argus-code-review 0.4.0__tar.gz → 0.4.2__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (162) hide show
  1. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/CHANGELOG.md +15 -1
  2. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/PKG-INFO +1 -1
  3. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/gemini_runner.py +169 -20
  4. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/graph.py +21 -14
  5. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_gemini_runner.py +120 -7
  6. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_get_llm_temperature.py +9 -8
  7. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_multi_round.py +43 -0
  8. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/.gitignore +0 -0
  9. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/LICENSE +0 -0
  10. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/README.md +0 -0
  11. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/__init__.py +0 -0
  12. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/bench.py +0 -0
  13. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/bench_default.toml +0 -0
  14. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/cli.py +0 -0
  15. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/config.py +0 -0
  16. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/coverage.py +0 -0
  17. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/dotenv_utils.py +0 -0
  18. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/gemini_cache.py +0 -0
  19. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/github_client.py +0 -0
  20. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/helpers.py +0 -0
  21. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/llm/models.py +0 -0
  22. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/llm/output_models.py +0 -0
  23. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/llm/pricing.py +0 -0
  24. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/llm/usage.py +0 -0
  25. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/models.py +0 -0
  26. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/openai_client.py +0 -0
  27. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/openai_runner.py +0 -0
  28. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/pipeline_models.py +0 -0
  29. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/__init__.py +0 -0
  30. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/actions_scanner.py +0 -0
  31. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/engine.py +0 -0
  32. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/eslint_bundle/.gitignore +0 -0
  33. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/eslint_bundle/README.md +0 -0
  34. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/eslint_bundle/eslint.config.js +0 -0
  35. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/eslint_bundle/package-lock.json +0 -0
  36. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/eslint_bundle/package.json +0 -0
  37. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/js_scanner.py +0 -0
  38. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/migration_scanner.py +0 -0
  39. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/rules/README.md +0 -0
  40. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/sarif.py +0 -0
  41. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/scanner_utils.py +0 -0
  42. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/secrets_scanner.py +0 -0
  43. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/shadow.py +0 -0
  44. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/terraform_scanner.py +0 -0
  45. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/precheck/workflow_lint_scanner.py +0 -0
  46. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/__init__.py +0 -0
  47. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-blocking-validator.md +0 -0
  48. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-coverage-check.md +0 -0
  49. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-cross-cutting.md +0 -0
  50. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-feedback-verifier.md +0 -0
  51. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-lite.md +0 -0
  52. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-planner.md +0 -0
  53. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-preflight-router.md +0 -0
  54. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-prior-art.md +0 -0
  55. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-scope.md +0 -0
  56. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-deployment.md +0 -0
  57. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-frontend.md +0 -0
  58. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-infra.md +0 -0
  59. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-llm-patterns.md +0 -0
  60. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-observability.md +0 -0
  61. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-orchestration.md +0 -0
  62. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-security.md +0 -0
  63. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-slackbot.md +0 -0
  64. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-specialist-sql.md +0 -0
  65. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-subagent.md +0 -0
  66. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-tests-and-docs.md +0 -0
  67. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts/pr-review-writer.md +0 -0
  68. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/prompts_runtime.py +0 -0
  69. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/repo_provision.py +0 -0
  70. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/review_tools.py +0 -0
  71. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/runners.py +0 -0
  72. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/scope.py +0 -0
  73. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/__init__.py +0 -0
  74. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/http.py +0 -0
  75. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/models.py +0 -0
  76. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/precheck.py +0 -0
  77. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/resolver.py +0 -0
  78. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/session.py +0 -0
  79. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/sql.py +0 -0
  80. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/argus/storage/sqlite.py +0 -0
  81. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/pyproject.toml +0 -0
  82. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/008_add_code_reviews.sql +0 -0
  83. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/009_add_reviewer_version.sql +0 -0
  84. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/010_add_review_patterns.sql +0 -0
  85. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/011_add_review_progress_columns.sql +0 -0
  86. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/015_create_agent_runs.sql +0 -0
  87. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/016_add_agent_runs_failure_reason.sql +0 -0
  88. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/017_add_precheck_rules.sql +0 -0
  89. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/schema/018_widen_agent_runs_failure_reason.sql +0 -0
  90. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/__init__.py +0 -0
  91. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/conftest.py +0 -0
  92. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/golden/review_response.schema.json +0 -0
  93. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/__init__.py +0 -0
  94. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/test_backend_contract.py +0 -0
  95. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/test_http.py +0 -0
  96. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/test_models.py +0 -0
  97. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/test_resolver.py +0 -0
  98. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/test_session.py +0 -0
  99. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/test_sql.py +0 -0
  100. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/storage/test_sqlite_backend.py +0 -0
  101. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_actions_scanner.py +0 -0
  102. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_argus_review_local.py +0 -0
  103. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_bench.py +0 -0
  104. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_bench_config_guard.py +0 -0
  105. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_catchup_gate.py +0 -0
  106. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_cli_args.py +0 -0
  107. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_cli_output_contract.py +0 -0
  108. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_cli_post_review.py +0 -0
  109. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_cli_preflight.py +0 -0
  110. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_cli_prompts.py +0 -0
  111. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_config.py +0 -0
  112. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_conftest.py +0 -0
  113. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_dotenv_utils.py +0 -0
  114. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_failure_diagnostics.py +0 -0
  115. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_gemini_cache.py +0 -0
  116. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_gemini_http_retry.py +0 -0
  117. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_github_client_checks_signal.py +0 -0
  118. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_github_client_write.py +0 -0
  119. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_fetch_diff.py +0 -0
  120. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_http_guards.py +0 -0
  121. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_precheck.py +0 -0
  122. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_preflight_image_bump.py +0 -0
  123. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_progress.py +0 -0
  124. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_storage_resolution.py +0 -0
  125. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_graph_timeout_surfacing.py +0 -0
  126. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_js_scanner.py +0 -0
  127. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_llm_models.py +0 -0
  128. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_llm_models_override.py +0 -0
  129. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_llm_pricing.py +0 -0
  130. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_migration_scanner.py +0 -0
  131. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_models.py +0 -0
  132. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_openai_client.py +0 -0
  133. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_openai_runner.py +0 -0
  134. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_output_models.py +0 -0
  135. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_packaging.py +0 -0
  136. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_plan_review.py +0 -0
  137. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_precheck_engine.py +0 -0
  138. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_precheck_engine_integration.py +0 -0
  139. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_precheck_sarif.py +0 -0
  140. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_precheck_shadow.py +0 -0
  141. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_precheck_shadow_integration.py +0 -0
  142. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_prompts.py +0 -0
  143. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_repo_provision.py +0 -0
  144. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_review_patterns_integration.py +0 -0
  145. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_review_tools.py +0 -0
  146. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_runners_context7.py +0 -0
  147. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_runners_context_usage.py +0 -0
  148. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_runners_helpers.py +0 -0
  149. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_runners_new.py +0 -0
  150. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_runners_turn_budget.py +0 -0
  151. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_scanner_utils.py +0 -0
  152. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_scope.py +0 -0
  153. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_scope_gate.py +0 -0
  154. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_sdk_bundled_cli_version.py +0 -0
  155. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_secrets_scanner.py +0 -0
  156. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_specialist_validation.py +0 -0
  157. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_stage_costs.py +0 -0
  158. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_storage_precheck.py +0 -0
  159. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_structured_output_methods.py +0 -0
  160. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_terraform_scanner.py +0 -0
  161. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_terraform_scanner_integration.py +0 -0
  162. {argus_code_review-0.4.0 → argus_code_review-0.4.2}/tests/test_workflow_lint_scanner.py +0 -0
@@ -7,6 +7,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
7
7
 
