Incremental code review, in ink. A reMarkable code review tool.

Prerequisites

Inkrement relies on having a couple of tools installed to work correctly. These are:

1. gh: Inkrement currently only works with GitHub, as it relies on the gh CLI to interact with pull requests assigned to your user and post back reviews.
2. claude: To convert handwritten comments to text as well as contextualize placement of those comments in the surrounding code (i.e. comment x was placed near line y), the PDFs are parsed using Anthropic's OCR tools.

Quick Start

Install

Right now the best way to install is through cargo:

cargo install --git https://github.com/duncanam/inkrement.git

Once the tool matures we'll add GitHub actions to publish binaries.

Use

Connect your reMarkable via USB. Under Settings > Storage, ensure USB communication is enabled. Then, run inkrement to launch the TUI.

Keybindings

The keybindings, for the most part, are described on the TUI. Here is a full description:

• [tab]: swap between PR selection and review publish tabs
• [space]: toggle a specific highlighted entry
• [a]: select all eligible entries
• [enter]: perform operation
  • PR selection tab: converts and uploads all selected PRs as review documents to the reMarkable.
  • Review tab: downloads review documents, trims off reference-only pages, performs OCR, and publishes review.
• [r]: refresh page
  • PR selection tab: fetches new PRs and re-scans the reMarkable for pushed PRs to review
  • Review tab: re-scans the reMarkable for review documents that have been uploaded and are eligible for publishing a review.
• [q]: quit Inkrement
• [Arrow Up]/[Arrow Down]/[j]/[k]: Navigation up/down for selecting items

Screenshots

PR Selection Tab

Review Publishing Tab

Motivation

In the world of AI, I've found myself reviewing more and more pull requests. I tried converting a code outline plan to PDF for my reMarkable Paper Pro, and it was a very pleasant experience. I had increased stamina for review due to an overall reduced fatigue. Additionally, it allowed me to be more mindful in the moment during my reviews, excusing myself from a notification-filled digital space and focus on the review to execute it more efficiently. Beyond the material benefits mentioned, it's also fun and engaging to literally draw on code.

In a rapidly-changing world of software and AI, I like finding ways to cultivate our humanity.

How it works

1. Inkrement will fetch your pending pull request review requests.
2. You get to then decide which PRs you'd like to pull down to review on your reMarkable.
3. Once you've selected your desired PRs to review, Inkrement parses the diff using the unidiff crate to split the work into tractable hunks.
4. These hunks are then rendered into a PDF using the native Typst Rust API. Full source is also provided for reference and is hyperlinked near the diffs.
5. The PDF is then transferred to your reMarkable over USB connection. Metadata is embedded in the filename.
6. Perform your markup and review offline.
7. Reconnect your reMarkable. Based on the metadata filename, you'll have the option of downloading specific reviews.
8. Inkrement will then strip off the reference full source code pages, leaving only your marked-up diff pages.
9. Once the stripped PDF is produced, it's fed to Claude to interpret your handwriting and positioning of your handwriting.
10. Inkrement will then take the resultant parsed handwriting and position data back from Claude and submit your PR review.

Roadmap and Outstanding Work

Here are some things I'd like to fix:

• overall tech debt and cleanup - I'd say maybe only 30-40% of the repo is idiomatic Rust that I've heavily combed through and corrected into proper Rusty patterns. The rest of the code is pathfinding "fast and loose" code to explore if stuff is possible.
• modules are a mess right now and need restructuring. tui.rs is a great example.
• only supports Claude as the OCR/contextualization backend. We should support more, including FOSS models.
• only supports GitHub through gh. Not all users use GitHub as a service.
• support iPads
• PDF splitting step is horrendously slow, because lopdf needs to parse the PDF and re-render it

LLM and AI Policy

LLMs have weaseled their ways into the workflow of many software engineers out there in order to maintain any semblance of competitiveness in today's world- it's unreasonable to outright ban all AI contribution. That said, here are some guidelines:

• All lines of code must be accounted for
• All lines of code must have human eyes on them
• Architecture decisions should be signed off by a human

Bonus points if you're still writing code the artisanal, old-fashioned way.

License

GPLv3. Some dependents are not (gh, claude). Fonts include their licenses, see those.