of them scans, with the same metric printed under three abbreviations and two units. The interesting failures are not “the model did not know LDL matters”. They are: the model read the wrong row, averaged two institutions that report different reference ranges, or quietly “corrected” a number that looked wrong.</p><p>Each of those failure modes has a fix, and none of the fixes is a better prompt. They are a conflict protocol, an append-only rule, a change gate, and evidence grading. That is what this skill is.</p><p>There is a second constraint. These records are the most sensitive files a person owns, and a hosted model means every file read is a file transmitted, irreversibly. So before the first record is read, an egress pre-check decides whether to work at all — not an analysis technique.</p><h3>Two layers, one authority chain</h3><p>The skill is split because the procedure and the local facts change at different rates.</p><pre>AGENTS.md (AUTHORITY)         ← project constitution: privacy, egress, evidence standard<br>  └─ project-layer skill      ← configuration, precedent table, language conventions<br>       └─ generic skill       ← procedure only, no patient data, no patient paths</pre><ul><li>The generic layer holds the workflow: intake, profile, registry, comparison, conflict protocol, evidence grading, output skeleton, merge and rename, compression, and a Traditional Chinese Medicine (TCM) sub-flow. It carries no patient values, no patient paths, no names — that is what makes it publishable.</li><li>The project layer holds what is true only for one record set: the nine config keys, whether in the project layer or in the AGENTS.md section that AUTHORITY names; the language conventions; and a precedent table of real mistakes and their rulings.</li><li>AGENTS.md outranks both. It is whatever rules file AUTHORITY names — usually the adopting project&#39;s own AGENTS.md, and never an active authority file this procedure ships. A patient-free starter ships beside the skill as <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/.opencode/skills/medical-record-review/references/AGENTS.example.md">references/AGENTS.example.md</a>; copy it, or point AUTHORITY at a rules file your agent already loads. The generic layer points at it for privacy and egress instead of restating them, because a restated rule drifts from its source.</li></ul><p>The split was also forced by a concrete failure. Before it, the generic skill, once specialized for the project, had patient demographics hard-coded into an analysis rule — violating its own “no patient data” clause. Genericising a rule forced the anchor out of the procedure and into configuration, which is where it belonged.</p><h3>The rules that carry the weight</h3><p>Ten rules do most of the work. Each one is anchored to an incident or a review finding, not to good taste. Two of them arrived during the review of this skill: one when an old analysis line could no longer be traced back to the model that wrote it, and one when a derived score could not be reproduced.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/915/1*czaRZBf6xTykXZSMfMQxMQ.png" /></figure><p>Two of these are unusual enough to spell out. Exemption scope exists because “let me clean this up” and “let me fix this number” arrive in the same sentence and deserve different answers. Attribution timeline exists because the most convincing wrong conclusion in the whole record was a symptom attributed to a new treatment that began weeks after the patient had already stopped the previous one.</p><p>The workflow itself is mapped in four phases — ingest, analyse, output, maintain — with a fast-path table beside it. A thirty-second answer may skip ASCII trends, re-reading the entire registry, and optional modules. It may never skip the egress check, the conflict gate, the evidence grading, the one-folder-per-person rule, or the ban on editing a data row.</p><h3>What the skill refuses to do</h3><ul><li>Diagnose, assign a disease subtype, prescribe, or adjust a dose.</li><li>Complete a gap by inference. Missing means missing, and it is named.</li><li>Subtract across institutions, or merge their trends.</li><li>Bulk-replace a filename string across files. A bulk replace once rewrote a historical log line, which is precisely what logs exist to prevent.</li><li>Cite a line number as a location. Section heading plus item name only; line numbers drift. Finding a line with grep -n is a different act.</li><li>Claim the analysis is official, or claim a review that did not happen. Every output is a draft that may be confidently wrong, and no “verified” or “approved” label attaches without a named clinician who read it.</li></ul><h3>Adopting it</h3><p>You need two things before you start: an agent that loads skills from a directory and can read and write files on the machine, and a rules file for it to read. The procedure assumes little beyond that.</p><p>Start in a folder that holds nothing else. No code, no unrelated notes, no second project — apart from the agent’s own skills directory, which is infrastructure, not project content. Put the rules file at its root and the records under it. A mixed folder fails three ways: the agent reads files it should never see, the registry absorbs documents that are not records, and another project’s instructions compete with the privacy rules you just wrote. Keep the folder out of any repository that pushes, or ignore the record directories and the profile directory in git.