againrom-pipeline — how AgainROM is built

AgainROM is built by AI agents under one owner's direction. The owner decides what to build and rules on conflicts. Agents research the original games, implement, review and release. Every rule they follow is written in instruction files in the repositories, not kept in a conversation.

Each piece of work runs as a lane: one agent, one branch, one worktree, one brief that names its scope, its proof and what it returns.

research	private. Evidence about the original games: observation of installed files and saves, and static analysis of the executables
knowledge	public snapshot of research: file formats and claims with confidence, executable locations replaced by opaque aliases · againrom-knowledge
engine	public. The game, the starter and the map editor, GPL-3.0 · againrom-engine
seat	private coordination: the queue, lane state, journals and review reports
factory	private. Agents build mods blind, from a release build and the public mod documentation only; every gap they hit becomes a fix or a feature

experiment	one falsifiable question about the original. Publishes claims as High, Medium or Unknown; Unknown is a valid result. One confidence review
story	one result a player can see. A story document, focused tests, one adversarial review by a fresh agent, at most one correction pass
hotfix	a small defect with a local cause: reproduce, test, fix, land
audit	after every fifth story: a read-only search for anything the engine builds in two ways

• Research is the authority on how the originals behave; the owner is the authority on what to build.
• Engine lanes send research questions, never the expected answer.
• Engine code is never evidence of how the original behaves.
• No game file, install byte, owner save or original code enters any repository. A pre-push hook scans every engine push.

Where the engine differs from the original, by the owner's choice or as known debt, the difference is one row in docs/divergences/, with the claim it departs from. A difference that lives only in a comment or a chat does not count.

Anything the engine builds more than once (a screen, a room, a save path, a widget, a loader) has one builder; a new instance is data for it. The first game is a reduced second game: shared screens are built in the second game's shape, with the first game as profile data.

fast gate	gofmt, build, game-asset scan and the fast tests; at most 60 s; before every commit
full chain	every test, with the release tests on the English and the Russian install; at each release
release	after five stories: builds for Windows, macOS (Intel, Apple Silicon) and Linux, with a player changelog

claims	4,500+ in the public knowledge snapshot
stories	700+ story documents in the engine
divergences	1,400+ rows
releases	60+ tags

mapedit(1): agents also build maps through the editor's MCP server.