Intro: From Writing Code to Reviewing It

Over the past year, AI agents writing and modifying code has stopped being a novelty. Claude Code, DeerFlow, and Ruflo have settled into daily workflows, and the older mental image of a developer hammering out lines on a keyboard is fading. A large slice of working hours now goes into reading, judging, accepting, or rolling back patches that an agent produced.

The tool that prompted this post is Hunk, recently surfaced on GeekNews. It is an interactive terminal diff viewer, explicitly positioned as "a way to review changes made by AI agents." This article is not a translation of the README. It is my own notes after browsing the repository and thinking through what the project signals.

The central question is plain — "In an era where agents write code, where is the human role moving, and where does a tool like Hunk fit into that shift?" Following this question, I cover the design of the tool, how it compares to IDE-based diff workflows, and what it means for the Korean development scene. This is the fourth entry in a series that already covered AI coding tools, DeerFlow, and Ruflo.

1. The Structural Weak Spots of Agent Code

Before judging whether a dedicated review tool is justified, it helps to name what is actually wrong with agent-generated code. After roughly a year of daily use across Claude Code, Cursor, and Copilot Agent, my notes converge on three patterns.

1.1 Hallucinations and Subtle Lies

Calls to APIs that do not exist, or functions whose signatures are plausibly close but actually wrong, are still common. In dynamic languages where the compiler cannot catch this, the risk is real.

1.2 Convention Violations

Team coding conventions, naming rules, and module boundaries are routinely ignored. Linters catch some of it, but the implicit "this project does things like that" agreement is not something an agent absorbs from a single prompt.

1.3 Burden of Bulk Changes

An agent can touch dozens of files in a single command. From a reviewer's seat that means a single PR drops dozens of diffs at once, and standard IDE diff viewers do not preserve focus at that scale. Reviews drift toward rubber-stamping, and bugs slip through.

2. What Hunk Proposes

Looking through the repository, Hunk's design boils down to a combination of two existing open-source building blocks. UI rendering leans on OpenTUI, and diff handling sits on Pierre-style libraries, together giving a terminal experience close to a GitHub PR screen.

2.1 Inline Comments and Keyboard Operations

Comments can be attached next to a diff line, and navigation across next hunk, previous hunk, and file switches is bound to single keystrokes. Mouse dependency is effectively zero, which is the most visible difference from IDE diffs.

2.2 Acting as a Gate for Agent Output

The core scenario is "review what the agent produced, then commit." In other words it slots in as a gate between the agent and git. That matters when automatic commits would be dangerous, such as operations infrastructure code.

2.3 Building Blocks, Not From Scratch

The repository assembles open-source pieces rather than reinventing everything. That widens the door to integrations with other tools later.

3. IDE Diff vs Terminal Diff: A Practical Comparison

The obvious objection here is "isn't VS Code's inline diff already enough?" My honest answer is: it depends on context.

AspectIDE Diff (VS Code etc.)Terminal Diff (Hunk etc.)
Visual richnessColors and iconsPlain text based
Keyboard efficiencyMouse neededShortcut driven
SSH remote workRemote-SSH setupWorks immediately
CI pipelinesHard to embedScript friendly
Large changesTab overloadFast key traversal
Agent workflowMixed with editingReview stays separate

The pivotal idea is separating review from editing. Looking at a diff inside an IDE almost guarantees you will start fixing code yourself before finishing the review. A terminal review tool keeps you in decision-maker mode — accept, reject, or comment. In the agent era, that separation matters more than it used to. I touched on related tooling differences in the AI Coding Tools 2026 Comparison.

4. What It Means for Korean Development Teams

Korean PR review culture varies sharply across companies. Large IT firms standardise on GitHub Enterprise or GitLab and review daily, while SMB and SI shops still treat reviews as a formality. Add AI agents and a few new dynamics show up.

4.1 Teams Where Over 50% of New Code Is AI-Authored

One startup I recently advised generates more than half of its new code via Claude Code. Review load is roughly two to three times higher than before, and the stock GitHub PR UI cannot keep up. The value proposition of a tool like Hunk is clearest in environments like this.

4.2 Extra Pressure in Finance and Public Sector

Where bringing external SaaS review tools inside is hard, a local terminal tool turns into the natural alternative. Data never leaves the box, and integrating audit logs is comparatively simple.

4.3 Entry Barrier

That said, CLI and TUI familiarity varies a lot among Korean developers. A junior who lives inside VS Code will need time to adopt a terminal review tool as a daily driver, and learning English-only shortcuts and messages adds friction. A company-level cheat sheet ends up being close to mandatory.

5. How the Developer Role Is Evolving

Stepping past the tool itself, the bigger trend deserves attention. Agent adoption is visibly moving the centre of gravity in developer skills.

5.1 From Author to Evaluator and Integrator

The value of writing code from scratch is sliding down slowly, while the value of judging an agent's output and integrating it cleanly is climbing fast. Review skill, architectural taste, and test design become the core competence.

5.2 The Senior/Junior Gap Widens

Counterintuitively, AI tools amplify the worth of senior engineers. When an agent emits code that "looks fine but is actually wrong," noticing it instantly comes from experience. I expect that gap to keep widening for another two to three years. The Ruflo post in Ruflo and Claude Code Multi-Agent Analysis covered some of the same dynamic.

5.3 New Job Titles Appearing

Roles like "AI code curator" or "agent workflow engineer" are starting to emerge. They cover prompt design, evaluation criteria definition, and review pipeline operations, not just code review in isolation.

6. Limits and Critique

Coming back to Hunk specifically, the downsides deserve equal space.

6.1 Early-Stage Project Risk

This is still a Show GN-stage early project. Stability, maintenance, and ecosystem integration have not been proven. Picking it as the primary tool in a mission-critical environment would be premature.

6.2 Handling Korean Comments

Like most TUI tools, there can be subtle issues with Korean input and rendering. If wide-character or emoji widths are miscomputed the layout can break, so verify before adopting at scale.

6.3 Not a Full GitHub PR Replacement

If your company standard is GitHub PR, review outcomes still need to flow back into GitHub. Realistically Hunk will land as a complementary tool rather than a replacement.

7. Conclusion

Treating Hunk as just one new tool misses the point. The bigger signal is the trend underneath — code review tools evolve alongside AI agents. As agents become routine, human time shifts further toward review, and the tools that make review fast and accurate rise in value sharply.

My practical suggestion is simple: try one or two terminal review tools, evaluate for a week against your real workflow, and share the "separate review from editing" idea inside your team. Tools are means, but the direction they point in is clear — code review competence is the developer's core skill. The next entry in the series will look at agent evaluation automation.

References