</p><ul><li>1. Copy the generic skill unchanged into the directory your agent loads skills from — for OpenCode, &lt;project&gt;/.opencode/skills/medical-record-review/SKILL.md. Do not edit the procedure to fit your records; that is what the project layer is for.</li><li>2. Start from the rules template. Copy <a href="https://github.com/j3ffyang/ai-thoughts/blob/main/.opencode/skills/medical-record-review/references/AGENTS.example.md">references/AGENTS.example.md</a>, which ships beside SKILL.md, to &lt;project&gt;/AGENTS.md if none exists; otherwise add the missing sections to the existing file. Fill it in this order:<br>— Set the nine keys first. AUTHORITY points at this file. PROFILE names the profile file. DOCS_DIR and IMGS_DIR name the record folder and the scan folder. REGISTRY names the trend registry, created on the first run. INDEX / INDEX_CMD name the index and its generator; with no generator, mark the index hand-maintained: true. LANG sets the writing language. MODULES enables analysis modules; start with western. UNWRAP_CMD is optional. A required key left unresolved means stop, not default; set the optional keys explicitly too, because the skill&#39;s table defaults DOCS_DIR to docs/, which may be a mixed folder.<br> — Point the profile away from the rules. The template carries rules and configuration only. Age, sex, diagnosis and duration, medication with dosing times, and measurement conditions go in the file that PROFILE names.<br> — Keep the five rule sections intact. Privacy, egress, evidence, measurement conditions, and data-change are the reason the skill can run at all. Do not trim them to save space.<br> — Run the completeness check. The template ends with five items. Fix any failure before the first record is read.<br> — Never create a second rules file. If the project already has AGENTS.md or a CLAUDE.md equivalent, point AUTHORITY at it and add the missing sections instead. Two rule files drift, and drift is what this skill exists to prevent.</li><li>3. Write a profile. Age, sex, diagnosis and duration, medication with dosing times, measurement conditions for weight, and two or three things this person actually cares about. Cite a source record for every value. Keep it free of names.</li><li>4. Start a precedent table. Three columns: what happened, what to do next time, where it came from. Check for an existing rule first; fold the precedent into that rule rather than duplicate it. This is the part that compounds.</li><li>5. Run it once and let it stop. The first real conflict is the test: if the agent reports it instead of resolving it, the skill is working.</li></ul><p>For a family, the rule is one folder per person. Configuration and profile are injected per folder, and folders never reference each other. Cross-person comparison is out of scope for this procedure and leaks context between folders.</p><h3>What the review changed, and what still strains</h3><p>The first version of this skill was reviewed against its own standards. Four gaps closed: two became rules in the table above, one tightened the index contract, and one became a schema in the registry section:</p><ul><li>The index could be hand-maintained, which reintroduced the second-copy risk the skill forbids elsewhere. Now a hand-maintained index must declare hand-maintained: true, and readers re-verify every field against the file.</li><li>Non-metric series had no schema. Symptoms, formulas, and imaging now get minimal columns — onset, attribution status; herb count, doses; modality, follow-up — instead of living wherever the last editor left them.</li><li>Derived scores could be silently miscalculated. Formula, inputs, and unit are now mandatory, and an incomplete one drops to low-confidence.</li><li>Model drift was unrecorded. Every analysis block now ends with provenance: model · date · skill version.</li></ul><p>Three things still strain, and probably always will:</p><ul><li>A disclaimer is not legal advice. Jurisdiction-specific terms are missing, and no lawyer has read the liability language.</li><li>The rules are prose, not a linter. Nothing enforces the provenance stamp or the derived-value triple; both depend on the agent remembering, which is exactly what a skill is supposed to fix.</li><li>The output skeleton carries no mandatory disclaimer. The disclaimer sits at the top of the skill, not in the output skeleton’s seven blocks. A generated review can omit it, so “label the output” remains advice, not an enforced step.</li></ul><h3>What would make the approach not worth using</h3><p>Three structural conditions would make it not worth using.</p><p>If the records are too sparse, the evidence discipline has nothing to stand on and the honest output is “insufficient data” — which users will read as a failure of the tool. If a hosted model is unacceptable, the workflow must run locally or by dictation; either way, throughput and quality drop. And if a clinician never sees the output, the skill’s careful separation of record from inference buys nothing.