argus-code-review 0.3.4__tar.gz → 0.4.1__tar.gz

{argus_code_review-0.3.4 → argus_code_review-0.4.1}/CHANGELOG.md

@@ -7,6 +7,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.4.1] - 2026-10-28
+
+### Fixed
+
+- Rounds that carry dismissals no longer abort with `temperature: Extra inputs are not permitted` (HTTP 400). The `dismiss_match` call (the small Claude Haiku call that maps `--dismiss` text to prior findings) no longer sends `temperature`: with the current `langchain-anthropic`, structured output rejects it, and a failure there aborts the whole review. First-round reviews were unaffected.
+
+## [0.4.0] - 2026-10-28
+
+### Changed
+
+- Gemini `caching = "auto"` (the packaged default for the bulk reviewers) no longer creates an explicit context cache; only `"on"` does. Explicit caching held just the system prompt and tools (about 1-3k tokens) and Gemini then reported only that slice as cached on every turn, while the growing conversation (the bulk of each request) was billed uncached. With no explicit cache, Gemini's automatic prefix caching applies to the whole repeated prefix. In paired live runs of the same PR the reported cached share of prompt tokens went from about 7% to about 90% (total prompt tokens about equal, verdicts and finding counts equivalent), and Argus's own cost estimate for the review fell from $12.66 to $3.20. That estimate is computed from the reported `cached_content_token_count`, not from a billing export, and Google does not document how automatic cache hits interact with `cached_content`, so treat the savings as unconfirmed. Because the default behavior of `auto` changes, this ships as a minor release (0.4.0). `caching = "on"` keeps the previous explicit-cache behavior.
+
 ## [0.3.4] - 2026-10-23
 
 ### Fixed
@@ -457,7 +469,9 @@ 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/v0.3.4...HEAD
+[Unreleased]: https://github.com/redesignhealth/argus-review/compare/v0.4.1...HEAD
+[0.4.1]: https://github.com/redesignhealth/argus-review/compare/v0.4.0...v0.4.1
+[0.4.0]: https://github.com/redesignhealth/argus-review/compare/v0.3.4...v0.4.0
 [0.3.4]: https://github.com/redesignhealth/argus-review/compare/v0.3.3...v0.3.4
 [0.3.3]: https://github.com/redesignhealth/argus-review/compare/v0.3.2...v0.3.3
 [0.3.2]: https://github.com/redesignhealth/argus-review/compare/v0.3.1...v0.3.2

{argus_code_review-0.3.4 → argus_code_review-0.4.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: argus-code-review
-Version: 0.3.4
+Version: 0.4.1
 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

{argus_code_review-0.3.4 → argus_code_review-0.4.1}/argus/gemini_runner.py

@@ -54,7 +54,7 @@ key so that function's own ``group_name`` fallback still applies
 (this runner has no access to the caller's ``group.name`` -- only to
 ``entry.role``, which is a different, bench-level identifier).
 
-Explicit context caching (``entry.caching in {"auto", "on"}``) is
+Explicit context caching (``entry.caching == "on"``) is
 strictly best-effort: Gemini's explicit-cache API has a minimum
 content-size requirement a short prompt plus this small a toolset might
 not clear, so any exception raised while creating/reusing a cache is
@@ -177,14 +177,22 @@ logger = logging.getLogger(__name__)
 
 _MAX_TURNS_GEMINI = 100  # raised from 45 -- see TECH-6558
 
-# Both "auto" and "on" attempt explicit caching today -- there is no
-# separate heuristic distinguishing them yet (Track 1 defined the
-# `caching` values without one; see argus.bench's `_VALID_CACHING`). The
-# graceful-degradation try/except in `_maybe_activate_cache` is what
-# actually protects a caller who set "auto" against a too-small
-# prompt+toolset that can't clear Gemini's minimum cache-content-size
-# requirement. "off" skips caching entirely.
-_CACHE_ENABLED_VALUES = frozenset({"auto", "on"})
+# Only "on" creates an explicit cache. "auto" (the packaged default) and "off"
+# both send the whole request inline and leave caching to Gemini's own
+# automatic prefix caching, which is enabled by default on current models.
+#
+# Why "auto" is not explicit: an explicit cache holds only what is placed in
+# it at creation time (here the system prompt, tools and tool config, about
+# 1-3k tokens), and a request that references it ("cached_content=") must
+# not also carry those three fields. Every turn's growing conversation
+# (the diff plus all prior tool results) is then sent as ordinary
+# `contents`. In live A/B runs of the same PR,
+# explicit caching reported ~7% cached prompt tokens (exactly the cache size
+# on every turn) while caching="off" reported ~90%, with the same total
+# prompt volume. The graceful-degradation try/except in
+# `_maybe_activate_cache` still protects a caller who sets "on" against a
+# prompt+toolset that cannot clear Gemini's minimum cache size.
+_CACHE_ENABLED_VALUES = frozenset({"on"})
 
