feat: reasoning gym agentic environments

environments/reasoning_gym_claude_code/README.md

@@ -0,0 +1,50 @@
+# Reasoning Gym — Claude Code Agent
+
+Claude Code agent harness for reasoning gym tasks.
+
+Source benchmark: https://github.com/open-thought/reasoning-gym
+
+## Configuration
+
+Set Anthropic credentials in `env.yaml`:
+
+```yaml
+anthropic_api_key: sk-ant-...
+anthropic_model_name: claude-sonnet-4-6
+anthropic_base_url: null
+```
+
+For a local vLLM or Ollama endpoint that serves the Anthropic Messages API:
+
+```yaml
+anthropic_api_key: EMPTY
+anthropic_model_name: Qwen/Qwen3-4B-Instruct-2507
+anthropic_base_url: http://localhost:8000
+```
+
+`anthropic_base_url` should not include `/v1`. Claude Code appends `/v1/messages` itself.
+
+See [`responses_api_agents/claude_code_agent`](../../responses_api_agents/claude_code_agent/README.md) for the full set of agent options (`thinking`, `max_thinking_tokens`, `allowed_tools`, `disallowed_tools`, `max_turns`, `timeout`, etc.).
+
+## Quick start
+
+```bash
+ng_run "+config_paths=[environments/reasoning_gym_claude_code/config.yaml]"
+```
+
+```bash
+ng_collect_rollouts \
+    +agent_name=reasoning_gym_claude_code_agent \
+    +input_jsonl_fpath=environments/reasoning_gym_claude_code/data/example.jsonl \
+    +output_jsonl_fpath=results/reasoning_gym_claude_code_rollouts.jsonl
+```
+
+## Prepare training data
+
+```bash
+python environments/reasoning_gym_claude_code/prepare.py --task knights_knaves --size 1000 --output environments/reasoning_gym_claude_code/data/train_knights_knaves.jsonl
+```
+
+See `prepare.py` for all available tasks, categories, and config options.
+
+Alternatively, a pre-built dataset is hosted on HuggingFace at [nvidia/Nemotron-RL-ReasoningGym-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-ReasoningGym-v1).

environments/reasoning_gym_claude_code/config.yaml

@@ -0,0 +1,34 @@
+reasoning_gym:
+  resources_servers:
+    reasoning_gym:
+      entrypoint: app.py
+      domain: knowledge
+      verified: true
+      description: Claude Code agent harness for reasoning gym tasks
+      value: Evaluate model capabilities in the Claude Code agent harness
+
+reasoning_gym_claude_code_agent:
+  responses_api_agents:
+    claude_code_agent:
+      entrypoint: app.py
+      resources_server:
+        type: resources_servers
+        name: reasoning_gym
+      concurrency: 32
+      model: ${anthropic_model_name}
+      anthropic_api_key: ${anthropic_api_key}
+      anthropic_base_url: ${anthropic_base_url}
+      max_turns: 30
+      timeout: 300
+      thinking: disabled
+      system_prompt: |
+        You are a precise reasoning assistant. You have access to Bash to run Python for calculations.
+        For every problem: think step by step, use code to verify when helpful, and state your final answer clearly.
+      datasets:
+      - name: example
+        type: example
+        jsonl_fpath: environments/reasoning_gym_claude_code/data/example.jsonl
+      - name: train
+        type: train
+        jsonl_fpath: environments/reasoning_gym_claude_code/data/train_knights_knaves.jsonl
+        license: Apache 2.0

environments/reasoning_gym_claude_code/data/.gitignore

@@ -0,0 +1,5 @@
+*train.jsonl
+*validation.jsonl
+*train_prepare.jsonl
+*validation_prepare.jsonl
+*example_prepare.jsonl

