glados

GLaDOS

Generative Logic and Documentation Operating System — yes, that GLaDOS. Unlike her namesake, this one writes everything down and lets you leave the testing chamber.

GLaDOS is a library of engineering workflows for AI coding agents — Claude Code, Gemini CLI (Google's open-source terminal agent), Google Antigravity (Google's agentic IDE; CLI agy; antigravity.google), and others. You install it into a project: a small compiler reads one configuration file in your repo (glados.yaml) and generates a set of slash commands — /glados:build-feature, /glados:fix-bug, /glados:review-mr — tailored to that project. When your agent runs one, it follows a disciplined, team-visible process instead of improvising: reviews get posted where the team can see them, decisions get written down with their reasons, and every run leaves a committed record in your repo.

What you get:

• Ready-made workflows for the whole development loop — plan, spec, implement, verify, review, bug-fix, epic runs, weekly cleanup — installed as slash commands your agent already knows how to invoke.
• Nothing important lives only in a chat transcript. Review verdicts land as merge-request comments (a merge request, or MR, is GitLab's name for a pull request), problems become issues, and every run writes a record file into .glados/runs/ that is committed alongside the work.
• One config file controls behavior. glados.yaml declares who may merge, which decisions the agent may make on its own, and where each kind of result gets posted. The installer refuses any configuration under which a result would silently disappear.

60-second quickstart

Coming from GLaDOS v1? Run python bin/glados.py migrate --target /path/to/your/project instead — MIGRATION.md is the guided path. Otherwise, from a clone of this repository:

cp glados.yaml.example /path/to/your/project/glados.yaml

Open the copy and set two keys. platform: is gitlab or github. phase: is a one-word honest answer to "who gets hurt if the agent gets something wrong?" — nascent (nobody; no users yet), evolving (early adopters), production (real users), or sunset (winding down) — describe reality, not ambition, because the word tunes how much caution gets compiled in.

Then install for your agent's runtime (here: Claude Code):

python bin/glados.py install --mode claude --target /path/to/your/project
# modes: claude | claude-plugin | direct | gemini | antigravity | aistudio

The installer prints an assembly report — every configuration value it resolved and where it came from — and writes the commands into your project. Now open your agent inside the project and run your first command:

/glados:adopt-codebase

The agent studies your codebase, writes what it learns (structure, conventions, candidate coding standards) into product-knowledge/, and leaves a record of the run in .glados/runs/ — ordinary files you review and commit like any other change.

What a run looks like

Say a user reports a crash and you run /glados:fix-bug. The agent does not jump straight to editing code: it first reproduces the bug and writes down how; it fixes the root cause rather than the symptom; then a second, fresh agent session — one that never saw the fix being written — independently verifies it. What you end up with is a merge request with the review verdict posted as a comment on it, plus a committed record in .glados/runs/ saying what was decided and why. Six months later, git log still knows the whole story.

fix-bug is one of fifteen commands. The everyday ones are build-feature (one feature from plan to a verified merge request in one sitting), review-mr and address-review (the review loop), and run-epic (a whole ticket queue). The full list, with what each does, is in docs/workflows.md.

Where things land in your repo

Everything is installed into your project workspace and committed to git: