<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Sam Stull</title>
    <description>The latest articles on DEV Community by Sam Stull (@samstull9007).</description>
    <link>https://dev.to/samstull9007</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4207497%2F4b85a085-0838-4d78-b526-5418ff578cb8.png</url>
      <title>DEV Community: Sam Stull</title>
      <link>https://dev.to/samstull9007</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/samstull9007"/>
    <language>en</language>
    <item>
      <title>Limitcast weekly update, 13 November: Codex and Grok catch up</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Sat, 14 Nov 2026 02:33:03 +0000</pubDate>
      <link>https://dev.to/samstull9007/limitcast-weekly-update-13-november-codex-and-grok-catch-up-2lm9</link>
      <guid>https://dev.to/samstull9007/limitcast-weekly-update-13-november-codex-and-grok-catch-up-2lm9</guid>
      <description>&lt;p&gt;&lt;a href="https://limitcast.dev/sam-stull/" rel="noopener noreferrer"&gt;Sam Stull&lt;/a&gt; · 13 November 2026&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Limitcast 0.19.4 is on npm and GitHub.&lt;/strong&gt; It and 0.19.3 bring the optimizer's pruning proposals, subagent receipts and alerts to Codex and Grok, and fix a Grok pruning defect in 0.19.3. Both were checked in live Codex and Grok sessions.&lt;/p&gt;

&lt;h2&gt;
  
  
  What shipped
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;New this week: Codex and Grok get the same optimizer, receipts and alerts as Claude Code.&lt;/strong&gt; Two releases, &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.19.3" rel="noopener noreferrer"&gt;0.19.3&lt;/a&gt; and &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.19.4" rel="noopener noreferrer"&gt;0.19.4&lt;/a&gt; (both on npm and GitHub, Nov 13 US evening), came out of an audit of why the three tools were handled differently. Claude Code was built first and many features read Claude-only data. The optimizer itself arrived in &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.19.0" rel="noopener noreferrer"&gt;0.19.0&lt;/a&gt; (Nov 9, GitHub only; that version never reached npm) and reached npm in &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.19.1" rel="noopener noreferrer"&gt;0.19.1&lt;/a&gt;. Seven releases have shipped since last week's update.&lt;/p&gt;

&lt;p&gt;What changed for a Codex or Grok user:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Pruning proposals reach Codex and Grok.&lt;/strong&gt; The optimizer proposes MCP servers that no session in this repository called, and &lt;code&gt;optimize capabilities --codex|--grok --apply &amp;lt;server&amp;gt;&lt;/code&gt; turns one off in the project's own config file (journaled, undoable, checked against &lt;code&gt;codex mcp list&lt;/code&gt; / &lt;code&gt;grok mcp list&lt;/code&gt; afterwards). Codex also keeps its three built-in candidates. Grok has no persistent setting that removes a built-in tool, so it proposes servers only.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A correction.&lt;/strong&gt; In 0.19.3 the Grok apply never took effect: Grok ignores a project server entry that has no command, so the command reported an error and rolled back (nothing was left changed). 0.19.4 writes an entry with a stub command that never runs, and copies nothing from your server's definition. Checked with the real CLI: &lt;code&gt;grok mcp list&lt;/code&gt; shows the server off and a headless Grok session does not start it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Subagent receipts for Codex and Grok&lt;/strong&gt;, in API-equivalent dollars (list prices, not a bill). Codex's is computed when the subagent stops. Grok writes a subagent's usage just after its stop hook, so its receipt is logged by the next hook.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fat-tool notice, fat-tool limit and starting-context alert for Codex and Grok&lt;/strong&gt; (&lt;code&gt;install --alerts&lt;/code&gt;), and &lt;strong&gt;Codex limit notifications&lt;/strong&gt; (&lt;code&gt;install --guard&lt;/code&gt;, then &lt;code&gt;notify on&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fixes the audit turned up:&lt;/strong&gt; a model pinned for Grok was applied to Claude spawns; the duplicate-spawn hold never fired for Codex (and Codex encrypts &lt;code&gt;spawn_agent&lt;/code&gt; arguments, so it now compares the task name); &lt;code&gt;results&lt;/code&gt; invented outcomes for past Grok blocks; Grok cost in &lt;code&gt;report&lt;/code&gt; was about 1.8x too high; &lt;code&gt;optimize capabilities --codex&lt;/code&gt; failed on Codex 0.169; &lt;code&gt;doctor&lt;/code&gt; now asks Codex which hooks it trusts.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;To pick up the new hooks after upgrading: run &lt;code&gt;limitcast install --guard --alerts&lt;/code&gt; again, restart Claude Code, Codex and Grok, and approve the new hooks in Codex (&lt;code&gt;/hooks&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;Also since last week: &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.18.2" rel="noopener noreferrer"&gt;0.18.2&lt;/a&gt; (mistyped options now fail with &lt;code&gt;Unknown option&lt;/code&gt; instead of running the plain command; settings and cache files are written atomically), &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.18.3" rel="noopener noreferrer"&gt;0.18.3&lt;/a&gt; (commands that change settings at the same moment no longer undo each other), and &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.19.2" rel="noopener noreferrer"&gt;0.19.2&lt;/a&gt; (optimizer fixes: learned command replacements keep the verification check they replace, and a map from a dirty working tree no longer contaminates the cached one). The optimizer is opt-in; see the &lt;a href="https://github.com/SamStull9007/limitcast/blob/v0.19.4/docs/optimizer.md" rel="noopener noreferrer"&gt;optimizer guide&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Checked in live sessions.&lt;/strong&gt; I ran real Codex 0.169 and Grok 1.0.77 sessions in isolated homes with limitcast's hooks recording what the tools actually send. That is how the Grok defect and the Codex encryption showed up; unit tests had passed. Not verified: whether Codex or Grok display a hook's message to you, and interactive Grok folder trust.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Re-checked on Nov 13&lt;/strong&gt; with a clean checkout of the v0.19.4 tag, using isolated home and state directories: all 293 tests pass when run one file at a time (in one parallel run on a heavily loaded machine, a timing-sensitive lock stress test timed out; it passes alone, and GitHub Actions passed on the release commit on Linux and macOS, Node 20, 22 and 24). The six-case deterministic replay behaves as documented on both the tag and the tarball downloaded from npm:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Replay case (30 simulated tool calls)&lt;/th&gt;
&lt;th&gt;Redirects&lt;/th&gt;
&lt;th&gt;Expected&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Same failing check, repeated&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Same failing check, every event delivered twice&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Failures that keep changing&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Passing check&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Read-only exploration&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Ambiguous exit status&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Replay uses simulated events, not a live model. It shows when the detector speaks up, not that a model follows the redirect. No token savings are claimed or measured.&lt;/p&gt;

&lt;h2&gt;
  
  
  One example
&lt;/h2&gt;

&lt;p&gt;Actual output, run on November 13, 2026 against the v0.19.4 tag in a &lt;strong&gt;synthetic task&lt;/strong&gt;: a real Codex 0.169 session in an isolated home was asked to start one subagent that reads a three-line text file. No real project, path or usage. This is the message limitcast's stop hook emits for you (never for the model); whether your terminal displays hook messages is one of the two things not verified above:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;limitcast: that default subagent used 53k tokens, 1k of them new (2 calls): about $0.01 at API list prices.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Most of those tokens are cached reads, which is why only 1k are new. The dollar figure is what the tokens would cost at OpenAI's list price for the model, not what your plan charges. Claude Code receipts show a share of your 5-hour window instead, because Claude has a calibrated window; Codex and Grok have none.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;From the maintainer's own machine.&lt;/strong&gt; One person's real usage: not synthetic, and not a user result. This is &lt;code&gt;limitcast optimize status&lt;/code&gt; in the maintainer's main working repository on November 13, 2026, counting sessions recorded since the optimizer was enabled in early November. Tool names, server names, paths and project names are left out.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Host&lt;/th&gt;
&lt;th&gt;Sessions recorded&lt;/th&gt;
&lt;th&gt;Capability-pruning proposals&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Claude Code&lt;/td&gt;
&lt;td&gt;194&lt;/td&gt;
&lt;td&gt;193&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Codex&lt;/td&gt;
&lt;td&gt;144&lt;/td&gt;
&lt;td&gt;2 (one built-in, one MCP server)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Grok&lt;/td&gt;
&lt;td&gt;122&lt;/td&gt;
&lt;td&gt;3 (all MCP servers)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;No proposal has been applied, and no check-command facts have been learned (0), even with learning turned on. The proposals are suggestions: a capability is proposed only after it was declared but unused across at least five sessions in this repository. They show what pruning could remove here, not what it saved. Nothing was disabled, so there is no savings figure. In the home directory, &lt;code&gt;--apply&lt;/code&gt; refuses on Codex and Grok, because the project config file there is your own user config; use &lt;code&gt;limitcast trim --codex --apply&lt;/code&gt; or &lt;code&gt;--grok --apply&lt;/code&gt; instead.&lt;/p&gt;

