Run many coding agents, merge one auditable result. Start with the end-to-end workflow, then dive into task-focused walkthroughs for using h5i with Claude Code and Codex: persistent memory and handoff, multi-agent coordination, AI code provenance, auditing AI-generated code, prompt-injection detection, and AI blame.
Wrap noisy commands with h5i capture run so only a structured summary reaches the agent; recover the full output on demand, and query captures by status, tool, branch, or file.
Snapshot Claude Code's memory into Git, version your reasoning, and resume the next session from a briefing instead of re-grounding.
Let two coding agents coordinate over a Git ref — ask, review, hand off, and reply as durable, replayable messages with no server.
Attach the prompt, model, agent, tokens, and test result to every AI-assisted commit — stored in Git, queryable like git log.
Rank AI commits by uncertainty, blind edits, churn, and scope so reviewers spend their attention on the riskiest changes first.
Keep the real API key on the host. The agent box authenticates through a credential-injecting proxy, so a prompt-injected agent has no reusable token to steal.
Scan Claude Code reasoning traces for prompt-injection and exfiltration patterns — deterministic, with no model in the detection path.
Per-line provenance for AI-written code: the prompt, the model, the agent, and the test result — same ergonomics as git blame.