Your AI agent is editing production code.

Git shows the final diff.

an agent, no box your real $HOME

What prompt produced this diff?

Did the agent run unsafe tools?

credentials leaked

important files deleted

h5i gives every agent a sealed, auditable workspace.

Sealed worktrees. Recorded evidence. Native Git.

🛡️ env/worker-env — sealed

$WORK — private worktree

.git/.h5i/env/you/worker-env/work

POLICY MANIFEST

fs.write$WORK only

net.egressapi.anthropic.com

✋ fs

✋ net

BLOCKED BY POLICY

A safe workspace for agents.

Dangerous operations are blocked.

terminal — youhost

Turn off agent prompts. Keep the guardrails. The box is the permission.

Every diff gets a receipt.

you — a week later, any clonemain

Prompt Quality

Commands

Test output

Denied actions

Final diff

One week later, you can still answer:
why did this diff happen? · was it built safely?

claude · env/worker-demo-claude sealed

codex · env/worker-demo-codex sealed

Works for one agent. Scales to many.

claude → codex · h5i msg review

jitter=0 shrinks to a flake — added a regression seed, please check

codex → claude · h5i msg review

backoff can overflow past 2^31 ms — clamp the upper bound too?

A conflict-free multi-agent orchestra — agents peer-review each other's work.

Git tracks the diff.

h5i tracks the run.

github.com/h5i-dev/h5i · h5i.dev

Auditable workspaces for AI coding agents

refs/h5i/* — the audit trail, in your repo