Add ACP (Agent Client Protocol) support

README.md

@@ -14,6 +14,7 @@ An Obsidian plugin that integrates with LLM CLI tools (Claude, Codex, OpenCode,
 - **Create Notes from Responses** - Save LLM responses as new notes in your vault
 - **Quick Prompts** - Commands for summarizing, explaining, and improving selected text
 - **Session Continuation** - Follow-up messages use session resumption for faster responses
+- **ACP Mode** - Optional persistent connection mode for faster multi-turn conversations
 
 ## Requirements
 
@@ -93,8 +94,13 @@ Copy `main.js`, `manifest.json`, and `styles.css` to your vault's `.obsidian/plu
 
 Each provider can be configured with:
 - Enable/disable
+- Model selection (dropdown with common models, or enter custom model ID)
 - Custom command (if CLI is named differently)
 - Timeout override
+- **ACP Mode** (experimental) - Use Agent Client Protocol for persistent connections
+- **Thinking Mode** (ACP only) - Control extended thinking level (none/low/medium/high)
+
+When ACP mode is enabled, the available models list is populated dynamically from the connected agent.
 
 ### System Prompt
 
@@ -153,28 +159,27 @@ The plugin captures session IDs from CLI tools and uses them for subsequent requ
 - **Claude**: `--resume <session_id>`
 - **OpenCode**: `--session <session_id>`
 - **Gemini**: `--resume <session_id>`
-- **Codex**: Uses `resume` subcommand (different pattern, not fully supported)
+- **Codex**: `--resume <session_id>`
 
 This improves response times for follow-up messages. Clearing the conversation resets the session.
 
-### Future Improvements
+### ACP Mode (Agent Client Protocol)
 
-**ACP (Agent Client Protocol)**: We're exploring ACP support for improved performance and capabilities. ACP is a standardized protocol (like LSP for AI agents) that allows:
-- Long-lived agent processes (no startup overhead per request)
-- Streaming responses via JSON-RPC
-- Terminal integration (agents can run commands)
-- Standardized session management
+ACP is a standardized protocol (like LSP but for AI agents) that provides:
+- **Persistent connections** - No startup overhead per request
+- **Streaming responses** - Real-time updates via JSON-RPC
+- **Dynamic model discovery** - Available models fetched from the agent
+- **Extended thinking** - Support for thinking mode levels
 
-CLI tools with ACP support:
-- `opencode acp` - OpenCode's ACP server
-- `gemini --experimental-acp` - Gemini CLI's ACP mode
-- Claude Code - via `@zed-industries/claude-code-acp` adapter
+All four providers support ACP mode:
+- **Claude** - via `@anthropic-ai/claude-code-acp` adapter
+- **OpenCode** - native `opencode acp` server
+- **Gemini** - via `gemini --experimental-acp` flag
+- **Codex** - via `@zed-industries/codex-acp` adapter
 
-See also: [obsidian-agent-client](https://github.com/RAIT-09/obsidian-agent-client) for an alternative ACP-based approach.
+Enable ACP in provider settings. The first message establishes a connection; subsequent messages reuse it for faster responses.
 
-**Long-lived CLI Process**: As an alternative to ACP, some tools support headless server modes:
-- `opencode serve` / `opencode attach` - Headless server mode
-- `codex mcp-server` - MCP server mode
+See also: [obsidian-agent-client](https://github.com/RAIT-09/obsidian-agent-client) for an alternative ACP-based approach
 
