Blog

If you have ever tested that two collections contain "the same stuff, order doesn't matter," you already know the workaround: sort both arrays, compare them element by element, and hope nobody adds a duplicate. It works, but it isn't what you meant to write. You meant "these are the same set."

Reasoning-capable models have been usable in BoxLang AI for a while. params passes straight through to the provider body, so params: { thinking: { type: "enabled", budget_tokens: 10000 } } for Claude, or params: { reasoning_effort: "high" } for OpenAI, already reached the API. What never worked was reading the reasoning back. It was parsed out on arrival and silently dropped. In 3.4.0, that's fixed, and it's fixed the same way for every provider.

TestBox 7.1.0 is here, and it is the largest single expansion of our assertion and expectation library since TestBox was born.

LLM applications face a class of attack traditional input validation was never built for: prompt injection. An attacker embeds instructions in user input, a retrieved document, a web page your tool fetched, or an MCP result, trying to override your system prompt, exfiltrate data, or hijack a tool call. BoxLang AI 3.4.0 ships four layered, configurable defenses against exactly this, plus one more that's on unconditionally.

You call writeDump() on an ORM entity or a very rich class graph . The browser locks up. Thirty seconds later you get a page with forty thousand rows on it, you scroll for a while, you give up, and you go edit your code to dump a sub-key instead. Or worse, you crash the server.

Everyone knows the datasource password should not be sitting in plain text in a config file. Everyone has also, at some point, shipped exactly that, because the alternative was a pile of environment variable plumbing that nobody wanted to build on a deadline. x

Here's a bug that's easy to miss until it bites someone in production: an agent turn asks for two tool calls at once, both need human approval, and only the first one actually suspends. The second one gets silently skipped. Not rejected, not queued, just gone. That's what happened before 3.4.0, and it's fixed now with batched tool-call approvals.

HumanInTheLoopMiddleware used to do everything itself: decide which tool calls needed approval, present the request, and wait for a decision. In 3.4.0, that logic has been pulled apart into a real subsystem, and the piece developers will feel the most is that a human's "always allow this" now actually means always.

Every serious language has a way to ask "is this even parseable?" without running it. bash -n script.sh. node --check file.js. python -m py_compile. As of 1.17.0, BoxLang has one too.

Every AI agent that touches something real eventually needs a human in the loop. Someone has to approve the delete, confirm the wire transfer, or just say "go ahead" before a tool call runs. The question BoxLang AI 3.4.0 answers is: approve it how? A terminal prompt? A webhook? A Slack button?