JosiahAllard/Agentic-Software-Factory

Run a whole team of Claude Code agents on one Windows machine, across every repository you own,
without them trampling each other's work or the machine itself.
## What works today
Live since 2026-10-15, with landings in three repositories, and covered by tests:
A managed engineering organisation. One PM agent across repositories, one coordinator per
repository, worker agents in parallel, each in its own worktree. Roles are written once and
rendered for Claude Code and Codex (content/roles/, tools/rendercontent.py). Task list:
the backlog store (tools/backlog.py); questions and to-dos for the owner: the Action Hub
(tools/hub.py); sessions: the launcher (tools/launch.py).
Quality lives in the pipeline. A branch needs a passing gate record and an independent
reviewer's verdict. The landing then merges it onto current main in a separate tree and gates
the merged result; only a green merge is pushed. On red, the merge is dropped and main is
untouched. A daily full run covers the rest (tools/land.py, tools/gate.py, docs/gate.md
"Landing", tools/dailyfull.py).
Guardrails stop risky actions and say what to do instead. The hook guard refuses a
destructive git command in a main checkout, a search from a drive root or a heavy run outside
the lock, and each refusal names the right form on an Instead: line
(kitlib/guard/rules.py, docs/guard.md). A pre-push hook refuses a non-fast-forward push to
main (githooks/).
Shared machine resources are scheduled. Heavy runs queue for CPU and GPU slots in the
machine lock, and landings queue behind each other in it (tools/machinelock.py,
docs/lock-api.md, docs/gate.md "The landing queue").
Also in use: worktree cleanup (tools/worktreeremove.py).
Not done: Linux support is in progress (the kit still refuses Linux; docs/linux-support-plan.md),
and the Codex bridge is not live yet. Codex is a supported host for the roles, hooks and commit
tools; the bridge between agents is the part that is not.
## The problem
One Claude Code session is easy to supervise. Ten are not. Two agents edit the same checkout,
three full test suites start at once and starve the machine, a destructive git command fires in
the wrong folder, a red branch slips onto main, and the questions only you can answer scatter
across a dozen terminal windows.
## How the kit solves it
One project manager (PM) Claude session runs the machine. One coordinator session runs
each repository: it splits the work into tasks and hands each to an agent, so many agents build,
test and land code at the same time. The kit lays the rails they run on:
Machine lock. Heavy runs (full test suites, builds, renders, training) queue for CPU and
GPU slots instead of fighting over them, so the machine never overloads. You see the queue at
a glance and can hold the machine for yourself.
Guard. A hook in each host stops risky commands before they run: a destructive git command
in a main checkout, a search from a drive root, a heavy program run outside the lock. Every
refusal names the command to use instead.
Safe landing. One tool carries a branch to main. It merges the branch, tests the merged
result, pushes, and checks the push arrived. A failed test undoes the merge. Behind it, a git