8
8
  ## [Unreleased]
9
9
 
10
+ ## [0.4.2] - 2026-10-29
11
+
12
+ ### Changed
13
+
14
+ - A Gemini `400 INVALID_ARGUMENT` no longer ends the reviewer session on the first occurrence. Argus now logs a key-free, text-free summary of the failing request (message count, approximate size, part kinds and sizes for the last four turns, and anomalies such as empty parts or function calls with no matching result) plus any error `details` the API returned, then retries the same request once after a short pause (`ARGUS_GEMINI_RETRY_INVALID_ARGUMENT_MIN_S` / `_MAX_S`, default 2-5 s). The log records whether the retry recovered, which separates a transient fault from a malformed request. Other 4xx errors still fail immediately.
15
+
16
+ ## [0.4.1] - 2026-10-28
17
+
18
+ ### Fixed
19
+
20
+ - 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.
21
+
10
22
  ## [0.4.0] - 2026-10-28
11
23
 
12
24
  ### Changed
@@ -463,7 +475,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
463
475
  packaged set.
464
476
  - `argus --version`, `argus prompts list`, and `argus prompts export`.
465
477
 
466
- [Unreleased]: https://github.com/redesignhealth/argus-review/compare/v0.4.0...HEAD
478
+ [Unreleased]: https://github.com/redesignhealth/argus-review/compare/v0.4.1...HEAD
479
+ [0.4.2]: https://github.com/redesignhealth/argus-review/compare/v0.4.1...v0.4.2
480
+ [0.4.1]: https://github.com/redesignhealth/argus-review/compare/v0.4.0...v0.4.1
467
481
  [0.4.0]: https://github.com/redesignhealth/argus-review/compare/v0.3.4...v0.4.0
468
482
  [0.3.4]: https://github.com/redesignhealth/argus-review/compare/v0.3.3...v0.3.4
469
483
  [0.3.3]: https://github.com/redesignhealth/argus-review/compare/v0.3.2...v0.3.3
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: argus-code-review
3
- Version: 0.4.0
3
+ Version: 0.4.2
4
4
  Summary: Self-orchestrated PR review agent using LangGraph + Claude Agent SDK
5
5
  Project-URL: Repository, https://github.com/redesignhealth/argus-review
6
6
  Project-URL: Issues, https://github.com/redesignhealth/argus-review/issues
@@ -426,6 +426,113 @@ def _failure_detail(exc: BaseException) -> str:
426
426
  return detail
427
427
 
