Four review specialists read every pull request in parallel — logic, security, performance, and tests. A synthesizer dedupes what they found into one inline review.
Every finding is grounded in the actual repository and arrives with the proof that earned it: the exact lines, a reproduction path, or the test that fails. Findings that cannot earn that proof get dropped before they reach your team.
Live demo
Paste any public GitHub pull request. Sigilix reviews it for real, right now — and shows you exactly what it found and why it believes it.
No account. No install. No repository permissions.
Specialists
Four narrow questions surface what a single broad pass can blur. Each specialist reads the same diff with a different question, and the synthesizer merges the answers, drops the duplicates, and posts a single review instead of four.
Traces the changed behavior path through the surrounding implementation instead of judging the diff as isolated text: broken invariants, missed edge cases, state that no longer survives the round trip.
Reads the diff for the openings it creates: injection paths, authorization gaps, secrets in the wrong place, and unsafe defaults that only look safe in the happy path.
Checks what the change costs at runtime: queries that multiply per row, allocation inside loops, and work that quietly moved onto the request path.
Judges the test surface against the new behavior: deleted assertions, weakened fixtures, and the failure mode the change introduces but no test covers.
Evidence
The cost of a review tool is not the findings it posts. It is the attention it spends. Sigilix treats every comment as a claim that needs proof.
A comment arrives with the exact lines, the reproduction path, or the test that demonstrates the problem. If the proof is missing, the finding says so instead of pretending.
When a finding can be checked by executing code, Sigilix executes it and attaches the output. The review separates what was proven from what the model inferred.
Each finding is scored on whether it earned confidence from the repository. Comments that cannot clear the bar are discarded before they cost a maintainer's attention.
Findings post inline on the diff, where the conversation about the code already happens, instead of piling into a summary the author has to map back by hand.
Intent
A diff is only right relative to what it was supposed to do. Sigilix reads the linked issue and the PR description, then reviews the change against that intent.
Input
Review by vibes
The diff, read as isolated text.
Review by intent
The diff, the linked ticket, the surrounding implementation, and prior review memory.
Question
Review by vibes
Does this code look wrong in general?
Review by intent
Does this change do what the ticket asked, without breaking what is around it?
Output
Review by vibes
A list of plausible-sounding observations.
Review by intent
One deduplicated inline review where each finding keeps its evidence.
In public
GitHub PR #26 · Sigilix/runner
Not a staged demo. A real pull request on one of our own public repositories, reviewed by Sigilix and merged. The inline comments, the evidence attached to each finding, and the author's responses are all still there.
For how we measure review quality — recall, precision, and false positives with the fixture boundaries visible — read the case study.
FAQ
Sigilix installs as a GitHub App and reviews every pull request with four specialists in parallel — logic, security, performance, and tests — grounded in your codebase. A synthesizer dedupes them into one inline review posted on the exact lines, where each finding carries the evidence that proves it: the lines, a reproduction path, or an executed check.
Yes. Once installed, Sigilix reviews pull requests automatically as they open and update — no manual trigger. Findings that cannot earn confidence from the repository are dropped before they post, so the automated review stays low-noise instead of flooding the PR with comments.
That is the core of Sigilix review. Every finding is anchored to exact lines and carries its proof — a reproduction path or an executed check whose output is attached — rather than an unbacked opinion. Findings that cannot clear a believability bar are discarded, not softened into hedges.
Yes. Paste a public GitHub pull request into the live demo on this page and watch a real Sigilix review run on it — four specialists, findings with evidence, and the file and line each lands on. No account, no install.
Put a reviewer on every pull request that has to prove its findings before your team reads them.