Added#
trazum from-otel: the cost lens over OpenTelemetry. The 1.71 plan generalises thefrom-claude-codepattern to the standard the ecosystem is converging on. A new coreotelRecords(text)reads the OTLP/JSON any GenAI exporter produces — one document or newline-delimited spans — and turns each LLM-call span into a usage-log record: the model (gen_ai.response.modelorgen_ai.request.model), the timestamp (startTimeUnixNano), a label (the span'sgen_ai.operation.nameor the resource'sservice.name, so a per-service bill falls out), and the input/output token counts. Spans that are not LLM calls are counted and skipped, never priced. The fortieth command walks a file or directory of.json/.jsonl/.ndjson, writes the usage log to-oor stdout, and reports on stderr how many spans were LLM calls, how many were skipped, and how many carried no cache data.--label-from-servicelabels by the resource'sservice.name. This makes Trazum complementary to every observability tool — LangSmith, Helicone, LiteLLM — reading whatever telemetry a team already emits rather than competing with the tool that emits it.The web Bill tab reads OTLP too. The folder drop gains a third arm: a dropped OpenTelemetry export is detected by
looksLikeOteland converted in the page withotelRecords, priced beside any transcripts and usage logs in the same drop — drag your OpenTelemetry export onto Trazum joins drag your ~/.claude/projects. No fetch, same invariant. The ingest banner names how many LLM spans were priced, how many non-LLM spans were skipped, and how many carried no cache data. Both locales.
Honest gaps, stated#
- No cache TTL split, because OTel has none. OpenTelemetry has not
standardised the cache-write TTL split, so an OTel-sourced record carries no
cache_creationobject and the cache verdicts read cannot tell rather than a fabricated one — the same refusal as inventing a price. Cache reads are taken only where agen_ai.usage.cache_read_input_tokens-shaped key is actually present. The converter says so on stderr and the documentation states it plainly. - Nothing but the numbers crosses. Prompt and completion content, trace ids
and every other span attribute stay in the span;
otel.test.jsplants a prompt and a trace id in a fixture and greps the whole conversion output for them, the same privacy proof the transcript converter carries. Thefrom-otelCLI suite andfolder-ingest.test.mjsextend it to the run and the web arm. - Vendor converters not guessed at. LangSmith, Helicone and LiteLLM are named as next, not built now: a converter for a documented-but-unseen format is an estimate wearing a parser's clothes. Each ships when a real export of it is seen.