Add afd-plugin adapter and knowledge routing

.env.template

@@ -35,20 +35,26 @@ CLAUDE_CODE_ATTREBUTION_HEADER=0
 
 # Repos: JSON map of repo name -> local path (machine-specific; .env is gitignored,
 # which is why absolute paths belong HERE and never in a committed file).
-REPO_PATHS={"vllm-omni": "/path/to/vllm-omni"}
+REPO_PATHS={"vllm-omni": "/path/to/vllm-omni", "afd-plugin": "/path/to/afd-plugin"}
 DEFAULT_REPO=vllm-omni
 
 # alias -> "owner/repo" GitHub identity, used to route pasted GitHub URLs to a
 # configured repo. Optional when the checkout's origin remote already points at
 # that owner/repo; without either, URL commands cannot route to the alias
 # (`doctor` warns about this).
-REPO_FULL_NAMES={"vllm-omni": "vllm-project/vllm-omni"}
+REPO_FULL_NAMES={"vllm-omni": "vllm-project/vllm-omni", "afd-plugin": "vllm-project/afd-plugin"}
+
+# MCP defaults to DEFAULT_REPO only. Add afd-plugin explicitly when this server
+# should expose its read-only knowledge/review workflows.
+MCP_REPO_ALLOWLIST=["vllm-omni", "afd-plugin"]
 
 # Fallback used by adapters/<repo>/manifest.yaml when no .env is in reach; the
-# manifest references these as ${VLLM_OMNI_REPO} / ${VLLM_UPSTREAM_REPO} so the
+# manifest references these as ${VLLM_OMNI_REPO} / ${VLLM_UPSTREAM_REPO} /
+# ${AFD_PLUGIN_REPO} so the
 # committed manifest stays portable. REPO_PATHS above takes precedence.
 VLLM_OMNI_REPO=/path/to/vllm-omni
 VLLM_UPSTREAM_REPO=/path/to/vllm
+AFD_PLUGIN_REPO=/path/to/afd-plugin
 
 # Push safety: pushes are dry-run unless this is 1
 ALLOW_PUSH=0

adapters/afd_plugin/manifest.yaml

@@ -0,0 +1,95 @@
+# Human-reviewed adapter for vllm-project/afd-plugin.
+name: afd_plugin
+status: active
+created_by: human
+schema_version: 1
+capabilities: [orchestrator.external]
+
+repo:
+  path: ${AFD_PLUGIN_REPO}
+  full_name: vllm-project/afd-plugin
+  aliases: [afd-plugin, afd_plugin, vllm-project/afd-plugin]
+  default_branch: main
+  language: python
+
+knowledge:
+  source: vllm-project/afd-plugin
+  repo_subdir: repos/afd-plugin
+  briefing_docs:
+  - repos/afd-plugin/rules.md
+  - repos/afd-plugin/_index.md
+
+modules:
+  compat:
+    local_paths: [afd_plugin/compat/]
+    wave: 1
+    risk: high
+  config_validation:
+    local_paths: [afd_plugin/config.py, afd_plugin/validation.py, afd_plugin/config_utils.py]
+    wave: 1
+    risk: high
+  connectors_distributed:
+    local_paths: [afd_plugin/connectors/, afd_plugin/distributed/]
+    wave: 1
+    risk: high
+  workers_runners:
+    local_paths: [afd_plugin/v1/worker/, afd_plugin/model_executor/]
+    wave: 1
+    risk: high
+  native_backends:
+    local_paths: [csrc/gpu/, csrc/npu/]
+    wave: 1
+    risk: high
+  tests:
+    local_paths: [tests/unit/, tests/e2e/]
+    wave: 2
+
+review_routes:
+- prefix: afd_plugin/compat/
+  owner: compat
+  doc: repos/afd-plugin/rules.md
+- prefix: afd_plugin/config.py
+  owner: config-validation
+  doc: repos/afd-plugin/rules.md
+- prefix: afd_plugin/validation.py
+  owner: config-validation
+  doc: repos/afd-plugin/rules.md
+- prefix: afd_plugin/config_utils.py
+  owner: config-validation
+  doc: repos/afd-plugin/rules.md
+- prefix: afd_plugin/connectors/
+  owner: connectors-distributed
+  doc: repos/afd-plugin/rules.md
+- prefix: afd_plugin/distributed/
+  owner: connectors-distributed
+  doc: repos/afd-plugin/rules.md
+- prefix: afd_plugin/v1/worker/
+  owner: workers-runners
+  doc: repos/afd-plugin/rules.md
+- prefix: afd_plugin/model_executor/
+  owner: workers-runners
+  doc: repos/afd-plugin/rules.md
+- prefix: csrc/gpu/
+  owner: native-backends
+  doc: repos/afd-plugin/rules.md
+- prefix: csrc/npu/
+  owner: native-backends
+  doc: repos/afd-plugin/rules.md
+- prefix: tests/unit/
+  owner: tests
+  doc: repos/afd-plugin/rules.md
+- prefix: tests/e2e/
+  owner: tests
+  doc: repos/afd-plugin/rules.md
+
+validation:
+  test_manifest_sources: [.github/workflows/]
+  precommit: true
+
+ci:
+  provider: github_actions
+
+push:
+  default_remote: origin
+  protected_branches: [main, master]
+  allowed: false