</p><p>The procedure is worth keeping only while someone acts on its output with a professional in the loop.</p><h3>Files</h3><ul><li><a href="https://github.com/j3ffyang/ai-thoughts/blob/main/.opencode/skills/medical-record-review/SKILL.md">medical-record-review/SKILL.md</a> — the generic procedure: config keys, workflow, evidence grading, appendices for trends, frontmatter, precedents, boundaries, locale profiles, and companion-article alignment. The article explains why, the skill defines how.</li><li><a href="https://github.com/j3ffyang/ai-thoughts/blob/main/.opencode/skills/medical-record-review/references/AGENTS.example.md">references/AGENTS.example.md</a> — the patient-free AUTHORITY starter: privacy red lines, egress policy, evidence standard, data-change rules, measurement conditions, and the nine keys, with placeholders only.</li><li>the project layer — not published. It is a small skill that supplies, or points to, the nine keys, the language conventions, the precedent table, and one profile per person.</li><li><a href="https://github.com/j3ffyang/ai-thoughts/blob/main/LICENSE">LICENSE</a> — MIT. Copy the procedure, keep the copyright notice.</li></ul><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=579179464ad6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Replace Cursor with a Minimal VSCodium AppImage]]></title>
            <link>https://j3ffyang.medium.com/replace-cursor-with-a-minimal-vscodium-appimage-3c919e8bfa1f?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c919e8bfa1f</guid>
            <category><![CDATA[vscode]]></category>
            <category><![CDATA[cursor-editor]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[vscodium]]></category>
            <category><![CDATA[open-code]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 14:51:01 GMT</pubDate>
            <atom:updated>2026-10-26T14:51:01.546Z</atom:updated>
            <content:encoded><![CDATA[<p>Goal: an editor that does 3 things and little else, self-contained, deletable in one step.</p><p>Verdict: VSCodium AppImage + a minimal Open VSX extension set covers all 3 functions with zero pacman/AUR additions, no account, and no forced agent panel. Chosen over a system package for the “delete it and it’s gone” model.</p><p>The 3 things it must do: (1) a file tree with git status tags; (2) view/edit source with Vim keys; (3) render Markdown with Mermaid + PlantUML diagrams. Everything else is deliberately out of scope — details are in “Mapping the 3 functions to tools” below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kTP9m1TTuE1uznHLQkbuKQ.png" /></figure><h3>Why leave Cursor</h3><p>The two things I dislike most, and everything else is downstream of them:</p><ol><li>Forced agent on launch — the agent is Cursor’s centre of gravity, so it opens itself; there is no “editor only” default landing state.</li><li>Telemetry &amp; privacy exposure — the editor is a data path by design (expanded in the next section).</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/837/1*ReCjiO9tGJzFh5audhtCuQ.png" /></figure><h3>Telemetry &amp; end-user privacy</h3><p>The structural difference: Cursor’s editor <em>is</em> a data path; VSCodium’s editor is not.</p><h3>VSCodium — telemetry crippled at build time, and its settings off by default</h3><p>Slightly weaker than “removed,” and VSCodium says so itself: they <em>“do not pass the telemetry build flags and go out of our way to cripple the baked-in telemetry,”</em> but warn that <em>“Microsoft still can track usage by default depending on your settings.”</em> The guarantee is a build without the MS telemetry flags plus these preferences, which VSCodium ships off by default (per its docs/telemetry.md):</p><pre>telemetry.telemetryLevel<br>telemetry.enableCrashReporter<br>telemetry.enableTelemetry<br>telemetry.editStats.enabled<br>workbench.enableExperiments<br>workbench.settings.enableNaturalLanguageSearch<br>workbench.commandPalette.experimental.enableNaturalLanguageSearch</pre><p>Microsoft’s <em>online services</em> are replaced, not merely disabled:</p><ul><li>Update check → updateUrl is repointed at the project&#39;s GitHub page, and on Linux the app update service is disabled at build time regardless of update.mode.</li><li>Extension gallery → replaced by Open VSX (Eclipse Foundation); no Microsoft marketplace calls.</li><li>Welcome page announcements are fetched from GitHub (workbench.welcomePage.extraAnnouncements: false to stop it).</li><li>The malicious/deprecated-extension list is fetched from an EclipseFdn URL (extensions.excludeUnsafes: false to stop it, at a safety cost).</li></ul><p>Remaining network surface is the extension-control list fetch above, plus (a) opt-in metadata and (b) whatever the extensions themselves do. PlantUML renders locally here — MPE’s builtin plantuml.jar + system java, no server — but two honest caveats: VSCodium can&#39;t police extension telemetry (some extensions (e.g. Microsoft&#39;s C#) phone home, and that is the author&#39;s choice), and MPE&#39;s preview webview can fetch assets from cdn.jsdelivr.net (a fetch for JS/CSS, not code egress). Audit what actually connects:</p><pre># while VSCodium is idle, expect zero established connections<br>ss -tpn | grep -i codium<br># or, per-socket, watch during extension install<br>lsof -i -a -c VSCodium</pre><p>No account either: VSCodium never asks you to log in to launch, edit, or install extensions — <em>“An account authentication occurs only when an extension is asking for it”</em> (docs/accounts-authentication.md). Its only sign-in is opt-in and token-based (Sign in with GitHub, needed only when an extension like GitLens requires it): no account wall, unlike Cursor&#39;s launch flow.