claude-connect 0.1.7 → 0.1.8

package/README.md

@@ -83,6 +83,7 @@ Al activar:
 - `Inception Labs` usa el gateway local sobre `https://api.inceptionlabs.ai/v1/chat/completions`
 - `OpenRouter` usa `openrouter/free` por gateway sobre `https://openrouter.ai/api/v1`
 - `Qwen` apunta al gateway local `http://127.0.0.1:4310/anthropic`
+- para algunos modelos con limites conocidos, el gateway ahora ajusta `max_tokens` y bloquea prompts sobredimensionados antes de que el upstream devuelva errores opacos
 
 ## Providers
 
@@ -123,6 +124,8 @@ Nota sobre `OpenAI`:
 Nota sobre `Inception Labs`:
 
 - esta primera integracion expone solo `mercury-2`, que es el modelo chat-compatible oficial en `v1/chat/completions`
+- `mercury-2` se trata como modelo solo texto en Claude Connect; si envias una imagen, la app ahora corta la peticion con un mensaje claro
+- Claude Connect aplica presupuesto preventivo de contexto para `mercury-2` usando ventana `128K` y salida maxima `16,384`
 - `Mercury Edit 2` no se publica todavia en Claude Connect porque usa endpoints `fim/edit` que no encajan con Claude Code en esta arquitectura
 - autenticacion soportada: `API key`
 - referencias oficiales:
@@ -130,6 +133,13 @@ Nota sobre `Inception Labs`:
   - https://docs.inceptionlabs.ai/get-started/authentication
   - https://docs.inceptionlabs.ai/get-started/models
 
+Nota sobre `DeepSeek`:
+
+- Claude Connect aplica presupuesto preventivo de contexto para `deepseek-chat` y `deepseek-reasoner`
+- referencias oficiales:
+  - https://api-docs.deepseek.com/quick_start/pricing/
+  - https://api-docs.deepseek.com/guides/reasoning_model
+
 Nota sobre `Ollama`:
 
 - la URL del servidor se define al crear la conexión

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-connect",
-  "version": "0.1.7",
+  "version": "0.1.8",
   "description": "CLI para configurar Claude Code con proveedores de modelos externos",
   "author": "wmcarlosv",
   "type": "module",

package/src/data/catalog-store.js

@@ -39,6 +39,7 @@ CREATE TABLE IF NOT EXISTS models (
   api_base_url TEXT,
   api_path TEXT,
   auth_env_mode TEXT NOT NULL DEFAULT 'auth_token',
+  supports_vision INTEGER NOT NULL DEFAULT 1,
   sort_order INTEGER NOT NULL DEFAULT 0,
   is_default INTEGER NOT NULL DEFAULT 0
 );
@@ -703,6 +704,7 @@ const seedProviders = [
         apiBaseUrl: 'https://api.inceptionlabs.ai/v1',
         apiPath: '/chat/completions',
         authEnvMode: 'auth_token',
+        supportsVision: false,
         sortOrder: 1,
         isDefault: 1
       }
@@ -838,10 +840,10 @@ function seedCatalog(db) {
     INSERT INTO models (
       id, provider_id, name, category, context_window, summary,
       upstream_model_id,
-      transport_mode, api_style, api_base_url, api_path, auth_env_mode,
+      transport_mode, api_style, api_base_url, api_path, auth_env_mode, supports_vision,
       sort_order, is_default
     )
-    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
     ON CONFLICT(id) DO UPDATE SET
       provider_id = excluded.provider_id,
       name = excluded.name,
@@ -854,6 +856,7 @@ function seedCatalog(db) {
       api_base_url = excluded.api_base_url,
       api_path = excluded.api_path,
       auth_env_mode = excluded.auth_env_mode,
+      supports_vision = excluded.supports_vision,
       sort_order = excluded.sort_order,
       is_default = excluded.is_default
   `);
@@ -924,6 +927,7 @@ function seedCatalog(db) {
           model.apiBaseUrl ?? null,
           model.apiPath ?? null,
           model.authEnvMode ?? 'auth_token',
+          model.supportsVision === false ? 0 : 1,
           model.sortOrder,
           model.isDefault
         );
@@ -995,6 +999,10 @@ function ensureSchemaMigrations(db) {
     alterStatements.push(`ALTER TABLE models ADD COLUMN auth_env_mode TEXT NOT NULL DEFAULT 'auth_token'`);
   }
 
