00local receipt for AI code changes

The merge button still needs a human.

DiffReceipt turns an AI-assisted coding run into a clear receipt: what was requested, what changed, what was checked, and where a human still needs to decide.

Open demo runs Memory example

For teams already using Cursor, Claude Code, Codex, Copilot, or similar tools. Early stage. Runs locally as a CLI. Demo data below.

example receiptneeds decision

one AI change1 file outside scope

RequestAdd CSV report exportclear

Files3 in scope, 1 nearbyreview

Checkstests passed, no migrationsclear

Decisionapprove session_policy.pyreview

example receiptAdd CSV report export

Requested: Add a CSV export button to the finance summary page.
Allowed to change: reports/export/*reports/ui/export-actions.tsxconfig/export.yaml
Actually changed: reports/export/csv_exporter.pyreports/ui/export-actions.tsxconfig/export.yamlAlso changed reports/auth/session_policy.py outside the requested scope.
Checks: export tests 12/12 passed
lint clean
no migrations
manual CSV check with Cyrillic text passed
Risks / limits: the out-of-scope change affects the session lifetime for file downloads
that behavior is not covered by the export tests
Needs you: Approve the session_policy.py change before merge.

Add CSV report exportneeds decision Fix login timeoutstopped by memory Add a clear empty import stateready Fix invoice sortingchecking Add file size checkqueued Update payment error copyready

01how it works

Not another dashboard. A receipt that leads to a decision.

The first screen shows one finished piece of work. If the reviewer needs more depth, they can open the run list or inspect the receipt for a specific change.

Requested

The original task stays at the top, so the reviewer does not reconstruct context from chat history.

Changed

Allowed scope and real touched files sit side by side. Anything outside the request is visible.

Checked

Tests, lint, and manual checks are separated from the agent promise.

Human decision

The receipt ends with one concrete action before merge.

02review format

A ready receipt first, all runs second.

The reviewer sees the useful summary first. The table stays as the second layer, not the first screen.

Open demo runs

Normal review	Find the task, diff, checks, and risk in different places.
DiffReceipt	One filled receipt with request, files, checks, and decision.
Memory	A known past failure is raised before a similar change moves forward.

03known-mistake memory

If a similar failure happened before, it appears in the receipt.

The example below shows a run where a known rule stops a risky change from passing silently.

The team wrote the rule after a previous failure. DiffReceipt raised it, stopped automatic acceptance, and handed the decision to a human.

Memory example

known-mistake memoryFix login timeout

Requested: Increase the allowed login wait time for users on slow connections.
Allowed to change: auth/login/*config/session.yaml
Actually changed: auth/login/timeout_policy.pyconfig/session.yaml
Checks: the old failure scenario was repeated
the regression was stopped before merge
the touched files stayed inside the allowed scope
Risks / limits: a similar change previously broke forced sign-out
the "forgotten open browser" scenario still needs a manual check
Needs you: Approve the new wait time after the manual safety check.

04FAQ

Questions before the first review.

Does it change code?

No. This demo is read-only: it packages a receipt and does not write back to the repository.

Why open with one receipt?

Because a reviewer first needs a finished example, not a table with no context.

Is the demo data real?

It is a demonstration set with human-readable task names. The receipt format is the same as the working version.

What does this NOT do?

DiffReceipt is not a sandbox and not a replacement for human review. The hard boundary is still commit and merge: a person decides what gets accepted.

What should I open next?

Open demo runs, choose a receipt, and check the decision that still needs a human.

for the person responsible for someone else's AI code

Start with one repository and one receipt.

It's early days. We'll set this up together with one or two teams: one repository, one workflow, a few days. Then you decide whether it actually helps your review.

Open demo runs Memory example

Contact: [email protected]

DiffReceipt makes AI changes visible and reviewable before they are accepted. It does not replace the reviewer.

The merge button still needs a human.