fix: address CodeRabbit review findings on PR #188

- Add openai_codex to the max_output_tokens kwarg-override map: it
  builds the same OpenAIResponses provider as openai_responses, so the
  compaction cap was silently no-op'ing (falling back to max_tokens)
  for ChatGPT/Codex-backed sessions.
- Add direct capped_chat_provider coverage for all provider-type ->
  kwarg mappings, including the openai_codex case above.
- Extract the duplicated Runtime(...) rebuild in
  test_pythinkersoul_stuck_loop.py into a shared helper.

Author: elkaix

src/pythinker_code/llm.py

@@ -66,6 +66,7 @@ def model_name(self) -> str:
 # providers, plus the first-party `pythinker` provider, all use `max_tokens`).
 _MAX_OUTPUT_TOKENS_KWARG_OVERRIDES: dict[str, str] = {
     "openai_responses": "max_output_tokens",
+    "openai_codex": "max_output_tokens",
     "google_genai": "max_output_tokens",
     "gemini": "max_output_tokens",
     "vertexai": "max_output_tokens",

tests/core/test_compaction_overflow.py

@@ -18,7 +18,7 @@
 from pythinker_core.chat_provider import APIStatusError
 from pythinker_core.message import Message
 
-from pythinker_code.llm import LLM
+from pythinker_code.llm import LLM, capped_chat_provider
 from pythinker_code.soul.compaction import SimpleCompaction
 from pythinker_code.wire.types import TextPart
 
@@ -46,6 +46,42 @@ def _fake_llm() -> LLM:
     return cast(LLM, SimpleNamespace(chat_provider=_FakeChatProvider(), provider_config=None))
 
 
+def _fake_llm_with_provider_type(provider_type: str) -> LLM:
+    return cast(
+        LLM,
+        SimpleNamespace(
+            chat_provider=_FakeChatProvider(),
+            provider_config=SimpleNamespace(type=provider_type),
+        ),
+    )
+
+
+@pytest.mark.parametrize(
+    ("provider_type", "expected_kwarg"),
+    [
+        ("openai_legacy", "max_tokens"),
+        ("anthropic", "max_tokens"),
+        ("pythinker", "max_tokens"),
+        ("openai_responses", "max_output_tokens"),
+        # ChatGPT/Codex sessions build the same OpenAIResponses provider as
+        # "openai_responses" (see create_llm's "openai_codex" case), so they
+        # take the same max_output_tokens kwarg, not the max_tokens default.
+        ("openai_codex", "max_output_tokens"),
+        ("google_genai", "max_output_tokens"),
+        ("gemini", "max_output_tokens"),
+        ("vertexai", "max_output_tokens"),
+    ],
+)
+def test_capped_chat_provider_picks_kwarg_by_provider_type(
+    provider_type: str, expected_kwarg: str
+) -> None:
+    llm = _fake_llm_with_provider_type(provider_type)
+
+    capped_chat_provider(llm, 4000)
+
+    assert cast(_FakeChatProvider, llm.chat_provider).generation_kwargs == {expected_kwarg: 4000}
+
+
 def _overflow_error() -> APIStatusError:
     return APIStatusError(400, "This model's maximum context length is exceeded")
 

tests/core/test_pythinkersoul_stuck_loop.py

@@ -128,11 +128,8 @@ async def __call__(self, params: _NoParams) -> ToolReturnValue:
         return ToolOk(output="ok", message="ok")
 
 
-def _make_soul(
-    runtime: Runtime, provider: _ScriptedToolCallProvider, tmp_path: Path
-) -> tuple[Context, PythinkerSoul]:
-    llm = LLM(chat_provider=provider, max_context_size=100_000, capabilities=set())
-    runtime = Runtime(
+def _rebuild_runtime_with_llm(runtime: Runtime, llm: LLM) -> Runtime:
+    return Runtime(
         config=runtime.config,
         llm=llm,
         session=runtime.session,
@@ -149,6 +146,13 @@ def _make_soul(
         skills_dirs=runtime.skills_dirs,
         role=runtime.role,
     )
+
+
+def _make_soul(
+    runtime: Runtime, provider: _ScriptedToolCallProvider, tmp_path: Path
+) -> tuple[Context, PythinkerSoul]:
+    llm = LLM(chat_provider=provider, max_context_size=100_000, capabilities=set())
+    runtime = _rebuild_runtime_with_llm(runtime, llm)
     agent = Agent(
         name="Stuck Test Agent",
         system_prompt="Stuck test prompt.",
@@ -166,23 +170,7 @@ def _make_soul_with_pythinker_toolset(
     """Like `_make_soul`, but with a real `PythinkerToolset` — required to exercise the
     identical-call repeat backstop, which is tracked on `PythinkerToolset` specifically."""
     llm = LLM(chat_provider=provider, max_context_size=100_000, capabilities=set())
-    runtime = Runtime(
-        config=runtime.config,
-        llm=llm,
-        session=runtime.session,
-        builtin_args=runtime.builtin_args,
-        denwa_renji=runtime.denwa_renji,
-        approval=runtime.approval,
-        labor_market=runtime.labor_market,
-        environment=runtime.environment,
-        notifications=runtime.notifications,
-        background_tasks=runtime.background_tasks,
-        skills=runtime.skills,
-        oauth=runtime.oauth,
-        additional_dirs=runtime.additional_dirs,
-        skills_dirs=runtime.skills_dirs,
-        role=runtime.role,
-    )
+    runtime = _rebuild_runtime_with_llm(runtime, llm)
     toolset = PythinkerToolset()
     toolset.add(_BoomTool())
     toolset.add(_OkTool())