CLI recipes still target v3 traces API; Cloud v4 needs observations v2 #107
lambi702
started this conversation in
Ideas & Improvements
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Describe your idea or feedback
The skill is solid for Python SDK v4 instrumentation (
references/sdk-upgrade.mdmatches the official v3→v4 guide:propagate_attributes(),start_observation(), noCallbackHandler(update_trace=...)). It is not aligned for reading data on Langfuse Cloud v4 (observations-first model, already live on Cloud).What we were doing: pull the latest
chat-turntraces from Langfuse Cloud (EU) to check generationmodel/ cost and TOOLERRORlevels, following the skill (npx langfuse-cli, then error-analysis recipes).What the skill did vs expected:
traces listis treated as the primary read path.references/error-analysis.mdstill has the gotcha--limit > 100 on traces list. On Cloud,npx langfuse-cli api traces list --helpnow describeslegacy-traces, deprecated, with a cut-off of 16 November 2026. It tells agents to useGET /api/public/v2/observations?fromStartTime=&toStartTime=instead. The skill never mentions that replacement, so the agent either hits a CLI that “cannot be called” or falls back to rawcurlon/api/public/traces.references/cli.mdsays “Preferobservations-v2soverobservations”. In Python SDK v4 those namespaces were remapped:api.observations_v_2→api.observations, legacy v1 →api.legacy.*. The CLI OpenAPI names and the SDK names now disagree; the skill does not say which world it is describing, or that Cloud v4 reads should start from Observations API v2 / root observations (Is Root Observation = true), notGET /api/public/traces.Auth env var drift. The skill exports
LANGFUSE_HOST. Current CLI docs documentLANGFUSE_BASE_URL(withLANGFUSE_HOSTas a leftover). Agents following the skill literally can miss the host the CLI actually reads.This is about the skill recipes, not a product bug. Ingestion via SDK v4 / OTel worked; only the “how to list/get a trace” instructions are stale.
What would the ideal outcome look like?
SKILL.md/cli.md/error-analysis.mddistinguish SDK v4 (write path) from Cloud / API v4 (read path).__schema, notlegacy-traces), then fetch children bytrace_id.traces listpagination tips,prefer observations-v2s.LANGFUSE_BASE_URLandLANGFUSE_HOST, and say which the current CLI honors.__schemafirst — that part already works.All reactions