</p><h3>Cursor — AI-first, so all traffic routes through its backend by design</h3><p>From Cursor’s own <em>Data Use &amp; Privacy Overview</em> (updated 2026–09–03):</p><ul><li>Privacy Mode off (a toggle, not the floor): <em>“we may use and store codebase data, prompts, editor actions, code snippets … to improve our AI features and train our models.”</em> With it on, training stops (zero-data-retention agreements with providers) — but risk/abuse classifiers may still run and, if triggered, <em>“your data may be stored for investigation.”</em></li><li>Your own API key does not bypass them: <em>“Even if you use your API key, your requests will still go through our backend! That’s where we do our final prompt building.”</em></li><li>File contents are cached on Cursor’s servers (encrypted, transient) to cut latency, and the client calls Cursor backend domains for API, update, and marketplace.</li></ul><p>You can opt out of <em>training</em> with Privacy Mode, but you cannot opt out of <em>routing — </em>the editor and the agent are one product, and the product is the backend.</p><h3>Net effect</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/838/1*x5S8-noUYXw6shh78Bib2w.png" /></figure><p>So dropping Cursor removes a standing data path you don’t control; with VSCodium the editor is inert and the only egress is OpenCode, invoked deliberately against a provider you chose.</p><h3>Mapping the 3 functions to tools</h3><ol><li>File navigator with git status tags → Explorer + the bundled git extension (git.decorations.enabled). No extra dependency — git is already installed.</li><li>View / edit source, including Markdown → built-in editor + Markdown grammar, driven by Vim keys (vscodevim.vim).</li><li>Markdown rendering + mermaid + plantuml diagrams → Mermaid is built into the default preview (shipped as the built-in mermaid-markdown-features extension); PlantUML comes from Markdown Preview Enhanced (shd101wyy.markdown-preview-enhanced), rendered locally via its builtin plantuml.jar + system java. MPE renders Mermaid too, so either preview works — the settings notes below cover the one MPE default that otherwise hides the built-in preview buttons.</li></ol><h3>Verified environment (read-only checks)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/844/1*PI_FuhZ0E-n5AFe9zGivRg.png" /></figure><h3>Install — AppImage, checksum verified</h3><p>Non-root, no pacman. AppImage lives outside the system package DB; removing it leaves no footprint except the config dir. Chosen over the AUR vscodium-bin / extra/code (drag in pacman deps) and Flatpak (drags in a whole runtime) precisely for that: self-contained and delete-and-go.</p><pre>mkdir -p ~/.local/bin<br>cd ~/.local/bin</pre><pre>VER=1.140.06880<br>BASE=https://github.com/VSCodium/vscodium/releases/download/$VER<br>curl -fLO &quot;$BASE/VSCodium-$VER-anylinux-x86_64.AppImage&quot;<br>curl -fLO &quot;$BASE/VSCodium-$VER-anylinux-x86_64.AppImage.sha256&quot;</pre><pre>sha256sum -c &quot;VSCodium-$VER-anylinux-x86_64.AppImage.sha256&quot;</pre><pre>chmod +x &quot;VSCodium-$VER-anylinux-x86_64.AppImage&quot;<br>ln -sf &quot;VSCodium-$VER-anylinux-x86_64.AppImage&quot; VSCodium.AppImage   # stable name, survives version bumps</pre><p>Note: AppImage anylinux builds work as-is on Arch&#39;s glibc; if one ever fails to self-mount, run it with --appimage-extract-and-run (no FUSE needed). FUSE is present here, so not expected.</p><p>Updating: download the new AppImage and re-point VSCodium.AppImage at it (ln -sf &quot;VSCodium-$VER-anylinux-x86_64.AppImage&quot; VSCodium.AppImage); nothing self-updates, and extensions refresh only on demand.</p><h3>Extensions (Open VSX, no Microsoft account)</h3><p>Open VSX is the Eclipse Foundation’s vendor-neutral extension registry — the free stand-in for Microsoft’s Visual Studio Marketplace, which is licensed for official VS Code only (so OSS builds like VSCodium can’t use it). VSCodium’s gallery points at it by default, so no marketplace flag or Microsoft account is required. The set is deliberately three — nothing the built-in core already does is re-added. Install counts are pulled from the Open VSX API (re-check any of them, e.g. curl -s https://open-vsx.org/api/shd101wyy/markdown-preview-enhanced | jq .downloadCount):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/844/1*4LorQSp6JOt5VwpGHMDDUA.png" /></figure><pre>V=~/.local/bin/VSCodium.AppImage<br><br>$V --install-extension vscodevim.vim<br>$V --install-extension shd101wyy.markdown-preview-enhanced<br>$V --install-extension yzane.markdown-pdf</pre><ul><li>Vim (vscodevim.vim) emulates Vim in-process; useNeovim is off by default, so no system vim/neovim dependency is added.