TKCode 0.1.4

Adds a Plan mode that ends in a plan you approve, project memory for the built-in agent, a context usage card that shows the prompt cache hit rate, automatic session titles with a separate fast model, configurable shells, interface zoom, and tray, Dock and jump-list menus with single-instance launch. Prompt caching now covers every Anthropic Messages endpoint and most Chat Completions endpoints. The app opens without a loading page and now builds on .NET 11 RC2, AtomUI 6.2.6 and Avalonia 12.1.4.

Added

• Plan mode with plan approval. In Plan mode the agent researches, writes its plan to a plan file (~/.tkcode/plans/<session>.md) and calls exit_plan_mode. An approval panel replaces the composer with these choices: approve and run automatically, approve and review each step, approve and hand off to ACP, or clear the context and run from the plan alone (built-in agent only). "Keep planning" sends your feedback back to the agent. You can edit the plan in the panel or open it in an external editor before approving. After approval the session continues in the same turn, and the plan stays in the transcript as a card that shows its outcome.
  • In Agent mode the model can call enter_plan_mode to propose planning first. You confirm in the question panel.
  • Plan mode is read-only. write and edit work only on the plan file, and subagents are limited to explore and plan. Commands that might change the workspace ask for approval in every permission mode, including Full access.
  • Claude Code's ExitPlanMode opens the same approval panel. When the CLI enters or leaves plan mode by itself, TKCode's mode follows it. Codex and ACP are not connected yet.
• Project memory (autoMemory, on by default). The built-in agent keeps notes in ~/.tkcode/projects/<project>/memory/, and all worktrees of a repository share them. The MEMORY.md index (first 200 lines or 25 KB) is loaded when a session starts and again after compaction. The agent reads and writes memory files without approval prompts. Subagents and external runtimes are not affected.
• @import in AGENTS.md. Context files can include other files with @path: relative, ~/ or absolute paths, up to 4 levels deep, ignored inside code. Project files can only import files inside the project. The global ~/.tkcode/AGENTS.md can import from anywhere.
• Context usage card. Hovering the context ring under the composer shows tokens used against the budget, a segmented bar, and a breakdown into system prompt, system tools, MCP tools, skills, messages and other. It also shows the session's average prompt cache hit rate and warns when auto-compaction is close. External runtimes show only the total and the hit rate.
• Automatic session titles. After the first message, a separate request generates a short title and saves it with the session. A manual rename always wins. If the request fails or takes more than 30 seconds, the first line of the message stays as the title. External runtimes with no fast model configured use a temporary read-only session of the same CLI.
• Fast model (fastModel, Settings → General → Model configuration). Subagents and session titles use it. By default it follows the session model. An explicit model passed to the agent tool still takes precedence. Only built-in API models can be selected.
• Prompt cache lifetime (promptCacheTtl: 5m or 1h, Settings → General). It applies to every Anthropic Messages endpoint and to Claude models over Chat Completions. Running sessions pick it up on their next request.
• Interface zoom (uiScale, 0.75–2). Text, icons and spacing scale together, popups included. Settings → General offers six steps from 80% to 150%. ⌘+ / ⌘− / ⌘0 (Ctrl on Windows and Linux) and the command palette step through the same levels. Tooltips and confirmation dialogs do not scale yet.
• Shell selection (agent.shell.profile, terminal.defaultProfile). Settings → General lists the shells detected on the machine, separately for the agent's bash tool and for new terminals, plus Auto and a field for a custom path or program name. A shell that is not available falls back to Auto and is marked in settings. A new agent shell applies from the next command and updates the shell named in the system prompt. A new terminal shell applies only to terminals opened afterwards.
• Tray, Dock and jump-list menus. The tray menu gains New session and a Projects submenu with recent projects. On macOS the Dock menu offers New session and recent projects. On Windows the taskbar jump list offers New session, Quit TKCode and recent projects.
• Single instance. Launching TKCode again brings up the running window, even one hidden in the tray, instead of starting a second copy. The running instance receives --new-session, --open-project <path> and --quit from the command line. --new-instance skips the check.
• File attachments. Dropping, pasting or picking a file that is not an image adds an attachment tag. The file is not uploaded: the agent gets its absolute path and reads it with its own tools. The attach picker accepts any file and has Images and Documents filters.
• Clickable file paths in replies. These open in the workbench preview:
  • Markdown links to local files, including path:line suffixes and Windows drive paths.
  • Inline code that names an existing file.
  • Code blocks whose fence names a file (for example ```cs src/App.cs), through an "Open preview" button.
• NextCoWork provider (nextcowork, key from COWORK_API_KEY). It uses the Responses API by default and the Messages API for Claude models. Once a key is set, its model list comes from /v1/models, minus image-generation models.
• Crash logs. Unhandled exceptions in the desktop app and the Agent Host are written to ~/.tkcode/logs/crash-*.log. The desktop app also writes a daily desktop-YYYYMMDD.log next to the host log.

