A worktree, a paper trail, and a referee for your AI agents

What it does

h5i gives every AI coding agent (Claude Code, Codex) an isolated, sandboxed Git worktree, then versions the prompts, commands, and context steps behind every change as Git objects. You review the diff before it lands on your branch, and the whole audit trail travels with your repo under refs/h5i/* — no SaaS involved. It also compresses noisy tool output (pytest failures only, raw logs recoverable), claiming up to 95% token reduction.

The interesting bit

The multi-agent orchestration is ordinary async code, not YAML soup: hire two agents, have them implement the same task in parallel, cross-review each other, verify in a neutral sandbox, and let a judge pick the smallest passing diff. The audit tooling is unusually honest for this space — h5i audit maturity will tell you your prompts are mediocre, with per-dimension scores.

Key highlights

• Per-agent sandboxed worktrees: no file, branch, or port clashes; agents run with permissions off inside the box
• Prompt and context versioning as Git objects, shareable across a team via h5i share
• PR summaries covering prompt quality, AI/human commit ratio, secret leaks, and prompt injection
• Orchestration SDK in Rust and Python with parallelism, loops, and conditionals
• Token-reduction filters derived from rtk and headroom (both Apache-2.0, credited properly)

Caveats

• The repo description talks up red-teaming and Lean 4 formal verification; the README covers none of that. If those are the features you came for, the docs don’t currently show them.
• The 95% token-reduction figure is a stated claim, not a benchmark you can inspect from the README.
• PR posting depends on the gh CLI being present.

Verdict

Worth a look if you’re running fleets of coding agents and need defensible review trails. If you’re a solo dev with one agent and a strong stomach, it’s probably more ceremony than you need.