</li><li>Markdown PDF (yzane.markdown-pdf) is an add-on beyond the three functions (export only); it renders through a Chromium via puppeteer-core, not the editor&#39;s Electron.</li><li>Mermaid is built in — no extension needed. VSCodium ships a built-in mermaid-markdown-features extension (verified inside the AppImage), so the default preview renders ```mermaid fences out of the box. bierner.markdown-mermaid was therefore dropped.</li><li>PlantUML via MPE, rendered locally. MPE renders it on-machine through a bundled plantuml.jar + your jre-openjdk 27, with no server. This replaces jebbs.plantuml, which was dropped.</li><li>CDN caveat (honest): MPE’s preview webview can load libraries from jsdelivrCdnHost (default cdn.jsdelivr.net) — a network <em>fetch for assets</em> when previewing, not code egress. For a fully offline preview you would have to vendor those assets; otherwise expect CDN calls.</li></ul><h3>Minimal settings.json</h3><p>Write to ~/.config/VSCodium/User/settings.json — nothing touches ~/.config/Cursor/. Everything else stays at defaults.</p><pre>{<br>  &quot;update.mode&quot;: &quot;none&quot;,<br>  &quot;extensions.autoCheckUpdates&quot;: false,<br>  &quot;extensions.autoUpdate&quot;: false,<br>  &quot;telemetry.telemetryLevel&quot;: &quot;off&quot;,<br>  &quot;telemetry.feedback.enabled&quot;: false,<br>  &quot;workbench.startupEditor&quot;: &quot;none&quot;,<br>  &quot;workbench.tips.enabled&quot;: false,<br>  &quot;workbench.enableExperiments&quot;: false,<br>  &quot;workbench.welcomePage.extraAnnouncements&quot;: false,<br>  &quot;workbench.settings.enableNaturalLanguageSearch&quot;: false,<br>  &quot;workbench.commandPalette.experimental.enableNaturalLanguageSearch&quot;: false,<br>  &quot;window.restoreWindows&quot;: &quot;none&quot;,<br>  &quot;editor.minimap.enabled&quot;: false,<br>  &quot;git.enabled&quot;: true,<br>  &quot;git.decorations.enabled&quot;: true,<br>  &quot;vim.useSystemClipboard&quot;: true,<br>  &quot;markdown-pdf.executablePath&quot;: &quot;/usr/bin/chromium&quot;,<br>  &quot;markdown-preview-enhanced.plantumlServer&quot;: &quot;&quot;,<br>  &quot;markdown-preview-enhanced.plantumlJarPath&quot;: &quot;/home/jeff/pool/Downloads/rog_drive/plantuml.jar&quot;,<br>  &quot;markdown-preview-enhanced.hideDefaultVSCodeMarkdownPreviewButtons&quot;: false<br>}</pre><ul><li>git.decorations.enabled gives the file-tree tags requested in function (1) — one letter per status (M/A/U), coloured, with no extra extension.</li><li>window.restoreWindows: none + workbench.startupEditor: none pin the landing state to Explorer + editor, so nothing (no agent pane, no restored panel) ever auto-opens.</li><li>vim.useSystemClipboard shares the system clipboard with Vim mode.</li><li>The MPE keys keep PlantUML local: empty plantumlServer = use the builtin jar, whose own description says <em>&quot;</em><em>java is required in system path&quot;</em>; plantumlJarPath points at your own jar (drop it to use MPE&#39;s bundled one); hideDefaultVSCodeMarkdownPreviewButtons: false keeps the built-in preview&#39;s button next to MPE&#39;s (its default true hides it).</li><li>markdown-pdf.executablePath pins the PDF exporter to the system Chromium, so puppeteer-core never auto-downloads Chrome.</li><li>The telemetry keys match VSCodium’s privacy defaults, except telemetry.feedback.enabled and workbench.welcomePage.extraAnnouncements, which VSCodium leaves on. The rest are deliberate departures from upstream defaults: update.mode, the two extension-update keys, workbench.startupEditor, workbench.tips.enabled, window.restoreWindows, editor.minimap.enabled, vim.useSystemClipboard, markdown-pdf.executablePath, plantumlJarPath, and hideDefaultVSCodeMarkdownPreviewButtons. Listing them makes the intent explicit and survives a settings import.</li></ul><h3>OpenCode outside the editor</h3><p>No extension, no integration. Run OpenCode in your own terminal emulator — here, Kitty, completely outside VSCodium:</p><pre>opencode</pre><p>VSCodium never opens an agent pane on its own, and OpenCode never runs inside it. This is the whole point of dropping Cursor: the editor and the agent are separate processes.</p><h3>Optional — launcher entry and $PATH</h3><p>Only if you want a menu/keybind launch. ~/.local/share/applications/VSCodium.desktop:</p><pre>[Desktop Entry]<br>Type=Application<br>Name=VSCodium<br>Exec=/home/jeff/.local/bin/VSCodium.AppImage %F<br>Terminal=false<br>Categories=Utility;TextEditor;Development;<br>MimeType=text/plain;text/markdown;inode/directory;</pre><p>The AppImage installs no icon, so omit Icon= (as above) or extract the real one with $V --appimage-extract &#39;code.svg&#39; — the icon is code.svg (also .DirIcon), and *.png pulls 170+ unrelated UI icons, not the app icon. Point Icon= at the extracted SVG. For shell use: ln -s &quot;$V&quot; ~/.local/bin/codium.</p><h3>What this is NOT (honest tradeoff)</h3><ul><li>It is still Electron — and the win is behaviour, not size. Extracted on disk ≈ 795 MiB (du), ~10 processes — not a Unix-frugal program. VSCodium is <em>Cursor minus the AI, telemetry, update server, and forced agent</em>: the AppImage is only ~34% smaller than Cursor&#39;s, and the extracted tree is the real cost (see Footprint below).