fix(model_runner): map Vertex AI headers via http_options and enable vertexai SDK extra

pyproject.toml

@@ -34,7 +34,7 @@ dependencies = [
     "discord-py>=2.6.4",
     "requests>=2.32.5",
     "python-dotenv>=1.2.1",
-    "any-llm-sdk[anthropic]>=1.8.0",
+    "any-llm-sdk[anthropic,vertexai]>=1.8.0",
 ]
 
 [project.urls]
@@ -73,9 +73,7 @@ paths = ["src"]
 
 [tool.mypy]
 files = ["src"]
-exclude = [
-    '^src/bub/skills/.*$',
-]
+exclude = ['^src/bub/skills/.*$']
 disallow_untyped_defs = false
 disallow_any_unimported = false
 no_implicit_optional = true
@@ -115,7 +113,8 @@ select = [
     # mccabe
     "C90",
     # pycodestyle
-    "E", "W",
+    "E",
+    "W",
     # pyflakes
     "F",
     # pygrep-hooks

src/bub/core/model_runner.py

@@ -165,13 +165,23 @@ async def _chat(self, prompt: str) -> _ChatResult:
         system_prompt = self._render_system_prompt()
         try:
             async with asyncio.timeout(self._model_timeout_seconds):
-                output = await self._tape.tape.run_tools_async(
-                    prompt=prompt,
-                    system_prompt=system_prompt,
-                    max_tokens=self._max_tokens,
-                    tools=self._tools,
-                    extra_headers=self.DEFAULT_HEADERS,
-                )
+                provider, _, _ = self._model.partition(":")
+                if provider.casefold() == "vertexai":
+                    output = await self._tape.tape.run_tools_async(
+                        prompt=prompt,
+                        system_prompt=system_prompt,
+                        max_tokens=self._max_tokens,
+                        tools=self._tools,
+                        http_options={"headers": self.DEFAULT_HEADERS},
+                    )
+                else:
+                    output = await self._tape.tape.run_tools_async(
+                        prompt=prompt,
+                        system_prompt=system_prompt,
+                        max_tokens=self._max_tokens,
+                        tools=self._tools,
+                        extra_headers=self.DEFAULT_HEADERS,
+                    )
                 return _ChatResult.from_tool_auto(output)
         except TimeoutError:
             return _ChatResult(

tests/test_model_runner.py

@@ -94,6 +94,7 @@ def all(self) -> list[object]:
 
     outputs: list[ToolAutoResult]
     calls: list[tuple[str, str, int]] = field(default_factory=list)
+    call_kwargs: list[dict[str, object]] = field(default_factory=list)
     query: _Query = field(default_factory=_Query)
 
     async def run_tools_async(
@@ -103,9 +104,10 @@ async def run_tools_async(
         system_prompt: str,
         max_tokens: int,
         tools: list[object],
-        extra_headers: dict[str, str] | None = None,
+        **kwargs: object,
     ) -> ToolAutoResult:
         self.calls.append((prompt, system_prompt, max_tokens))
+        self.call_kwargs.append(kwargs)
         return self.outputs.pop(0)
 
 
@@ -382,3 +384,71 @@ async def test_model_runner_refreshes_skills_from_provider_between_runs() -> Non
     _, second_system_prompt, _ = tape.tape.calls[1]
     assert "<basic_skills>" in second_system_prompt
     assert "friendly-python" in second_system_prompt
+
+
+@pytest.mark.asyncio
+async def test_model_runner_passes_extra_headers_for_openrouter() -> None:
+    tape = FakeTapeService(FakeTapeImpl(outputs=[ToolAutoResult.text_result("assistant-only")]))
+    runner = ModelRunner(
+        tape=tape,  # type: ignore[arg-type]
+        router=SingleStepRouter(),  # type: ignore[arg-type]
+        tool_view=FakeToolView(),  # type: ignore[arg-type]
+        tools=[],
+        list_skills=lambda: [],
+        model="openrouter:test",
+        max_steps=1,
+        max_tokens=512,
+        model_timeout_seconds=90,
+        base_system_prompt="base",
+        get_workspace_system_prompt=lambda: "",
+    )
+
+    await runner.run("hi")
+    kwargs = tape.tape.call_kwargs[0]
+    assert kwargs.get("extra_headers") == ModelRunner.DEFAULT_HEADERS
+
+
+@pytest.mark.asyncio
+async def test_model_runner_maps_headers_for_vertexai() -> None:
+    tape = FakeTapeService(FakeTapeImpl(outputs=[ToolAutoResult.text_result("assistant-only")]))
+    runner = ModelRunner(
+        tape=tape,  # type: ignore[arg-type]
+        router=SingleStepRouter(),  # type: ignore[arg-type]
+        tool_view=FakeToolView(),  # type: ignore[arg-type]
+        tools=[],
+        list_skills=lambda: [],
+        model="vertexai:test",
+        max_steps=1,
+        max_tokens=512,
+        model_timeout_seconds=90,
+        base_system_prompt="base",
+        get_workspace_system_prompt=lambda: "",
+    )
+
+    await runner.run("hi")
+    kwargs = tape.tape.call_kwargs[0]
+    assert "extra_headers" not in kwargs
+    assert kwargs["http_options"] == {"headers": ModelRunner.DEFAULT_HEADERS}
+
+
+@pytest.mark.asyncio
+async def test_model_runner_uses_extra_headers_for_unknown_provider() -> None:
+    tape = FakeTapeService(FakeTapeImpl(outputs=[ToolAutoResult.text_result("assistant-only")]))
+    runner = ModelRunner(
+        tape=tape,  # type: ignore[arg-type]
+        router=SingleStepRouter(),  # type: ignore[arg-type]
+        tool_view=FakeToolView(),  # type: ignore[arg-type]
+        tools=[],
+        list_skills=lambda: [],
+        model="custom:test",
+        max_steps=1,
+        max_tokens=512,
+        model_timeout_seconds=90,
+        base_system_prompt="base",
+        get_workspace_system_prompt=lambda: "",
+    )
+
+    await runner.run("hi")
+    kwargs = tape.tape.call_kwargs[0]
+    assert kwargs.get("extra_headers") == ModelRunner.DEFAULT_HEADERS
+    assert "http_options" not in kwargs