&lt;p&gt;Codex and Grok used to show 1 and 0 here (0.19.2): the optimizer only looked at three Codex built-ins and nothing on Grok. The Claude column is unchanged at 193 even though the Claude tool catalog now holds 317 tools (the cap was raised from 200 to 400); proposals need five sessions of evidence, so newly seen tools cannot qualify yet.&lt;/p&gt;

&lt;p&gt;For limitcast's own repository (140 tracked files, v0.19.4) the repository map fills its 6,000-character default budget. Injection stays off until you run &lt;code&gt;limitcast optimize map on&lt;/code&gt;, so weigh that recurring context against the searching it saves.&lt;/p&gt;

&lt;h2&gt;
  
  
  One useful command
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;limitcast optimize capabilities --grok
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Use &lt;code&gt;--codex&lt;/code&gt; for Codex. Read-only preview: it lists the MCP servers (and, on Codex, built-in capabilities) the optimizer would propose turning off in this repository, and nothing else changes until you pass &lt;code&gt;--apply &amp;lt;name&amp;gt;&lt;/code&gt;. Needs Node.js 20+, a Git repository, and at least five recorded sessions and ten prompts with the optimizer on (&lt;code&gt;limitcast install --optimize&lt;/code&gt;, restart your coding tool; Codex asks you to review and trust the new hooks). &lt;code&gt;limitcast optimize status&lt;/code&gt; shows the same proposals for all hosts, plus which features are on.&lt;/p&gt;

&lt;p&gt;Output from limitcast 0.19.4 on November 13, 2026, in a &lt;strong&gt;synthetic one-file repository&lt;/strong&gt; with an isolated home and state directory, so it shows a fresh setup with nothing recorded yet:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ limitcast optimize capabilities --grok
{
  "host": "grok",
  "proposed": [],
  "applied": [],
  "keep": []
}
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;proposed&lt;/code&gt; is empty here because there is no evidence yet. A preview is a suggestion, not proof a hook fired live. Anything you apply can be undone with &lt;code&gt;--undo&lt;/code&gt;, and &lt;code&gt;LIMITCAST_ALLOW=1&lt;/code&gt; bypasses the optimizer hooks.&lt;/p&gt;

&lt;h2&gt;
  
  
  One question
&lt;/h2&gt;

&lt;p&gt;On Codex and Grok the optimizer now proposes idle MCP servers. If you try &lt;code&gt;limitcast optimize capabilities --codex&lt;/code&gt; or &lt;code&gt;--grok&lt;/code&gt;, did it propose servers you would really switch off? And would you rather it applied per project (what &lt;code&gt;--apply&lt;/code&gt; does) or for all projects (what &lt;code&gt;trim&lt;/code&gt; does)? Tell us your coding tool and what made you decide, without sharing prompts, logs, private names or paths.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/weekly-update-2026-11-13/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Limitcast 0.19.2: safer learned checks and clean source maps</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Thu, 12 Nov 2026 14:40:19 +0000</pubDate>
      <link>https://dev.to/samstull9007/limitcast-0192-safer-learned-checks-and-clean-source-maps-4a03</link>
      <guid>https://dev.to/samstull9007/limitcast-0192-safer-learned-checks-and-clean-source-maps-4a03</guid>
      <description>&lt;p&gt;Originally published on &lt;a href="https://limitcast.dev/blog/limitcast-0-19-2/" rel="noopener noreferrer"&gt;the Limitcast blog&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;An optimizer should preserve the checks you intended to run and the source you actually have. An audit of Limitcast found two places where that could go wrong: learning an unrelated command as a replacement, and reusing a source map generated during temporary edits.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Limitcast 0.19.2 fixes both issues and is available on npm, GitHub and Homebrew.&lt;/strong&gt; This is a correctness release. It also adds the Limitcast icon to Codex plugin settings and pins every plugin command to 0.19.2.&lt;/p&gt;

&lt;h2&gt;
  
  
  A successful build does not replace a lint check
&lt;/h2&gt;

&lt;p&gt;The old learning logic could observe a failed &lt;code&gt;npm run lint&lt;/code&gt;, followed by a successful &lt;code&gt;npm run build&lt;/code&gt;, across three sessions and recommend the build command as a replacement. Both were grouped as non-test commands, even though they verify different things.&lt;/p&gt;

&lt;p&gt;The fix checks the identity of the verification command before learning a replacement. Equivalent launchers for the same Python test framework can still qualify, as can &lt;code&gt;npm test&lt;/code&gt; and &lt;code&gt;npm run test&lt;/code&gt;. Lint, build, type-check and unrelated test runners do not substitute for one another.&lt;/p&gt;

&lt;p&gt;The same compatibility check runs when saved facts are loaded. That means an incompatible recommendation learned before this patch is excluded from the context supplied to a later session.&lt;/p&gt;

&lt;h2&gt;
  
  
  Temporary edits should not survive in a cached map
&lt;/h2&gt;

&lt;p&gt;Previously, a source map generated from a dirty working tree could be stored against the unchanged Git commit. Restoring the original source could then leave temporary function names in the map returned for the clean repository.&lt;/p&gt;

&lt;p&gt;Limitcast now writes map caches only for clean working trees. Maps can still be generated while you work, but that output cannot overwrite the clean cache. The cache format also changes, so legacy entries are regenerated rather than trusted after the upgrade.&lt;/p&gt;

&lt;h2&gt;
  
  
  What was checked
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;All 258 tests passed, including regression coverage for command identity and map-cache behavior. Lint and privacy/plugin checks passed.&lt;/li&gt;
&lt;li&gt;GitHub CI passed on Linux and macOS with Node 20, 22 and 24.&lt;/li&gt;
&lt;li&gt;Packaged and installed CLI checks confirmed that incompatible saved replacements are suppressed, valid replacements still appear, and a reverted source edit returns the original map.&lt;/li&gt;
&lt;li&gt;Downloaded npm and GitHub packages matched the tested release artifact. Installed Claude, Codex and Grok plugin files matched the release commit.&lt;/li&gt;
&lt;li&gt;Isolated native adapter checks passed for Claude, Codex, Grok, Copilot and Pi. Fresh Codex discovery found six trusted, enabled Limitcast hooks.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Those checks establish the repaired behavior and installation state. They do not measure real-world token savings or replace testing in full signed-in conversations.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the packing algorithm stayed the same
&lt;/h2&gt;

&lt;p&gt;The audit also compared the source-map selector with a knapsack reference. At the default 6,000-character budget, all 89 eligible files already fit. A smaller budget improved a synthetic coverage score, but the score alone did not show that developers would get a more useful map. The selector stays unchanged in this release.&lt;/p&gt;

&lt;h2&gt;
  
  
  Upgrade to 0.19.2
&lt;/h2&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;npm install -g limitcast@0.19.2
limitcast --version
limitcast doctor
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;For an existing Homebrew installation:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;brew update
brew upgrade samstull9007/limitcast/limitcast
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Update the Limitcast plugin in your agent host too. Restart existing sessions to load refreshed plugin commands and extensions; a CLI upgrade alone does not reload an already-running session.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.19.2" rel="noopener noreferrer"&gt;Read the release notes&lt;/a&gt; · &lt;a href="https://www.npmjs.com/package/limitcast/v/0.19.2" rel="noopener noreferrer"&gt;Get 0.19.2 on npm&lt;/a&gt; · &lt;a href="https://github.com/SamStull9007/limitcast/blob/v0.19.2/docs/optimizer.md" rel="noopener noreferrer"&gt;Optimizer setup guide&lt;/a&gt;&lt;/p&gt;

</description>
      <category>opensource</category>
    </item>
    <item>
      <title>I audited my own tool: what limitcast got wrong and what I fixed</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Sun, 08 Nov 2026 14:43:16 +0000</pubDate>
      <link>https://dev.to/samstull9007/i-audited-my-own-tool-what-limitcast-got-wrong-and-what-i-fixed-35eo</link>
      <guid>https://dev.to/samstull9007/i-audited-my-own-tool-what-limitcast-got-wrong-and-what-i-fixed-35eo</guid>
      <description>&lt;p&gt;I read limitcast the way I would read a stranger's tool: corrupt its state, feed it bad input, time it on a large account, and read the code that writes to disk. It passed most of that. Here is what it did not pass, what I changed, and what the measurements say.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;limitcast held up where it matters most, and had four real defects.&lt;/strong&gt; No hook can block a tool call on bad state, but a mistyped option was silently ignored, settings were written in place, and &lt;code&gt;results&lt;/code&gt; walked my transcripts twice, and a cold &lt;code&gt;report --why&lt;/code&gt; ran 26 seconds in silence. All four are fixed in 0.18.2. The speedup turned out smaller than I first measured.&lt;/p&gt;

&lt;h2&gt;
  
  
  What held up
&lt;/h2&gt;

