AI observability

This is a proposed direction, not an available feature or a delivery commitment. No release date is assigned.

Operator problem

Operators need to separate model/provider failures from query-tool failures and missing evidence.

Current alpha boundary

The alpha Gateway supports controlled read-only investigation, stored conversation replay and explicit failure/cancellation states. LM Studio/Ollama monitor templates inspect model-service metadata, not model quality. See the current operator guide.

Proposed capability

Explore declared model-request telemetry and evaluation references without exposing prompts or credentials by default.

Evidence required

Use real protocol success/refusal/timeout/cancel paths, secret redaction and independent request isolation; keep model-quality evaluation separate.

Non-goals

No universal LLM tracing, automatic quality score, external MCP activation or model-performance benchmark claim.

Contribution entry

Bring a concrete workflow and a reproducible failure to the contribution process. Agree the smallest contract and acceptance evidence before implementation. Return to the roadmap.