OpenTelemetry GenAI export
blackbox export turns the ledger into OTLP/JSON that follows the OpenTelemetry GenAI semantic conventions where one exists, so the record of what an agent did can be read by any OTel tool.
Local by default. With no flags it writes traces.json and logs.json to a folder and sends nothing. It sends data only to the URL you pass with --endpoint, and never anywhere else.
blackbox export [--session ID] [--out dir] [--endpoint URL]Metadata only. Session, tool and rule names, decisions, timestamps and chain hashes. Never prompts, commands, paths, tool arguments, results or summaries; those stay in the ledger, sealed under the session key.
Mapping
| Ledger | OTel |
|---|---|
PreToolUse .. PostToolUse / PostToolUseFailure (same tool_use_id) | span execute_tool <tool> |
session_id | gen_ai.conversation.id, and the trace id (sha256 of it) |
tool_name, tool_use_id, agent_id | gen_ai.tool.name, gen_ai.tool.call.id, gen_ai.agent.id |
| (constant) | gen_ai.operation.name=execute_tool, gen_ai.provider.name=anthropic |
decision records | span event and log record blackbox.decision; a deny or ask sets the span status to error |
taint, intent, purge, settings, genesis | log records blackbox.<kind> |
Attributes the conventions do not define use the blackbox.* namespace: blackbox.decision, blackbox.rule, blackbox.taint, blackbox.record.seq, blackbox.record.hash (the ledger position, so an exported span can be checked against blackbox verify).
Trace and span ids are derived from session and tool-call ids, so exporting the same ledger twice gives the same ids.
The GenAI conventions are still in development; the attribute names live in one table (GENAI in src/otel-genai.js) so they can follow upstream.
Not covered yet: model-call spans (chat, with model and token usage). They need the telemetry payloads, which are sealed, so they will come as an explicit opt-in.