+  if (!modelColumns.has('supports_vision')) {
+    alterStatements.push(`ALTER TABLE models ADD COLUMN supports_vision INTEGER NOT NULL DEFAULT 1`);
+  }
+
   if (!modelColumns.has('upstream_model_id')) {
     alterStatements.push(`ALTER TABLE models ADD COLUMN upstream_model_id TEXT`);
   }
@@ -1037,6 +1045,7 @@ function mapModelRow(row) {
     apiBaseUrl: row.api_base_url,
     apiPath: row.api_path,
     authEnvMode: row.auth_env_mode,
+    supportsVision: Boolean(row.supports_vision),
     sortOrder: Number(row.sort_order),
     isDefault: Boolean(row.is_default)
   };

package/src/gateway/server.js

@@ -24,6 +24,7 @@ import {
 } from './state.js';
 import { resolveClaudeConnectPaths } from '../lib/app-paths.js';
 import { readSwitchState } from '../lib/claude-settings.js';
+import { enforceModelTokenBudget } from '../lib/model-budget.js';
 import { readOAuthToken, refreshOAuthToken } from '../lib/oauth.js';
 import { readProfileFile } from '../lib/profile.js';
 import { readManagedProviderTokenSecret, readManagedTokenSecret } from '../lib/secrets.js';
@@ -161,6 +162,48 @@ function getUpstreamModelId(profile) {
   return profile?.model?.upstreamModelId ?? profile?.model?.id ?? 'unknown';
 }
 
+function requestContainsImageInput(body) {
+  return Array.isArray(body?.messages)
+    && body.messages.some((messageItem) => Array.isArray(messageItem?.content)
+      && messageItem.content.some((part) => part?.type === 'image'));
+}
+
+function profileSupportsImageInput(profile) {
+  if (typeof profile?.model?.supportsVision === 'boolean') {
+    return profile.model.supportsVision;
+  }
+
+  if (profile?.provider?.id === 'inception') {
+    return false;
+  }
+
+  return true;
+}
+
+function stringifyUpstreamMessage(value) {
+  if (typeof value === 'string') {
+    return value;
+  }
+
+  if (value == null) {
+    return '';
+  }
+
+  if (typeof value === 'object') {
+    if ('message' in value && typeof value.message === 'string') {
+      return value.message;
+    }
+
+    try {
+      return JSON.stringify(value);
+    } catch (_error) {
+      return String(value);
+    }
+  }
+
+  return String(value);
+}
+
 function resolveGatewayUpstreamConfig(profile) {
   if (profile?.provider?.id === 'ollama') {
     return {
@@ -335,9 +378,9 @@ async function forwardUpstreamRequest({ targetUrl, headers, payload, context, re
     });
   }
 
-  const message = responsePayload?.error?.message
-    || responsePayload?.message
-    || responsePayload?.error
+  const message = stringifyUpstreamMessage(responsePayload?.error?.message)
+    || stringifyUpstreamMessage(responsePayload?.message)
+    || stringifyUpstreamMessage(responsePayload?.error)
     || `HTTP ${response.status}`;
   const providerName = context?.profile?.provider?.name ?? context?.profile?.provider?.id ?? 'El proveedor';
   const containsImageInput = Array.isArray(payload?.messages)
@@ -453,9 +496,20 @@ async function handleCountTokens(request, response) {
 }
 
 async function handleMessages(request, response) {
-  const body = await readJsonBody(request);
+  const rawBody = await readJsonBody(request);
   const context = await resolveGatewayContext();
 
+  if (requestContainsImageInput(rawBody) && !profileSupportsImageInput(context.profile)) {
+    const providerName = context.profile.provider.name;
+    const modelName = context.profile.model.name;
+    throw new Error(`${providerName} no admite imagenes con el modelo ${modelName} en esta integracion. Usa un proveedor o modelo con soporte visual.`);
+  }
+
+  const body = enforceModelTokenBudget({
+    body: rawBody,
+    profile: context.profile
+  });
+
   if (context.upstreamApiStyle === 'anthropic') {
     const upstreamResponse = await forwardAnthropicMessage({
       requestBody: body,

