日本語

Teaspoon IDE News

New features and improvements in each release. Newest first.
Product page / Settings guide

v1.4.0 2026-10-16

  • Your API key never reaches the app's UI anymore The key was already encrypted in the OS keychain, but it used to be read back into the app window when sending requests. Now the app sends requests from behind the scenes (the Electron main process) and the UI only ever learns that a key is saved. The key field stays empty at all times and only replaces the stored key when you type a new one. If a saved key gets rejected by the API, a warning tells you, so expired or mistyped keys are easy to spot.
  • Model list fetched live from the API — older generations move to the bottom Once a key is set, the model picker shows the live list returned by Google's API (falling back to the built-in list when unreachable). The API has no deprecation flag, so models a full major version behind the newest are marked "(legacy)" and sorted to the end. Custom model IDs still work as before.
  • Internals: provider registry LLM calls now go through a provider registry, and each model carries capability metadata (streaming, prompt variant, history window). The unused native function-calling path and the SDK dependency were removed. Ollama, organization mode, and the file-operation commands are unchanged.

v1.3.1 2026-10-15

  • Gemini model list refreshed — now featuring Gemini 4 Argon Google's flagship Gemini 4 Argon and Gemini 3 Pro (Preview) are now in the model picker. Retired models (the Gemini 1.5 line, etc.) were removed and the preview model IDs were corrected to their official names. The default stays Gemini 3.8 Flash, with Gemini 3.5 Flash-Lite as the low-cost option. Models not in the list can still be typed into the Custom Model field.
  • Dependency updates and a security fix Updated the dependency stack, including the move to React 19, and resolved a security advisory (a shell-quote vulnerability). The fallback model used when the selected one is unavailable is now gemini-3.8-flash.

v1.3.0 2026-10-11

  • Bubbles dressed for the theme — five themes get dedicated colors The chat bubbles no longer just borrow the accent color: five themes now wear colors chosen for their own world. Chaya gets a three-color-dango pink, Otegami a kraft envelope for you with a white envelope for the AI, Soda Float a white ice-cream foam floating on soda blue, Muted Ocean a muted bubble blue, and Rich Wine a wine-cork brown. All other themes are unchanged.
  • Internal library update React, which renders the UI, was updated to 18.3.1 — the final React 18 release, easing a future React 19 migration. No visible or behavioral changes.
  • For developers: window-size flag and a dev-mode fix You can now launch with an exact window size (e.g. --window-size=1372x734), handy for screenshots. Also fixed a dev-mode bug where Electron could open a different project's dev server instead of Teaspoon.

v1.2.0 2026-10-10

  • Stay responsive while generating — CPU Threads setting You can now cap how many CPU threads Ollama may use for inference, so a local model no longer pins every core while it generates and the PC stays usable for other work. Leave it empty to let Ollama decide. Found under Settings > LLM Provider. Read the measured-effect article.
  • Thinking on/off for reasoning models For models that support a thinking mode (qwen3.5, gpt-oss, ...), you can force reasoning On, Off, or leave it to the model ("Model default"). Turning it off makes replies dramatically faster on slower hardware.
  • Ollama connection moved to the native API Teaspoon now talks to Ollama's native /api/chat endpoint instead of the OpenAI-compatible /v1 API — the two settings above exist only on the native endpoint. Side effect: the Ollama endpoint field must point at a real Ollama instance; generic OpenAI-only servers (e.g. LiteLLM) are no longer supported there.

v1.1.2 2026-10-09

  • Cleaner settings storage (security scan follow-up) All settings writes now funnel through a single storage helper, so API keys can only ever be saved through the OS keychain — there is structurally no path that writes them as plain text. No visible or behavioral changes.
  • Future-proofed the build config The development build config (vite.config) was converted to a format compatible with the next major Vite release's new loader. No effect on the app itself.

v1.1.1 2026-10-08

  • API keys are now stored encrypted Both the personal Gemini key and the organization-managed session key are now encrypted with the OS keychain before being saved (they used to sit as plain text in the app's storage area). Keys saved by older versions migrate automatically on first launch — no need to re-enter them.
  • All known library vulnerabilities resolved Cleared all 49 known vulnerabilities found in the npm packages the app is built from. No visible or behavioral changes.
  • Security reporting and automated monitoring The GitHub repository now has a private vulnerability reporting channel (SECURITY.md) and Dependabot, which watches dependencies for vulnerabilities and updates automatically.

v1.1.0 2026-10-08

  • No reading or writing outside your project Even if an AI instruction carries a wrong or malicious path, the app's core (the main process) now blocks access to files outside the opened project — a second layer of defense on top of the chat-side check.
  • Truncated write commands are no longer missed If an AI response is cut short by a timeout and a file-write command comes out incomplete, it is now detected and the AI resumes the file from the fragment. Previously, a truncated write went unnoticed when other files in the same response succeeded, leaving a half-written file behind.
  • No more "I'll continue in my next reply" If the AI tries to defer remaining work to a future response, the app automatically keeps it working until the task is done.
  • Unresolved edits get fixed first If you try to open the result in a browser while a file edit or run failure is still unresolved, the app finishes the fix first instead of launching a half-finished app.
  • The AI no longer launches apps on its own Creating files no longer triggers an automatic run or preview. Ask explicitly when you want something started.

