Built for TestSprite Season 4

See what the AI builds.Know when it breaks it.

The black-box recorder and time-travel autopsy for AI coding agents.

01Records every turn
02Blames the exact prompt
03Rolls the file back in one click

01The problem

AI moved the bottleneckfrom writing codeto understanding it.

Agentic coding made writing code cheap. Understanding what was written is the new bottleneck. A pull request can be green on types and still leak memory, spin the CPU, or slow p95 latency by 400 ms.

CI / types
01
Does it compile?
A green check on types says nothing about memory, CPU or latency.
TestSprite
02
Does the PR break the app?
Red or green at the merge, tested against the deployed app.
Ghost
03
Which turn broke it, and how do I undo it?
Names the exact turn behind a regression and rolls the file back.

Ghost is the layer behind the maker, not another checker at the merge.

02The loop

Record. Attribute.Reverse.

Three steps run behind every agent turn. Ghost watches the filesystem and the process, so facts come from the app itself, not from what the agent says it did.

01Record

Every change, on the record.

Ghost takes a shadow git commit for every burst of changes, without moving HEAD or touching your branch. It samples runtime metrics from your app and probes HTTP for latency and errors.

Shadow commit per change burst
cpu, rss, heap and uptime from /api/health
HTTP probe for latency and errors

02Attribute

The turn that did it, named.

A threshold check, a trend slope and a per-turn bisect name the exact turn that introduced the regression. In the demo session memory climbs about 38 MB per sample for two turns, and the autopsy blames turn 4.

Threshold check
Trend slope
Per-turn bisect

03Reverse

Undo exactly that turn.

Roll the blamed file back to the pre-regression commit, re-measure, and watch the curve flatten. In the demo, invoices.ts goes back to its t3 version.

Blamed file rolled back
Re-measured after the rollback
Curve flat again

01Record

Every change, on the record.

Ghost takes a shadow git commit for every burst of changes, without moving HEAD or touching your branch. It samples runtime metrics from your app and probes HTTP for latency and errors.

Shadow commit per change burst
cpu, rss, heap and uptime from /api/health
HTTP probe for latency and errors

02Attribute

The turn that did it, named.

A threshold check, a trend slope and a per-turn bisect name the exact turn that introduced the regression. In the demo session memory climbs about 38 MB per sample for two turns, and the autopsy blames turn 4.

Threshold check
Trend slope
Per-turn bisect

03Reverse

Undo exactly that turn.

Roll the blamed file back to the pre-regression commit, re-measure, and watch the curve flatten. In the demo, invoices.ts goes back to its t3 version.

Blamed file rolled back
Re-measured after the rollback
Curve flat again

03The autopsy

Everything the agent did,on one screen.

The playground replays a recorded session: the turns, the diff that did it, the metric that caught it, and a fix prompt you can hand straight back to the agent.

Timeline

7 turns · shadow commits

t1Add pagination to the invoices list endpoint.app/api/invoices/route.tshealthy
t2Format amounts in the customer's currency.lib/format.tshealthy
t3Add a status filter to the invoice table.components/InvoiceTable.tsxhealthy
t4Speed up invoice lookups.lib/invoices.tsblamed
t5Add a PDF export for a single invoice.lib/pdf.ts, app/api/invoices/[id]/pdf/route.tsmemory climbing
Autopsy blames t4. Ghost rolls lib/invoices.ts back to t3.re-measured: memory curve flat
t6Show an empty state when there are no invoices.app/invoices/page.tsxhealthy
t7Add tests for the invoices endpoints.tests/invoices.test.tshealthy

Autopsy

Regression found

The regression first appears at t4. lib/invoices.ts is the blamed file.

Blamed turn: t4 · Speed up invoice lookups.
File: lib/invoices.ts
Cause: Unbounded in-memory cache

Threshold checkMemory crossed the session threshold during t5.
Trend slopeAbout +38 MB per sample across t4 and t5.
Per-turn bisectThe regression first appears at t4.