428
428
 
429
+ _INVALID_ARGUMENT_RETRY_MIN_S = 2.0
430
+ _INVALID_ARGUMENT_RETRY_MAX_S = 5.0
431
+ _SHAPE_TAIL_CONTENTS = 4
432
+
433
+
434
+ def _is_invalid_argument(exc: BaseException) -> bool:
435
+ """True for a Gemini ``400 INVALID_ARGUMENT`` API answer."""
436
+ code, status = _error_code_status(exc)
437
+ return code == 400 and (status or "").upper() == "INVALID_ARGUMENT"
438
+
439
+
440
+ def _part_shape(part: Any) -> str:
441
+ """Kind and size of one content part -- never its text."""
442
+ call = getattr(part, "function_call", None)
443
+ if call is not None:
444
+ size = len(json.dumps(call.args or {}, default=str))
445
+ return f"call:{call.name}({size})"
446
+ resp = getattr(part, "function_response", None)
447
+ if resp is not None:
448
+ # An empty result reports size 0 so _describe_request can flag it. The
449
+ # runner wraps every tool output as {"result": output}, so unwrap it.
450
+ payload = resp.response
451
+ if isinstance(payload, dict) and set(payload) == {"result"}:
452
+ payload = payload["result"]
453
+ size = len(json.dumps(payload, default=str)) if payload else 0
454
+ return f"result:{resp.name}({size})"
455
+ text = getattr(part, "text", None)
456
+ if text is not None:
457
+ tag = "thought" if getattr(part, "thought", False) else "text"
458
+ return f"{tag}({len(text)})"
459
+ return type(part).__name__
460
+
461
+
462
+ def _describe_request(contents: list[Any], config: Any = None) -> str:
463
+ """Key-free, text-free summary of a request for diagnosing a 400.
464
+
465
+ Gemini's INVALID_ARGUMENT body carries no field name, so this records the
466
+ structure that could make a request malformed: roles, part kinds and sizes
467
+ for the last few turns, plus anomalies (empty parts, unanswered function
468
+ calls, consecutive same-role turns).
469
+ """
470
+ total = 0
471
+ biggest = 0
472
+ anomalies: list[str] = []
473
+ prev_role = None
474
+ for i, content in enumerate(contents):
475
+ parts = list(content.parts or [])
476
+ if not parts:
477
+ anomalies.append(f"empty-content@{i}")
478
+ calls = results = 0
479
+ for part in parts:
480
+ shape = _part_shape(part)
481
+ size = int(m.group(1)) if (m := re.search(r"\((\d+)\)$", shape)) else 0
482
+ total += size
483
+ biggest = max(biggest, size)
484
+ if shape.startswith("call:"):
485
+ calls += 1
486
+ elif shape.startswith("result:"):
487
+ results += 1
488
+ if size == 0 and shape.startswith(("text", "thought", "result")):
489
+ anomalies.append(f"zero-size-{re.split(r'[:(]', shape)[0]}@{i}")
490
+ if prev_role is not None and content.role == prev_role:
491
+ anomalies.append(f"repeated-role-{content.role}@{i}")
492
+ prev_role = content.role
493
+ nxt = contents[i + 1] if i + 1 < len(contents) else None
494
+ if calls and nxt is not None:
495
+ answered = sum(
496
+ 1 for p in (nxt.parts or []) if getattr(p, "function_response", None) is not None
497
+ )
498
+ if answered != calls:
499
+ anomalies.append(f"calls({calls})!=results({answered})@{i}")
500
+ tail = contents[-_SHAPE_TAIL_CONTENTS:]
501
+ first = len(contents) - len(tail)
502
+ tail_desc = "; ".join(
503
+ f"#{first + j} {c.role}[{', '.join(_part_shape(p) for p in (c.parts or []))}]"
504
+ for j, c in enumerate(tail)
505
+ )
506
+ cfg = ""
507
+ if config is not None:
508
+ cfg = (
509
+ f" tools={len(getattr(config, 'tools', None) or [])}"
510
+ f" system_instruction={bool(getattr(config, 'system_instruction', None))}"
511
+ f" cached_content={bool(getattr(config, 'cached_content', None))}"
512
+ )
513
+ return (
514
+ f"contents={len(contents)} approx_chars={total} largest_part={biggest}{cfg} "
515
