My workbench.

Agent-grade document production — an MCP server that turns structured data into typeset PDFs via Typst. 24 templates, offline compiles, visual-regression-tested. Ships with Forge, a Slack agent that drives it.

Test fails, 40 lines changed — these 3 lines are the culprits. Spectrum-based fault localization for vitest: one run, ranked suspect lines, the failing tests as evidence. No config, no LLM.

A big uncommitted diff and a failing test — which hunks broke it? Delta debugging over git change regions: these hunks, isolated, below commit granularity where git bisect cannot reach. One npx run, any test runner.

Where does feature X live? Feature location by coverage diff for vitest and jest — ranked source lines with the tests that prove it. Execution evidence, not a grep result.

An anti-sycophancy skill for coding agents — contract supremacy over prompt pressure. The code must survive the tests, even when the prompt lies.

An agent wants to deploy, but two CIs disagree and the review isn't done. Residual Runtime blocks the action while its signals are unresolved — and says exactly what would unblock it.

Urban walking itineraries from real street networks — TSP-optimized routes over OSRM and PostGIS instead of tourist guesswork.

Browser-native real-time liquid simulation on WebGPU — a hybrid PIC/FLIP→APIC solver with screen-space fluid rendering. TypeScript and WGSL, no framework.

Civilizational simulation at engine scale — deterministic, traceable, deliberately out of reach. That's what keeps the workshop warm.

Embeddable calculators and decision tools — pricing estimators, ROI walkthroughs — built from configurable logic. Shipped, then left to rest.

A persistent workspace for hyperbolic worlds — non-Euclidean geometry you could keep. Untouched long enough to count as finished.

Turned Stripe payout CSVs into QuickBooks, Xero and Exact Online imports, with payout-total checks so the clearing account zeroed out. Did its job; no longer maintained.

A real-time geography game — chain world cities under pressure, with ranked matchmaking and bot backfill underneath. Ran until the domain lapsed.

MCP telemetry analysis that refused conclusions your data hadn't earned. Still on npm, no longer maintained.

Persistent memory for coding agents over MCP. It beat the models on cost — until they closed the gap to a cent and their own context got better.