feat(agents): dynamic model list from /v1/llm_models at agent startup

src/agents/core/session/SessionStore.ts

@@ -7,7 +7,7 @@
 
 import { readFile, writeFile, mkdir } from 'fs/promises';
 import { existsSync } from 'fs';
-import { dirname } from 'path';
+import { dirname, basename, join } from 'path';
 import type { Session } from './types.js';
 import { getSessionPath } from './session-config.js';
 import { logger } from '../../../utils/logger.js';
@@ -41,13 +41,24 @@ export class SessionStore {
 
   /**
    * Load session from disk
+   *
+   * Falls back to the 'completed_' prefixed filename when the primary path is
+   * not found. This handles the race where handleSessionEnd renames
+   * {sessionId}.json → completed_{sessionId}.json before the final SSO sync runs.
    */
   async loadSession(sessionId: string): Promise<Session | null> {
-    const sessionPath = getSessionPath(sessionId);
+    let sessionPath = getSessionPath(sessionId);
 
     if (!existsSync(sessionPath)) {
-      logger.debug(`[SessionStore] Session file not found: ${sessionId}`);
-      return null;
+      // Fallback: session may have been renamed with 'completed_' prefix by handleSessionEnd
+      const completedPath = join(dirname(sessionPath), `completed_${basename(sessionPath)}`);
+      if (existsSync(completedPath)) {
+        sessionPath = completedPath;
+        logger.debug(`[SessionStore] Using completed session file: ${sessionId}`);
+      } else {
+        logger.debug(`[SessionStore] Session file not found: ${sessionId}`);
+        return null;
+      }
     }
 
     try {

src/agents/plugins/__tests__/codemie-code-plugin.test.ts

@@ -84,7 +84,7 @@ vi.mock('../opencode/opencode.session.js', () => ({
   }),
 }));
 
-// Mock getModelConfig and getAllOpenCodeModelConfigs
+// Mock getModelConfig and model config helpers
 vi.mock('../opencode/opencode-model-configs.js', () => ({
   getModelConfig: vi.fn(() => ({
     id: 'gpt-5-2-2025-12-11',
@@ -102,7 +102,13 @@ vi.mock('../opencode/opencode-model-configs.js', () => ({
     cost: { input: 2.5, output: 10 },
     limit: { context: 1048576, output: 65536 },
   })),
-  getAllOpenCodeModelConfigs: vi.fn(() => ({})),
+  getChatCompletionsModelConfigs: vi.fn(() => ({})),
+  getResponsesApiModelConfigs: vi.fn(() => ({})),
+}));
+
+// Mock dynamic model fetcher so tests don't make real API calls
+vi.mock('../opencode/opencode-dynamic-models.js', () => ({
+  fetchDynamicModelConfigs: vi.fn(() => Promise.resolve({})),
 }));
 
 // Mock fs

src/agents/plugins/__tests__/codemie-code-reasoning.test.ts

@@ -89,7 +89,7 @@ vi.mock('../opencode/opencode.session.js', () => ({
   }),
 }));
 
-// Mock getModelConfig and getAllOpenCodeModelConfigs
+// Mock getModelConfig and model config helpers
 vi.mock('../opencode/opencode-model-configs.js', () => ({
   getModelConfig: vi.fn(() => ({
     id: 'gpt-5-2-2025-12-11',
@@ -98,7 +98,13 @@ vi.mock('../opencode/opencode-model-configs.js', () => ({
     tool_call: true,
     reasoning: true,
   })),
-  getAllOpenCodeModelConfigs: vi.fn(() => ({})),
+  getChatCompletionsModelConfigs: vi.fn(() => ({})),
+  getResponsesApiModelConfigs: vi.fn(() => ({})),
+}));
+
+// Mock dynamic model fetcher so tests don't make real API calls
+vi.mock('../opencode/opencode-dynamic-models.js', () => ({
+  fetchDynamicModelConfigs: vi.fn(() => Promise.resolve({})),
 }));
 
 // Mock fs

src/agents/plugins/codemie-code.plugin.ts

@@ -2,7 +2,8 @@ import type { AgentMetadata, AgentConfig } from '../core/types.js';
 import { join } from 'path';
 import { existsSync } from 'fs';
 import { logger } from '../../utils/logger.js';
-import { getModelConfig, getAllOpenCodeModelConfigs } from './opencode/opencode-model-configs.js';
+import { getModelConfig, getChatCompletionsModelConfigs, getResponsesApiModelConfigs } from './opencode/opencode-model-configs.js';
+import { fetchDynamicModelConfigs } from './opencode/opencode-dynamic-models.js';
 import { BaseAgentAdapter } from '../core/BaseAgentAdapter.js';
 import type { SessionAdapter } from '../core/session/BaseSessionAdapter.js';
 import type { BaseExtensionInstaller } from '../core/extension/BaseExtensionInstaller.js';
@@ -73,6 +74,10 @@ function resolveOllamaBaseUrl(baseUrl: string, provider: string | undefined): st
 
