argus-code-review 0.2.0__tar.gz → 0.2.2__tar.gz

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/.gitignore

@@ -12,3 +12,4 @@ build/
 .env.local
 .env.*.local
 !.env.example
+.argus/

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/CHANGELOG.md

@@ -7,6 +7,34 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.2.2] - 2026-09-11
+
+### Added
+
+- `docs/BENCH.md`, documenting the reviewer bench's config-driven platform,
+  model, and caching selection for leaf-reviewer roles, including the override
+  chain, valid values, platform credentials, and a worked Gemini configuration.
+  README and `docs/BUILDING_A_REVIEW_LOOP.md` now point to the new guide.
+
+### Fixed
+
+- Updated the `gemini-mini` alias from the superseded `gemini-3-flash-preview`
+  model to the stable `gemini-3.8-flash` release. The approved-model policy
+  table's Google default now matches, and a regression test pins the alias to
+  the approved value.
+
+## [0.2.1] - 2026-09-10
+
+### Added
+
+- `GOOGLE_BASE_URL` proxy support for the Gemini runner (`argus/gemini_runner.py`),
+  mirroring `OPENAI_BASE_URL`: closes a gap where the Gemini platform could only
+  be used with a raw, unproxied `GOOGLE_API_KEY` talking directly to
+  `generativelanguage.googleapis.com`, unlike the OpenAI (`OPENAI_BASE_URL`) and
+  Claude (`ANTHROPIC_BASE_URL`) paths that the `argus-review-loop` skill relies
+  on to route through a short-lived, PR-scoped credential proxy instead of a
+  standing API key.
+
 ## [0.2.0] - 2026-09-10
 
 ### Added
@@ -175,7 +203,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   packaged set.
 - `argus --version`, `argus prompts list`, and `argus prompts export`.
 
-[Unreleased]: https://github.com/redesignhealth/argus-review/compare/v0.2.0...HEAD
+[Unreleased]: https://github.com/redesignhealth/argus-review/compare/v0.2.2...HEAD
+[0.2.2]: https://github.com/redesignhealth/argus-review/compare/v0.2.1...v0.2.2
+[0.2.1]: https://github.com/redesignhealth/argus-review/compare/v0.2.0...v0.2.1
 [0.2.0]: https://github.com/redesignhealth/argus-review/compare/v0.1.5...v0.2.0
 [0.1.5]: https://github.com/redesignhealth/argus-review/compare/v0.1.4...v0.1.5
 [0.1.4]: https://github.com/redesignhealth/argus-review/compare/v0.1.3...v0.1.4

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: argus-code-review
-Version: 0.2.0
+Version: 0.2.2
 Summary: Self-orchestrated PR review agent using LangGraph + Claude Agent SDK
 Project-URL: Repository, https://github.com/redesignhealth/argus-review
 Project-URL: Issues, https://github.com/redesignhealth/argus-review/issues
@@ -437,6 +437,8 @@ see the [write-up](https://www.redesignhealth.com/content/agentic-code-review-ha
   contract for self-hosters.
 - [`docs/BUILDING_A_REVIEW_LOOP.md`](docs/BUILDING_A_REVIEW_LOOP.md) — how to
   wrap Argus in a review → fix → re-review automation loop gated on CI.
+- [`docs/BENCH.md`](docs/BENCH.md) — configure the static reviewer bench to
+  choose platforms and models for leaf reviewers.
 - [`docs/PRECHECKS.md`](docs/PRECHECKS.md) — the deterministic, non-LLM gate
   that runs before the LLM pipeline: CI-status routing signal and custom
   semgrep rules mined from Argus's own recurring findings.

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/README.md

@@ -385,6 +385,8 @@ see the [write-up](https://www.redesignhealth.com/content/agentic-code-review-ha
   contract for self-hosters.
 - [`docs/BUILDING_A_REVIEW_LOOP.md`](docs/BUILDING_A_REVIEW_LOOP.md) — how to
   wrap Argus in a review → fix → re-review automation loop gated on CI.
+- [`docs/BENCH.md`](docs/BENCH.md) — configure the static reviewer bench to
+  choose platforms and models for leaf reviewers.
 - [`docs/PRECHECKS.md`](docs/PRECHECKS.md) — the deterministic, non-LLM gate
   that runs before the LLM pipeline: CI-status routing signal and custom
   semgrep rules mined from Argus's own recurring findings.

