3.22.0

Added: jumbo decision show command to display the complete architectural decision record by ID. Fixed: goal review verification now discovers and runs project-supported verification instead of assuming a specific package manager or ecosystem.

3.21.0

Added

• TUI goal creation result screen: Cockpit and Goals-screen authoring now show a bounded pending, success, or failure result
  after submission. Successful results include the created goal ID and remain visible until Enter is pressed; failures retain every
  entered value for retry or allow cancellation with Escape.

Fixed

• TUI goal scope authoring: Scope-in and scope-out values are now collected as independent repeatable items, preserving spaces,
  item order, and entered values when navigating backward or forward through the goal wizard.
• TUI goal authoring layout: Long objective values now wrap within the bounded goal wizard without expanding its background
  into horizontal bands, and the input cursor remains after the final wrapped character.

3.20.0

Added

• Relation graph audits: Added jumbo relations audit with selectable dangling, isolated, inactive-only, ambiguous-ID, and summary checks, optional entity-type filtering, deterministic text output, and a stable structured result for relation coverage diagnostics.

3.19.0

[3.19.0] - 2026-07-24

Added

• Shortest relation paths: Added jumbo relations path with typed or uniquely inferred endpoints, deterministic bounded breadth-first search, direction and relation filters, provenance-preserving text output, and a stable structured result for connected or disconnected memories.

3.18.0

Added

• Bounded relation traversal: Added jumbo relations traverse with deterministic breadth-first traversal, typed root inference, directed graph semantics, depth and edge limits, relation filters, hop-grouped text output, and stable structured output.

Changed

• Relation list filters: Expanded jumbo relations list with direction, relation type, related entity type, strength, and entity ID filters while preserving active-only status as the default.

3.17.0

Changed

• Domain primitive identities: Newly created audiences, audience pains, components, decisions, dependencies, goals, guidelines, invariants, projects, relations, sessions, value propositions, and workers now receive domain-owned UUID identities. Existing persisted identifiers remain supported during replay and through CLI operations without migration.

3.16.1

[3.16.1] - 2026-07-10

Fixed

• TUI daemon reliability: Cockpit refiner, reviewer, and codifier workers now stream bounded activity and errors while running, retain meaningful outcomes and elapsed-time heartbeats, survive navigation between TUI screens, and stop exactly once on request or application shutdown.

3.16.0

Changed

• Antigravity agent support: Replaced Gemini CLI initialization and worker-agent support with Antigravity CLI support using the documented \�gy\ executable, .agents/skills, and .agents/hooks.json\ hook surface.

Removed

• Gemini agent support: Removed Gemini as a selectable initialization and worker agent target, including obsolete Gemini settings fragments, configurers, command mappings, and logo assets.
• Co-located evals project: Removed the \�vals/\ directory from this repository; the eval framework now lives in its own repository rather than as a nested npm project here.

3.15.0

Added

• Rejected goal issues on start: 'jumbo goal start' now renders a 'Review Issues' section when restarting a goal that was previously rejected during QA review, instructing the developer to address the issues before resubmitting.

3.14.0

Added

• Evals validity scenarios: Added the
  ull-hypothesis-rate-limiter\ and \ile-reconstruction-expense-tracker\ scenarios to validate the eval harness itself, and \�val scenario create\ templates now carry structural assertions through to the created scenario.

Changed

• Codex initialization: \jumbo init, repair, and \jumbo evolve --yes\ now install Codex-managed skills to the documented repository skill location .agents/skills\ while keeping .codex\ limited to Codex hooks/configuration.

Fixed

• Codex managed artifact preservation: Obsolete Jumbo-managed Codex skill copies under .codex/skills\ are removed only when byte-identical to the current managed templates, preserving customized skill content, user files, hook comments, unknown hook settings, and unrelated .codex\ configuration.
• Evals headless permissions: The Claude Code eval harness adapter now runs sessions with --dangerously-skip-permissions, since headless -p\ runs in isolated temp workdirs have no human to approve tool use and were leaving both A/B arms unable to write files.