Updates

vendor/agent-harness/lib/qvac.js

@@ -11,6 +11,7 @@ const events = require('./events.js');
 const paths = require('./paths.js');
 const completeWatch = require('./complete-watch.js');
 const toolParse = require('./tool-parse.js');
+const repeatLoop = require('./repeat-loop.js');
 
 let sdk = null;
 let initError = null;
@@ -273,6 +274,7 @@ async function load(opts, onProgress) {
   const wantVision =
     opts.vision !== false && (entry ? entry.vision === true : catalog.isVisionModel(opts.model || opts.modelSrc));
   const mmprojSrc = wantVision ? await resolveMmproj(s, entry) : null;
+  const loadParams = catalog.generationParamsForModel((entry && entry.id) || opts.model);
   const modelConfig = Object.assign(
     {
       device: dev.device,
@@ -281,6 +283,7 @@ async function load(opts, onProgress) {
       tools: !!toolsOn,
       'mmproj-use-gpu': !!mmprojGpu,
     },
+    loadParams.reasoning_budget != null ? { reasoning_budget: loadParams.reasoning_budget } : {},
     mmprojSrc ? { projectionModelSrc: mmprojSrc } : {}
   );
   const loadOpts = {
@@ -317,7 +320,7 @@ async function load(opts, onProgress) {
   } catch (err) {
     const msg = flattenError(err);
     log('load failed: ' + msg);
-    if (dev.device === 'gpu' && String(opts.device || 'auto').toLowerCase() !== 'cpu') {
+    if (dev.device === 'gpu' && String(opts.device || 'auto').toLowerCase() === 'auto' && process.env.JARVIS_GPU_REQUIRED !== '1') {
       log('retrying load on cpu');
       modelConfig.device = 'cpu';
       modelConfig.gpu_layers = 0;
@@ -340,8 +343,8 @@ async function load(opts, onProgress) {
     tools: !!toolsOn,
     vision: !!mmprojSrc,
     device: dev.device,
-    backend: bl.backend,
-    backendId: bl.backendId,
+    backend: dev.device === 'gpu' ? bl.backend : 'cpu',
+    backendId: dev.device === 'gpu' ? bl.backendId : 0,
     deviceName: bl.deviceName,
     vram: res && (res.vram || res.vramBytes),
     ctxSize: modelConfig.ctx_size,
@@ -371,9 +374,13 @@ async function complete(opts, onEvent) {
   const history = prepareVisionHistory(rawHistory);
   const tools = toTools(opts && opts.tools);
   const dialect = (opts && opts.toolDialect) || catalog.toolDialectFor(loaded.friendlyId || loaded.constant);
+  const generationParams = catalog.generationParamsForModel(
+    loaded.friendlyId || loaded.constant,
+    opts && opts.generationParams
+  );
   const params = {
     modelId: (opts && opts.modelId) || loaded.modelId,
-    history,
+    history: toolParse.prepareToolHistory(history, dialect),
     stream: opts && opts.stream === false ? false : true,
     captureThinking: true,
   };
@@ -381,7 +388,7 @@ async function complete(opts, onEvent) {
     params.tools = tools;
     params.toolDialect = dialect;
   }
-  if (opts && opts.generationParams) params.generationParams = opts.generationParams;
+  if (Object.keys(generationParams).length) params.generationParams = generationParams;
   const run = s.completion(params);
   const requestId = run.requestId || (opts && opts.requestId) || null;
   loaded.requestId = requestId;
@@ -416,9 +423,19 @@ async function complete(opts, onEvent) {
         if (n.type === 'contentDelta') {
           text += n.delta;
           if (onEvent) onEvent(n);
+          if (repeatLoop.isRepeating(text)) {
+            abortRun();
+            settleTimeout();
+            return;
+          }
         } else if (n.type === 'thinkingDelta') {
           thinking += n.delta;
           if (onEvent) onEvent(n);
+          if (repeatLoop.isRepeating(thinking)) {
+            abortRun();
+            settleTimeout();
+            return;
+          }
         } else if (n.type === 'toolCall') {
           toolCalls.push(n.call);
           if (onEvent) onEvent(n);
@@ -432,6 +449,11 @@ async function complete(opts, onEvent) {
         watch.bump();
         text += token;
         if (onEvent) onEvent({ type: 'contentDelta', delta: token });
+        if (repeatLoop.isRepeating(text)) {
+          abortRun();
+          settleTimeout();
+          return;
+        }
       }
       if (run.toolCallStream) {
         for await (const evt of run.toolCallStream) {
@@ -476,6 +498,8 @@ async function complete(opts, onEvent) {
     } else {
       text = toolParse.stripToolMarkup(text);
     }
+    text = repeatLoop.collapseRepeats(text);
+    thinking = repeatLoop.collapseRepeats(thinking);
     return {
       text,
       thinking,