feat(acp): visual env editor with per-endpoint key suggestions

mateclaw-server/src/main/java/vip/mate/acp/service/AcpRuntimeSupport.java

@@ -84,22 +84,6 @@ public class AcpRuntimeSupport {
      * and the cost of a false positive (a slightly more verbose error
      * banner) is much smaller than a false negative (user staring at a
      * raw JSON-RPC blob).
-     *
-     * <p>Special case: a claude-code endpoint returning {@code 403
-     * "Request not allowed"} is almost always the keychain-hijack
-     * scenario rather than a wrong API key. The third-party
-     * {@code @zed-industries/claude-agent-acp} package wraps
-     * {@code @anthropic-ai/claude-agent-sdk}, whose auth dispatcher
-     * checks the macOS keychain ({@code Claude Code-credentials}) /
-     * {@code ~/.claude/credentials.json} BEFORE the
-     * {@code ANTHROPIC_API_KEY} env var. So a host that's done
-     * {@code claude login} silently shadows whatever API key the user
-     * configured in the endpoint env, and Anthropic's API rejects the
-     * subscription OAuth token (first-party-only) with the very
-     * specific {@code "Request not allowed"} error string. We detect
-     * that exact combination and surface the keychain-clearing remedy
-     * instead of the generic "set ANTHROPIC_API_KEY" hint, which
-     * doesn't apply here.
      */
     public String translateAuthError(AcpEndpointEntity endpoint, String originalMessage) {
         if (originalMessage == null) return null;
@@ -115,32 +99,8 @@ public class AcpRuntimeSupport {
                 || lower.contains("authenticate");
         if (!looksLikeAuth) return null;
 
-        String name = endpoint != null && endpoint.getName() != null ? endpoint.getName() : "(unknown)";
-        String slug = lower(name);
-        String command = endpoint != null ? lower(endpoint.getCommand()) : "";
-
-        // Keychain-hijack detection — must come before the generic env-
-        // missing branch because both would superficially match.
-        boolean keychainHijack = lower.contains("request not allowed")
-                && (slug.contains("claude") || command.contains("claude-agent-acp"));
-        if (keychainHijack) {
-            StringBuilder sb = new StringBuilder();
-            sb.append("ACP endpoint '").append(name).append("' upstream auth failed with ");
-            sb.append("'Request not allowed' — almost always means the host CLI's OAuth ");
-            sb.append("credentials are hijacking the SDK auth path. ");
-            sb.append("The Claude Agent SDK reads ~/.claude/ / macOS keychain BEFORE the ");
-            sb.append("ANTHROPIC_API_KEY env var, so the API key you configured here is ");
-            sb.append("never sent — Anthropic rejects the subscription OAuth token because ");
-            sb.append("third-party processes aren't allowed to use it. ");
-            sb.append("To fix: ");
-            sb.append("(macOS) run `claude logout`, or `security delete-generic-password ");
-            sb.append("-s \"Claude Code-credentials\"`; ");
-            sb.append("(Linux / Windows) delete ~/.claude/credentials.json. ");
-            sb.append("Then click Test connection again. Original: ").append(originalMessage);
-            return sb.toString();
-        }
-
         String envVar = expectedAuthEnvVar(endpoint);
+        String name = endpoint != null && endpoint.getName() != null ? endpoint.getName() : "(unknown)";
         StringBuilder sb = new StringBuilder();
         sb.append("ACP endpoint '").append(name).append("' upstream auth failed. ");
         sb.append("Most likely the endpoint env has no API key. ");

mateclaw-server/src/main/java/vip/mate/agent/binding/service/AgentBindingService.java

