How Server-Sent Events Solved the Long-Running AI Job Problem in the Browser

·1 views

A developer rebuilding the spectr-ai contract audit tool found that users were refreshing the page after 90 seconds of seeing a spinner, inadvertently cancelling long-running AI audits. To fix this, the pipeline was rebuilt using Server-Sent Events (SSE), which streams real-time updates from server to client over plain HTTP without requiring WebSockets or bidirectional connections. A fixed vocabulary of four typed events — progress, partial-finding, done, and error — replaced an ad-hoc stream of raw model output, giving the frontend a stable, predictable API contract. Individual findings are now streamed to the browser as soon as they are validated server-side, reducing perceived wait time even when total processing time remains unchanged. The architecture also accounts for network interruptions by assigning monotonic IDs to every event and maintaining a replayable server-side log, allowing EventSource's built-in reconnection to resume mid-audit seamlessly.

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)

Git Worktrees Let You Manage Multiple Branches Without Duplicate Clones

Developers often end up with multiple cloned copies of the same repository for different branches or experiments, leading to duplicated history and inconsistent remote tracking. Git worktrees solve this by allowing multiple checked-out working directories to share a single underlying repository. Each worktree maintains its own branch, staged changes, and working files, while commits, tags, remotes, and fetched objects remain common across all of them. A single fetch from any worktree instantly updates remote references for all others, eliminating the need to sync each clone separately. The approach also reduces disk usage on large repositories by avoiding redundant copies of the full Git object database.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Solo developer launches 27 security tools and AI-powered on-chain escrow arbitration system

A solo developer behind pigfox.com has built 27 tools spanning OSINT, email security, TLS analysis, and document forensics, with 16 available for free. Alongside the toolset, three live smart-contract demos have been deployed on the Base Sepolia testnet, including an AI arbiter that resolves escrow disputes by selecting from a fixed set of contract-defined outcomes. The developer emphasizes a evidence-first approach, ensuring all tools query real infrastructure before any AI model is involved, and only five of the 27 tools invoke a language model at all. On the blockchain side, the AI arbiter's decision space is deliberately constrained by the contract, preventing it from inventing outcomes or making unauthorized transfers. Nine disputes have already been settled end-to-end on-chain, with every transaction publicly inspectable on the testnet block explorer.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Strong Backend Systems Are Built on Foundations, Not Frameworks

A seasoned backend developer argues that great backend systems are defined not by the technologies they use, but by the foundational principles underlying their design. Key among these principles are deeply understanding the problem before writing any code, clearly separating system responsibilities to avoid tangled dependencies, and carefully designing data models since data outlasts most software changes. The author also emphasizes that APIs should be treated as contracts, promising clients consistent, predictable behavior rather than merely serving as routes that return data. These insights stem from years of building REST APIs, authentication systems, and real-world backend services. The core message is that engineering begins with understanding, while programming is merely its implementation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Understanding AI Agent Cycles as State Machines for Better Design

A technical article on DEV Community explains how LLM-based AI agents can be modeled explicitly as Finite State Machines (FSMs), with defined states such as thinking, acting, observing, done, and error. The piece argues that many developers build agents through vibe coding — copying AI-generated code without grasping the underlying concepts — which can lead to unnecessarily complex implementations. By treating the agent loop as a state machine, developers gain clearer control over transitions, tool execution, and error handling. The article further introduces Hierarchical State Machines (HSMs) as a structured solution for when agents grow complex enough that a flat FSM becomes difficult to manage. The core goal is to help developers understand what they are actually building, enabling more deliberate and scalable agent design decisions.

0 comments Read more at DEV Community