Issue → Fix → Review → Deploy

Ghost Foundry

Turns GitHub issues into reviewed pull requests, with a human holding the merge button.

Role: Design, architecture, and implementation
Status: Working · in daily use

Ghost Foundry workspace overview with recent runs and repository readiness

Ghost Foundry is a local, single-user software lab. You describe a change in a GitHub issue; a fixer agent reproduces the problem, implements the change, and runs the checks. A separate reviewer agent reviews the result, and the run collects evidence (logs, diffs, before/after screenshots) before opening a linked PR.

Fixer and reviewer are configured independently. Each can be Codex, Claude Code, OpenRouter, or a local Ollama model, with its own fallback chain, so one provider having a bad day doesn't stall the queue.

Engineering highlights

Agents run in throwaway Docker containers with dropped capabilities, capped CPU and memory, and no access to the host's home directory or Docker socket.
Durable job queue with atomic claims, leases, and crash recovery, so no run is ever silently lost or done twice.
Production and development instances run side by side; development is fixture-only by default so nothing real gets touched by accident.
Read-only upstream repositories are handled through a writable fork.

A Ghost Foundry run report showing each pipeline stage completed

Ghost Foundry approval inbox with a change ready for review

Next project

KLAWT →

Build something like this