Release history

Limitcode release notes.

A version-by-version record of visible features, behavior changes, and fixes in Limitcode Pro.

MCP 2026-07-28 stateless specification and native OAuth 2.1 flow

Native support for the MCP 2026-07-28 stateless HTTP transport and end-to-end OAuth 2.1 + PKCE authorization, allowing remote MCP servers like Cloudflare to connect directly without external Node bridges.

Added

  • MCP 2026-07-28 stateless HTTP protocol support (server/discover, MCP-Protocol-Version, Mcp-Method, Mcp-Name, _meta metadata, ttlMs caching, and pagination) with automatic fallback to stateful MCP servers
  • Native OAuth 2.1 + PKCE authorization flow with RFC 9728 Protected Resource Metadata, RFC 8414 Authorization Server discovery, Dynamic Client Registration (/register), local callback capture, and automatic token refresh

Changed

  • Simplified MCP configuration that auto-detects remote or local servers from url or command without requiring an explicit type field

Fixed

  • Standard User-Agent headers on remote MCP requests to avoid Cloudflare WAF Error 1010 (browser_signature_banned)

Prompt cache stability and Send to Agent context fix

Sending a selection or active file to the agent now injects the context once in the user turn instead of duplicating it into the system prompt, keeping the system prompt stable across turns for provider prompt caching.

Changed

  • System prompt stays stable across turns when attaching editor context, preserving provider prompt cache hits

Fixed

  • Removed duplicate Current Selection/Context injection from the system prompt when using Limitcode: Send to Agent

GitHub Copilot Auto session and Responses API support

Full support for GitHub Copilot Auto model routing and the /responses transport, so Copilot Free, Student, and Pro accounts work out of the box.

Added

  • Full GitHub Copilot auto session flow (/models/session, /models/session/intent, and Copilot-Session-Token management)
  • /responses endpoint support for Auto-selected models with streaming SSE, instructions, input items, and function_call round-trips
  • 26 automated regression tests for GitHub Copilot Auto and Responses API flows

Changed

  • Copilot model picker now parses structured capabilities metadata and scopes the list to models authorized for the signed-in account (defaulting Free and Student accounts to auto)

Fixed

  • Preserved assistant text alongside tool calls in Copilot history for accurate Responses API turn reconstruction

Permission engine fixes

Trust and always-allow decisions now work everywhere, with full visibility into every permission prompt and decision.

Added

  • Every permission prompt and decision is logged with its reason ([PERMISSION] ask/decision/denied)
  • Six permission regression tests

Changed

  • Trust Agent applies instantly mid-run and covers all tools, outside_project access and the active session; Untrust reverts everything

Fixed

  • Always-allow decisions and Trust Agent are now honored outside the project boundary (no more prompt loops)
  • permission.always_allowed entries from Limitcode.sublime-settings are now loaded (previously ignored)
  • git-bash /c/... paths are normalized to Windows form only when terminal_shell is git-bash
  • File denies (.env, *.pem, ...) still win over Trust

Stable LimitcodePro package and installers

The installed package is now LimitcodePro with a single version source of truth and one-click installers per platform.

Added

  • VERSION.txt as the single source of truth for the release version
  • build_release_bundle.py producing one Gumroad bundle for both py38 and plugin hosts
  • install.ps1 (Windows) and install.sh (macOS/Linux) with automatic host detection
  • Manual installation instructions in dist/INSTALL.md
  • Lifecycle, provider, and settings-command automated tests

Changed

  • Installed package renamed to LimitcodePro.sublime-package, fixing a Python module import error

Fixed

  • Settings resource paths corrected to Packages/LimitcodePro in commands and menus

Console cleanup and LSP diagnostics

Chat stays quiet during typing, and diagnostics are read through the official LSP package API.

Added

  • New debug_log setting (off by default) gating the sanitized JSON log file and status-bar error messages

Changed

  • Diagnostics are now read through the LSP package session API with guarded optional imports and polling instead of a fixed sleep

Fixed

  • Removed a leftover print in on_query_completions that flooded the console on every keystroke in the chat view

Chat style customization

The chat status bar can now follow your Sublime Text theme without changing the rest of the editor.

Added

  • New chat_style settings for chip background, border, text, link color, muted text, opacity, radius and padding
  • Advanced per-user override through Packages/User/Limitcode/chat.css
  • Automated coverage for chat style loading, fallbacks and invalid values

Fixed

  • Chip rows no longer overlap when the chat panel is narrow
  • Improved legibility of chips, links and muted text on dark themes such as Monokai Pro

Menu integration fix

Fixed the ordering of Limitcode entries in the Sublime Text main menu.