@@ -228,41 +228,25 @@ public class AgentBindingService {
      * {@link #getEffectiveToolNames} allowlist completely.
      */
     private static final Set<String> SYSTEM_LEVEL_TOOLS = Set.of(
-            // Structured memory primitives — used by every agent regardless
+            // Memory write/read primitives — every agent needs these
-            // of skill bindings, otherwise the self-evolution path collapses
+            // regardless of skill bindings, otherwise the self-evolution
-            // (§11.3 / §11.4).
+            // path collapses (§11.3 / §11.4).
             "record_lesson",
             "remember",
             "remember_structured",
             "recall_structured",
             "forget_structured",
-            // Workspace memory file CRUD (PROFILE.md / MEMORY.md / SOUL.md /
+            // Workspace memory file CRUD (PROFILE.md / MEMORY.md / SOUL.md)
-            // memory/YYYY-MM-DD.md). Prior versions whitelisted
+            "read_workspace_file",
-            // "read_workspace_file" / "write_workspace_file" /
+            "write_workspace_file",
-            // "list_workspace_files" — those names match no @Tool bean; the
+            "list_workspace_files",
-            // actual function names carry the "_memory" segment, so the
-            // earlier carve-out was silently dead.
-            "list_workspace_memory_files",
-            "read_workspace_memory_file",
-            "write_workspace_memory_file",
-            "edit_workspace_memory_file",
             // Skill discovery / dispatch — skills are docs, not callables;
             // these helpers let the LLM read SKILL.md / run scripts.
             "readSkillFile",
             "runSkillScript",
-            "listSkillFiles",
+            // Date/time + delegate — fundamental cross-skill utilities
-            "listAvailableSkills",
+            "datetime",
-            // Date / time — prior whitelist had a fictional "datetime"; the
+            "delegate_agent",
-            // real DateTimeTool exposes three separate methods.
-            "getCurrentDate",
-            "getCurrentDateTime",
-            "getCurrentTime",
-            // Multi-agent delegation — prior whitelist had "delegate_agent",
-            // but DelegateAgentTool's @Tool methods are delegateToAgent /
-            // delegateParallel / listAvailableAgents. Same dead-name bug.
-            "delegateToAgent",
-            "delegateParallel",
-            "listAvailableAgents",
             // Document / media generation — agent-wide capabilities, never
             // declared inside any skill manifest. Pre-Phase-2b these were
             // universally visible; the new gate silently strips them whenever
@@ -276,26 +260,7 @@ public class AgentBindingService {
             "renderDocxFromFiles",
             "image_generate",
             "music_generate",
-            "video_generate",
+            "video_generate"
-            // Universal capabilities the global system prompts (SOUL.md /
-            // AGENTS.md / "Web Search Capability" / "File Reading Guidelines")
-            // explicitly tell the LLM exist. Pre-Phase-2b they were globally
-            // available; the new gate silently hid them on any agent with
-            // skills bound, so the prompt promises a tool the registry then
-            // refuses ("Tool not found: search"). Observed 2026-05-01 on the
-            // Code Reviewer agent — the model called search → got
-            // not-found → gave up before ever reaching renderDocx.
-            "search",
-            "browser_use",
-            "read_file",
-            "write_file",
-            "edit_file",
-            "execute_shell_command",
-            "detect_file_type",
-            "extract_document_text",
-            "extract_pdf_text",
-            "extract_docx_text",
-            "readMateClawDoc"
     );
 
     private ResolvedSkill findResolvedSkillById(Long skillId) {

mateclaw-ui/src/i18n/locales/en-US.ts

@@ -1168,7 +1168,7 @@ export default {
         openai: 'OpenAI API key (sk-…). Get one at https://platform.openai.com/api-keys',
         dashscope: 'Aliyun DashScope key — same value as MateClaw .env DASHSCOPE_API_KEY',
         google: 'Google AI Studio API key. Get one at https://aistudio.google.com/apikey',
-        claudeOauth: 'Heads up: claude-code OAuth login (the one stored in ~/.claude/) does NOT work here, and worse — the Claude Agent SDK reads it BEFORE this env var, silently shadowing your API key. If you have ever run `claude login` on this host, clear the keychain first: (macOS) `claude logout` or `security delete-generic-password -s "Claude Code-credentials"`; (Linux/Windows) delete ~/.claude/credentials.json. Then put a real API key (sk-ant-…) above. Or switch to qwen-code / opencode to reuse a CLI login.',
+        claudeOauth: 'Heads up: claude-code OAuth login (the one stored in ~/.claude/) does NOT work here. The third-party Zed wrapper uses the public Anthropic API, which only accepts a real API key. Get one at https://console.anthropic.com/settings/keys, or switch to qwen-code / opencode if you want to reuse a CLI login.',
       },
     },
     modal: {

