[giladsamuels-cognition]

Summary

Adds unit tests for TrackAPIPricingMixin.get_effective_cost and its Anthropic/OpenAI variants using non-zero cached tokens, and fixes the bug the tests caught. Closes code-scan finding sfind-ba28615afd004752ba3b8515683639d9.

Production change (Anthropic path only): the code read cache-read tokens from a field that doesn't exist, so the 0 default applied and cache-read cost was always dropped from effective_cost. anthropic.types.Usage has cache_read_input_tokens, cache_creation_input_tokens, input_tokens, output_tokens and no cache_input_tokens. Both cache fields are Optional[int], so a null value now falls back to 0.

- cache_read_tokens = getattr(usage, "cache_input_tokens", 0)
- cache_write_tokens = getattr(usage, "cache_creation_input_tokens", 0)
+ cache_read_tokens = getattr(usage, "cache_read_input_tokens", 0) or 0
+ cache_write_tokens = getattr(usage, "cache_creation_input_tokens", 0) or 0

The create_mock_anthropic_response helper in tests/llm/test_response_api.py also set the old field name. It now sets the real one, otherwise the MagicMock would return a mock for the missing attribute.

New tests in tests/llm/test_tracking.py build usage objects with the real SDK types (anthropic.types.Usage, openai CompletionUsage / ResponseUsage) and fixed prices (input_cost=2e-6, output_cost=1e-5). They assert exact costs:

• Anthropic: 100 new input + 50 output + 1000 cache-read (0.1x) + 200 cache-write (1.25x) → 1.4e-3. Also checks that cache reads are billed at the discount, and that missing or null (None) cache fields fall back to 0.
• OpenAI chat completion (prompt_tokens_details present): 1000 prompt tokens, 600 of them cached (0.5x), + 100 output → 2.4e-3. Also covers prompt_tokens_details=None.
• OpenAI Responses API (input_tokens_details branch): same numbers → 2.4e-3.
• Fallbacks: usage is None or missing → 0.0; unsupported provider (openrouter, unknown, None) → 0.0; litellm delegates to completion_cost (patched).
• End-to-end through TrackAPIPricingMixin.__call__: stats["effective_cost"], usage_cached_tokens and n_api_calls add up across calls. The tracker cost still bills all prompt tokens at the full price.

Without the field-name fix, the Anthropic cache-read tests fail. With it, uv run pytest tests/llm passes and black --check src/ is clean.

Devin-Org: engineering

Link to Devin session: https://app.devin.ai/sessions/190c921121af47b2956a84e336f23eef
Open in Devin Desktop: https://app.devin.ai/desktop/session/190c921121af47b2956a84e336f23eef?variant=devin
Requested by: @giladsamuels-cognition

---

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring