Kizunax

kizunax-plugin-cc

A Claude Code plugin that runs AI code reviews against the KizunaX API. Single Go binary, no extra CLI.

A personal research and learning project, built to study how LLM APIs and Claude Code plugins fit together.

How it works

Claude Code  ──/kizunax:review──▶  plugins/kizunax/bin/kizunax
                                          │
                                          ▼
                                  llmreviewkit engine
                                  (diff → prompt → JSON → markdown)
                                          │
                                          ▼
                                  KizunaX API
                                  (OpenAI- or Anthropic-compatible
                                   chat completions)
                                          │
                                          ▼
                                  markdown back to Claude

Engine details live in llmreviewkit. The plugin adds: slash commands, config with OpenAI- and Anthropic-compatible slots and a shared key pool, background jobs, hooks (SessionStart / SessionEnd cleanup + optional Stop review-gate), and a setup wizard.

Compared to codex-plugin-cc

Mechanism, not features:

	kizunax-plugin-cc	codex-plugin-cc
Execution	single LLM request	multi-turn agent loop
Transport	direct HTTP to the model API	spawns the Codex CLI subprocess
Diff handling	pre-packed bundle (diff + optional refs)	agent reads files on demand
Output enforcement	enforced JSON schema / tool_use	Codex CLI structured output
Token budget	capped at EnrichBudget (32 / 96 KiB)	no cap
Latency	~one inference call (or N parallel calls with --strategy=fanout on large diffs)	varies with task; can be minutes

Use kizunax for fast, predictable reviews of a specific change; use codex when the model needs to read beyond the diff.

Install

macOS / Linux:

git clone https://github.com/thanhhaudev/kizunax-plugin-cc.git
cd kizunax-plugin-cc
./install.sh

Windows (PowerShell):

git clone https://github.com/thanhhaudev/kizunax-plugin-cc.git
cd kizunax-plugin-cc
.\install.ps1

Requires git, curl, and one of jq / python3 (macOS / Linux) or PowerShell 5.1+ (Windows). Go 1.21+ only if no pre-built binary matches your platform.

Setup

From the repo root:

./setup.sh

Or, inside Claude Code:

/kizunax:setup

Both open a local browser form for provider, base URL, model, and API key(s). Saved to ~/.kizunax/config.json (%USERPROFILE%\.kizunax\config.json on Windows).

Other flags ./setup.sh accepts: --status, --enable-stop-gate, --disable-stop-gate.

Commands

Command	What it does
/kizunax:review	Review changes (working tree, branch base, commit, or range). Flags: --strategy auto|single|fanout, --paths, --focus, --model, --add-context-prompt. Accepts trailing free-form text as focus.
/kizunax:adversarial-review	Same as review (all flags + trailing focus text apply), but the model challenges the design and looks for risks.
/kizunax:context	Build or refresh .kizunax/review-context.md by synthesizing CLAUDE.md + memory + conversation. v0.27.0+.
/kizunax:status	List background review jobs.
/kizunax:result <id>	Print the result of a finished job.
/kizunax:cancel <id>	Cancel a running background job.
/kizunax:setup	Re-configure provider, model, or API keys.
/kizunax:usage	Print per-key quota table.
/kizunax:index <sub>	Manage the workspace AST index.
/kizunax:expansion <sub>	Toggle bundle expansion strategies.

--strategy=auto (default since v0.26.0) routes large diffs through the binary's internal fan-out (parallel review buckets merged into one report), keeping the slash-command bash count to 1 per review. Full flag detail in docs/flag-reference.md.

Examples

# Review the working tree (default target)
/kizunax:review

# Review the current branch against develop (git-flow base)
/kizunax:review --base develop

# Review only specific paths
/kizunax:review --paths internal/auth,internal/handlers --focus "race conditions"

# Adversarial review with trailing focus text
/kizunax:adversarial-review --base main challenge the cache invalidation strategy

# Force fan-out for a large diff (parallel buckets + Anthropic)
/kizunax:review --base develop --strategy=fanout

# Override the configured model for one invocation
/kizunax:review --model claude-opus-4-7

# Per-review inline context hint (asks once via AskUserQuestion)
/kizunax:review --base develop --add-context-prompt

# Build .kizunax/review-context.md from CLAUDE.md + memory + conversation
/kizunax:context

Configuration

Three files / one env var shape kizunax's behavior on top of ~/.kizunax/config.json (provider + API keys, written by /kizunax:setup):

File	Purpose
.kizunax/review-context.md	Auto-injected above the system prompt: intentional patterns the reviewer should not flag, suppressed categories, business notes. Generate via /kizunax:context. v0.27.0+.
.kizunax/glossary.md	Project-specific vocabulary auto-injected as the "Project glossary" section. v0.11+.
CLAUDE.md (workspace root)	Already a Claude Code convention; /kizunax:context reads it as one of the sources when synthesizing review context.

Environment variable:

• KIZUNAX_PHP_EXTRACTOR=auto|phpsyms|treesitter|regex — override the PHP extraction strategy. Default auto prefers Go-native phpsyms and falls back to tree-sitter on empty result. v0.25.0+.

Uninstall

macOS / Linux:

./uninstall.sh

Windows (PowerShell):

.\uninstall.ps1

Both remove the plugin from settings.json and delete the binary. Your config and job history under ~/.kizunax/ stay; delete by hand for a full reset.

---

Inspired by codex-plugin-cc.