Add kensai multi-agent code review plugin

.claude-plugin/marketplace.json

@@ -0,0 +1,28 @@
+{
+	"name": "yatai",
+	"owner": {
+		"name": "Gaijin Entertainment",
+		"email": "a.zinoviev@gaijin.team"
+	},
+	"metadata": {
+		"description": "Gaijin Entertainment Claude Code plugin marketplace",
+		"pluginRoot": "./claude"
+	},
+	"plugins": [
+		{
+			"name": "kensai",
+			"description": "Multi-agent code review pipeline with adversarial falsification, orchestrated by a stateful MCP server",
+			"source": "kensai",
+			"version": "0.1.0",
+			"license": "MIT",
+			"keywords": [
+				"code-review",
+				"multi-agent",
+				"review-pipeline",
+				"falsification",
+				"proving",
+				"mcp"
+			]
+		}
+	]
+}

.claude/settings.json

@@ -0,0 +1,19 @@
+{
+  "enabledPlugins": {
+    "ai-helpers@foundry": true,
+    "git-commit@foundry": true,
+    "skill-enforcer@foundry": true,
+    "the-coder@foundry": true,
+    "the-workflow@foundry": true,
+    "javascript@foundry": true
+  },
+  "extraKnownMarketplaces": {
+    "foundry": {
+      "source": {
+        "source": "github",
+        "repo": "xobotyi/cc-foundry"
+      }
+    }
+  },
+  "alwaysThinkingEnabled": true
+}

.github/workflows/kensai-review-mcp.yml

@@ -0,0 +1,58 @@
+name: Kensai Review MCP
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - "claude/kensai/mcp/review/**"
+      - ".github/workflows/kensai-review-mcp.yml"
+  pull_request:
+    paths:
+      - "claude/kensai/mcp/review/**"
+      - ".github/workflows/kensai-review-mcp.yml"
+
+defaults:
+  run:
+    working-directory: claude/kensai/mcp/review
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: pnpm/action-setup@v4
+        with:
+          package_json_file: claude/kensai/mcp/review/package.json
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "26"
+          cache: pnpm
+          cache-dependency-path: claude/kensai/mcp/review/pnpm-lock.yaml
+
+      - run: pnpm install --frozen-lockfile
+
+      - run: pnpm lint
+
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        node-version: ["lts/*", "26"]
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: pnpm/action-setup@v4
+        with:
+          package_json_file: claude/kensai/mcp/review/package.json
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: ${{ matrix.node-version }}
+          cache: pnpm
+          cache-dependency-path: claude/kensai/mcp/review/pnpm-lock.yaml
+
+      - run: pnpm install --frozen-lockfile
+
+      - run: pnpm test