&lt;p&gt;I started from 222 passing tests, a clean Biome run, npm's latest version equal to the repository head (0.18.1), and no open issues or pull requests. Then I overwrote all 21 state files on my account with &lt;code&gt;{garbage&lt;/code&gt; and ran every hook: &lt;code&gt;guard&lt;/code&gt;, &lt;code&gt;prompt-check&lt;/code&gt;, &lt;code&gt;tool-limit&lt;/code&gt;, &lt;code&gt;tool-check&lt;/code&gt;, &lt;code&gt;subagent-receipt&lt;/code&gt;, &lt;code&gt;drift&lt;/code&gt; and &lt;code&gt;statusline&lt;/code&gt;. All seven exited 0, so none can block a tool call. Out-of-range input (&lt;code&gt;budget subagents 150&lt;/code&gt;, &lt;code&gt;plan set bogus&lt;/code&gt;, &lt;code&gt;--days -1&lt;/code&gt;) prints the usage and exits 1.&lt;/p&gt;

&lt;h2&gt;
  
  
  A typo that looked like success
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;limitcast trim --bogus-flag&lt;/code&gt; printed the normal report and exited 0. The CLI only looked for the options it knew, so &lt;code&gt;install --gaurd&lt;/code&gt; would have run a plain install and looked fine. Mistyped options now fail with &lt;code&gt;Unknown option --gaurd&lt;/code&gt; and exit 1. The hooks are exempt, so they still fail open. A test checks the list of options against the ones the code reads, so a new option cannot be forgotten.&lt;/p&gt;

&lt;h2&gt;
  
  
  Settings written in place
&lt;/h2&gt;

&lt;p&gt;Seven places wrote &lt;code&gt;config.json&lt;/code&gt; with &lt;code&gt;writeFileSync&lt;/code&gt;, which empties the file before it writes. A hook that reads at that moment sees an empty file and treats budgets and the keep-list as unset for that call. In a stress test (a 3.5 MB file rewritten in a loop while another process parsed it), 25 of 1,710 reads failed with the in-place write and 0 of 1,639 with a temp file and a rename. Real config files are a few hundred bytes, so the window is far narrower than that. It is still a window, and the fix is one helper, now used for the settings, the usage cache, &lt;code&gt;last.json&lt;/code&gt; and the pruned &lt;code&gt;samples.jsonl&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reading the same transcripts twice
&lt;/h2&gt;

&lt;p&gt;On my account there are 6,492 transcript files and a 17.3 MB usage cache. &lt;code&gt;results&lt;/code&gt; made 15,030 directory reads in the transcript folder, including two full walks of it, and parsed and rewrote that cache twice. It now reads usage once and makes 9,666 reads, one walk fewer. The usage and &lt;code&gt;report --why&lt;/code&gt; caches are rewritten only when a transcript changed. The first &lt;code&gt;report --why&lt;/code&gt; on a cold cache took 25.8 seconds with no output; it now says "reading N transcripts for the first time" first.&lt;/p&gt;

&lt;p&gt;I timed the committed version against the new one, alternating runs, median of five:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;command&lt;/th&gt;
&lt;th&gt;before&lt;/th&gt;
&lt;th&gt;after&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;results&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;5.34 s&lt;/td&gt;
&lt;td&gt;4.16 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;trim&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;3.48 s&lt;/td&gt;
&lt;td&gt;3.51 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;plan&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;2.37 s&lt;/td&gt;
&lt;td&gt;2.66 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;report --why --days 7&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;1.30 s&lt;/td&gt;
&lt;td&gt;1.33 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;status&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;0.89 s&lt;/td&gt;
&lt;td&gt;0.86 s&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Only &lt;code&gt;results&lt;/code&gt; got faster, by about a fifth. The rest is noise: my own session was writing transcripts the whole time, so the caches had changed anyway. My first timings, taken against a warm cache, suggested a much bigger win, and I had written that into the changelog before I re-measured. I corrected it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Smaller things
&lt;/h2&gt;

&lt;p&gt;The changelog had an "Unreleased" block that had already shipped in 0.15.0, sitting between 0.15.0 and 0.14.0. Nine comments still labelled features "Pro" after everything became free. The structured data on the home page said 0.18.0 while npm had 0.18.1. All three are fixed.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I did not fix
&lt;/h2&gt;

&lt;p&gt;About 0.6 seconds of every &lt;code&gt;report --why&lt;/code&gt; is the transcript walk itself, roughly 5,400 directory reads per pass here. Skipping old directories would be faster, but it could miss a subagent file that was appended to after its folder last changed, so I left it. Two commands changing the same setting at the same instant can still lose one update, because that needs a lock, not a rename. The fixes shipped in &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.18.2" rel="noopener noreferrer"&gt;0.18.2&lt;/a&gt; (&lt;a href="https://github.com/SamStull9007/limitcast/commit/38b3302" rel="noopener noreferrer"&gt;commit 38b3302&lt;/a&gt;); the 228 tests, six of them new, pass.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measure your own numbers
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;npm install -g limitcast
limitcast doctor
limitcast report --why --days 7
limitcast results
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If one of those is slow on your machine, open an issue with the transcript count and the time. The &lt;a href="https://limitcast.dev/guide/" rel="noopener noreferrer"&gt;guide&lt;/a&gt; lists every command, and the &lt;a href="https://limitcast.dev/blog/limitcast-0-18/" rel="noopener noreferrer"&gt;0.18 post&lt;/a&gt; covers the last release.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/limitcast-audit-2026-11-07/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Limitcast weekly update, Nov 6: safer trimming and clearer results</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Fri, 06 Nov 2026 22:18:17 +0000</pubDate>
      <link>https://dev.to/samstull9007/limitcast-weekly-update-nov-6-safer-trimming-and-clearer-results-20mf</link>
      <guid>https://dev.to/samstull9007/limitcast-weekly-update-nov-6-safer-trimming-and-clearer-results-20mf</guid>
      <description>&lt;h2&gt;
  
  
  One improvement
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Shipped this week:&lt;/strong&gt; &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.18.1" rel="noopener noreferrer"&gt;Limitcast 0.18.1&lt;/a&gt; is available on &lt;a href="https://www.npmjs.com/package/limitcast/v/0.18.1" rel="noopener noreferrer"&gt;npm&lt;/a&gt; and GitHub. It makes trimming more conservative: &lt;code&gt;trim --apply&lt;/code&gt; now disables only MCP servers that have been quiet for 30 days. A server unused during the report window but used within those 30 days stays enabled. An explicitly empty keep-list requires &lt;code&gt;--yes&lt;/code&gt; before applying changes.&lt;/p&gt;

&lt;p&gt;The release also makes &lt;code&gt;results&lt;/code&gt; lead with the net change in the starting context you control, and fixes the Codex guard's lookback around local midnight. These changes are in the published package; &lt;code&gt;main&lt;/code&gt; currently has no additional source-only changes beyond the release.&lt;/p&gt;

&lt;p&gt;Also shipped since last week's update: &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.18.0" rel="noopener noreferrer"&gt;0.18.0&lt;/a&gt; added duplicate-subagent holds, optional tool-result limits, and the &lt;code&gt;trim --share&lt;/code&gt; privacy fixes previously available only on &lt;code&gt;main&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  One example
&lt;/h2&gt;

&lt;p&gt;Actual public npm-registry output checked for this update on November 6, 2026; no synthetic usage figures or private session data:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight console"&gt;&lt;code&gt;&lt;span class="gp"&gt;$&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;npm view limitcast version
&lt;span class="go"&gt;0.18.1
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  One useful command
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;limitcast results &lt;span class="nt"&gt;--days&lt;/span&gt; 7
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Use it for a weekly review of starting-context changes, forecast evidence, and recorded protection events. Install or update with &lt;code&gt;npm install -g limitcast&lt;/code&gt; on macOS or Linux with Node.js 20+. For Claude usage readings, run &lt;code&gt;limitcast install&lt;/code&gt; and restart Claude Code; useful comparisons need recorded history, so a new installation may have little to report.&lt;/p&gt;

&lt;p&gt;Context estimates are not measured cash savings or proof that Limitcast caused a reduction. Recorded holds are events, not saved-token totals. A configuration check alone does not prove a live guard hold.&lt;/p&gt;

&lt;h2&gt;
  
  
  One question
&lt;/h2&gt;

&lt;p&gt;Does the 30-day inactivity threshold keep the MCP servers you still need enabled, or is it too conservative? Tell us which coding tool you use and the cutoff you would find useful, without sharing private logs.&lt;/p&gt;

&lt;p&gt;Follow shipped releases with &lt;strong&gt;Watch → Custom → Releases&lt;/strong&gt; and enable email delivery in your GitHub notification settings if you want release emails.&lt;/p&gt;

