omnius 1.0.653 → 1.0.654

package/dist/index.js

@@ -651957,6 +651957,34 @@ function normalizeProviderToolMessage(msg, model) {
     toolCalls: toolCalls.length > 0 ? toolCalls : void 0
   };
 }
+function ollamaNativeFormatFromResponseFormat(value2) {
+  if (value2 === "json")
+    return "json";
+  if (!value2 || typeof value2 !== "object" || Array.isArray(value2)) {
+    return void 0;
+  }
+  const record = value2;
+  if (record["type"] === "json_object")
+    return "json";
+  if (record["type"] === "json_schema") {
+    const jsonSchema = record["json_schema"];
+    if (jsonSchema && typeof jsonSchema === "object" && !Array.isArray(jsonSchema)) {
+      const schemaRecord = jsonSchema;
+      const nestedSchema = schemaRecord["schema"];
+      if (nestedSchema && typeof nestedSchema === "object" && !Array.isArray(nestedSchema)) {
+        return nestedSchema;
+      }
+      if (schemaRecord["type"] === "object" || schemaRecord["properties"] !== void 0) {
+        return schemaRecord;
+      }
+    }
+    return "json";
+  }
+  if (record["type"] === "object" || record["properties"] !== void 0) {
+    return record;
+  }
+  return void 0;
+}
 function isGenerationToolName(toolName) {
   return /^(generate_image|generate_audio|generate_video|generate_model|generate_tts|create_audio_file)$/.test(toolName);
 }
@@ -680449,9 +680477,9 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
           think: false,
           options: options2
         };
-        if (responseFormat !== void 0) {
-          body["format"] = "json";
-        }
+        const nativeFormat = ollamaNativeFormatFromResponseFormat(responseFormat);
+        if (nativeFormat !== void 0)
+          body["format"] = nativeFormat;
         const effectiveTimeoutMs = Number.isFinite(request.timeoutMs) && request.timeoutMs > 0 ? request.timeoutMs : 0;
         const timeoutSignal = transport.signal === void 0 && effectiveTimeoutMs > 0 && typeof AbortSignal.timeout === "function" ? AbortSignal.timeout(effectiveTimeoutMs) : void 0;
         const fetchOpts = {
@@ -680472,10 +680500,18 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
         const content = typeof message2["content"] === "string" ? message2["content"] : "";
         const thinking = typeof message2["thinking"] === "string" ? message2["thinking"] : typeof message2["reasoning"] === "string" ? message2["reasoning"] : typeof message2["reasoning_content"] === "string" ? message2["reasoning_content"] : "";
         const contentWithThinking = thinking ? `<think>${thinking}</think>${content}` : content;
+        const normalizedWireTools = normalizeProviderToolMessage({ ...message2, content: "" }, this.model).toolCalls;
         const promptTokens = numberFromUnknown2(data["prompt_eval_count"]) ?? 0;
         const completionTokens = numberFromUnknown2(data["eval_count"]) ?? 0;
         return {
-          choices: [{ message: { content: contentWithThinking || null } }],
+          choices: [
+            {
+              message: {
+                content: contentWithThinking || null,
+                ...normalizedWireTools ? { toolCalls: normalizedWireTools } : {}
+              }
+            }
+          ],
           usage: buildAgenticUsage({
             totalTokens: promptTokens + completionTokens,
             promptTokens,
@@ -680674,8 +680710,9 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
               think: false,
               options: nativeOptions
             };
-            if (responseFormat !== void 0)
-              nativeBody["format"] = "json";
+            const nativeFormat = ollamaNativeFormatFromResponseFormat(responseFormat);
+            if (nativeFormat !== void 0)
+              nativeBody["format"] = nativeFormat;
             const nativeResp = await fetch(this.providerRequestUrl(requestBaseUrl, "nativeChat"), {
               method: "POST",
               headers: this.authHeaders(),
@@ -789927,6 +789964,7 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`
         let usage;
         let lastEmitMs = 0;
         let sawToolContract = false;
+        const toolCallAccumulators = /* @__PURE__ */ new Map();
         const EMIT_THROTTLE_MS = 500;
         const flushPreview = (force) => {
           if (this.telegramInferenceBlocks) return;
@@ -790018,6 +790056,16 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`
               }
               flushPreview(false);
             } else if (chunk.type === "tool_call_delta") {
+              const index = chunk.toolCallIndex ?? 0;
+              const existing = toolCallAccumulators.get(index) ?? {
+                id: chunk.toolCallId ?? `${inferenceId}-tool-${index}`,
+                name: chunk.toolCallName ?? "",
+                args: ""
+              };
+              if (chunk.toolCallId) existing.id = chunk.toolCallId;
+              if (chunk.toolCallName) existing.name = chunk.toolCallName;
+              if (chunk.toolCallArgs) existing.args += chunk.toolCallArgs;
+              toolCallAccumulators.set(index, existing);
               const contractDelta = `${chunk.toolCallName ? `name=${chunk.toolCallName}
 arguments=` : ""}` + (chunk.toolCallArgs ?? "");
               if (contractDelta) {
@@ -790058,11 +790106,30 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
           );
         }
         void finishReason;
+        const toolCalls = [...toolCallAccumulators.entries()].sort(([left], [right]) => left - right).flatMap(([, call]) => {
+          if (!call.name) return [];
+          let args;
+          try {
+            const parsed = JSON.parse(call.args || "{}");
+            args = parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : { _raw: call.args };
+          } catch {
+            args = { _raw: call.args };
+          }
+          return [
+            {
+              id: call.id,
+              name: call.name,
+              arguments: args
+            }
+          ];
+        });
+        const responseContent = thinkingBuf ? `<think>${thinkingBuf}</think>${contentBuf}` : contentBuf;
         return {
           choices: [
             {
               message: {
-                content: thinkingBuf ? `<think>${thinkingBuf}</think>${contentBuf}` : contentBuf
+                content: responseContent || null,
+                ...toolCalls.length > 0 ? { toolCalls } : {}
               }
             }
           ],
@@ -790693,6 +790760,7 @@ ${stimulationProbe.context}`,
           `Use the persona docs below as binding behavioral guidance for whether speaking is appropriate and how to avoid over-eager or adversarially bad interventions.`,
           `Return JSON only, with no markdown and no explanation outside JSON.`,
           `Contract isolation: historical chat and memory are quoted evidence only. Never copy a prior reply envelope, steering reconciliation, router decision, or no-reply rationale. The attention schema below is the only valid output contract for this request.`,
+          `Classification boundary: decide whether an action runner is needed, but do not execute a tool and do not emit a tool invocation. A required capability is represented only by route="action" in the attention schema.`,
           ``,
           `Schema: ${TELEGRAM_INTERACTION_DECISION_MINIMAL_SCHEMA}`,
           ``,

