Issues

• openai_model: a reasoning item without summary fails validation with a 500, ending Claude Code sessions

  qa_bug
  NVIDIA-NeMo/Gym#5663 · Filter by author qiaochuz-nv opened on Nov 5, 2026

• codex_agent: trajectory tool calls carry Codex's executed shell command, not the model's call

  qa_bug
  NVIDIA-NeMo/Gym#5651 · Filter by author qiaochuz-nv opened on Nov 5, 2026

• trajectory: ATIF export drops cached input tokens

  qa_bug
  NVIDIA-NeMo/Gym#5644 · Filter by author qiaochuz-nv opened on Nov 4, 2026

• claude_code_agent: the shipped config cannot start a sandbox session

  qa_bug
  NVIDIA-NeMo/Gym#5643 · Filter by author qiaochuz-nv opened on Nov 4, 2026

• harbor: a verifier that exits non-zero is recorded as a measured result

  qa_bug
  NVIDIA-NeMo/Gym#5642 · Filter by author qiaochuz-nv opened on Nov 4, 2026

• harbor: reward.json without a reward key is scored as the sum of its fields

  qa_bug
  NVIDIA-NeMo/Gym#5641 · Filter by author qiaochuz-nv opened on Nov 4, 2026

• skills: skills.path is refused for task-folder targets, with no harness-level alternative

  qa_bug
  NVIDIA-NeMo/Gym#5640 · Filter by author qiaochuz-nv opened on Nov 4, 2026

• harbor: tasks that declare [environment] skills_dir are refused for Claude Code, Codex and OpenCode

  qa_bug
  NVIDIA-NeMo/Gym#5639 · Filter by author qiaochuz-nv opened on Nov 4, 2026

• openclaw_agent test_exec_transport deadline tests fail intermittently under four workers per shard

  NVIDIA-NeMo/Gym#5638 · Filter by author AdamRajfer opened on Nov 4, 2026
  

• Model server base class is named for one API but serves several

  area:model
  NVIDIA-NeMo/Gym#5615 · Filter by author cwing-nvidia opened on Nov 4, 2026

• token capture: terminal attribution misses OpenCode answers split from a Gym reply's reasoning block

  NVIDIA-NeMo/Gym#5609 · Filter by author gwarmstrong opened on Nov 4, 2026

• BYO Hermes agent

  NVIDIA-NeMo/Gym#5588 · Filter by author hwolff99 opened on Nov 4, 2026

• Add a chess environment (Gymnasium-style, built-in opponent)

  community-request
  NVIDIA-NeMo/Gym#5566 · Filter by author Pd172944 opened on Nov 4, 2026
  1

• claude_code_agent: eval output folds the model's thinking into the graded answer

  NVIDIA-NeMo/Gym#5565 · Filter by author gwarmstrong opened on Nov 4, 2026
  

• Session harnesses: close the gaps with the deprecated sandboxed agents for evaluation frameworks

  NVIDIA-NeMo/Gym#5552 · Filter by author gwarmstrong opened on Nov 3, 2026

• MLflow post-run export: take the tracking URI and token from the environment like W&B

  area:evaluationfeature
  NVIDIA-NeMo/Gym#5499 · Filter by author prokotg opened on Nov 3, 2026
  

• Print complete next-step commands from gym dataset init in CLI

  NVIDIA-NeMo/Gym#5477 · Filter by author hwolff99 opened on Nov 3, 2026

• docs: split the harness guide and other overlong pages

  area:docsdocs
  NVIDIA-NeMo/Gym#5472 · Filter by author adil-a opened on Nov 2, 2026

• gym eval submit: head node IP from getent hosts can be a multi-line list of IPv6 link-local addresses, breaking ray start --address

  NVIDIA-NeMo/Gym#5454 · Filter by author phtran8 opened on Nov 2, 2026

• gym eval submit: #SBATCH --gpus-per-node is always emitted, so the shipped multi-node example cannot be submitted on a cluster without GRES

  NVIDIA-NeMo/Gym#5453 · Filter by author phtran8 opened on Nov 2, 2026

• Flaky test: terminal_bench_4 test_verifier_boundary wall-clock bound

  NVIDIA-NeMo/Gym#5434 · Filter by author AdamRajfer opened on Nov 2, 2026

• Flaky test: test_session_of_an_exited_worker_gets_an_explicit_error fails in its precondition

  NVIDIA-NeMo/Gym#5433 · Filter by author AdamRajfer opened on Nov 1, 2026

• simpleqa, omniscience, ragtruth, imo_proofbench_judge: lenient parsing gives reward 1.0 to non-answers

  community-request
  NVIDIA-NeMo/Gym#5430 · Filter by author jimmy-holdout opened on Nov 1, 2026

• Flaky test: mimo_music test_real_timeout_reaps_renderer reads the renderer pid file before it exists

  NVIDIA-NeMo/Gym#5429 · Filter by author AdamRajfer opened on Nov 1, 2026

• keywords:keyword_specific_position counts punctuation as words, so a comma or quote before the target word fails a compliant response

  community-request
  NVIDIA-NeMo/Gym#5414 · Filter by author sven-rankov opened on Oct 31, 2026