Add /imupdate release knowledge workflow

.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
     {
       "name": "infermatrix-copilot",
       "source": "./plugins/infermatrix-copilot",
-      "description": "Maintainer-grounded vLLM-Omni review knowledge over MCP."
+      "description": "InferMatrixCopilot — governed PR review and release-driven knowledge maintenance over MCP and skills."
     }
   ]
 }

.cursor-plugin/marketplace.json

@@ -10,7 +10,7 @@
     {
       "name": "infermatrix-copilot",
       "source": "plugins/infermatrix-copilot",
-      "description": "Maintainer-grounded vLLM-Omni review knowledge over MCP."
+      "description": "Maintainer-grounded vLLM-Omni review and release knowledge updates."
     }
   ]
 }

.github/workflows/vllm-omni-release-audit.yml

@@ -7,6 +7,7 @@ on:
       - adapters/vllm_omni/manifest.yaml
       - adapters/vllm_omni/release_baseline.yaml
       - doc/VLLM_OMNI_RELEASE_MAINTENANCE.md
+      - plugins/infermatrix-copilot/skills/imupdate/**
       - knowledge/repos/vllm-omni/**
       - tools/audit_vllm_omni_release.py
       - tools/vllm_omni_release_audit.py

README.md

@@ -15,7 +15,7 @@ vLLM-Omni 的兼容性、正确性和性能问题。
 ## 快速开始
 
 私有仓库过渡期，克隆后运行一个入口。脚本会自动识别本机的 Codex、Claude Code
-和 Cursor，并同时安装 MCP 与 `imreview` Skill。
+和 Cursor，并同时安装 MCP、`imreview` 和 `imupdate` Skill。
 
 ```text
 git clone git@github.com:JiusiServe/InferMatrixCopilot.git
@@ -35,9 +35,13 @@ install.cmd
 
 ```text
 /imreview https://github.com/vllm-project/vllm-omni/pull/5172
+/imupdate D:\path\to\vllm-omni
+/imupdate vllmomni
 ```
 
 不带地址时，`/imreview` 会审查当前 PR 或本地工作区。
+`/imupdate` 接受本地 Git 路径，也接受仓库名、别名或 URL。本地路径直接走机器审计；
+名字或别名先由 Agent 找到权威仓库和目标版本，再尽量获取临时 checkout 运行机器审计。
 
 公共市场上线后的安装方法和通用 MCP 配置见 [`doc/MCP.md`](doc/MCP.md)。
 
@@ -49,6 +53,13 @@ install.cmd
   → InferMatrixCopilot 返回知识库入口
   → Agent 按改动选择相关模型、组件和通用规则
   → Agent 输出带文件和行号的审查结论
+
+你发起 /imupdate <仓库>
+  → 本地路径：直接读取 baseline 和目标仓库 HEAD
+  → 仓库名/别名：Agent 先解析权威仓库、版本和可用 checkout
+  → 只读审计器对比 Git 对象、注册表、pipeline、deploy 和路径路由
+  → Agent 根据报告更新结构事实，不自动生成 owner 规则
+  → enforce、知识检查和相关测试确认更新完整
 ```
 
 代码理解和推理由 Agent 当前模型完成。InferMatrixCopilot 只负责提供知识地图和维护者

doc/MCP.md

@@ -3,7 +3,7 @@
 InferMatrixCopilot 的核心发布物保持一份：
 
 - `plugins/infermatrix-copilot/.mcp.json`：标准 stdio MCP 描述；
-- `plugins/infermatrix-copilot/skills/imreview/SKILL.md`：开放 Agent Skills；
+- `plugins/infermatrix-copilot/skills/`：开放 Agent Skills；
 - Claude、Codex、Cursor 的市场文件只负责把同一个插件展示出来，不包含各自的安装逻辑。
 
 ## 私有仓库过渡安装
@@ -32,6 +32,7 @@ install.cmd
 
 ```text
 /imreview <PR URL>
+/imupdate <local path, repository name, alias, or URL>
 ```
 
 市场只安装同一份 MCP 与 Skill；不会把逐 Agent 逻辑带回核心包。
@@ -44,6 +45,7 @@ install.cmd
 
 ```text
 plugins/infermatrix-copilot/skills/imreview/SKILL.md
+plugins/infermatrix-copilot/skills/imupdate/SKILL.md
 ```
 
