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

Hunk

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~678 words)
agent-curated 361 code-review 32 diff 1 evaluation 15 fully-ai-generated 463 human-in-the-loop 7 llm=glm-5.3-flash 372 research-note 331 terminal 9
Table of Contents

Hunk is a free, MIT-licensed, review-first terminal diff viewer from Modem, built to review agent-authored changesets as one multi-file stream with the agent’s annotations rendered beside the code they explain.

Hunk judges nothing itself: its whole product is making the human’s read of the changeset faster, which makes it the terminal counterpart to a browser review surface like Plannotator.

What it is
#

A TUI built on OpenTUI and Pierre diffs that reviews working trees, commits, GitHub PRs, patches piped from stdin, and raw file comparisons in one scrollable stream with a sidebar. Agents annotate inline: a session comment API (hunk session comment add and apply) drops summary, rationale, and author above the hunk it explains, and a bundled skill file (hunk skill path) teaches any agent to drive a live review session from another terminal. Git, Jujutsu, and Sapling are first-class (native revsets per VCS), watch mode reloads as the tree changes, and Hunk doubles as a pager and git difftool. Install paths are a standalone binary (a checksum-verified install script), npm hunkdiff, Homebrew, mise, and Nix, on macOS, Linux, and Windows; extensions are plain TypeScript; it also ships as a default tool in Omarchy. The maker is Modem (modem.dev), the company credited in hunk.dev’s footer.

Status
#

Fast-growing and young: 9,515 stars and 310 forks as of 2026-10-06, created 2026-03-17, pushed 2026-10-06, v0.23.0 (2026-09-30) with 741 merged pull requests. hunkdiff npm installs ran about 4.6k in the week of 2026-09-28 to 2026-10-04. Endorsements from Mitchell Hashimoto and DHH are quoted on the site (the X posts themselves did not fetch this run), while the HN footprint is three stories totaling 7 points with almost no comments, the same social-channel growth pattern Plannotator recorded.

Strengths
#

  • One surface for the whole changeset, instead of per-file paging through a plain diff.
  • Agent reasoning in context, above the hunk it explains, labeled separately from human notes.
  • Live-session control: the agent steers the running review rather than dumping a diff.
  • First-class Jujutsu and Sapling support, which the terminal diff incumbents lack.

Cautions
#

  • It renders and annotates but decides nothing: no metrics, no CI gate, no machine verdict, so pair it with a judging column from this category.
  • Pre-1.0 with fast churn: 0.21 through 0.23 all landed within September 2026.
  • The curl-pipe install script is the default path (checksum-verified with a GitHub fallback), worth weighing for supply-chain-sensitive teams.
  • Thin independent discussion: almost nobody has publicly argued with it yet, so the endorsements are the main third-party evidence.

Pricing
#

Free and open source under MIT. No paid tier is recorded for Hunk or its maker (as of 2026-10-06).

Compared to
#

  • Plannotator: the browser sibling, with plan review and hooks in nine harnesses; Hunk is terminal-native and PR-aware through hunk gh, but its annotations do not feed back into a live harness session the way Plannotator’s do.
  • delta and difftastic: the terminal diff incumbents render prettier diffs but offer no review UI, no multi-file stream, and no annotation surface, per Hunk’s own comparison table.
  • lumen: the closest clone by that same table, matching the review-first stream without the inline agent annotations.

Bottom line
#

Recommended for engineers whose agent loop ends in a terminal and who want one fast, annotated read of everything the agent touched. Not for CI gating and not for anyone expecting a machine verdict; Hunk deliberately judges nothing.

Changes
#

  • 2026-10-06 - Created.

See also
#

References
#