</description>
      <category>github</category>
      <category>javascript</category>
      <category>node</category>
      <category>npm</category>
    </item>
    <item>
      <title>limitcast weekly update, 30 October: what shipped and one week of real usage</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Fri, 30 Oct 2026 21:28:05 +0000</pubDate>
      <link>https://dev.to/samstull9007/limitcast-weekly-update-30-october-what-shipped-and-one-week-of-real-usage-49pk</link>
      <guid>https://dev.to/samstull9007/limitcast-weekly-update-30-october-what-shipped-and-one-week-of-real-usage-49pk</guid>
      <description>&lt;p&gt;Nothing new reached npm this week after 0.17.3 on 28 October. Here is what changed on &lt;code&gt;main&lt;/code&gt;, and one week of my own usage numbers with the caveats that come with them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;My total token usage fell 68% week over week, but limitcast can only vouch for an estimated 11.9M Claude context tokens not re-sent, about 1.3% of one 5-hour window.&lt;/strong&gt; The two figures measure different things, and neither proves limitcast caused the drop.&lt;/p&gt;

&lt;h2&gt;
  
  
  What shipped
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;0.17.3 (28 October, npm and GitHub):&lt;/strong&gt; the Grok status line prints one line per tool, so Claude and Grok windows no longer share one unlabeled row. &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.17.3" rel="noopener noreferrer"&gt;Release notes&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;On &lt;code&gt;main&lt;/code&gt;, not in npm 0.17.3:&lt;/strong&gt; &lt;code&gt;trim --share&lt;/code&gt; keeps configuration names and paths private when no measured startup transcript is available, and honors totals-only output for Codex and Grok. Sharing cannot be combined with keep-list edits, &lt;code&gt;--apply&lt;/code&gt; or &lt;code&gt;--undo&lt;/code&gt;. &lt;code&gt;trim --share --out FILE&lt;/code&gt; now prints "Wrote shared summary." instead of the file path. Until the next release, use &lt;code&gt;results --share&lt;/code&gt; for voluntary check-ins and read the output before posting.&lt;/li&gt;
&lt;li&gt;Also on &lt;code&gt;main&lt;/code&gt;: a Discussions space with tester guidance and issue templates, and a README with a CLI showcase built from synthetic demo sessions.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  One week of my own usage
&lt;/h2&gt;

&lt;p&gt;This is one person on one machine, real usage, not synthetic data and not a user result. It contains no prompts, project names or account IDs.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Usage recap:&lt;/strong&gt; Claude Bar's weekly recap shows 28.8M tokens (Codex 16.5M, Claude 12.3M) across 108 sessions and 86 h 25 min, &lt;strong&gt;down 68% from last week&lt;/strong&gt;. That is total usage, which depends on how much work I did as well as how I worked. Claude Bar counts tokens its own way, so its figures are not comparable with limitcast's totals.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;limitcast, same week&lt;/strong&gt; (&lt;code&gt;limitcast results --days 7&lt;/code&gt;): an estimated 11.9M Claude context tokens not re-sent, across 7,132 calls in 40 sessions.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What limitcast counts as saved
&lt;/h2&gt;

&lt;p&gt;limitcast reports one saved-token figure: context tokens not re-sent, estimated from characters divided by 4 against my first recorded starting context. Most of them would have been cached input, so 11.9M is about 1.3% of one 5-hour window. It is an estimate. It can include project or configuration changes, it does not prove limitcast caused the reduction, and it is not a lower bill.&lt;/p&gt;

&lt;p&gt;Codex and Grok context savings are unavailable, and holds and model changes are not converted into saved tokens. There is no larger total to add up.&lt;/p&gt;

&lt;h2&gt;
  
  
  Protection and forecasts
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;2 usage limits reached, both on 24 October. The 5-hour window was already at 100% in the first reading limitcast saw. For the weekly window, limitcast warned 40 minutes ahead.&lt;/li&gt;
&lt;li&gt;Applied events: 8 tight-window model changes (subagents ran on Sonnet instead of my main model) and 2 cold-cache holds. The guard did not need to stop a subagent. These are events, not token savings.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Starting context
&lt;/h2&gt;

&lt;p&gt;My Claude Code sessions start at 109.4k tokens before I type. About 7.2k (7%) went unused for 7 days: 7 MCP servers and 68 skills. Claude Code's own prompt and tools are 68.5k of the total.&lt;/p&gt;

&lt;p&gt;The part I control (hooks, MCP servers, skills, agents, instruction files) grew from about 39.1k to about 40.9k tokens (+1.8k) since 24 October. The 11.9M estimate covers my whole starting context against the first recorded one, so it can reflect changes beyond that part.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measure your own numbers
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
limitcast onboard
limitcast results &lt;span class="nt"&gt;--days&lt;/span&gt; 7
limitcast trim
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;To share a summary, run &lt;code&gt;limitcast results --share&lt;/code&gt;, read every line, and remove anything you do not want public. Nothing uploads automatically. The weekly updates and a one-week tester check-in are in &lt;a href="https://github.com/SamStull9007/limitcast/discussions" rel="noopener noreferrer"&gt;Discussions&lt;/a&gt;.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/weekly-update-2026-10-30/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>limitcast 0.17: Grok weekly usage percentage, off by default</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Wed, 28 Oct 2026 02:24:11 +0000</pubDate>
      <link>https://dev.to/samstull9007/limitcast-017-grok-weekly-usage-percentage-off-by-default-1chm</link>
      <guid>https://dev.to/samstull9007/limitcast-017-grok-weekly-usage-percentage-off-by-default-1chm</guid>
      <description>&lt;p&gt;Until now limitcast could show what Grok cost you, but not how close you were to its limit. 0.17 adds Grok's weekly usage percentage. It is the first limitcast feature that uses a login, so it is off by default.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Run &lt;code&gt;limitcast grok-usage on&lt;/code&gt; and Grok's weekly window is forecast like Claude's and Codex's.&lt;/strong&gt; On my account it reads 9% used, resetting on Sunday at 19:34.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Grok had no percentage
&lt;/h2&gt;

&lt;p&gt;Grok writes what each session cost to disk, not a share of your limit. So limitcast could show dollars, and could only forecast once you set a dollar cap with &lt;code&gt;limitcast budget grok&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;I first tried asking the Grok CLI itself. It speaks JSON-RPC over stdio, and its billing method, &lt;code&gt;x.ai/billing&lt;/code&gt;, answers "Method not found" in Grok 1.0.70. The percentage is held by grok.com's billing service, so that is where limitcast asks. I found the call by reading how a usage monitor I already run gets the same number, and my figure matched its 9%.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it does
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ limitcast grok-usage
Grok weekly: 9% used, resets 11/1/2026, 7:34:33 PM

$ limitcast status
window     used  burn     runs out   resets     today     status
grok 7d    9%    +0.0%/h  -          Sun 19:34  16% left  OK
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;About every 5 minutes, limitcast reads the login Grok keeps in &lt;code&gt;~/.grok/auth.json&lt;/code&gt; and asks grok.com for your billing cycle. The refresh runs in the background, so a status line never waits on the network. The login goes only to grok.com, a redirect to any other host is refused, and it is never stored or logged. The readings are kept in &lt;code&gt;~/.local/state/limitcast/grok-quota.json&lt;/code&gt;, which is what lets limitcast work out a burn rate.&lt;/p&gt;

&lt;p&gt;The request identifies itself as limitcast. 0.17.0 sent a Grok client's User-Agent instead; 0.17.1 fixes that, and grok.com gave the same answer either way.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it adds
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;A &lt;code&gt;grok 7d&lt;/code&gt; row in &lt;code&gt;limitcast status&lt;/code&gt;, with runout time, reset and today's share, and the same figure on Grok's status line.&lt;/li&gt;
&lt;li&gt;The guard can hold a new Grok subagent when the weekly window is on pace to run out, as it does for Claude and Codex.&lt;/li&gt;
&lt;li&gt;A line in &lt;code&gt;limitcast doctor&lt;/code&gt;: &lt;code&gt;✓ Grok weekly: 9% used (checked 0 min ago)&lt;/code&gt;. If the login has expired or the reply changes, it says why instead.&lt;/li&gt;
&lt;li&gt;A weekly dollar cap from &lt;code&gt;limitcast budget grok&lt;/code&gt; steps aside for the real percentage. A 5-hour cap still works as before.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It tracks weekly plans. If your plan reports monthly credits, limitcast shows the figure but does not forecast it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why it is off by default
&lt;/h2&gt;

