H

h5i

h5i.dev·Infrastructure

h5i provides a sandboxed, auditable workspace layer designed specifically for AI coding agents such as Claude Code, Codex, and agent teams. The tool assigns each AI coding agent a dedicated sandboxed Git worktree, enabling isolated execution and auditable tracking of all changes made by agents. h5i records detailed information for every modification, including prompts, commands, logs, policies, and reviews, ensuring traceability and accountability in agent-driven code development. One of its core features is prompt versioning, where every commit in the codebase is linked with the prompt, model, and agent responsible for producing it. This information is versioned alongside the code, providing a comprehensive audit trail. h5i also supports persistent context and memory: goals, reasoning traces, and memory snapshots are preserved across sessions, allowing agents to resume work without redundant re-grounding. For multi-agent workflows, h5i enables conflict-free orchestration of multiple agents working in parallel, with the ability to merge their independent contributions into a single, auditable outcome. Security is addressed through a supervised sandboxed environment that utilizes Landlock, seccomp, and Linux namespaces to confine each agent run, with optional network egress allowlisting. No changes are committed to the main codebase until explicitly applied, adding a layer of review and control. To optimize efficiency, h5i reduces token usage by up to 95% by presenting agents with compact summaries of failures, counts, and paths rather than raw tool output. The platform operates entirely within the user's own repository, leveraging Git for version control and storing all data in refs/h5i/, with no reliance on external SaaS infrastructure. This makes h5i suitable for users seeking secure, reproducible, and auditable workflows for AI coding agents, particularly in environments where code provenance and agent activity tracking are critical.