Replay in playground

Diff that did it

t4 · Speed up invoice lookups.

lib/invoices.ts· change in t4+5−0

1	1		import { db } from './db'
2	2
	3	added	const cache = new Map<string, Invoice[]>()
	4	added
3	5		export async function listInvoices({ page, pageSize }: ListOptions) {
	6	added	const key = `${page}:${pageSize}:${Date.now()}`new key on every request
	7	added	if (cache.has(key)) return cache.get(key)!
4	8		const invoices = await db.invoice.findMany({
5	9		skip: (page - 1) * pageSize,
6	10		take: pageSize,
7	11		include: { lines: true, customer: true }
8	12		})
	13	added	cache.set(key, invoices)
9	14		return invoices
10	15		}

Memory

from /api/health: cpu, rss, heap, uptime

Facts come from the app, not from what the agent says it did. The heal flattens the curve.

Agent-agnostic

Ghost watches the filesystem and the process, not the agent.

Every one of these produces the same recording. Nothing to install inside the agent, nothing to trust about its own report.

04Proof

Three questions.Three different tools.

TestSprite tells you the PR is red. Ghost tells you which prompt turned it red. They answer different questions, so they work together.

CI, TestSprite and Ghost compared by the question each one answers
	CI / types	TestSprite	Ghost
The question	Does it compile?	Does the PR break the app?	Which turn broke it, and how do I undo it?
Where it sits	At the merge	At the merge, against the deployed URL	Behind the maker, on every agent turn
What you get back	Pass or fail	A red or green suite	The blamed turn and a one-click rollback

CI / types
Does it compile?
Where
At the merge
Returns
Pass or fail
TestSprite
Does the PR break the app?
Where
At the merge, against the deployed URL
Returns
A red or green suite
Ghost
Which turn broke it, and how do I undo it?
Where
Behind the maker, on every agent turn
Returns
The blamed turn and a one-click rollback

TestSprite guards this repo

Required check

13/13

backend tests green after PR #2

A TestSprite 3.0 project with 13 backend tests
Runs against the deployed URL on every pull request
Required check: a red suite blocks the merge

The suite in action

PR #1Suite red
Shipped a BOLA hole: one owner could read another owner's invoice.
PR #213/13 green
Fixed the cross-owner read. The suite went green, 13 of 13.

05Quickstart

Three commands.One recording.

Install, start your app, and point the recorder at its health endpoint. Then open the playground and replay what happened.

01npm installInstall dependencies.
02npm run devStart the app on port 3000.
03npm run recordRecord a session against the app’s health endpoint.

/: Landing
/playground: Timeline, diff, metrics, 3D city, copy fix prompt

ghost · zsh

npm install
npm run dev
npm run record -- --repo . --health http://localhost:3000/api/health --duration 60

Then open the playground

TestSprite tells you the PR is red.

Ghost tells youwhich promptturned it red.

Open playground

Turn	Sample	Change since previous sample
t1	1	baseline
t1	2	0 MB
t2	1	0 MB
t2	2	0 MB
t3	1	0 MB
t3	2	0 MB
t4	1	38 MB
t4	2	38 MB
t5	1	38 MB
t5	2	38 MB
t6	1	0 MB
t6	2	0 MB
t7	1	0 MB
t7	2	0 MB

See what the AI builds.Know when it breaks it.

AI moved the bottleneckfrom writing codeto understanding it.

Does it compile?

Does the PR break the app?

Which turn broke it, and how do I undo it?

Record. Attribute.Reverse.

Every change, on the record.

The turn that did it, named.

Undo exactly that turn.

Every change, on the record.

The turn that did it, named.

Undo exactly that turn.

Everything the agent did,on one screen.

Timeline

Diff that did it

Memory

Ghost watches the filesystem and the process, not the agent.

Three questions.Three different tools.

Three commands.One recording.

Ghost tells youwhich promptturned it red.