Agent-kit – a size-based mandatory review chain for Claude Code

What happened

The GitHub page returned raw HTML without the README content being parsed. Let me grep for the actual README text in that file. The GitHub page returned JavaScript-rendered HTML without the README content. Let me try the raw README URL and also the GitHub API. Now I have the actual README content. Let me write the article based on these concrete facts.

---

Agent-kit: a size-based mandatory review chain for Claude Code

A GitHub repository published by the user horilla98 under the name agent-kit has surfaced on Hacker News, describing itself as the free, open-source core of a size-based workflow gate for Claude Code. The tool inserts a mandatory review chain into agentic coding sessions: it reminds Claude Code of a ticket's size category and the required review steps at the start of every session, then tracks via a bot comment on the pull request which chain members have actually left a trace. The post received 1 point and 0 comments at time of writing.

How it works

This piece covers what agent-kit is, how it works in practice, why it matters for teams adopting Claude Code at scale, and who should be paying attention. It is aimed at developers already using or evaluating Claude Code and at engineering leads who want structured oversight over AI-generated pull requests.

What happened

The repository horilla98/agent-kit appeared on Hacker News, submitted by user danielhorilla on August 14, 2026. It presents itself as the free, open-source core of a larger paid product called agent-kit, now at version 2.0. The free core contains 8 scripts under .github/scripts/, 2 GitHub Actions workflows (ci.yml and pr-jelzok.yml), 3 agent charters out of a full set of 11, and 45 automated tests that pass independently of the paid portion. The project carries zero npm dependencies and requires only Node.js built-ins, with a minimum runtime of Node.js 22.6.

The free package is described as a verified subset of the agent-kit v2.0 source. It exposes the chain-table, workflow-gate, workflow-gate-logic, chain-trace, chain-trace-logic, gh-api, project-config, language-labels, and messages scripts. The 3 bundled agent charters are reka, handling code review; tibor, handling test engineering; and gergo, handling security and privacy with explicit veto power. All three share a file called _protokoll.md that governs their shared behavioral rules.

Why it matters

How it works

At the center of agent-kit is a size classification that determines which review steps are mandatory before a pull request can proceed. The workflow-gate and workflow-gate-logic scripts read that size signal and surface the required chain steps to Claude Code at the beginning of each session. When the agent opens a new coding session, it is told both the ticket's size category and which chain members must leave a trace. The chain-trace and chain-trace-logic scripts then monitor the pull request for evidence that each required reviewer has acted.

The bot comment mechanism on the pull request is the audit surface. Rather than blocking merges through branch protection alone, agent-kit writes a structured comment that records which of the named chain members have fulfilled their step. Installation requires copying the contents of .github/scripts/, .github/workflows/, and .claude/agents/ into the target repository, then adding an agent-kit.config.json file at the repo root. That configuration file accepts four keys: projekt for the project name, repo for the GitHub username and repository slug, nyelv for language (the README shows "english" as the example value), and teszt for the test command. One manual step remains: the three agent charters contain placeholder text for the project name and a one- to two-sentence project description that must be filled in by hand because the free package ships without the installer script included in the paid tier.

Why it matters

Who is affected

Claude Code and similar agentic coding tools can author and commit large diffs autonomously. The core oversight problem that emerges is that the normal code-review process was designed around humans writing code incrementally; an agent can produce a pull request that spans hundreds of lines in a single session, making size-blind review workflows inadequate. Agent-kit's approach of gating on ticket size before a session even starts addresses this upstream rather than at merge time, which is the moment when catching an oversized or under-reviewed diff is most disruptive.

The veto power granted to the gergo charter is a notable design choice. Giving a security and privacy agent the ability to block a chain, rather than just comment, encodes a specific risk hierarchy directly into the workflow. Whether that hierarchy is appropriate depends on the team, but the structure itself is unusual for an open-source agentic workflow tool. The bilingual template tree referenced in the paid feature list (the README shows the config key nyelv suggesting language as a first-class concern) also hints at a design intended for teams working across language boundaries, a situation where automated review chains carry additional coordination overhead.

Who is affected

The immediate audience is any team running Claude Code inside a GitHub-hosted repository that wants structured, auditable review rather than relying entirely on the agent's own judgment. The free core's 3 agent charters and 45 tests make it usable as a starting point without a paid commitment, though teams that need the full 11 charters, the 7 slash commands, the 13-step development lifecycle, or the bilingual template tree will find the free tier deliberately constrained. Teams with no Node.js 22.6 environment in their CI pipeline will need to address that requirement before anything else.

What to watch next

Engineering organizations that have already built their own Claude Code conventions through CLAUDE.md files or custom hooks will want to evaluate whether agent-kit's charter model conflicts with or complements what they have. The three shipped charters, reka, tibor, and gergo, each have a named responsibility domain, and teams that have already assigned those responsibilities differently will need to reconcile the naming and the behavioral rules in _protokoll.md with their own conventions before deploying.

What to watch next

The most concrete thing to verify before adopting agent-kit's free core is whether the 45 automated tests cover the workflow-gate behavior you actually need. The README states they are green on their own and carry no dependency on the paid part, but any size-threshold logic not exercised by those tests will be opaque until you read the scripts directly. A builder considering this tool should also check how pr-jelzok.yml interacts with existing branch protection rules, since adding a bot-comment-based trace mechanism on top of required status checks can produce conflicting merge conditions if the two are not configured in agreement.

The paid tier's scope, 11 agent charters, 7 slash commands, the Backlog import workflow, the installer script, and the full bilingual template tree, represents a substantially larger surface area than what is publicly inspectable. Until that code is released or independently reviewed, teams evaluating agent-kit for production use should treat the free core as a working prototype of the review-chain concept rather than as a complete solution. The HN submission had 0 comments at time of writing, so community feedback on edge cases has not yet accumulated; watching that thread for questions about the workflow-gate threshold logic and the veto mechanism will be more informative than the README alone.