diff --git a/MAKING_A_COGAME.md b/MAKING_A_COGAME.md
index 22589c2db..475e19798 100644
--- a/MAKING_A_COGAME.md
+++ b/MAKING_A_COGAME.md
@@ -187,7 +187,7 @@ HardOreMineMission = OreMineMission(
 ```bash
 cogames tutorial train \
     --mission my_module.OreMineMission \
-    --timesteps 2000000
+    --steps 2000000
 ```
 
 CoGames uses [PufferLib](https://github.com/PufferAI/PufferLib) for training — PPO with parallel vectorized environments. The trained checkpoint loads back for play or eval:
diff --git a/pyproject.toml b/pyproject.toml
index f9b4383db..c9ed337d9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -9,7 +9,7 @@ readme = "README.md"
 requires-python = ">=3.12,<3.13"
 classifiers = ["Programming Language :: Python :: 3", "Programming Language :: Python :: 3.12"]
 dependencies = [
-  "mettagrid==0.26.12",
+  "mettagrid==0.26.13",
   "softmax-cli==0.26.5",
   "packaging>=24.0.0",
   "pydantic>=2.11.5",
@@ -75,8 +75,8 @@ testpaths = ["tests"]
 source = ["cogames"]
 
 [tool.uv.sources]
-mettagrid = { workspace = true }
-softmax-cli = { workspace = true }
+mettagrid = {git = "https://github.com/Metta-AI/mettagrid.git"}
+softmax-cli = {git = "https://github.com/Metta-AI/softmax-cli.git"}
 diplomacog = { git = "https://github.com/Metta-AI/cogame-diplomacog.git" }
 hungercog = { git = "https://github.com/Metta-AI/cogame-hungercog.git" }
 overcogged = { git = "https://github.com/Metta-AI/cogame-overcogged.git" }
diff --git a/src/cogames/docs/AMONGTHEM_POLICY.md b/src/cogames/docs/AMONGTHEM_POLICY.md
index b94f3dc74..62808e537 100644
--- a/src/cogames/docs/AMONGTHEM_POLICY.md
+++ b/src/cogames/docs/AMONGTHEM_POLICY.md
@@ -17,13 +17,14 @@ that will run in the tournament worker, uploading it, and checking whether it sc
 cogames tutorial make-policy --amongthem -o amongthem_policy.py
 ```
 
-Edit `AmongThemPolicy._choose_actions()` in the generated file. The template receives raw BitWorld observations and must
-write integer action indices from the BitWorld trainable action set into `raw_actions`.
+Edit `AmongThemPolicy._choose_actions()` in the generated file. The template receives raw BitWorld observations and
+must return an `np.ndarray` of integer action indices from the BitWorld trainable action set; the parent
+`step_batch()` writes those into `raw_actions`.
 
 ## 2. Log in and pick a season
 
 ```bash
-cogames login
+cogames auth login
 cogames season list
 cogames season show <season>
 ```
diff --git a/tutorials/02_TRAIN.ipynb b/tutorials/02_TRAIN.ipynb
index 58f9d4c86..6d958cb6b 100644
--- a/tutorials/02_TRAIN.ipynb
+++ b/tutorials/02_TRAIN.ipynb
@@ -78,7 +78,7 @@
     "Training on mission: arena\n",
     "...progress logs...\n",
     "Training complete. Checkpoints saved to: ./train_dir\n",
-    "Checkpoint saved to: ./train_dir/<run_id>/model_000001.pt\n",
+    "Final checkpoint: ./train_dir/<run_id>/model_000001.pt\n",
     "```\n",
     "\n",
     "Replace `<run_id>` with your actual run ID from your training output.\n",
diff --git a/tutorials/02_TRAIN.md b/tutorials/02_TRAIN.md
index af67fe118..77eab6476 100644
--- a/tutorials/02_TRAIN.md
+++ b/tutorials/02_TRAIN.md
@@ -40,7 +40,7 @@ Expected terminal output (example):
 Training on mission: arena
 ...progress logs...
 Training complete. Checkpoints saved to: ./train_dir
-Checkpoint saved to: ./train_dir/<run_id>/model_000001.pt
+Final checkpoint: ./train_dir/<run_id>/model_000001.pt
 ```
 
 Replace `<run_id>` with your actual run ID from your training output.
diff --git a/tutorials/02_TRAIN.py b/tutorials/02_TRAIN.py
index 1cb0ea89b..e4d3bd409 100644
--- a/tutorials/02_TRAIN.py
+++ b/tutorials/02_TRAIN.py
@@ -61,7 +61,7 @@
 # Training on mission: arena
 # ...progress logs...
 # Training complete. Checkpoints saved to: ./train_dir
-# Checkpoint saved to: ./train_dir/<run_id>/model_000001.pt
+# Final checkpoint: ./train_dir/<run_id>/model_000001.pt
 # ```
 #
 # Replace `<run_id>` with your actual run ID from your training output.