How branchdiff grew up.

git diff compares commit paths. branchdiff compares content.

Two branches can hold identical bytes and still produce a phantom diff — because Git's two-dot diff compares commit history, not file content. A rebase, a cherry-pick, a squash: Git reports a "change" where there is none. That phantom diff is the whole reason branchdiff exists.

Same SHA-1 blob hash on both tips → no diff. Comparing maps of hashes is O(n) on files, not on history.

Deliberately small. Four flags.

The default compared actual content — the thing Git's own diff gets wrong. Tab-completion came free from the built-in readline; the whole thing was one local page reading the diff from disk.

No roadmap. Each version was a toe I kept stubbing.

The release notes read less like marketing and more like a diary of things that irritated me. Each feature starts with "I keep having to…"

• Comments had to land on real lines. Reviewing meant typing into a disconnected text box. So: inline threads that pin to exact lines — and an AI pass that drops comments on those same lines. v1.1.0
• Local work had to reach the remote PR. I'd write twelve careful comments, then retype them into GitHub. So: two-way sync — local pushes up, PR pulls back down. v1.2.0
• "I don't have Node installed." Every time I told someone to try it, that was the reply. So: six install methods — npm, PyPI, Homebrew, Scoop, apt, binaries. v1.4.0
• Approve and merge without leaving the terminal. The full PR lifecycle — approve, request changes, merge, close, reopen, mark ready — as CLI commands. v1.5.0

The unit of work is "my open PRs", not "this checkout's PRs".

I keep about ten side projects going at once. Each has its own repo, each occasionally has an open PR. Pointing auto at a directory replaced an entire morning of context-switching:

Each flag is a sentence that ends with "…so I added a flag".

The same command that reviews one PR reviews a hundred.

The only difference is which directory you point it at. That is what makes it a self-hosted review bot a team can actually run.

A review for PR #2124 landed on PR #2125.

Two sessions on the same repo, two ports. One shared current-session pointer — rewritten on every browser status poll. The other PR's tab, polling in the background, flipped the pointer mid-review. The fix: stop the flip at its source, pin the reviewer's identity explicitly, force one chokepoint.

explicit identity beats ambient context — the reviewer checks its own pin before it ever reads the shared pointer.

Each bug left a habit behind.

• It hung people's machines. Synchronous Git calls blocked the single Node thread; on big repos the server froze, and on some setups the device hung on sleep. Habit: never block the hot path — go async, pass arg arrays (also closes a shell-injection door).
• cron silently did nothing on macOS. Since Catalina, macOS blocks cron unless it has Full Disk Access — and fails silently. Habit: distrust the platform's defaults; "works on my Linux box" is not a test. On macOS it now writes launch agents instead.
• Export and import were both broken — and silent. Server shipped bundles tagged v2; the import screen rejected anything but v1. The only live path was dead end to end, and nobody reported it. Habit: version your contracts out loud.
• "Request Changes" did nothing on Bitbucket. It posted to the comments endpoint instead of the dedicated request-changes endpoint — a wrong URL that looked like it worked. Habit: verify the side effect, not the call.

The v1 → v2 leap: twenty-one releases in four months.

the leap — from reviewing one PR to reviewing a directory of repos unattended

Same wall of hunks. Different reader this time.

Once auto was reviewing unattended, the AI was still cold-reading every diff — no orientation, just a wall of hunks. That is the exact problem that started this whole project, except now it was the AI stubbing its toe instead of me.

One diagram, per wired section, before a single line of review.

• Which areas moved — and by how much.
• Which areas are wired together by imports — labeled with the actual new symbols the diff introduces, not just an import count.
• One coherent change, or several bundled into a PR — told apart automatically.
• The AI's own review comment now opens with that same diagram before its line-by-line findings.

same shape, two readers: the human's wall of hunks, then the AI's

Then it moved to where you already were.

The last itch was not a feature — it was a doorway. Every workflow in this story starts with a command and a URL, and the people who stuck around were the ones already living in a terminal. So both groups got their own front door.

The real payoff is the habits, not the features.

• Ship the itch you actually have. Features that scratched a real, recurring annoyance are the ones people thank you for; the clever planned ones mostly didn't survive contact with how I really work.
• Explicit identity beats ambient context. Any component that figures out "who am I" by reading a shared mutable thing has a race condition pretending to be a feature.
• Never block the hot path. If a request waits for a subprocess, it waits asynchronously — or not at all.
• Version your contracts out loud. Silent disagreement between two parts is the worst bug: it looks alive and is dead.
• Distrust the platform's defaults. macOS cron, sync APIs, 0.0.0.0 bindings — the convenient default is frequently the wrong one.
• Meet people where they are. "I don't have Node" killed adoption; six package managers was tedious and also the difference between a toy and a tool.
• Reuse your own commands. v2 was largely built by composing v1 commands as child processes — less code, fewer bugs, every fix inherited for free.