Free hosted GitHub App · Public repositories only · Apache-2.0 source
LlamaPReview reviews a pull request at its exact head commit, explains findings against the code, and reports known gaps in the evidence.
Do not merge until the nap instruction names the runnable driver path, matching the activation scheme this PR ships; currently the printed instruction tells agents to run summem, not .summem/summem.
Exact-head CI remains unresolved (1 pending); no CI-dependent merge-safety claim is made.
In this public review, LlamaPReview flagged a mismatch between the command an agent was told to run and the executable path shipped by the PR. It also recorded unresolved CI separately.
One historical example, not a measure of overall accuracy. Read the full review or inspect another example.
A substantive review summary contains findings, evidence and suggested next steps. Zero or more inline comments appear where findings can be placed safely. An optional diagram can help explain the change. Known evidence gaps remain visible. Output contract →
The app does not review every push. Reviews use GitHub's COMMENT state, not APPROVE or REQUEST_CHANGES, and are not a required merge check.
Free hosted capacity: up to 3 admitted review attempts per repository per UTC day, subject to shared daily capacity and available model budget. These are admission limits, not a guaranteed allocation or a promise of a completed review. Over-capacity requests are skipped, not queued.
Reviews can miss issues or be wrong; they supplement your tests and review process. Evidence retrieval is bounded, and missing coverage remains a gap.
The model supplies engineering judgment. Code enforces evidence boundaries, output validation, safe placement and publication.
This repository contains the Webhook and Pipeline source used by the hosted service, under Apache-2.0. Inspect the evidence retrieval, output validation and publication path, or read the architecture documentation.
Run the public-repository review pipeline in your own AWS account. Supply your own GitHub App and model-provider credentials, and pay your own AWS and provider costs.
The reference Terraform stack disables hosted quota counters by default and leaves one-time head succession enabled. New source installations default to an OpenRouter GPT-6 Luna max trial with documented qualification limits. The existing DeepSeek profile remains selectable. Direct OpenAI is also an explicit option with a stage profile and verified budget policy. These source choices are separate from the hosted service, which uses the OpenAI API directly.
New private-repository events stop after signature verification and minimum visibility parsing, before durable product storage, provider calls or GitHub product API calls. For eligible public requests, selected evidence and prompts are sent directly to OpenAI; generated responses can be included in subsequent review-stage requests.
The Privacy Policy separates OpenAI processing from LlamaPReview’s AWS storage and dates the last verified retention settings. Neither prompt caching nor store=false is a zero-retention guarantee. Read the Privacy Policy and security model.
Found a reproducible problem or have a public review worth discussing? Open an Issue, join Discussions, or read the contributing guide.
Install for a public repository