 核心 MCP 配置：
@@ -71,6 +73,8 @@ plugins/infermatrix-copilot/skills/imreview/SKILL.md
 
 ```text
 /imreview https://github.com/vllm-project/vllm-omni/pull/5172
+/imupdate D:\path\to\vllm-omni
+/imupdate vllmomni
 ```
 
 默认 Direct 模式只提供知识库，由 Agent 当前模型完成审查；不会自动评论或推送代码。

doc/VLLM_OMNI_RELEASE_MAINTENANCE.md

@@ -4,7 +4,29 @@ InferMatrixCopilot 的模型清单、代码 owner 路由和源码引用会随着
 vLLM-Omni 发版而过期。这个审计只读取两个 Git 提交和本仓库声明，不会 checkout
 目标版本，也不会自动修改知识规则。
 
-## 本地运行
+## 用户入口
+
+安装 InferMatrixCopilot 后，在 Codex、Claude Code 或 Cursor 中运行：
+
+```text
+/imupdate D:\path\to\vllm-omni
+/imupdate vllm-omni
+/imupdate vllmomni
+```
+
+支持三种用法：
+
+1. 本地 Git 路径：直接用 baseline 的 `audited_sha` 对比仓库 `HEAD`，机器审计后
+   由 Agent 更新知识。
+2. 仓库名、别名或 URL：Agent 先解析权威仓库与目标版本，寻找已配置 checkout
+   或建立临时 clone，再尽量进入同一机器审计流程；无法运行机器审计时必须明确标记
+   为模型分析，不能宣称 `CLEAN`。选择版本时不得因为 GitHub “latest release”
+   忽略预发布版本而退回到比当前 baseline 更旧的版本。
+3. 下方完整 Python 命令：纯机器规则对比，不调用 LLM，也不修改文件。
+
+`/imupdate` 的可选第二个参数可以指定目标 tag 或 SHA。
+
+## 底层命令
 
 先在 vLLM-Omni checkout 中 fetch 需要比较的 tag 或 SHA，然后运行：
 
@@ -30,6 +52,13 @@ python tools/audit_vllm_omni_release.py `
 默认 `--mode enforce`：存在未解释漂移时退出 1，输入或 Git 失败时退出 2。
 `--mode report-only` 仍报告 `DRIFT`，但退出 0，供定时巡检使用。
 
+底层实现不 checkout 版本：它用 `git rev-parse` 固定两个提交，用
+`git diff --name-status -M` 找新增、修改、删除和重命名，再用
+`git show <sha>:<path>` 直接读取 Git 对象。Python AST 负责提取模型和
+pipeline registry，deploy YAML 直接按提交列举；结果做排序和哈希后与 baseline、
+owner 路由、manifest、知识 `sources:` 和 pin 对账，最后输出稳定 JSON。
+这个命令只报告证据，不编辑任何文件。
+
 ## 更新一个 release
 
 1. 用当前 `release_baseline.yaml` 的 audited SHA 作为 `--from`，新 tag 作为

plugins/infermatrix-copilot/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "infermatrix-copilot",
   "version": "0.1.0",
-  "description": "Maintainer-grounded vLLM-Omni review knowledge over MCP.",
+  "description": "Maintainer-grounded vLLM-Omni review and release knowledge updates.",
   "author": {
     "name": "JiusiServe",
     "url": "https://github.com/JiusiServe"

plugins/infermatrix-copilot/.codex-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "infermatrix-copilot",
   "version": "0.1.0",
-  "description": "Maintainer-grounded vLLM-Omni review knowledge over MCP.",
+  "description": "Maintainer-grounded vLLM-Omni review and release knowledge updates.",
   "author": {
     "name": "JiusiServe",
     "url": "https://github.com/JiusiServe"
@@ -18,17 +18,19 @@
   "mcpServers": "./.mcp.json",
   "interface": {
     "displayName": "InferMatrixCopilot",
-    "shortDescription": "Review vLLM-Omni changes with maintainer knowledge",
-    "longDescription": "Use maintainer-grounded vLLM-Omni knowledge to review pull requests and local changes.",
+    "shortDescription": "Review and update vLLM-Omni knowledge",
+    "longDescription": "Use maintainer-grounded vLLM-Omni knowledge to review changes and update release-driven facts.",
     "developerName": "JiusiServe",
     "category": "Developer Tools",
     "capabilities": [
       "MCP",
-      "Code review"
+      "Code review",
+      "Knowledge maintenance"
     ],
     "websiteURL": "https://github.com/JiusiServe/InferMatrixCopilot",
     "defaultPrompt": [
-      "Review this PR with InferMatrixCopilot."
+      "Review this PR with InferMatrixCopilot.",
+      "Update InferMatrixCopilot knowledge from this repository."
     ]
   }
 }