Changed

• No loading page at startup. The main window appears right away instead of after the loading page, and does not wait for the Agent Host. The quips from the old loading page now appear on the new-session welcome screen.
• Full access is now remembered. Each project remembers its permission mode for new sessions, Full access included. Reopened sessions and sessions continued in a new session keep their mode. Before, Full access always reset to Ask. The confirmation dialog now says this.
• The primary model starts every new conversation. The model setting, now labelled "Primary model", is the starting model for every new conversation, ahead of the model you last picked. Changing it does not switch existing sessions.
• New default shells on Windows.
  • The agent's bash tool now tries the latest stable PowerShell 7 first, then Git Bash, then Windows PowerShell. Before, Git Bash came first.
  • New terminals open PowerShell 7, then Windows PowerShell, then cmd. Before, they opened cmd.
  • Output from PowerShell commands the agent runs is forced to UTF-8, which avoids garbled text on non-English Windows.
• Prompt caching.
  • Anthropic Messages requests now carry one cache breakpoint on the system prompt and one on the last message. This replaces the top-level automatic cache_control, which is still sent to Kimi. Breakpoints go to every Messages endpoint, not only api.anthropic.com. Requests also carry a per-session metadata.user_id, so relays keep a session on one upstream account.
  • Chat Completions requests carry a per-session prompt_cache_key on every endpoint except Gemini's OpenAI-compatible one. Claude models (through OpenRouter and relays) and Qwen on DashScope get cache_control breakpoints. xAI and OpenRouter also get their session-affinity headers.
  • Compaction summary requests now read the cached conversation prefix instead of paying full price for it again.
  • If an endpoint rejects these fields, set compat.supportsCacheControl: false or compat.supportsPromptCacheKey: false for that provider.
• The session list shows 20 sessions. A Show more button adds 20 at a time, replacing loading on scroll.
• User messages longer than 8 lines are collapsed, with Show more / Show less.
• The per-turn changes card lists the first five files and folds the rest behind "Show N more".
• Request errors appear in a card that wraps, can be selected and copied, and scrolls when long, for example a JSON error body.
• The send and stop buttons are now solid accent circles that fade when disabled.
• Text and thinking blocks are marked finished as soon as the block ends, not when the whole message or tool run ends.
• Removing a provider now asks for confirmation, and every provider has a remove button.
  • A built-in provider detected from an environment variable is hidden with "hidden": true. The variable itself is not changed.
  • An unsaved draft is discarded without asking.
• StreamMark is updated to 0.1.0-alpha.23, built against Avalonia 12.1.4. Fallback fonts that only have a synthetic italic face render upright instead of loading a synthesized copy of the font.
• TKCode now runs on .NET 11 RC2 with Runtime Async, AtomUI 6.2.6 and Avalonia 12.1.4. On x64, .NET 11 requires a CPU with x86-64-v2 support.

Fixed

• A tool error in a Claude Code session no longer crashes the app. The crash happened because Claude Code reports some tool results as a plain string.
• Selected text in replies is readable in the light theme. It used to be black on black.
• The copy and fullscreen buttons on tables in replies are no longer covered by the header text. They now sit in a bar above the table (#7).
• originalImages, defaultRole and skills in settings.json no longer produce "Unknown field" warnings.
• The context ring no longer flickers during external-runtime turns. Those runtimes report zero usage on intermediate messages, and that zero used to replace the last real value.
• Plan mode could run shell commands that change the workspace without asking when the permission mode was Auto-approve or Full access. Plan mode is now read-only (see Added).
• Claude Code's ExitPlanMode went to the generic approval card as raw JSON, or was allowed without asking, and could leave TKCode in Plan mode after the CLI had left it. It now goes through the plan approval panel.
• Dropping files that are not images no longer also attaches the drag preview icon as an image.
• Subagents that run on a different model now use the closest thinking level that model supports, instead of inheriting one it does not.
• Terminal: a zero-width character such as U+200B added to the previous cell no longer leaves a stale cell on screen. This is fixed with a libghostty-vt overlay patch.
• The branch picker under the composer follows branch changes made outside the app. A branch created, switched or deleted in the terminal or by the agent now shows up without switching projects. The list is also reloaded each time the picker opens.

Build

• Building requires the .NET 11 RC2 SDK (11.0.100-rc.2.26504.105 or a later feature band; global.json allows prereleases). It also requires a .NET 10 Runtime, because AtomUI 6.2.6's build tools need it. All projects target net11.0 and enable Runtime Async, NativeAOT builds included.
• AtomUI 6.2.6 changes how the app starts. AtomUI is configured once in Program.UseTKCodeAtomUI(), and every AppBuilder that starts App must call it: desktop, headless tests and smoke mode. See docs/architecture.md §7.1.
• Desktop and the memory harness set UseOptimizedPublish=false. Without it, .NET 11's optimized publish can succeed but leave out the compiled XAML. Check published builds with --smoke.
• The new TKCode.Shell library holds the shell detection shared by the agent and the pty host. It has its own test project, which the terminal workflow runs on all three platforms.
• Local builds are single-instance too. To run one next to an installed TKCode, pass --new-instance or use a different TKCODE_HOME. --crash-test throws an unhandled exception so you can check crash logging.
• --startup-bench renders the first frame of a published build headlessly, and tools/startup-bench.py measures startup with it.
• tools/streammark/pack.sh packs the packages under their local StreamMark.* IDs. On nuget.org the packages are named TKCode.StreamMark.*.
• In a build without the terminal native library, a terminal tab now says the library is missing and points to tools/terminal/build-ghostty.sh.
• Release CI installs the .NET 10 SDK alongside .NET 11. On every platform it runs the Agent Host --self-test and a headless Desktop --smoke on the published build.

Full Changelog: v0.1.3...v0.1.4