Expanding Terminus Slicing PR (#597)

Expanding PR to include reward logic for string similarity and schema
validation

---------

Signed-off-by: Khushi Bhardwaj <kbhardwaj@nvidia.com>

Author: kbhardwaj-nvidia

resources_servers/terminus_judge/README.md

@@ -1,11 +1,65 @@
 # Description
 
-Data links: ?
+This is a resource server for verifying terminal-based agent actions. It evaluates agent responses that represent terminal command sequences against expected answers. The server supports two different schema formats (`terminus_1` and `terminus_2`) for terminal interaction tasks.
+
+For each verification request, the agent's JSON output is validated through multiple checks:
+1. **JSON Parsing**: The model output must be valid JSON
+2. **Schema Validation**: The response must conform to the specified harness schema (`terminus_1` or `terminus_2`)
+3. **Task Completion**: If the expected answer requires task completion, the agent must also indicate completion
+4. **Command Correctness**: The predicted keystrokes must exactly match the expected keystrokes in order
+  - This is evaluated via string similarity and equivalency llm-as-judge
+
+
+## Supported Schemas
+
+### terminus_1
+- `state_analysis`: Description of the current terminal state
+- `explanation`: Brief explanation of what the commands will do
+- `commands`: List of command objects with `keystrokes`, `is_blocking`, and `timeout_sec`
+- `is_task_complete`: Boolean indicating if the task is complete
+
+### terminus_2
+- `analysis`: Analysis of the current state based on terminal output
+- `plan`: Description of the plan for next steps
+- `commands`: List of command objects with `keystrokes` and optional `duration`
+- `task_complete`: Boolean indicating if the task is complete (optional)
+
+
+# Example usage
+
+## Running servers
+
+The following command can be used to run this resource server, along with the simple agent and a policy model:
+
+```bash
+config_paths="resources_servers/terminus_judge/configs/terminus_judge.yaml,\
+responses_api_models/openai_model/configs/openai_model.yaml"
+
+ng_run "+config_paths=[$config_paths]" \
+  +terminus_judge_resources_server.resources_servers.terminus_judge.judge_responses_create_params.max_output_tokens=512
+```
+
+Then, rollouts can be collected using a command such as the following:
+
+```bash
+ng_collect_rollouts +agent_name=terminus_judge_simple_agent \
+    +input_jsonl_fpath=resources_servers/terminus_judge/data/example.jsonl \
+    +output_jsonl_fpath=resources_servers/terminus_judge/example_rollouts.jsonl
+```
+
+## Expected Data Format
+
+Each data sample should include:
+- `expected_answer`: A JSON string containing the expected terminal commands
+- `metadata.harness`: Either `"terminus_1"` or `"terminus_2"` to specify the schema format
+- `threshold`: threshold for string similarity to calculate the reward
 
 # Licensing information
-Code: ?
-Data: ?
 
-Dependencies
+Code: Apache 2.0<br>
+Data: TBD
+
+## Dependencies
+
 - nemo_gym: Apache 2.0
-?
+- openapi-schema-validator: BSD-3-Clause