omnius 1.0.135 → 1.0.136

package/dist/index.js

@@ -1374,7 +1374,7 @@ var init_model_broker = __esm({
     DEFAULT_IDLE_EVICT_MS = 5 * 60 * 1e3;
     DEFAULT_POLL_MS = 4e3;
     DEFAULT_INFLIGHT_WAIT_MS = 6e4;
-    DEFAULT_SLOT_CAPACITY = 4;
+    DEFAULT_SLOT_CAPACITY = 8;
     DEFAULT_QUEUE_CAPACITY = 50;
     THROUGHPUT_EMA_ALPHA = 0.2;
     THROUGHPUT_INITIAL_TPS = 25;
@@ -1977,6 +1977,33 @@ var init_model_broker = __esm({
           return Promise.resolve(slot);
         }
         return new Promise((resolve55, reject) => {
+          if (this._slotQueue.length >= this.queueCapacity) {
+            const newPrio = spec.priority ?? 0;
+            let victim = -1;
+            let victimPrio = Infinity;
+            for (let i2 = this._slotQueue.length - 1; i2 >= 0; i2--) {
+              const p2 = this._slotQueue[i2].spec.priority ?? 0;
+              if (p2 < victimPrio) {
+                victimPrio = p2;
+                victim = i2;
+              }
+              if (victimPrio === 0)
+                break;
+            }
+            if (victim >= 0 && victimPrio < newPrio) {
+              const dropped = this._slotQueue.splice(victim, 1)[0];
+              if (dropped.onSignalAbort && dropped.spec.signal) {
+                dropped.spec.signal.removeEventListener("abort", dropped.onSignalAbort);
+              }
+              try {
+                dropped.reject(new Error("broker queue shed: capacity reached, lower-priority entry displaced"));
+              } catch {
+              }
+            } else {
+              reject(new Error(`broker queue full (capacity=${this.queueCapacity}); caller priority ${newPrio} insufficient to displace`));
+              return;
+            }
+          }
           const entry = { spec, resolve: resolve55, reject, enqueuedAt: Date.now() };
           if (spec.signal) {
             const onAbort = () => {
@@ -618138,6 +618165,32 @@ function estimatePromptTokensFromRequest(request) {
   }
   return Math.ceil(chars / 4);
 }
+function isLikelyTruncatedRouterJson(text) {
+  if (typeof text !== "string") return false;
+  const stripped = text.replace(/^\s*<think>[\s\S]*?<\/think>\s*/i, "").trim();
+  if (!stripped.startsWith("{")) return false;
+  let depth = 0;
+  let inString = false;
+  let escape2 = false;
+  for (let i2 = 0; i2 < stripped.length; i2++) {
+    const ch = stripped[i2];
+    if (escape2) {
+      escape2 = false;
+      continue;
+    }
+    if (inString) {
+      if (ch === "\\") escape2 = true;
+      else if (ch === '"') inString = false;
+      continue;
+    }
+    if (ch === '"') inString = true;
+    else if (ch === "{") depth++;
+    else if (ch === "}") depth--;
+  }
+  if (depth <= 0) return false;
+  const hits = (stripped.includes('"route"') ? 1 : 0) + (stripped.includes('"should_reply"') ? 1 : 0) + (stripped.includes('"confidence"') ? 1 : 0) + (stripped.includes('"reason"') ? 1 : 0) + (stripped.includes('"silent_disposition"') ? 1 : 0) + (stripped.includes('"mental_note"') ? 1 : 0);
+  return hits >= 3;
+}
 function telegramRouterTimeoutMs(configTimeoutMs, _minMs, _legacyMaxMs) {
   void _minMs;
   void _legacyMaxMs;
@@ -623949,7 +624002,8 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
               ],
               tools: [],
               temperature: 0,
-              maxTokens: 650,
+              // Reflection has 12 string fields; 650 was tight enough to truncate.
+              maxTokens: 1500,
               timeoutMs: telegramRouterTimeoutMs(timeoutMs),
               think: false
             },
@@ -624039,7 +624093,8 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
         const promptTokens = estimatePromptTokensFromRequest(request);
         const broker = getModelBroker();
         const trainCtx = await broker.getNctxTrain(model).catch(() => null);
-        const targetCtx = trainCtx && trainCtx > 0 ? Math.min(trainCtx, Math.max(2048, promptTokens + 1024)) : Math.max(2048, promptTokens + 1024);
+        const completionHeadroom = 4096;
+        const targetCtx = trainCtx && trainCtx > 0 ? Math.min(trainCtx, Math.max(2048, promptTokens + completionHeadroom)) : Math.max(2048, promptTokens + completionHeadroom);
         const requestWithCtx = { ...request, numCtx: targetCtx };
         const slot = await broker.acquireInferenceSlot({
           model,
@@ -624049,10 +624104,12 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
           promptTokens,
           priority: kind === "router" || kind === "router-repair" || kind === "router-strict-retry" ? 1 : 0
         });
-        this.tuiWrite(() => renderTelegramSubAgentEvent(
-          sessionKey,
-          `inference admitted [${kind}] model=${model} prompt~${promptTokens}t num_ctx=${targetCtx} slot=${slot.info.id}${slot.info.reserved ? " reserved" : ""}`
-        ));
+        if (process.env["OMNIUS_BROKER_TRACE"] === "1") {
+          this.tuiWrite(() => renderTelegramSubAgentEvent(
+            sessionKey,
+            `inference admitted [${kind}] model=${model} prompt~${promptTokens}t num_ctx=${targetCtx} slot=${slot.info.id}${slot.info.reserved ? " reserved" : ""}`
+          ));
+        }
         const streamFn = backend.chatCompletionStream;
         const id = this.registerTelegramInference(kind, sessionKey, model);
         let completionTokens = 0;
@@ -624274,7 +624331,7 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
       getTelegramThinkingVisible() {
         return this.telegramThinkingVisible;
       }
-      async repairTelegramInteractionDecision(backend, rawOutput, forcedRoute, timeoutMs, diagnostics) {