mateclaw-ui/src/i18n/locales/zh-CN.ts

@@ -277,7 +277,7 @@ export default {
     mcpConnections: 'MCP 连接',
     toolsCatalog: '工具目录',
     activity: '活动记录',
-    acpEndpoints: 'ACP 端点',
+    acpEndpoints: 'ACP Endpoints',
     settingsGroup: '设置',
     agents: '智能体',
     security: '安全',
@@ -1121,23 +1121,23 @@ export default {
   },
   acp: {
     kicker: '外部 Agent',
-    title: 'ACP 端点',
+    title: 'ACP Endpoints',
     desc: '通过 stdio 委派给外部编码 Agent（codex / claude-code / opencode / qwen-code）。安装对应 CLI 后开启该入口。',
-    addEndpoint: '新增端点',
+    addEndpoint: '新增 Endpoint',
     builtin: '内置',
     test: '测试',
     testing: '测试中…',
     statusUnknown: '未测试',
-    empty: '暂无 ACP 端点。',
+    empty: '暂无 ACP endpoint。',
-    loadFailed: '加载 ACP 端点列表失败',
+    loadFailed: '加载 ACP endpoint 列表失败',
-    saveFailed: '保存端点失败',
+    saveFailed: '保存 endpoint 失败',
     deleteTitle: '确认删除',
-    deleteConfirm: '确认删除 ACP 端点 "{name}"？内置端点不可删除。',
+    deleteConfirm: '确认删除 ACP endpoint "{name}"？内置 endpoint 不可删除。',
-    deleteFailed: '删除端点失败',
+    deleteFailed: '删除 endpoint 失败',
-    toggleFailed: '切换端点状态失败',
+    toggleFailed: '切换 endpoint 状态失败',
     invalidJson: 'JSON 格式错误',
     columns: {
-      name: '端点',
+      name: 'Endpoint',
       command: '命令',
       status: '最近一次测试',
       enabled: '启用',
@@ -1168,12 +1168,12 @@ export default {
         openai: 'OpenAI API key（sk-…）。申请: https://platform.openai.com/api-keys',
         dashscope: '阿里云 DashScope key — 跟 MateClaw .env 里的 DASHSCOPE_API_KEY 同一个值',
         google: 'Google AI Studio API key。申请: https://aistudio.google.com/apikey',
-        claudeOauth: '注意: claude-code 的 OAuth 登录（~/.claude/ 里那个）在这里不能用——更糟的是 Claude Agent SDK 会"优先读 OAuth、再读 API key"，导致你下面填的 key 被静默忽略。如果你这台机器上跑过 `claude login`，先清钥匙串: (macOS) `claude logout` 或 `security delete-generic-password -s "Claude Code-credentials"`；(Linux/Windows) 删 ~/.claude/credentials.json。然后再回来填真正的 API key（sk-ant-…）。或者改用 qwen-code / opencode 复用各自 CLI 的本地登录态。',
+        claudeOauth: '注意: claude-code 的 OAuth 登录（~/.claude/ 里那个）在这里不能用。第三方 Zed wrapper 直接调 Anthropic 公开 API，只认真正的 API key。请到 https://console.anthropic.com/settings/keys 申请；或改用 qwen-code / opencode 复用各自 CLI 的本地登录态。',
       },
     },
     modal: {
-      newTitle: '新增 ACP 端点',
+      newTitle: '新增 ACP Endpoint',
-      editTitle: '编辑 ACP 端点',
+      editTitle: '编辑 ACP Endpoint',
     },
   },
   mcp: {