diff --git a/README.md b/README.md
index d8b1505c46..bc7885eeb0 100644
--- a/README.md
+++ b/README.md
@@ -1,5 +1,15 @@
 # Nemo Skills
 
+> [!NOTE]
+> **Benchmarks and rollouts are moving to [NeMo-Gym](https://github.com/NVIDIA-NeMo/Gym).** We
+> encourage you to move your rollout workflows over to NeMo-Gym, and to implement new benchmarks
+> there rather than in Nemo-Skills.
+>
+> Nemo-Skills will continue to add support for running NeMo-Gym benchmarks, so you can run them with
+> the same cluster configs, server configurations, and slurm setup you already use. The benchmark
+> implementation itself should live in NeMo-Gym. An early version of this exists today as
+> `ns nemo_gym_rollouts`, but the interface is experimental and will likely change.
+
 Nemo-Skills is a collection of pipelines to improve "skills" of large language models (LLMs). We support everything needed for LLM development, from synthetic data generation, to model training, to evaluation on a wide range of benchmarks. Start developing on a local workstation and move to a large-scale Slurm cluster with just a one-line change.
 
 
diff --git a/core/requirements.txt b/core/requirements.txt
index 0d0fdd2c58..ff5ff67111 100644
--- a/core/requirements.txt
+++ b/core/requirements.txt
@@ -23,9 +23,13 @@ koifeval @ git+https://github.com/bzantium/koifeval.git@7a27889c0cf285e60b1a7b47
 langcodes
 langdetect
 language-data
-litellm[caching]==1.83.14
+# 1.83.14 hard-pins click==8.1.8, which caps the whole tree below click 8.2.
+# 1.84.10 relaxes that to click>=8.0.0,<9.0 (and clears GHSA-4xpc-pv4p-pm3w). From #1507.
+litellm[caching]==1.84.10
 math-verify[antlr4_9_3]
-mcp
+# mcp 2.0 removed the streamablehttp_client alias that nemo_skills/mcp/clients.py imports.
+# Unpin once those call sites move to streamable_http_client and the 2.0 API is verified.
+mcp<2.0
 numpy
 openai
 openpyxl>=3.1.0
@@ -43,4 +47,6 @@ sympy
 torchcodec
 tqdm
 transformers
-wandb
+# Floors click to >=8.2 (wandb 0.26.x only requires click>=8.0.1, so the resolver
+# was settling on click 8.1.8). Taken from #1507.
+wandb>=0.27.1
diff --git a/docs/index.md b/docs/index.md
index fb9edd25ed..ddec0a3f03 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -4,6 +4,17 @@ hide:
   - toc
 ---
 
+!!! note "Benchmarks and rollouts are moving to NeMo-Gym"
+
+    We encourage you to move your rollout workflows over to
+    [NeMo-Gym](https://github.com/NVIDIA-NeMo/Gym), and to implement new benchmarks there rather
+    than in Nemo-Skills.
+
+    Nemo-Skills will continue to add support for running NeMo-Gym benchmarks, so you can run them
+    with the same cluster configs, server configurations, and slurm setup you already use. The
+    benchmark implementation itself should live in NeMo-Gym. An early version of this exists today
+    as `ns nemo_gym_rollouts`, but the interface is experimental and will likely change.
+
 [Nemo-Skills](https://github.com/NVIDIA-NeMo/Skills) is a collection of pipelines to improve "skills" of large language models (LLMs). We support everything needed for LLM development, from synthetic data generation, to model training, to evaluation on a wide range of benchmarks. Start developing on a local workstation and move to a large-scale Slurm cluster with just a one-line change.
 
 Here are some of the features we support:
diff --git a/pyproject.toml b/pyproject.toml
index f6f24aca81..617e40e6e9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -73,7 +73,7 @@ dev = { file = ["requirements/common-tests.txt", "requirements/common-dev.txt"]
 # rely on pip's looser resolver instead.
 override-dependencies = [
     # leptonai (pulled in by nemo_run) hard-pins httpx[http2]==0.27.2, which
-    # conflicts with litellm==1.83.14's httpx==0.28.1. Relax the pin so the
+    # conflicts with litellm's httpx>=0.28.0 floor. Relax the pin so the
     # resolver can pick a single version that satisfies litellm.
     "httpx[http2]>=0.28.1",
     # urllib3<2 has open CVEs (e.g. CVE-2024-37891 / CVE-2025-50182). torchx
diff --git a/requirements/pipeline.txt b/requirements/pipeline.txt
index 8cc278c7ed..00e5a19f06 100644
--- a/requirements/pipeline.txt
+++ b/requirements/pipeline.txt
@@ -1,7 +1,11 @@
 # Pipeline/orchestration dependencies (CLI, cluster management, experiment tracking).
 # These are additional to core/requirements.txt.
 
-click < 8.2.0  # https://github.com/ai-dynamo/dynamo/issues/1039
 nemo-evaluator-launcher<0.1.47
 nemo_run @ git+https://github.com/NVIDIA-NeMo/Run
-typer >= 0.13
+# >=0.16 is click 8.2 compatible (fixes the make_metavar break behind
+# https://github.com/ai-dynamo/dynamo/issues/1039).
+# <0.27 because typer 0.27.0 generates dash-separated option names (--output-dir)
+# instead of the underscore names (--output_dir) that every `ns` command, test and
+# doc example uses. Lift the cap only together with a repo-wide rename.
+typer >= 0.16, < 0.27