Two-Phase C++ Executor Guards Against Unsafe AI-Proposed Build Commands

·2 views

A developer built a maintenance agent for a C++ repository that accepts natural-language requests and uses a free model endpoint to map them to one of four predefined tool calls. The agent's first prototype validated only the JSON shape of the model's response, exposing a critical flaw when the model proposed deleting the build directory while another job was still writing to it. To address this, the developer introduced a two-phase C++ executor that parses model proposals into strongly typed structs and checks them against a hardcoded whitelist before any action is taken. Each destructive tool is subjected to a non-mutating dry run that tests the same target without altering state, and only after all invariant checks pass does the executor commit to a side effect. The article notes it was prepared as part of promotional outreach for MonkeyCode, whose free model access and hosting were used during development.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

High Google Rankings No Longer Guarantee Visibility in AI-Powered Search Results

A large-scale study by Fractl analyzing 8,090 keywords across 25 verticals and 22,410 domains finds that strong Google search rankings do not reliably predict whether a brand will appear in AI-generated responses. Unlike traditional SEO, which targets search engine results pages, generative AI systems synthesize answers based on entity clarity, source authority, and structured data signals. Supporting research from the GEO 2026 barometer by Reworld MediaConnect and ELMARQ similarly highlights reputation and authoritative sourcing as key factors in LLM visibility. Experts stress that technical SEO foundations such as crawlability and content quality remain important but are insufficient on their own for AI search presence. The findings suggest brands need to treat generative engine optimization as a distinct discipline separate from conventional search strategy.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Cuts LLM Memory Recall Testing from 30 Minutes to 3 Seconds Using pytest and FAISS

A developer built an automated testing suite after repeatedly debugging AI memory recall failures that only surfaced through user complaints in the middle of the night. The core problem was that stored memories, such as dietary restrictions, existed in the vector store but were silently dropped during top-k retrieval due to issues like unnormalized vectors or embedding model drift. Using pytest and an in-memory FAISS index with deterministic hash-based embeddings, the developer created a reproducible, dependency-free test pipeline that runs in roughly three seconds. The setup uses pytest fixtures for index isolation per test and parametrized queries to assert that specific memories consistently appear in recall results. The approach deliberately avoids external vector databases and higher-level frameworks like LangChain to enable precise, fast regression testing directly at the vector layer.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Evidence Gating Outperforms Embeddings for Reliable Docs Chatbot Answers

A two-stage architecture for document-based chatbots first retrieves candidate policy passages, then applies a separate evidence-gating step to decide whether the system should generate an answer at all. When evidence quality or scope is insufficient, the system abstains and routes the query to human review with a machine-readable reason rather than producing an unsupported response. The approach distinguishes four distinct failure modes — retrieval, scope, evidence, and generation — arguing that treating all errors as generic hallucination causes teams to fix the wrong component. Every generated answer must cite versioned policy text that directly supports the classification, and conflicting passages must never be silently merged into a confident label. Abstention is treated as a successful outcome, prioritising decision quality over lower latency on straightforward queries.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Meta patents Ray-Ban smart glasses with facial recognition, raising privacy alarms

Meta filed patent US 2024/0189456 in 2024 for smart eyewear that uses hidden micro-cameras, a dedicated neural processor, and cloud connectivity to identify nearby individuals and overlay augmented reality content in real time. The glasses can locally cache up to 500 facial profiles and sync with remote databases via 5G. The patent disclosure triggered a surge in public interest, with Google searches in Brazil rising 215% and the hashtag #SmartGlasses generating over 12,000 tweets. Privacy experts warn the technology risks violating Brazil's LGPD and Europe's GDPR, which require clear user consent before biometric data is collected. Companies intending to commercialize such devices in these jurisdictions would need to complete a Data Protection Impact Assessment prior to launch.

0 comments Read more at DEV Community