</li><li>If Electron ever grates, the terminal-native fallback is nvim + gitsigns/neo-tree + render-markdown.nvim + plantuml.jar — lighter, but N plugins and no browser-grade preview.</li></ul><h3>Footprint</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/723/1*vUJ4_t7rZDI1iviYxd01oA.png" /></figure><h3>Rollback / uninstall</h3><p>Everything is contained:</p><pre>rm -f ~/.local/bin/VSCodium.AppImage ~/.local/bin/VSCodium-*.AppImage*<br>rm -rf ~/.config/VSCodium ~/.vscode-oss ~/.cache/VSCodium ~/.local/share/VSCodium<br>rm -f ~/.local/bin/codium ~/.local/share/applications/VSCodium.desktop</pre><p>No pacman packages were added; nothing else references it.</p><h3>If you are also retiring Cursor (manual)</h3><p>Cursor is removed by hand and is separate from the VSCodium uninstall. Delete the app and any launcher, then the profile it leaves behind (see Footprint):</p><pre># adjust the AppImage path to wherever you kept Cursor<br>rm -f ~/Downloads/Cursor-*.AppImage ~/.local/bin/cursor ~/.local/share/applications/cursor.desktop<br>rm -rf ~/.config/Cursor ~/.cursor<br>ls -d ~/.config/Cursor ~/.cursor 2&gt;/dev/null || echo &quot;Cursor purged&quot;</pre><p>No pacman packages are involved either way — Cursor was never in the system package DB.</p><h3>Verification checklist</h3><ul><li>sha256sum -c passes before first run</li><li>Explorer shows coloured M/A/U git tags in a dirty repo</li><li>Opening a .md file and pressing Ctrl+Shift+V renders Markdown</li><li>```mermaid block renders in the built-in preview (no extension)</li><li>```plantuml renders in MPE&#39;s preview with no PlantUML-server connection (ss -tpn | grep -i codium; Wi-Fi-off is <em>not</em> a clean test, since MPE&#39;s webview also fetches JS/CSS from cdn.jsdelivr.net)</li><li>No update prompt after 24 h</li><li>No login/sign-in prompt on launch or during extension install</li><li>opencode runs in Kitty, outside VSCodium</li><li>While idle, ss -tpn | grep -i codium shows no established connections</li></ul><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c919e8bfa1f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where Does Brave’s Leo AI Keep Your Data?]]></title>
            <link>https://j3ffyang.medium.com/where-does-braves-leo-ai-keep-your-data-7b8446f38cb5?source=rss-78472ebe2e19------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b8446f38cb5</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[privacy]]></category>
            <category><![CDATA[brave-browser]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[open-source]]></category>
            <dc:creator><![CDATA[j3ffyang]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 15:01:05 GMT</pubDate>
            <atom:updated>2026-10-23T15:01:05.628Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NB6JW89XgQBtqL54bPmKvg.png" /></figure><p><em>Arch Linux. Opening </em><em>AIChat — the one SQLite file behind Brave&#39;s Leo sidebar — to see where the conversations, the pages it read, and the search results actually live.</em></p><p>Sometimes I use Brave’s Leo AI for nothing more than reading the page in front of me: highlight, ask Leo to summarize, close the sidebar, move on. It is the laziest kind of AI use — no account, no chat to save, no prompt engineering. That casualness is exactly what made me curious. If I never “save” anything, where does it go? Is my little pile of questions, pages, and answers sitting on Brave’s servers, or is it on my own disk? So I went looking, in my own profile, without cloning or installing anything.</p><h3>One file, on my own disk</h3><p>The answer is my own disk, and it is almost anticlimactic. Everything Leo keeps about your conversations lives in a single SQLite database inside the browser profile:</p><pre>~/.config/BraveSoftware/Brave-Browser/Default/AIChat<br>~/.config/BraveSoftware/Brave-Browser/Default/AIChat-journal</pre><p>At the moment I looked, AIChat was a little over 300 KB, and file called it a &quot;SQLite 3.x database.&quot; The -journal file is the rollback journal: it sits at zero bytes while the browser is idle and grows only while a write is in flight. There is no leo/ folder, no per-conversation JSON, no synced document. It is one database.</p><p>That single fact is the whole article in miniature: your AI history is as local, and as portable, as any other browser file.</p><p>One exception worth knowing: the facts Leo is asked to remember across chats are not in this file at all. They are appended as plain strings to brave.ai_chat.user_memories in the profile&#39;s Preferences — unencrypted JSON, no v10, no key required. That list is empty on my profile, but the setting ships on by default. Conversations live in the database; memories live in a preference.</p><h3>What is actually in there</h3><p>Brave had the file locked, so I opened a copy. The schema reads like a plain-English description of what that sidebar does:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*ugElCtAbjJOtXvHXofRa1A.png" /></figure><p>On my machine, when I looked: about a dozen conversations and a few dozen turns, a dozen associated pages, a handful of search and inline-search events, and zero uploaded files.