adapters/vllm_omni/manifest.yaml

@@ -11,6 +11,8 @@ repo:
   # takes precedence; this is only the fallback when no .env is in reach. Never
   # hardcode a machine-specific absolute path in this committed file.
   path: ${VLLM_OMNI_REPO}
+  full_name: vllm-project/vllm-omni
+  aliases: [vllm-omni, vllm_omni]
   default_branch: main
   rebase_branch: dev/vllm-align
   language: python
@@ -66,6 +68,26 @@ modules:
     local_paths: [benchmarks/]
     wave: 2
 
+review_routes:
+- prefix: vllm_omni/config/
+  owner: config
+  doc: repos/vllm-omni/components/configuration/rules.md
+- prefix: vllm_omni/core/
+  owner: scheduler
+  doc: repos/vllm-omni/components/scheduler/rules.md
+- prefix: vllm_omni/entrypoints/
+  owner: serving
+  doc: repos/vllm-omni/components/serving/rules.md
+- prefix: vllm_omni/model_executor/
+  owner: model-executor
+  doc: repos/vllm-omni/components/model-executor/rules.md
+- prefix: vllm_omni/models/
+  owner: models
+  doc: repos/vllm-omni/models/_index.md
+- prefix: vllm_omni/diffusion/
+  owner: diffusion
+  doc: repos/vllm-omni/components/diffusion/rules.md
+
 validation:
   test_manifest_sources: [.buildkite/]
   precommit: true

knowledge/SCHEMA.md

@@ -35,7 +35,7 @@ contradictions: [相对路径]         # 可选：与本页冲突的页面
 
 新标签必须先加入此表再使用（防止标签蔓延）：
 
-- 归属：`general`、`vllm-omni`
+- 归属：`general`、`vllm-omni`、`afd-plugin`
 - 工作主题：`review`、`ci`、`docs`、`git`、`debug`、`benchmark`、`environment`、
   `remote`、`agents`、`planning`、`dev`、`rebase`
 - 代码/模型轴：`components`、`models`、`diffusion`、`model-executor`、`serving`、

knowledge/repos/_index.md

@@ -3,7 +3,7 @@ title: "仓库经验入口"
 created: 2026-07-10
 updated: 2026-07-10
 type: index
-tags: [vllm-omni]
+tags: [vllm-omni, afd-plugin]
 sources: []
 ---
 
@@ -24,4 +24,5 @@ sources: []
 | 仓库 | 上游地址 | 查看哪里 |
 |---|---|---|
 | vLLM-Omni | `vllm-project/vllm-omni` | [vllm-omni](vllm-omni/_index.md) |
+| afd-plugin | `vllm-project/afd-plugin` | [afd-plugin](afd-plugin/_index.md) |
 | Jianghan roleplay data pipeline | 私有项目 | [jianghan-roleplay-data-pipeline](jianghan-roleplay-data-pipeline/_index.md) |

