OpenInference-shaped views at egress, not OTel emission #4

New issue

Open

opened 2026-05-23 20:55:34 +00:00 by coilysiren · 0 comments

coilysiren commented

2026-05-23 20:55:34 +00:00

Owner

Originally filed by @coilysiren on 2026-05-17T17:30:52Z - https://github.com/coilysiren/session-lattice/issues/13

Problem

The substrate's pivot from a four-tier OTel-spans pipeline to a stateless consumer of session-lattice views (per agentic-os-kai/docs/substrate.md) removes any OTel surface from the stack. That is the right call for the ingest path. It is the wrong call for egress, for two reasons:

bug-finder needs a frozen corpus that can be visualized somewhere. Without an OpenTelemetry / OpenInference-shaped view, bug-finder either re-implements a trace UI or ships as numbers-only.
The OpenInference ecosystem (Phoenix, Arize, and a growing list of evals tooling) already speaks a stable column shape for agent traces. A session-lattice view in that shape interoperates with the ecosystem at near-zero cost.

This is not a reversion to the retired OTel pipeline. Primary sources stay Claude session JSONL, git log, GitHub. The OpenInference view sits at egress, on top of repo-recall-fed views, not as an emitter upstream of repo-recall.

Proposal

Add a Feldera SQL pack to session-lattice that maintains materialized views in OpenInference's column shape. Specifically:

Columns: trace_id, span_id, parent_span_id, start_time, end_time, name, status_code, status_message, attributes (JSON).
OpenInference semantic-convention attributes broken out: openinference.span.kind (LLM / TOOL / CHAIN / RETRIEVER / AGENT / EMBEDDING), input.value, output.value, llm.model_name, llm.token_count.prompt, llm.token_count.completion, tool.name, tool.parameters.
View name pins the semconv version: openinference_spans_v0_1_x. A future bump adds a new view rather than mutating the existing one.
Trace ID generation: deterministic map from Claude session UUID. Span IDs deterministic from (session_id, turn_index, tool_index) so re-running the materialization is stable.
Span hierarchy: parent-child links populated for tool invocations under an assistant turn. Subagent / nested-invocation handling is an open sub-question.

Out of scope for v1

No live OTLP gRPC/HTTP exporter. The view is queryable via the existing HTTP egress; anyone who wants a push exporter can poll + POST themselves until there is real demand.
No commitment to keep the column shape stable across OpenInference major versions. The version pin in the view name is the contract.
No re-introduction of OTel into the ingest path.

Why this earns its keep

The single strongest argument is the bug-finder hook. With this view in place, bug-finder's frozen-corpus capture writes Parquet in OpenInference schema and an installer drops it into a local Phoenix instance for a trace UI. Without it, bug-finder needs its own visualization layer.

Interop with the broader OpenInference ecosystem (Arize, hosted Phoenix, OpenLLMetry) is a free secondary benefit.

Open design sub-questions

Subagent / nested-invocation parent-span linking. Claude session JSONL may not surface this cleanly; needs explicit attention.
Whether attributes JSON should be a single column or normalized into typed attribute columns. Phoenix loaders work fine with JSON; SQL ergonomics are better with typed columns. Probably both, via two adjacent views.
Whether to also emit OpenInference Traces over OTLP for runtime push later. Reserve the decision until someone asks.

Origin

Conversation 2026-05-17. Pairs with the substrate framing in agentic-os-kai/docs/substrate.md, the OpenInference standard, and the Arize "agent telemetry needs standards" piece.

_Originally filed by @coilysiren on 2026-05-17T17:30:52Z - [https://github.com/coilysiren/session-lattice/issues/13](https://github.com/coilysiren/session-lattice/issues/13)_ **Problem** The substrate's pivot from a four-tier OTel-spans pipeline to a stateless consumer of session-lattice views (per `agentic-os-kai/docs/substrate.md`) removes any OTel surface from the stack. That is the right call for the ingest path. It is the wrong call for **egress**, for two reasons: 1. **bug-finder** needs a frozen corpus that can be visualized somewhere. Without an OpenTelemetry / OpenInference-shaped view, bug-finder either re-implements a trace UI or ships as numbers-only. 2. The OpenInference ecosystem (Phoenix, Arize, and a growing list of evals tooling) already speaks a stable column shape for agent traces. A session-lattice view in that shape interoperates with the ecosystem at near-zero cost. This is **not** a reversion to the retired OTel pipeline. Primary sources stay Claude session JSONL, git log, GitHub. The OpenInference view sits at egress, on top of repo-recall-fed views, not as an emitter upstream of repo-recall. **Proposal** Add a Feldera SQL pack to session-lattice that maintains materialized views in OpenInference's column shape. Specifically: - Columns: `trace_id`, `span_id`, `parent_span_id`, `start_time`, `end_time`, `name`, `status_code`, `status_message`, `attributes` (JSON). - OpenInference semantic-convention attributes broken out: `openinference.span.kind` (LLM / TOOL / CHAIN / RETRIEVER / AGENT / EMBEDDING), `input.value`, `output.value`, `llm.model_name`, `llm.token_count.prompt`, `llm.token_count.completion`, `tool.name`, `tool.parameters`. - View name pins the semconv version: `openinference_spans_v0_1_x`. A future bump adds a new view rather than mutating the existing one. - Trace ID generation: deterministic map from Claude session UUID. Span IDs deterministic from `(session_id, turn_index, tool_index)` so re-running the materialization is stable. - Span hierarchy: parent-child links populated for tool invocations under an assistant turn. Subagent / nested-invocation handling is an open sub-question. **Out of scope for v1** - No live OTLP gRPC/HTTP exporter. The view is queryable via the existing HTTP egress; anyone who wants a push exporter can poll + POST themselves until there is real demand. - No commitment to keep the column shape stable across OpenInference major versions. The version pin in the view name is the contract. - No re-introduction of OTel into the ingest path. **Why this earns its keep** The single strongest argument is the bug-finder hook. With this view in place, bug-finder's frozen-corpus capture writes Parquet in OpenInference schema and an installer drops it into a local Phoenix instance for a trace UI. Without it, bug-finder needs its own visualization layer. Interop with the broader OpenInference ecosystem (Arize, hosted Phoenix, OpenLLMetry) is a free secondary benefit. **Open design sub-questions** - Subagent / nested-invocation parent-span linking. Claude session JSONL may not surface this cleanly; needs explicit attention. - Whether `attributes` JSON should be a single column or normalized into typed attribute columns. Phoenix loaders work fine with JSON; SQL ergonomics are better with typed columns. Probably both, via two adjacent views. - Whether to also emit OpenInference Traces over OTLP for runtime push later. Reserve the decision until someone asks. **Origin** Conversation 2026-05-17. Pairs with the substrate framing in `agentic-os-kai/docs/substrate.md`, the OpenInference standard, and the Arize "agent telemetry needs standards" piece.

coilysiren added the

label

2026-05-31 01:55:09 +00:00

coilysiren added

and removed

labels

2026-05-31 07:01:22 +00:00

No labels

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

coilyco-flight-deck/session-lattice#4

No description provided.

Rows
Columns