GitHub Action for kyosei code review from konoka marketplace.

Kyosei is a multi-perspective AI code review plugin that analyzes pull requests for code quality, performance, security, test coverage, and documentation accuracy.

The Claude Code Review workflow installed via install-github-app does not re-review a PR after subsequent pushes. Once the initial review is posted, pushing fixes in response to feedback does not trigger a new review, so there is no automated way to verify that review comments have been properly addressed.

Using claude-code-action directly enables per-push reviews, but introduces other problems:

• Pushing to the same PR repeatedly causes the same comments to be posted over and over
• Comments that have already been answered with "this is intentional" or "by design" are re-posted on each push

The kyosei plugin solves these problems. It collects existing PR conversations (comments, inline comments, and review comments) before each review, excludes already-posted feedback, resolved comments, and comments that have been acknowledged as intentional, so only genuinely new feedback is provided. It also removes project-specific coding conventions embedded in claude-code-action's default review agents. For example, claude-code-action's code-quality-reviewer includes the instruction "Prefer type over interface as per project standards", which is applied even to projects that do not use TypeScript. kyosei strips such opinionated defaults and expects project-specific conventions to be specified in CLAUDE.md instead.

kyosei-action wraps the kyosei plugin as a GitHub Action, making it easy to run these reviews automatically in CI.

This repository provides a Composite Action and a Reusable Workflow. The Composite Action is a low-level building block that requires the caller to handle checkout and permissions. For simpler setup, use the Reusable Workflow.

At least one of the following is required. Typically you only need one; if multiple are provided they are passed through to claude-code-action as-is.

• claude_code_oauth_token - Claude Code OAuth token
• anthropic_api_key - Anthropic API key
• A cloud provider (use_bedrock, use_vertex, or use_foundry)

claude --bare setup-token

Copy the output token.

gh secret set CLAUDE_CODE_OAUTH_TOKEN

Paste the token when prompted.

gh secret set CLAUDE_CODE_OAUTH_TOKEN --app dependabot

Paste the token when prompted.

This project follows immutable releases: once a version tag is published, it is never moved or overwritten. Version tags such as kyosei-action@v2.3.0 are safe to use as-is.

If your policy requires pinning to a commit hash rather than a tag, you need the commit SHA, not the tag object SHA. Annotated tags have their own object SHA which differs from the commit SHA. GitHub Actions requires the commit SHA.

Use git ls-remote with ^{} to fetch the dereferenced commit SHA without cloning the repository:

git ls-remote https://github.com/ncaq/kyosei-action.git 'refs/tags/v2.3.0^{}'

Do not query refs/tags/v2.3.0 without ^{}. For annotated tags it returns the tag object SHA, which GitHub Actions cannot resolve.

Handles checkout and timeout internally. Permissions must be declared by the caller since the reusable workflow is constrained by the caller's permissions.

name: Kyosei

on:
  pull_request:
    # Only opened and synchronize to avoid duplicate reviews
    # on the same revision from ready_for_review or reopened events.
    types: [opened, synchronize]

permissions: {}

jobs:
  workflow:
    # Reusable workflows are constrained by the caller's permissions,
    # so they must be explicitly declared here.
    # Claude GitHub App manages its own token, so only minimal permissions are needed.
    permissions:
      contents: read # Read repository contents for checkout
      id-token: write # GitHub App token exchange via OIDC (needed regardless of auth method)
    uses: ncaq/kyosei-action/.github/workflows/review.yml@v2.3.0
    secrets:
      claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

Most Composite Action inputs can be passed via with:. The Reusable Workflow additionally accepts the following inputs:

runs-on accepts a plain string, a JSON-quoted string, a JSON array, or a JSON object:

# Single label (plain string)
with:
  runs-on: ubuntu-24.04

# Single label (JSON string — also works)
with:
  runs-on: '"ubuntu-24.04"'

# Multiple labels (JSON array)
with:
  runs-on: '["self-hosted", "linux"]'

# Runner group with labels (JSON object)
with:
  runs-on: '{"group":"my-group","labels":["x64"]}'

See the Composite Action section below for the full input list.

Examples below use version tags.

name: Kyosei

on:
  pull_request:
    # Only opened and synchronize to avoid duplicate reviews
    # on the same revision from ready_for_review or reopened events.
    types: [opened, synchronize]

permissions: {}

jobs:
  review:
    runs-on: ubuntu-24.04
    # Claude GitHub App manages its own token, so only minimal permissions are needed.
    permissions:
      contents: read # Read repository contents for checkout
      id-token: write # GitHub App token exchange via OIDC (needed regardless of auth method)
    timeout-minutes: 30
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
        with:
          persist-credentials: false
          fetch-depth: 50
      - uses: ncaq/kyosei-action@v2.3.0
        with:
          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

All inputs are optional.

Default: ""

Claude Code OAuth token.

Default: ""

Anthropic API key (alternative to OAuth token).

Default: false

Use Amazon Bedrock with OIDC authentication.

Default: false

Use Google Vertex AI with OIDC authentication.

Default: false

Use Microsoft Foundry with OIDC authentication.

Default: ""

GitHub token for API access. If omitted, claude-code-action uses Claude GitHub App token (claude[bot]). Provide explicitly to use a custom token or github.token instead. Named custom_github_token to avoid the reserved github_token input name.

Default: *

Comma-separated list of allowed bot usernames, or * to allow all bots. Defaults to * because bots are not inherently more dangerous than humans.

Default: ""

Comma-separated list of users without write permission who are allowed to trigger Claude, or * to allow all. Only works when custom_github_token is provided. Enables bubblewrap sandbox and env scrubbing for safety.

