Switch to stronger model in CI builds

.github/workflows/nightly_examples.yml

@@ -18,7 +18,7 @@ on:
       model:
         description: "Model to run the examples against"
         required: false
-        default: "gpt-4o-mini"
+        default: "gpt-4.1-mini"
 
 jobs:
   examples:
@@ -43,7 +43,7 @@ jobs:
       # example is a subprocess, so xdist gets real parallelism across them.
       - name: Run every example
         env:
-          EFFECTFUL_LLM_MODEL: ${{ github.event.inputs.model || 'gpt-4o-mini' }}
+          EFFECTFUL_LLM_MODEL: ${{ github.event.inputs.model || 'gpt-4.1-mini' }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
         run: |

.github/workflows/test_llm.yml

@@ -15,7 +15,7 @@ jobs:
       fail-fast: false
       matrix:
         python-version: ["3.13", "3.14"]
-        model: ["gpt-4o-mini"]
+        model: ["gpt-4.1-mini"]
         mypy: ["mypy", "not mypy"]
     name: "test-llm (py ${{ matrix.python-version }}, ${{ matrix.mypy }})"
     steps:
@@ -66,7 +66,7 @@ jobs:
 
       - name: Run the basics examples
         env:
-          EFFECTFUL_LLM_MODEL: gpt-4o-mini
+          EFFECTFUL_LLM_MODEL: gpt-4.1-mini
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
         run: |

docs/source/llm_examples/choreographies/library.py

@@ -113,7 +113,7 @@ async def build_codebase(project_spec, architect, coder, reviewer):
 Calling it is all there is to it -- the skill calls inside need a model behind
 them, which the module launcher supplies::
 
-    python -m effectful.handlers.llm.harness your_choreography.py --model gpt-4o-mini
+    python -m effectful.handlers.llm.harness your_choreography.py
 
 ``multi_agent_choreography.py``, alongside this module, is a complete, runnable
 version: agents with tools, a review-and-fix loop, and resumption.

docs/source/llm_examples/choreographies/multi_agent_choreography.py

@@ -29,7 +29,7 @@
 left off::
 
     python -m effectful.handlers.llm.harness \\
-        docs/source/llm_examples/choreographies/multi_agent_choreography.py --model gpt-4o-mini
+        docs/source/llm_examples/choreographies/multi_agent_choreography.py
 
 Use ``--restart`` to forget the recorded steps and build from scratch, and pass
 ``--persist-db PATH`` to the harness to checkpoint each agent's own

docs/source/llm_examples/optimization/ds1000.py

@@ -31,22 +31,6 @@
 - the data/skill module split that keeps the held-out test honest: each
   problem's ``code_context`` contains its reference solution, and lives in a
   module the solver's system prompt only names
-
-Run with::
-
-    python -m effectful.handlers.llm.harness \\
-        docs/source/llm_examples/optimization/ds1000.py \\
-        --model gpt-5-mini --reasoning-effort low --tool-choice none
-
-Expect a few minutes: ~10 solver calls plus a backward call per training
-problem and one accumulation per parameter. Results are stochastic run to run,
-and the model matters: the run above flipped the held-out problem FAIL -> PASS
-(the direct attempt returned the whole ``OptimizeResult`` where the test wants
-``res.x``; training on ``scipy_716`` taught exactly that contract). A weaker
-model (gpt-4o-mini) visibly *learns* -- its trained attempts fix the
-L-BFGS-B bounds format its direct attempt crashed on -- but tends to keep
-failing the held-out problem on some other of its four simultaneous
-requirements.
 """
 
 import argparse

tests/conftest.py

@@ -12,7 +12,7 @@
 from effectful.handlers.llm.harness.legibility.lexical import _tools_in_scope
 from effectful.ops.syntax import ObjectInterpretation, implements
 
-EFFECTFUL_LLM_MODEL = os.environ.get("EFFECTFUL_LLM_MODEL", "gpt-4o-mini")
+EFFECTFUL_LLM_MODEL = os.environ.get("EFFECTFUL_LLM_MODEL", "gpt-4.1-mini")
 
 _HAS_LLM_API_KEY = litellm.validate_environment(model=EFFECTFUL_LLM_MODEL)[
     "keys_in_environment"