[CodexVeritas]

Summary

CombinedLlmArchetype.invoke added response.cost to the active MonetaryCostManagers after its direct call. That direct call goes through GeneralLlm._mockable_direct_call_to_model, which already adds the same cost, so every call through the deprecated model classes (Gpt4o, Gpt4oVision, Perplexity, DeepSeekR1, ...) was counted twice.

This is separate from the litellm-callback double count addressed in #358. The two PRs touch different files.

Changes

• invoke no longer adds the cost a second time. The raise_error_if_limit_would_be_reached() check before the call stays. It is the only pre-call limit check on this path, because litellm catches and logs exceptions raised from a callback's log_pre_api_call.
• The hard-limit tests in test_models_incurring_cost.py mock one level lower for archetypes. They mocked the archetype's direct call, which also skips the cost counting inside GeneralLlm. For archetypes they now mock GeneralLlm's litellm call instead, via the new AiModelMockManager.mock_general_llm_litellm_call_with_value. The cost is then counted by the real code path. Other models (ExaSearcher) keep the direct-call mock.
• New unit test test_invoke_counts_call_cost_once. It invokes Gpt4o with a mocked litellm call and checks the manager total equals one call's cost. Before this change it was double (0.001 vs 0.0005).

Testing

• code_tests/unit_tests: the new test passes. The only failures are 14 order-dependent tests that also fail on main and pass when run alone.
• The mocked hard-limit tests in test_models_incurring_cost.py (-k error_thrown) pass for all 5 models (15 cases). Removing the line without moving the mocks made the 12 archetype cases fail.

🤖 Generated with Claude Code