Read the diff. Then ask why.
WHY is a local-first desktop app for reviewing code your coding agents wrote. It shows the live diff of every repo and worktree you add, and what Claude Code and Codex did, turn by turn. Read-only, early, Linux only.

Because git diff tells you what, not why.
Coding agents write code fast. A human still has to understand and approve it, so review is now the bottleneck.
git diff shows what changed. It does not show the task the agent was given, why it made a change, whether a block was necessary, or whether the agent drifted from scope. Today people paste diffs back into the agent's chat and ask “why did you do that?”.
WHY aims to answer three questions in one place: what changed, how it changed, and why. An agent's own explanation can be an after-the-fact rationalisation, so WHY labels every explanation by how well it is backed. It never replaces reading the code.
Today, in v0.2.0
What WHY does
The live diff of every worktree
Add your repositories and WHY lists every git worktree, including the ones your agents create. See what changed against the base branch, uncommitted or staged, and watch it update while an agent works.
What the agents did, turn by turn
With hooks you install yourself, the inspector shows each Claude Code and Codex session: the task, the model, and a live timeline of turns, tool calls, edits, shell commands, test runs and subagents.
Which turns explain each changed file
Each changed file is marked as explained by agent turns, as “Agent tried, not applied”, or as “No captured agent edit”. WHY never guesses who made a change that nothing explains.
“Why?” in the agent's own words
Ask on a turn or a tool call to read what the agent said it was doing, straight from its own transcript, labelled Declared. Codex exit codes are recovered from its session log.
Every explanation says how it is backed
WHY labels each piece of evidence, and never presents a guess as fact. In v0.2 the labels you will see are Observed, Declared and Unknown.
- Observed
- The harness recorded it: a tool call, an edit, an exit code.
- Declared
- The agent said it. Shown as its own claim, never as proof.
- Inferred
- Reconstructed from the available records. Part of the model; v0.2 does not produce it yet.
- Unknown
- There is not enough evidence. WHY says so instead of guessing.
Unknown beats a confident guess. A tool call that started and never reported a result is shown as unknown, never as success. A change that no captured agent edit explains is shown as exactly that, because WHY cannot tell a human from a shell command or a lost event.
Safe to run next to your agents
WHY runs git against repositories that agents are using right now, so it is built never to disturb them.
A passive observer
WHY never writes to your repositories: no checkout, reset, add or commit. It reads them, and treats their config as untrusted.
Local only
No network requests, no telemetry. Nothing WHY captures leaves your machine.
Hooks are opt-in
The Agent hooks dialog shows the exact change to your user-level Claude Code or Codex config and writes nothing until you confirm. Uninstall restores your original file, or removes only WHY's entries if you have edited it since.
It never stalls your agent
The hook is fail-open: it always exits successfully within a short timeout, even if WHY is closed or broken.
It never reads agent credentials
WHY does not read your agents' credential files, and collects no account, plan or rate-limit data.
Programs in a repo need your OK
A repository can define git filter programs that run while git reads it. WHY refuses to read such a repository until you choose to trust it.
It is early software and has had no external security review. The security policy lists its known limits and how to report a problem.
What it is not, yet
These are plans, not shipped features. Everything from v0.3 on is a plan, not a promise, and nothing here has a date.
v0.3 · planned, not shipped
Line provenance
Attribute individual lines to the edits that produced them.
v0.4 · planned, not shipped
Attention
Signals such as unknown origin, churn and untested code, and a review queue ordered by them.
v0.5 · planned, not shipped
Close the loop
Anchored comments, and feedback sent back to the agent that wrote the code.
Limits today
- Linux only. Releases are an AppImage and a .deb; macOS and Windows have no builds.
- Unsigned builds and no auto-update. Check downloads against the published SHA256SUMS.
- Needs Git 2.45.1 or newer.
- Attribution is per file, not per line.
- Session context covers Claude Code and Codex. Codex asks you to approve the hook in its own /hooks screen, and WHY cannot tell whether you did.
Get WHY
WHY v0.2.0 ships as a Linux AppImage and a .deb. Download one from the latest release.
Requirements
- Git 2.45.1 or newer on your
PATH. Check withgit --version. - webkit2gtk 4.1 and GTK 3. The
.debinstalls them for you; the AppImage bundles its own copy.
Check the download
Builds are unsigned. Download SHA256SUMS from the same release and run this next to the files:
sha256sum --check --ignore-missing SHA256SUMSAppImage, any distribution
chmod +x WHY_0.2.0_amd64.AppImage
./WHY_0.2.0_amd64.AppImageIf it fails with a FUSE error, install libfuse2 (libfuse2t64 on Ubuntu 24.04).
Debian, Ubuntu and derivatives
sudo apt install ./WHY_0.2.0_amd64.debThen start WHY from your application menu, or run why.
Uninstall the agent hooks from the Agent hooks dialog before you remove the app, so your agent config no longer points at the hook.