This candidate consolidates architecture and capability improvements subsequently included in public releases.
Added
- Added global Chat outside project Tasks, the runtime Skill Market and Inspect AI benchmark integration.
- Added Light for read-only consultation, evidence investigation, comparison and planning.
Changed
- Unified responsibility/composition boundaries for Task, Mission, Session, Provider, MCP, dispatch, persistence and desktop workspace, removing duplicate state and retired entries.
- Tasks, requirements, acceptance, artifacts and tool results exchange typed facts and exact sources; specialized capabilities use authorized references.
Fixed
- Fixed shared issues in cross-project initialization/ownership, native OpenAI tool-call serialization, MCP session runtime ownership and Mission acceptance/wake recovery.
- Synchronized native SDK, util, plugin and dependency projections with the candidate version; installation preflight checks actual packaged output.