Product

It reads what the agent said, and checks it against what ran.

ThruLine organises your work around a Change: one feature, fix or refactor you are deciding whether to accept, even across several agent sessions. Six stages take it from “what counts as done” to approving one version — one exact commit of your code. Every fact on screen traces to a line in your own logs: the session files Claude Code and Codex already write on your machine.

Open to beta users. Built by one engineer. Nothing leaves your machine — how, on Security.

Define → Prepare → Build → Verify → Review → Approve

One Change, six stages

Each stage answers one plain question. The six are views of the same Change, so moving between them settles nothing.

Every screen below is rebuilt from the product’s own words, on the checkout-web demo record. Where part of a screen is still in build, its row says so.

The state vocabulary

Thirteen marks, one encoding

Every state is a word, a glyph and a line style, so the screen still reads with the colour off. The glyph is hidden from screen readers; the word is what they read.

Twelve things · eight built · four in build

What only ThruLine does

Every reviewer judges the pull request; ThruLine judges the agent’s claims against what actually ran, on the version you are about to accept. All twelve follow from one fact: the agent’s runs and words, in order, bound to one version.

Built

• #1

  The agent’s own words are what gets reviewed.

  Each thing the agent said it did, like “Tests pass.”, is read out of the session and checked on its own.

• #2

  “Proven” means a recorded run on this exact version, not a re-read.

  A line turns Proven only when a run in the record passed on the version you are looking at. A model reading the code again does not count.

• #3

  It knows the test ran before the last edit.

  The Brief names the file that changed after the run, and the step.

• #4

  Numbers that cannot drift from the record.

  Every count comes from fields in the record, not from prose, and says which way it counts: 1 failed, 3 passed (4 tests).

  The count that counted the passes →

• #5

  Test-integrity signals are joined to the claim.

  Signals about the test itself, such as a run that reports success (exit 0) having checked nothing, sit beside the claim they bear on. “Ran, checked nothing” is not a pass.

  Ran, checked nothing →

• #6

  Re-check on this version, with honest outcomes.

  A check ends one of six ways, and only a pass on this version counts.

• #9

  Approval names one version, and refuses a moved one.

  If the code moves while the screen is open, approving is refused and the file that changed is named. An approval never carries over to a later version.

• #10

  “Done” is proven by running, not by reading a ticket.

  What counts as done is a list you accept. Each line is proven by a recorded run, not by a model reading the ticket’s text.

• #7

  The repair loop closes on proof, not on a thread.

  Partly built; needs engine work.

• #8

  The agent hears receipts — records of what actually ran — before it claims.

  Needs wiring into the app.

• #11

  It says what it could not see.

  Mostly built; the rest needs engine work.

• #12

  Deterministic, tamper-evident, local and unmetered.

  Deterministic: the same record always gives the same answer, and no model decides. Tamper-evident: any change to the record shows, because it is hash-chained — each entry carries a fingerprint of the one before, so a changed byte breaks the chain. Local: nothing leaves your machine. Unmetered: no charge per review.

  Mostly engine work: the verdict path and the hash-chained record are built. How the record is kept →

The field · fetched 25 Sep 2026

What the field does, and what we do

Twelve AI code reviewers — tools that read a pull request (PR) and comment on it — as fetched on 25 September 2026: what each one is, what it costs, and the idea it does best. Every reviewer here reads the diff — the lines that changed — at a commit, and some run the code themselves, in a walled-off sandbox.

ThruLine reads something else: the agent’s own record of what it said and what it ran, in order, on the version you are about to accept.

Prices and descriptions as each source stated them on the fetch date. Nothing here has been re-checked since.

Sources checked for this table: docs.coderabbit.ai · walkthroughs · change-stack/findings · change-stack/snapshots · triage · reference/configuration · greptile.com · docs · cursor.com · docs/bugbot

Refused on purpose

What we deliberately do not build

Twelve of the twenty-one patterns the field ships that ThruLine declines on principle, each with what it does instead and why.

Try it on one Change.

ThruLine runs on macOS and Linux and opens with one command, thruline ui, in your browser on your own machine. Ctrl+C stops it. Open to beta users. Built by one engineer.