OpenCrabs v0.5.15 is out. This one is what happens when you actually live in the tool you ship and write down everything that made you swear. The headline fix is one you could feel. Press Ctrl+O to expand a long answer and the viewport threw you to the top of the turn, every single time the header was off screen. The anchor code recorded the screen row of the turn header, and when that header was above the viewport the subtraction saturated to 0, so the renderer believed the header sat at the top of the screen. Any turn whose answer is longer than your display has its header off screen when you are at the bottom, which is the normal state after a real answer. The fix is a checked subtraction that returns "no row" instead of a fake one, and while verifying I found Ctrl+O was also silently disarming auto-scroll, so streaming output stopped following the bottom. That is pinned now too. The second thing is honesty in the interface. The queue row that appears while a turn runs had been restyled to luma 201, brighter than every label on screen, screaming for attention against a thinking line it was supposed to sit under. It is back on a gray hierarchy, all four spans under the thinking line, and a new test reads actual rendered cell colours, so nobody can ship amber in that row again without a red build. Same spirit elsewhere: the input box now says Enter queues while a turn runs, the history marker points at PageUp instead of a key that folds the conversation, and a caught render panic hands the terminal back instead of leaving keystrokes echoing into the status bar. Provider resilience got scoped properly. The built-in free endpoint now gets a 120 second handshake instead of 60, and stops the retry ladder after 2 silent timeouts, because a timeout means the endpoint never answered and five more tries will not change that. Worst case on a dead endpoint dropped from over five minutes to about four, and a slow-but-alive handshake at 70 or 100 seconds now survives instead of dying. While hardening it we found endpoint detection matched a substring, so any URL merely containing the hostname was misclassified, and every special case keyed off that one check. It parses the host now. The rest is connective tissue: the Telegram dispatcher recovers from a failed startup instead of panicking, /evolve retries dropped downloads, and the docs stop claiming 3 retries where the code does 5. There is also a new work-shape guard on the phantom language scan, mutation-tested red twice, so when the Korean, Japanese and German packs land next release, the scan cannot quietly turn input-proportional. Grab it at opencrabs.com/download

My openCode Go subs barely survives 5 days with the paid models enabled in the subs. @opencrabs free provider with DeepSeek v4.1 Flash running 24/7 all my flows for 13 days non stop. Life is good sirs!

Bros 🦀

Bro is on a mission here

Free money, literally.

We have got a bot attack a few days ago. Sadly. We did not have 3,8k users at ATH, most was bots. We added a few safe guards plus deduplication for a cleaner tracking. Regarding the FREE AI. We will be able to support the free AI for quite sometime still. No breaks. So yes go and break something! 🦀 Sidenote: This meme was fully built by @opencrabs btw

OpenCrabs patch release v0.5.14 is out. This one is about reaching more people and breaking fewer files. Two new languages joined the lie-catching packs, and the edit tools got a fix for a bug class that has been quietly corrupting files. The headline is the language coverage. OpenCrabs now speaks Arabic and Chinese natively: 732 intent phrases for Arabic built from verb families, and 1,222 for Chinese where every phrase exists in both Simplified and Traditional, so 我会检查 and 我會檢查 both land. The detection is script-aware, which is the subtle part: Japanese and Korean also write with Han characters, so the Chinese branch only fires when there is no kana or hangul in the text. At the same time the existing Spanish, French, Indonesian, Russian and Portuguese packs jumped from around a hundred intent phrases each to 436-545, the same order of magnitude as English. A pack with a third of the phrases catches a third of the lies, and that gap is closed. The edit bug came from a user report and it was a good one. Ask edit_file to delete a line and it rebuilt the whole file from a line split: your final newline vanished, and every CRLF ending in the file turned into LF. The diff view even hid it, because it diffs the same rebuilt strings. The fix splices bytes instead of rebuilding, and the same CRLF defect was found and fixed in hashline_edit while verifying. Nineteen new tests now pin the byte-level behavior, where before this there were zero. Infrastructure got stiffer too. A new CI lint fails the build the moment anyone introduces a byte slice that can panic on a multi-byte character, with 910 existing sites baselined. Every workflow action is pinned to an immutable commit, the Rust toolchain is pinned to 1.98.1, and the test suite now runs on Linux as well as macOS. 2 features. 10 fixes. 11,407 tests. Grab it at opencrabs.com/download Thanks to @syafriadit1 @iamloonix and the 3 new contributors @echoxiangchen, @yimengzhiyan and @VincentFeliciano.