argus-code-review 0.3.4__tar.gz → 0.4.0__tar.gz

{argus_code_review-0.3.4 → argus_code_review-0.4.0}/CHANGELOG.md

@@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [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 +463,8 @@ 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/v0.4.0...HEAD
+[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-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: argus-code-review
-Version: 0.3.4
+Version: 0.4.0
 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-0.4.0}/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-0.4.0}/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()