Add Cost Review
to your pull requests.

P2Reasoning repair measures the gap against raw input, re-billing cached input as completion for additive-cache providers · Cost Review

When an additive-cache provider (Anthropic/Bedrock — providers NOT in SUBSET_PROVIDERS, whose `input_tokens` excludes the cached tokens) reports `total_tokens` that includes the cached prompt tokens, `resolveCompletionTokens` returns `total - input = output + cache`, so the cached tokens are charged at the completion rate in addition to the cache read/write rate that `spendStructuredTokens`/`prepareStructuredTokenSpend` already bills from the same entry. The repo itself proves cache is separate from `input_tokens` for these providers: the additive branch computes `totalInput: rawInput + cacheCreation + cacheRead`, yet the new gap check uses raw `usage.input_tokens` only. Example entry that mis-bills: `{ input_tokens: 100, output_tokens: 50, total_tokens: 190, cache_read_input_tokens: 40, provider: 'anthropic' }` → completion 90 instead of 50, with the 40 cache tokens billed twice. For the subset providers the same line is harmless because `input_tokens` already contains the cache, which is why the new tests (all OpenAI/Vertex) pass.

P1Anthropic cached prompt tokens are recorded but excluded from input/total, so cached calls understate spend · Cost Review

Anthropic's `usage.input_tokens` counts only the tokens that were neither read from nor written to the prompt cache; the real prompt size is `input_tokens + cache_read_input_tokens + cache_creation_input_tokens`. The diff now records the cache read/write counts as `gen_ai.usage.input_tokens.cached` and `gen_ai.usage.input_tokens.cache_write`, but still passes `input_tokens` unchanged and lets the shared `record_token_usage` derive `total_tokens = input_tokens + output_tokens`. So for a cached call the record reports input=100, cached=80, cache_write=20, output=50, total=150, while the actual prompt was 200 tokens and the actual total 250. Every cached Anthropic call therefore bills its cached prompt tokens as zero in the input/total the ledger reports, which is exactly the spend the cache fields were added to expose. This differs from the repo's OpenAI/Google records, where the provider's prompt-token count already includes the cached subset and the total comes from the provider, so the new fields sit on top of a complete total.

P2Cache tokens dropped when inputTokens is absent (Langfuse and PostHog) · Cost Review

When a MODEL_GENERATION span's `usage` has `inputDetails.cacheRead`/`inputDetails.cacheWrite` but no total `inputTokens`, `formatUsageMetrics` no longer emits `cache_read_input_tokens`/`cache_write_input_tokens` (PostHog: `$ai_cache_read_input_tokens`/`$ai_cache_creation_input_tokens`). The PR moved the cache extraction from a top-level check into the `if (usage.inputTokens !== undefined)` block in both exporters, so cache fields are now gated on the unrelated presence of the input total. `UsageStats` declares `inputTokens?` and `inputDetails?` as independently optional (`packages/core/src/observability/types/tracing.ts:134` and `:138`), and the repo's sibling exporters record cache fields without gating on `inputTokens` (e.g. `observability/datadog/src/metrics.ts:25-28`, `observability/otel-exporter/src/gen-ai-semantics.ts:89-92`), so the cached portion of a span's spend becomes invisible for such spans. The same change is present at `observability/posthog/src/tracing.ts:39` (and its cacheWrite branch at `:44`).

P1message_delta overwrites usage from message_start, dropping input_tokens · Cost Review

The RawMessageDeltaEvent branch rebuilds usage_metadata from scratch with `usage_metadata = { ... }`, replacing the input/output counts captured at message_start. On a standard Anthropic stream the message_delta `usage` object reports the cumulative output_tokens and does not repeat the prompt's input_tokens, so the final chunk's record (the only chunk whose raw usage the core TokenCountingHandler reads) ends up with input_tokens taken from the delta event (absent/None) and the real input count captured at line 610 is lost. Every streamed call therefore misstates prompt-side spend.