Pull-request reviews you can install, inspect, and verify.
LlamaPReview reviews the exact pull-request head using bounded, sourced evidence and deterministic publication logic. The active Webhook and Pipeline source that powers the hosted service is public.
The hosted App is free for public repositories and does not review private repositories.
Open source, actively
This is not a demo or a mirror. The public repository is the production source.
Read the reviewer before you trust it
The Webhook and Pipeline code in JetXu-LLM/LlamaPReview powers the official hosted service under the Apache-2.0 license.
- Inspect how evidence is retrieved, bounded, and tied to an exact commit.
- Trace the code-owned safety, accounting, recovery, and publication path.
- Open an Issue, join a Discussion, or contribute a focused improvement.
Public repositories
One production path from a signed webhook to an exact-head GitHub review.
- Evidence with provenance and honest coverage gaps
- High-confidence inline comments
- Sanitized Mermaid diagrams when useful
- Idempotent exact-head publication
Why it is different
Model judgment stays focused on engineering decisions. Code owns the boundaries that make the result safe to publish.
Evidence with boundaries
Every review uses a bounded set of public, exact-head evidence with provenance. Missing coverage is reported instead of guessed away.
Judgment and publication, separated
DeepSeek supplies engineering judgment. Deterministic code decides the public shape, safety, placement, and publication identity.
Pinned to the exact head
Admission, rereads, and publication bind to the same pull-request commit. Recovery reuses the prepared request instead of posting blindly.
How a review works
One production path, from a signed event to a review on the same verified head.
Verify and admit
Verify the webhook signature, enforce the public-only boundary, and pin the exact pull-request head.
Retrieve evidence
Gather bounded repository evidence with provenance, coverage, and explicit gaps.
Judge and present
Use model judgment for causal engineering risk, then compress it into owner actions and placement requests.
Project and publish
Sanitize, place, account, recover, and publish idempotently to the exact head that was reviewed.
Privacy and trust, plainly
The hosted service reviews public repositories only. Its active source and data boundaries are inspectable.
Private events stop early
After signature verification and minimum visibility parsing, a private event is acknowledged and discarded before product storage, provider processing, or GitHub mutation.
Public-run retention is stated
Run records, artifacts, provider traces, and logs follow documented retention windows. Historical private records were not deleted.
Provider processing is disclosed
Selected public evidence, prompts, and output are sent to DeepSeek. Read the full privacy and retention details.
Join the project
Inspect the source, report a reproducible problem, or help sharpen what useful engineering judgment looks like.