knowledge/repos/afd-plugin/_index.md

@@ -0,0 +1,34 @@
+---
+title: "afd-plugin"
+created: 2026-08-05
+updated: 2026-08-05
+type: index
+tags: [afd-plugin]
+sources: ["vllm-project/afd-plugin:AGENTS.md", "vllm-project/afd-plugin:README.md", "vllm-project/afd-plugin:pyproject.toml", "vllm-project/afd-plugin:.github/workflows/cpu-only-ci.yml"]
+---
+
+# afd-plugin
+
+- 上游仓库：`vllm-project/afd-plugin`
+- 默认分支：`main`
+- 适用范围：AFD 插件的兼容 patch、配置校验、connector/distributed、worker/runner、GPU/NPU native backend、单测和硬件 E2E review
+- 仓库自己的 `AGENTS.md`、`CLAUDE.md`、`.agents/skills/run-e2e/SKILL.md`、CI 和脚本保持权威；这里仅保存 InferMatrixCopilot 的最小路由规则
+
+## 什么时候查这里
+
+- 当前 Git 仓库、GitHub URL 或 PR 目标是 `vllm-project/afd-plugin`。
+- 需要判断 AFD PR 的 changed files 应读哪些仓库专属 review 规则。
+
+## 不放什么
+
+- 不放 vLLM-Omni 的模型、远端、benchmark 或 rebase 经验。
+- 不复制 AFD 仓库完整 README、AGENTS 或 E2E skill；运行细节以目标仓库当前文件为准。
+- 不记录本机 checkout 路径、硬件账号、模型 cache 或凭据。
+
+## 当前入口
+
+| 遇到什么 | 查看哪里 | 说明 |
+|---|---|---|
+| 开始任何 afd-plugin review、issue answer 或仓库初始化 | [硬门禁](rules.md) | AFD 仓库级默认规则 |
+| 需要跨仓库 review 基础方法 | [general review](../../general/review/_index.md) | 通用审查入口 |
+

knowledge/repos/afd-plugin/rules.md

@@ -0,0 +1,40 @@
+---
+title: "afd-plugin 硬门禁"
+created: 2026-08-05
+updated: 2026-08-05
+type: rule
+tags: [afd-plugin, review, config, distributed, model-executor]
+sources: ["vllm-project/afd-plugin:AGENTS.md", "vllm-project/afd-plugin:README.md", "vllm-project/afd-plugin:pyproject.toml", "vllm-project/afd-plugin:.github/workflows/cpu-only-ci.yml", "vllm-project/afd-plugin:.agents/skills/run-e2e/SKILL.md"]
+---
+
+# afd-plugin 硬门禁
+
+只在当前任务明确属于 `vllm-project/afd-plugin` 时应用本页。先按 [afd-plugin 入口](_index.md) 确认仓库身份，再结合 [通用 review 入口](../../general/review/_index.md) 做常规审查。
+
+## 1. 仓库身份和权威来源
+
+- AFD-1a：review 或 issue 回答前必须确认目标是 canonical `vllm-project/afd-plugin`，不能把 vLLM-Omni 的知识树、模型规则、远端策略或 rebase 假设带入 AFD。
+- AFD-1b：AFD 仓库自己的 `AGENTS.md`、`CLAUDE.md`、`.agents/skills/run-e2e/SKILL.md`、CI workflow 和仓库脚本是权威来源；InferMatrixCopilot 只做只读路由、提示和证据组织。
+- AFD-1c：默认路径保持只读；不能自动发 PR 评论、push、改 protected branch，不能要求 AFD runtime 行为因为本 adapter 初始化而改变。
+
+## 2. 兼容范围和插件边界
+
+- AFD-2a：兼容结论必须精确写当前声明的 runtime：vLLM `0.19.1`；Ascend NPU baseline 是 vLLM-Ascend `0.19.1rc1`、CANN/NNAL `8.5.1`、torch/torch-npu `2.9.0`。不得推断更宽版本范围。
+- AFD-2b：AFD 通过 `vllm.general_plugins` entry point、`VLLM_PLUGINS`、`additional_config["afd"]`、自动 worker 选择、plugin-owned model wrapper 和窄 compatibility shim 生效；不得建议新增独立 `--afd-config` 之类的公开入口。
+- AFD-2c：`additional_config["afd"]` 之外的 connector 私有字段必须进入 `connector_extra_config`，并由对应 connector parser 校验；未知 AFD 顶层字段应暴露为错误，而不是静默透传。
+- AFD-2d：没有 `additional_config["afd"]` 时必须保持非 AFD 路径隔离；非 AFD 请求、worker、模型和上游 vLLM 行为不能被 AFD patch 或自动选择污染。
+
+## 3. Patch、connector 和 worker review
+
+- AFD-3a：`afd_plugin/compat/` patch 必须 upstream-first：复制目标 tag 的上游函数，保持函数签名和返回类型，使用 `# ### PATCH START` / `# ### PATCH END` 标出 AFD 差异，并写清 patch reason、功能变化和移除或 upstream 计划。
+- AFD-3b：不得用宽泛 `getattr`、`hasattr`、`Any`、`object` 或 `_original_*` delegation 掩盖上游 API 漂移；只有 AGENTS 明确允许的例外才可使用，并必须在 patch 注释说明原因。
+- AFD-3c：`afd_plugin/connectors/`、`afd_plugin/distributed/` 改动必须审 rank 布局、attention/ffn 数量约束、同步/异步语义、端口/host、进程组生命周期、资源 cleanup 和跨 role 同步失败路径。
+- AFD-3d：`afd_plugin/v1/worker/`、`afd_plugin/model_executor/` 改动必须审自动 worker 选择、role-specific runner、Attention/FFN 组件加载边界、scheduler-driven FFN fail-fast、CUDA/ACL graph 模式和 DBO/ubatch 条件。
+- AFD-3e：`csrc/gpu/`、`csrc/npu/` 改动必须区分 CUDA 与 Ascend CANN/ACLNN toolchain；NPU op 构建只能在明确 Ascend 环境或显式 `AFD_BUILD_ASCEND_OPS=1` 下作为证据。
+
+## 4. 验证和证据
+
+- AFD-4a：CPU-safe 默认检查是 `uv run pytest`、`uv run pytest -q tests/unit -m "not gpu and not vllm_runtime"`、`uv run ruff check .`、`uv run ruff format --check .`；它们不能证明 GPU/NPU runtime、模型精度或性能。
+- AFD-4b：GPU/NPU E2E 必须通过 AFD 仓库 `.agents/skills/run-e2e/SKILL.md` 或仓库测试入口运行，记录 backend、设备数、模型路径、vLLM/vLLM-Ascend/CANN/torch 版本、拓扑、connector、graph/DBO/async 配置和 skipped 原因。
+- AFD-4c：性能、精度或硬件支持 claim 必须给可复现证据：硬件型号、软件版本、driver/toolchain、rank 拓扑、命令、日志或 pytest node id；CPU import/config smoke 不能替代硬件证据。
+- AFD-4d：changed files 没有命中 adapter route 时，必须把 unmatched path 明确留给 host reviewer 继续人工审，不得把它当作已覆盖。

