traceroot-ai/traceroot

Summary When a span reports Anthropic's usage the way Anthropic returns it, input_tokens is only the tokens after the last cache breakpoint; the cached prefix is reported separately as cache_read_input_tokens / cache_creation_input_tokens. The worker assumes the opposite convention (input already includes the cache) and computes uncached = input − cache_read − cache_write, floored at 0. For these emitters that is always negative, so every uncached token is floored away: the span stores input_tokens = cache_read + cache_write (a flat number equal to the cache size), the usage breakdown shows uncached 0, and cost is short by exactly input_tokens × input rate on every call. The gap grows on every turn of an agent loop, because the new tool results and messages are precisely the uncached part. Anthropic's documentation (prompt caching, "Tracking cache performance"): source: https://platform.claude.com/docs/en/build-with-claude/prompt-caching input_tokens: Number of input tokens which

Part of #2357 Summary Check whether the OpenRouter integration page and its example still work with the latest OpenRouter release. If they do, report the result and close this issue. If not, update them. Where it lives Docs page: docs/integrations/openrouter.mdx Example: examples/python/openrouter-tool-agent Example: examples/typescript/openrouter Instrumentation: traceroot-py, traceroot-ts Step 1: Check Find the latest OpenRouter release and its changelog. Compare it with the version pinned in the example and the model ids used on the page. Run 2–3 examples from OpenRouter's own official docs (quickstart, tool calling, and anything the page mentions), with TraceRoot initialized exactly as our page says. Run our example as-is from a fresh environment (see its README). In TraceRoot, compare the traces to the page's What Gets Captured table: span names and nesting, inputs and outputs, tokens, cost, session and user ids. If everything matches, comment with the versions you tested, t

Part of #2357 Summary Check whether the OpenAI integration page and its example still work with the latest OpenAI release. If they do, report the result and close this issue. If not, update them. Where it lives Docs page: docs/integrations/openai.mdx Example: examples/python/openai-tool-agent Example: examples/typescript/openai Instrumentation: traceroot-py, traceroot-ts Step 1: Check Find the latest OpenAI release and its changelog. Compare it with the version pinned in the example and the model ids used on the page. Run 2–3 examples from OpenAI's own official docs (quickstart, tool calling, and anything the page mentions), with TraceRoot initialized exactly as our page says. Run our example as-is from a fresh environment (see its README). In TraceRoot, compare the traces to the page's What Gets Captured table: span names and nesting, inputs and outputs, tokens, cost, session and user ids. If everything matches, comment with the versions you tested, the upstream examples you ran

Part of #2357 Summary Check whether the Mistral integration page and its example still work with the latest Mistral release. If they do, report the result and close this issue. If not, update them. Where it lives Docs page: docs/integrations/mistral.mdx Example: examples/python/mistral-tool-agent Instrumentation: traceroot-py Step 1: Check Find the latest Mistral release and its changelog. Compare it with the version pinned in the example and the model ids used on the page. Run 2–3 examples from Mistral's own official docs (quickstart, tool calling, and anything the page mentions), with TraceRoot initialized exactly as our page says. Run our example as-is from a fresh environment (see its README). In TraceRoot, compare the traces to the page's What Gets Captured table: span names and nesting, inputs and outputs, tokens, cost, session and user ids. If everything matches, comment with the versions you tested, the upstream examples you ran, and trace screenshots, and the issue can b

Part of #2357 Summary Check whether the Anthropic integration page and its example still work with the latest Anthropic release. If they do, report the result and close this issue. If not, update them. Where it lives Docs page: docs/integrations/anthropic.mdx Example: examples/python/anthropic-tool-agent Example: examples/typescript/anthropic Instrumentation: traceroot-py, traceroot-ts Step 1: Check Find the latest Anthropic release and its changelog. Compare it with the version pinned in the example and the model ids used on the page. Run 2–3 examples from Anthropic's own official docs (quickstart, tool calling, and anything the page mentions), with TraceRoot initialized exactly as our page says. Run our example as-is from a fresh environment (see its README). In TraceRoot, compare the traces to the page's What Gets Captured table: span names and nesting, inputs and outputs, tokens, cost, session and user ids. If everything matches, comment with the versions you tested, the upst