clef-mcp 0.1.3 → 0.1.5

package/dist/index.d.ts

@@ -191,6 +191,12 @@ interface LlamaCppRuntimeOptions {
     loadTimeoutMs?: number;
     /** Per-request timeout for /v1/systemone. */
     requestTimeoutMs?: number;
+    /**
+     * llama.cpp physical batch size (`-b`/`-ub`). The clef decision path scores
+     * all questions in one sequence, so multi-question requests need more than
+     * the default 512; the clef-mcp limit is 64 questions (~6-8k tokens).
+     */
+    batchSize?: number;
     log?: Logger;
 }
 declare class LlamaCppRuntime implements ClefRuntime {
@@ -202,6 +208,7 @@ declare class LlamaCppRuntime implements ClefRuntime {
     private readonly requestTimeoutMs;
     /** Serializes decide() calls: the llama.cpp clef path handles one sequence per batch. */
     private queue;
+    private readonly batchSize;
     constructor(opts: LlamaCppRuntimeOptions);
     get endpoint(): string;
     get running(): boolean;
@@ -426,6 +433,6 @@ declare function clefDecideOutputSchema(): z.ZodObject<{
 /** `state` cannot be size-checked inside zod (it is `unknown`), so it is validated separately. */
 declare function validateStateSize(state: unknown, limits: Limits): void;
 
-declare const VERSION = "0.1.3";
+declare const VERSION = "0.1.5";
 
 export { CLEF_FLASH, type ClefConfig, type ClefDecideInput, type ClefDecideOutput, type ClefDecisionAnswer, ClefError, ClefErrorCode, type ClefQuestion, type ClefRuntime, type ClefServerDeps, DEFAULT_LIMITS, type DownloadProgress, type InstallModelResult, type Limits, LlamaCppRuntime, type LlamaCppRuntimeOptions, type LogLevel, type Logger, MODELS, type ModelManifest, type ModelSpec, type PlatformInfo, type ProbeResult, type QuantSpec, type QuestionType, type ResolvedBinary, type RunningClefServer, type RuntimeHealth, SERVER_INSTRUCTIONS, StderrLogger, type SystemOneRequest, type SystemOneResponse, TOOL_NAME, VERSION, assertSupportedPlatform, clefDecideInputSchema, clefDecideOutputSchema, createClefServer, createLogger, deepVerifyChecksum, defaultClefHome, detectPlatform, downloadFile, downloadManagedLlama, ensureLlamaRuntime, errorJson, exitCodeFor, fetchFileInfo, findInstalledManifest, formatBytes, fromSystemOneResponse, getModelSpec, getQuantSpec, installModel, limitsFromEnv, loadConfig, manifestPathFor, modelQuantDir, pathsFor, pickQuant, probeInference, readModelManifest, removeManagedRuntime, resolveLlamaServerBinary, toSystemOneRequest, validateStateSize, verifyModelFiles, writeModelManifest };

package/dist/index.js

@@ -394,6 +394,7 @@ var LlamaCppRuntime = class {
     this.host = opts.host ?? "127.0.0.1";
     this.loadTimeoutMs = opts.loadTimeoutMs ?? 12e4;
     this.requestTimeoutMs = opts.requestTimeoutMs ?? 6e4;
+    this.batchSize = opts.batchSize ?? (Number.parseInt(process.env.CLEF_LLAMA_BATCH ?? "", 10) || 8192);
   }
   opts;
   proc;
@@ -403,6 +404,7 @@ var LlamaCppRuntime = class {
   requestTimeoutMs;
   /** Serializes decide() calls: the llama.cpp clef path handles one sequence per batch. */
   queue = Promise.resolve();
+  batchSize;
   get endpoint() {
     if (!this.resolvedPort) throw new ClefError(ClefErrorCode.RUNTIME_INIT_FAILED, "Runtime is not started");
     return `http://${this.host}:${this.resolvedPort}`;
@@ -422,6 +424,10 @@ var LlamaCppRuntime = class {
       "--port",
       String(port),
       "--no-webui",
+      "-b",
+      String(this.batchSize),
+      "-ub",
+      String(this.batchSize),
       ...this.opts.alias ? ["--alias", this.opts.alias] : []
     ];
     this.opts.log?.debug("spawning llama-server", { bin: this.opts.serverBin, args });
@@ -539,7 +545,8 @@ ${tail}` : ""),
       }
       throw new ClefError(
         ClefErrorCode.CLEF_INFERENCE_FAILED,
-        `Inference failed with HTTP ${res.status}: ${apiMessage.slice(0, 2e3)}`
+        `Inference failed with HTTP ${res.status}: ${apiMessage.slice(0, 2e3)}`,
+        /physical batch size/i.test(apiMessage) ? "Too many/too large questions for one request. Split into several calls, reduce criteria text, or raise the batch with CLEF_LLAMA_BATCH." : void 0
       );
     }
     if (!body || typeof body !== "object") {
@@ -1425,7 +1432,7 @@ function createClefServer(deps = {}) {
 }
 
 // src/version.ts
-var VERSION = "0.1.3";
+var VERSION = "0.1.5";
 export {
   CLEF_FLASH,
   ClefError,