Skip to content

Limitcast 0.19.3: Codex and Grok parity

Choose a tag to compare

@SamStull9007 SamStull9007 released this 14 Nov 00:50

Brings Codex and Grok in line with Claude Code where each tool allows it, from an audit of how each host was handled.

  • Optimizer: proposes MCP servers that no session in the repository called, for Codex and Grok, and switches them off per project (.codex/config.toml, .grok/config.toml). Same safeguards as Claude: five sessions of evidence, a keep list, a journaled write, a post-write check against codex mcp list / grok mcp list, and an undo that restores exact bytes. Grok has no persistent setting for built-in tools, so it proposes servers only.
  • Codex 0.169: optimize capabilities --codex no longer fails because multi_agent is listed as "deprecated".
  • Guard: a model pinned for Grok no longer applies to Claude spawns; the duplicate-spawn hold now fires for Codex (spawn_agent sends message); a Grok resume_from re-run is not held as a duplicate; Codex and Grok send SubagentStop to clear the hold (install --guard adds the hook; approve it in Codex).
  • results: past Grok blocks are judged from windows rebuilt from the spend ledger instead of invented outcomes; percentages round.
  • Mix-ups: Grok's status line no longer overwrites the Claude feed file that doctor checks; notifications say "Grok"; burn-report session labels use the end of the id.
  • after-reset: accepts --src grok and uses the shortest window a tool reports (a Codex plan with only a 7-day window failed on the default 5h).
  • Grok cost in report uses the cost Grok recorded per turn (about 1.8x too high before).

Validation: 277 tests, lint, and the Codex apply, preview and undo path run against the real codex CLI in an isolated home. Not verified: the Grok apply path in a trusted folder (an untrusted folder is refused and rolled back with the real grok CLI), and a live run of the new SubagentStop hooks. No real-world token-savings claim is made.