&lt;p&gt;limitcast's promise so far was that it reads files already on your machine. This feature sends a login to a server, which is a different kind of thing, so you turn it on yourself and &lt;code&gt;limitcast grok-usage off&lt;/code&gt; turns it off again. It also depends on a billing endpoint that xAI can change. If that happens, &lt;code&gt;doctor&lt;/code&gt; says "Grok usage unavailable" with the reason, and the dollar-cap forecast keeps working.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measure your own numbers
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
limitcast grok-usage on
limitcast status
limitcast doctor
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The &lt;a href="https://limitcast.dev/guide/" rel="noopener noreferrer"&gt;guide&lt;/a&gt; has the rest, the &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.17.1" rel="noopener noreferrer"&gt;0.17.1 release&lt;/a&gt; has the tarball, and the &lt;a href="https://limitcast.dev/blog/limitcast-0-16/" rel="noopener noreferrer"&gt;0.16 post&lt;/a&gt; covers the previous release.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/limitcast-0-17/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>limitcast 0.16: find the notes Claude Code never loads, and split big memory files</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Wed, 28 Oct 2026 01:46:29 +0000</pubDate>
      <link>https://dev.to/samstull9007/limitcast-016-find-the-notes-claude-code-never-loads-and-split-big-memory-files-3fei</link>
      <guid>https://dev.to/samstull9007/limitcast-016-find-the-notes-claude-code-never-loads-and-split-big-memory-files-3fei</guid>
      <description>&lt;p&gt;Claude Code keeps notes for each project in a &lt;code&gt;MEMORY.md&lt;/code&gt; index plus topic files, and it reads only the start of that index at the beginning of every session. limitcast 0.16 shows how much of yours fits, and adds two Claude Code commands to tidy the rest.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Anything past line 200 or 25 KB of MEMORY.md is never loaded, and 0.16 tells you when you are there.&lt;/strong&gt; &lt;code&gt;limitcast memory&lt;/code&gt; and &lt;code&gt;limitcast doctor&lt;/code&gt; report it, and &lt;code&gt;/limitcast:memory-audit&lt;/code&gt; and &lt;code&gt;/limitcast:memory-split&lt;/code&gt; help you fix it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Claude Code loads
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://docs.claude.com/en/docs/claude-code/memory" rel="noopener noreferrer"&gt;memory documentation&lt;/a&gt; says the first 200 lines of &lt;code&gt;MEMORY.md&lt;/code&gt;, or the first 25 KB, whichever comes first, are loaded at the start of every conversation, and that content beyond that is not loaded at session start. It also says Claude Code checks the file after Claude writes to it, and asks Claude to shorten an index that is near or over a limit.&lt;/p&gt;

&lt;p&gt;That covers an index Claude just wrote. An index that grew over months, or was edited by hand, can still be past the cap with nobody told. The documentation describes that check for the index; it does not describe one for broken links, files nothing points at, or topic files that have grown large.&lt;/p&gt;

&lt;h2&gt;
  
  
  limitcast memory
&lt;/h2&gt;

&lt;p&gt;Run it in a project and it reads that project's memory folder and changes nothing. This is a made-up folder with a few problems built in:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ limitcast memory
Memory: ~/.claude/projects/-home-you-app/memory
Index: MEMORY.md 21 / 200 lines, 0.8 / 25 KB
Topic files: 16
Not linked from the index (2): old-notes.md, scratch.md
Broken links (1): MEMORY.md -&amp;gt; deploy-notes.md
Over 150 lines or 24 KB (1): api-conventions.md (312 lines, 8 KB)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The index line is used lines and bytes against the cap. Below it are a link in the index that points at a missing file, two topic files nothing links to, and a file long enough that every task that opens it reads all 312 lines.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;limitcast doctor&lt;/code&gt; shows the same as one line. It fails only when the index is past the cap, because that is the one case where Claude is missing notes you wrote. Here it is for that folder, and for an index of 214 lines:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;✓ Memory: 16 files, index 21/200 lines, 1/25 KB; 2 not linked from the index, 1 broken link, 1 file over 150 lines. Run /limitcast:memory-audit, then /limitcast:memory-split
✗ Memory index 214/200 lines, 6/25 KB: Claude Code loads only the first 200 lines or 25 KB of MEMORY.md, so the rest is never read. Fix: /limitcast:memory-audit
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The daily background check leaves the memory line out, because memory belongs to one project. Run &lt;code&gt;limitcast doctor&lt;/code&gt; inside the project to see it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Audit and split from inside Claude Code
&lt;/h2&gt;

&lt;p&gt;The plugin adds two commands. Both start from the same sweep and change nothing until you agree.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;/limitcast:memory-audit&lt;/code&gt; goes through the index cap, broken links, files nothing links to, notes that name a file, flag or command that no longer exists, and notes that contradict each other. You get two lists: mechanical fixes first, then judgement calls with the evidence. It never deletes a note unasked.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/limitcast:memory-split&lt;/code&gt; splits one oversized file into a short hub plus topic files. First it decides whether a split is worth it: a file that is one continuous procedure stays whole, because splitting would turn one read into two. If it is worth it, it shows you the table of contents and waits. Then it moves the text verbatim, keeps the hub's filename so nothing that links to it breaks, and checks that whatever cited the old file still finds what it cited.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  On my own memory folder
&lt;/h2&gt;

&lt;p&gt;I ran it on mine: 300 topic files, and an index of 54 of 200 lines and 7.0 of 25 KB, so nothing is cut off. It also lists 116 files that nothing links to with a Markdown link. That comes from how I route: my index points at a few catalog files, and the notes behind them are found by filename pattern. So the list is a prompt to check, not a verdict. It is why the audit command searches for references before it calls a file an orphan, and why only the index cap makes &lt;code&gt;doctor&lt;/code&gt; fail.&lt;/p&gt;

&lt;h2&gt;
  
  
  Free and MIT
&lt;/h2&gt;

&lt;p&gt;Since 0.15.0 all of limitcast is free and MIT licensed, and the commands above need no key. The &lt;a href="https://github.com/SamStull9007/limitcast/releases/tag/v0.16.0" rel="noopener noreferrer"&gt;0.16.0 release&lt;/a&gt; has the tarball, and &lt;code&gt;npm install -g limitcast&lt;/code&gt; gets it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measure your own numbers
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
&lt;span class="nb"&gt;cd &lt;/span&gt;your-project
limitcast memory
limitcast doctor
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In Claude Code, run &lt;code&gt;/plugin marketplace add SamStull9007/limitcast&lt;/code&gt;, then &lt;code&gt;/plugin install limitcast@limitcast&lt;/code&gt;, and try &lt;code&gt;/limitcast:memory-audit&lt;/code&gt; in the project. The &lt;a href="https://limitcast.dev/guide/" rel="noopener noreferrer"&gt;guide&lt;/a&gt; has the rest, and the &lt;a href="https://limitcast.dev/blog/limitcast-0-14/" rel="noopener noreferrer"&gt;0.14 post&lt;/a&gt; covers the previous release.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/limitcast-0-16/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>limitcast 0.14: pace your week by the day, and keep what trim must not touch</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Tue, 27 Oct 2026 18:31:03 +0000</pubDate>
      <link>https://dev.to/samstull9007/limitcast-014-pace-your-week-by-the-day-and-keep-what-trim-must-not-touch-3aii</link>
      <guid>https://dev.to/samstull9007/limitcast-014-pace-your-week-by-the-day-and-keep-what-trim-must-not-touch-3aii</guid>
      <description>&lt;p&gt;limitcast 0.14 adds three things: a keep-list so &lt;code&gt;trim&lt;/code&gt; never touches what you rely on, a daily share of your weekly limit that can now hold new subagents, and a status line that says which limit you will hit first.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;0.14 makes limitcast's advice safer to apply and harder to ignore.&lt;/strong&gt; &lt;code&gt;trim&lt;/code&gt; can be told what to leave alone, the daily pace acts instead of only showing, and &lt;code&gt;status&lt;/code&gt; names the window that runs out first.&lt;/p&gt;

&lt;h2&gt;
  
  
  A keep-list for trim
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;limitcast trim --apply&lt;/code&gt; turns off MCP servers and hides skills you have not used. Some are rare but essential. Put them on the keep-list and no &lt;code&gt;--apply&lt;/code&gt; will touch them, in Claude Code, Codex or Grok:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ limitcast trim keep ego-browser
Kept by `trim --apply`: ego-browser
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Spelling does not matter: &lt;code&gt;ego-browser&lt;/code&gt; and &lt;code&gt;ego_browser&lt;/code&gt; match. &lt;code&gt;limitcast trim&lt;/code&gt; shows what it would keep, &lt;code&gt;limitcast trim keep --remove &amp;lt;name&amp;gt;&lt;/code&gt; drops one, and &lt;code&gt;limitcast trim keep&lt;/code&gt; on its own lists them. It is free.&lt;/p&gt;

&lt;h2&gt;
  
  
  Your week, paced by the day
&lt;/h2&gt;

&lt;p&gt;The &lt;code&gt;today&lt;/code&gt; column in &lt;code&gt;limitcast status&lt;/code&gt; is your share of what is left of a weekly window, split evenly over the days until it resets. In 0.14 it acts (Pro). With notifications on, you get one message a day when you have used 80% of today's share:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;limitcast: Claude 7d: 80% of today's share used, 1% left today.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Turn on &lt;code&gt;limitcast budget today on&lt;/code&gt; and the subagent guard also holds new subagents once today's share is used up, until local midnight:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;limitcast: today's share of your Claude 7d limit is used up (6% of 5%).
New subagents are held until midnight. Turn this off:
`limitcast budget today off`, or set LIMITCAST_ALLOW=1 to spawn anyway.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The &lt;code&gt;(6% of 5%)&lt;/code&gt; reads as 6% used against a 5% share for today. The hold is off by default. A window already at 100% is left to the tool, and &lt;code&gt;LIMITCAST_ALLOW=1&lt;/code&gt; overrides it. Free installs record what Pro would have held, and &lt;code&gt;limitcast results&lt;/code&gt; shows it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which window runs out first
&lt;/h2&gt;

