Cortex Code changelog

This page documents notable changes to Cortex Code CLI.

1.1.52

Added

  • cortex agent-studio commands: The commands for managing Semantic Views and Cortex Agents (read, write, deploy, query, and evaluate) are now available in the CLI without an experimental flag.
  • MCP OAuth client secrets: MCP server OAuth configuration now accepts a client_secret, so you can connect to servers that use a pre-registered confidential OAuth client.
  • Inherit a subagent’s model: The task tool accepts model: "inherit" so a subagent runs with the same model as the session that spawned it. You can also set model: inherit in an agent definition’s front matter.
  • More object types in object search: cortex search object can now search for streams, stages, tasks, and pipes with --types, in addition to the object types it already supported.

Fixed

  • Proxy support for model streaming: The CLI again honors the HTTP_PROXY and HTTPS_PROXY environment variables when it streams responses from the model. Proxy support on this path regressed after Cortex Code moved to the Node.js runtime.
  • Selection prompts: Fixed rendering problems in interactive selection prompts, such as permission approvals, plan review, and clarifying questions. Long option lists could hide choices, the prompt could overwrite earlier transcript output, and text could be truncated even when space was available.
  • Windows sign-in: Fixed a “Cannot find module” error during connection setup on Windows that could interfere with OAuth sign-in, including authorization for MCP servers.

Changed

  • Fewer prompts for read-only system functions: Expanded the read-only SQL allowlist so that more read-only SYSTEM$ functions run without a permission prompt. For example, SYSTEM$GET_SERVICE_STATUS, SYSTEM$CLUSTERING_INFORMATION, and SYSTEM$READ_YAML_FROM_SEMANTIC_VIEW.
  • Catalog skills install globally: Skills that you install from the catalog now always install to ~/.snowflake/cortex/skills/, so they’re available from any Cortex Code session no matter which directory you start from. Skills installed by earlier versions continue to load.
  • Live output for background subagents: When you wait on a background subagent, its output now streams live as it runs instead of appearing all at once after it finishes.

1.1.47

Added

  • MCP OAuth for remote and headless machines: HTTP MCP servers that require OAuth can now be authorized by opening the printed authorization URL in any browser and pasting the redirect URL back into the terminal, so sign-in completes even when the OAuth callback can’t reach a cloud workspace or remote development box.
  • FIPS 140 mode for government and FedRAMP accounts: The CLI now bundles the CMVP-validated OpenSSL 3.1.2 FIPS provider across all supported platforms and enables FIPS 140 mode automatically on government and FedRAMP deployments.
  • Background monitors view: Added the /monitors command and monitor detail panels to inspect long-running background processes, plus a footer hint that shows the number of active monitors.
  • Warehouse credit attribution: Added the COCO_ADDITIONAL_QUERY_TAGS environment variable to inject custom key-value pairs into the QUERY_TAG on every Snowflake connection, so you can attribute warehouse credits by application or team.

Fixed

  • Plan mode commands: Bash commands in plan mode again defer to the normal permission system instead of being blocked. A previous change restricted plan mode to a handful of commands, which prevented routine read-only commands and produced worse plans.
  • Model picker: The /model picker now shows a clear warning when it can’t reach the model service and falls back to a default list, and it no longer hangs on a stalled request because of a new 15-second timeout.
  • Windows connection sync: Sync from Snowflake no longer hangs on Windows hosts where localhost resolves to IPv6 first.
  • Profile handling: The CLI no longer crashes when a profile name is empty or undefined.

Changed

  • Faster Cortex Extension plugin installs: cortex plugin install now downloads extension files in parallel, which greatly reduces install time for large extensions. For example, a 108-file extension dropped from about 50 seconds to about 12 seconds.
  • Project settings source: Cortex Code no longer reads .claude/settings.json as project settings. Only its own .cortex/settings.json is used, so move any project-level configuration into .cortex/settings.json.

1.0.77

Added

  • /changelog command: View release notes directly in the CLI.
  • /mcp manager overhaul: Split-pane view with search and filter, add/edit/delete server controls, source badges, OAuth state display, invalid-tool diagnostics, and correct disable semantics.
  • SQL read-only mode: --sql-read-only launch flag and /sql-writes on%off%status slash command to restrict write operations.
  • Project skill discovery: Added .agents/skills as a project-level skill discovery directory alongside the existing .snowflake/cortex/skills.
  • Clipboard shortcut: Added ESC+C to copy the current input to the clipboard.
  • Fast in-repo search: Added /index command backed by an FTS5 trigram index for instant grep across large codebases.
  • ACP session identity: Editor integrations can now supply a session ID via _meta.sessionId in ACP requests.
  • GPT 5.5 model support: Added GPT 5.5 to available coding and reasoning models.
  • Mac workspace acceleration: ws create --hv=vz uses the Apple Virtualization framework directly for ~1-second boots without Lima.
  • Rule management: Session-scoped rule suppression and inline editing of rule text, category, and tags.
  • Agent SDK: Support for custom MCP tool definitions and hook callbacks.

