Factwall

Bulkheads for multi-agent AI systems.

Like bulkheads on a ship: a breach in one compartment does not sink the whole vessel.

• Agent = separate identity, separate workspace, separate context
• Domains exchange facts only - never raw data, both ways through the hub, with a human consent valve where you want one
• Walls enforced by boring servers (git + Postgres), not by prompts
• Audit trail for free: every crossing is an append-only row, every workspace and brief a git history

The full pattern - problem, five principles, architecture, limitations: PATTERN.md

Reference implementation running daily in my homelab (personal pilot).

Try it - one command, no AI key

cd demo && ./run.sh

Spins up a disposable Forgejo + Postgres in Docker, builds the walls, runs two fake agents + a hub, shows the human consent valve, prints a morning brief - and ends with the servers refusing two forbidden crossings. Under a minute. Details in demo/README.md.

Needs git 2.31 or newer: the engine passes bot credentials to git through GIT_CONFIG_COUNT, which older versions ignore. Building it needs Go 1.25+ (the database driver asks for it).

Layout

PATTERN.md      the pattern: problem, principles, architecture
cmd/factwall/   the engine: one Go binary, wall tests included
internal/core/  what it is made of - walls, bridges, workspaces
demo/           self-contained demo on generated data

Commands

provision                lay the foundations: org, brief, hub, schema
join <domain>            add one domain: workspace repo + bridge role
sync / save              an agent's workspace repo, as its own bot
facts / send / inbox / collect
                         the two-way fact contract
pending / approve        the human consent valve
route add|rm|list        relay edges: whose facts reach whose inbox
relay                    carry every declared edge, one cycle at a time
credential               git asks it for the bot token

🚧 Status: early preview. The personal pilot is running; the pattern and the demo are stable, the engine is still moving. Questions and feedback: open an issue.