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/argus/config.py

@@ -135,6 +135,10 @@ class Settings(BaseSettings):
             ``platform = "gemini"``. Only required if you actually
             opt a role into that platform (see ``argus.bench``'s
             override chain) -- the packaged default bench never does.
+        GOOGLE_BASE_URL: Optional override for the Gemini API endpoint (e.g. for
+            proxying). Mirrors ``OPENAI_BASE_URL`` above. Note that GOOGLE_API_KEY
+            will be forwarded to whatever endpoint is configured here; point only
+            at trusted proxy hosts.
         ARGUS_BENCH_FILE: Highest-priority bench-config override file. See
             ``argus.bench`` for the full override search chain.
         ARGUS_NO_BENCH_OVERRIDES: Set truthy to ignore every bench override
@@ -154,6 +158,7 @@ class Settings(BaseSettings):
     OPENAI_API_KEY: str
     OPENAI_BASE_URL: str | None = None
     GOOGLE_API_KEY: str | None = None
+    GOOGLE_BASE_URL: str | None = None
 
     ARGUS_DB_URL: str | None = None
     SUPABASE_DB_URL: str | None = None

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/argus/gemini_runner.py

@@ -607,7 +607,21 @@ async def _run_turns(
     # clause catches and translates into a clean `failure_reason="timeout"` result
     # -- rather than racing the outer deadline so closely that the raw
     # httpx exception could escape uncaught instead.
-    http_options = types.HttpOptions(timeout=int(timeout_s * _HTTP_TIMEOUT_FRACTION * 1000))
+    #
+    # `base_url` is only set when truthy; `None` safely defaults to the
+    # standard Gemini endpoint (`generativelanguage.googleapis.com`) --
+    # mirrors `OPENAI_BASE_URL`'s exact same optional-proxy pattern in
+    # `argus.openai_runner`/`argus.openai_client`. This same `http_options`
+    # instance is reused as-is by `_maybe_activate_cache`'s own
+    # short-lived `genai.Client` (see that function's docstring for why it
+    # needs a separate client instance), so threading `base_url` through
+    # here covers both `genai.Client(...)` construction sites in this
+    # module without a second read of `settings.GOOGLE_BASE_URL`.
+    base_url = getattr(settings, "GOOGLE_BASE_URL", None)
+    http_options = types.HttpOptions(
+        timeout=int(timeout_s * _HTTP_TIMEOUT_FRACTION * 1000),
+        base_url=base_url if isinstance(base_url, str) and base_url else None,
+    )
     client = genai.Client(api_key=api_key, http_options=http_options)
     logger.info(
         "Gemini session started: %s model=%s caching=%s timeout=%ss",

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/argus/llm/models.py

@@ -105,7 +105,12 @@ ALIAS_MAP: Final[dict[str, str]] = {
     # TODO: upgrade to stable gemini-3.1-pro (non-preview) when GA;
     # re-eval by 2026-11-01.
     "gemini-frontier": "gemini-3.1-pro-preview",
-    "gemini-mini": "gemini-3-flash-preview",
+    # gemini-3-flash-preview has been superseded by gemini-3.8-flash, the
+    # stable (non-preview) release -- same reasoning as gemini-frontier's
+    # move off a deprecated preview above: prefer the stable successor over
+    # a preview model once one exists. test_llm_models.py pins this exact
+    # value as a regression guard.
+    "gemini-mini": "gemini-3.8-flash",
 }
 
 # Short-lived pins for evals / preview models. Anything in here is a known

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/argus/prompts/pr-review-specialist-llm-patterns.md

@@ -9,7 +9,7 @@ Only approved model families. Check model strings in code against this table:
 |----------|----------|---------|
 | Anthropic | claude-opus-4, claude-sonnet-4, claude-haiku-4 (use `CLAUDE_MINI`), claude-fable-5 (use `CLAUDE_FRONTIER`), claude-opus-5 (use `CLAUDE_OPUS`), claude-sonnet-4-6 (use `CLAUDE_DEFAULT`) families; claude-sonnet-5 (the previous `CLAUDE_DEFAULT` pin) remains valid as an explicit `ARGUS_SPECIALIST_MODEL`/`--specialist-model` override value, not as a new hardcoded default | claude-sonnet-4-6 |
 | OpenAI | gpt-5.4 family (gpt-5.4, gpt-5.4-mini) and gpt-5.6 family | gpt-5.6-luna |
-| Google | gemini-3 family | gemini-3-flash |
+| Google | gemini-3 family | gemini-3.8-flash |
 
 Flag any use of: o3, o1, gpt-4 family, gpt-5-mini, claude-3/3.5 family, gemini-1.5/2.5 family.
 

