Limitcast 0.20.1: safer trimming and fresher forecasts
Limitcast 0.20.1 is available on npm, GitHub and Homebrew. This patch fixes cases where incomplete activity logs could produce unsafe trim decisions, and where an old forecast could keep holding new subagents after an idle period. It also adds clearer evidence for MCP pruning and optimizer learning.
Limitcast forecasts usage limits and provides optional controls for Claude Code, Codex and Grok. All features remain free under the MIT License. Version 0.20.1 requires Node.js 22 or later on macOS or Linux.
Trimming needs readable evidence
A tool that changes your configuration needs to distinguish an unused server from a server whose activity history could not be read. The latest audit found that Claude activity parsing could miss calls when JSON used different spacing. Incomplete logs could also leave too little evidence to justify a cut.
Version 0.20.1 parses the records structurally and tracks whether activity was observed and read completely, including when results come from the cache. It also recognizes slash commands in both string messages and array-based text blocks.
When history is missing or incomplete,
trim --apply
refuses to change configuration. Unverified servers and skills stay
on. Unsupported nested settings and malformed undo entries are
rejected before writing, so the command cannot report a successful
change that the settings format could not save.
Idle time should not preserve an old hold
An unchanged usage percentage is still a fresh observation. Previously, discarding repeated readings could leave an old burst of activity looking current. A forecast based on that burst could continue holding new subagents after work had gone quiet.
Limitcast now retains one latest timestamped observation for each usage series without appending every unchanged reading to history. Short-window burn rates require recent observations, and a predicted runout time already in the past cannot authorize a forecast hold. A genuine fresh burst can still trigger the guard.
Each server needs its own observation history
A newly configured MCP server should not inherit an unused verdict from sessions that happened before it was available. Codex and Grok proposals now require five distinct completed sessions that observed that particular server as available at both the beginning and end of the session, alongside the existing usage and keep-list protections.
When removal is observed, the server's evidence is reset; re-adding it starts a new observation window. Duplicate stop events do not count as additional sessions.
Claude transcript ingestion now shares the optimizer's 400-tool catalog limit. It reports incomplete bounded reads, and partial catalogs do not count as complete observations toward pruning eligibility. Used tools remain protected when the bounded usage history fills up.
See why learning has no facts
If learning has produced no facts, it helps to know what evidence it actually saw. From your repository, run:
limitcast optimize status
The report now includes per-host counts for eligible, passed, failed, unsupported and ambiguous verification checks, plus the number of active facts. These counters begin with 0.20.1; earlier observations are not backfilled. Excluded command text is not retained.
The safety rules remain in place: learning uses allowlisted checks and requires evidence from three independent sessions before activating a fact. Optional model learning and judging use your authenticated Codex CLI, consume your Codex allowance, and remain off by default.
These diagnostics explain an evidence gap. They do not establish task success or net token savings.
What was verified
The release passed 316 tests and all eight GitHub CI jobs. The matrix covers Node 22, 24 and 26 on macOS and Linux, alongside lint and focused Semgrep checks. A production-only coverage gate now enforces floors of 95% lines, 82% branches and 91% functions.
Packed-package checks exercise the installed Claude, Codex and Grok hooks, the Copilot executable contract and the Pi extension. Release verification also compared the published npm and GitHub downloads with the tested tarball and checked the Homebrew installation.
Those checks verify specific behavior and adapter contracts. They do not replace signed-in model conversations or a real-world savings study.
Upgrade to 0.20.1
Check your Node version, then use the package manager you installed Limitcast with. For npm:
node --version
npm install -g limitcast@0.20.1
For an existing Homebrew installation:
brew update
brew upgrade samstull9007/limitcast/limitcast
Check the result with:
limitcast --version
limitcast doctor
If doctor reports a moved launcher,
limitcast doctor --fix
refreshes the integration using your existing feature choices.
If you already installed the plugin, update it in each host you use. For Claude Code:
claude plugin marketplace update limitcast
claude plugin update limitcast@limitcast --scope user
For Codex:
codex plugin marketplace upgrade limitcast
codex plugin add limitcast@limitcast
For Grok:
grok plugin update limitcast
Restart open Claude Code, Codex and Grok sessions
to load the refreshed plugin definitions and version-pinned commands.
If you use the separate Copilot or Pi optimizer adapter, refresh its
installed reference with
limitcast optimize install --copilot
or
limitcast optimize install --pi, then restart that host.
Read the 0.20.1 release notes, the optimizer guide, and the previous 0.20.0 update for more context.