kwrkb · Sep 19, 2026 · Sep 19, 2026 · Sep 19, 2026 · Sep 19, 2026
diff --git a/.github/workflows/build-verify.yml b/.github/workflows/build-verify.yml
@@ -14,6 +14,7 @@ on:
       - 'settings-advisor/**'
       - 'go-lsp/**'
       - 'worktree-manager/**'
+      - 'test-runner/**'
       - 'build.sh'
       - 'build.ps1'
       - '.go-version'
@@ -29,6 +30,7 @@ on:
       - 'settings-advisor/**'
       - 'go-lsp/**'
       - 'worktree-manager/**'
+      - 'test-runner/**'
       - 'build.sh'
       - 'build.ps1'
       - '.go-version'
@@ -244,3 +246,36 @@ jobs:
             exit 1
           fi
         working-directory: .
+
+  test-runner:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+    runs-on: ${{ matrix.os }}
+    defaults:
+      run:
+        working-directory: test-runner/src
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: .go-version
+      - run: go vet ./...
+      - run: go test -count=1 ./...
+      - name: govulncheck
+        if: runner.os == 'Linux'
+        run: |
+          go install golang.org/x/vuln/cmd/govulncheck@latest
+          "$GITHUB_WORKSPACE/.github/scripts/govulncheck-gate.sh" "$GITHUB_WORKSPACE/.go-version"
+      - name: Deterministic rebuild
+        if: runner.os == 'Linux'
+        run: ./build.sh test-runner
+        working-directory: .
+      - name: Verify committed binaries are up to date
+        if: runner.os == 'Linux'
+        run: git diff --exit-code -- test-runner/bin/test-runner-linux-amd64 test-runner/bin/test-runner-darwin-arm64 test-runner/bin/test-runner.exe
+        working-directory: .
+      - name: MCP smoke test through the distributed dispatcher
+        if: runner.os == 'Linux'
+        run: TEST_RUNNER_BINARY="$GITHUB_WORKSPACE/test-runner/bin/test-runner" go test -count=1 -run '^TestMCPStdio$' ./...
diff --git a/AGENTS.md b/AGENTS.md
@@ -13,6 +13,7 @@
 - `retro-status/`: リポジトリ情報をレトロゲーム風に表示する Go 製 MCP サーバー
 - `settings-advisor/`: ワークスペースに適した agy 設定を提案する Go 製 MCP サーバー
 - `worktree-manager/`: Git worktree の作成・削除・一覧・prune を行う Go 製 MCP サーバー
+- `test-runner/`: Go テストを実行し、失敗ログと再実行引数を返す Go 製 MCP サーバー
 - `agy-plugin-kit/`: プラグイン作成用の command、skill、template、および Go 製 validator
 
 Go 製プラグインは原則として `<plugin>/src/` に独立した Go モジュール、`<plugin>/bin/` に配布用バイナリを持ちます。validator のみ `agy-plugin-kit/validator/{src,bin}/` 配下です。
@@ -71,12 +72,13 @@ go-lsp/src
 retro-status/src
 settings-advisor/src
 worktree-manager/src
+test-runner/src
 ```
 
 Go ソースを変更したら、リポジトリルートで対象を決定論的に再ビルドします。
 