 _DEFAULT_READ_LIMIT = 2000  # mirrors argus.review_tools._DEFAULT_READ_LIMIT
 

{argus_code_review-0.3.4 → argus_code_review-0.4.1}/argus/graph.py

@@ -522,8 +522,12 @@ async def _apply_dismissals(
         "or -1 if it doesn't match any finding."
     )
 
+    # No `temperature`: with langchain-anthropic 1.7.7 (anthropic 2.3.0), structured output on this
+    # call sends it in a way the Anthropic API rejects ("temperature: Extra inputs are not
+    # permitted", HTTP 400), and a failure here aborts the whole review. The matching task
+    # is deterministic enough without it.
     llm = _get_llm(
-        f"anthropic:{CLAUDE_MINI}", "dismiss_match", max_tokens=1024, temperature=0
+        f"anthropic:{CLAUDE_MINI}", "dismiss_match", max_tokens=1024
     ).with_structured_output(DismissMatches)
     result = await llm.ainvoke([{"role": "user", "content": prompt}])
 
@@ -1037,7 +1041,8 @@ _TEMPERATURE_UNSUPPORTED_MODELS: frozenset[str] = (
             # set on the assumption Anthropic's temperature deprecation
             # applies across the Sonnet line, not per-checkpoint; re-verify
             # empirically and correct this comment if that assumption turns
-            # out wrong. claude-haiku-4-5 (CLAUDE_MINI) accepts it fine, so
+            # out wrong. claude-haiku-4-5 (CLAUDE_MINI) accepts it fine on plain
+            # requests (not with structured output, see _apply_dismissals), so
             # is deliberately not included here -- see the explicit
             # exclusion below for why "not included" alone isn't enough.
             ALIAS_MAP["claude-frontier"],
@@ -1046,18 +1051,20 @@ _TEMPERATURE_UNSUPPORTED_MODELS: frozenset[str] = (
             CLAUDE_DEFAULT,
         }
     )
-    # Explicit exclusion, not just an absent literal: without this, an
-    # override that happens to land CLAUDE_DEFAULT/CLAUDE_FRONTIER on
-    # CLAUDE_MINI's own resolved value (e.g. `--specialist-model
-    # claude-haiku-4-5`, used directly by this suite's own tests) would pull
-    # that value into the union above via the CLAUDE_DEFAULT/CLAUDE_FRONTIER
-    # terms, and then _apply_dismissals's `_get_llm(f"anthropic:{CLAUDE_MINI}",
-    # "dismiss_match", temperature=0)` call would have ITS temperature silently stripped too
-    # -- a real, silent determinism regression in a call site that has
-    # nothing to do with the override, caught in Argus round 3 review of
-    # this PR. `- {CLAUDE_MINI}` guarantees CLAUDE_MINI's resolved value can
-    # never end up in this set regardless of what any other alias happens to
-    # resolve to.
+    # Explicit exclusion, not just an absent literal: without this, an override that
+    # happens to land CLAUDE_DEFAULT/CLAUDE_FRONTIER on CLAUDE_MINI's own resolved value
+    # (e.g. `--specialist-model claude-haiku-4-5`, used directly by this suite's own
+    # tests) would pull that value into the union above via the
+    # CLAUDE_DEFAULT/CLAUDE_FRONTIER terms, silently stripping `temperature` from any plain
+    # `_get_llm(f"anthropic:{CLAUDE_MINI}", ..., temperature=...)` call that has nothing
+    # to do with the override. `- {CLAUDE_MINI}` guarantees CLAUDE_MINI's resolved value
+    # can never end up in this set.
+    #
+    # `_apply_dismissals`, the former user of such a call, no longer passes `temperature`
+    # at all: structured output on Haiku is rejected with langchain-anthropic 1.7.7
+    # (HTTP 400, "temperature: Extra inputs are not permitted"). `langchain-anthropic` is
+    # unpinned above 1.2.0, so re-check this before relying on `temperature` together
+    # with structured output.
     - {CLAUDE_MINI}
 )
 

{argus_code_review-0.3.4 → argus_code_review-0.4.1}/tests/test_gemini_runner.py

@@ -606,6 +606,36 @@ class TestCacheActivation:
         assert request_config.tools is None
         assert request_config.tool_config is None
 
+    async def test_cache_auto_never_calls_caches_create(self) -> None:
+        # "auto" leaves caching to Gemini's automatic prefix caching: it must
+        # NOT create or reference an explicit cache, because a request that
+        # references one carries only the cache's static prefix and the
+        # growing conversation is then billed uncached.
+        entry = _make_entry(caching="auto")
+        settings = _make_settings()
+
+        response = _make_response(calls=[("finish_review", {"files_explored": []})])
+        fake_client = _make_fake_client([response], cache_name="cachedContents/should-not-be-used")
+
+        with patch("argus.gemini_runner.genai.Client", return_value=fake_client):
+            await run_session_gemini(
+                entry=entry,
+                system_prompt="sys",
+                user_message="msg",
+                settings=settings,
+                repo_root="/tmp/does-not-need-to-exist",
+            )
+
+        fake_client.caches.create.assert_not_called()
+        generate_kwargs = fake_client.aio.models.generate_content.await_args.kwargs
+        assert generate_kwargs["config"].system_instruction == _with_turn_budget_line("sys")
+        assert generate_kwargs["config"].cached_content is None
+        # With no cache to carry them, tools and the forced tool_config must be
+        # sent inline on the request.
+        assert generate_kwargs["config"].tools is not None
+        assert generate_kwargs["config"].tool_config is not None
+        assert generate_kwargs["config"].tool_config.function_calling_config.mode == "ANY"
+
     async def test_cache_off_never_calls_caches_create(self) -> None:
         entry = _make_entry(caching="off")
         settings = _make_settings()

