Skip to content

Limitcast 0.20.0: safe limits and recoverable configuration

Latest

Choose a tag to compare

@SamStull9007 SamStull9007 released this 16 Nov 03:47

Requires Node.js 22 or later. Node 20 users must upgrade before installing.

  • Fix: MCP arguments are never rewritten or denied by the generic result limiter. Large-output notices remain; only native Read/Grep controls are enforced.
  • Fix: live settings locks cannot be evicted based on age. Dead-owner recovery and release check unique ownership. Malformed existing settings are preserved and refused.
  • Fix: Claude, Codex and Grok trim/apply/undo persist a write-ahead recovery record before changing configuration. Partial writes recover, and conflicting user edits are preserved for reconciliation.
  • Fix: drift checks preserve explicit user choices, including fully disabled skills and re-enabled servers.
  • Fix: Codex usage parsing accepts flat and namespaced call records, independent of JSON whitespace. Missing or incomplete usage history cannot authorize pruning.
  • Validation: security scans and isolated packed-artifact smoke checks now run in CI. Added persistence, ownership, schema-safety and native model subprocess regressions, including timeout and cancelled-worker behavior.
  • Measurement: optimizer status reports injected context overhead and actual model usage receipts without inferring task success or net savings. The privacy disclosure includes optional Codex-backed model calls.
  • Compatibility: Node.js 22+ is now required; CI tests maintained LTS versions 22, 24 and 26 on macOS and Linux. Node 20 users must upgrade their runtime.

Validation: 307 tests; lint, focused Semgrep, privacy and packed-artifact checks. CI covers Node 22/24/26 on macOS and Linux.

After upgrading, re-run limitcast install --guard --alerts --optimize if those features were already enabled, then restart your host sessions and review any Codex hook approval requests. Refresh Copilot/Pi adapters with limitcast optimize install --copilot / --pi if installed. No real-world task-success or net-savings claim is made.