&lt;p&gt;When a window is on pace to run out, &lt;code&gt;limitcast status&lt;/code&gt; now ends with one line naming the one that hits the wall first and where else there is room:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;You'll hit Claude 5h first (16:40). Codex 5h has 78% left.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you split work across Claude Code, Codex and Grok, that line tells you where to send the next task.&lt;/p&gt;

&lt;h2&gt;
  
  
  Upgrade
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
limitcast status
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Local only, never uses your Claude login. &lt;a href="https://limitcast.dev/guide/" rel="noopener noreferrer"&gt;Feature guide&lt;/a&gt; · &lt;a href="https://limitcast.dev/#pricing" rel="noopener noreferrer"&gt;Pro plans&lt;/a&gt;&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/limitcast-0-14/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Do subagents burn your Claude Code limit faster?</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Tue, 27 Oct 2026 18:27:09 +0000</pubDate>
      <link>https://dev.to/samstull9007/do-subagents-burn-your-claude-code-limit-faster-4ke5</link>
      <guid>https://dev.to/samstull9007/do-subagents-burn-your-claude-code-limit-faster-4ke5</guid>
      <description>&lt;p&gt;Subagents are the usual suspect when a Claude Code limit runs out early. I measured what they actually cost on my account, using the same fit as in &lt;a href="https://limitcast.dev/blog/tokens-per-percent/" rel="noopener noreferrer"&gt;How many tokens is 1% of your Claude Code limit?&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Per model call, no.&lt;/strong&gt; A subagent call cost about the same share of my 5-hour limit as a main-thread call: roughly 0.07%. Subagents burn the limit because of how many calls they make. A typical subagent cost about 2% of a 5-hour window, so five in parallel cost about 10%.&lt;/p&gt;

&lt;h2&gt;
  
  
  Per call: about the same
&lt;/h2&gt;

&lt;p&gt;Over two days (23–25 October), on one account:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Model calls&lt;/th&gt;
&lt;th&gt;New tokens per call&lt;/th&gt;
&lt;th&gt;Cached tokens per call&lt;/th&gt;
&lt;th&gt;Limit per call&lt;/th&gt;
&lt;th&gt;Share of the limit used&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Main thread&lt;/td&gt;
&lt;td&gt;4,234&lt;/td&gt;
&lt;td&gt;6.6k&lt;/td&gt;
&lt;td&gt;353k&lt;/td&gt;
&lt;td&gt;0.072%&lt;/td&gt;
&lt;td&gt;≈ 60%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Subagents&lt;/td&gt;
&lt;td&gt;2,895&lt;/td&gt;
&lt;td&gt;8.4k&lt;/td&gt;
&lt;td&gt;223k&lt;/td&gt;
&lt;td&gt;0.071%&lt;/td&gt;
&lt;td&gt;≈ 40%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;These two days end a few hours later than the ones in the first post, so its subagent share (42%) and this one (40%) differ slightly.&lt;/p&gt;

&lt;p&gt;The two cancel out. The main thread re-reads a bigger context on every call (353k cached tokens against 223k), but cached tokens count about 1/67 as much as new ones. Each subagent writes its own context into the cache from scratch, so its calls carry more new tokens (8.4k against 6.6k), and new tokens are what move the meter.&lt;/p&gt;

&lt;h2&gt;
  
  
  Per subagent: about 2% of a window
&lt;/h2&gt;

&lt;p&gt;The same two days had 93 subagents. Priced against the 5-hour limit:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Typical subagent (median)&lt;/td&gt;
&lt;td&gt;16 calls&lt;/td&gt;
&lt;td&gt;≈ 1.9% of a 5-hour window&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Average&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;td&gt;≈ 2.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;1 in 10 cost more than&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;td&gt;≈ 4.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Largest&lt;/td&gt;
&lt;td&gt;86 calls&lt;/td&gt;
&lt;td&gt;≈ 10.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The costliest tenth of subagents used 29% of everything subagents used. Most subagents are cheap; a few long ones are not, and a batch of them started at once is what ends a window early.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do with this
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Count before you fan out.&lt;/strong&gt; Budget roughly 2% of a 5-hour window per subagent. Ten parallel subagents near the end of a window are a fifth of it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Watch the long ones.&lt;/strong&gt; A subagent that keeps working for 80+ calls costs as much as five typical ones. Give it a narrow task and a clear stopping point.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Don't count on the model alone.&lt;/strong&gt; 97% of my subagent calls already ran on Sonnet, and they were still 40% of the limit. The number of calls matters more.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stop new ones when the window is nearly gone.&lt;/strong&gt; That's what limitcast's guard does: when a window is on pace to run out within 30 minutes, it blocks new subagents in Claude Code and Codex, and while a window is tight it starts them on Sonnet.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The caveats
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;One account, one plan, about two days. Your mix of work will give different numbers.&lt;/li&gt;
&lt;li&gt;The fit doesn't separate models. My main thread ran a lot of Opus and my subagents almost all Sonnet. If Anthropic counts Opus tokens more heavily, the main thread's real share is higher than shown and the subagents' lower.&lt;/li&gt;
&lt;li&gt;"Limit per call" applies the measured weights to each call's tokens; it isn't read off the meter call by call. The readings are whole percentages.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Measure your own
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
limitcast &lt;span class="nb"&gt;install&lt;/span&gt;      &lt;span class="c"&gt;# records your limit readings from the status line&lt;/span&gt;
&lt;span class="c"&gt;# use Claude Code for a day or two, then:&lt;/span&gt;
limitcast results      &lt;span class="c"&gt;# what moves your meter: new vs cached tokens&lt;/span&gt;
limitcast report       &lt;span class="c"&gt;# Pro: tokens by day, project, model and session, and the subagent share&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The free version is local-only and never uses your Claude login.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/subagents/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;. limitcast is a local-only command-line tool that forecasts when your Claude Code, Codex and Grok usage limits run out.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Why a long idle pause costs you: Claude Code's prompt cache</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Mon, 26 Oct 2026 03:39:26 +0000</pubDate>
      <link>https://dev.to/samstull9007/why-a-long-idle-pause-costs-you-claude-codes-prompt-cache-3jp4</link>
      <guid>https://dev.to/samstull9007/why-a-long-idle-pause-costs-you-claude-codes-prompt-cache-3jp4</guid>
      <description>&lt;p&gt;You leave a long Claude Code session for lunch, come back, type one short message, and the 5-hour meter jumps. Nothing went wrong:&lt;br&gt;
the prompt cache expired while you were away.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;After a pause, your next message re-sends the whole conversation as new tokens.&lt;/strong&gt; New tokens count about 67 times as much&lt;br&gt;
against the limit as cached ones (on my account), so by those rates the first message back into a 300k-token session costs about 1.8% of a 5-hour&lt;br&gt;
window, where the same message a minute earlier would have cost about 0.03%.&lt;/p&gt;
&lt;h2&gt;
  
  
  What the prompt cache does
&lt;/h2&gt;

&lt;p&gt;Every Claude Code model call sends the whole conversation: instructions, tool definitions, every message and tool result so far. The&lt;br&gt;
part that hasn't changed since the last call is read back from Anthropic's prompt cache instead of being processed again. Cache reads&lt;br&gt;
are cheap, which is why a long session costs far less than its size suggests.&lt;/p&gt;

&lt;p&gt;The cache doesn't last forever. Anthropic's docs: "By default, the cache has a 5-minute lifetime. The cache is refreshed for no&lt;br&gt;
additional cost each time the cached content is used", and a 1-hour lifetime is available as an option&lt;br&gt;
(&lt;a href="https://platform.claude.com/docs/en/build-with-claude/prompt-caching" rel="noopener noreferrer"&gt;Prompt caching&lt;/a&gt;). Each call keeps it warm. Stop&lt;br&gt;
for longer than its lifetime and it's gone.&lt;/p&gt;
&lt;h2&gt;
  
  
  What the first message back costs
&lt;/h2&gt;

&lt;p&gt;With the cache gone, the next call writes the whole context to the cache again. Cache writes are new tokens, and new tokens are what&lt;br&gt;
the usage limit mostly counts. From my own fit (&lt;a href="https://limitcast.dev/blog/tokens-per-percent/" rel="noopener noreferrer"&gt;How many tokens is 1% of your limit?&lt;/a&gt;): 1% of&lt;br&gt;
the 5-hour window was about 166k new tokens or 11.2M cached ones.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Context re-sent&lt;/th&gt;
&lt;th&gt;Cache still warm&lt;/th&gt;
&lt;th&gt;Cache expired&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;100k tokens&lt;/td&gt;
&lt;td&gt;≈ 0.01%&lt;/td&gt;
&lt;td&gt;≈ 0.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;300k tokens&lt;/td&gt;
&lt;td&gt;≈ 0.03%&lt;/td&gt;
&lt;td&gt;≈ 1.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;600k tokens&lt;/td&gt;
&lt;td&gt;≈ 0.05%&lt;/td&gt;
&lt;td&gt;≈ 3.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Share of the 5-hour window for the context alone (the reply adds a little), from one account's measurement: 6.04% per million new&lt;br&gt;
tokens and 0.09% per million cached ones. Your plan's numbers differ; the ratio is what matters.&lt;/p&gt;

