@otto-code/brain 0.8.19 → 0.9.2

package/dist/models/scan.js

@@ -9,12 +9,95 @@ import os from "node:os";
 import path from "node:path";
 import * as gguf from "../gguf.js";
 export const LMSTUDIO_MODELS_DIR = path.join(os.homedir(), ".lmstudio", "models");
-// The quant is the terminal GGUF filename suffix, optionally preceded by a
-// source-defined qualifier such as Muse's `UD-`. Matching a known substring
-// truncates newer labels (for example Q2_K_XL -> Q2_K) and misses them when
-// their exact spelling is not in a hard-coded list.
-const QUANT_SUFFIX = /(?:^|[-_])((?:UD-)?(?:IQ[1-4]|Q[2-8])(?:_[A-Z0-9]+)*|NVFP\d+|MXFP\d+|BF16|F16|F32)(?:-(?:MTP|IMATRIX|DISTILL))?(?:-\d{5}-OF-\d{5})?\.GGUF$/i;
 const MULTIPART = /-(\d{5})-of-(\d{5})\.gguf$/i;
+const QUANT_TRAILING_MARKERS = ["-MTP", "-IMATRIX", "-DISTILL"];
+function isDigit(char) {
+    return char >= "0" && char <= "9";
+}
+function isUppercaseLetterOrDigit(char) {
+    return (char >= "A" && char <= "Z") || isDigit(char);
+}
+function isQuant(value) {
+    if (["BF16", "F16", "F32"].includes(value))
+        return true;
+    if (value.startsWith("IQ")) {
+        return value.length === 3 && value[2] >= "1" && value[2] <= "4";
+    }
+    if (value.startsWith("NVFP") || value.startsWith("MXFP")) {
+        const digits = value.slice(4);
+        return digits.length > 0 && [...digits].every(isDigit);
+    }
+    if (!value.startsWith("Q") || value.length < 2 || value[1] < "2" || value[1] > "8") {
+        return false;
+    }
+    let separatorExpected = false;
+    for (const char of value.slice(2)) {
+        if (char === "_") {
+            if (separatorExpected)
+                return false;
+            separatorExpected = true;
+        }
+        else if (isUppercaseLetterOrDigit(char)) {
+            separatorExpected = false;
+        }
+        else {
+            return false;
+        }
+    }
+    return !separatorExpected;
+}
+function startsQuant(value, index) {
+    const first = value[index]?.toUpperCase();
+    const second = value[index + 1]?.toUpperCase();
+    if (first === "Q")
+        return second >= "2" && second <= "8";
+    if (first === "I")
+        return second === "Q" && value[index + 2] >= "1" && value[index + 2] <= "4";
+    if (first === "N")
+        return value.slice(index, index + 4).toUpperCase() === "NVFP";
+    if (first === "M")
+        return value.slice(index, index + 4).toUpperCase() === "MXFP";
+    if (first === "B")
+        return value.slice(index, index + 4).toUpperCase() === "BF16";
+    if (first === "F") {
+        const prefix = value.slice(index, index + 3).toUpperCase();
+        return prefix === "F16" || prefix === "F32";
+    }
+    return false;
+}
+function stripQuantSuffix(filename) {
+    let stem = path.basename(filename);
+    if (!stem.toLowerCase().endsWith(".gguf"))
+        return null;
+    stem = stem.slice(0, -".gguf".length);
+    const multipart = stem.slice(-15);
+    if (multipart.length === 15 &&
+        multipart[0] === "-" &&
+        multipart[6] === "-" &&
+        multipart.slice(7, 9).toLowerCase() === "of" &&
+        multipart[9] === "-" &&
+        [...multipart.slice(1, 6), ...multipart.slice(10)].every(isDigit)) {
+        stem = stem.slice(0, -multipart.length);
+    }
+    const upperStem = stem.toUpperCase();
+    for (const marker of QUANT_TRAILING_MARKERS) {
+        if (upperStem.endsWith(marker)) {
+            stem = stem.slice(0, -marker.length);
+            break;
+        }
+    }
+    let candidateStart = -1;
+    for (let index = 0; index < stem.length; index += 1) {
+        if (index > 0 && stem[index - 1] !== "-" && stem[index - 1] !== "_")
+            continue;
+        if (startsQuant(stem, index))
+            candidateStart = index;
+    }
+    if (candidateStart === -1)
+        return null;
+    const candidate = stem.slice(candidateStart).toUpperCase();
+    return isQuant(candidate) ? candidate : null;
+}
 function walk(dir, out = []) {
     let entries;
     try {
@@ -33,9 +116,10 @@ function walk(dir, out = []) {
     return out;
 }
 export function detectQuant(filename) {
-    const match = path.basename(filename).match(QUANT_SUFFIX);
-    // `UD-` is a source filename qualifier, not part of the user-facing quant.
-    return match?.[1]?.replace(/^UD-/i, "").toUpperCase() ?? null;
+    // A source qualifier such as `UD-` is not part of the user-facing quant.
+    // Scan terminal candidates rather than matching a nested filename regex:
+    // model filenames are host input and this remains linear for malformed names.
+    return stripQuantSuffix(filename);
 }
 export function isProjectorFile(filename) {
     return /^mmproj/i.test(path.basename(filename));

