Developer tooling, plus the memory and MCP plumbing underneath AI agents. Mostly TypeScript, some Python.

claude-glass didn't work on Windows at all. Backslash path separators left every internal link unrewritten, and serve returned 403 for every file, because a containment check compared a backslash path against a forward-slash prefix. I reported it as #28 and fixed it in #31. That PR also carries two dependency XSS advisories, because the project's pre-push tests can't go green on Windows while those path bugs are there, so they had to land together. It shipped in v0.7.8

Backslash path separators no longer break Windows builds (community contribution by @voyagi, #31, fixes #28)

The maintainer added CI in the same release, running the tests on Ubuntu and Windows. The Windows job failed on pre-fix main and passed on the branch carrying the fix. Both runs are red overall on a separate dependency-audit job.

Claude Code is closed source, so what I send Anthropic is bug reports. Two became shipped fixes: deferred tools not reaching skills that run in a forked context (#46654, fixed in 2.1.126), and the same CLAUDE.md loading twice on Windows when drive letter casing differs (#19516, fixed in 2.1.47, where I brought the cost measurements and someone else brought the diagnosis).

• claude-usage turns Claude Code's local session logs into a VS Code dashboard: tokens, cost, and how close you are to the rate limit, with no second API key to manage.
• throughline is an auditable memory layer for incident response, where every recall comes back with a receipt, so "no prior incidents" can't quietly mean "the search broke".
• datahub-blast-radar scores what a data change is about to break, by walking lineage through DataHub's MCP server and writing the verdict back into the graph.

The rest is side projects, templates I reuse, and forks I keep around. taranity.com is the studio I run.