&lt;p&gt;One message is a few percent. It adds up when you come back to several long sessions after a meeting, or when every subagent you&lt;br&gt;
start after a break has to rebuild its context too.&lt;/p&gt;
&lt;h2&gt;
  
  
  What to do about it
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Compact before you leave.&lt;/strong&gt; &lt;code&gt;/compact&lt;/code&gt; replaces the conversation with a summary, so the cold rebuild is small.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Or clear it.&lt;/strong&gt; If the next task is unrelated, &lt;code&gt;/clear&lt;/code&gt; and start fresh instead of paying to rebuild a context you don't need.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Come back to one session at a time.&lt;/strong&gt; Each expired session rebuilds separately.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Watch your context size.&lt;/strong&gt; limitcast Pro puts a &lt;code&gt;/compact&lt;/code&gt; nudge in the status line once a session re-sends more than ~150k tokens per call.&lt;/p&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  limitcast 0.11: a warning before the expensive message
&lt;/h2&gt;

&lt;p&gt;limitcast Pro now has a cold-cache warning. When a session that re-sends 100k+ tokens has sat idle past the cache lifetime (60 minutes&lt;br&gt;
by default, &lt;code&gt;LIMITCAST_CACHE_MINUTES&lt;/code&gt; to change it), it holds your first message once and tells you what it will cost against&lt;br&gt;
your own measured limit. Send it again to go ahead, or &lt;code&gt;/compact&lt;/code&gt; first:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;limitcast: this session sat idle for 75 minutes, so its prompt cache has
expired. Your message re-sends 300k tokens of context: ≈1.8% of your
5-hour window (≈0.03% if the cache were still warm). Send it again to go
ahead, or run /compact first to shrink it.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It comes with &lt;code&gt;limitcast install --guard&lt;/code&gt; (&lt;a href="https://limitcast.dev/#install" rel="noopener noreferrer"&gt;install&lt;/a&gt;); free installs record what it would have held.&lt;br&gt;
Already installed with &lt;code&gt;--guard&lt;/code&gt;? Run &lt;code&gt;limitcast doctor --fix&lt;/code&gt; to add it. &lt;a href="https://limitcast.dev/#f-cold" rel="noopener noreferrer"&gt;More on the warning&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Measure your own numbers
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
limitcast &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="c"&gt;# records your limit readings from the status line&lt;/span&gt;
&lt;span class="c"&gt;# use Claude Code for a day or two, then:&lt;/span&gt;
limitcast results &lt;span class="c"&gt;# what 1% of your 5-hour window is, in new and cached tokens&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Local only, never uses your Claude login. &lt;a href="https://limitcast.dev/#install" rel="noopener noreferrer"&gt;Install details&lt;/a&gt; · &lt;a href="https://limitcast.dev/#pricing" rel="noopener noreferrer"&gt;Pro plans&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;limitcast forecasts when each Claude Code and Codex window runs out, and its Pro plan stops runaway subagents before they burn the rest of it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://limitcast.dev/" rel="noopener noreferrer"&gt;See what limitcast does&lt;/a&gt; · &lt;a href="https://limitcast.dev/blog/claude-code-usage-limits/" rel="noopener noreferrer"&gt;How the 5-hour and weekly limits work&lt;/a&gt; ·&lt;br&gt;
&lt;a href="https://limitcast.dev/blog/tokens-per-percent/" rel="noopener noreferrer"&gt;How many tokens is 1%?&lt;/a&gt; · &lt;a href="https://limitcast.dev/blog/subagents/" rel="noopener noreferrer"&gt;Subagents and your limit&lt;/a&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Claude Code usage limit reached: how the 5-hour and weekly limits work</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Mon, 26 Oct 2026 03:38:33 +0000</pubDate>
      <link>https://dev.to/samstull9007/claude-code-usage-limit-reached-how-the-5-hour-and-weekly-limits-work-3mn1</link>
      <guid>https://dev.to/samstull9007/claude-code-usage-limit-reached-how-the-5-hour-and-weekly-limits-work-3mn1</guid>
      <description>&lt;p&gt;"You've hit your limit" usually arrives in the middle of something. Here is how Claude Code's limits work on a Pro, Max or Team plan,&lt;br&gt;
what actually moves them, and how to see the end of a window coming instead of finding out from the error.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;There are two limits.&lt;/strong&gt; A session limit that resets every five hours, and a weekly limit that resets at a fixed time each week.&lt;br&gt;
What fills them is mostly &lt;strong&gt;new&lt;/strong&gt; tokens: on my account, 1% of the 5-hour limit was about 166k new tokens but about 11 million&lt;br&gt;
cached ones.&lt;/p&gt;
&lt;h2&gt;
  
  
  The two windows
&lt;/h2&gt;

&lt;p&gt;-&lt;br&gt;
&lt;strong&gt;The 5-hour session limit.&lt;/strong&gt; Anthropic's help center: paid plans have "a session limit that resets every five hours, and weekly&lt;br&gt;
limits that reset once every week"&lt;br&gt;
(&lt;a href="https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work" rel="noopener noreferrer"&gt;How do usage and length limits work?&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;-&lt;br&gt;
&lt;strong&gt;The weekly limit.&lt;/strong&gt; It "reset[s] at a fixed time each week that is assigned to your account", whenever you started using Claude&lt;br&gt;
(&lt;a href="https://support.claude.com/en/articles/8325606-what-is-the-pro-plan" rel="noopener noreferrer"&gt;What is the Pro plan?&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;-&lt;br&gt;
&lt;strong&gt;Where to see them.&lt;/strong&gt; Settings → Usage on claude.ai shows both as progress bars, with the time left in the current session&lt;br&gt;
(&lt;a href="https://support.claude.com/en/articles/9797557-usage-limit-best-practices" rel="noopener noreferrer"&gt;Usage limit best practices&lt;/a&gt;). Claude Code also&lt;br&gt;
passes the same readings to your status line, which is how a tool can show them while you work.&lt;/p&gt;

&lt;p&gt;When either one is used up you wait for its reset, or keep going on paid extra usage if you've turned that on. A full weekly limit&lt;br&gt;
stops you even with room left in the 5-hour window.&lt;/p&gt;
&lt;h2&gt;
  
  
  What fills them: new tokens, not cached ones
&lt;/h2&gt;

&lt;p&gt;The meter is a percentage, and Anthropic doesn't say how many tokens a percent is. I measured it on my own account by comparing each&lt;br&gt;
rise of the 5-hour reading with the tokens my sessions used in between (61 observations, two days; the method is in&lt;br&gt;
&lt;a href="https://limitcast.dev/blog/tokens-per-percent/" rel="noopener noreferrer"&gt;How many tokens is 1% of your Claude Code limit?&lt;/a&gt;):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Tokens per 1% of the 5-hour limit&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;New tokens (input, cache writes, output)&lt;/td&gt;
&lt;td&gt;≈ 166k&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cached tokens (cache reads)&lt;/td&gt;
&lt;td&gt;≈ 11.2M&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;One account, one plan. A fit on the previous day gave about 200k and 10.7M, so read these as a range.&lt;/p&gt;

&lt;p&gt;A cached token counted roughly 1/67 as much as a new one. Claude Code re-sends your whole conversation on every model call, but most&lt;br&gt;
of it comes back from the cache, so a long session is cheaper than its size suggests. What runs a window out is anything that sends&lt;br&gt;
&lt;strong&gt;new&lt;/strong&gt; tokens:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Subagents.&lt;/strong&gt; Each one writes its own context from scratch. A typical one cost me about 2% of a 5-hour window (&lt;a href="https://limitcast.dev/blog/subagents/" rel="noopener noreferrer"&gt;Do subagents burn your limit faster?&lt;/a&gt;).&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Coming back after a long pause.&lt;/strong&gt; Once the prompt cache expires, the next message re-sends the whole context as new tokens (&lt;a href="https://limitcast.dev/blog/prompt-cache-idle/" rel="noopener noreferrer"&gt;Why a long idle pause costs you&lt;/a&gt;).&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Big tool output and file reads.&lt;/strong&gt; Everything new that enters the conversation is new tokens once.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;A large starting context.&lt;/strong&gt; Instructions, tool lists and skills are written once per session and every subagent, then re-read on every call.&lt;/p&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  How to see the runout coming
&lt;/h2&gt;