 /**
  * Build the OpenCode config object that gets passed to the whitelabel binary.
+ *
+ * Models are split into two groups:
+ * - chatModels: routed via codemie-proxy/litellm (Chat Completions API)
+ * - responsesApiModels: routed via OpenCode's built-in openai CUSTOM_LOADER (Responses API)
  */
 function buildOpenCodeConfig(params: {
   proxyBaseUrl: string | undefined;
@@ -83,10 +88,13 @@ function buildOpenCodeConfig(params: {
   modelId: string;
   timeout: number;
   providerOptions?: any;
-  allModels: Record<string, unknown>;
+  chatModels: Record<string, unknown>;
+  responsesApiModels: Record<string, unknown>;
+  responsesApiBaseUrl: string | undefined;
 }): Record<string, unknown> {
+  const hasResponsesApiModels = Object.keys(params.responsesApiModels).length > 0;
   return {
-    enabled_providers: ['codemie-proxy', 'ollama', 'amazon-bedrock', 'litellm'],
+    enabled_providers: ['codemie-proxy', 'openai', 'ollama', 'amazon-bedrock', 'litellm'],
     share: 'disabled',
     provider: {
       ...(params.proxyBaseUrl && {
@@ -99,7 +107,24 @@ function buildOpenCodeConfig(params: {
             timeout: params.timeout,
             ...(params.providerOptions?.headers && { headers: params.providerOptions.headers })
           },
-          models: params.allModels
+          models: params.chatModels
+        }
+      }),
+      // OpenCode's built-in openai CUSTOM_LOADER — uses @ai-sdk/openai sdk.responses()
+      // which calls POST /v1/responses instead of /v1/chat/completions
+      ...(params.responsesApiBaseUrl && hasResponsesApiModels && {
+        openai: {
+          name: 'CodeMie SSO',
+          // whitelist: suppress the built-in openai model list (GPT-4, GPT-4o, etc.)
+          // OpenCode merges user models with models.dev — whitelist restricts to ours only
+          whitelist: Object.keys(params.responsesApiModels),
+          options: {
+            baseURL: `${params.responsesApiBaseUrl}/`,
+            apiKey: 'proxy-handled',
+            timeout: params.timeout,
+            ...(params.providerOptions?.headers && { headers: params.providerOptions.headers })
+          },
+          models: params.responsesApiModels
         }
       }),
       ...(params.litellmBaseUrl && {
@@ -111,7 +136,7 @@ function buildOpenCodeConfig(params: {
             apiKey: params.litellmApiKey || 'not-required',
             timeout: params.timeout,
           },
-          models: params.allModels
+          models: params.chatModels
         }
       }),
       ollama: {
@@ -226,10 +251,21 @@ export const CodeMieCodePluginMetadata: AgentMetadata = {
         return env;
       }
 
+      // Fetch live model catalogue from the CodeMie API.
+      // Falls back to the static OPENCODE_MODEL_CONFIGS on any error.
+      const allModels = await fetchDynamicModelConfigs(
+        baseUrl,
+        env.CODEMIE_URL,
+        env.CODEMIE_JWT_TOKEN,
+      );
+
+      // Model selection priority: env var > config > default
+      // Use dynamic catalogue first, then fall back to static getModelConfig for unknown IDs.
       const selectedModel = env.CODEMIE_MODEL || config?.model || 'gpt-5-2-2025-12-11';
-      const modelConfig = getModelConfig(selectedModel);
+      const modelConfig = allModels[selectedModel] ?? getModelConfig(selectedModel);
       const { providerOptions } = modelConfig;
-      const allModels = getAllOpenCodeModelConfigs();
+      const chatModels = getChatCompletionsModelConfigs(allModels);
+      const responsesApiModels = getResponsesApiModelConfigs(allModels);
 
       const isBedrock = provider === 'bedrock';
       const isLiteLLM = provider === 'litellm';
@@ -241,11 +277,21 @@ export const CodeMieCodePluginMetadata: AgentMetadata = {
         ? toBedrockModelId(modelConfig.id, env.AWS_REGION || env.CODEMIE_AWS_REGION)
         : modelConfig.id;
 
+      // Responses API base URL: use proxyBaseUrl for SSO/bearer-auth, or baseUrl for LiteLLM.
+      // Always set regardless of selected model — fixes model-switching bug where switching
+      // from a Claude model to a GPT model mid-session would miss the CUSTOM_LOADER.
+      const responsesApiBaseUrl = proxyBaseUrl || (isLiteLLM ? baseUrl : undefined);
+      if (responsesApiBaseUrl && Object.keys(responsesApiModels).length > 0) {
+        env.OPENAI_API_KEY = 'proxy-handled';
+        logger.debug('[codemie-code] Enabling openai CUSTOM_LOADER for Responses API models');
+      }
+
       const openCodeConfig = buildOpenCodeConfig({
         proxyBaseUrl,
         litellmBaseUrl: isLiteLLM ? baseUrl : undefined,
         litellmApiKey: isLiteLLM ? env.CODEMIE_API_KEY : undefined,
-        ollamaBaseUrl, activeProvider, modelId, timeout, providerOptions, allModels
+        ollamaBaseUrl, activeProvider, modelId, timeout, providerOptions,
+        chatModels, responsesApiModels, responsesApiBaseUrl
       });
 
       // --- Hooks injection ---

src/agents/plugins/opencode/node-sqlite.d.ts

@@ -0,0 +1,24 @@
+/**
+ * Minimal type declarations for node:sqlite (Node.js 22.5+, experimental)
+ *
+ * node:sqlite lacks @types declarations in @types/node@20.x.
+ * These ambient declarations provide compile-time safety for the sqlite reader.
+ * At runtime, Node.js 22.5+ provides the module natively.
+ */
+declare module 'node:sqlite' {
+  interface StatementSync {
+    all(...params: unknown[]): Record<string, unknown>[];
+  }
+
+  interface DatabaseSyncOptions {
+    open?: boolean;
+  }
+
+  class DatabaseSync {
+    constructor(path: string, options?: DatabaseSyncOptions);
+    prepare(sql: string): StatementSync;
+    close(): void;
+  }
+
+  export { DatabaseSync };
+}