docs-refreshclaude-1
→ mainh5i env shell docs-refresh
Run Claude Code, Codex, or whole agent teams inside sandboxes. h5i records the prompts, commands, and reviews behind every change.
curl -fsSL https://raw.githubusercontent.com/h5i-dev/h5i/main/install.sh | sh
h5i (pronounced high-five) gives every AI coding agent a sandboxed Git worktree, and records the prompts, commands, logs, policies, and reviews behind every change. Run one agent safely, scale to many via a conflict-free multi-agent orchestra, then merge one auditable result. It all lives in your repo, carried by Git, with no SaaS.
apply.A live record of what each agent did, why it changed the code, and what evidence supports merging it.
Agents now produce more patches than any team can safely review. Git shows you the diff, but not the prompt, the commands, the sandbox policy, the failed attempts, the peer reviews, or the verifier evidence behind it. The scarce resource is no longer writing code. It is accepting it, whether you run one agent or many:
Isolate the run, record the evidence, then review it before anything merges. It works the same for one Claude Code session or a whole team, and everything writes to Git-native sidecar refs under refs/h5i/* that travel with your repo.
apply.refs/h5i/*. Raw logs stay out-of-band, so agents see compact summaries and context stays lean, up to 95% less token waste.audit review, then merge one evidence-backed result. For agent teams, sealed candidates can also be replayed by a neutral verifier.Powered by h5i env, h5i capture, h5i audit, h5i team, and refs/h5i/*.
import asyncio from h5i.orchestra import Conductor async def main(): async with Conductor(".", "fix-auth") as c: claude = await c.hire("claude-agent", runtime="claude") codex = await c.hire("codex-agent", runtime="codex") # Both agents implement the task in parallel claude_work, codex_work = await asyncio.gather( claude.work(task), codex.work(task)) # Seal the round: no agent saw a peer's attempt await c.freeze() # Each agent reviews the other's work await asyncio.gather( codex.review(claude_work), claude.review(codex_work)) # Replay each submission in a fresh, neutral sandbox await c.verify(claude_work, ["pytest", "--quiet"]) await c.verify(codex_work, ["pytest", "--quiet"]) # Smallest diff among submissions that pass all tests verdict = await c.judge() print("winner:", verdict.selected_submission) asyncio.run(main("implement quicksort in python with unit test"))
The same auditable workspace that contains one agent scales to a whole ensemble, and the workflow itself is a program you write. With h5i-orchestra, parallel fan-out, loops, and conditionals are ordinary control flow in Python or Rust: hire a roster of personas, let them work sealed and in parallel, cross-review, verify in a neutral sandbox, and let an explainable verdict pick the winner. Every step is journaled to refs/h5i/team, so an interrupted run resumes instead of re-paying agent turns.
pip3 install h5i-orchestra
The Python SDK lives at github.com/h5i-dev/h5i-python. Prefer Rust? The same define-by-run eDSL ships in the main repo as h5i-orchestra.
refs/h5i/*, so it travels with your repo with no hosted control plane.h5i team runs several agents on the same task, each in its own isolated h5i env, through a phased protocol: create a roster of personas (runtime, model, role) → dispatch the task to all → each agent works sealed and independently → freeze immutable submissions → permissioned review → a neutral sandboxed verify → apply one winner. Run state is an append-only event log in refs/h5i/team that travels with the repo. It's a thin coordination layer, not a group chat, not a daemon.h5i team verify replays each frozen candidate at the shared base in a fresh, sandboxed worktree and runs the declared command itself, so finalization never trusts an agent's own captures. Hard gates (tests pass, applies cleanly) come only from that run; smallest-diff breaks ties only among gate-passers, so nothing wins by deleting tests or stubbing features. The verdict is explainable, apply is gated, and if nothing clears the gates it records no_verdict instead of merging a loser.prelude, sync, and finish commands. Both agents restore the same shared context when a session starts and check their work back in when it ends.refs/h5i/* refs that travel with the repo without touching your commit history.push/pull, no SaaS. It makes agent work reproducible, reviewable, and accountable without adding a new system of record.h5i audit scan applies deterministic regex rules to every OBSERVE/THINK/ACT entry in an agent's reasoning trace to flag prompt-injection signals, with no model in the audit path, and can rank which commits most need human review.h5i env gives an agent a disposable, confined environment, a git worktree plus a policy that limits what code can read, write, and reach over the network. It picks the strongest isolation the host can enforce (Landlock + seccomp + namespaces, up to a rootless network egress allowlist), records every command and blocked access, and nothing reaches your branch until you apply. No root and no VM required.Spin up a roster of agents, seal their independent attempts, let a neutral verifier pick the winner, and merge one auditable result. Versioned in your Git. Apache 2.0. No SaaS, no lock-in.