plugins/infermatrix-copilot/.cursor-plugin/plugin.json

@@ -2,7 +2,7 @@
   "name": "infermatrix-copilot",
   "displayName": "InferMatrixCopilot",
   "version": "0.1.0",
-  "description": "Maintainer-grounded vLLM-Omni review knowledge over MCP.",
+  "description": "Maintainer-grounded vLLM-Omni review and release knowledge updates.",
   "author": {
     "name": "JiusiServe"
   },

plugins/infermatrix-copilot/skills/imupdate/SKILL.md

@@ -0,0 +1,89 @@
+---
+name: imupdate
+description: Resolve an upstream repository from a local path, repository name, alias, or URL, then update InferMatrixCopilot's release-driven knowledge. Use when the user invokes /imupdate or $imupdate, asks to sync knowledge after an upstream release, compare vLLM-Omni tags or SHAs, or refresh the release baseline, model catalog, source maps, and manifests.
+---
+
+# InferMatrix update
+
+```text
+/imupdate <path-or-repository> [target-tag-or-sha]
+```
+
+The installed InferMatrixCopilot root is:
+
+```text
+{{INFERMATRIX_COPILOT_ROOT}}
+```
+
+If that placeholder was not replaced, locate or temporarily clone the
+InferMatrixCopilot checkout containing `tools/audit_vllm_omni_release.py`.
+
+## Select the input mode
+
+### Local Git path
+
+When the argument exists and is a Git checkout, use the deterministic audit.
+Read `adapters/vllm_omni/release_baseline.yaml`; use
+`upstream.audited_sha` as the old revision and the optional target or checkout
+`HEAD` as the new revision. Refuse an accidental downgrade when the target is
+older than the audited revision. Do not fetch or checkout unless explicitly
+asked.
+
+### Repository name, alias, or URL
+
+When the argument is not a local path, use the host model to resolve it:
+
+1. Treat `vllm-omni`, `vllmomni`, `vllm omni`, and
+   `vllm-project/vllm-omni` as `https://github.com/vllm-project/vllm-omni`.
+2. Check the configured repo map and current workspace for a matching checkout.
+   If found, continue as the local-path mode.
+3. Otherwise verify the canonical repository from authoritative GitHub
+   metadata. Choose the supplied target, otherwise the newest release or tag
+   that is not older than the audited revision, otherwise the default-branch
+   HEAD; state that choice. Never silently downgrade because GitHub's "latest
+   release" can exclude newer prereleases.
+4. Prefer a temporary clone and run the deterministic audit. If repository
+   access or cloning is unavailable, inspect the authoritative source with the
+   host model, update only source-backed facts, and clearly report that machine
+   audit/enforce validation was unavailable. Never call that result `CLEAN`.
+5. Ask for clarification only when multiple plausible canonical repositories
+   remain after resolution.
+
+## Audit and update
+
+1. Resolve old and new revisions to full SHAs.
+2. Create a temporary JSON path and run the read-only audit in report mode when
+   a local or temporary Git checkout is available:
+
+   ```text
+   python <infermatrix-root>/tools/audit_vllm_omni_release.py \
+     --from <audited-sha> \
+     --to <target-sha> \
+     --repo <upstream-repo> \
+     --mode report-only \
+     --json-output <temporary-json>
+   ```
+
+3. Explain the registry, pipeline, deploy, path-routing, source, and pin drift.
+   Read the relevant upstream diff before changing InferMatrixCopilot.
+4. Update only facts proven by the report and source: the release baseline,
+   model catalog, component source maps, pins, or adapter manifest. Never
+   generate or rewrite owner rules automatically; report semantic-rule
+   candidates for human review.
+5. When machine audit is available, run the same command with `--mode enforce`,
+   then run:
+
+   ```text
+   python knowledge/tools/check_knowledge_tree.py
+   python knowledge/tools/check_wiki_lint.py
+   python -m pytest test/test_release_audit.py test/test_adapters.py \
+     test/test_routing.py test/test_knowledge_source.py \
+     test/test_skills_scope.py test/test_capabilities.py -q
+   ```
+
+6. Delete temporary reports and clones. Return the canonical repository, input
+   mode, old and new SHAs, structural deltas, files changed, validation
+   results, and any unresolved or model-only findings.
+
+The audit is evidence collection, not the updater. Codex performs the bounded
+knowledge edits and must not commit, push, or open a PR unless asked.

