We can't wait to launch our new feature: a brand-new human-AI interaction mode where the AI proactively asks questions whenever it's in doubt. It fills in missing details, catches edge-case logic gaps you'd normally overlook, and continuously aligns on specifications to eliminate endless back-and-forth rework.
I believe this is crucial—until now, too many people vibe coding have had no real idea what they're actually doing, completely blind to the underlying execution pipelines and core project logic.
https://t.co/TxgOudyrNf
#VibeCoding #DevSec #BuildInPublic #OpenSource #AIWorkflow
The gap between GPT and Claude is surprisingly vast—and it always reveals itself in the subtlest details.
Recently, I built a web dashboard and implemented multi-language (i18n) support across English and several other languages. Due to varying string lengths, certain layouts looked far less aesthetic than expected.
I tested multiple models on this: • Astra and GPT-6.1 barely had a clue about the visual imbalance. • Opus 5.5 immediately spotted the issue and proactively refined the layout aesthetics to keep the UI clean.
True visual sensibility in code generation is still where Claude stands apart.
#Claude #VibeCoding #BuildInPublic #Frontend
Breezell does not collect user data. Recently, I received feedback regarding data security concerns—specifically highlighting a deep distrust of closed-source applications in the Vibe Coding era. I suspect this is a ripple effect from the recent ZCode incident where local code and full Git histories were silently uploaded.
Based on our investigation: about 90% of AI applications from large Chinese tech companies collect user data, including your code and workflows. This isn't fear-mongering—when they have enough resources, they always manage to pull off unexpected moves. And frankly, I wouldn't rule out coercion.
We simply don't have the storage capacity to collect user info... honestly, we don't even have the storage infrastructure for it...
Learn more about us: https://t.co/eq3qoRdrhz
#VibeCoding #ZCode #DataPrivacy #OpenSource #DevSec
We can't stop the models themselves from collecting data—like GPT collecting user data—that has nothing to do with the tool itself (not that I'm saying GPT actually does). The truth is, being a toolmaker is a really awkward spot to be in. It's out of most people's hands because you just can't avoid using their products; all we can do is focus on doing our part right... Our stance hasn't changed at all, simply because collecting user information was never something we ever planned to do.
@thsottiaux I have canceled my ChatGPT subscription. You’ve made too many changes recently—the model is painfully slow, often runs for hours without completing tasks, and constantly generates numerous errors. I've decided to switch to Claude.
1.4.1 · 2026-10-01
New Features
Memory pre-analysis (experimental) in Box: Relevant memories are selected before each reply. The memory analysis page shows streaming progress. The model slot is named MemoryPreflight.
Export memory packs directly from Box.
Desktop sign-in now confirms in the browser before returning to the editor.
Performance monitoring toggle in Box: Performance monitoring can be switched on or off. Metrics group windows by remote server, and leftover extension hosts are cleaned up.
Terminal buffers are cleaned up automatically.
Quote earlier conversations into the current session, marked with a golden robot icon, and cleared together with the input tray.
AI browser operations now run in hidden windows by default; page snapshots switch to reading the accessibility tree, configurable in Box feature options.
Added support for AMD model provider, GPT-6.1 Sol, and Gemini 4 Argon.
Fixes
Keep All / Reject All only applies to the current conversation. Committed renames and new files settle properly, allowing the action bar to clear.
Tool steps hold at the loading row until results arrive, then transition to the completed state.
Streaming code fences now close according to CommonMark specifications. Thinking effort sent to Gemini is restricted to tiers actually supported by Code Assist.
Memory analysis no longer fails when encountering an empty {}. Search planning is more accurate and consumes fewer tokens.
Script cards now only display the count of modified files (for example, "1 edited") instead of relisting every filename.
Clear error details are provided when subscription validation fails. Subprocesses stopped early terminate immediately without lingering.
Resolved layout jitter and twitching issues on the conversation page.
Various minor bug fixes and stability improvements.
Improvements
Diffs in Code Review are now mounted on demand, layout measurements are batched, and memory retrieval is noticeably faster.
#AI #DevTools #Coding #Gemini #AMD #Changelog #Breezell
https://t.co/TxgOudyrNf
1.3.9 · 2026-09-22
New Features
Preconfigured models: Claude Opus 5.5, GPT 6 Sol, GPT 6 Terra, GPT 6 Luna, Gemini 4 Pro, Qwen4 Max, Qwen4 Plus, Qwen4 Flash, and Qwen4 27B.
China Telecom Xingchen: Xing4.0-29B.
Xiaomi: MiMo-V2.6 Pro, Pro UltraSpeed, and Flash. Also MiMo-V2.5 and V2.5 Pro.
Tongyi: Qwen3.8-2.4T-A95B, Qwen-Doc-Turbo, and Qwen-Long.
Sign in to Gemini with a Google account (Antigravity Code Assist).
Image generation now includes Hy Image 3.5 and Qwen Image 2.1.