+ f"anomalies={anomalies or 'none'} tail: {tail_desc}"
516
+ )
517
+
518
+
519
+ def _safe_describe_request(contents: list[Any], config: Any) -> str:
520
+ """``_describe_request`` that can never mask the error being diagnosed."""
521
+ try:
522
+ return _describe_request(contents, config)
523
+ except Exception as exc: # noqa: BLE001 - diagnostics must not replace the 400
524
+ return f"unavailable ({type(exc).__name__})"
525
+
526
+
527
+ def _error_details(exc: BaseException) -> str:
528
+ """Scrubbed ``details`` payload of an API error (may name the bad field)."""
529
+ details = getattr(exc, "details", None)
530
+ try:
531
+ return _scrub(json.dumps(details, default=str)) if details else ""
532
+ except (TypeError, ValueError):
533
+ return ""
534
+
535
+
429
536
  # Matches the fenced-json extraction argus.helpers.parse_review_result
430
537
  # already applies for the Claude path.
431
538
  _FENCED_JSON_RE = re.compile(r"```(?:json)?\s*(\{.*?\})\s*```", re.DOTALL)
@@ -965,26 +1072,68 @@ async def _run_turns(
965
1072
  _turn,
966
1073
  _scrub(str(exc)),
967
1074
  )
968
- raise
969
- logger.warning(
970
- "Gemini cache [%s] appears invalid upstream (role=%s); "
971
- "invalidating the local record and retrying this turn "
972
- "uncached: %s",
973
- label or "unlabeled",
974
- entry.role,
975
- exc,
976
- )
977
- await asyncio.to_thread(
978
- _invalidate_cache_record,
979
- entry=entry,
980
- model=model,
981
- system_prompt=system_prompt,
982
- )
983
- cached_content_name = None
984
- config = _build_config(None)
985
- response = await client.aio.models.generate_content(
986
- model=model, contents=contents, config=config
987
- )
1075
+ if not _is_invalid_argument(exc):
1076
+ raise
1077
+ # A 400 INVALID_ARGUMENT names no field, so record the
1078
+ # request's shape, then retry this SAME request once:
1079
+ # whether it recovers tells a transient fault from a
1080
+ # deterministic one.
1081
+ logger.error(
1082
+ "Gemini 400 request shape [%s] (turn=%d) details=%s: %s",
1083
+ label or "unlabeled",
1084
+ _turn,
1085
+ _error_details(exc) or "none",
1086
+ _safe_describe_request(contents, config),
1087
+ )
1088
+ lo = _env_float(
1089
+ "ARGUS_GEMINI_RETRY_INVALID_ARGUMENT_MIN_S",
1090
+ _INVALID_ARGUMENT_RETRY_MIN_S,
1091
+ )
1092
+ hi = _env_float(
1093
+ "ARGUS_GEMINI_RETRY_INVALID_ARGUMENT_MAX_S",
1094
+ _INVALID_ARGUMENT_RETRY_MAX_S,
1095
+ )
1096
+ await asyncio.sleep(
1097
+ random.uniform(lo, max(lo, hi)) # noqa: S311 - jitter, not security
1098
+ )
1099
+ try:
1100
+ response = await client.aio.models.generate_content(
1101
+ model=model, contents=contents, config=config
1102
+ )
1103
+ except Exception as retry_exc:
1104
+ logger.error(
1105
+ "Gemini 400 retry failed [%s] (turn=%d): %s",
1106
+ label or "unlabeled",
1107
+ _turn,
1108
+ _scrub(str(retry_exc)),
1109
+ )
1110
+ raise
1111
+ logger.warning(
1112
+ "Gemini 400 recovered on retry [%s] (turn=%d): the failure "
1113
+ "was transient, not a malformed request",
1114
+ label or "unlabeled",
1115
+ _turn,
1116
+ )
1117
+ else:
1118
+ logger.warning(
1119
+ "Gemini cache [%s] appears invalid upstream (role=%s); "
1120
+ "invalidating the local record and retrying this turn "
1121
+ "uncached: %s",
1122
+ label or "unlabeled",
1123
+ entry.role,
1124
+ exc,
1125
+ )
1126
+ await asyncio.to_thread(
1127
+ _invalidate_cache_record,
1128
+ entry=entry,
1129
+ model=model,
1130
+ system_prompt=system_prompt,
1131
+ )
1132
+ cached_content_name = None
1133
+ config = _build_config(None)
1134
+ response = await client.aio.models.generate_content(
1135
+ model=model, contents=contents, config=config
1136
+ )
988
1137
 