{argus_code_review-0.3.4 → argus_code_review-0.4.1}/tests/test_graph_get_llm_temperature.py

@@ -188,14 +188,15 @@ def test_temperature_guard_does_not_swallow_claude_mini_when_specialist_model_co
     was over-inclusive. Setting ARGUS_SPECIALIST_MODEL=claude-haiku-4-5 makes
     CLAUDE_DEFAULT resolve to the SAME value as CLAUDE_MINI, pulling
     CLAUDE_MINI's own resolved value into _TEMPERATURE_UNSUPPORTED_MODELS via
-    the CLAUDE_DEFAULT term of the union. _match_dismissals
-    (argus/graph.py) calls `_get_llm(f"anthropic:{CLAUDE_MINI}",
-    temperature=0)` for deterministic dismissal-index matching -- that
-    temperature=0 would then be silently stripped, breaking the exact
-    invariant this set's own comment documents ("claude-haiku-4-5
-    (CLAUDE_MINI) accepts it fine, so is deliberately not included here").
-    The explicit `- {CLAUDE_MINI}` exclusion in the set's construction is
-    what prevents this.
+    the CLAUDE_DEFAULT term of the union. A plain
+    `_get_llm(f"anthropic:{CLAUDE_MINI}", ..., temperature=0)` call (this was
+    the dismissal-index matcher, which no longer passes temperature because
+    Haiku structured output rejects it) would then have its temperature
+    silently stripped, breaking the invariant this set's own comment
+    documents ("claude-haiku-4-5 (CLAUDE_MINI) accepts it fine on plain
+    requests, so is deliberately not included here"). The explicit
+    `- {CLAUDE_MINI}` exclusion in the set's construction is what prevents
+    this.
     """
     overridden_set = _temperature_unsupported_models_under_override(CLAUDE_MINI)
     assert CLAUDE_MINI not in overridden_set

{argus_code_review-0.3.4 → argus_code_review-0.4.1}/tests/test_multi_round.py

@@ -7,6 +7,7 @@ round 2 writer messages, and model contracts.
 from __future__ import annotations
 
 import importlib
+import inspect
 import json
 import sys
 from unittest.mock import AsyncMock, MagicMock, patch
@@ -1174,6 +1175,48 @@ class TestApplyDismissals:
         assert matched[0].description == "except Exception"
         assert matched[0].dismissed_by == "user1"
 
+    @pytest.mark.asyncio
+    async def test_dismiss_match_call_sends_no_temperature(self) -> None:
+        """The dismiss_match LLM call must not pass `temperature`.
+
+        With structured output, the current langchain-anthropic request is rejected by the
+        Anthropic API when a temperature is set ("temperature: Extra inputs are not
+        permitted", HTTP 400), and a failure here aborts the whole review.
+        """
+        _reload_graph()
+        from argus.graph import _apply_dismissals
+        from argus.graph import _get_llm as _get_llm_fn
+
+        from pydantic import BaseModel
+
+        class DismissMatch(BaseModel):
+            dismiss_index: int
+            finding_index: int
+
+        class DismissMatches(BaseModel):
+            matches: list[DismissMatch]
+
+        findings = [PriorFinding(severity="BLOCKING", description="issue")]
+        dismissals = [DismissedFinding(description="issue", dismissed_by="u", reason="r")]
+
+        mock_structured = MagicMock()
+        mock_structured.ainvoke = AsyncMock(
+            return_value=DismissMatches(matches=[DismissMatch(dismiss_index=0, finding_index=0)])
+        )
+        mock_llm = MagicMock()
+        mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
+
+        with patch("argus.graph._get_llm", return_value=mock_llm) as mock_get_llm:
+            await _apply_dismissals(findings, dismissals)
+
+        mock_get_llm.assert_called_once()
+        mock_llm.with_structured_output.assert_called_once()
+        bound = inspect.signature(_get_llm_fn).bind(
+            *mock_get_llm.call_args.args, **mock_get_llm.call_args.kwargs
+        )
+        assert bound.arguments["stage"] == "dismiss_match"
+        assert "temperature" not in bound.arguments
+
     @pytest.mark.asyncio
     async def test_out_of_range_indices_ignored(self) -> None:
         """Matches with out-of-range indices are silently skipped."""