 ## License
 

main.ts

@@ -197,6 +197,29 @@ export default class LLMPlugin extends Plugin {
     if (leaf) {
       workspace.revealLeaf(leaf);
     }
+
+    return leaf;
+  }
+
+  /**
+   * Get the ChatView instance if it exists
+   */
+  getChatView(): ChatView | null {
+    const leaves = this.app.workspace.getLeavesOfType(CHAT_VIEW_TYPE);
+    if (leaves.length > 0) {
+      return leaves[0].view as ChatView;
+    }
+    return null;
+  }
+
+  /**
+   * Add a message exchange to the chat view
+   */
+  addToChatView(userMessage: string, assistantMessage: string, provider: LLMProvider) {
+    const chatView = this.getChatView();
+    if (chatView) {
+      chatView.addMessageExchange(userMessage, assistantMessage, provider);
+    }
   }
 
   async loadSettings() {
@@ -258,8 +281,10 @@ export default class LLMPlugin extends Plugin {
   /**
    * Update the status bar with current provider and model info
    * @param provider Optional provider to display (uses default if not specified)
+   * @param actualModelName Optional actual model name from ACP session (overrides configured model display)
+   * @param status Optional status: "idle" (default), "connecting", "connected"
    */
-  updateStatusBar(provider?: LLMProvider) {
+  updateStatusBar(provider?: LLMProvider, actualModelName?: string, status?: "idle" | "connecting" | "connected") {
     if (!this.statusBarEl) return;
 
     const displayProvider = provider ?? this.settings.defaultProvider;
@@ -278,8 +303,13 @@ export default class LLMPlugin extends Plugin {
 
     // Build status text with provider and model
     let statusText = providerNames[displayProvider] || displayProvider;
-    if (providerConfig?.model) {
-      // Show abbreviated model name
+    if (status === "connecting") {
+      statusText += " (connecting...)";
+    } else if (actualModelName) {
+      // Use actual model name from ACP session
+      statusText += ` (${this.formatModelName(actualModelName)})`;
+    } else if (providerConfig?.model) {
+      // Show configured model name
       statusText += ` (${this.formatModelName(providerConfig.model)})`;
     } else {
       // Indicate CLI default is being used
@@ -291,8 +321,10 @@ export default class LLMPlugin extends Plugin {
       cls: "llm-status-text",
     });
 
-    // Check if provider is enabled
-    if (providerConfig?.enabled) {
+    // Set indicator state based on status
+    if (status === "connecting") {
+      indicator.addClass("connecting");
+    } else if (providerConfig?.enabled) {
       indicator.addClass("active");
     }
   }
@@ -301,7 +333,7 @@ export default class LLMPlugin extends Plugin {
    * Format model name for display (abbreviate long names)
    */
   private formatModelName(model: string): string {
-    // Common abbreviations
+    // Common abbreviations for model IDs
     const abbreviations: Record<string, string> = {
       "claude-3-5-haiku-latest": "haiku",
       "claude-3-5-sonnet-latest": "sonnet-3.5",
@@ -319,8 +351,33 @@ export default class LLMPlugin extends Plugin {
       "gpt-5": "5",
       "claude-sonnet": "sonnet",
       "claude-haiku": "haiku",
+      // ACP display names (from Claude ACP adapter)
+      "default": "opus",
+      "Default (recommended)": "opus",
+      "Sonnet": "sonnet",
+      "Haiku": "haiku",
     };
 
-    return abbreviations[model] || model;
+    // Check for exact match first
+    if (abbreviations[model]) {
+      return abbreviations[model];
+    }
+
+    // Try case-insensitive match
+    const lowerModel = model.toLowerCase();
+    for (const [key, value] of Object.entries(abbreviations)) {
+      if (key.toLowerCase() === lowerModel) {
+        return value;
+      }
+    }
+
+    // If model name is long, try to extract a shorter name
+    // Remove text in parentheses and trim
+    const simplified = model.replace(/\s*\([^)]*\)\s*/g, "").trim();
+    if (simplified !== model && simplified.length > 0) {
+      return this.formatModelName(simplified);
+    }
+
+    return model;
   }
 }

package.json

@@ -42,5 +42,8 @@
     "typescript": "^5.0.0",
     "wdio-obsidian-reporter": "^2.2.1",
     "wdio-obsidian-service": "^2.2.1"
+  },
+  "dependencies": {
+    "@agentclientprotocol/sdk": "^0.13.1"
   }
 }