+      async repairTelegramInteractionDecision(backend, rawOutput, forcedRoute, timeoutMs, diagnostics, sessionKey = "__router__") {
         const rawPreview = telegramRouterRawPreview(rawOutput, 4e3);
         if (!rawPreview || telegramDecisionOutputHasDanglingJson(rawOutput)) {
           if (diagnostics) {
@@ -624309,10 +624366,10 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
             ],
             tools: [],
             temperature: 0,
-            maxTokens: 500,
+            maxTokens: 1500,
             timeoutMs: telegramRouterTimeoutMs(timeoutMs, 8e3, 2e4),
             think: false
-          });
+          }, diagnostics, "router-repair", sessionKey);
           const repairedText = result.choices[0]?.message?.content ?? "";
           if (telegramDecisionRecoverableFlag(repairedText) === false) {
             if (diagnostics) diagnostics.repairStatus = "no-recoverable-output";
@@ -624344,7 +624401,7 @@ ${repairedText}`,
           return null;
         }
       }
-      async retryTelegramInteractionDecisionStrict(backend, userPrompt, rawOutput, forcedRoute, timeoutMs, diagnostics) {
+      async retryTelegramInteractionDecisionStrict(backend, userPrompt, rawOutput, forcedRoute, timeoutMs, diagnostics, sessionKey = "__router__") {
         const invalidPreview = telegramRouterRawPreview(rawOutput, 1200) ?? "(empty assistant content)";
         const routeInstruction = forcedRoute ? `The operator selected Telegram mode "${forcedRoute}". The route field must be "${forcedRoute}", but should_reply must still be inferred from context.` : `Infer route live from context.`;
         const trimmedUserPrompt = userPrompt.length > 4e3 ? `…
@@ -624376,10 +624433,10 @@ ${userPrompt.slice(-4e3)}` : userPrompt;
             ],
             tools: [],
             temperature: 0,
-            maxTokens: 1200,
+            maxTokens: 2400,
             timeoutMs: telegramRouterTimeoutMs(timeoutMs, 1e4, 3e4),
             think: false
-          });
+          }, diagnostics, "router-strict-retry", sessionKey);
           const retryText = result.choices[0]?.message?.content ?? "";
           if (diagnostics) diagnostics.strictRetryPreview = telegramRouterRawPreview(retryText, 320);
           const parsed = parseTelegramInteractionDecision(retryText, forcedRoute, {
@@ -624762,10 +624819,14 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`
             ],
             tools: [],
             temperature: 0,
-            maxTokens: 1e3,
+            // Router JSON schema has ~18 string-valued fields when reflection is
+            // embedded (consolidated mode). 1000 tokens was the documented cause
+            // of truncated JSON → repair → strict-retry cascade. 2400 is enough
+            // for normal verbose values without slowing the call appreciably.
+            maxTokens: 2400,
             timeoutMs: telegramRouterTimeoutMs(config.timeoutMs),
             think: false
-          }, diagnostics);
+          }, diagnostics, "router", sessionKey);
           const text = result.choices[0]?.message?.content ?? "";
           const routerLatencyMs = Date.now() - routerStartMs;
           try {
@@ -624788,12 +624849,40 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`
           if (parsed) {
             return this.applyTelegramSilentReflectionNotes(parsed, reflectionNotes);
           }
+          if (isLikelyTruncatedRouterJson(text)) {
+            if (diagnostics) diagnostics.repairStatus = "skipped-truncation-rerun";
+            try {
+              const reissued = await this.telegramRouterJsonCompletion(backend, {
+                messages: [
+                  {
+                    role: "system",
+                    content: "You perform live Telegram route and stimulation inference. Output strict JSON only."
+                  },
+                  { role: "user", content: userPrompt }
+                ],
+                tools: [],
+                temperature: 0,
+                maxTokens: 4096,
+                timeoutMs: telegramRouterTimeoutMs(config.timeoutMs),
+                think: false
+              }, diagnostics, "router", sessionKey);
+              const reissuedText = reissued.choices[0]?.message?.content ?? "";
+              const reparsed = parseTelegramInteractionDecision(reissuedText, forcedRoute, {
+                defaultShouldReply: false
+              });
+              if (reparsed) {
+                return this.applyTelegramSilentReflectionNotes(reparsed, reflectionNotes);
+              }
+            } catch {
+            }
+          }
           const repaired = await this.repairTelegramInteractionDecision(
             backend,
             text,
             forcedRoute,
             config.timeoutMs ?? 3e4,
-            diagnostics
+            diagnostics,
+            sessionKey
           );
           if (repaired) {
             return this.applyTelegramSilentReflectionNotes(repaired, reflectionNotes);
@@ -624804,7 +624893,8 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`
             text,
             forcedRoute,
             config.timeoutMs ?? 3e4,
-            diagnostics
+            diagnostics,
+            sessionKey
           );
           if (strictRetry) {
             return this.applyTelegramSilentReflectionNotes(strictRetry, reflectionNotes);

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.135",
+  "version": "1.0.136",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.135",
+      "version": "1.0.136",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.135",
+  "version": "1.0.136",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",