fix(ci): pull the ollama model the tests actually use

.github/workflows/package-ci.yml

@@ -72,9 +72,9 @@ jobs:
       MELLEA_REF: ${{ inputs.mellea_ref }}
     steps:
       - uses: actions/checkout@v4
-      # Ubuntu runners fill up: Ollama tarball + granite4:micro model + uv
-      # envs exhaust the disk (seen as "No space left on device"). Reclaim
-      # space first, same as quality-generic.yml / build-package.yml.
+      # Ubuntu runners fill up: Ollama tarball + the pulled model + uv envs
+      # exhaust the disk (seen as "No space left on device"). Reclaim space
+      # first, same as quality-generic.yml / build-package.yml.
       - name: Free disk space
         uses: ./.github/actions/free-disk-space
       - uses: actions/setup-python@v5
@@ -100,9 +100,16 @@ jobs:
         if: env.SKIP_OLLAMA != 'true'
         run: nohup ollama serve &
 
-      - name: Pull granite4:micro
+      # Must match the model the tests actually request. They call bare
+      # `start_session()`, so that is whatever mellea defaults to — NOT a tag we
+      # get to choose here. mellea v0.8.0 defaults to granite4.2:3b, and pulling the stale
+      # tag meant the real model was downloaded mid-test instead.
+      #
+      # Keep this in step when bumping mellea. Deriving it from the installed
+      # mellea instead of hardcoding is tracked as follow-up work.
+      - name: Pull granite4.2:3b
         if: env.SKIP_OLLAMA != 'true'
-        run: ollama pull granite4:micro
+        run: ollama pull granite4.2:3b
 
       - name: Sync dependencies
         working-directory: ${{ inputs.package }}

_integration_core/pyproject.toml

@@ -56,7 +56,10 @@ ignore_errors = true
 ignore_missing_imports = true
 
 [tool.mellea-contribs.ci]
-skip_ollama = false
+# No Ollama: these tests never construct a session or a backend — they are
+# pure `unittest.mock`. Matches what legacy-ci.yml already did for this
+# package before the new-flat migration moved CI flags in here.
+skip_ollama = true
 timeout_minutes = 30
 python_versions = ["3.12"]
 

agent-utilities/tests/test_double_round_robin.py

@@ -1,6 +1,7 @@
 import logging
 
 from mellea import start_session
+from mellea.backends import ModelOption
 
 from mellea_contribs.agent_utilities.core.double_round_robin import double_round_robin
 
@@ -32,7 +33,12 @@
 
 
 def test_generic_double_round_robin():
-    m = start_session()
+    # Thinking off: each pairwise comparison only needs a single "A"/"B" token,
+    # and a double round robin over N items issues N*(N-1) of them. Granite 4.2
+    # (mellea's default since v0.8.0) reasons unless told not to, which costs
+    # ~190x the output tokens per comparison and made these calls exceed the
+    # backend's 300s timeout on CPU-only CI runners (httpx.ReadTimeout).
+    m = start_session(model_options={ModelOption.THINKING: False})
 
     comparison_prompt = """
         Select which option is more likely to be the primary root-cause

agent-utilities/tests/test_top_k.py

@@ -1,4 +1,5 @@
 from mellea import start_session
+from mellea.backends import ModelOption
 
 from mellea_contribs.agent_utilities.core.top_k import top_k
 
@@ -28,7 +29,11 @@
 
 
 def test_top_k_selection():
-    m = start_session()
+    # Thinking off: top_k's contract is a bare JSON array, so a reasoning trace
+    # is pure overhead. Granite 4.2 (mellea's default since v0.8.0) reasons
+    # unless told not to, which costs ~36x the output tokens here and pushes
+    # each generation past the backend's 300s timeout on CPU-only CI runners.
+    m = start_session(model_options={ModelOption.THINKING: False})
 
     comparison_prompt = """
     Select which items are the most severe issues.

langchain/pyproject.toml

@@ -62,7 +62,10 @@ ignore_errors = true
 ignore_missing_imports = true
 
 [tool.mellea-contribs.ci]
-skip_ollama = false
+# No Ollama: these tests never call `start_session()` — every one of them uses
+# a hand-written double (MockMelleaSession / FakeMelleaSession /
+# ThreadSafeMelleaSession), so no model is ever reached under any selector.
+skip_ollama = true
 timeout_minutes = 30
 python_versions = ["3.12"]