mcp-data-platform-v1.88.0

mcp-data-platform v1.88.0

This release introduces canonical Knowledge Pages as a first-class layer of the platform and makes knowledge bidirectionally linked to the entities it describes. Memory, Insights, and Knowledge are now unified behind a single hub, and the agent-facing write path (apply_knowledge) promotes captured knowledge into the right place automatically.

Canonical Knowledge Pages

A new portal-native knowledge layer: durable, human-readable markdown pages that hold business and domain knowledge (vocabulary, definitions, runbooks, context) rather than per-row metadata.

• Pages are stored in the platform database, version-tracked (every save is a revision), and semantically indexed (768-dim embeddings, hybrid lexical + vector search) so the agent recalls them through the unified search tool.
• Pages render full markdown including Mermaid diagrams.
• Each page carries a feedback thread, so reviewers can comment on and improve canonical knowledge in place.

Unified Knowledge hub

Memory, Insights, and Knowledge are no longer three separate areas. They are unified under one Knowledge hub that teaches the lifecycle they share: a memory is captured, promoted to an insight proposed for review, and finally promoted to knowledge that is shared and canonical. Review and promotion are gated by apply_knowledge tool access rather than an admin role.

Entity references (bidirectional linking)

Knowledge pages, and the insights they are synthesized from, now reference the entities they are about, end to end. This closes the loop the platform's cross-enrichment thesis depends on: knowledge is no longer a dead end for entity linkage.

• A page (or insight) can reference any entity type: an asset, a prompt, a collection, a connection, a DataHub dataset/glossary term, or another knowledge page. References are stored as typed rows with referential integrity and delete-cascade, not as opaque strings.
• Inline references written in page or insight prose as mcp: / urn:li: links render as live chips showing the entity's current name (so renaming an asset updates the chip with no edit) and deep-link to the entity's viewer.
• A Related panel lists a page's references grouped by type; a manual-reference picker lets editors search assets, collections, prompts, and pages to add references by hand.
• Reverse lookup / backlinks: entity views surface "the knowledge pages that reference this," via an indexed reverse lookup.
• Page-to-page linking is supported, enabling incremental disclosure (an overview page that links to deeper pages).
• When a business_knowledge insight is promoted, all of its reference types are carried onto the page, not just DataHub URNs. Inline references in the insight's own prose are carried too.
• A one-time startup backfill re-derives references for pages that predate this feature, from their source insights, so existing knowledge gets its links without a manual re-save.
• References are access-filtered: a reference is only resolved and shown when the viewer (or agent) can access the target, and inaccessible or deleted targets never leak their existence. A deleted but accessible reference renders as a broken-reference chip rather than a dead link.

apply_knowledge: the unified write path / sink router

apply_knowledge is now a sink router for promoting captured knowledge. It classifies each promotion by its sink class and routes it to the correct destination: schema_entity knowledge goes to DataHub, business_knowledge becomes a canonical knowledge page, and entity references are preserved through the promotion (and unioned across repeated promotions to the same page). Rolling back a promotion restores the page's prior state, including its references, while leaving manually-authored references intact.

Feedback on knowledge pages

Knowledge pages participate in the feedback/review loop: a page carries a thread bridged into apply_knowledge, so feedback on canonical knowledge flows back into the capture-to-promotion cycle.

Terminology: "enrichment", not "injection"

The cross-service context feature is now consistently called cross-enrichment throughout the codebase, APIs, and docs (previously "injection", which carried a misleading security connotation). Genuine security terms (SQL injection, prompt injection) are unchanged. This is a naming and concept clarification with no behavioral change.

Fixes and improvements

• Search: entity_urns lookups now fan out across every entity-keyed source, so an entity-scoped search returns matches from all sources that key on it rather than a single one.
• Portal: thumbnail prose CSS is scoped per color scheme, so a dark code background no longer bleeds into light-scheme captures.
• Dev: make dev preflight now names the actual process holding a contended port.

Upgrade notes

This release adds database migrations 000070-000075 (knowledge pages, page versions, page feedback threads, entity-reference table, reverse-lookup indexes, and the backfill sentinel). They run automatically on startup when a database is configured. The reference backfill runs once in the background at first startup and is idempotent; it never blocks boot.

Full delta: v1.87.1...v1.88.0