&lt;p&gt;The percentage alone doesn't say whether you'll make it to the reset. That depends on your pace: 70% with an hour left is fine at a&lt;br&gt;
slow pace and a lockout in twenty minutes during a burst of subagents. limitcast does that arithmetic for you:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ limitcast status
window used burn runs out resets today status
claude 5h 71% +18.2%/h 16:40 17:00 - AT_RISK
claude 7d 41% +0.6%/h - Sun 03:00 6% left OK
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;&lt;p&gt;The 5-hour window is forecast from the last hour's pace; &lt;strong&gt;AT_RISK&lt;/strong&gt; means it runs out before it resets.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;The weekly window uses your pace over the last full day, nights included, and shows today's even share of the week, so one busy hour doesn't cry wolf.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;The same line sits in your Claude Code status line, from the readings Claude Code already passes to it. No login, nothing leaves your machine.&lt;/p&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The free version gives you the forecast and the status line. &lt;a href="https://limitcast.dev/#pricing" rel="noopener noreferrer"&gt;Pro&lt;/a&gt; also stops new subagents when a window is on&lt;br&gt;
pace to run out within 30 minutes, moves them to Sonnet while it's tight, and sends a desktop notification before you hit the limit.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do when you're close
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Don't start a fan-out near the end of a window.&lt;/strong&gt; Ten subagents at about 2% each are a fifth of it.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Compact long sessions&lt;/strong&gt; before a pause, not after: &lt;code&gt;/compact&lt;/code&gt; shrinks what the next call has to re-send.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Queue the heavy job for after the reset&lt;/strong&gt; instead of starting it with 10% left.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;&lt;strong&gt;Check what loads before you type.&lt;/strong&gt; Unused MCP servers and skills are re-sent on every call and to every subagent: &lt;code&gt;limitcast trim&lt;/code&gt; measures them.&lt;/p&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  See your own numbers
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
limitcast &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="c"&gt;# adds the forecast to your status line; backs up settings.json&lt;/span&gt;
limitcast status &lt;span class="c"&gt;# every window: used, burn rate, runout time, reset&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Node 20+, macOS or Linux. &lt;a href="https://limitcast.dev/#install" rel="noopener noreferrer"&gt;Install details&lt;/a&gt; · &lt;code&gt;limitcast uninstall&lt;/code&gt; restores your settings.&lt;/p&gt;

&lt;p&gt;limitcast forecasts when each Claude Code and Codex window runs out, and its Pro plan stops runaway subagents before they burn the rest of it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://limitcast.dev/" rel="noopener noreferrer"&gt;See what limitcast does&lt;/a&gt; · &lt;a href="https://limitcast.dev/#pricing" rel="noopener noreferrer"&gt;Pricing&lt;/a&gt; ·&lt;br&gt;
&lt;a href="https://limitcast.dev/blog/tokens-per-percent/" rel="noopener noreferrer"&gt;How many tokens is 1%?&lt;/a&gt; · &lt;a href="https://limitcast.dev/blog/subagents/" rel="noopener noreferrer"&gt;Subagents and your limit&lt;/a&gt; ·&lt;br&gt;
&lt;a href="https://limitcast.dev/blog/prompt-cache-idle/" rel="noopener noreferrer"&gt;Idle pauses and the prompt cache&lt;/a&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>How many tokens is 1% of your Claude Code limit?</title>
      <dc:creator>Sam Stull</dc:creator>
      <pubDate>Mon, 26 Oct 2026 01:26:25 +0000</pubDate>
      <link>https://dev.to/samstull9007/how-many-tokens-is-1-of-your-claude-code-limit-232n</link>
      <guid>https://dev.to/samstull9007/how-many-tokens-is-1-of-your-claude-code-limit-232n</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://limitcast.dev/blog/tokens-per-percent/" rel="noopener noreferrer"&gt;limitcast.dev&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Anthropic shows how much of your 5-hour and weekly limit you've used, as a percentage. It doesn't say how many tokens that percentage is, and it doesn't say how different kinds of tokens count. So I measured it on my own account.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;1% of the 5-hour limit was about 166,000 new tokens, or about 11 million cached tokens.&lt;/strong&gt; A cached token counted roughly 1/67 as much as a new one.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That ratio changes what's worth optimizing. Over the two days I measured, new tokens were 2.5% of everything Claude Code sent, and they used about 63% of my limit.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Share of tokens&lt;/th&gt;
&lt;th&gt;Share of the limit they used (estimated)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;New tokens (input, cache writes, output)&lt;/td&gt;
&lt;td&gt;2.5%&lt;/td&gt;
&lt;td&gt;≈ 63%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cached tokens (cache reads)&lt;/td&gt;
&lt;td&gt;97.5%&lt;/td&gt;
&lt;td&gt;≈ 37%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;My Claude Code usage, 23–25 October 2026: 53.7M new tokens and 2,103.5M cached tokens. The limit share applies the fit below to those totals.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  How I measured it
&lt;/h2&gt;

&lt;p&gt;Claude Code passes your limit readings to the status line: percent used of each window and when it resets. Its transcripts record every model call's tokens: input, cache writes, cache reads and output. Put the two side by side and you can see how many tokens it took to move the meter.&lt;/p&gt;

&lt;p&gt;Each time the 5-hour reading rose between two status-line updates, I took the rise and the tokens my sessions on that account used in between, split into &lt;strong&gt;new tokens&lt;/strong&gt; (input, cache writes and output) and &lt;strong&gt;cached tokens&lt;/strong&gt; (cache reads). Then a least-squares fit with no intercept:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;rise in % ≈ a × new tokens + b × cached tokens
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;With 61 such observations, the fit gave:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;% of the 5-hour limit per million&lt;/th&gt;
&lt;th&gt;Tokens per 1%&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;New tokens&lt;/td&gt;
&lt;td&gt;6.04%&lt;/td&gt;
&lt;td&gt;≈ 166k&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cached tokens&lt;/td&gt;
&lt;td&gt;0.090%&lt;/td&gt;
&lt;td&gt;≈ 11.2M&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A fit on the previous day's readings gave about 200k and 10.7M, so treat these as a range rather than constants: roughly 165–200k new tokens, or 10.7–11.2M cached tokens, per 1%.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it means in practice
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Subagents are expensive.&lt;/strong&gt; Each subagent starts its own context and writes it into the cache from scratch. Subagents sent about a third of my tokens (33%), but nearly half of my new tokens (48%), and used an estimated 42% of my limit.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Your starting context is cheap per call, but it adds up.&lt;/strong&gt; My sessions start at about 100k tokens before I type: system prompt, tools, MCP servers, skills, CLAUDE.md. Written once into the cache, that's about 0.6% of the limit. After that, each model call re-reads it from the cache for about 0.009%. A session with 300 model calls spends about 3% of a window re-reading that starting context. Trimming 4k tokens of unused skills and MCP servers from mine saved about 0.3% of a window over 1,100 calls: real, but small.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Cache misses are where it gets expensive.&lt;/strong&gt; This one is an inference from the ratio rather than something I measured directly: when the prompt cache expires (after a pause, for example), the next call writes the whole context again as new tokens. In a session that has grown to 200k tokens, that single call is more than 1% of the limit.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;So the order of what to watch is:&lt;/strong&gt; how many subagents you start, how long your sessions run between pauses, and only then how big your CLAUDE.md is.&lt;/p&gt;

&lt;h2&gt;
  
  
  The caveats
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;This is one person's account on one plan, over about two days. Your numbers will differ.&lt;/li&gt;
&lt;li&gt;The fit doesn't separate models. If Anthropic weights Opus and Sonnet differently, my mix of the two is baked into these numbers.&lt;/li&gt;
&lt;li&gt;Cache writes are counted as new tokens. They may be weighted differently from input and output.&lt;/li&gt;
&lt;li&gt;The readings are whole percentages, so each observation is noisy. It takes dozens of them before the fit settles.&lt;/li&gt;
&lt;li&gt;Anthropic can change how limits are counted at any time; a measurement like this is a snapshot.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Measure your own
&lt;/h2&gt;

&lt;p&gt;I built this measurement into limitcast, a free command-line tool for Claude Code and Codex. It reads only the files Claude Code already writes on your machine and sends nothing anywhere.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-g&lt;/span&gt; limitcast
limitcast &lt;span class="nb"&gt;install&lt;/span&gt;      &lt;span class="c"&gt;# records your limit readings from the status line&lt;/span&gt;
&lt;span class="c"&gt;# use Claude Code as normal for a day or two, then:&lt;/span&gt;
limitcast results
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Once it has about 30 observations, &lt;code&gt;limitcast results&lt;/code&gt; prints your own numbers: how many new and cached tokens move your meter by 1%. The fit is in &lt;a href="https://github.com/SamStull9007/limitcast/blob/main/src/calibrate.mjs" rel="noopener noreferrer"&gt;&lt;code&gt;src/calibrate.mjs&lt;/code&gt;&lt;/a&gt; if you want to check it.&lt;/p&gt;

&lt;p&gt;I'd like to hear whether other people's ratios look similar, especially on different plans.&lt;/p&gt;

</description>
      <category>cli</category>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
