Git Why

git blame tells you who changed the code. git why finds the history that explains it.

Local-first semantic + full-text search over Git history. It returns the actual commit, its author's real words, and the relevant diff. It retrieves evidence; it does not generate an explanation of its own.

Real output from mise run demo, which rebuilds a small fixture repository from bench/fixtures/demo/build.mjs and runs these two queries against it. Nothing here is hand-written, and you can reproduce it in one command.

For output against real repositories — curl, zod, redis — including a worked case where this is the wrong tool, see docs/examples.md or the recorded terminal sessions.

$ git why "that bizarre bug where reconnecting subscribed twice"

1. 7deab42  Stop duplicate subscriptions after reconnect
   2025-11-20 · Maya Chen

   Reconnecting re-ran the subscribe handler without clearing the previous
   registration, so every reconnect doubled the delivered events.

   src/net/socket.ts
   -export function connect(url) { return new Socket(url); }
   +export function connect(url) {
   +  const s = new Socket(url);
   +  s.on('reconnect', () => resubscribeOnce(s));
   +  return s;
   +}

$ git why "why do we keep the session when the refresh token is empty?"

1. f17db20  Fix infinite token-refresh loop
   2025-11-03 · Maya Chen

   Provider X can return an empty refresh token while the current access
   token remains valid. Retrying here puts clients into an infinite loop.

   src/auth/refresh.ts
    export function refresh(session, refreshToken) {
   -  if (!refreshToken) throw new InvalidTokenError();
   +  if (!refreshToken) return session;
      return exchange(refreshToken);
    }

Install

curl -fsSL https://alliecatowo.github.io/git-why/install.sh | sh

npm install -g @alliecatowo/git-why

The package installs a git-why executable, which Git dispatches as the subcommand git why. No alias setup needed. More install paths — pinned versions, a GitHub release tarball, building from source, uninstalling — are in docs/install.md.

Use it with an agent

Two plugins ship in plugins/:

OpenCode users get opencode/ — an opencode.json with the MCP registration and an AGENTS.md fragment.

What the skill actually teaches is when not to reach for history. It tells an agent to use git log -S when it can name the symbol, because that is a case this tool measurably loses (Hit@10 0.950 against 0.350), and to use code search rather than history for the current state of the code. A skill that claims its own tool is always best makes an agent worse at its job.

See docs/plugin.md.

Make it faster (optional)

git why server on

Holds the index, the embedding model and the lineage table open between queries. On curl — 30,000 commits — that is 569 ms a query down to 286 ms. Searches use it automatically once it is running.

It can never be the reason a search fails: if no daemon is running, or it is unreachable, or your index moved under it, the query runs directly instead. --daemon=direct opts out, --daemon=server requires one. See docs/daemon.md.

Why

git log --grep only matches words you already know. The reason code changed usually lives in a commit message, but you rarely remember its exact wording — you remember the problem, in your own words, months later.

Look again at the first example above: the query is "reconnecting subscribed twice" and the commit is titled "Stop duplicate subscriptions after reconnect." They share almost no vocabulary. That's not a coincidence Git Why is showing off — it's the actual point of a semantic index. It carries the meaning of the change, not just its words, alongside an ordinary keyword index for when you do know the exact term.

Git Why retrieves historical evidence: the commit, the author's actual words, and the relevant diff. Historical commit messages are assertions by their authors, not infallible accounts of intent, and some reasons were never committed at all — Git Why will not manufacture those.

Measured, not claimed

Every number below is written into this file by bench/report.mjs from raw run data, and CI fails if it drifts — no figure here was typed by hand. Full methodology, limits, and the negative results are in docs/report.md.

Against the tools you would otherwise use

174 recall questions — you remember a problem but cannot name anything in the commit that fixed it — derived mechanically from 6 pinned real repositories (curl, redis, requests, ripgrep, caddy, zod). Every question is verified unanswerable by keyword search before it enters the set: if git log --grep or git log -S finds the answer from the question's own words, the case is discarded.

10.2x zg and 23x the best Git-native strategy — and the only approach that answers nearly every question rather than returning an empty set.

Where it loses

When you can name the symbol, use pickaxe search instead. On cross-file causal questions, git log -S scores Hit@10 0.950 against git why's 0.350. Semantic search has no advantage over a tool you can hand the exact literal.

