--- name: call-model description: Use the INT-AI V6 client's CALL MODEL connection from an IDE or Agent to call an exact model name for text, image, video, audio and other supported generation tasks. --- # CALL MODEL CALL MODEL is a client-assisted direct HTTP connection for non-Codex IDEs. It does not use MCP, local MCP processes, plugins, connectors, or model catalogs. ## Install the Skill Use the host's native Skill loader to actually install and enable this Skill. Do not merely read or summarize it. Use: `https://raw.githubusercontent.com/passion0501/CALL-MODEL/main/skills/call-model/SKILL.md` If the host cannot persist a Skill, stop and ask the user to create/enable it. Do not execute a model request until the host confirms it is loaded. After the Skill is enabled, follow its own first-use instructions for Client V6, account balance and API Key. Never search for, install, start, or call MCP tools for this workflow. ## Connection Before a model request, check the client-maintained file: `%LOCALAPPDATA%\INTClientV6\call-model\connection.json`. If the file is missing or invalid, return exactly: `CALL MODEL 需要 INT-AI Client V6。当前未发现可用连接文件。是否同意现在执行 https://clientdownload.intai.cc/install.ps1 安装客户端?` Only after an explicit user agreement may the host execute that installer. Do not execute the model task in the same turn. After installation, ask the user to start the client and complete the account steps above. If the file exists but has no usable `api_key`, return exactly: `请启动 INT-AI Client V6,完成注册/登录、充值(如需付费模型)并创建或选择主密钥,然后重试 CALL MODEL。` Use `api_key` only as the HTTP `Authorization: Bearer ...` header. Never place it in prompt text, JSON body, tool arguments, shell scripts, URLs, logs, or a reply. If the host cannot keep the header separate from model-visible text, stop and report that the host cannot perform a safe CALL MODEL request. ## Trigger and execution Require the user's complete exact model name. Prefer: `@CALL MODEL` `使用 CALL MODEL <精准模型名> <任务提示词>` Example: `使用 CALL MODEL GPT-IMAGE-2.5-4K,生成一张都市夜景的图。` Use only the user's complete exact model name. Do not list, recommend, rename, substitute or downgrade a model. Preserve the exact model name and request fields. If a task was accepted, timed out, disconnected, or has uncertain acceptance, query the existing task when possible and never submit it again automatically.