Cortex Code Desktop v1.14.0 release notes

Headlining: Plugins

Cortex Code now has Plugins — portable packages that extend what the agent can do. A plugin can contain skills (custom instructions the agent follows), hooks (scripts that run on events like session start), MCP servers (external tool integrations), and agent definitions (specialized subagents). Install plugins from GitHub or create your own — both flows are agent-led: just ask the agent to import a plugin from a repo URL or scaffold a new one, and it walks you through the whole process conversationally. Plugins show up in a card-grid view with on/off toggles, sync buttons, and live file-watching so new plugins appear instantly without refreshing. Cross-platform (Windows path bugs squashed), auto-trusted MCP, path variable substitution — the works. Think of it as the extension marketplace, but for agent behaviors.

Highlights

  • App Gallery + Snowflake Apps Runtime: The App Gallery is a new UI page for discovering and managing your Snowflake apps. The Apps Runtime is the engine behind it — it builds, deploys, and runs Snowflake apps (Streamlit + Next.js) directly from CoCo. One click and the agent vibe-codes a near-production-grade app for you. Run local apps with live preview and see status chips per card.

  • Chat Reliability: Three wins here: (1) mid-stream SSE socket drops and auth failures now auto-retry instead of killing your response, (2) VPN/network blips auto-reconnect without needing to quit and relaunch, (3) SSO auth now surfaces a fallback URL when the browser blocks the app from opening — this was a login blocker for some users and they can now always complete authentication. No more losing long responses to flaky networks.

  • Workspace Hooks: Hooks are now fully workspace-aware. See and edit workspace hooks directly in the Agent Settings panel, with per-hook “Open Config” buttons. Hooks from untrusted workspaces are blocked at the executor level.

  • Security and Workspace Trust: Write/edit tools now prompt when targeting paths outside your workspace (multi-root aware). Protected path checks use the tool’s actual working directory. Agent Mode auto-trusts, Editor Mode gates properly. 10 PRs hardening the trust boundary.

  • Smarter Subagents + Identity Awareness: Subagents are significantly smarter and produce higher-quality results. Thinking blocks, persistence, follow-up history, and turn limits all fixed. The main agent is now identity-aware — it knows your name, so no more hallucinated names in responses.

  • Snow CLI in Agent Terminals: The agent’s Bash tool terminals automatically get your active Snowflake connection injected, so snow sql just works without extra setup.

  • Platform and Security: Electron bumped to 39.8.10, addressing an OS-level edge case on macOS where auto-update could silently fail. 100% Windows binary signing — every single file is now signed, addressing security compliance requirements.

Notable fixes

  • GrepTool glob patterns now match recursively (no more root-only matches).
  • SQL Playground multi-statement execution works correctly.
  • SQL Playground autocomplete no longer clips on narrow panels.
  • Removed workspace reappearing after window refresh — fixed.
  • “Other” option in ask_user_question auto-selects when you click the text input.
  • Chat message padding fixed (no more overlap with input bar).
  • Remote notebook path resolution hardened against traversal.
  • Snowflake identifiers properly quoted in catalog view SQL.
  • File explorer rename/delete works on gitignored files.
  • MDX extension now allowed from marketplace.