Fixed

  • Limitcode menu entries are now ordered consistently under Tools

Chat typography controls

The chat can now use its own font size without changing the editor or other Sublime views.

Added

  • Configurable chat_font_size setting with auto inheritance or a fixed numeric size
  • Increase, decrease, and reset chat font size commands in the Command Palette
  • Optional chat-only keybindings with no default Ctrl+Plus or Ctrl+Minus shortcuts
  • Automated coverage for chat font size behavior

Native OpenAI Responses transport

OpenAI now has a native Responses API path while the existing provider compatibility paths remain available for other services.

Added

  • Native OpenAI Responses provider with streaming and tool-call support
  • Improved OpenAI tool and reasoning support through the Responses API
  • Dedicated automated coverage for the OpenAI Responses transport

GPT-5.6 tool-call compatibility

Improved compatibility and reliability when using GPT-5.6 models with tools.

Added

  • Additional coverage for GPT-5.6 tool use

Fixed

  • GPT-5.6 tool calls now complete reliably with supported reasoning settings

More reliable tool history

The agent now preserves the actions it completed so follow-up turns retain a clearer picture of what happened.

Fixed

  • Preserved completed tool actions across chat turns
  • Improved continuity when the model needs to remember earlier tool work

Updated Antigravity model discovery

Google Antigravity now uses the verified Gemini Flash identifiers returned by its discovery catalog.

Added

  • Real Gemini 3.8 and 3.7 Flash model identifiers from Antigravity discovery

Fixed

  • Improved model listing and selection for the Antigravity provider

Reasoning, cards, and packaging polish

A broader release focused on richer agent feedback, cross-platform chat behavior, and a cleaner Package Control layout.

Added

  • Streaming reasoning display and richer tool, subagent, and batch cards
  • Clipboard image support across macOS and Linux

Changed

  • Shared runtime modules moved into lib/ to improve Package Control packaging

Fixed

  • Improved plugin loading and command registration after the package restructure

Stable tool status

Long-running tools now keep their progress indicator and Stop control in one stable place while output streams.

Added

  • Animated spinner inside the active tool card, so long commands show they are still working
  • Stop control moved into the running tool card instead of a separate floating status line

Fixed

  • Fixed the jumping/flashing status line during long command output (e.g. pnpm build)
  • Tool cards are restored in chronological order when reloading a chat history

History and loader fixes

Chat history restores tool calls where they happened, and running tools animate while they work.

Added

  • Persisted tool-call position inside assistant responses, so reloaded chats show calls in the order the agent made them

Fixed

  • Long-running tools now show an animated loader instead of a static dot

First stable release

The first packaged Limitcode Pro release, with license activation and image support across platforms.

Added

  • Clipboard image paste on macOS and Linux, matching the existing Windows support
  • Multimodal image attachments in chat, supported by image-capable models
  • License activation against the Limitcode Pro license service with device limits and offline grants
  • Separate downloads for current and next-generation Sublime plugin hosts

V1 release candidate

The first packaged Limitcode Pro V1 candidate brings the full coding-agent loop into a side-by-side Sublime Text view.

Added

  • Side-by-side Sublime chat with streaming responses, reasoning display, cancellation, and visible tool results
  • Code, Architect, Ask, and Debug modes with mode-specific tool access
  • Project tools for file listing, reading, searching, creation, targeted edits, patches, web access, and shell commands
  • Configurable permission rules: reads and searches can stay frictionless while mutating tools ask by default
  • Project-scoped conversation history, prompt history, and automatic long-context compaction
  • Agent edit snapshots with Limitcode Undo and Redo during the active session
  • Provider, model, reasoning-effort, context, and estimated cost controls in the chat status bar
  • OpenAI, Anthropic, Gemini, OpenRouter, Mistral AI, Copilot, DeepSeek, and OpenAI-compatible or local provider paths
  • Account sign-in for GitHub Copilot, OpenAI Codex, and Google Antigravity with a shared logout command
  • Web search and URL fetching as permission-gated tools with a configurable search backend
  • LSP diagnostics collected after edits to surface syntax and type errors back to the agent
  • Repository-level instructions that can join task context automatically
  • Reusable skills, configured MCP connections, checklists, batches, and focused subagents
  • Dynamic model discovery with per-provider caches and reasoning-effort controls
  • Session cost statistics and token tracking in the chat status bar

Changed

  • Separate downloads now support both current and next-generation Sublime plugin hosts

Fixed

  • Improved startup and module loading from the installed package