fix: remove excessive rllm-model-gateway reference

Author: luyuzhe111

examples/strands_appworld_agent/README.md

@@ -82,7 +82,7 @@ curl -X POST http://localhost:8080/invocations \
 
 You should see the rollout and reward saved as a json file in s3 folder `s3://agentcore-rl/appworld_test/prompt_123`.
 
-> **Note:** The `_rollout` config must include `base_url` and `model_id`, which tell the agent which inference server to use. The remaining fields (`exp_id`, `s3_bucket`, `session_id`, `input_id`) control S3 result storage and are optional — if omitted, S3 save will be skipped. During RL training, `base_url` can point to [rllm-model-gateway](https://github.com/rllm-org/rllm/tree/main/rllm-model-gateway) for automatic token capture. For evaluation, `base_url` points directly to any OpenAI-compatible inference endpoint.
+> **Note:** The `_rollout` config must include `base_url` and `model_id`, which tell the agent which inference server to use. The remaining fields (`exp_id`, `s3_bucket`, `session_id`, `input_id`) control S3 result storage and are optional — if omitted, S3 save will be skipped.
 
 ## Docker
 

examples/strands_math_agent/README.md

@@ -121,8 +121,6 @@ curl -X POST http://localhost:8080/invocations \
 You should see the rollout and reward saved to `s3://agentcore-rl/test/prompt_123_session_123.json`.
 
 > **Note:** The `_rollout` config must include `base_url` and `model_id`, which tell the agent which inference server to use. The remaining fields (`exp_id`, `s3_bucket`, `session_id`, `input_id`) control S3 result storage and are optional — if omitted, S3 save will be skipped. During training, the full `_rollout` config is injected automatically by the training engine. We recommend testing the full flow locally.
->
-> During RL training, `base_url` can point to [rllm-model-gateway](https://github.com/rllm-org/rllm/tree/main/rllm-model-gateway) for automatic token capture. For evaluation, `base_url` points directly to any OpenAI-compatible inference endpoint (vLLM, SGLang, LiteLLM, etc.).
 
 
 ## Run RL App Hosted on ACR

examples/strands_math_agent/rl_app.py

@@ -26,10 +26,6 @@ def invoke_agent(payload: dict):
     - answer: ground truth (str)
     - _rollout: rollout config with base_url and model_id
 
-    During RL training, base_url can point to rllm-model-gateway for automatic
-    token capture. For evaluation, base_url points directly to any
-    OpenAI-compatible inference endpoint (vLLM, SGLang, LiteLLM, etc.).
-
     The @rollout_entrypoint decorator automatically:
     - Executes the function in the background for non-blocking processing
     - Saves results to S3 with a predictable key