h5i (pronounced "high-five") is an open-source Git sidecar designed for teams where AI agents assist in the development process. While Git is the standard for versioning code, it fails to record the intent, reasoning, and context behind changes made by AI models like Claude Code or Codex. h5i fills these gaps by storing AI provenance, agent reasoning traces, and cross-agent communication alongside your existing Git repository without altering the actual source tree or commit history. By utilizing Git refs, h5i ensures that the AI's versioned context travels with the code, making it accessible to team members and other AI agents.

Some of the key features are:

• AI Provenance Tracking: Automatically records the prompt, model, agent, test results, and audit signals for every AI-generated commit.
• Agent-to-Agent Messaging (i5h): Facilitates typed work handoffs—such as review requests, risks, and task handoffs—between different agents over a shared Git ref.
• Token-Saving Object Store: Captures raw tool output—such as massive build logs or test results—out-of-band and surfaces structured summaries to keep the AI's context window clear.
• Auditable Agent Sandbox: Provides tiered isolation (Landlock, seccomp, namespaces) and network egress allowlisting to run agents in a confined environment before applying changes.
• Versioned Reasoning: Keeps a trace of OBSERVE, THINK, and ACT steps so sessions can resume with full context instead of re-grounding from scratch.
• Deterministic Audit: Applies rule-based, reproducible checks to identify risks like prompt injection, secrets, or blind edits without needing a model-in-the-loop.
• Sticky PR Briefs: Generates idempotent GitHub PR comments that render the full reasoning, risk signals, and provenance so reviewers can evaluate AI work efficiently.

Operating h5i is simple and mimics standard Git workflows. Once initialized in a repository with h5i init, you can wrap commands with h5i capture to store metadata, use h5i recall to view historical reasoning or blame information, and leverage h5i share to sync this information across your team. Because it lives in Git, no new SaaS services or daemons are required. It operates as a local tool, ensuring your data remains private within your own infrastructure.

Some common use cases include:

• AI Session Resumption: Resuming complex tasks by restoring the full reasoning state from a previous session, preventing repetitive re-derivation.
• Multi-Agent Coordination: Orchestrating complex refactors by letting two distinct AI agents, such as Claude and Codex, exchange review requests and feedback through the repository.
• Evidence-Based Code Review: Providing human reviewers with a comprehensive brief that includes the agent's intent, the files it accessed, and the integrity of the test runs that validated the changes.
• Secure Agent Sandboxing: Running untrusted or experimental AI agents inside a confined environment where network access and local file system modifications are strictly controlled and logged.