 ```sh
-./build.sh <github|validator|ast-grep|go-lsp|retro-status|settings-advisor|worktree-manager>
+./build.sh <github|validator|ast-grep|go-lsp|retro-status|settings-advisor|worktree-manager|test-runner>
 ```
 
 Windows PowerShell では同じ target を `./build.ps1` に渡します。決定論ビルドが前提とする Go のパッチ版は `.go-version` に定義され、`build.sh`/`build.ps1` が `GOTOOLCHAIN` で強制します（未取得なら Go が自動ダウンロード）。素の `go build` で焼いたバイナリはローカル Go 版によって別物になるため、必ずビルドスクリプト経由で再ビルドしてください。`.go-version` を独断で書き換えないこと。引き上げ条件は CLAUDE.md「Go バージョンの引き上げ方針」に定義され、条件1（固定中のマイナー内のパッチで直る stdlib 脆弱性）は CI の govulncheck が fail で通知します。

diff --git a/LESSONS.md b/LESSONS.md
@@ -458,3 +458,8 @@ response = json.loads(proc.stdout.readline())
 
 #### 判断 14: bot 誤指摘（`master`→`main`）の却下
 - Codex GitHub bot が「既定ブランチは `main`」と P2 指摘したが、`gh repo view`=`master`・`main` ref 不在・PR #7 が master マージ済みで**前提が誤り**と検証。鵜呑みにすると実ブランチでゲートが無効化される。却下根拠を PR コメントとコミットに残した（#28）。
+
+## 2026-09-19: test-runner はログの保持上限と結果イベントの集計を分離する
+- 却下した案: `CombinedOutput` でテスト出力全体を保持してから解析する、またはログ上限に達した時点で出力の読み取りを止める。
+- 決め手: 回帰テストでは 64 KiB のログを40回（2.5 MiB）出力した後に成功・失敗イベントを送る。逐次解析とログ保持上限を分けることで、保持1 MiB以内・返却ログ64 KiB以内を守りながら末尾の成功1件・失敗1件を集計できた。読み取りを止めると末尾の結果を失い、子プロセスのパイプも詰まる。
+- 覆す条件: 完全なログをディスク上の成果物として保存・取得する機能が必要になった場合。応答サイズ制限と結果集計は引き続き独立させる。
diff --git a/PLAN.md b/PLAN.md
@@ -209,3 +209,10 @@ settings-advisor のスキャン精度（CI/CD、言語、除外ディレクト
 - [x] ステップ1: スキャン検知機能と除外ロジックの改善・テスト追加
 - [x] ステップ2: models.json 更新と traits 駆動モデル選定ロジックの実装・テスト追加
 - [x] ステップ3: ドキュメント同期（README.md / README_ja.md）とビルド・品質検証
+
+## フェーズ13: Go テストランナー (test-runner) の追加
+
+現在地: 実装・Linuxローカル検証完了。macOS/Windowsのネイティブ実行はCI確認待ち。次フェーズには進まない。
+
+- [x] MCP実行・結果集計・失敗テストの再実行引数が実装され、境界とキャンセルを検証済み。
+- [x] 日英ドキュメント・skill・3環境の配布物・CIが揃い、ローカル検証が完了。
diff --git a/README.md b/README.md
@@ -14,6 +14,7 @@ A collection of Model Context Protocol (MCP) plugins designed for AI assistants,
 | **[retro-status](./retro-status/README.md)** | MCP server scanning repositories to render RPG-style retro status ASCII art |
 | **[settings-advisor](./settings-advisor/README.md)** | MCP server analyzing workspace scale, languages, and settings to recommend optimal models and sandbox configurations |
 | **[worktree-manager](./worktree-manager/README.md)** | MCP server providing safe git worktree operations (list, add, remove, prune) for isolated parallel workflows |
+| **[test-runner](./test-runner/README.md)** | MCP server running Go tests with bounded failure logs and targeted rerun arguments |
 
 ## Installation
 
@@ -41,6 +42,9 @@ agy plugin install https://github.com/kwrkb/agy-plugins/settings-advisor
 
 # worktree-manager plugin
 agy plugin install https://github.com/kwrkb/agy-plugins/worktree-manager
+
+# test-runner plugin
+agy plugin install https://github.com/kwrkb/agy-plugins/test-runner
 ```
 
 Please refer to the README in each plugin directory for specific prerequisites (required CLI binaries on PATH / authentication).
@@ -61,12 +65,13 @@ Each plugin includes an agent guide in `skills/<name>/SKILL.md` (loaded upon too
 | retro-status | `git` (recommended), `rg` (optional) | None |
 | settings-advisor | None (prebuilt binaries included) | None |
 | worktree-manager | `git` (on PATH) | None |
+| test-runner | `go` >= 1.24 (on PATH) | None |
 
 ### Bundled Platforms & Self-Build
 
-✅ **Windows native execution verified across all plugins**
+Existing plugins have been verified on native Windows. `test-runner` includes CI tests for Linux, macOS, and Windows; local verification is on Linux.
 
