Knowledge Share
Technical articles, tutorials, and insights
LLM-Driven Automated Testing Series (14): Production Engineering — Why Testing Turns Out to Be the Cheapest Line Item in This Whole Series
A code-generation workflow I documented earlier burned thousands of dollars in tokens during debugging. The testing scenarios verified across this series — AutoRestTest testing a 15-operation API for about a dime, pytest-triage capping itself at 10 model calls per run by default — are off by orders of magnitude. This article puts the evidence from 13 prior articles side by side to answer the three questions the planning doc left for last: why testing-scenario cost is what it is, where the human-in-the-loop gate actually belongs, and which scenarios should explicitly avoid LLMs altogether.
LLM-Driven Automated Testing Series (13): Flaky Tests Aren't a Rerun-Count Problem, They're a Classification Problem
pytest-rerunfailures (477 stars) and flaky (397 stars) handle test instability with the same one-liner: if it fails, rerun it a few times. That doesn't answer what the planning doc actually asks — was this failure a real bug, environment jitter, or a badly designed test? The only project actually attempting that root-cause classification with an LLM is a 2-star, less-than-two-months-old repo called pytest-triage — and its most instructive feature turns out to be the four safety invariants it wraps around AI judgment, guaranteeing AI is never allowed to affect whether a test actually passes or fails.
LLM-Driven Automated Testing Series (12): Beyond Pixel Diffs — Who's Actually Making LLMs 'Understand' UI Changes
BackstopJS, at 7.2k stars, handles rendering noise with exactly four numeric config knobs — misMatchThreshold, requireSameDimensions, ignoreAntialiasing, usePreciseMatching — which is still just fighting font-rendering jitter with pixel-level tolerance. The project that actually wires a VLM/LLM into the visual-regression judgment loop is a 23-star repo called vlmkit, whose own internal benchmark notes state outright that one model hallucinates red into red. Putting these two side by side answers, concretely, what semantic visual regression actually solves — and what new problems it introduces.
LLM-Driven Automated Testing Series (11): Why API Testing Is Actually the Most Conservative Battleground for LLMs
Keploy has 18.5k stars and its README says 'Expand API Coverage using AI' — but grep the entire open-source Go codebase for LLM/AI keywords and you'll find zero implementation code. Schemathesis, at 3.6k stars, uses no LLM at all — it finds real server 500 errors purely through Hypothesis-based property-based fuzzing. The projects that actually call an LLM API in their core logic turn out to be two sub-100-star projects: AutoRestTest and api-automation-agent. This article pins down why: structured API input/output is naturally suited to traditional fuzzing, so what can LLMs actually do here that traditional tools can't — and why does this domain stay cautious instead of going all-in the way UI automation has?
Open Source Project #249: Hallmark — A Design Skill That Refuses to Look AI-Generated, Built on 48 Material Themes and a Browser Linter
Hallmark is a design skill from Together AI, open-sourced by Nutlope, for Claude Code, Codex, Cursor and OpenCode. It holds a hard floor (contrast, focus, reduced motion, no invented facts, no eyebrows), deals three themes from a catalog of 48 materials, builds the page around what the visitor came to do, lints the result in a real browser with sloplint, and hands the last word to a fresh-context reviewer. MIT licensed, nearly 38k stars.
Open Source Project #250: Cindy — An Out-of-the-Box Open Source AI Agent That Links Your Computers and Phone, and Brings Claude Code, Codex and Pi Together
Cindy is an open-source desktop and mobile AI agent client (Electron and Expo). It doesn't build its own engine; it puts Claude Code, Codex and Pi in one interface, then adds multi-device handoff, cross-computer model sharing, phone remote desktop, companions and group chat, Orca multi-agent collaboration, and a plugin and Skill ecosystem. Your computer is your server: no rental or self-hosting. Apache-2.0, 7k+ stars; the client is free but model usage is still billed by providers.
Open Source Project #248: CleanYourTV — Let an AI Coding Assistant Debloat Your Android TV Over Wi-Fi and Give It an Apple TV-Style Home
CleanYourTV is an open-source toolkit for debloating Android TV and Google TV. After you enable USB debugging, a coding assistant such as Claude Code, Codex or Gemini CLI backs up the TV over Wi-Fi, disables vendor ads and telemetry, and installs a native launcher of just 523 KB. No root, nothing uninstalled, and a one-line restore. Nine days after launch it has nearly 38k GitHub stars and community-tested lists for 300+ TV models.
LLM-Driven Automated Testing Series (10): Self-Healing Locators — Where Heuristic Scoring Hits the Ceiling of Semantic Understanding
Healenium is the most established open-source project in self-healing locators, and its README claims it 'leverages machine learning' — but reading the code directly reveals the core mechanism is a weighted scoring formula over DOM tree similarity, with hardcoded constants for weights and thresholds, nothing to do with machine learning. More notably, its 2026 addition of an AI-based XPath generation endpoint is gated behind a hard-coded exception: 'you must have a paid hlm-ai service.' This article first cracks open exactly how Healenium's scoring formula works, then examines how MarketSquare/robotframework-selfhealing-agents — a genuinely LLM-based self-healing project — designs its multi-agent architecture, and finally answers the question the planning doc raised: where exactly does the tradeoff between healing success rate and false-positive rate actually bite.
Teardown of Anthropic's Official Evaluation Methodology: From Defining Success Criteria to claude plugin eval
Anthropic published two evaluation docs: 'Define success criteria and build evaluations' on the Claude developer platform (methodology) and 'Test plugins with evals' in Claude Code (engineering). This article reads them together: SMAR success criteria, how to choose among six grading methods, the case/grader/baseline structure of claude plugin eval, the with-vs-without Δ design, sandbox isolation and CI exit-code conventions, and what they mean for anyone building their own evaluation platform.
A Skill Evaluation Post-Mortem: I Overturned an Existing Model Selection, Then Built My Own Eval Platform
A teammate's skill-model adaptation report looked polished, but on closer inspection it had real problems: one case per scenario, and AI judging its own correctness. I redesigned the evaluation for the Jira Skill's most frequent task — fetching a bug and downloading attachments — with 100 cases × 3 samples, scored by scripts that diff against ground truth. The result was counter-intuitive: the top-scoring sonnet-5-5 was not the best choice; ds-v4.1-flash won on success rate, cost and speed combined. The post ends with the Skill evaluation platform behind the re-test: eval plans, cases, runs, metrics, AI-written reports, plus version-over-version and model-vs-model comparison.
Open Source Project #245: Raven — The 'Harness of Harnesses' That Orchestrates Claude Code, Codex and More, and Improves Itself Recursively
Raven, from EverMind-AI, is an open-source Host Agent: four built-in agents (Research / Code / Design / Oncall), 13 preset third-party agents (Claude Code, Codex, Hermes Agent, etc.) connected over ACP, CLI or OpenAI-compatible APIs, an Evolver that improves the harness itself through propose → evaluate → adopt, and cross-session memory from EverOS. 5.7k Stars, Apache-2.0.
Open Source Project #246: rec — Recursive's Purpose-Built Local Inference Engine for Running Flint on a Mac, Claiming 1.4–2.6x Over llama.cpp
rec is a local inference engine from recursive-oss built to run Recursive's Flint language model on Apple Silicon Macs, aimed at single-session coding-agent use. One command installs it, rec launch pi / opencode starts it, and it exposes local OpenAI- and Anthropic-compatible APIs. The project claims 1.4–2.6x the throughput of llama.cpp on an M4 Pro. Apache-2.0, 151 Stars.