package/dist/runtime/managed.js

@@ -249,7 +249,11 @@ function slug(spec) {
         .replace(/(^-|-$)/g, "");
 }
 function displayNameForManagedRuntime(label, version) {
-    return `${label.replace(/\s*\(managed\)$/iu, "")} · ${version} (Otto managed)`;
+    const suffix = "(managed)";
+    const withoutSuffix = label.toLowerCase().endsWith(suffix)
+        ? label.slice(0, -suffix.length).trimEnd()
+        : label;
+    return `${withoutSuffix} · ${version} (Otto managed)`;
 }
 function legacyManagedDisplayName(dirName, version) {
     const cuda = /^cuda-(\d+)-(\d+)-managed(?:-|$)/iu.exec(dirName);

package/dist/service/host-api.js

@@ -355,6 +355,11 @@ export function createHostApi(deps) {
         sendError(res, 403, "remote configuration is disabled on this brain; enable it with `otto brain share --allow-config`");
         return false;
     };
+    /** Keep implementation details in the local service log, never the HTTP response. */
+    const sendOperationError = (res, status, operation, error) => {
+        deps.log?.("server", `${operation}: ${errorMessage(error)}`);
+        sendError(res, status, operation);
+    };
     const inventory = async () => {
         const [gpu, store] = [await deps.queryGpuInfo(), deps.getProfilesStore()];
         const defaults = deps.getProfileDefaults();
@@ -445,7 +450,7 @@ export function createHostApi(deps) {
                     });
                 }
                 catch (error) {
-                    sendError(res, 400, errorMessage(error));
+                    sendOperationError(res, 400, "could not update the model profile", error);
                 }
             })();
         });
@@ -537,7 +542,7 @@ export function createHostApi(deps) {
                 });
             }
             catch (error) {
-                sendError(res, 400, errorMessage(error));
+                sendOperationError(res, 400, "could not calculate the model budget", error);
             }
         })();
     };
@@ -630,7 +635,7 @@ export function createHostApi(deps) {
             });
         }
         catch (error) {
-            sendError(res, 409, errorMessage(error));
+            sendOperationError(res, 409, "could not remove the model component", error);
         }
     };
     const handleLogs = (res, params) => {
@@ -778,7 +783,7 @@ export function createHostApi(deps) {
                     });
                 }
                 catch (error) {
-                    sendError(res, 409, errorMessage(error));
+                    sendOperationError(res, 409, "could not start the benchmark", error);
                 }
             });
             return true;
@@ -916,7 +921,7 @@ export function createHostApi(deps) {
                     });
                 }
                 catch (error) {
-                    sendError(res, 400, errorMessage(error));
+                    sendOperationError(res, 400, "could not start the job", error);
                 }
             });
             return true;
@@ -1091,7 +1096,7 @@ export function createHostApi(deps) {
                     }));
                 }
                 catch (error) {
-                    sendError(res, 500, errorMessage(error));
+                    sendOperationError(res, 500, "could not build the model details", error);
                 }
             })();
             return true;

package/dist/service/log-format.js

@@ -1,12 +1,20 @@
-const TAGGED_LINE = /^\[(?:brain|llama-server)\]/u;
-const SOURCE_AND_AREA = /^(\[(?:brain|llama-server)\])(?:\s+(\[(?:library|model|api|server)\]))?\s*(.*)$/u;
 const LLAMA_SERVER_PREFIX = /^\d+(?:\.\d+){3}\s+[A-Z]\s+\S+\s+(?:\S+:\s+)?(.+)$/u;
+const SOURCES = ["brain", "llama-server"];
+const AREAS = ["library", "model", "api", "server"];
+function taggedValue(line, values) {
+    for (const value of values) {
+        const tag = `[${value}]`;
+        if (line.startsWith(tag))
+            return tag;
+    }
+    return null;
+}
 /**
  * Every service-owned event carries both its process source and operation area.
  * llama-server output is separately marked by `formatLlamaServerLog`.
  */
 export function formatBrainLog(area, message) {
-    return TAGGED_LINE.test(message) ? message : `[brain] [${area}] ${message}`;
+    return taggedValue(message, SOURCES) ? message : `[brain] [${area}] ${message}`;
 }
 /**
  * Remove llama.cpp's elapsed-time, level and component columns. Otto owns the
@@ -18,15 +26,21 @@ export function stripLlamaServerPrefix(message) {
 }
 /** Preserve the useful llama.cpp message while making its process boundary explicit. */
 export function formatLlamaServerLog(message) {
-    return TAGGED_LINE.test(message) ? message : `[llama-server] ${stripLlamaServerPrefix(message)}`;
+    return taggedValue(message, SOURCES)
+        ? message
+        : `[llama-server] ${stripLlamaServerPrefix(message)}`;
 }
 /** Place source tags ahead of the timestamp so they are scannable in a dense log. */
 export function timestampBrainLogLine(timestamp, line) {
     const tagged = formatBrainLog("server", line);
-    const match = SOURCE_AND_AREA.exec(tagged);
-    if (!match)
+    const source = taggedValue(tagged, SOURCES);
+    if (!source)
         return `${timestamp} ${tagged}`;
-    const [, source, area, message] = match;
+    let remainder = tagged.slice(source.length).trimStart();
+    const area = taggedValue(remainder, AREAS);
+    if (area)
+        remainder = remainder.slice(area.length).trimStart();
+    const message = remainder;
     return `${source}${area ? ` ${area}` : ""} ${timestamp}${message ? ` ${message}` : ""}`;
 }
 //# sourceMappingURL=log-format.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@otto-code/brain",
-  "version": "0.8.19",
+  "version": "0.9.2",
   "description": "Otto Brain - self-contained host for local GGUF models, with measured VRAM budgeting and reasoning-budget control",
   "license": "AGPL-3.0-or-later",
   "bin": {