989
1138
  progress.responses += 1
990
1139
  usage = response.usage_metadata
@@ -522,8 +522,12 @@ async def _apply_dismissals(
522
522
  "or -1 if it doesn't match any finding."
523
523
  )
524
524
 
525
+ # No `temperature`: with langchain-anthropic 1.7.7 (anthropic 2.3.0), structured output on this
526
+ # call sends it in a way the Anthropic API rejects ("temperature: Extra inputs are not
527
+ # permitted", HTTP 400), and a failure here aborts the whole review. The matching task
528
+ # is deterministic enough without it.
525
529
  llm = _get_llm(
526
- f"anthropic:{CLAUDE_MINI}", "dismiss_match", max_tokens=1024, temperature=0
530
+ f"anthropic:{CLAUDE_MINI}", "dismiss_match", max_tokens=1024
527
531
  ).with_structured_output(DismissMatches)
528
532
  result = await llm.ainvoke([{"role": "user", "content": prompt}])
529
533
 
@@ -1037,7 +1041,8 @@ _TEMPERATURE_UNSUPPORTED_MODELS: frozenset[str] = (
1037
1041
  # set on the assumption Anthropic's temperature deprecation
1038
1042
  # applies across the Sonnet line, not per-checkpoint; re-verify
1039
1043
  # empirically and correct this comment if that assumption turns
1040
- # out wrong. claude-haiku-4-5 (CLAUDE_MINI) accepts it fine, so
1044
+ # out wrong. claude-haiku-4-5 (CLAUDE_MINI) accepts it fine on plain
1045
+ # requests (not with structured output, see _apply_dismissals), so
1041
1046
  # is deliberately not included here -- see the explicit
1042
1047
  # exclusion below for why "not included" alone isn't enough.
1043
1048
  ALIAS_MAP["claude-frontier"],
@@ -1046,18 +1051,20 @@ _TEMPERATURE_UNSUPPORTED_MODELS: frozenset[str] = (
1046
1051
  CLAUDE_DEFAULT,
1047
1052
  }
1048
1053
  )
1049
- # Explicit exclusion, not just an absent literal: without this, an
1050
- # override that happens to land CLAUDE_DEFAULT/CLAUDE_FRONTIER on
1051
- # CLAUDE_MINI's own resolved value (e.g. `--specialist-model
1052
- # claude-haiku-4-5`, used directly by this suite's own tests) would pull
1053
- # that value into the union above via the CLAUDE_DEFAULT/CLAUDE_FRONTIER
1054
- # terms, and then _apply_dismissals's `_get_llm(f"anthropic:{CLAUDE_MINI}",
1055
- # "dismiss_match", temperature=0)` call would have ITS temperature silently stripped too
1056
- # -- a real, silent determinism regression in a call site that has
1057
- # nothing to do with the override, caught in Argus round 3 review of
1058
- # this PR. `- {CLAUDE_MINI}` guarantees CLAUDE_MINI's resolved value can
1059
- # never end up in this set regardless of what any other alias happens to
1060
- # resolve to.
1054
+ # Explicit exclusion, not just an absent literal: without this, an override that
1055
+ # happens to land CLAUDE_DEFAULT/CLAUDE_FRONTIER on CLAUDE_MINI's own resolved value
1056
+ # (e.g. `--specialist-model claude-haiku-4-5`, used directly by this suite's own
1057
+ # tests) would pull that value into the union above via the
1058
+ # CLAUDE_DEFAULT/CLAUDE_FRONTIER terms, silently stripping `temperature` from any plain
1059
+ # `_get_llm(f"anthropic:{CLAUDE_MINI}", ..., temperature=...)` call that has nothing
1060
+ # to do with the override. `- {CLAUDE_MINI}` guarantees CLAUDE_MINI's resolved value
1061
+ # can never end up in this set.
1062
+ #
1063
+ # `_apply_dismissals`, the former user of such a call, no longer passes `temperature`
1064
+ # at all: structured output on Haiku is rejected with langchain-anthropic 1.7.7
1065
+ # (HTTP 400, "temperature: Extra inputs are not permitted"). `langchain-anthropic` is
1066
+ # unpinned above 1.2.0, so re-check this before relying on `temperature` together
1067
+ # with structured output.
1061
1068
  - {CLAUDE_MINI}
1062
1069
  )
1063
1070
 
@@ -33,7 +33,12 @@ from google.genai import errors as genai_errors # noqa: E402
33
33
  from google.genai import types # noqa: E402
34
34
 