</p><p>So “the search results” are not one thing. The queries Leo ran are in search_queries; the sources it found are in web_sources; and the page it summarized for me is in associated_content, right next to content_used_percentage — how much of the page the model actually read. That last column is quietly useful: it tells you whether the summary saw the whole page or only part of it. On my pages that column reads 100 across the board — none of them hit the limit; it drops below that only when Leo truncates a long page to fit the model&#39;s context.</p><h3>Local storage is not local processing</h3><p>Before I get sentimental about privacy, one caveat: stored locally is not the same as processed locally. The history, the pages, and the answers sit on disk — but the prompt and the page text still travel to Brave’s servers, and on to the third-party model provider serving the model you picked (Mixtral, Anthropic’s Claude, Meta’s Llama), while the search queries go to Brave Search. Brave’s own page is explicit: the browser shares “your latest prompt, the context of your current conversation, and — when the use case requires it — the necessary context from the page you’re viewing.”</p><p>If that bothers you — and it should, if data security and privacy are the priority — Leo can run a local model instead, via BYOM (Bring Your Own Model): Ollama on your own machine, for example. That is the way to keep the entire conversation on your local disk with nothing sent outside — the prompt never leaves, the page context never leaves, and the answer is generated on your own hardware. The local database is the receipt of what you did; a local model is what makes the receipt the whole story.</p><h3>The “encryption” is real — the protection may not be</h3><p>Every message and page column — titles, prompts, selected text, completions, page URLs and content, search queries and sources — is a BLOB that begins with the bytes v10:</p><pre>76 31 30 ...   # &quot;v10&quot; + AES-128-CBC ciphertext</pre><p>The structural columns are the exception: uuids, dates (Windows FILETIME — 100-nanosecond ticks, not Unix time), token counts, and — worth knowing — model_key are stored in cleartext, so the shape of your history and which model actually answered are readable without any key.</p><p>That v10 is Chromium&#39;s OSCrypt — and Brave, being Chromium-based, inherits it. It is AES-128-CBC, the same mechanism that protects saved cookies and passwords. Normally the key comes from the system keyring (libsecret, GNOME Keyring, KWallet). But look at my machine:</p><ul><li>Local State&#39;s os_crypt.portal reports prev_init_success: false.</li><li>secret-tool reports no Secret Service to talk to.</li></ul><p>When a keyring is unavailable, Chromium falls back to a hardcoded key: the v10 prefix is the fallback provider, whose key derives from the string peanuts, while v11 marks the keyring-backed provider. On a machine with no Secret Service like mine, the fallback is what you get — so the Leo database is &quot;encrypted&quot; in the sense that a text editor shows garbage, but not in the sense that it is safe: anyone who obtains a copy of the file can read it, with no login, no keyring, and no key. I decrypted my own store with that public key while writing this; titles, page URLs, and every search query Leo ran came back in clear text. It is obfuscation wearing the word <em>encrypted</em>.</p><p>The lesson is the same one I keep relearning: <em>encrypted</em> means nothing without knowing where the key lives. On a desktop with a working keyring, Leo’s history is protected against another OS user and against a stolen disk — though anything already running as you can still ask the keyring, or drive the browser, to decrypt it. On a machine like mine it is exposed to anyone who copies the file.</p><p>Two controls make most of this moot, and both are in the settings: temporary chats are never written to AIChat at all — the code refuses to persist them — and local chat storage can be switched off entirely. Worth knowing before you type anything into that sidebar.</p><h3>What else this could be used for</h3><p>This is the part that turned a curiosity into a small project. A local SQLite database is not just where data collects; it is a resource waiting for a use.</p><ul><li>Back it up. One file to copy — but copy it cleanly: sqlite3 AIChat &quot;.backup leo.db&quot; with the browser open, or just close the browser first, so you never catch a write mid-flight. On a setup like mine the copy is readable anywhere; on a keyring-backed profile the key stays in the keyring, so a reinstall without the old keyring leaves the backup as opaque bytes. Export the key too, or a backup is not a backup.</li><li>Read your own attention. Because it is SQL, this is a query about what you asked AI to read, and how much of it the model digested — with one catch: url is itself a v10 blob, so raw SQL hands back ciphertext. The honest version queries the cleartext columns:</li></ul><pre>SELECT content_type, content_used_percentage FROM associated_content;</pre><p>To get the URLs back you have to decrypt the column first — on a peanuts setup, a few lines of Python with the same public key.</p><ul><li>Audit what left the machine. The queries and sources are recorded. If you care what you asked and where it went, the receipt is already written.