-Go-based plugins (`github`, `ast-grep`, `go-lsp`, `retro-status`, `settings-advisor`, `worktree-manager`, and `agy-plugin-kit` validator) include native binaries for **`linux/amd64`**, **`darwin/arm64` (Apple Silicon)**, and **`windows/amd64`** in `bin/`. The extensionless `bin/<name>` dispatcher script detects the platform via `uname` and executes `<name>-<goos>-<goarch>` (on Windows, agy directly invokes `<name>.exe`).
+Go-based plugins (`github`, `ast-grep`, `go-lsp`, `retro-status`, `settings-advisor`, `worktree-manager`, `test-runner`, and `agy-plugin-kit` validator) include native binaries for **`linux/amd64`**, **`darwin/arm64` (Apple Silicon)**, and **`windows/amd64`** in `bin/`. The extensionless `bin/<name>` dispatcher script detects the platform via `uname` and executes `<name>-<goos>-<goarch>` (on Windows, agy directly invokes `<name>.exe`).
 
 **Other platforms (e.g. `linux/arm64` / `darwin/amd64`) are not bundled by default**, but can be built on the target machine without modifying any scripts:
 
@@ -89,5 +94,6 @@ Each plugin acts as a wrapper for existing tools/servers and complies with their
 | settings-advisor | Custom Go implementation | MIT |
 | agy-plugin-kit | Custom Go implementation | MIT |
 | worktree-manager | `git` CLI | MIT |
+| test-runner | `go test` (Go toolchain) | BSD-3-Clause |
 
 Plugins delegate execution to user-installed binaries/CLIs on PATH (no redistribution of external binaries).
diff --git a/README_ja.md b/README_ja.md
@@ -14,6 +14,7 @@
 | **[retro-status](./retro-status/README_ja.md)** | リポジトリをスキャンし、RPGのレトロステータス画面風のAA（アスキーアート）で出力する MCP サーバー |
 | **[settings-advisor](./settings-advisor/README_ja.md)** | ワークスペースの規模・言語・機密/CI/本番設定を解析し、最適なモデル・サンドボックス・許可モードを控えめに提案する MCP サーバー |
 | **[worktree-manager](./worktree-manager/README_ja.md)** | `git worktree` の安全な操作（一覧・作成・削除・クリーンアップ）を提供し、ブランチ隔離や並行作業を支援する MCP サーバー |
+| **[test-runner](./test-runner/README_ja.md)** | Goテストの実行・失敗ログの整理・再実行引数の生成を行う MCP サーバー |
 
 ## インストール方法
 
@@ -41,6 +42,9 @@ agy plugin install https://github.com/kwrkb/agy-plugins/settings-advisor
 