package/src/lib/model-budget.js

@@ -0,0 +1,73 @@
+import { estimateTokenCountFromAnthropicRequest } from '../gateway/messages.js';
+
+function getModelIdentity(profile) {
+  return profile?.model?.upstreamModelId ?? profile?.model?.id ?? '';
+}
+
+export function getModelTokenLimits(profile) {
+  const providerId = profile?.provider?.id;
+  const modelId = getModelIdentity(profile);
+
+  if (providerId === 'inception' && modelId === 'mercury-2') {
+    return {
+      contextWindowTokens: 128_000,
+      defaultOutputTokens: 8_192,
+      maxOutputTokens: 16_384
+    };
+  }
+
+  if (providerId === 'deepseek' && modelId === 'deepseek-chat') {
+    return {
+      contextWindowTokens: 128_000,
+      defaultOutputTokens: 4_000,
+      maxOutputTokens: 8_000
+    };
+  }
+
+  if (providerId === 'deepseek' && modelId === 'deepseek-reasoner') {
+    return {
+      contextWindowTokens: 128_000,
+      defaultOutputTokens: 32_000,
+      maxOutputTokens: 64_000
+    };
+  }
+
+  return null;
+}
+
+export function enforceModelTokenBudget({ body, profile, safetyMarginTokens = 1024 }) {
+  const limits = getModelTokenLimits(profile);
+
+  if (!limits) {
+    return body;
+  }
+
+  const estimatedInputTokens = estimateTokenCountFromAnthropicRequest(body);
+  const availableForOutput = limits.contextWindowTokens - estimatedInputTokens - safetyMarginTokens;
+
+  if (availableForOutput <= 0) {
+    throw new Error(
+      `La conversacion actual excede el contexto aproximado de ${limits.contextWindowTokens.toLocaleString('en-US')} tokens para ${profile.model.name}. Usa /compact o /clear antes de continuar.`
+    );
+  }
+
+  const requestedOutputTokens = typeof body?.max_tokens === 'number'
+    ? body.max_tokens
+    : limits.defaultOutputTokens;
+  const clampedOutputTokens = Math.min(
+    requestedOutputTokens,
+    limits.maxOutputTokens,
+    availableForOutput
+  );
+
+  if (clampedOutputTokens < 256) {
+    throw new Error(
+      `Queda muy poco margen de salida para ${profile.model.name} dentro de su contexto aproximado de ${limits.contextWindowTokens.toLocaleString('en-US')} tokens. Usa /compact o /clear antes de continuar.`
+    );
+  }
+
+  return {
+    ...body,
+    max_tokens: clampedOutputTokens
+  };
+}

package/src/lib/profile.js

@@ -32,7 +32,8 @@ export function buildProfile({ provider, model, authMethod, profileName, apiKeyE
       apiStyle: model.apiStyle,
       apiBaseUrl: model.apiBaseUrl,
       apiPath: model.apiPath,
-      authEnvMode: model.authEnvMode
+      authEnvMode: model.authEnvMode,
+      supportsVision: model.supportsVision ?? true
     },
     auth: {
       method: authMethod.id

package/src/lib/terminal.js

@@ -19,7 +19,26 @@ export function openAppScreen() {
 }
 
 export function closeAppScreen() {
-  process.stdout.write('\x1b[?25h\x1b[?1049l');
+  process.stdout.write('\x1b[?25h\x1b[?1049l\r\n');
+}
+
+function beginKeyboardCapture() {
+  readline.emitKeypressEvents(process.stdin);
+  process.stdin.resume();
+  process.stdin.setRawMode(true);
+}
+
+function restoreKeyboardState(onKeypress) {
+  if (typeof onKeypress === 'function') {
+    process.stdin.removeListener('keypress', onKeypress);
+  }
+
+  if (process.stdin.isTTY && process.stdin.isRaw) {
+    process.stdin.setRawMode(false);
+  }
+
+  process.stdin.pause();
+  process.stdout.write('\x1b[?25h');
 }
 
 export function clearScreen() {
@@ -64,13 +83,11 @@ export function buildFrame({ eyebrow, title, subtitle, body = [], footer = [] })
 
 export function waitForAnyKey(message = 'Presiona una tecla para continuar.') {
   return new Promise((resolve, reject) => {
-    readline.emitKeypressEvents(process.stdin);
-    process.stdin.setRawMode(true);
+    beginKeyboardCapture();
     let escapePending = false;
 
     const cleanup = () => {
-      process.stdin.removeListener('keypress', onKeypress);
-      process.stdin.setRawMode(false);
+      restoreKeyboardState(onKeypress);
     };
 
     const onKeypress = (_input, key = {}) => {
@@ -124,12 +141,10 @@ export function selectFromList({
     let selectedIndex = 0;
     let escapePending = false;
 
-    readline.emitKeypressEvents(process.stdin);
-    process.stdin.setRawMode(true);
+    beginKeyboardCapture();
 
     const cleanup = () => {
-      process.stdin.removeListener('keypress', onKeypress);
-      process.stdin.setRawMode(false);
+      restoreKeyboardState(onKeypress);
     };
 
     const render = () => {
@@ -247,12 +262,10 @@ export function promptText({
     let value = '';
     let escapePending = false;
 
-    readline.emitKeypressEvents(process.stdin);
-    process.stdin.setRawMode(true);
+    beginKeyboardCapture();
 
     const cleanup = () => {
-      process.stdin.removeListener('keypress', onKeypress);
-      process.stdin.setRawMode(false);
+      restoreKeyboardState(onKeypress);
     };
 
     const render = () => {