v1.0.0 2026-10-05

  • Six new themes — 14 in total Violet Fizz (a violet cocktail), Otegami (bright washi paper with ink and vermilion), Soda Float (sky-blue soda with pale foam), Modern Syntax eXtensible (a retro MSX-blue screen), Chaya (deep tea-field green), and Coquette (pale blush ribbons and lace). Pick any of them under Settings > Appearance > Theme.
  • Localization, down to the last corner The remaining English in menus, the right-click context menu, native dialogs, and error messages now follows the UI language. Switch languages anytime in Settings > Appearance > Language.
  • A tidier Settings dialog Options are now grouped into collapsible panels — Appearance, AI Context, LLM Provider, and more — and the open/closed state is remembered across restarts.
  • Safer API-key removal "Clear API Key" now asks for a confirming second click and removes only the key itself. Your model selection, custom model name, and proxy URL stay put, so rotating a key no longer resets your preferences.

v0.9.1 2026-10-04

  • Commands work reliably even with small models The app now accepts more of the mistakes weaker models make — a missing "//" prefix or commands written in JSON format still execute. Conversely, command examples inside code blocks never run, so you can safely ask the AI how commands work.
  • Opening URLs in the browser is fixed Asking the AI to open a URL used to fail because the address was cut off at the "//" inside it. Targets given as markdown links now open correctly too.
  • Close-project command Ask the AI to "close the project" and it can now do so. Your running conversation and history are kept.
  • Warning before using an existing folder When creating a new project, choosing a folder that already contains files now shows a warning — preventing new files from being silently mixed into an existing directory.
  • Closing a project no longer disrupts the chat Fixed an issue where closing a project could lose an in-flight AI reply or switch to a different conversation. Reopening the project restores the right conversation.
  • Privacy and display improvements Command results now show project-relative paths instead of full paths, so your OS user name no longer appears in chat history or gets sent to cloud providers. Also fixed tables overflowing the chat bubble and header buttons being pushed out of view.

v0.9.0 2026-10-03

  • Chat focus mode A button in the chat header (or Ctrl+Shift+B) hides the sidebar, editor, and terminal so the chat fills the window. Pane sizes are remembered and panes come back automatically when an action needs them. The mode persists across restarts.
  • Conversation list In focus mode, a rail beside the chat lists your saved conversations with title, project, and timestamp. Click to reopen one, delete it, or start a fresh chat with "+ New chat". Conversations without an open project are saved too and restored on the next run.
  • Create-project prompt If the AI wants to create a file while no project is open, a dialog now offers to create a project folder on the spot (under Documents by default) instead of failing.
  • Automatic summaries for long chats When a conversation gets long, older messages are summarized automatically before being sent to the AI, so smaller models can keep working (can be turned off in Settings).
  • New theme "Rich Wine" A dark, bar-lit wine palette where no code color looks like an error message.
  • Easier-to-read AI answers AI replies now render with headings, lists, tables, and code blocks.
  • More reliable chat history History is now stored differently, fixing the issue where conversations could stop being saved once they hit the storage limit.
  • Tighter security The AI can no longer read or write files outside the currently open project.

v0.8.0 2026-10-03

  • The AI understands command failures better When a command fails, the AI is now told what kind of failure it was (failing test, command not found, etc.). It used to misread failures as environment problems and waste many steps probing.
  • No more confusion over pre-existing test failures Tests that were already failing before the AI made any changes are no longer mistaken for something it broke.
  • Never reports "done" over a failed edit A failed file write or edit stays on the AI's radar until it succeeds. If the task still ends with the change unapplied, a warning tells you it may not exist on disk.
  • A tidier terminal Finished commands collapse to a one-line status row automatically; click to expand the output when you need it.
  • Dialog layout fix Fixed the Open Project dialog, where buttons could overflow and overlap the content below.

v0.7.0 2026-10-01

  • Four new themes Muted Ocean (deep-sea blue), Ancient Console (eye-friendly green CRT), Walnut (dark wood), and Heritage (retro-PC light). "Quiet Light" is renamed "Organic Light" — your existing selection keeps working.
  • Edits are verified by tests When the project has a test suite, the AI now runs it after editing and fixes regressions before giving its final answer.
  • Readable text in light-accent themes Fixed text that was hard to read on buttons and chat bubbles in themes with light accent colors.
  • More reliable editing with small models The app tolerates small mistakes in the AI's edit format, and if edits keep failing it tells the AI to rewrite the whole file instead.
  • Finishes promptly when done The AI now knows how many work steps remain, so it stops re-checking finished work and wraps up with a plain answer.

v0.6.0 2026-09-29

  • Renamed: Forger → Teaspoon IDE Renamed to avoid confusion with similarly named projects. The name reflects the goal: getting real work done on a teaspoon of tokens.
  • Explorer auto-refresh When files are added, deleted, or renamed outside the app (another editor, File Explorer, git commands), the Explorer now updates automatically — the Git panel refreshes too. A manual refresh button (⟳) was also added.