Limitcast maintainer Sam Stull shipped versions 0.19.3 and 0.19.4 to npm and GitHub on the evening of November 13, 2026, closing the feature gap between coding agents. The updates bring the optimizer's pruning proposals, subagent receipts, and alert systems to OpenAI's Codex and xAI's Grok, matching the capabilities previously exclusive to Claude Code. This release cycle was driven by an audit revealing that the tool's architecture had inadvertently treated these platforms as second-class citizens due to early development priorities.

Closing the Platform Gap

The core of this update is parity in capability management. Codex and Grok users can now use the optimize capabilities command to propose disabling idle MCP servers within their project configurations. For Codex, the tool can also target built-in capabilities, while Grok is limited to MCP servers because it lacks a persistent setting to remove built-in tools. The --apply flag now writes journal entries that are undoable and verifiable against the native CLI lists, ensuring that configuration changes are transparent and reversible for developers.

Critical Fixes and Defect Resolution

Version 0.19.3 initially failed to apply changes for Grok users because the platform ignores project server entries lacking a command field. The 0.19.4 release rectifies this by writing entries with a stub command that never executes, a fix verified in live headless sessions. Beyond the main feature, the audit uncovered and resolved several defects: a model pinning error that affected Claude spawns, a failure in the duplicate-spawn hold for Codex due to encrypted arguments, and a significant overestimation of Grok costs by approximately 1.8x in reports. These fixes ensure that the local forecasting data is accurate across all supported environments.

Real-World Verification and Metrics

Stull emphasized that these features were validated in live sessions using isolated home directories, revealing issues that unit tests missed. The update also introduces API-equivalent dollar receipts for subagent usage, calculated at list prices rather than actual plan costs. For Codex, receipts are computed when the subagent stops, whereas Grok's receipts are logged by the next hook due to its usage reporting timing. The maintainer's own repository data shows that while Claude Code generated 193 pruning proposals, Codex and Grok produced only a handful, reflecting the different ways these tools handle capability declarations and session histories.

Key Takeaways

  • Limitcast 0.19.4 is the first version to provide full optimizer parity for Codex and Grok users.
  • A critical bug in 0.19.3 that prevented Grok configuration changes has been fixed with a stub command workaround.
  • Users must run limitcast install --guard --alerts again and restart their tools to activate the new hooks.
  • The tool now provides subagent receipts in API-equivalent dollars for Codex and Grok, though these are estimates based on list prices.

The Bottom Line

This release is a necessary infrastructure cleanup that finally treats Codex and Grok as first-class citizens in the Limitcast ecosystem. For developers juggling multiple AI coding agents, having consistent visibility into token usage and idle capabilities across all three platforms is a massive quality-of-life win. The audit-driven approach to fixing these platform-specific quirks demonstrates a commitment to reliability that is rare in the fast-moving dev tool space.