0.21.0

Baseplate 0.21.0 — See the plan, keep control

Baseplate makes it easier to update an existing Roblox game with your AI tools and see
what is happening along the way.

• Choose your AI clearly. A custom keyboard-accessible chooser shows each installed
  tool, account state and supported modes. Live updates preserve an available selection;
  model choices stay inside the CLI.
• One useful improvement. Improve my game guides you through a bug fix, a small feature,
  a free local scan or Open Cloud setup. It explains the next step, preserves project task
  drafts and keeps guidance available when an AI account needs attention. Tool help maps
  common problems to Map, Board and Labs. A passed playtest belongs to its actual task.
• Guided Open Cloud setup. Find the right IDs and read permissions, save locally, then
  test real game details and DataStore-list access. Saved settings are not presented as a
  verified connection; write, publish and Analytics access need separate checks.
• Open your existing game. The first-run action adds your source folder without
  scaffolding a new project or sending you straight into a diagnostic workflow.
• Watch the agent's plan. Chat tiles and the Studio panel show steps from Baseplate's
  plan tool. A new turn clears the previous plan; completed plans stay visible while idle.
  Chat is the default where the selected engine supports it.
• Review protected structural edits. The bridge asks before supported restructuring
  operations. Declining, timing out, disconnecting or cancelling does not authorize the edit.
  Switching to another named place invalidates pending approvals and the inspection state.
• Reliable Revert. Bridge 0.18.1 keeps instance identities after renames and moves,
  including the original parent. Revert does not target a same-named replacement;
  failed or incomplete Studio results remain unconfirmed. Updating the bridge requires
  restarting Studio. Older bridges do not offer exact Revert for path-changing edits.
• Explore the Game map. An index of your source groups mechanics and proposes next moves.
  These are suggestions to investigate, not proof that a feature is missing. Free can run
  a suggestion with one agent; eligible paid plans can use a Crew.
• More reliable startup. Claude readiness requires a successful real response rather
  than trusting a credentials file. Map requests ignore stale responses after project
  switches. Map code and Labs styles load when their views are opened.
• Maintenance. Updated vulnerable dependencies, including Electron, and aligned
  marketing comparisons with current Roblox and Ropilot capabilities.
  Updated MCP SDK/WebSocket dependencies. License checks cannot revoke a newly
  activated key using a delayed result for the previous key; validation has a bounded timeout.
• Current integrations. Updated Studio MCP instance selection and Play arguments;
  restored current Codex replies, MCP results and reliable resume. Codex configuration
  stays bound to its session. Cursor has Windows launcher support and a sign-in flow;
  malformed project MCP settings are refused. Grok tool events are translated; Gemini
  receives project MCP settings in Native mode.
  Windows npm launchers resolve the real Node executable so Gemini Native opens without
  the packaged-app "File not found" error. Grok trusts the selected project to discover
  its MCP servers and displays actual wrapped MCP results. Claude organization restrictions
  have a specific explanation. The main launch bar uses each CLI's own model defaults.
• Personal Google accounts. Antigravity Native tiles have project-scoped MCP setup.
  Gemini CLI requires an API key or a supported organization license after Google's
  retirement of individual-account access. Baseplate shows those requirements.
  Antigravity does not offer structured Chat or Playtest Copilot.
• Start building spacing. Wrapped project context keeps its own height and separates
  the primary actions at compact window sizes. The CLI bar has no empty model wrapper.
• Evidence and live data. Playtest PASS requires actual Start/Play/Stop results.
  DataOps writes require ETag concurrency protection and preserve metadata. Asset uploads
  return a completed asset ID or an explicit pending operation, with a status tool.

Baseplate does not include AI inference. Your provider's account requirements and usage
limits apply. Approval cards cover supported bridge operations, not every possible agent
action; Revert is available only when the app has enough information to reverse a change.

Validation includes clean Windows installation and upgrades from 0.20.0, and actual
packaged macOS apps on Intel and Apple Silicon with a visible welcome screen and native
terminal. The 277-test suite and automated UI checks pass; guidance loads on demand.
Performance checks passed in the earlier tested configuration. An additional empty-profile
check exceeded the startup and transfer budgets; first-run optimization remains a follow-up,
and the source of that overhead is still under investigation. macOS installers do not have
Apple Developer ID signing or notarization.

Studio, bridge and Rojo workflows were tested on disposable projects. Cursor and Codex
completed clean Playtests; authenticated Cursor, Grok and Antigravity responses/MCP were
verified. Claude requires an eligible account with organization access where applicable;
Gemini API-key/organization inference was not verified in this release check. Provider
account restrictions are explained separately from installation or launcher failures.

Open Cloud reads and Safe DataOps migration/rollback were tested on a private experience,
including a stale-ETag rejection. Open Cloud requires your own valid key and permissions;
Analytics, asset upload and place publishing were not included in those live checks.
The Stripe test-mode checkout, license activation and cancellation cycle passed in an
isolated environment without real funds. Production billing, email delivery and hosted
success-page redirects were outside that test.