Fixed

  • Role switching: USE ROLE via sql_execute no longer causes 403 errors on subsequent agent API calls (dedicated connection isolation).
  • Large command output: Bash commands producing more than 30K characters are no longer silently truncated; output between 30K–50K is preserved inline and output over 50K is offloaded to a file.
  • Plan mode: Plan mode is now correctly disabled immediately after the user approves a plan.
  • Memory search: No longer silently disabled when the snowmem embed endpoint returns a malformed response.
  • Snowflake OAuth retry: Retrying OAuth no longer purges all stored credentials.
  • data_diff: Now works correctly on user machines after bundling the required PBS Python runtime.
  • Snowmem embed parsing: Fixed incorrect 1-dimensional response handling; responses are now correctly parsed as 1024-dimensional vectors.

Changed

  • Permission prompts: tgrep and cron_create now prompt for permission approval instead of auto-approving.

1.0.65

Added

  • Managed plugins and marketplace installs: Added registry-backed plugin install, update, enable, disable, validate, and uninstall flows, plus marketplace-backed installs and a fullscreen plugin manager.
  • Runtime plugin refresh: Added in-process plugin reloads so lifecycle changes can refresh commands, skills, agents, hooks, and MCP servers without restarting Cortex Code.

Changed

  • Disabled plugin handling: Switched disabled plugins to explicit activation metadata so inactive installs stay discoverable without loading their full runtime contents.

1.0.59

Added

  • Postgres connections and SQL workflows: Added Postgres connection management, Snowflake Postgres token authentication, and dialect-aware /sql execution across Snowflake and Postgres.
  • ACP editor integration: Added cortex acp serve with streaming editor sessions, native permission routing, clarifying-question prompts, and diff-based file edit previews.

Changed

  • Postgres safety and credential handling: Hardened Postgres permission analysis, secured Snowflake Postgres token generation, and removed shell-exposed password flags in favor of secure prompts or ~/.pgpass.

1.0.28+173700.0975210405a4 (2026-03-02)

Added

  • Faster guardrails: Reduced guardrail latency so protected requests feel less disruptive during streaming.
  • Browser workflows: Added cortex browser and improved browser recovery flows.
  • Team workflows: Expanded swarm and background-agent collaboration flows.

Changed

  • Sandbox coverage: Expanded sandbox support across more platforms and workflows.
  • Profiles and tool controls: Added richer profiles and easier tool allow or deny controls.

1.0.17+201538.00c735fc7e3a (2026-02-19)

Added

  • Windows platform support: Added broader Windows compatibility across shell, file, and CLI workflows.
  • Background agents: Improved reconnect, resume, and navigation flows for /agents.
  • Permission rules: Added configurable permission rules for safer command execution.

Fixed

  • Auth and credential resilience: Improved token storage, migration, and recovery for authentication flows.
  • Shell safety and controls: Strengthened shell safeguards and expanded usage, model, and view controls.

1.0.11+011422.721be00264d0 (2026-02-12)

Added

  • Skill updates and remote plugins: Added versioned skill updates and support for plugins hosted in GitHub repositories.
  • Inter-agent coordination: Added agent-to-agent messaging and stronger multi-agent task coordination.
  • Model discovery: Improved model discovery and request routing.

Changed

  • Resume and sandbox controls: Improved session resume behavior and sandbox testing controls.

1.0.10+223843.7a31e1a1441a (2026-02-11)

Added

  • Skill and plugin packaging: Bundled the Airflow plugin and added tarball-based skill installs.
  • Performance and output handling: Improved startup time and large tool-output handling.
  • Notebook execution robustness: Improved notebook execution permissions, reliability, and diagnostics.

Changed

  • Inter-agent orchestration: Expanded agent coordination and task flows.
  • Model and skill management: Improved model discovery and skill update flows.

1.0.9+154353.e3dd198d62ee (2026-02-09)

Added

  • Instruction management: Added instruction settings and clearer visibility into loaded instruction files.
  • Fork and rewind reliability: Made /fork and /rewind more reliable and easier to follow.
  • Custom slash commands: Expanded custom slash command naming and collision handling.

Changed

  • MCP workflow: Improved live updates when changing MCP server configurations.
  • Rendering polish: Reduced noisy spinners and improved transcript rendering behavior.

1.0.7+015551.0d10922ac7fe (2026-02-06)

Added

  • Background agents and /agents: Added background-agent launch flows and stronger plan handling for interactive sessions.
  • Settings and fullscreen controls: Added tabbed settings views and a more capable fullscreen settings experience.
  • Session browsing and docs: Added session search and a /docs shortcut to open documentation.

Changed

  • Model selection: Expanded /model options and improved default model behavior for subagents.
  • Semantic and dbt helpers: Improved semantic and dbt assistant workflows.

1.0.6+020057.75a0ba5e6a69 (2026-02-04)

Added

  • Snowflake config and path handling: Improved Snowflake home and config resolution plus safer connection setup workflows.
  • Secret and connection tooling: Added cortex secret and better support for local connections.
  • Worktree and remote-source improvements: Expanded worktree creation and improved remote skill and repository sourcing.

Changed

  • Notebook actions: Improved notebook permissions and notebook-action reliability.
  • CLI and session resilience: Improved interrupt handling, screen clearing, and table rendering in the terminal.