Architecture and security review, checked against the code

Know what you are taking on before you take it on.

Repo Review reads a GitHub repository and tells you how healthy it is: a verdict, seven ratings out of ten, the findings that matter and what to fix first. Every claim is checked against the source, and by default an architect reviews the report before you see it.

Get a free Snapshot How private repositories stay private

Read-only GitHub AppHosted in SydneyReviewed by a person

What you get

Verdicts, not descriptions, and every one of them traceable to the code

How healthy is it, what is wrong, and what do we do about it.

Written from a structural graph of your repository built by real parsers, so every module, function and link a report names is one that exists in your source.

A verdict and seven ratings

Sound, Adequate, At risk or Critical, and engineering standards, security, configuration, data, resilience, availability and performance, each out of ten with its reasons.

Findings and a fix list

A risk register rated High, Medium and Low, each finding tied to the file and function it is about, and a remediation roadmap in the order to do it.

Architecture and evidence

An architecture document with diagrams generated from the code, an inventory of every outbound call the code makes, and a searchable index of every function and class with its exact source.

  • Same code, same scores: rerun unchanged code and the ratings do not move
  • Your README corrected against the code, as a patch you can review
  • Self-contained HTML that opens anywhere, offline

How it works

Parsers decide what the code contains. Checks hold the writing to it. By default, a person signs it off.

  1. Graph

    A structural pass over every source file maps modules, classes, functions and calls. No model involved.

    01
  2. Exact source

    Every function gets its start and end line from the language's own parser, never a guess. Anything it cannot range is labelled approximate.

    02
  3. Checked writing

    The report is written against the graph, then 20 to 50 automated checks: every named function must exist, every rating must carry its reasons, every link must resolve.

    03
  4. Architect review

    By default, a person reads every report before it is released. You get an email with links when it is.

    04
01

Graph

The first pass reads the repository the way a compiler would, not the way a language model would. A parser walks every committed source file and records what is there: modules, classes, functions, and which one calls or imports which.

  • No model and no API key. It is a structural pass, so the same code always gives the same graph.
  • Only what git tracks. Uncommitted experiments, notes and vendored libraries are left out, so the report describes the code you ship.
  • JavaScript, TypeScript, Python, PHP, Go, Rust, Java, C#, Ruby, Swift, Kotlin, Terraform and CloudFormation.
  • Everything after this is written from the graph, so a report can only talk about code that exists.
api/routes.pystripe_webhook()payments/stripe.pyhandle_webhook()orders/models.pyOrder.mark_paid()payments/stripe.pyverify_signature()callscallsnever callsno callersdefined, never used

Part of the graph for the booking API in the Snapshot above Illustrative

payments/stripe.py
  handle_webhook()     called by  api/routes.py · stripe_webhook()
                       calls      orders/models.py · Order.mark_paid()
  verify_signature()   called by  nothing

verify_signature() exists, but nothing calls it. That line of the graph is where the High finding “Webhook signatures are never verified” starts; the report then shows the source to prove it.

02

Exact source

When a report names a function, you can open it and read its code. For that, every function and class needs its exact first and last line, and those come from the language's own parser.

  • @babel/parser for JavaScript and TypeScript, Python's own ast module, PHP's own tokenizer, and tree-sitter grammars for Go, Rust, Java, C#, Ruby, Swift and Terraform.
  • Never a guess. No counting blank lines or looking for the next function: a wrong end line would put the wrong code next to a finding.
  • Where a body cannot be closed with certainty (a Kotlin expression body, a PHP arrow function, or a language with no extractor such as C++), the symbol is labelled approximate and shown without a body.
4142def handle_webhook(request):43 payload = request.body.. ...6162 event = json.loads(payload).. ...87 return HttpResponse(status=200)88start 42from Python's own parsera guess would stop here:line 61 is blank, not the endend 87from the same parser

Three rows from a symbol index Illustrative

handle_webhook()    payments/stripe.py   42-87     exact
Order.mark_paid()   orders/models.py     118-131   exact
formatTotal         billing/Money.kt     line 12   approximate

The symbol index lists every function and class like this, searchable, and each one opens on its source.

03

Checked writing

A language model writes the report, but it writes against the graph and the source, and then automated checks read what it wrote. A section that fails is sent back with the exact failure, rewritten and checked again.

  • Every function, class and file the report names must exist in the source.
  • Every rating must carry its reasons and its weakest point.
  • Every link must resolve, between the documents and into the code.
  • Unchanged code gets the same scores: a rerun of the same commit is held to the ratings it was first given.
  • Between 20 and 50 checks, depending on the report. A report with a failed check is never released without a person looking at it.
Graph and sourcewhat the code isWrite a sectionone section at a time20 to 50 checksnames, ratings, linksReportassembledpassfail: sent back with the exact failure

Checks on one section of a first draft Illustrative

fail  named symbols exist    "validate_stripe_event()" is not in the source
      section 6 sent back with that line, rewritten
pass  named symbols exist    41 of 41
pass  ratings carry reasons  7 of 7
pass  links resolve          64 of 64
pass  ratings match the fixed scores

The function the draft invented never reaches you: the check caught it, and the rewrite named the one that is really there.