 # worktree-manager プラグイン
 agy plugin install https://github.com/kwrkb/agy-plugins/worktree-manager
+
+# test-runner プラグイン
+agy plugin install https://github.com/kwrkb/agy-plugins/test-runner
 ```
 
 各プラグインの前提条件（PATH に入れるバイナリ / 認証設定）については、各ディレクトリの README を参照してください。
@@ -61,12 +65,13 @@ agy plugin install https://github.com/kwrkb/agy-plugins/worktree-manager
 | retro-status | `git`（推奨）、`rg`（任意） | 不要 |
 | settings-advisor | 不要（バイナリ同梱） | 不要 |
 | worktree-manager | `git`（PATH 上） | 不要 |
+| test-runner | `go` >= 1.24 （PATH 上） | 不要 |
 
 ### 同梱プラットフォームと self-build
 
-✅ **全プラグインで Windows ネイティブ動作確認済み**
+既存プラグインは Windows ネイティブ動作確認済み。`test-runner` は Linux・macOS・Windows のCIテストを追加し、ローカルではLinuxで検証しています。
 
-Go 製プラグイン（`github`、`ast-grep`、`go-lsp`、`retro-status`、`settings-advisor`、`worktree-manager`、および `agy-plugin-kit` の validator）は、**`linux/amd64`**、**`darwin/arm64`（Apple Silicon）**、**`windows/amd64`** のネイティブバイナリを `bin/` に同梱しています。拡張子なしの `bin/<name>` ディスパッチャスクリプトが `uname` で実行環境を判別し、`<name>-<goos>-<goarch>` を起動します（Windows では agy が `<name>.exe` を直接起動）。
+Go 製プラグイン（`github`、`ast-grep`、`go-lsp`、`retro-status`、`settings-advisor`、`worktree-manager`、`test-runner`、および `agy-plugin-kit` の validator）は、**`linux/amd64`**、**`darwin/arm64`（Apple Silicon）**、**`windows/amd64`** のネイティブバイナリを `bin/` に同梱しています。拡張子なしの `bin/<name>` ディスパッチャスクリプトが `uname` で実行環境を判別し、`<name>-<goos>-<goarch>` を起動します（Windows では agy が `<name>.exe` を直接起動）。
 
 **それ以外の環境（`linux/arm64` や `darwin/amd64` など）は標準同梱していません**が、スクリプトの書き換えなしで手元でビルドして動かせます：
 
@@ -89,5 +94,6 @@ cd <plugin>/src && CGO_ENABLED=0 go build -o "../bin/<name>-$(go env GOOS)-$(go
 | settings-advisor | 独自 Go 実装 | MIT |
 | agy-plugin-kit | 独自 Go 実装 | MIT |
 | worktree-manager | `git` CLI | MIT |
+| test-runner | `go test` (Go toolchain) | BSD-3-Clause |
 
 各プラグインはユーザーの PATH 上にあるインストール済みバイナリ/CLI に処理を委譲します（外部バイナリの再配布は行っていません）。
diff --git a/build.ps1 b/build.ps1
@@ -75,6 +75,7 @@ switch ($Target) {
     'settings-advisor' { Build 'settings-advisor' 'settings-advisor' }
     'go-lsp' { Build 'go-lsp' 'go-lsp' }
     'worktree-manager' { Build 'worktree-manager' 'worktree-manager' }
+    'test-runner' { Build 'test-runner' 'test-runner' }
     'all' {
         Build 'github' 'github'
         Build 'agy-plugin-kit/validator' 'validator'
@@ -83,9 +84,10 @@ switch ($Target) {
         Build 'settings-advisor' 'settings-advisor'
         Build 'go-lsp' 'go-lsp'
         Build 'worktree-manager' 'worktree-manager'
+        Build 'test-runner' 'test-runner'
     }
     default {
-        Write-Error "unknown target: $Target (expected: github | validator | ast-grep | retro-status | settings-advisor | go-lsp | worktree-manager | all)"
+        Write-Error "unknown target: $Target (expected: github | validator | ast-grep | retro-status | settings-advisor | go-lsp | worktree-manager | test-runner | all)"
         exit 2
     }
 }
diff --git a/build.sh b/build.sh
@@ -57,6 +57,7 @@ case "$target" in
 	settings-advisor) build settings-advisor settings-advisor ;;
 	go-lsp)       build go-lsp go-lsp ;;
 	worktree-manager) build worktree-manager worktree-manager ;;
+	test-runner) build test-runner test-runner ;;
 	all)
 		build github github
 		build agy-plugin-kit/validator validator
@@ -65,9 +66,10 @@ case "$target" in
 		build settings-advisor settings-advisor
 		build go-lsp go-lsp
 		build worktree-manager worktree-manager
+		build test-runner test-runner
 		;;
 	*)
-		echo "unknown target: $target (expected: github | validator | ast-grep | retro-status | settings-advisor | go-lsp | worktree-manager | all)" >&2
+		echo "unknown target: $target (expected: github | validator | ast-grep | retro-status | settings-advisor | go-lsp | worktree-manager | test-runner | all)" >&2
 		exit 2
 		;;
 esac
diff --git a/test-runner/README.md b/test-runner/README.md
@@ -0,0 +1,57 @@
+# test-runner — Go Test Runner
+
+A Go-only MCP plugin for agy. Runs one module at a time, summarizes test/build failures, and returns arguments for explicitly rerunning failed tests. Calls wait for completion; there is no background job, persistent history, or automatic retry.
+
+## Requirements and installation
+
+- Go 1.24 or later on `PATH`, with the toolchain/dependencies required by the target module available.
+- Prebuilt servers: Linux amd64, macOS arm64, Windows amd64. Windows uses the OS-provided `taskkill.exe` to stop a process tree.
+
+```sh
+agy plugin install https://github.com/kwrkb/agy-plugins/test-runner
+```
+
+## Tool: `go_test`
+
+| Argument | Meaning |
+| :-- | :-- |
+| `module_path` | Required directory containing `go.mod`. Prefer an absolute path: relative paths use the MCP server's working directory. |