That boundary is the honest positioning, and the shipped skill tells agents both halves:

• cannot name the term → git why
• can name the term → git log -S
• current code, not history → zg

Scale and cost

Does it help an agent?

Retrieval quality is not the product. The question is whether an agent answering a real question does it more accurately, or in fewer turns, with the tool than without. Four arms over the same frozen tasks, paired per task, with token counts reconciled against the provider's own accounting database.

Results are mixed across models. At single-digit paired n per model this is descriptive, not significant, and it is reported that way deliberately — the direction is consistent, the magnitude is not established. Full method and per-arm figures in the benchmark report.

Method, per-arm figures and the registered hypothesis: docs/report.md.

What works, and what does not

Eight optimisations were implemented and measured. One survived.

The one that did: lexical overlap between the question and the commit's own message, reranked over a candidate pool deeper than the result list. Weight chosen on a dev half, evaluated once on the held-out half — +26% MRR and +44% Hit@1 on cases it never saw. Shipped.

The seven that did not: a prose-tuned embedding model, pseudo-relevance feedback, a prose-commit penalty, caller-side query restatement, wider result windows, structural expansion, and phrase fusion. None improved MRR.

What made the difference was not a better idea but a different question. The seven all asked "does this rank better". The one that worked started by asking where the right commit actually is — and found that a quarter of the corpus is retrieved but ranked below 5, which is a reordering problem, while 42% is never retrieved at all, which is not. See docs/decisions.md.

The remaining headroom is that 42%: a recall problem in the embedding itself. That is now measured rather than guessed — docs/embedding.md compares ten models on a fixed pool. Transformer embedders score +41% (jina-v2-small) and +64% (jina-v2-base) MRR over the shipped static one, at 191x and 1287x the indexing time plus an ONNX runtime. Both are available opt-in; the default stays fast and dependency-free, because a tool that installs in seconds and indexes in under a minute should not quietly become one that needs fifteen hours.

It is also wrong most of the time. Hit@5 of 0.374 means the right commit is outside the top five on 62.6% of these questions. It beats every alternative on them and still fails on most. Treat a result as a lead to verify with git show, never as established fact.

When to use ordinary Git instead

Git Why is ranked retrieval over a semantic index. That's the wrong tool for some jobs, and Git already has the right one:

• A known exact string. git log -S and git log -G are exhaustive. Keyword mode here is ranked, not exhaustive.
• An exhaustive search. ripgrep over a checkout, or git grep.
• Verifying causality. Similarity is not a timeline. Confirm ancestry with git merge-base, git log --ancestry-path and git show.

Git Why is for the case where you remember what happened but not what it was called.

Learn more

• Documentation site — guided install, how retrieval works, CLI reference, FAQ.
• ROADMAP.md — pull requests, gh why, wikis, and what would need measuring before any of it ships.
• docs/embedding.md — nine embedding models measured, the +41% one you can opt into, and why the default did not change.
• docs/daemon.md — the optional daemon: what it holds warm, why it can never break a search, and its security posture.
• docs/indexes.md — where indexes live, worktrees, submodules, monorepos, disk use, and the shared model cache.
• docs/examples.md — real output on real repositories, including a case where this is the wrong tool.
• docs/plugin.md — both plugins, the MCP tools, the skills, and why there is no hook.
• docs/decisions.md — what was tried and rejected, with the measurements. Seven optimisations that did not work.
• docs/install.md — every install path and uninstall, including where the index and model cache live on disk.
• docs/operations.md — the full operational contract: durability, concurrency, exit codes, history scope, coverage limits.
• docs/report.md — the benchmark report this README's numbers come from.
• docs/spec.md — the build specification this was written against. Source comments cite its sections, so it is kept as provenance rather than as user documentation.

Development

mise setup             # install dependencies, then doctor
mise check             # format, lint, typecheck, protocol hash, unit tests
mise test:integration  # real repositories, real storage, real locks
mise test:package      # pack, install into a clean prefix, invoke through Git
mise site:build        # build the docs/marketing site in site/
mise tasks             # everything else, including every benchmark

Numbers in README.md, site/index.md and docs/report.md are written by bench/report.mjs from raw run data — CI fails if you edit one by hand. The man page, the site's CLI reference and all three shell completions are checked against git why -h, which is the only place the flag set is defined.

See docs/contributing.md for module ownership and how to cut a release.

License

Apache-2.0. See LICENSE and NOTICE.