Developer Uses AI Agent to Raise Test Coverage from 38% to 71% in 30 Days

·2 views

A software developer ran a 30-day experiment using an open-source AI tool called 'the-agent' to automate test writing for a TypeScript project, replacing manual test creation entirely. The tool generates, runs, and maintains tests from natural-language descriptions, integrating with CI pipelines to target only changed files per pull request. Over the trial period, code coverage rose from 38% to 71%, new-feature test time dropped from half a day to around 40 minutes, and three potential regressions were caught before shipping. However, the developer noted key pitfalls including incorrect tests from vague descriptions, legacy compatibility failures, and occasional async timing gaps. The conclusion was that the tool delivers real gains but requires active tuning and human review of business logic — it is not a set-and-forget solution.

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)

dotdotgod Query Uses Local AI Embeddings to Match Questions to Relevant Docs

dotdotgod's query tool lets software agents find relevant documentation passages by semantic meaning rather than keyword or filename matching, bridging gaps when a user's language differs from document terminology. The tool searches Markdown files under a docs/ directory, applying exclusion policies that omit active plans and archived records from its default retrieval scope. Documents are split along their heading hierarchy into fragments of up to 1,600 characters, each tagged with its file path and heading context to aid interpretation. Retrieval runs entirely on-device using the Xenova/multilingual-e5-small model via Hugging Face Transformers, meaning document content is never sent to an external API. Similarity between a question and all document passages is measured using cosine distance on 384-dimensional vectors, with results ranked and capped by a configurable file-level limit.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Autocorrelation Outperforms FFT for Browser-Based Erhu Tuner Accuracy

A developer building a browser-based tuner for the Erhu, a traditional two-stringed Chinese bowed instrument, found that standard FFT-based pitch detection failed due to the complex noise produced by bowing. Bow friction generates high-frequency scratchiness, transients, and strong harmonics that cause FFT algorithms to misread the fundamental pitch. To solve this, the developer switched to a time-domain autocorrelation algorithm, which identifies pitch by measuring how a waveform repeats over time rather than analyzing frequency peaks. Because bow noise is random and non-periodic, autocorrelation naturally filters it out and locks onto the true fundamental frequency with sub-1Hz precision. This approach proved essential for fretless instruments like the Erhu, where even a fraction of a Hertz in tuning error can undermine a player's intonation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Observer Pattern Explained: How YouTube's Bell Icon Models a Core Design Concept

The Observer design pattern allows one object, called the Subject, to automatically notify multiple dependent objects, called Observers, whenever a change or event occurs. A YouTube channel notifying subscribers on a new upload is a real-world analogy for this pattern. In software, a practical example is an order placement system where actions like sending emails, updating inventory, and alerting a warehouse are decoupled from the core order class and handled by independent observer components. Adding new reactions, such as SMS alerts, requires no changes to the original subject class — only a new observer needs to be written and subscribed. In C#, this pattern is natively supported through events and delegates, and underpins common constructs like UI click handlers, MVVM data binding via INotifyPropertyChanged, and real-time systems like SignalR.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds governed AI agent platform with 27 skills and full audit trails

A developer has built a production-grade AI agent platform featuring 27 skills across seven functional groups, including publishing, testing, communication, and knowledge management. Every tool call is routed through a central dispatch layer that enforces a consistent audit trail and trace ID, eliminating untracked or ad-hoc script execution. The platform integrates with real services such as GitHub, Jira, Gmail, Slack, LinkedIn, and Google Drive, with all integrations managed as versioned, governed skills scoped to a workspace. A built-in Test Engineering Architecture suite ships six workflows covering risk-matrix test design, baseline diffing, and requirement-to-coverage tracing, with all outputs stored as JSON or markdown. The developer notes that four of the 27 skills remain blocked due to missing services, and highlights that risky commands are pattern-blocked, underscoring a pragmatic approach to safe, verifiable agent deployment.

0 comments Read more at DEV Community