When Super Memory is used up, a notice card appears above the input box.
Editor notifications can show images and links; documents open on click.
Conversation export and import now have a progress panel and can be canceled mid-way.
Grok shows your current subscription tier.
In Box, broken compatible services and models get a badge and can be retested individually.
Fixes
Canceling generation no longer wipes the reply that already streamed.
Streaming output no longer jitters line height; tool cards stay clickable.
Right-clicking the title bar no longer crashes the window.
Abnormal values or looping data no longer leave the sidebar blank.
Chat sidebar failed to load because of a CSS comment syntax error.
Backslashes in Windows paths are preserved; if a file read fails, Breezell retries with forward slashes.
Looking up a Git commit no longer walks the entire history.
Improvements
Long conversations scroll more stably.
Editor performance is faster.
https://t.co/TxgOudyrNf
#GPT6 #GPT6Sol #Claude #Opus55 #Gemini4 #Qwen #Xiaomi #MiMo
1.3.8 · 2026-09-19
More models in one editor, cleaner switching mid-conversation, and a tighter Plan loop — same thread, same context.
New Features
Providers & models: StepFun and AiMass are now first-class providers. Qwen 3.8 lands with the 27B and Omni-Flash variants. Zhipu GLM-5.3-FlashX is in the list.
Doubao Seed 2.1: Pro 0915 and Evolving are available.
Official protocol wiring: OpenCode, Zhipu, Kimi, and DeepSeek now follow each vendor’s official plan tiers and protocol formats (Chat / Responses / Anthropic).
OpenRouter: New slugs appear in the model list. Free models show a FREE ribbon; retired models are stamped Unavailable.
Model picker in chat: Switching models opens an in-panel picker with Models and Compatible tabs.
Models page: Jump back to top, or jump to the latest additions.
Fixes
Paste image: Images are sent to the model as pixels. Breezell no longer looks for image.png in the repo.
Composer: Line breaks are preserved. Shift+Enter is hinted in the input.
Plan / Todo: Task status in .plan.md stays in sync as work moves. Resuming an old plan pulls todos with it. update_todo_list no longer re-marks items while a plan is executing.
Codebase Exploring: Fold groups rebuilt.
Edit cards: Completed cards collapse by default — controlled from the box feature toggle.
Citation pills: Styling matches the editor selection helper.
Account pool: Rotation is per request, not pinned to a single conversation.
Built-in browser: Tabs are isolated per workbench window.
Explorer: Project folders named dev or run are no longer hidden.
PDF preview: Scroll, zoom, and drag work.
Video preview: Duplicate selectors removed. Send to chat is available.
Minimize hang: The unresponsive dialog waits 15 seconds before appearing, so fewer false alarms.
Context windows: Qwen 3.8-27B can switch 256K / 1M. Short-context windows and list prices follow the official tables.
Thinking intensity: Official tier names aligned across providers, including DeepSeek.
Compatible slots: Shared config-column width. Capability tags sit under the token line.
Admin: Model-management table header stays pinned. Fetching the model list is fixed. Reference price and long-cache fields line up.
Sidebar: The overflow panel on the top activity bar no longer paints outside the sidebar.
Improvements
Retired GPT-5.3-Codex-Spark removed from the catalog.
Conversation body color matches tool-card filenames.
Electron updated to 43.7.3.😀
https://t.co/TxgOudyrNf
What Zcode did is highly unethical. Although they've issued a response, we have no way of knowing whether their claims are actually true. We simply cannot completely prevent data leaks. As long as your data is routed through the cloud, there is inevitably a risk of leakage. The only alternative is for your enterprise to opt for a comprehensive self-hosted solution, utilizing your own tools and deploying local models, but the cost is simply too high.
I'd use Cursor in a heartbeat, with Claude Code as a close second. Cursor's dedicated API line is blazing fast—unlike Codex, which is a bit underwhelming—and it lets me rapidly review everything within just 24 hours. Meanwhile, Claude Code is incredibly powerful when paired with a Claude subscription.
Breezell 1.3.6 is officially out! We’ve moved the indexing model to run entirely on your local machine, and real-world testing confirms its footprint is incredibly minimal. Looking back at all the versions we’ve shipped so far... could this be the future? We truly believe this is shaping up to be the ultimate code editor!
https://t.co/TxgOudyrNf
https://t.co/zNsScULJwH
#AICoding #DevTools #LocalLLM #CodingAgents #IndieHackers
I feel like AI has actually made things worse. Whether it's open-source or private projects, people seem to constantly look for the easiest way to bypass testing. Or they'll let AI generate a few lines of code and then spam authors with PRs without even verifying the architectural integrity or checking if it fits the existing structure.
@sahildarz I totally get how you feel. We only have two people on the team, and after releasing version 1.3.4 yesterday, opening a new project resulted in a black screen. We had to pull an emergency overtime shift for two hours just to rebuild the package—it was an absolute nightmare!💤