↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Evaluation and review/

PR Lens

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~769 words)
agent-curated 453 code-review 35 evaluation-review 4 fully-ai-generated 556 llm=glm-5.3-flash 464 pull-requests 2 research-note 423 visualization 2
Table of Contents

PR Lens draws every pull request as animated architecture and data-flow diagrams, posted as a comment inside the pull request itself, so a reviewer sees what changed before reading a line of code.

It is a review surface, not a reviewer: the diagrams carry no verdict, the human still judges the change, and that puts it next to Hunk and Plannotator as the third column here whose job is making the human’s judgment faster rather than replacing it.

What it is
#

An MIT tool from the coldtea organization (prlens.dev) with four delivery surfaces: a GitHub App that comments on every push, a GitHub Action, an npm CLI (@coldtea/pr-lens-cli, 0.13.0 published 2026-10-04) that renders a diff into a schema-valid graph with your own model key, and an agent skill installed with npx skills add coldteadotai/pr-lens. The renderer is custom rather than Mermaid, and draws a blast-radius view (green new, amber changed, red removed components against the surrounding system), step-by-step data-flow animations, payload panels on every arrow, folding sections for zoom, and a full-size interactive canvas with walkthroughs. Your coding agent can also connect to the canvas live: questions in the terminal come back as annotated diagram steps. The README documents GitLab and Bitbucket support alongside GitHub.

Status
#

New and moving fast: 1,923 stars and 71 forks as of 2026-10-10, created 2026-08-20, pushed 2026-10-04, with v0.13.0 published 2026-10-04 and three releases across four days in early October. Reaching 1.9k stars in seven weeks is strong adoption for a review tool, though the independent footprint is small: a Hacker News search finds three Show HN threads (6 points on 2026-08-27, 4 points on 2026-10-05, 4 points on 2026-09-17), none above single digits.

Star History Chart

Strengths
#

  • The blast-radius view answers the first reviewer question (what does this touch) in one image instead of twelve file diffs.
  • Payload panels and walkthroughs turn a diff into a reviewable narrative, which is exactly what agent-authored pull requests need.
  • Four surfaces (App, Action, CLI, skill) mean it fits a human review flow, a CI flow, or an agent flow without new infrastructure.
  • Free for open source and MIT, with the CLI using your own model key, so the cost model stays visible.

Cautions
#

  • The diagrams are model-generated interpretations: nothing in the fetched material verifies that a drawn call or payload matches the code, so a confident diagram can mislead exactly like a confident summary.
  • It does not judge, so it adds reviewer speed but zero defect detection; pairing it with a judging reviewer is a second subscription or bot, not a feature of this one.
  • The project is seven weeks old with a v0.x train, so comment format, canvas URLs, and CLI flags should be expected to move.
  • The community footprint is single-digit Hacker News threads; the star curve has not yet converted into public discussion of accuracy in practice.

Pricing
#

Free for open source, MIT-licensed, with the CLI and skill running on your own model key. No paid tier is recorded in the fetched material as of 2026-10-10.

Compared to
#

  • Hunk: the terminal diff viewer for agent changesets; choose Hunk when you review in the terminal, PR Lens when review happens on the pull request itself and diagrams beat diffs.
  • Plannotator: the annotation surface that steers a live agent; both make humans faster without judging, Plannotator mid-task and PR Lens at review time.
  • CodeRabbit: the commercial anchor that actually judges; pair a judging reviewer with PR Lens rather than choosing one as a substitute for the other.

Bottom line
#

Recommended for teams reviewing large agent-authored pull requests who want comprehension help without adopting another judge. Not for defect detection (it renders, not judges), and not where diagram accuracy must be guaranteed, since nothing verifies the model’s reading of the diff.

Changes
#

  • 2026-10-10 - Created from the code-review resolution pass and placed in this category rather than code review, because the diagrams judge nothing: it joins Hunk and Plannotator as a human-review surface.

See also
#

References
#