04

Architect review

Before a report reaches you, a person with architecture experience reads the review copy against the code, with the check results beside it.

  • They release it, or withhold it with a reason and have it put right.
  • A report with any failed check always waits for a person.
  • When it is released you get an email with a link to each document. Each one opens in the browser and can be downloaded to keep, and your reports are also in the customer portal.
Checked reportevery check runArchitect reads itreview copy, with checksReleased to youemail and portalWithheld with a reasonthen put rightreleasewithholdback through the checks

The email you get Illustrative

Subject: booking-api: repository documentation

Your repository documentation is ready. Each document opens
in the browser and can be downloaded to keep.

  Snapshot                How healthy the repository is
  Architecture document   How the system is designed
  Symbol index            Every function and class, with its source

Questions about a report go to support@reporeview.com.au, and a person answers.

Two steps

Start with a free Snapshot. Go to the Full Audit when you want the whole story.

1. Snapshot

Free one per repository

How healthy is it? A quick read with the verdict, the seven ratings and the three findings that matter most. About five minutes, then review.

  • Health verdict and seven ratings, each with its weakest point
  • The three findings that matter most, named
  • What the Full Audit of this repository would cover
  • Symbol index: every claim one click from its source
  • Public GitHub repositories: sign up with just an email
Get a free Snapshot

Prices in Australian dollars, including GST. Very large repositories are quoted before anything runs. Private repositories: see how access works.

Examples

Real reports on public open-source projects, released as they were delivered. Open the Snapshot, or the Full Audit's architecture document and symbol index

  1. PythonAdequate

    Flask

    The Python web framework.

    pallets/flask · main · 49943b5

    Snapshot and Full Audit

  2. PythonAdequate

    browser-use

    Lets AI agents drive a web browser.

    browser-use/browser-use · main · 4b5b1b0

    Snapshot and Full Audit

  3. TypeScriptAdequate

    Excalidraw

    The open-source virtual whiteboard.

    excalidraw/excalidraw · master · 82d5847

    Snapshot and Full Audit

  4. PythonAdequate

    crewAI

    A framework for teams of AI agents.

    crewAIInc/crewAI · main · 2817a95

    Snapshot and Full Audit

  5. SwiftSound

    Alamofire

    HTTP networking for Swift.

    Alamofire/Alamofire · master · b3e74e8

    Snapshot

Adequate

Flask

The Python web framework. pallets/flask (opens in a new tab), main at commit 49943b5, Python. The verdict is the Full Audit's.

Snapshot

How healthy the repository is: the verdict, all seven ratings with the weakest point of each, and the three findings that matter most.

Open the Snapshot (opens in a new tab)

Full Audit

How the system is designed, with diagrams generated from the code, and every function and class with its exact source.

Architecture document (opens in a new tab) Symbol index (opens in a new tab)

Each report opens in a new tab, exactly as it was delivered. The technical assessment is shared with the project's maintainers first, so it is not published here.

Adequate

browser-use

Lets AI agents drive a web browser. browser-use/browser-use (opens in a new tab), main at commit 4b5b1b0, Python. The verdict is the Full Audit's.

Snapshot

How healthy the repository is: the verdict, all seven ratings with the weakest point of each, and the three findings that matter most.

Open the Snapshot (opens in a new tab)

Full Audit

How the system is designed, with diagrams generated from the code, and every function and class with its exact source.

Architecture document (opens in a new tab) Symbol index (opens in a new tab)

Each report opens in a new tab, exactly as it was delivered. The technical assessment is shared with the project's maintainers first, so it is not published here.

Adequate

Excalidraw

The open-source virtual whiteboard. excalidraw/excalidraw (opens in a new tab), master at commit 82d5847, TypeScript. The verdict is the Full Audit's.

Snapshot

How healthy the repository is: the verdict, all seven ratings with the weakest point of each, and the three findings that matter most.

Open the Snapshot (opens in a new tab)

Full Audit

How the system is designed, with diagrams generated from the code, and every function and class with its exact source.

Architecture document (opens in a new tab) Symbol index (opens in a new tab)

Each report opens in a new tab, exactly as it was delivered. The technical assessment is shared with the project's maintainers first, so it is not published here.

Adequate

crewAI

A framework for teams of AI agents. crewAIInc/crewAI (opens in a new tab), main, Python: the Snapshot at commit 364b879, the Full Audit a few hours later at 2817a95. The verdict is the Full Audit's.

Snapshot

How healthy the repository is: the verdict, all seven ratings with the weakest point of each, and the three findings that matter most.

Open the Snapshot (opens in a new tab)

Full Audit

How the system is designed, with diagrams generated from the code, and every function and class with its exact source.