plugins/infermatrix-copilot/skills/imupdate/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "InferMatrix Update"
+  short_description: "Audit and update release-driven repository knowledge"
+  default_prompt: "Use $imupdate to resolve a repository path or name and update InferMatrix knowledge."

scripts/install_mcp.py

@@ -15,9 +15,7 @@
 
 
 PROJECT_ROOT = Path(__file__).resolve().parent.parent
-SKILL_SOURCE = (
-    PROJECT_ROOT / "plugins" / "infermatrix-copilot" / "skills" / "imreview"
-)
+SKILLS_SOURCE = PROJECT_ROOT / "plugins" / "infermatrix-copilot" / "skills"
 SERVER_NAME = "infermatrix-copilot"
 PACKAGE = (
     "infermatrix-copilot[mcp] @ "
@@ -46,9 +44,23 @@ def _run_quiet(command: Sequence[str]) -> None:
     )
 
 
-def _install_skill(destination: Path) -> None:
-    destination.parent.mkdir(parents=True, exist_ok=True)
-    shutil.copytree(SKILL_SOURCE, destination, dirs_exist_ok=True)
+def _install_skills(destination_root: Path) -> None:
+    destination_root.mkdir(parents=True, exist_ok=True)
+    for source in sorted(path for path in SKILLS_SOURCE.iterdir() if path.is_dir()):
+        destination = destination_root / source.name
+        shutil.copytree(source, destination, dirs_exist_ok=True)
+        for path in destination.rglob("*"):
+            if not path.is_file():
+                continue
+            content = path.read_text(encoding="utf-8")
+            if "{{INFERMATRIX_COPILOT_ROOT}}" in content:
+                path.write_text(
+                    content.replace(
+                        "{{INFERMATRIX_COPILOT_ROOT}}",
+                        str(PROJECT_ROOT),
+                    ),
+                    encoding="utf-8",
+                )
 
 
 def _install_codex(config_root: Path) -> None:
@@ -60,7 +72,7 @@ def _install_codex(config_root: Path) -> None:
         [codex, "mcp", "add", SERVER_NAME, "--", *SERVER_COMMAND],
         "Codex MCP registration failed.",
     )
-    _install_skill(config_root / ".codex" / "skills" / "imreview")
+    _install_skills(config_root / ".codex" / "skills")
 
 
 def _install_claude(config_root: Path) -> None:
@@ -83,7 +95,7 @@ def _install_claude(config_root: Path) -> None:
         ],
         "Claude Code MCP registration failed.",
     )
-    _install_skill(config_root / ".claude" / "skills" / "imreview")
+    _install_skills(config_root / ".claude" / "skills")
 
 
 def _install_cursor(config_root: Path) -> None:
@@ -120,18 +132,15 @@ def _install_cursor(config_root: Path) -> None:
         json.dumps(config, ensure_ascii=False, indent=2) + "\n",
         encoding="utf-8",
     )
-    _install_skill(cursor_root / "skills" / "imreview")
+    _install_skills(cursor_root / "skills")
 
 
 def _cursor_installed(config_root: Path) -> bool:
     if shutil.which("cursor") or shutil.which("cursor-agent"):
         return True
     candidates = [
         config_root / ".cursor",
-        Path(os.environ.get("LOCALAPPDATA", ""))
-        / "Programs"
-        / "cursor"
-        / "Cursor.exe",
+        Path(os.environ.get("LOCALAPPDATA", "")) / "Programs" / "cursor" / "Cursor.exe",
         Path("/Applications/Cursor.app"),
     ]
     return any(path.exists() for path in candidates)
@@ -201,7 +210,7 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
         output = PROJECT_ROOT / "infermatrix-copilot.mcp.json"
         _write_generic_config(output)
         print(f"No known Agent detected. MCP config written to:\n  {output}")
-        print(f"Portable Skill:\n  {SKILL_SOURCE}")
+        print(f"Portable Skills:\n  {SKILLS_SOURCE}")
         return 0
 
     failures = []
@@ -223,7 +232,7 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
             print(f"  {failure}", file=sys.stderr)
         return 1
 
-    print("Restart your Agent, then run: /imreview <PR URL>")
+    print("Restart your Agent, then run: /imreview <PR URL> or /imupdate <repository>")
     return 0