comfyui-mcp/dist/services/llamacpp-probe.js

Local-first, agent-native control plane for ComfyUI — MCP server + autonomous sidebar agent that drives your live graph in natural language on ANY LLM: Claude/ChatGPT/Gemini on your subscription (no API key), free local models via Ollama (fully offline),

// llama.cpp (llama-server) probes for the panel orchestrator — the two
// launch-time gotchas the provider must surface instead of failing cryptically:
//
//   * `--jinja` is MANDATORY for tool calling: without it llama-server rejects
//     any request carrying `tools` (and can never emit finish_reason=
//     tool_calls). A server that "connects fine" then errors on the first real
//     message reads as a wedge — detect it at connect time instead.
//   * context is fixed at LAUNCH (`-c`), not per request: the comfyui-mcp tool
//     payload needs ~16K minimum, so a small n_ctx silently truncates. /props
//     reports the server's actual window.
//
// Best-effort: failures return "unknown" rather than blocking a connect.
import { logger } from "../utils/logger.js";
/** The REST root (strip a trailing /v1 — native endpoints hang off the root). */
function serverRoot(host) {
    return host.replace(/\/+$/, "").replace(/\/v1$/, "");
}
export async function llamacppProps(host) {
    try {
        const res = await fetch(`${serverRoot(host)}/props`, { signal: AbortSignal.timeout(4000) });
        if (!res.ok)
            return {};
        const body = (await res.json());
        return {
            nCtx: body.default_generation_settings?.n_ctx,
            modelPath: body.model_path,
        };
    }
    catch {
        return {};
    }
}
/** Can this server take tool-calling requests?
 *
 *  Primary signal: /props `chat_template_caps.supports_tools` — authoritative
 *  on current builds (which enable jinja BY DEFAULT; verified live on b9945:
 *  a plain launch reports supports_tools:true and accepts tools). Fallback for
 *  OLD builds (where --jinja was opt-in and tools requests get REJECTED): a
 *  minimal max_tokens-1 tools request, "no" on the tools/jinja error signature. */
export async function llamacppToolsReady(host, model) {
    try {
        const res = await fetch(`${serverRoot(host)}/props`, { signal: AbortSignal.timeout(4000) });
        if (res.ok) {
            const body = (await res.json());
            const caps = body.chat_template_caps;
            if (caps && (caps.supports_tools === false || caps.supports_tool_calls === false)) {
                logger.warn("[llamacpp] /props reports the chat template does NOT support tools");
                return "no";
            }
            if (caps?.supports_tools || caps?.supports_tool_calls)
                return "yes";
        }
    }
    catch {
        // fall through to the request probe
    }
    try {
        const res = await fetch(`${host.replace(/\/+$/, "")}/chat/completions`, {
            method: "POST",
            headers: { "Content-Type": "application/json" },
            body: JSON.stringify({
                model,
                messages: [{ role: "user", content: "ok" }],
                tools: [
                    {
                        type: "function",
                        function: { name: "noop", description: "noop", parameters: { type: "object", properties: {} } },
                    },
                ],
                max_tokens: 1,
            }),
            // generous: this may also be the JIT/model warm-up request
            signal: AbortSignal.timeout(120000),
        });
        if (res.ok)
            return "yes";
        // unknown-ok: "" cannot match either marker, so an unreadable body falls
        // through to the `return "unknown"` below rather than to `return "no"` — the
        // three states are already distinct and a failed read lands on the right one.
        const text = (await res.text().catch(() => "")).toLowerCase();
        if (text.includes("jinja") || text.includes("tools")) {
            logger.warn(`[llamacpp] tools probe rejected (server without --jinja?): ${text.slice(0, 160)}`);
            return "no";
        }
        return "unknown";
    }
    catch {
        return "unknown";
    }
}
//# sourceMappingURL=llamacpp-probe.js.map