Default: ""

Filter to include only comments from specific actors. Supports wildcards (e.g. *[bot]). Empty means include all.

Default: ""

Filter to exclude comments from specific actors. Supports wildcards. Exclusion takes precedence over inclusion.

Default: ""

Additional GitHub permissions for the App token (newline-separated key: value). Example: actions: read enables CI/CD failure analysis. Only effective with OIDC token exchange (ignored when custom_github_token is set).

Default: ""

Claude Code settings as a JSON string or path to a JSON file. Merged with existing settings (input takes precedence). Can configure hooks, env, MCP settings, etc.

Default: opus

Claude model to use.

The Claude Code CLI is installed by this action itself instead of the one claude-code-action would install, so aliases such as opus resolve to the model of the version pinned in the Install Claude Code step of action.yml.

Default: medium

Reasoning effort level for the top-level orchestrator session. Valid values: low, medium, high, xhigh, max.

The kyosei skill itself only dispatches reviewer subagents and posts JSON, so it does not need deep reasoning. Lowering this level reduces cost and latency without affecting review quality.

Reviewer subagents (code-quality-reviewer, security-reviewer, etc.) have their own effort: high in agent frontmatter, which overrides the session-level value. The actual reviews still run at high regardless of this setting.

Set to an empty string to omit the flag and use the model default.

Default: see the allowed_tools default in action.yml.

Allowed tools for Claude Code (newline-separated, replaces default set). Restricted to read-only commands and non-destructive tools.

Mutations such as posting reviews are performed by the kyosei skill via its Node.js implementation rather than by Claude directly.

List each GitHub MCP tool individually as mcp__github__<tool_name>, because claude-code-action only starts the GitHub MCP server when a tool name begins with mcp__github__ (the bare mcp__github prefix does not work).

The default also allows what the bundled plugins need: the Agent tool and the MCP servers used by the research survey agent.

• mcp__backlog__* (only the read-only tools the survey agent uses)
• mcp__plugin_nix-tasuke_nixos
• mcp__plugin_research_*

The * here is conceptual shorthand for explanation, not a literal value. claude-code-action does not necessarily honor wildcards in allowed_tools, so the actual default enumerates each tool or server individually (for example mcp__backlog__get_issue and mcp__plugin_research_cloudflare).

The mcp__plugin_* servers ship with the bundled plugins and work out of the box. Backlog (mcp__backlog__*) is not bundled, so it only works if you configure the Backlog MCP server yourself (for example via .mcp.json).

Default: ""

Additional allowed tools to append to allowed_tools (newline-separated).

Default: ""

Additional arguments to pass to Claude CLI (appended after --model and --allowed-tools).

Default: true

Include "Fix this" deep links in PR review feedback.

Default: auto

Display Claude Code Report in GitHub Step Summary. Useful for understanding what Claude did during the review. auto (default) enables it only for private repositories. always enables it regardless of repository visibility. never disables it entirely. true/false are accepted as aliases for backward compatibility. When GitHub Actions debug logging is enabled (ACTIONS_STEP_DEBUG), the report is forced on regardless of this value, since it only summarizes what Claude did.

Default: false

Show full Claude Code JSON output in Actions logs. May contain secrets in tool results; use only for debugging. Ignored on public repositories to prevent secret leakage. When GitHub Actions debug logging is enabled (ACTIONS_STEP_DEBUG), the full output is shown automatically, but it remains restricted to private repositories.

Default: https://github.com/ncaq/konoka.git

Git URL of the plugin marketplace.

Default: kyosei@konoka and research@konoka (newline-separated)

Plugin identifier within the marketplace.

Default:

self_hosted_packages: |
  curl
  gawk
  gh
  git
  jq
  zstd

Newline-separated list of packages to install via apt-get on self-hosted runners. When the runner is self-hosted and apt-get is available, the listed packages are installed automatically. On runners without apt-get (e.g. Fedora, NixOS), installation is skipped and required commands must be pre-installed.

After installation (or skip), the action verifies that curl, gh, git, and node are available, and fails with an error if any are missing.

Node.js is set up separately via actions/setup-node and does not need to be included in this list.

To skip automatic installation entirely, pass an empty string:

with:
  self_hosted_packages: ""

On GitHub-hosted runners the apt-get step is always skipped because runner.environment is not self-hosted.

Please see the allowed_tools section of action.yml.

The full list of GitHub MCP tools includes all read-only tools from github-mcp-server v0.17.1:

• Actions
• Code
• Context
• Dependabot
• Discussions
• Issues
• Pull
• Repos
• Secret
• Users

To add tools without replacing the defaults, use additional_allowed_tools:

- uses: ncaq/kyosei-action@v2.3.0
  with:
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
    additional_allowed_tools: |
      Bash(npm test)
      Edit

When custom_github_token is omitted (default), Claude GitHub App manages its own token, so each job only needs minimal permissions:

permissions:
  contents: read # Read repository contents for checkout
  id-token: write # GitHub App token exchange via OIDC (needed regardless of auth method)

id-token: write is required even when using claude_code_oauth_token or anthropic_api_key. claude-code-action internally exchanges this GitHub OIDC token at Anthropic's endpoint for a short-lived GitHub App token that posts PR comments. This is separate from Claude API authentication. Only when custom_github_token is provided is this OIDC exchange skipped.

If you explicitly pass custom_github_token, the token needs additional permissions such as pull-requests: write for posting review comments. If the token lacks pull-requests: write (e.g. due to workflow file changes in the PR or fork PRs), the action will skip gracefully with a warning instead of failing.

direnv allow

nix fmt

nix-fast-build --option eval-cache false --no-link --skip-cached

Apache License 2.0