Summary
The MCP server never uses sampling/createMessage to ask the connected client model for help. For suggest_improvement, this is a natural fit: when a user submits raw feedback, the server could ask the client model "extract a 1-line title and a structured tag set from this body" before persisting — improving downstream search/triage quality at no cost to the user. Sampling is a server→client request defined by the MCP spec and is the protocol's way to delegate model-side work without forcing the server to embed an LLM client.
Where
src/CodeIndex/Mcp/McpToolHandlers.cs (suggest_improvement handler, no sampling call)src/CodeIndex/Mcp/McpServer.cs (no sampling capability negotiation)
Suggested approach
(1) Negotiate sampling capability in the initialize response (declare it, then check the client's capabilities.sampling before relying on it). (2) In suggest_improvement, when sampling is available, send a sampling/createMessage request to extract structured fields from the user's body, then store both raw and structured versions. (3) Fall back to the current raw-only path when client doesn't support sampling. (4) Add a feature flag so users can opt out if they don't want server→model traffic.
Summary
The MCP server never uses
sampling/createMessageto ask the connected client model for help. Forsuggest_improvement, this is a natural fit: when a user submits raw feedback, the server could ask the client model "extract a 1-line title and a structured tag set from this body" before persisting — improving downstream search/triage quality at no cost to the user. Sampling is a server→client request defined by the MCP spec and is the protocol's way to delegate model-side work without forcing the server to embed an LLM client.Where
src/CodeIndex/Mcp/McpToolHandlers.cs(suggest_improvement handler, no sampling call)src/CodeIndex/Mcp/McpServer.cs(no sampling capability negotiation)Suggested approach
(1) Negotiate sampling capability in the initialize response (declare it, then check the client's
capabilities.samplingbefore relying on it). (2) Insuggest_improvement, when sampling is available, send asampling/createMessagerequest to extract structured fields from the user's body, then store both raw and structured versions. (3) Fall back to the current raw-only path when client doesn't support sampling. (4) Add a feature flag so users can opt out if they don't want server→model traffic.