src/auth.ts. Three sentences across your README and
CLAUDE.md quietly became wrong, and nobody will notice until a teammate — or an
AI agent — trusts one of them. Run hibi diff --since origin/main and Hibi names
the exact sentences your change broke, then hands you the command to fix each.
Hibi (日々, “day after day”) tracks claims: sentences in your docs and
AI-agent instruction files that assert how the code behaves. You anchor each
claim to the code it describes. When either side drifts (you change the code, or
you edit the documented sentence itself), hibi check flags the claim and can
stamp a status banner into the doc, so no reader, and no agent, acts on a page
that has fallen out of sync with the source.
Run it in CI, in a git hook, or as a pre-edit lookup an agent makes before it
trusts a doc.
The mental model
A claim is a binding between a documented sentence and the code it describes. You record one withhibi record, pointing at a span of prose and
the code span(s) it asserts something about. Hibi stores a bidirectional
anchor (pointers to both sides, not a copy of the prose) in a committed
store beside your docs (.claims/, one file per claim).
On every hibi check, Hibi resolves the doc side first: it re-reads the
live sentence from the current document. If that sentence is gone or has been
edited, Hibi flags the claim and stops. It will not verify a stale
sentence against code as if it still existed. If the sentence is intact, Hibi
localizes the code side and grades whether the code still matches.
The result is a verdict with two independent axes, read like
doc:unchanged · code:changed · behavior:at-risk. A flag is a request to
re-verify, not a claim that the doc is wrong: the evidence under the
sentence moved, and a human or agent decides what the prose should now say.
Hibi never rewrites prose. It flags, and can stamp a banner.
Hibi is deterministic: no model runs in the check loop. The same working
tree always yields the same verdicts. An optional semantic resolver may advise
out-of-process, but it never decides a verdict and never gates.
Start here
Quickstart
Install Hibi and record your first claim in a couple of minutes.
How it works
Claims, anchors, and the doc-first flow that grades verdicts deterministically.
Why Hibi
The problem, the threat model, and why no model sits on the verdict path.
CLI reference
Every command, flag, exit code, and strictness level.
Resolvers & SDKs
Extend drift detection with out-of-process resolvers in any language.
Claude Code skill
Install the official Agent Skill so coding agents use Hibi correctly.
What you can rely on
Deterministic
No model runs in the check loop, so the same working tree yields the same
verdicts every time. The optional semantic resolver only advises.
A flag means re-verify
Hibi reports that the evidence under a claim moved: suspect, not false. It
never declares a doc wrong on its own.
Any file format
Hibi treats docs as text, so Markdown, plain text, AsciiDoc, or anything
else works without a per-format parser in the core.
Offline & shallow-clone safe
The anchor is its own baseline, so
check reads your files, never git
history. It stays correct even under git clone --depth=1.