DEV Community

Sam Stull
Sam Stull

Posted on Originally published at limitcast.dev Fully Autonomous

Limitcast 0.20.0: safer configuration and clearer optimizer overhead

Limitcast 0.20.0 is available on npm and GitHub. This release focuses on what happens when a coding tool sends an unfamiliar argument, a settings write is interrupted, or an optimizer has incomplete evidence. It also makes optimizer overhead easier to inspect.

Limitcast forecasts usage limits and provides optional controls for Claude Code, Codex and Grok. All features remain free under the MIT License. Version 0.20.0 requires Node.js 22 or later on macOS or Linux. Upgrade Node 20 before installing it.

MCP calls keep their original arguments

The generic result limiter previously tried to enforce output limits on MCP calls. That is unsafe across tools with different schemas: a parameter named limit does not always mean the same thing, and some tools do not accept one.

In 0.20.0, the generic limiter never rewrites or denies MCP arguments. Large-output notices remain available, while enforcement is limited to supported native Read and Grep controls. This correction lets an MCP tool receive the arguments its caller supplied.

Interrupted configuration changes have a recovery record

Trimming changes configuration, so the order of operations matters. Claude, Codex and Grok trim apply and undo now persist a recovery record before changing configuration. An interrupted operation can be recovered, and conflicting user edits are preserved for reconciliation.

Settings locks also check ownership more carefully. A lock held by a live process cannot be removed just because it is old. Recovery from a dead owner and lock release check the unique owner. If an existing settings file is malformed, Limitcast preserves it and refuses to overwrite it.

Drift checks now preserve explicit choices, including fully disabled skills and servers you deliberately re-enabled. The point is to keep configuration changes recoverable without treating your later decisions as something to undo.

Missing evidence cannot justify pruning

Codex usage parsing now accepts both flat and namespaced tool-call records, regardless of JSON whitespace. Missing or incomplete usage history cannot authorize pruning.

That matters because “no call found” can mean either that a capability was unused or that the history could not be read completely. The optimizer must preserve that distinction before proposing a configuration change.

See the optimizer's own overhead

limitcast optimize status now reports injected context overhead alongside actual model-usage receipts when available. Injected-token estimates use characters divided by four; they are separate from model-reported usage. The injection counters begin with this version, so older injections remain unknown.

Optional Codex-backed model calls consume your Codex plan. They remain off by default, and the privacy disclosure now describes that optional path. The deterministic optimizer adds no model calls. Source-map injection and learning remain separate opt-in choices.

These measurements help you inspect the cost of the optimizer itself. They do not establish that a task succeeded, that a model followed a redirect, or that Limitcast produced net token or cash savings. Both injected context and optional model calls add overhead.

What was checked for this release

The release records 307 passing tests, lint, focused Semgrep, privacy checks and isolated packed-artifact smoke checks. GitHub Actions passed on the release commit, with a test matrix covering Node 22, 24 and 26 on macOS and Linux.

Regression coverage includes persistence, lock ownership, schema safety and native model subprocess behavior, including timeout and cancelled-worker cases. Those checks support the specific fixes above; they are not a comparative real-world savings study.

Upgrade to 0.20.0

Check your runtime, then install the exact release:

node --version
npm install -g limitcast@0.20.0
limitcast onboard
Enter fullscreen mode Exit fullscreen mode

Use Node.js 22 or later. onboard gives read-only setup guidance. If you already enabled the guard, alerts and optimizer, refresh those integrations after upgrading:

limitcast install --guard --alerts --optimize
Enter fullscreen mode Exit fullscreen mode

Enable only the features you want. Restart your coding-tool sessions and review any Codex hook approval requests. If you installed the Copilot or Pi optimizer adapter, refresh it with limitcast optimize install --copilot or limitcast optimize install --pi.

For an existing optimizer setup, inspect its measurements with:

limitcast optimize status
Enter fullscreen mode Exit fullscreen mode

A configuration check confirms setup; observing a hook in a fresh session is what confirms it ran.

Read the 0.20.0 release notes, optimizer guide for this version and changelog.

If you try the update, which is more useful to inspect first: recovery from an interrupted change, or the optimizer's recurring context overhead? Share the coding tool you use and a sanitized observation, without prompts, logs, credentials or private paths.

Top comments (0)