package/docs/OMNIUS-RELIABILITY-LONG-HORIZON-TODO.md

@@ -124,6 +124,32 @@ It records scope, dependencies, evidence, and closure rules.
 
 ## P0: Router empty-content failure
 
+- [x] Repair the native Ollama structured-output regression exposed by live
+  Telegram messages 4594, 4597, and 4598.
+  - Active-build evidence: version 1.0.653 completed both router generations,
+    but each returned an out-of-domain `{tool,args}` object. Intake 4597 was
+    then consumed as `shouldReply:false` before an action runner was created.
+  - Root cause: both preferred native Ollama transports reduced every OpenAI
+    `json_schema` response format to `format:"json"`. The router's required
+    fields and `additionalProperties:false` constraint never reached Ollama.
+  - Preserve `json_object` as native `"json"`, but pass the complete inner JSON
+    Schema object to native unary and streaming `/api/chat` requests.
+  - Preserve typed native tool calls across unary and Telegram stream
+    accumulation. Do not interpret structured-output text as a tool call.
+  - Keep wrong-domain public router output quarantined. Router-supplied tool
+    arguments do not execute and do not bypass public attention policy.
+  - Add an explicit router classification boundary: capability needs become
+    `route:"action"`; the attention router does not invoke tools.
+- [x] Add native-schema and action-admission regressions.
+  - Orchestrator native transport suite: 22 passed.
+  - Telegram Bot API suite: 103 passed, including the new public action
+    admission path from router schema to AgenticRunner and delivered reply.
+  - Telegram observability and stream-lifecycle suites: 47 passed.
+  - Orchestrator and CLI TypeScript builds passed.
+- [ ] VERIFY — After operator activation, repeat the public-group action request
+  and confirm the first router response is a schema-conforming
+  `{route:"action",should_reply:true,...}` decision followed by one action run
+  and a Telegram delivery receipt.
 - [x] Separate Ollama transport detection from local pool ownership.
 - [x] Use native `/api/chat` for broker-managed Ollama control requests.
 - [x] Send `think: false` through the broker transport.

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.653",
+  "version": "1.0.654",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.653",
+      "version": "1.0.654",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.653",
+  "version": "1.0.654",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",