An open-source project that spawns Claude Code CLI directly and parses the full NDJSON protocol, with session fork/resume/worktree, Split deployment, file explorer, and Git integration. Why interactive mode matters after the June 15 Agent SDK billing change, and how the architecture differs from other approaches.
When you have multiple GitHub accounts or need a specific SSH key for a repo, here are three approaches: GIT_SSH_COMMAND for one-off use, git config core.sshCommand for a single repo, and ~/.ssh/config Host aliases with url.insteadOf for multi-account management.
An open-source proxy that lets you keep using Claude Code CLI, VS Code, and JetBrains interfaces while routing API calls to NVIDIA NIM, OpenRouter, DeepSeek, or local models.
Continuing the cache TTL audit series, this time breaking down by model. The server silently switched main agent models three times (Opus 4.6 → Sonnet 4.6 → Opus 4.7 → Sonnet 4.6), and Claude Code autonomously assigns sub-agent models. Comparing cost per million main output tokens across 7 periods reveals an 11.5x efficiency gap.
Introducing JetBrains Air, an agentic development environment that orchestrates Claude, Codex, Gemini, and Junie to run tasks concurrently across Local, Git Worktree, and Docker execution environments.
Introducing Knip, a dead code detection tool for JavaScript/TypeScript projects. Learn why ESLint and depcheck each have blind spots, and how Knip uses a full module graph to find unused files, exports, and dependencies in one pass.
Introducing vitest-fail-on-console — why console.warn/error appearing silently in tests is a code smell, and how this package forces you to deal with it.
A deep dive into testing a filesystem service using memfs to replace fs and a hand-written Fake to replace chokidar, making filesystem tests fast, deterministic, and security-aware.
After Opus 4.7 auto-upgrade, the old version vanished from /model. Tested availableModels, modelOverrides, and ANTHROPIC_CUSTOM_MODEL_OPTION — with pitfalls and GitHub community discussion.
$127K equivalent cost, 127K turns, four models, three months. After turning myself into a dataset, "long sessions are the culprit" and "too many skills" were debunked by data. Only two levers actually worked.
