atlasbot: call ollama chat directly

services/comms/atlasbot-deployment.yaml

@@ -16,7 +16,7 @@ spec:
       labels:
         app: atlasbot
       annotations:
-        checksum/atlasbot-configmap: manual-atlasbot-33
+        checksum/atlasbot-configmap: manual-atlasbot-34
         vault.hashicorp.com/agent-inject: "true"
         vault.hashicorp.com/role: "comms"
         vault.hashicorp.com/agent-inject-secret-turn-secret: "kv/data/atlas/comms/turn-shared-secret"
@@ -80,7 +80,7 @@ spec:
             - name: BOT_MENTIONS
               value: atlasbot,aatlasbot
             - name: OLLAMA_URL
-              value: http://chat-ai-gateway.bstein-dev-home.svc.cluster.local/
+              value: http://ollama.ai.svc.cluster.local:11434
             - name: OLLAMA_MODEL
               value: qwen2.5:14b-instruct-q4_0
             - name: OLLAMA_TIMEOUT_SEC

services/comms/scripts/atlasbot/bot.py

@@ -155,6 +155,37 @@ def _ensure_confidence(text: str) -> str:
     return "\n".join(lines)
 
 
+def _ollama_endpoint() -> str:
+    url = (OLLAMA_URL or "").strip()
+    if not url:
+        return ""
+    if url.endswith("/api/chat"):
+        return url
+    return url.rstrip("/") + "/api/chat"
+
+
+def _history_to_messages(lines: list[str]) -> list[dict[str, str]]:
+    messages: list[dict[str, str]] = []
+    for line in lines:
+        raw = (line or "").strip()
+        if not raw:
+            continue
+        role = "user"
+        content = raw
+        lowered = raw.lower()
+        if lowered.startswith("atlas:"):
+            role = "assistant"
+            content = raw.split(":", 1)[1].strip()
+        elif lowered.startswith("user:"):
+            role = "user"
+            content = raw.split(":", 1)[1].strip()
+        elif ":" in raw:
+            content = raw.split(":", 1)[1].strip()
+        if content:
+            messages.append({"role": role, "content": content})
+    return messages
+
+
 # Mention detection (Matrix rich mentions + plain @atlas).
 MENTION_TOKENS = [m.strip() for m in BOT_MENTIONS.split(",") if m.strip()]
 MENTION_LOCALPARTS = [m.lstrip("@").split(":", 1)[0] for m in MENTION_TOKENS]
@@ -1837,25 +1868,33 @@ def _ollama_call(hist_key, prompt: str, *, context: str) -> str:
         "If the answer is not grounded in the provided context or tool data, say you do not know. "
         "End every response with a line: 'Confidence: high|medium|low'."
     )
-    transcript_parts = [system]
+    endpoint = _ollama_endpoint()
+    if not endpoint:
+        raise RuntimeError("ollama endpoint missing")
+    system_content = system
     if context:
-        transcript_parts.append("Context (grounded):\n" + context[:MAX_CONTEXT_CHARS])
+        system_content += "\n\nContext (grounded):\n" + context[:MAX_CONTEXT_CHARS]
-    transcript_parts.extend(history[hist_key][-24:])
-    transcript_parts.append(f"User: {prompt}")
-    transcript = "\n".join(transcript_parts)
 
-    payload = {"model": MODEL, "message": transcript}
+    messages: list[dict[str, str]] = [{"role": "system", "content": system_content}]
+    messages.extend(_history_to_messages(history[hist_key][-24:]))
+    messages.append({"role": "user", "content": prompt})
+
+    payload = {"model": MODEL, "messages": messages, "stream": False}
     headers = {"Content-Type": "application/json"}
     if API_KEY:
         headers["x-api-key"] = API_KEY
-    r = request.Request(OLLAMA_URL, data=json.dumps(payload).encode(), headers=headers)
+    r = request.Request(endpoint, data=json.dumps(payload).encode(), headers=headers)
     lock = _OLLAMA_LOCK if OLLAMA_SERIALIZE else None
     if lock:
         lock.acquire()
     try:
         with request.urlopen(r, timeout=OLLAMA_TIMEOUT_SEC) as resp:
             data = json.loads(resp.read().decode())
-            raw_reply = data.get("message") or data.get("response") or data.get("reply") or data
+            msg = data.get("message") if isinstance(data, dict) else None
+            if isinstance(msg, dict):
+                raw_reply = msg.get("content")
+            else:
+                raw_reply = data.get("response") or data.get("reply") or data
             reply = _normalize_reply(raw_reply) or "I'm here to help."
         history[hist_key].append(f"Atlas: {reply}")
         return reply