35
35
  from argus.bench import BenchEntry # noqa: E402
36
- from argus.gemini_runner import _MAX_TURNS_GEMINI, run_session_gemini # noqa: E402
36
+ from argus.gemini_runner import ( # noqa: E402
37
+ _MAX_TURNS_GEMINI,
38
+ _describe_request,
39
+ _safe_describe_request,
40
+ run_session_gemini,
41
+ )
37
42
  from argus.llm.models import resolve as resolve_alias # noqa: E402
38
43
  from argus.runners import _TURN_BUDGET_SYSTEM_PROMPT_LINE # noqa: E402
39
44
 
@@ -65,7 +70,7 @@ def _isolate_gemini_cache_dir(tmp_path: Any, monkeypatch: pytest.MonkeyPatch) ->
65
70
 
66
71
  monkeypatch.setenv("ARGUS_GEMINI_CACHE_DIR", str(tmp_path / "gemini-cache"))
67
72
  # Existing tests must not sleep through the real retry cooldown.
68
- for name in ("RATE_LIMIT", "COOLDOWN"):
73
+ for name in ("RATE_LIMIT", "COOLDOWN", "INVALID_ARGUMENT"):
69
74
  monkeypatch.setenv(f"ARGUS_GEMINI_RETRY_{name}_MIN_S", "0")
70
75
  monkeypatch.setenv(f"ARGUS_GEMINI_RETRY_{name}_MAX_S", "0")
71
76
  clear_settings_cache()
@@ -1601,10 +1606,13 @@ class TestCacheInvalidatedUpstreamDegradesGracefully:
1601
1606
  repo_root="/tmp/does-not-need-to-exist",
1602
1607
  )
1603
1608
 
1604
- # Not retried uncached (that would mean 2 awaits, as in the
1605
- # cache-invalid-recovery test above) -- the error propagated
1606
- # straight out of the cache-retry logic on the first attempt.
1607
- assert fake_client.aio.models.generate_content.await_count == 1
1609
+ # Not retried uncached: the one retry a 400 INVALID_ARGUMENT gets
1610
+ # re-sends the exact same request (cache reference included), unlike
1611
+ # the cache-invalid-recovery test above, which drops the cache.
1612
+ calls = fake_client.aio.models.generate_content.await_args_list
1613
+ assert len(calls) == 2
1614
+ assert calls[0].kwargs["config"].cached_content == "cachedContents/abc123"
1615
+ assert calls[1].kwargs["config"].cached_content == "cachedContents/abc123"
1608
1616
  assert result.failure_reason == "worker_crashed"
1609
1617
  assert result.result_text == ""
1610
1618
 
@@ -1807,7 +1815,7 @@ class TestTransientRetry:
1807
1815
  [
1808
1816
  (401, "UNAUTHENTICATED"),
1809
1817
  (403, "PERMISSION_DENIED"),
1810
- (400, "INVALID_ARGUMENT"),
1818
+ (400, "FAILED_PRECONDITION"),
1811
1819
  (404, "NOT_FOUND"),
1812
1820
  ],
1813
1821
  )
@@ -1818,6 +1826,49 @@ class TestTransientRetry:
1818
1826
  assert result.failure_reason == "worker_crashed"
1819
1827
  assert result.failure_detail == f"{code} {status}, turn 0"
1820
1828
 
