langfuse-mcp: Skill for Codex

Langfuse Trace Export

Download complete Langfuse traces (detail_level=full) directly to local JSON files with one script command. The script talks to the fixed gateway HTTP endpoint by itself — it does NOT require any MCP server to be registered in the agent, and it keeps the payload out of the agent context, returning only a small JSON summary.

Zero-interaction policy

Minimize interaction. When the user invocation already contains an ID and an environment (e.g. /langfuse-trace-export SES_2083147782008475648 test or 帮我导出测试环境 session SES_xxx), run the export command immediately without asking anything.

• IDs starting with SES_ (or explicitly called session/会话) are session IDs → use --session-id.
• Other IDs are trace IDs → use --trace-id.
• Environment words map as: 测试→test，开发→dev，预发→pre，生产/线上→onl.

Only ask when something required is truly missing, and ask for all missing values in a single question:

• missing environment → ask once with the four choices test / dev / pre / onl;
• missing ID → ask for the session_id or trace_id in the same question.

Never infer an environment. Never ask twice.

Fixed environment → gateway mapping (do not accept a custom URL):

Environment	Gateway URL
test, dev	http://mcp-gw.test.maxiot-inc.com/mcp_langfuse_test/mcp
pre, onl	https://mcp-gw.pre.maxiot-inc.com/mcp_langfuse/mcp

Session export (preferred)

Export every trace of a session in one command, run from the target project root:

python3 <langfuse-trace-export skill directory>/scripts/export_trace.py \
  --environment "<test|dev|pre|onl>" \
  --session-id "<session_id>" \
  --project-root "<absolute project root>"

The script:

1. Uses only the Python 3.9+ standard library; no project or third-party dependencies.
2. Connects directly to the gateway with the fixed non-secret placeholder header and completes the MCP initialize handshake itself (no agent-side MCP registration needed).
3. Discovers the get_session_trace_list and get_trace tools (exact name or -<name> suffixed).
4. Pages through get_session_trace_list(detail_level=compact) internally to collect all trace IDs — pagination stays inside the script, never in the agent.
5. Fetches each trace with get_trace(detail_level=full). If Langfuse rejects a trace because its observations exceed the server's single-response limit, fetches trace fields with list_traces(fields=core,io,scores,metrics) and all observations with paged list_observations(detail_level=full, limit=100), then validates and reconstructs the complete trace.
6. Writes each trace atomically to: