## Summary This epic will track the needs of MCP Server instrumentation. --- #### Instrumentation use cases The below table tracks the use cases to solve, **in priority order**, for when instrumenting the MCP Server: | Priority | Issue | Use case| Details| Related/Dependency| |---|---|-------|---------|------| | 1 | gitlab-org/gitlab#630113+s | Origin marker on DAP sessions created over MCP | Sessions created through MCP marked with their origin at creation, so existing DAP session outcomes are filterable by MCP without a join | Flows workstream, https://gitlab.com/gitlab-org/gitlab/-/work_items/607619 | | 2 | gitlab-org/gitlab#630114+s | Which DAP flow or agent was invoked | Identity and version on every invocation, plus whether it is foundational or a catalog item | Flows workstream, https://gitlab.com/gitlab-org/gitlab/-/work_items/607619 | | 3 | gitlab-org/gitlab#630105+s | Namespace on tool-call events | 39.4% of calls carry none, which blinds the flagship-account target. Likely a resolution-timing problem on start_mcp_tool_call | MCP Server. No dependency | | 4 | gitlab-org/gitlab#630115+s | Credit consumption attributable to MCP origin | Credit-bearing events attributable to MCP Server origin | Depends on 1. Sequence after | | 5 | gitlab-org/gitlab#630108+s | Client identity on every tools/call | Carried from the initialize handshake onto each call, or the session made joinable to it | MCP Server. No dependency | | 6 | gitlab-org/gitlab#630109+s | Alias-resolved vs canonical-name flag | The only direct measure of whether the alias map is working | Alias map, https://gitlab.com/gitlab-org/gitlab/-/work_items/604609 | | 7 | gitlab-org/gitlab#630110+s | Aggregated tool dispatched sub-tool | Which sub-tool a composite tool dispatched to. Resolved sub-tool name and resource type, not the route or the query string | MCP Server. No dependency | | 8 | gitlab-org/gitlab#630107+s | Protocol methods: initialize, tools/list, tools/call | Decides whether renames break existing connections or only affect new ones | MCP Server. Informs 6 | | 9 | gitlab-org/gitlab#630111+s | Response payload size | Token proxy for context cost | MCP Server. Confirm `gsc_extra` first | | 10 | gitlab-org/gitlab#631372+s | Dropped MCP Connections | Connection/reconnect failures (expired token, registration failure, missing `mcp` scope) happen before tool dispatch, so they are invisible to `start_mcp_tool_call`/`finish_mcp_tool_call` and to the failure-rate epic (gitlab-org#23705). Also a measurement-integrity risk: re-registration on reconnect can look like usage growth in "new OAuth applications created" and per-client connection-volume counts | Server-side only, no client change needed. Registration endpoint ownership unconfirmed, likely Authentication group | Four more issues came out of the breakdown. They are children of this epic but do not map to a row above, because they were not in the original use-case list: - gitlab-org/gitlab#630106+s - the two MCP events use none of the flat Snowflake columns (`label`, `property`, `value`), so every dimension needs JSON extraction to query - gitlab-org/gitlab#630112+s - `failure_reason` on `finish_mcp_tool_call` still records a wrapper class name, while the log now carries a real error category - gitlab-org/gitlab#630116+s - `duo_add_on` is declared on both events but never populated - gitlab-org/gitlab#630117+s - the Orbit MCP endpoint (`/api/v4/orbit/mcp`) emits no instrumentation at all, so every measurement here misses it --- ### Scope Instrumentation should send events for: 1. To Snowflake for internal dashboards and data analysis 2. To GitLab's analytics dashboard for customer-facing reference 3. To GitLab audit events for comprehensive events coverage ([related](https://gitlab.com/groups/gitlab-org/-/work_items/22496)) 4. To `mcp.log` through `Gitlab::Mcp::Logger`, for operational debugging in Kibana