Back

Foundry

Dev

The forge I build everything else with.

A Claude Code and Codex plugin that steers a project from idea to merge — decisions kept in memory, precise tickets, and CI proven green before every merge.

Claude CodePythonGitHub Actions

About

From idea to merge, dropping nothing on the way

Foundry is the forge I build everything else with. Orfeo, Luchamon, Souffle: every ticket, every decision and every merge on those projects goes through it.

It is a plugin for Claude Code and Codex. It does not code for me: it holds the course. It knows what has been decided, what is left to do, what is blocked, and it lets nothing ship until it is clean. And it is used the same way whatever the model and the tracking tool.

Since I started working with it, what I ship is clearly better. That is the only measurement I care about.

A day with Foundry

/foundry:frame        a vague idea becomes decisions and precise tickets
/foundry:roadmap      the smallest slice that proves it is worth it
/foundry:next-issue   the most useful ticket right now, with the reason
/foundry:start-issue  the branch is created, the ticket moves to in progress
/foundry:open-pr      the PR goes out, linked to the ticket
/foundry:merge-pr     review, CI, merge

Sixteen commands in all, the same on Claude Code and on Codex.

It thinks where it should

Picking the next ticket is not sorting a list. Foundry gives the model the project's whole graph — dependencies, acceptance criteria, progress — and asks for a reasoned recommendation. "Take this one: it unblocks three others" beats "it is first by priority".

It does not give way where it should not

The rules that matter are written in Python, not in instructions. The merge requires at least one check that really succeeded, nothing pending, nothing red, and it is pinned to the exact commit that was verified. A hand-typed gh pr merge is refused. So is a direct push to the main branch.

No session in a hurry can merge red. Neither can I.

It remembers

Every architecture decision is recorded. Before reopening a topic, Foundry rereads the ones that concern it and takes them as settled. An idea that was already ruled out is refused, citing the decision. No more fresh session cheerfully suggesting what I dropped a month ago.

A team of four

  • Lupin explores the repository. He gets in everywhere and touches nothing.
  • Eiffel builds, one ticket at a time.
  • Maigret reviews the diff before the merge, knowing nothing of the conversation that produced it.
  • Vauban is called in for architecture decisions.

Each works on the model tier it needs. If tests or review block, the role moves up a tier on its own.

My tracking tool, not one imposed on me

YouTrack, Linear or GitHub Projects: each repository picks its own, and switching leaves the rest of the forge untouched. I wore out four tracking tools in four months before getting there. Now I have the choice, and that is exactly what I wanted.

In numbers

  • 16 skills, 4 agents
  • 3 trackers, 2 environments
  • 3,416 tests
  • 8 projects steered with it, Foundry included

Try it

/plugin marketplace add patobiskoto/patolabs-plugins
/plugin install foundry@patolabs

The code is open, under the Apache 2.0 license. In the same repository, its little brother ship-ios takes over after the merge and carries an app to the App Store — it is the one that shipped Souffle 1.1.

Roadmap

  1. 1.1 — Models and cache

    Shipped2026-10

    Updated the model mapping for each tier, and took the first prompt-cache measurements.

  2. 1.0 — Three trackers to choose from

    Shipped2026-10

    YouTrack, Linear or GitHub Projects, chosen repository by repository, on Claude Code and on Codex.

  3. Public repository

    Shipped2026-09

    The code is published under the Apache 2.0 license, from an audited copy of the private repository.

  4. Agents and model-tier routing

    Shipped2026-08

    Four roles — Lupin, Eiffel, Maigret, Vauban — each on the model tier it needs, with deterministic escalation when tests or review block.

  5. Claude Code and Codex

    Shipped2026-07

    One implementation, two environments. The skills have the same names on both sides.

  6. The pipeline, from idea to merge

    Shipped2026-07

    Framing, roadmap, picking the next ticket, PR, blank-context review, and a merge that refuses to go through until CI is proven green.

More from patolabs