workshop
A multi-lane render, judge, and dispatch system
The machinery behind the Design Workshop board — run the board there; this page is the engine note. This system is how one person’s game gets a studio’s worth of review: renders, judgments and dispatches flow through it so the design decisions land as verdicts, not vibes.
An event-driven, multi-lane design-review system: seven lanes (balance, battle feel, flow, mechanics, menus, playthrough, VFX), state kept as plain JSON/JSONL in the repository, served by a local Python service. Writes are locked per file path, not behind one global lock, so one lane’s write can’t stall another’s; a delta-notification command prints only the difference against a stored snapshot, and an agent’s own writes are tagged by source so they never re-trigger the monitor that reads them.
Its flow-mapping tool cross-checks three independently-produced views of the game’s screen-flow graph — a static parse, a hand-authored intended map, and a runtime trace from live play. The runtime trace caught the static parser wrongly classifying reachable game states as dead code; every disagreement gets filed automatically as a de-duplicated work item. A separate autonomous-review mode grades five safety dimensions green/yellow/red from two independently recorded trace trails against ground truth — logged across six dated audit rounds, one of which found two regressions its own previous round had introduced.
Private personal-project repository; the manifests and review data are the legible artifact here, in addition to whatever video capture follows.