1829
+ async def test_invalid_argument_logs_request_shape_and_retries_once(
1830
+ self, caplog: pytest.LogCaptureFixture
1831
+ ) -> None:
1832
+ ok = _make_response(calls=[], usage=_make_usage(1, 1), text="fine")
1833
+ client = _make_fake_client([_api_error(400, "INVALID_ARGUMENT"), ok])
1834
+ with caplog.at_level("INFO", logger="argus.gemini_runner"):
1835
+ result = await _run(client)
1836
+ assert result.failure_reason is None
1837
+ assert client.aio.models.generate_content.await_count == 2
1838
+ messages = [r.message for r in caplog.records]
1839
+ assert any("Gemini 400 request shape [retry-test] (turn=0)" in m for m in messages)
1840
+ assert any("Gemini 400 recovered on retry [retry-test]" in m for m in messages)
1841
+
1842
+ async def test_describer_failure_does_not_skip_the_retry(self) -> None:
1843
+ ok = _make_response(calls=[], usage=_make_usage(1, 1), text="fine")
1844
+ client = _make_fake_client([_api_error(400, "INVALID_ARGUMENT"), ok])
1845
+ with patch("argus.gemini_runner._describe_request", side_effect=ValueError("odd args")):
1846
+ result = await _run(client)
1847
+ assert result.failure_reason is None
1848
+ assert client.aio.models.generate_content.await_count == 2
1849
+
1850
+ async def test_inverted_retry_bounds_do_not_crash(
1851
+ self, monkeypatch: pytest.MonkeyPatch
1852
+ ) -> None:
1853
+ monkeypatch.setenv("ARGUS_GEMINI_RETRY_INVALID_ARGUMENT_MIN_S", "0.02")
1854
+ monkeypatch.setenv("ARGUS_GEMINI_RETRY_INVALID_ARGUMENT_MAX_S", "0")
1855
+ ok = _make_response(calls=[], usage=_make_usage(1, 1), text="fine")
1856
+ client = _make_fake_client([_api_error(400, "INVALID_ARGUMENT"), ok])
1857
+ result = await _run(client)
1858
+ assert result.failure_reason is None
1859
+
1860
+ async def test_invalid_argument_failing_twice_fails_session_after_one_retry(
1861
+ self, caplog: pytest.LogCaptureFixture
1862
+ ) -> None:
1863
+ err = _api_error(400, "INVALID_ARGUMENT")
1864
+ client = _make_fake_client([err, err, err])
1865
+ with caplog.at_level("INFO", logger="argus.gemini_runner"):
1866
+ result = await _run(client)
1867
+ assert client.aio.models.generate_content.await_count == 2
1868
+ assert result.failure_reason == "worker_crashed"
1869
+ assert result.failure_detail == "400 INVALID_ARGUMENT, turn 0"
1870
+ assert any("Gemini 400 retry failed [retry-test]" in r.message for r in caplog.records)
1871
+
1821
1872
  async def test_httpx_connect_error_is_retried(self) -> None:
1822
1873
  ok = _make_response(calls=[], usage=_make_usage(1, 1), text="fine")
1823
1874
  client = _make_fake_client([httpx.ConnectError("reset"), ok])
@@ -1988,3 +2039,65 @@ class TestTransientRetry:
1988
2039
  text = " ".join(r.getMessage() for r in caplog.records)
1989
2040
  assert secret not in text
1990
2041
  assert len(text) < 1200
2042
+
2043
+
2044
+ class TestDescribeRequest:
2045
+ async def test_summarises_shape_without_leaking_text(self) -> None:
2046
+ contents = [
2047
+ types.Content(role="user", parts=[types.Part.from_text(text="SECRETBODY" * 50)]),
2048
+ types.Content(
2049
+ role="model",
2050
+ parts=[types.Part.from_function_call(name="grep", args={"pattern": "x"})],
2051
+ ),
2052
+ types.Content(role="user", parts=[types.Part.from_text(text="")]),
2053
+ ]
2054
+ out = _describe_request(contents)
2055
+ assert "SECRETBODY" not in out
2056
+ assert "contents=3" in out
2057
+ assert "call:grep(" in out
2058
+ assert "calls(1)!=results(0)@1" in out # the call was never answered
2059
+ assert "zero-size-text@2" in out
2060
+
2061
+ async def test_clean_request_reports_no_anomalies(self) -> None:
2062
+ contents = [
2063
+ types.Content(role="user", parts=[types.Part.from_text(text="hi")]),
2064
+ types.Content(
2065
+ role="model",
2066
+ parts=[types.Part.from_function_call(name="grep", args={"pattern": "x"})],
2067
+ ),
2068
+ types.Content(
2069
+ role="user",
2070
+ parts=[types.Part.from_function_response(name="grep", response={"output": "ok"})],
2071
+ ),
2072
+ ]
2073
+ assert "anomalies=none" in _describe_request(contents)
2074
+
2075
+ async def test_empty_tool_result_is_flagged(self) -> None:
2076
+ contents = [
2077
+ types.Content(
2078
+ role="model",
2079
+ parts=[types.Part.from_function_call(name="grep", args={"pattern": "x"})],
2080
+ ),
2081
+ types.Content(
2082
+ role="user",
2083
+ parts=[types.Part.from_function_response(name="grep", response={})],
2084
+ ),
2085
+ ]
2086
+ assert "zero-size-result@1" in _describe_request(contents)
2087
+
2088
+ async def test_empty_wrapped_tool_result_is_flagged(self) -> None:
2089
+ contents = [
2090
+ types.Content(
2091
+ role="user",
2092
+ parts=[types.Part.from_function_response(name="grep", response={"result": ""})],
2093
+ ),
2094
+ ]
2095
+ assert "zero-size-result@0" in _describe_request(contents)
2096
+
2097
+ async def test_first_turn_without_role_is_not_a_repeat(self) -> None:
2098
+ contents = [types.Content(role=None, parts=[types.Part.from_text(text="hi")])]
2099
+ assert "repeated-role" not in _describe_request(contents)
2100
+
2101
+ async def test_safe_describe_never_raises(self) -> None:
2102
+ out = _safe_describe_request(object(), None)
2103
+ assert out.startswith("unavailable (")
@@ -188,14 +188,15 @@ def test_temperature_guard_does_not_swallow_claude_mini_when_specialist_model_co
188
188
  was over-inclusive. Setting ARGUS_SPECIALIST_MODEL=claude-haiku-4-5 makes
189
189
  CLAUDE_DEFAULT resolve to the SAME value as CLAUDE_MINI, pulling
190
190
  CLAUDE_MINI's own resolved value into _TEMPERATURE_UNSUPPORTED_MODELS via
191
- the CLAUDE_DEFAULT term of the union. _match_dismissals
192
- (argus/graph.py) calls `_get_llm(f"anthropic:{CLAUDE_MINI}",
193
- temperature=0)` for deterministic dismissal-index matching -- that
194
- temperature=0 would then be silently stripped, breaking the exact
195
- invariant this set's own comment documents ("claude-haiku-4-5
196
- (CLAUDE_MINI) accepts it fine, so is deliberately not included here").
197
- The explicit `- {CLAUDE_MINI}` exclusion in the set's construction is
198
- what prevents this.
191
+ the CLAUDE_DEFAULT term of the union. A plain
192
+ `_get_llm(f"anthropic:{CLAUDE_MINI}", ..., temperature=0)` call (this was
193
+ the dismissal-index matcher, which no longer passes temperature because
194
+ Haiku structured output rejects it) would then have its temperature
195
+ silently stripped, breaking the invariant this set's own comment
196
+ documents ("claude-haiku-4-5 (CLAUDE_MINI) accepts it fine on plain
197
+ requests, so is deliberately not included here"). The explicit
198
+ `- {CLAUDE_MINI}` exclusion in the set's construction is what prevents
199
+ this.
199
200
  """
200
201
  overridden_set = _temperature_unsupported_models_under_override(CLAUDE_MINI)
201
202
  assert CLAUDE_MINI not in overridden_set
@@ -7,6 +7,7 @@ round 2 writer messages, and model contracts.
7
7
  from __future__ import annotations
8
8
 
