Highlights
- Autonomous
/goal: approval gates in AGENTS.md, CLAUDE.md, memories, rules and skills count as satisfied by the goal. A goal pauses only for an explicit user block, missing user-only access, a dependency owned by someone else, or an irreversible production action (money, billing, production data deletion, destructive migrations) the objective does not authorize. - Reasoning model takes over stalled goals: lack of progress never pauses a goal. At the first stall the reasoning model runs the goal's rounds in the same conversation until the goal ends; if a reasoning round fails, the main model takes the goal back instead of pausing.
- Caveman output mode: terse prose for the main agent and subagents, on by default (
full). Configure with[agent] caveman,GROK_CAVEMAN, or/caveman lite|full|ultra|off. The/goalharness roles keep normal prose. - Cheaper tool-output compression: the utility-model extractive pass only runs on outputs of at least 4 KB after deterministic reduction.
Full Changelog: v2.0.18...v2.0.19