</li><li>Build a personal knowledge base. Conversations plus the pages they touched are a reading journal — structured, timestamped, and queryable. Feed it to a local model, or reduce it into notes.</li><li>Sync it on your own terms. That ai_chat_sync_metadata table exists because Leo can sync; but a sync chain moves the data too. Since it is one database, you can sync it your way — an encrypted folder, a git-ignored directory — instead of only through Brave.</li><li>Keep the loop offline. With a local model, the same database becomes a fully offline cycle: prompt, page, answer, all on disk, nothing out.</li></ul><p>None of this needs a new feature from Brave. It needs the data to be local and structured — and it already is.</p><h3>Would this work for a whole team? I don’t think so</h3><p>Everything above is one user on one machine. The obvious next question is corporate: if a whole team uses Brave and Leo, is it worth collecting every conversation for commercial and business use — setting privacy aside on the assumption that all of it is team business anyway?</p><p>My answer, having looked at the file, is no: the assumption is the whole argument, and it does not hold. Leo stores the prompt and any selected text, the full page content it summarized in associated_content, and any uploaded files with their extracted text. People paste credentials, customer data, HR and legal matters, and personal questions into that sidebar. What you would be collecting is not &quot;team business&quot; — it is incidental personal and third-party content promoted into a company record, which then becomes your regulated data to retain, produce on request, and defend after a breach.</p><p>There is also a practical wall: the data is not centralized. It lives in each user’s own profile, and Brave does not retain or share it, so there is nothing to switch on. To collect it you would either sweep every machine (a management policy or MDM) or route Leo through a company-run BYOM gateway — a self-hosted OpenAI-compatible endpoint — and log there. The gateway is the only defensible collection point, and it is an engineering project, not a setting.</p><p>And even if you build it, the legal and cultural bill is real: employee monitoring needs disclosure and, in many places, a lawful basis plus a DPIA or a works-council agreement; collecting customer or partner content can breach the very NDAs that content arrived under; and the stored answers are model output — hallucinations — so treating them as authoritative business records is a mistake. Then there is the trust cost: people self-censor, or move to personal devices, and you end up seeing less, not more.</p><p>Where the value actually lives is the derived, minimized layer — not the raw transcripts: metadata and spend, a few voluntarily contributed prompts and answers, risk signals like leaked secrets or unsafe uploads, and reusable templates. Collect that, on disclosure, minimized, with a retention clock — not everyone’s AIChat.</p><p>So: locally, this is a personal record worth keeping. As a company-wide harvest — everything treated as team business, privacy waved away — I do not think it is worth it. The artifact is wrong, the assumption is wrong, and the cheapest part, storing it, is the least of the cost.</p><h3>Closing</h3><p>What started as a lazy question — <em>where does this go?</em> — ended where most of my terminal curiosity ends: a file I can open, a schema I can read, and a habit sharpened. Leo is not a black box to me anymore. It is one SQLite file, a handful of tables, and an “encryption” whose strength I can now measure. The sidebar gives me less than the word <em>encrypted</em> implies, and more than I expected — because the whole thing is just data, on my own disk.</p><h3>Sources</h3><ul><li>Brave, <em>Brave Leo AI</em> — features, local chat history, BYOM, and exactly what data is sent to Leo’s servers — <a href="https://brave.com/leo/">https://brave.com/leo/</a></li><li>Chromium, components/os_crypt — the OSCrypt implementation that encrypts saved cookies, passwords, and this database — <a href="https://chromium.googlesource.com/chromium/src/+/refs/heads/main/components/os_crypt/">https://chromium.googlesource.com/chromium/src/+/refs/heads/main/components/os_crypt/</a></li><li>Chromium, posix_key_provider.cc — the v10 fallback provider and its hardcoded peanuts key — <a href="https://chromium.googlesource.com/chromium/src/+/refs/heads/main/components/os_crypt/async/browser/posix_key_provider.cc">https://chromium.googlesource.com/chromium/src/+/refs/heads/main/components/os_crypt/async/browser/posix_key_provider.cc</a></li><li>Chromium, freedesktop_secret_key_provider.cc — the v11 keyring-backed provider — <a href="https://chromium.googlesource.com/chromium/src/+/refs/heads/main/components/os_crypt/async/browser/freedesktop_secret_key_provider.cc">https://chromium.googlesource.com/chromium/src/+/refs/heads/main/components/os_crypt/async/browser/freedesktop_secret_key_provider.cc</a></li></ul><p>btw, i use arch</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b8446f38cb5" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>