9
9
  import importlib
10
+ import inspect
10
11
  import json
11
12
  import sys
12
13
  from unittest.mock import AsyncMock, MagicMock, patch
@@ -1174,6 +1175,48 @@ class TestApplyDismissals:
1174
1175
  assert matched[0].description == "except Exception"
1175
1176
  assert matched[0].dismissed_by == "user1"
1176
1177
 
1178
+ @pytest.mark.asyncio
1179
+ async def test_dismiss_match_call_sends_no_temperature(self) -> None:
1180
+ """The dismiss_match LLM call must not pass `temperature`.
1181
+
1182
+ With structured output, the current langchain-anthropic request is rejected by the
1183
+ Anthropic API when a temperature is set ("temperature: Extra inputs are not
1184
+ permitted", HTTP 400), and a failure here aborts the whole review.
1185
+ """
1186
+ _reload_graph()
1187
+ from argus.graph import _apply_dismissals
1188
+ from argus.graph import _get_llm as _get_llm_fn
1189
+
1190
+ from pydantic import BaseModel
1191
+
1192
+ class DismissMatch(BaseModel):
1193
+ dismiss_index: int
1194
+ finding_index: int
1195
+
1196
+ class DismissMatches(BaseModel):
1197
+ matches: list[DismissMatch]
1198
+
1199
+ findings = [PriorFinding(severity="BLOCKING", description="issue")]
1200
+ dismissals = [DismissedFinding(description="issue", dismissed_by="u", reason="r")]
1201
+
1202
+ mock_structured = MagicMock()
1203
+ mock_structured.ainvoke = AsyncMock(
1204
+ return_value=DismissMatches(matches=[DismissMatch(dismiss_index=0, finding_index=0)])
1205
+ )
1206
+ mock_llm = MagicMock()
1207
+ mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
1208
+
1209
+ with patch("argus.graph._get_llm", return_value=mock_llm) as mock_get_llm:
1210
+ await _apply_dismissals(findings, dismissals)
1211
+
1212
+ mock_get_llm.assert_called_once()
1213
+ mock_llm.with_structured_output.assert_called_once()
1214
+ bound = inspect.signature(_get_llm_fn).bind(
1215
+ *mock_get_llm.call_args.args, **mock_get_llm.call_args.kwargs
1216
+ )
1217
+ assert bound.arguments["stage"] == "dismiss_match"
1218
+ assert "temperature" not in bound.arguments
1219
+
1177
1220
  @pytest.mark.asyncio
1178
1221
  async def test_out_of_range_indices_ignored(self) -> None:
1179
1222
  """Matches with out-of-range indices are silently skipped."""