environments/reasoning_gym_claude_code/data/example.jsonl

@@ -0,0 +1,5 @@
+{"responses_create_params": {"input": [{"role": "user", "content": "A very special island is inhabited only by sages and fools. Sages always tell the truth, and fools always lie. You meet 2 inhabitants: Zoey, and Riley. Zoey commented, \"Riley is a fool\". In Riley's words: \"Zoey is a sage or Riley is a sage\". So who is a sage and who is a fool? (Format your answer like: \"Zoey is a sage/fool, and Riley is a sage/fool\")"}]}, "question": "A very special island is inhabited only by sages and fools. Sages always tell the truth, and fools always lie. You meet 2 inhabitants: Zoey, and Riley. Zoey commented, \"Riley is a fool\". In Riley's words: \"Zoey is a sage or Riley is a sage\". So who is a sage and who is a fool? (Format your answer like: \"Zoey is a sage/fool, and Riley is a sage/fool\")", "answer": "Zoey is a fool, and Riley is a sage.", "metadata": {"source_dataset": "knights_knaves", "source_index": 0, "statements": [["lying", 1], ["or", ["telling-truth", 0], ["telling-truth", 1]]], "solution": [false, true], "names": ["Zoey", "Riley"], "knight_knave_terms": {"knight": "sage", "knave": "fool", "a_knight": "a sage", "a_knave": "a fool", "Knight": "Sage", "Knave": "Fool"}, "difficulty": {"n_people": 2, "depth_constraint": 2, "width_constraint": 2}}}
+{"responses_create_params": {"input": [{"role": "user", "content": "A very special island is inhabited only by pioneers and laggards. Pioneers always tell the truth, and laggards always lie. You meet 2 inhabitants: Riley, and Olivia. \"if Riley is a pioneer then Olivia is a laggard\" - Riley. Olivia stated, \"Olivia is a pioneer and Riley is a laggard\". So who is a pioneer and who is a laggard? (Format your answer like: \"Riley is a pioneer/laggard, and Olivia is a pioneer/laggard\")"}]}, "question": "A very special island is inhabited only by pioneers and laggards. Pioneers always tell the truth, and laggards always lie. You meet 2 inhabitants: Riley, and Olivia. \"if Riley is a pioneer then Olivia is a laggard\" - Riley. Olivia stated, \"Olivia is a pioneer and Riley is a laggard\". So who is a pioneer and who is a laggard? (Format your answer like: \"Riley is a pioneer/laggard, and Olivia is a pioneer/laggard\")", "answer": "Riley is a pioneer, and Olivia is a laggard.", "metadata": {"source_dataset": "knights_knaves", "source_index": 1, "statements": [["->", ["telling-truth", 0], ["lying", 1]], ["and", ["telling-truth", 1], ["lying", 0]]], "solution": [true, false], "names": ["Riley", "Olivia"], "knight_knave_terms": {"knight": "pioneer", "knave": "laggard", "a_knight": "a pioneer", "a_knave": "a laggard", "Knight": "Pioneer", "Knave": "Laggard"}, "difficulty": {"n_people": 2, "depth_constraint": 2, "width_constraint": 2}}}
+{"responses_create_params": {"input": [{"role": "user", "content": "A very special island is inhabited only by saints and sinners. Saints always tell the truth, and sinners always lie. You meet 2 inhabitants: Samuel, and Jacob. Samuel expressed that if Samuel is a saint then Jacob is a sinner. Jacob was heard saying, \"if Samuel is a saint then Samuel is a sinner\". So who is a saint and who is a sinner? (Format your answer like: \"Samuel is a saint/sinner, and Jacob is a saint/sinner\")"}]}, "question": "A very special island is inhabited only by saints and sinners. Saints always tell the truth, and sinners always lie. You meet 2 inhabitants: Samuel, and Jacob. Samuel expressed that if Samuel is a saint then Jacob is a sinner. Jacob was heard saying, \"if Samuel is a saint then Samuel is a sinner\". So who is a saint and who is a sinner? (Format your answer like: \"Samuel is a saint/sinner, and Jacob is a saint/sinner\")", "answer": "Samuel is a saint, and Jacob is a sinner.", "metadata": {"source_dataset": "knights_knaves", "source_index": 2, "statements": [["->", ["telling-truth", 0], ["lying", 1]], ["->", ["telling-truth", 0], ["lying", 0]]], "solution": [true, false], "names": ["Samuel", "Jacob"], "knight_knave_terms": {"knight": "saint", "knave": "sinner", "a_knight": "a saint", "a_knave": "a sinner", "Knight": "Saint", "Knave": "Sinner"}, "difficulty": {"n_people": 2, "depth_constraint": 2, "width_constraint": 2}}}
+{"responses_create_params": {"input": [{"role": "user", "content": "A very special island is inhabited only by pioneers and laggards. Pioneers always tell the truth, and laggards always lie. You meet 2 inhabitants: Olivia, and Lily. \"Lily is a laggard and Lily is a pioneer\" - Olivia. Lily noted, \"if Olivia is a pioneer then Lily is a pioneer\". So who is a pioneer and who is a laggard? (Format your answer like: \"Olivia is a pioneer/laggard, and Lily is a pioneer/laggard\")"}]}, "question": "A very special island is inhabited only by pioneers and laggards. Pioneers always tell the truth, and laggards always lie. You meet 2 inhabitants: Olivia, and Lily. \"Lily is a laggard and Lily is a pioneer\" - Olivia. Lily noted, \"if Olivia is a pioneer then Lily is a pioneer\". So who is a pioneer and who is a laggard? (Format your answer like: \"Olivia is a pioneer/laggard, and Lily is a pioneer/laggard\")", "answer": "Olivia is a laggard, and Lily is a pioneer.", "metadata": {"source_dataset": "knights_knaves", "source_index": 3, "statements": [["and", ["lying", 1], ["telling-truth", 1]], ["->", ["telling-truth", 0], ["telling-truth", 1]]], "solution": [false, true], "names": ["Olivia", "Lily"], "knight_knave_terms": {"knight": "pioneer", "knave": "laggard", "a_knight": "a pioneer", "a_knave": "a laggard", "Knight": "Pioneer", "Knave": "Laggard"}, "difficulty": {"n_people": 2, "depth_constraint": 2, "width_constraint": 2}}}
+{"responses_create_params": {"input": [{"role": "user", "content": "A very special island is inhabited only by altruists and egoists. Altruists always tell the truth, and egoists always lie. You meet 2 inhabitants: Mason, and Jack. Mason expressed that if Jack is an egoist then Mason is an altruist. In Jack's words: \"Mason is an egoist\". So who is an altruist and who is an egoist? (Format your answer like: \"Mason is a altruist/egoist, and Jack is a altruist/egoist\")"}]}, "question": "A very special island is inhabited only by altruists and egoists. Altruists always tell the truth, and egoists always lie. You meet 2 inhabitants: Mason, and Jack. Mason expressed that if Jack is an egoist then Mason is an altruist. In Jack's words: \"Mason is an egoist\". So who is an altruist and who is an egoist? (Format your answer like: \"Mason is a altruist/egoist, and Jack is a altruist/egoist\")", "answer": "Mason is an altruist, and Jack is an egoist.", "metadata": {"source_dataset": "knights_knaves", "source_index": 4, "statements": [["->", ["lying", 1], ["telling-truth", 0]], ["not", ["telling-truth", 0]]], "solution": [true, false], "names": ["Mason", "Jack"], "knight_knave_terms": {"knight": "altruist", "knave": "egoist", "a_knight": "an altruist", "a_knave": "an egoist", "Knight": "Altruist", "Knave": "Egoist"}, "difficulty": {"n_people": 2, "depth_constraint": 2, "width_constraint": 2}}}