mcp-data-platform-v1.88.1

mcp-data-platform v1.88.1

A focused follow-up to v1.88.0's canonical Knowledge Pages: this release makes the knowledge loop something a fresh agent can actually understand and use, and makes platform upgrades reach connected agents cleanly.

Agents understand the memory to knowledge loop

The apply_knowledge contract previously stated outcomes ("business and domain facts become knowledge pages") but never taught an agent the loop or how to perform the review step well. A context-free agent (Claude Code, Claude Desktop) now learns it from the contract itself:

• The apply_knowledge tool description is reframed to read as what it does, "review and apply knowledge," and teaches the expert workflow inline: discover existing knowledge first (so every decision is update-vs-create, never blind-create), compare each insight against what exists, synthesize related insights into one coherent artifact rather than promoting one-to-one, route entity-tied facts to DataHub and business or domain knowledge to a canonical knowledge page, update in place over duplicating, and mark insights applied, rejected, or superseded.
• It now documents the sink and page parameters in prose, so an agent reading the description (not just the JSON schema) discovers how to promote a business_knowledge or operational_rule capture to a knowledge page (sink=knowledge_page with a page object, found-or-created by slug so repeated promotions consolidate one living page).
• A new knowledge_apply_guidance prompt teaches the full lifecycle: what memory, insight, and knowledge each are, why durable knowledge matters (no user re-teaches the same fact; answers stay grounded in established truth), and the six-step expert review workflow with routing examples.

The net effect: the memory to insight to knowledge loop, and expert synthesis of insights into durable knowledge, are now first-class, self-explaining parts of the agent contract rather than tribal knowledge.

Smooth upgrades for connected agents

The platform ships frequently, and an upgrade changes the tool contract. Previously a connected agent kept its stale tool list for the full shutdown grace period, because MCP runs over long-lived SSE and streamable-HTTP streams that never go idle, so the connection only dropped at process exit.

On shutdown the server now closes live MCP sessions (after the load-balancer drain delay and a brief settle for in-flight requests). Dropping the stream lets the client reconnect to the new build and re-fetch the tool list promptly. For Claude Code this is fully automatic: it auto-reconnects HTTP/SSE servers and re-handshakes against the new pod, so it picks up new tools and parameters on the next deploy with no user action.

The close is graceful and bounded: an idle session drops immediately, a session with an in-flight tool call is allowed to finish within the grace budget, and the drain can never outlive the pod's termination grace period (so a long-running call cannot cause a hard kill mid-call).

Operating notes

A new "Upgrades and connected agents" section in the deployment guide documents how each client picks up a changed tool contract:

• Claude Code: automatic on reconnect after a deploy.
• Claude Desktop: requires a full app restart (no in-session tool-list refresh today).
• claude.ai managed web connector: caches the schema at the connector level; needs a re-sync.

Keep tool and schema changes additive across releases: a client briefly on a cached contract safely ignores a new optional parameter, but renaming or removing a parameter or tool breaks it mid-session, so deprecate across a release before removing.

No database migrations in this release.

Full delta: v1.88.0...v1.88.1