Export: OpenTelemetry GenAI spans

deedglass export --format otel-genai writes each sealed record as an OpenTelemetry span, using the GenAI semantic conventions’ execute_tool span, so a trace backend can show what an agent’s tool calls did next to the rest of its traces. The output is OTLP/JSON, one ExportTraceServiceRequest per line (the format the OpenTelemetry file exporter writes and the Collector’s otlpjsonfile receiver reads). Overview and the other formats: EXPORTS.md.

Pinned versions

What	Version	Date	Status
Semantic conventions, `gen_ai.*` attributes and the `execute_tool` span	v1.41.0 (`schemaUrl` `https://opentelemetry.io/schemas/1.41.0`)	2026-04-28	Development
OTLP/JSON encoding	opentelemetry-proto v1.11.1, `trace/v1`	2026-09-29	Stable

v1.41.0 is the last semantic-conventions release that defines the GenAI attributes in the main repository. Later releases mark them as moved to open-telemetry/semantic-conventions-genai, which had published no release on 2026-10-19. When it does, the export moves to it in one change: the attribute registry file, the goldens and this page. The GenAI conventions are at Development status, so names can still change.

Mapping

Span field or attribute	From the record	Notes
`traceId`	`action.id`	the first 16 bytes of SHA-256 over a fixed tag and the action id: every record about one action lands in one trace
`spanId`	`deed_id`	the first 8 bytes, the same way
`parentSpanId`	`amends` (amendments only)	the amended record’s span
`name`	`action.tool`	`execute_tool {tool}` for an action; `deedglass.check {tool}` for an amendment
`kind`	fixed	`INTERNAL` (1)
`startTimeUnixNano`, `endTimeUnixNano`	`execution.started_at`, `finished_at`	`issued_at` for both when the call never ran, and for amendments
`status`	`execution`, `policy`, `approval`	unset when the call ran and succeeded; `ERROR` (2) when it was denied, never ran, failed or its outcome is unknown
`gen_ai.operation.name`	fixed	`execute_tool` (actions only)
`gen_ai.tool.name`	`action.tool`
`gen_ai.tool.call.id`	`action.id`
`gen_ai.tool.type`	`action.mode`	`extension` when Deedglass ran the tool (`gateway`); `function` when the agent ran it and reported back (`proxy`)
`gen_ai.agent.id`, `gen_ai.agent.name`	`agent.id`, `agent.name`
`gen_ai.conversation.id`	`session`	when present
`error.type`	the reason it didn’t run cleanly	`policy_denied`, `approval_denied`/`approval_expired`/`approval_cancelled`, `not_run`, the record’s `error_code`, `outcome_unknown`, or `_OTHER`
`deedglass.record.*`	`deed_id`, `kind`, `seq`, `issued_at`, `body_hash`, `seal`, `seal_key_id`, `amends`	to find the sealed record the span came from
`deedglass.on_behalf_of`	`on_behalf_of`	the opaque reference, as recorded
`deedglass.action.*`	`mode`, `connector`, `request_digest`, `intent_hash`, `intent_signed`
`deedglass.policy.*`	`decision`, `mode`, `rule_id`, `id`, `version_id`, `version_hash`
`deedglass.approval.*`	`id`, `outcome`, `decided_by` (an array), `decided_at`
`deedglass.execution.*`	`status`, `latency_ms`, `result_digest`, `error_code`
`deedglass.check.*`	`witness.level`, `verdict`, `method`, `detail_digest`	the independent check: `VERIFIED`, `MISMATCH` or `UNVERIFIABLE`
`deedglass.cost.*`	`cost.amount`, `currency`
resource `service.name`	fixed	`deedglass`
resource `deedglass.org.id`	`org`
scope	fixed	`deedglass.export`, with the Deedglass version that wrote the file

Every attribute is either in the pinned semantic-conventions registry (and not deprecated there) or under the deedglass. namespace. CI checks that, the execute_tool requirements (the required attributes, the span name, error.type exactly when the span ended in an error) and the OTLP/JSON shape (hex ids, integer enums, 64-bit integers as decimal strings) on every span of a real release’s bundle.

What is lost or approximated

The verdict is not the span status. OpenTelemetry’s status describes the call. A call that ran and succeeded but whose effect didn’t match what the agent reported has an unset status and deedglass.check.verdict = MISMATCH. Alert on that attribute, not on the status.
A later check is a second span. An amendment can’t change a span already exported, so it becomes a child span of the action’s span with the new verdict. A reader takes the latest one.
Trace context. The ids are derived from Deedglass ids, not taken from the agent’s own trace: a record doesn’t carry the caller’s W3C traceparent, so these spans don’t join the agent’s trace by themselves. Correlate on gen_ai.tool.call.id (the action id).
Admin records (agent stopped, member invited, ledger pruned, and so on) have no GenAI span and are skipped; the count is printed.
No arguments or results. gen_ai.tool.call.arguments and gen_ai.tool.call.result are never written: the record holds only their digests, which are in deedglass.action.request_digest and deedglass.execution.result_digest.
gen_ai.tool.type is approximate. The conventions’ three kinds don’t include “a gateway ran it for the agent”; extension is the closest.
gen_ai.conversation.id is the agent’s own session reference, which may not be a conversation in the model provider’s sense.
No model, provider or token counts. Deedglass sees the tool call, not the model call.
The seal can’t be checked from the span. The attributes don’t reproduce the record’s canonical body, so deedglass.record.seal is there to find the record, not to check it. Check the record with deedglass verify.
Development status. Attribute names may change in the GenAI conventions; a new pin regenerates the export from the records.