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/tests/test_config.py

@@ -149,6 +149,7 @@ def test_optional_fields_default_to_none(monkeypatch: pytest.MonkeyPatch) -> Non
         "ARGUS_CONTEXT7_LIBRARY_ID",
         "ARGUS_CONTEXT7_BASE_URL",
         "GOOGLE_API_KEY",
+        "GOOGLE_BASE_URL",
         "ARGUS_BENCH_FILE",
         "ARGUS_GEMINI_CACHE_DIR",
     ):
@@ -160,6 +161,7 @@ def test_optional_fields_default_to_none(monkeypatch: pytest.MonkeyPatch) -> Non
     assert settings.ARGUS_CONTEXT7_BASE_URL is None
     assert settings.ARGUS_PROMPTS_DIR is None
     assert settings.GOOGLE_API_KEY is None
+    assert settings.GOOGLE_BASE_URL is None
     assert settings.ARGUS_BENCH_FILE is None
     assert settings.ARGUS_GEMINI_CACHE_DIR is None
 
@@ -252,3 +254,17 @@ def test_openai_base_url_override(monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setenv("OPENAI_BASE_URL", "https://proxy.example.com/v1")
     settings = get_settings()
     assert settings.OPENAI_BASE_URL == "https://proxy.example.com/v1"
+
+
+def test_google_base_url_defaults_to_none(monkeypatch: pytest.MonkeyPatch) -> None:
+    _set_required(monkeypatch)
+    monkeypatch.delenv("GOOGLE_BASE_URL", raising=False)
+    settings = get_settings()
+    assert settings.GOOGLE_BASE_URL is None
+
+
+def test_google_base_url_override(monkeypatch: pytest.MonkeyPatch) -> None:
+    _set_required(monkeypatch)
+    monkeypatch.setenv("GOOGLE_BASE_URL", "https://proxy.example.com/v1beta")
+    settings = get_settings()
+    assert settings.GOOGLE_BASE_URL == "https://proxy.example.com/v1beta"

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/tests/test_gemini_runner.py