src/infermatrix_copilot/adapters/base.py

@@ -108,6 +108,25 @@ def repo_path(self) -> str:
         (`${VAR}` / `~`) instead of hardcoding one machine's absolute layout."""
         return expand_path(self.manifest.get("repo", {}).get("path", ""))
 
+    @property
+    def aliases(self) -> set[str]:
+        """Names that may identify this adapter at the boundary.
+
+        The manifest's canonical `name` remains the primary id, but callers may
+        receive dashed aliases, GitHub `owner/repo` strings, or explicit aliases
+        declared in `repo.aliases`. Matching is case-insensitive.
+        """
+        repo = self.manifest.get("repo", {}) or {}
+        values = {self.name, self.name.replace("_", "-")}
+        values.update(str(alias) for alias in repo.get("aliases", []) or [])
+        if repo.get("full_name"):
+            values.add(str(repo["full_name"]))
+        return {value.casefold() for value in values if value}
+
+    def matches_name(self, name: str) -> bool:
+        """Return true when `name` is this adapter's canonical name or alias."""
+        return str(name).casefold() in self.aliases
+
     @property
     def protected_branches(self) -> list[str]:
         """Branches a push may never target (defaults to `["main"]`) — read by the
@@ -129,6 +148,15 @@ def module_for_path(self, path: str) -> str | None:
                     return module
         return None
 
+    @property
+    def review_routes(self) -> list[dict]:
+        """Changed-file review routes declared by the manifest.
+
+        Each item is data only, typically `{prefix, doc, owner}`. Runtime code
+        validates scope before reading the referenced docs.
+        """
+        return list(self.manifest.get("review_routes") or [])
+
     @property
     def high_risk_modules(self) -> list[str]:
         """Modules declared `risk: high` — feeds the patch-review trigger."""
@@ -284,7 +312,7 @@ def resolve(self, *, name: str | None = None, repo_path: str | None = None) -> R
         adapters = self.all()
         if name:
             for p in adapters:
-                if p.name == name:
+                if p.matches_name(name):
                     return p
             raise AdapterError(f"no adapter named {name!r}")
         if repo_path:
@@ -301,7 +329,8 @@ def _git(repo: Path, *args: str) -> str:
     """Run `git <args>` in `repo` and return its stripped stdout — the read-only
     git helper the deterministic fingerprint uses."""
     out = subprocess.run(["git", *args], cwd=str(repo), capture_output=True,
-                         text=True, encoding="utf-8", errors="replace", timeout=30)
+                         text=True, encoding="utf-8", errors="replace", timeout=30,
+                         check=False)
     return out.stdout.strip()