Tools for working with AI agents
Roomi's Grillhouse
A ready-to-use project for working with Claude Code: a grill that settles the architecture before the code, then tickets carried out by development agents and reviewed before they close.
active version 0.24.13, 28 October 2026 licence MIT
Roomi’s Grillhouse is the framework I built for BPscript, FaustScript and Kanopi. These projects are too complex for vibe coding focused on the solution: without structure, coding agents work against the architecture, and a complex project ends up as a plate of spaghetti. Grillhouse holds that the architecture decides and the code follows: every decision is written into the architecture before it is coded. BPscript is the pilot project: 99 of the 163 changelog entries, from 0.16.0 to 0.24.13, cite an incident observed on it, with its ticket. Its main limitation: it assumes Linux, Claude Code, Node.js 22 or later and a TypeScript or JavaScript project, and its skills speak French.
No playable demo: the framework runs in a terminal, on the project’s machine.
The grill first, the code second
Before any code, a session questions the project owner, turn by turn: goals, mature reference products, consumers, axes of change, components and interfaces. It writes the project’s charter and reference documents. The grill stays open: any structural question raised along the way goes back to it. What a rule decides, the pipeline decides; the owner only receives the real decisions, with a recommendation.
A CI pipeline whose stages are agents
The README itself compares it to continuous integration: each stage applies its rules without asking, and calls on a person only for what no rule settles. A ticket moves from stage to stage: 1 Plan, 2 Tests, 3 Code, 4 Final pass and review, 5 Integration (packages/grillhouse/scripts/etapes.mjs). Each stage is held by a fresh agent, a Claude Code session launched for that role alone.
The roles are named like a kitchen brigade. The supervisor (pitmaster) keeps the ticket queue. The explorer writes the plan, and the questioner challenges it once. The testwriter writes the tests before the code, the developer makes them pass without being able to touch them, and the reviewer reviews without writing anything. Each session returns a structured sheet, validated against a schema; the pipeline decides on these sheets and on the test reports, never on an agent’s free text.
Getting started
From the README (I did not rerun these commands for this page):
npx @roomi-fields/grillhouse init ~/dev/my-project [ticket-prefix]
cd ~/dev/my-project && claude
init creates the project from the templates/projet/ template, installs Beads (the ticket store, kept in git) and CodeGraph, indexes the code, sets up the git hooks, installs the declared plugins and ends with a commit. It refuses a directory that is not empty. An existing project adopts the framework with npm install -D @roomi-fields/grillhouse, then npx grillhouse maj.
What it does
- The initialisation grill. At every session start, a hook lists what is still empty: arbitration criteria, charter fields, architecture, frame, interfaces, lexicon. The session offers to grill them, round by round: goals, mature reference products, owner, consumers, data representations, axes of change, components, interfaces. Each question comes with its recommendation and its evidence. The grill then writes
CLAUDE.md,docs/ARCHITECTURE.md,docs/CADRE.md,docs/INTERFACE.mdandCONTEXT.md(mods/grillhouse/skills/grill/SKILL.md). - The structure guard. On every message, a hook sends any question about packages, modules, interfaces or boundaries back to the grill.
- The locks. Each agent runs inside a bubblewrap envelope: it reads the whole copy and writes only the packages its plan names. Permission hooks forbid the developer from changing the tests.
- Supervision by tickets. A stop in the pipeline ends on a question to the supervisor, with its fact and its evidence (file and line, or command and output). The supervisor resumes, extends the plan or stops, through a pipeline command.
- The board. The
/grillhousepane, ornpm run tableau, shows where each ticket stands, and each agent’s time and tokens. - The night.
grillhouse nuit, run from the crontab, replays the full pass; each red task adds to its open ticket or opens one.
A real example, from version 0.24.12: the pipeline filed each discovery made by an agent as a closed ticket. BPScript had 161 of them since 12 October, none read one by one; on a single ticket, three out of nine said the same thing. Since then, a stage 5.3 sorts each discovery with a verdict (handled, duplicate, no action, to carry forward, unconfirmed) and its evidence, replayed on the main branch.
Common issues
- Outside Linux: the envelope relies on bubblewrap, which is Linux only. Without
bwrap, launching an agent stops on « ENVELOPPE IMPOSSIBLE ». The README lists it among the requirements (sudo apt install bubblewrap);initdoes not install it, since the framework never runs sudo for you, and the session start reports it while it is missing. - The night no longer runs: a crontab line older than 0.24.0 points at
nuit.sh, which no longer exists. The session start reports it and proposes the replacement line. - The pane does not scroll: it docks beside the conversation in fullscreen mode (
/tui fullscreen, at least 110 columns), and the mouse wheel needsCLAUDE_CODE_DISABLE_MOUSEto stay unset.
Where it stands
Version 0.24.13 of 28 October 2026; 509 commits since 7 October 2026, 44 releases published on GitHub since 0.16.0 of 18 October. The npm package is @roomi-fields/grillhouse. 1 star, no forks as of 28 October 2026.
The vision: one framework, identical in every project that installs it, which each project improves by proposing its deviations. The aim is for each project to build a mature product, on what exists and on what the references of its field do.
The board computes the time and cost of each ticket; I do not publish any figure of gain until a comparative measurement exists. What is new goes in the repository’s changelog.
Repository · npm package · rtfm · bpscript
Grillhouse wires together existing tools: Beads, CodeGraph, rtfm, the mattpocock-skills skills, Turborepo, Vitest. The reviewer’s passes (edge cases, verification gaps) are taken from BMAD-METHOD, under the MIT licence (CREDITS.md).
How it was made. Claude Code (Claude Opus 5.5), from the repository, its README and its release notes. Method