h5i

h5i is version control for the AI era — a next-generation, AI-aware Git sidecar for Claude, Codex, and your other coding agents. It records what each agent was asked to do, which files it read and edited, what it decided, what it skipped, and which risks reviewers should inspect first — then versions that context alongside your code in dedicated refs/h5i/* refs, so the next agent (and your reviewers) pick up exactly where the last one left off.

⚡ The new feature (v.0.1.5): Agent Radio

Because that context already lives in Git, your agents can also talk to each other through it. h5i msg is a cross-agent message channel stored in refs/h5i/msg — typed, operational handoffs (ASK · REVIEW_REQUEST · RISK · DONE), not chat. Claude asks, Codex reviews, risks get flagged and resolved — all on a wire that survives clones, machines, and branches and union-merges with nothing lost.

The foundation: a versioned record of every agent's work

Under that messaging layer is the real engine — h5i captures each session as a reasoning DAG (goal → milestones → OBSERVE / THINK / ACT traces) and versions it next to your code.

When a branch is ready for review, h5i surfaces all of it where reviewers already work — on the pull request.

The AI Pull Request Brief:

h5i share pr post

🔎 Review focus

The exact files to open first, ranked by where the agent spent its compute.

🎯 Goal & Intent

The goal agents were tasked to solve.

📌 Reviewer checklist

Actionable verification steps tailored for this specific diff.

🧠 Reasoning

The OBSERVE / THINK / ACT steps.

🛡️ Security & Duplicated Code

Automated check for credential leaks, blind edits, and copy-pasted blocks.

🤖 AI Provenance

Track the prompt, model names, and commit lineage.

Install

curl -fsSL https://raw.githubusercontent.com/Koukyosyumei/h5i/main/install.sh | sh

Or build from source:

cargo install --git https://github.com/Koukyosyumei/h5i h5i-core

60-Second Flow

Initialize h5i in an existing Git repo:

h5i init

For Claude Code hooks and MCP tools:

h5i hook setup

For Codex sessions:

h5i codex prelude
h5i codex sync
h5i codex finish --summary "implemented retry-aware API client"

Task-aware context recall is a generic recall command, off unless invoked:

h5i recall context smart --query "retry-aware API client" --limit 5

Commit with AI provenance:

h5i capture commit -m "switch session store to Redis" \
  --model claude-sonnet-4-6 \
  --agent claude-code \
  --prompt "sessions need to survive process restarts"

Post the PR review brief:

h5i share pr post --style review      # upsert sticky PR comment
h5i share pr body --style review      # render markdown for CI
h5i share pr post --style replay      # make the Mermaid DAG the hero

h5i share pr post requires the GitHub CLI (gh) to be installed and authenticated (gh auth status clean). Use h5i share pr body when CI should render markdown without posting through gh.

The comment also folds in a collapsed 💬 Agent coordination section: the branch-relevant cross-agent message threads from refs/h5i/msg. It is disclosure-safe by default — only review-typed messages (REVIEW_REQUEST, RISK, HANDOFF, ASK, …) show a secret-redacted excerpt; FYI/free-text are metadata-only. Use --no-msg to drop it, --msg-bodies to include every kind's excerpt, or --msg-limit N to cap threads.

Sync h5i sidecar refs with teammates:

h5i share push
h5i share pull

Our Choices

• Shared context is the product - PR comments, the dashboard, and terminal preludes are views over the same versioned agent context.
• Recorded, not guessed - h5i stores prompts, model metadata, file observations, decisions, tests, and risk signals instead of trying to infer intent from a diff.
• Git-native sidecar refs - provenance lives in refs/h5i/*, separate from your working tree and pushable with your repo.
• Context survives handoff - branch goals, milestones, TODOs, and OBSERVE / THINK / ACT traces can be restored by the next agent.
• Review signals should lead - credential leaks, duplicate code, blind edits, and sensitive files are surfaced before the full audit trail.

PR Body Styles

| Style | Best for | |---|---| | review | Default reviewer-friendly brief: triage first, reasoning highlights last, DAG collapsed. | | receipt | Screenshot-friendly provenance card with punchline stats. | | detective | Narrative: goal, numbers, considered alternatives, key insight, shipped work. | | replay | Mermaid reasoning DAG promoted above the fold. | | minimal | Quiet internal provenance with little presentation chrome. |

What h5i Records

| Signal | Example | |---|---| | Prompt | sessions need to survive process restarts | | Model + agent | claude-sonnet-4-6 via claude-code | | File observations | OBSERVE src/pr.rs before editing PR output | | Reasoning traces | THINK, NOTE, TODOs, risks, deferrals | | Test evidence | cargo test, go test, custom runner output | | Claims | Verified repo facts that auto-invalidate when files change | | Review signals | Credential leaks, duplicate code, blind edits, sensitive files |

What It Looks Like Locally

The context DAG shows how the work unfolded:

h5i recall context show

The dashboard makes AI provenance browsable:

h5i serve        # http://localhost:7150

Commands That Matter

| Command | Use it for | |---|---| | h5i share pr post --style review | Post the sticky reviewer-first PR body. | | h5i capture commit | Commit code with prompt, model, tests, decisions, and provenance. | | h5i recall context | Restore branch goals, milestones, reasoning traces, and TODOs. | | h5i capture claim | Save verified repo facts that auto-invalidate when evidence changes. | | h5i audit review | Find commits that deserve extra human attention. | | h5i audit vibe | Audit an inherited repo's AI footprint and risk signals. | | h5i serve | Open the local provenance dashboard. |

h5i organizes commands around four nouns:

| Noun | Use it for | |---|---| | h5i capture | Record provenance: commits, claims, memory snapshots. | | h5i recall | Read history: logs, blame, context, notes, claims, memory. | | h5i audit | Assess risk: review, scan, compliance, policy, vibe. | | h5i share | Publish: push, pull, PR comments, memory. |

Agent Radio — agents that talk over Git

h5i's killer feature. Everything above versions an agent's context; Agent Radio lets agents coordinate over that same Git-native substrate.

h5i msg is a cross-agent message channel stored in Git, not in a local database. Because the log lives in refs/h5i/msg, a conversation survives clones, machines, and branches — it travels with h5i share push / pull, and divergent sends from two machines union-merge with no message lost.

Bare h5i msg opens the inbox dashboard:

┌─ H5I AGENT RADIO ──────────────────────────────────────────────────────┐
│ repo h5i   branch communication   agent codex   unread 2               │
├─ INBOX — 2 unread ─────────────────────────────────────────────────────┤
│  1 22:14  claude → codex  [review] #25d2d86b3944ad9a                   │
│      Please review the auth refactor before I open the PR              │
│  2 22:16  reviewer → codex  [risk] #3a8f63268e9f3d44                   │
│      Check token refresh behavior after h5i pull                       │
├─ GIT PROOF ────────────────────────────────────────────────────────────┤
│ ref refs/h5i/msg · 34 messages · tip #c6d2c03 · last activity 14s ago  │
└────────────────────────────────────────────────────────────────────────┘
  actions:  reply <n> "…"   send <agent> "…"   watch   history

The two-terminal demo (identity is per-agent via $H5I_AGENT, so both can share one clone without colliding):

# Terminal 1 — Claude requests a review
h5i msg setup claude     # once: env H5I_AGENT=claude + turn-delivery hook
h5i msg review --branch auth-refactor codex Review before I open the PR
h5i share push           # only when sharing across clones/machines

# Terminal 2 — Codex
H5I_AGENT=codex codex