@@ -1395,3 +1395,122 @@ class TestGeminiRedactionHooks:
         inputs = {"label": "test"}
         redacted = _redact_gemini_inputs(inputs)
         assert redacted == {"label": "test"}
+
+
+# ---------------------------------------------------------------------------
+# GOOGLE_BASE_URL threading (mirrors OPENAI_BASE_URL in
+# tests/test_openai_runner.py / tests/test_openai_client.py): a proxy base
+# URL configured via settings must reach the underlying `genai.Client`'s
+# `http_options.base_url`, at every `genai.Client(...)` construction site in
+# this module -- the session's own client (`_run_turns`) AND the dedicated,
+# short-lived client `_maybe_activate_cache`'s `_create_fn` builds for
+# `caches.create()`.
+# ---------------------------------------------------------------------------
+
+
+class TestGoogleBaseUrlOverride:
+    async def test_settings_google_base_url_reaches_session_client_http_options(self) -> None:
+        """GOOGLE_BASE_URL on settings is threaded through to the session's
+        own `genai.Client`'s `http_options.base_url`."""
+        entry = _make_entry(caching="off")
+        settings = _make_settings()
+        settings.GOOGLE_BASE_URL = "https://proxy.example.com/v1beta"
+        response = _make_response(calls=[("finish_review", {"files_explored": []})])
+        fake_client = _make_fake_client([response])
+
+        with patch("argus.gemini_runner.genai.Client", return_value=fake_client) as mock_ctor:
+            result = await run_session_gemini(
+                entry=entry,
+                system_prompt="sys",
+                user_message="msg",
+                settings=settings,
+                repo_root="/tmp/does-not-need-to-exist",
+            )
+
+        assert result.failure_reason is None
+        mock_ctor.assert_called_once()
+        http_options = mock_ctor.call_args.kwargs["http_options"]
+        assert http_options.base_url == "https://proxy.example.com/v1beta"
+
+    async def test_no_google_base_url_defaults_to_sdk_default(self) -> None:
+        """Without an override, `http_options.base_url` stays `None` -- the
+        SDK's own default (`generativelanguage.googleapis.com`) applies,
+        exactly like the OPENAI_BASE_URL/AsyncOpenAI pattern this mirrors."""
+        entry = _make_entry(caching="off")
+        settings = _make_settings()  # GOOGLE_BASE_URL never set on this MagicMock
+        response = _make_response(calls=[("finish_review", {"files_explored": []})])
+        fake_client = _make_fake_client([response])
+
+        with patch("argus.gemini_runner.genai.Client", return_value=fake_client) as mock_ctor:
+            result = await run_session_gemini(
+                entry=entry,
+                system_prompt="sys",
+                user_message="msg",
+                settings=settings,
+                repo_root="/tmp/does-not-need-to-exist",
+            )
+
+        assert result.failure_reason is None
+        http_options = mock_ctor.call_args.kwargs["http_options"]
+        assert http_options.base_url is None
+
+    async def test_google_base_url_composes_with_existing_timeout(self) -> None:
+        """A `base_url` override must compose with the existing `timeout=`
+        `HttpOptions` field -- both set together on the same instance --
+        rather than one replacing the other."""
+        from argus.gemini_runner import _HTTP_TIMEOUT_FRACTION
+
+        entry = _make_entry(caching="off")
+        settings = _make_settings(session_timeout=300)
+        settings.GOOGLE_BASE_URL = "https://proxy.example.com/v1beta"
+        response = _make_response(calls=[("finish_review", {"files_explored": []})])
+        fake_client = _make_fake_client([response])
+
+        with patch("argus.gemini_runner.genai.Client", return_value=fake_client) as mock_ctor:
+            result = await run_session_gemini(
+                entry=entry,
+                system_prompt="sys",
+                user_message="msg",
+                settings=settings,
+                repo_root="/tmp/does-not-need-to-exist",
+                timeout_s=10.0,
+            )
+
+        assert result.failure_reason is None
+        http_options = mock_ctor.call_args.kwargs["http_options"]
+        assert http_options.base_url == "https://proxy.example.com/v1beta"
+        assert http_options.timeout == int(10.0 * _HTTP_TIMEOUT_FRACTION * 1000)
+
+    async def test_google_base_url_also_reaches_the_cache_creation_client(self) -> None:
+        """`_maybe_activate_cache`'s `_create_fn` builds its own, separate
+        `genai.Client` for `caches.create()` (see that function's docstring)
+        -- it must carry the same `base_url` override as the session's own
+        client, not silently fall back to the real API host."""
+        entry = _make_entry(caching="on")
+        settings = _make_settings()
+        settings.GOOGLE_BASE_URL = "https://proxy.example.com/v1beta"
+        response = _make_response(calls=[("finish_review", {"files_explored": []})])
+        fake_client = _make_fake_client([response], cache_name="cachedContents/abc123")
+
+        constructed_http_options: list[types.HttpOptions] = []
+
+        def _new_client(*args: Any, **kwargs: Any) -> MagicMock:
+            constructed_http_options.append(kwargs["http_options"])
+            return fake_client
+
+        with patch("argus.gemini_runner.genai.Client", side_effect=_new_client):
+            result = await run_session_gemini(
+                entry=entry,
+                system_prompt="sys",
+                user_message="msg",
+                settings=settings,
+                repo_root="/tmp/does-not-need-to-exist",
+            )
+
+        assert result.failure_reason is None
+        # Two `genai.Client` constructions: the session's own, and the
+        # dedicated cache-creation client -- both must carry the override.
+        assert len(constructed_http_options) == 2
+        assert all(
+            ho.base_url == "https://proxy.example.com/v1beta" for ho in constructed_http_options
+        )

{argus_code_review-0.2.0 → argus_code_review-0.2.2}/tests/test_llm_models.py

@@ -14,6 +14,7 @@ from argus.llm.models import (
     CLAUDE_DEFAULT,
     EXPERIMENTAL_MODELS,
     GEMINI_FRONTIER,
+    GEMINI_MINI,
     GPT_FRONTIER,
     GPT_MINI,
     estimate_cost_usd,
@@ -163,6 +164,20 @@ class TestEstimateCostUsd:
         cost = estimate_cost_usd(GPT_MINI, input_tokens=1000, output_tokens=500)
         assert cost > 0.0
 
+    def test_gemini_mini_pinned_to_approved_model(self) -> None:
+        """Regression guard for gemini-mini, alongside gpt-frontier and
+        gpt-mini's guards above: gemini-3-flash-preview has been superseded by
+        gemini-3.8-flash, the stable (non-preview) release on the approved model
+        list (see pr-review-specialist-llm-patterns.md). Pin the alias to that
+        approved value so an accidental regression to a preview model or
+        unapproved string is caught here instead of at review time.
+
+        Also exercises estimate_cost_usd(GEMINI_MINI, ...) directly."""
+        assert ALIAS_MAP["gemini-mini"] == "gemini-3.8-flash"
+        assert GEMINI_MINI == "gemini-3.8-flash"
+        cost = estimate_cost_usd(GEMINI_MINI, input_tokens=1000, output_tokens=500)
+        assert cost > 0.0
+
 
 class TestResolveOverrides:
     def test_resolve_claude_default_honors_env_override(