.gitignore

@@ -0,0 +1,3 @@
+.idea
+node_modules/
+.DS_Store

CLAUDE.md

@@ -0,0 +1,86 @@
+# yatai
+
+Gaijin Entertainment plugin repository. Houses plugins for two ecosystems: Claude Code (`claude/`) and Pi (`pi/`).
+
+## Structure
+
+```
+yatai/
+├── claude/               # Claude Code plugins
+│   └── kensai/           # Multi-agent code review pipeline
+└── pi/                   # PI plugins (future)
+```
+
+Each ecosystem has its own instruction file (`CLAUDE.md`, `AGENTS.md`, etc.) with ecosystem-specific conventions.
+
+## Conventions
+
+<conventions>
+
+**Plugin context:**
+
+- Each plugin has an agent-facing instruction file (`CLAUDE.md`, `AGENTS.md`, etc.) explaining its components
+- Each plugin has `README.md` for humans
+
+**Plugin documentation style:**
+
+- **Agent-facing file** — brief prose introduction, then structured content: component tables, dependency diagrams,
+  bullet-list conventions. Optimize for LLM compliance — terse bullets over explanatory paragraphs.
+- **README.md** — human-facing documentation. Explanatory prose that frames the plugin around the problem it solves,
+  explains what each component does and when to use it.
+
+**Documentation maintenance:** Update plugin docs in the same work session as the code change:
+
+- **Plugin agent-facing file** — update when: adding/removing/renaming components; changing purpose, scope, or
+  dependencies; adding/removing conventions
+- **Plugin README.md** — update when: any change that affects what users see or install
+- **Ecosystem instruction file** — update plugin table when: adding/removing a plugin
+- **Root CLAUDE.md** — update structure diagram when: adding/removing a plugin or ecosystem
+
+**Installation instructions:**
+
+- All plugin READMEs must use marketplace commands:
+  ```
+  /plugin marketplace add GaijinEntertainment/yatai
+  /plugin install <plugin-name>
+  ```
+- Do not use manual cp/ln installation methods
+
+**Version management:**
+
+- Plugin versions must be synchronized between each plugin's `.claude-plugin/plugin.json` and the repo-root
+  `.claude-plugin/marketplace.json` (which lists all plugins in a single file)
+- Update both files when bumping versions
+
+**Licensing:**
+
+- Every plugin must contain a copy of the root `LICENSE` file in its directory
+- When adding a new plugin, copy `LICENSE` from the repository root into the plugin directory
+
+**npm packages:**
+
+- Use `vp` for all toolchain operations — `vp install`, `vp check`, `vp test`, `vp build`. Never invoke `npm`,
+  `yarn`, or bare `pnpm` commands directly in instructions or CI
+- Every `package.json` must include: `license` (matching the plugin license), `author`, `repository`, `type` (`module`),
+  `engines`, and `packageManager` fields
+- Publishable packages must also include: `description`, `bin` or `exports`, `engines`, and `publishConfig`
+- Publishable packages must have a `.npmignore` excluding tests, configs, lock files, and agent-facing docs
+- Lock files (`pnpm-lock.yaml`) must be committed
+- Pin `packageManager` in `package.json` to the exact pnpm version used by the project
+- Use `--frozen-lockfile` in CI to catch uncommitted dependency changes
+
+**npm supply chain protection:**
+
+- Every package must have an `.npmrc` with `ignore-scripts=true` (blocks post-install script attacks) and
+  `minimum-release-age=7d` (rejects packages published less than 7 days ago)
+- All publishable packages must set `"publishConfig": { "access": "public" }` in `package.json`
+- The `@gaijin` npm scope must have 2FA enforced at the org level (manual npm setting)
+
+</conventions>
+
+<git-commit-config>
+<extra-instructions>
+Since the project is about plugins, the scope in the commit message, if defined, must not contain the ecosystem path
+(`claude/` or `pi/`). Use the plugin name as the scope.
+</extra-instructions>
+</git-commit-config>

claude/CLAUDE.md

@@ -0,0 +1,37 @@
+# Claude Code Plugins
+
+Claude Code plugin ecosystem. Each subdirectory is a self-contained plugin.
+
+## Plugins
+
+| Plugin | Purpose |
+|--------|---------|
+| [kensai](kensai/) | Multi-agent code review pipeline with adversarial falsification |
+
+## Conventions
+
+<conventions>
+**Formatting:**
+- Agent instruction files and prompts use the MDX parser (block-level XML tag support)
+- README.md files use GFM parser
+
+**Skill structure:**
+
+- Router pattern: SKILL.md routes to `references/` for detailed content
+- Keep SKILL.md under 500 lines; move depth to references
+
+**Installation instructions:**
+
+- All plugin READMEs must use marketplace commands:
+  ```
+  /plugin marketplace add GaijinEntertainment/yatai
+  /plugin install <plugin-name>
+  ```
+- Do not use manual cp/ln installation methods
+
+**Version management:**
+
+- Plugin versions must be synchronized between each plugin's `.claude-plugin/plugin.json` and the repo-root
+  `.claude-plugin/marketplace.json` (which lists all plugins in a single file)
+- Update both files when bumping versions
+</conventions>

claude/kensai/.claude-plugin/plugin.json

@@ -0,0 +1,19 @@
+{
+	"name": "kensai",
+	"version": "0.1.0",
+	"description": "Multi-agent code review pipeline with adversarial falsification, orchestrated by a stateful MCP server",
+	"author": {
+		"name": "xobotyi",
+		"email": "a.zinoviev@gaijin.team"
+	},
+	"repository": "https://github.com/GaijinEntertainment/yatai",
+	"license": "MIT",
+	"keywords": [
+		"code-review",
+		"multi-agent",
+		"review-pipeline",
+		"falsification",
+		"proving",
+		"mcp"
+	]
+}

claude/kensai/.mcp.json

@@ -0,0 +1,10 @@
+{
+  "mcpServers": {
+    "kensai": {
+      "type": "stdio",
+      "command": "node",
+      "args": ["${CLAUDE_PLUGIN_ROOT}/mcp/review/src/index.ts"],
+      "alwaysLoad": true
+    }
+  }
+}

