feat: Add 'The user requested' after thought tags in prompt seeding

codai/models/templates.py

@@ -23,16 +23,16 @@ class AgenticTemplateManager:
     # Family-specific prefixes for Prompt Seeding (force reasoning start)
     # These templates end with the thought tag to force the model to start reasoning
     REASONING_PREFIXES = {
-        "qwen": "<|im_start|>system\n{sys}<|im_end|>\n<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n<think>\n",
-        "deepseek": "<|begin_of_sentence|><|im_start|>system\n{sys}<|im_end|>\n<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n<think>\n",
-        "llama3": "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{sys}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{user}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n<thought>\n",
-        "mistral": "[INST] {sys}\n\n{user} [/INST] Thought:\n",
-        "anthropic": "\n\nSystem: {sys}\n\nHuman: {user}\n\nAssistant: <thinking>\n",
-        "command-r": "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>{sys}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>{user}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><thought>\n",
-        "gemma": "<bos><start_of_turn>user\n{sys}\n\n{user}<end_of_turn>\n<start_of_turn>model\n<thought>\n",
-        "phi3": "<|system|>\n{sys}<|end|>\n<|user|>\n{user}<|end|>\n<|assistant|>\n<|thought|>\n",
-        "yi": "<|im_start|>system\n{sys}<|im_end|>\n<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n<think>\n",
-        "generic": "System: {sys}\nUser: {user}\nAssistant: <think>\n"
+        "qwen": "<|im_start|>system\n{sys}<|im_end|>\n<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n<think>The user requested ",
+        "deepseek": "<|begin_of_sentence|><|im_start|>system\n{sys}<|im_end|>\n<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n<think>The user requested ",
+        "llama3": "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{sys}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{user}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n<thought>The user requested ",
+        "mistral": "[INST] {sys}\n\n{user} [/INST] Thought: The user requested ",
+        "anthropic": "\n\nSystem: {sys}\n\nHuman: {user}\n\nAssistant: <thinking>The user requested ",
+        "command-r": "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>{sys}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>{user}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><thought>The user requested ",
+        "gemma": "<bos><start_of_turn>user\n{sys}\n\n{user}<end_of_turn>\n<start_of_turn>model\n<thought>The user requested ",
+        "phi3": "<|system|>\n{sys}<|end|>\n<|user|>\n{user}<|end|>\n<|assistant|>\n<|thought|>The user requested ",
+        "yi": "<|im_start|>system\n{sys}<|im_end|>\n<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n<think>The user requested ",
+        "generic": "System: {sys}\nUser: {user}\nAssistant: <think>The user requested \n"
     }
     
     # Stop tokens for each family (used to stop reasoning generation)