Threshold

Threshold

Nothing crosses without being seen.

A desktop review inbox that lets you actually review code again.

A change arrives cut into chapters, the idea first and the mechanical edits last, with the few lines that matter already flagged. Ask about any line and the answer comes back in a thread on that line. It works for your agent's own changes and for colleagues' PRs.

The problem

"Shall I proceed?"

An agent finishes something wide-reaching: a schema migration, a mass rename, a bulk delete. It asks permission in a line of chat.

That line asks you to approve a change you cannot see. Four hundred lines, summarized in one sentence, by the author, who is sure it's fine.

So you type y. Everyone does.

agent
Done. I moved every book cover into storage, deleted the local copies, and updated 22 templates to use the new URLs. Tests pass. Shall I proceed with the push?
you
y
// one of the 22 was in an email template outside this repo

The insight

Git tells a change in alphabetical order.
Someone who understands it can tell it in chapters.

An agent that wrote the change, or has just read it, knows which part is the idea, which parts follow from it, and which are 22 identical call-site edits. Threshold has it cut the diff that way: chapters in the order that explains the change, flags on the three lines that could bite, and a thread on any line you ask about.

A review

A diff, cut into chapters.

The brief

What the change is, what it implies, and what's worth a second look. Before a single line of code.

Chapters

The diff in the order that explains it: the idea, then what follows from it, then the mechanical edits, labelled "skim this".

Flags

Out of four hundred lines, the three that matter: attention, caution, context.

Threads

Ask on any line, chapter or the whole change. The answer lands in a thread on that line, and stays there.

The verdict

Approve, or reject with a reason. Either names the exact diff that was on your screen.

Two ways in

Your agent's change. Your colleague's PR.

a · before it lands

Your agent hands you its work.

It's finished something wide-reaching, like a migration, a mass rename or a refactor. It commits to a branch and writes the review. When you ask about a line, it answers in a thread. If you find a problem, it fixes it and commits, and the review moves to the new commits.

b · someone else's change

An agent reviews a colleague's PR with you.

Ask it to read the PR and hand it to you in Threshold, with chapters, flags on what you should look at, and what it couldn't work out from the code. You read it, ask about anything, and take what you've learned back to the PR. The agent didn't write the change, so it can explain it but not revise it.

How it works

How reviews work.

coverage

Review the whole diff

Threshold checks that each hunk of the diff is in one of the agent's chapters. If any are left out, the count turns red and the missing hunks get a section of their own.

pinning

Reviews don't drift

When a review arrives, Threshold records the two commits it compares. Later commits on the branch or on main don't change it. Your approval includes a SHA-256 hash of that diff, so the agent can check it before shipping.

~/.threshold/reviews/

Reviews are files

Each review is a JSONL file the agent appends to. There's no server or SDK: any agent that can run shell commands can use Threshold, and you can read the file yourself.

Away from the desk

The doorway follows you.

An agent waiting on your answer shouldn't be stuck until you're back at a desk. Threshold serves your reviews to your phone over your own tailnet: read, ask, mark chapters, approve or reject, and get a notification when an agent has said something you're waiting on.

There is nothing to download for the phone. Pair it from the Mac by scanning a code that works once, for one minute.

how it stays safe
  • Listens on 127.0.0.1; tailscale serve puts it on your tailnet, never the public internet.
  • A verdict from the phone names the diff it read, and is refused if that diff has changed.
  • Every request carries a device credential. Revoke a phone from the Mac.
  • Notifications are encrypted end to end; the push service relays what it can't read.

Download · v0.1.0

Pick up Threshold.

macOS

Apple silicon (M1 and later). In daily use.
Threshold_0.1.0_aarch64.dmg
Download .dmg

macOS

Intel Macs.
x86_64 · not built yet
Download .dmg

Linux

Builds and passes its tests; not packaged yet.
.AppImage / .deb · not built yet
Download

Windows

Not tried yet.
 
Download
first open

The builds aren't notarized yet, so the first launch stops with a warning that offers only Done or Move to Trash. Click Done, then go to System Settings → Privacy & Security, scroll down, and click Open Anyway. After that it opens like anything else. Step by step →

then

Give your agent the threshold-review skill, so it knows how to hand you a review. Hooking up an agent →

Source · MIT

The source code.

github.com/thirdbearsolutions/threshold.app
View source