claude/kensai/CLAUDE.md

@@ -0,0 +1,92 @@
+# Kensai
+
+Multi-agent code review pipeline with adversarial falsification. A stateful MCP server orchestrates four phases —
+grounding, surfacing, proving, filing — with dedicated agents per phase.
+
+## Pipeline
+
+```
+/kensai-review [mode] [path]
+  → mcp__kensai__session_start            IDLE → GROUNDING
+  → mcp__kensai__grounding_store
+  → mcp__kensai__grounding_complete       GROUNDING → SURFACING
+  → mcp__kensai__finding_surface / surface_clean
+  → mcp__kensai__surfacing_complete       SURFACING → PROVING (or FILING if 0 findings)
+  → mcp__kensai__finding_verdict
+  → mcp__kensai__proving_complete         PROVING → FILING
+  → mcp__kensai__filing_complete          FILING → COMPLETE
+```
+
+## Components
+
+| Component | Purpose |
+|-----------|---------|
+| `mcp/review/src/index.ts` | MCP server — phase state machine, context persistence, handler-level phase enforcement |
+| `skills/review/SKILL.md` | `/kensai-review` entry point — mode/path resolution, pipeline orchestration |
+| `agents/review-grounder.md` | Grounding agent — maps the change, builds three-layer model, stores context via MCP |
+| `agents/review-surfacer.md` | Surfacing agent — parallel read-only reviewer per dimension, records findings via MCP |
+| `agents/review-prover.md` | Proving agent — adversarial falsification of surfaced findings, verdicts via MCP |
+
+## MCP Tools
+
+### Session lifecycle
+
+| Tool | Purpose |
+|------|---------|
+| `session_start` | Initialize session (IDLE -> GROUNDING) |
+| `session_state` | Current phase, mode, refs, file/finding counts |
+| `session_priming` | Paginated full context: metadata, changed files, file stats, agent-instructions, diffs. Call with `page=N` for large changes. |
+| `session_end` | Destroy session |
+
+### Filesystem and search (session-scoped)
+
+| Tool | Purpose |
+|------|---------|
+| `read_file` | Read file with line numbers, binary detection |
+| `list_dir` | Directory listing with depth control |
+| `find_files` | Fuzzy file-name search |
+| `grep` | Regex content search via ripgrep |
+
+### Git (session-scoped)
+
+| Tool | Purpose |
+|------|---------|
+| `changed_files` | File list with status (A/M/D) and +/- counts |
+| `diff_file` | Annotated unified diff for one file |
+| `log` | Recent commit history |
+
+### Review pipeline
+
+| Tool | Phase | Purpose |
+|------|-------|---------|
+| `observation_create` | GROUNDING | Record an observation |
+| `observation_cancel` | GROUNDING | Retract an observation |
+| `grounding_store` | GROUNDING | Persist grounding context |
+| `grounding_get` | SURFACING+ | Retrieve grounding context |
+| `grounding_complete` | GROUNDING -> SURFACING | Advance phase |
+| `finding_surface` | SURFACING | Record a finding |
+| `surface_clean` | SURFACING | Report a clean dimension |
+| `finding_cancel` | SURFACING | Retract a finding |
+| `surfacing_complete` | SURFACING -> PROVING/FILING | Advance (skip proving if 0 findings) |
+| `findings_list` | PROVING+ | List findings with filter |
+| `finding_get` | PROVING+ | Get single finding detail |
+| `finding_verdict` | PROVING | Verdict a finding (pass/reject) |
+| `proving_complete` | PROVING -> FILING | Advance phase |
+| `filing_complete` | FILING -> COMPLETE | Mark review done |
+
+## Requirements
+
+- Agent teams: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
+- Node.js 26+ (MCP server)
+
+## Design
+
+- **MCP-first** — agents access codebase exclusively through `mcp__kensai__*` tools; no direct Read/Grep/Bash
+- **Self-serve context** — agents call `session_priming` to receive full review context; no data threading from lead
+- **Handler-level phase enforcement** — all tools always registered; invalid-phase calls return hard error via `#require()`
+- **Dedicated completion tools** — `grounding_complete`, `surfacing_complete`, `proving_complete` (not generic `transition`)
+- **0-findings shortcut** — `surfacing_complete` jumps to FILING when no findings, skipping PROVING
+- **Recall over precision** — surfacers cast a wide net; prover applies falsification gates to filter
+- **Read-only agents** — grounder, surfacers, and prover never edit files
+- **Stable finding IDs** — F1, F2, ... carried across phase boundaries
+- **Soft errors** — not-found/binary/empty return `ok("[marker]")` with path suggestions, not `err()`

claude/kensai/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Gaijin Entertainment
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.