+| `packages` | Nonempty string array of Go package patterns; default `["./..."]`. Local paths and import paths are accepted only when all resolved packages belong to the selected module and are physically inside it. |
+| `run` | Optional Go `-run` expression, including slash-separated subtest expressions. |
+| `timeout_seconds` | Integer 1–300, default 60. Covers discovery, compilation, and tests. Process cleanup can take a few additional seconds. |
+
+Example arguments:
+
+```json
+{"module_path":"/code/agy-plugins/test-runner/src","packages":["./..."],"timeout_seconds":60}
+```
+
+The server resolves packages with `go list -e -json`, then runs `go test -json -count=1 -timeout=…` using an argument array, without a shell. It always disables successful-test caching. It inherits the Go environment (including `go.work`, `GOFLAGS`, build settings, and toolchain selection); it does not recursively discover other modules. A workspace containing several modules can be tested through separate calls with explicit module directories.
+
+No arbitrary command, extra flags, `.go` file list, race/coverage option, fuzzing session, or benchmark option is exposed. Standard `go test` behavior, including examples and fuzz seed cases, remains in effect. Tests **execute project code** with the server's permissions and may change files or access the network. Go may use/download caches, dependencies, or a toolchain according to the inherited environment. This plugin is not an execution sandbox; use the host's execution permissions.
+
+## Results and retry
+
+Each MCP result contains a JSON text object:
+
+- `status`: `passed`, `failed`, `no_tests`, `timeout`, `cancelled`, or `error`. Test/build failures are ordinary results; execution errors, timeouts, and cancellation set MCP `isError`.
+- `exit_code`: Go process exit code, or `null` when no process exit code is available. Discovery errors can report the discovery process's exit code.
+- `elapsed_seconds`: Total elapsed time; `packages[].elapsed_seconds` is Go's per-package duration.
+- `tests`: `passed`, `failed`, `skipped` terminal-event counts, **including parent tests and subtests**. Package events are not counted as tests. All-skipped tests are distinguished by these counts; `no_tests` means no test terminal events and no failures.
+- `packages`: Sorted package results with Go status `pass`, `fail`, `skip`, or `incomplete`; failing tests include their names, logs, and `rerun` arguments. Package-level failures such as a panic or `TestMain` exit may have no named failure; inspect the package log.
+- `build_failures`: Build package IDs and diagnostics, separate from failed test counts. IDs may include a test variant suffix.
+- `stderr`, `error`: Diagnostics and execution errors, when present.
+- `logs_truncated`: Some log bytes were omitted. Retained event logs and stderr together are bounded to 1 MiB; returned log text is bounded to 64 KiB, prioritizing named failures. JSON escaping can make the serialized result larger.
+- `incomplete`: Partial output, interrupted execution, or missing terminal events. Never interpret this as complete validation.
+
+Use a failure's `rerun` object as the next `go_test` arguments. It retains the absolute module path and timeout, selects one package, and anchors/escapes each subtest-name segment. Go still runs parent setup to discover a selected subtest. Parent failures can overlap child failures; choose the parent to rerun a group or the child to narrow the retry. There is no server-side retry state.
+
+Malformed/oversized JSON events (over 1 MiB), over 100,000 result records, and discovery output over 8 MiB produce explicit execution errors. Output continues to drain after a parsing error to avoid blocking the child; the overall timeout still applies. Timeout/cancellation stops the Unix process group or the Windows process tree and bounds pipe shutdown. Programs that deliberately detach from the process tree are outside this mechanism.
+
+## Development and validation
+
+In `test-runner/src`, run `gofmt`, `go vet ./...`, and `go test -count=1 ./...`. Tests use temporary modules without third-party dependencies and verify the real Go JSON stream, exact subtest retries, child-process termination, and MCP stdio calls.
+
+From the repository root, run `./build.sh test-runner` (or `./build.ps1 test-runner`). These use the pinned `.go-version`; do not replace distributed binaries with an unpinned build. CI runs tests on Linux, macOS, and Windows and checks vulnerability policy and deterministic artifacts on Linux.
+
+The bundled agent guide is [test-runner-gemini](skills/test-runner-gemini/SKILL.md).