Architecture document (opens in a new tab) Symbol index (opens in a new tab)

Each report opens in a new tab, exactly as it was delivered. The technical assessment is shared with the project's maintainers first, so it is not published here.

Sound

Alamofire

HTTP networking for Swift. Alamofire/Alamofire (opens in a new tab), master at commit b3e74e8, Swift. The verdict is the Snapshot's.

Snapshot

How healthy the repository is: the verdict, all seven ratings with the weakest point of each, and the three findings that matter most.

Open the Snapshot (opens in a new tab)

Full Audit

This example is a Snapshot only. Meanwhile, every function and class with its exact source:

Symbol index (opens in a new tab)

Each report opens in a new tab, exactly as it was delivered. The technical assessment is shared with the project's maintainers first, so it is not published here.

Private repositories

Read through a GitHub App you control, one repository at a time, never with your password or a long-lived token

Three steps, all on your terms

  1. Install the Architecture Automation app

    On GitHub, choose "Only select repositories" and pick the one to review. It asks for two permissions, both read-only: contents and metadata.

  2. We issue a key for that repository

    Email support@reporeview.com.au. The key works for that one repository and branch and nothing else, and expires.

  3. Start the run from the portal

    Paste the key into the portal and press Start. The report comes back to you by email once it is released, after an architect has reviewed it unless you have agreed otherwise with us.

View the app on GitHub
  • Read-only, and only what you chose

    The app cannot push, open issues, read your organisation or act as any person. It sees only the repositories you selected.

  • One repository, one hour, one run

    When your run starts, GitHub issues a token for that single repository, valid for at most an hour. It is used for one shallow clone of one branch and deleted straight after.

  • Revoke at any time, in your own settings

    Uninstall the app, or remove the repository from it, and the next request is refused. You never have to ask us.

  • The app's own key is locked down

    It has its own encryption key, only the parts of the service that start runs can use it, never the containers that handle your code, and any other use alerts us at once.

  • Nobody can sign up for your repository

    Free sign-up takes public repositories only. A private repository is only ever set up with you directly, so naming it gets a stranger nothing.

Your code is read, reviewed and deleted.

Everything runs in our own AWS account in Sydney. Your code is copied once, analysed in a container that is destroyed afterwards, and the copy is deleted the next day, whatever happens to the run.

Read access, nothing more

No production systems, no credentials, no secrets of yours. Nothing is ever committed or pushed to your repository.

Read our data handling document (opens in a new tab)
Deleted on a schedule

The copy of your code after one day, the reports after thirty, or both at once whenever you ask. Retention is enforced by configuration, not by memory.

A sealed analysis

The analysis can reach only the model and the AWS services that run it, through a filter that logs every connection. During the alpha, your code is read and parsed, never run.

Keys that cannot wander

Each key is bound to one repository and branch, stored only as a hash, handed to you once, and expires. Rotate it from the portal whenever you like.

Expiring links

Reports are reached through links that expire after seven days, and the pages are sandboxed so nothing in them can reach your account.

One model provider

The parts of your code the analysis needs are sent to Anthropic's Claude. Nothing is sent anywhere else.

An audit trail

Every key issued or revoked, every run started, released or deleted, and every link handed out is logged and kept for a year.

Questions

Before you ask

1Who is this for?

Anyone who has to decide about a codebase they did not write: buying a software business, investing in one, inheriting a system, or signing off an agency's or vendor's work. It answers in an hour what otherwise takes weeks of reading or a due-diligence engagement.

2How do I get a free Snapshot?

Open the portal, enter your email and a public GitHub repository, and we email you a key for one Snapshot of it. Paste the key back into the portal and press Start. One free Snapshot per repository, and three per email address.

3My repository is private. What do you need?

Only that you install the read-only Architecture Automation GitHub App on that repository and email support@reporeview.com.au. No password, no personal access token, no deploy key. See how access works. GitLab and Bitbucket are possible too: ask us.

4What stops the AI making things up?

The structure comes from parsers, not the model, and every report has to pass 20 to 50 automated checks against it: every function it names must exist, every rating must carry its reasons, every link must resolve. Then an architect reads it before it is released. And the same code always gets the same scores, so a score that moves means the code changed.

5Which languages do you cover?

Exact source ranges for JavaScript and TypeScript, Python, PHP, Go, Rust, Java, C#, Ruby, Swift, Terraform and CloudFormation, and Kotlin for most functions. Other languages get the full report with every symbol labelled approximate, never given a guessed body.

6How long does it take?

A Snapshot runs in about five minutes and a Full Audit in about an hour. Then an architect reviews it, and you get an email with links as soon as it is released.

7What do you keep, and for how long?

The copy of your code for one day, the reports for thirty, and a few numbers (your scores, never code) so a rerun of unchanged code scores the same. You can delete a run and everything it left from the portal at any time.