Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 2 min read

Context Hydration: Abstracting Merge Request Metadata for LLM Consumption

The Hook Feed a raw, unprojected enterprise GraphQL Merge Request graph directly into an autonomous coding agent's context window, and you will watch 40% of your token budget vanish into deeply nested pagination edges,

The Hook

Feed a raw, unprojected enterprise GraphQL Merge Request graph directly into an autonomous coding agent's context window, and you will watch 40% of your token budget vanish into deeply nested pagination edges, redundant color hex codes, and internal global IDs before the model ever reasons about a single line of diff.

The Deep Dive

In enterprise AI tooling, context hydration is the architectural boundary between how a relational database or GraphQL schema stores domain entities and how an LLM consumes them via the Model Context Protocol (MCP). A human browser client rendering a Merge Request dashboard needs UI-specific presentation metadataβ€”avatar URLs, label background colors, text contrast ratios, and GraphQL __typename nodes. An autonomous DevSecOps agent triaging a review queue, by contrast, requires high-entropy operational state: who is assigned to review (reviewers), semantic classification (labels), readiness gates (draft, detailedMergeStatus), temporal freshness (createdAt, updatedAt), and exact cryptographic commit state (diffHeadSha).

While authoring MR !270521 (feat(mcp): add merge request read fields and reviewer filter) to resolve Issue #639870 in gitlab-org/gitlab, I engineered the context hydration layer for GitLab's list_merge_requests and get_merge_request MCP tools. Exposing reviewers, labels, and merge-status fields across both GraphQL-backed (ListMergeRequestsService) and REST-backed (get_merge_request) tool pipelines revealed a severe token-economics bottleneck. When querying 100 merge requests on gitlab-org/gitlab, projecting full label objects (id, title, color, description, textColor) inflated the serialized JSON response from 80 KB to 132 KBβ€”an extra 52 KB (~13,000 tokens) of pure structural noise on every agent turn.

To solve this without violating GitLab's strict GraphQL linter (@graphql-eslint/require-selections), I restricted the GraphQL selection set in app/graphql/queries/mcp/merge_requests/list_merge_requests.query.graphql to the minimal scalar leaf labels { nodes { title } } and implemented a post-query hydration transform in format_merge_request that collapses the nested { "nodes": [{ "title": "..." }] } graph into a flat scalar array of label strings (["backend", "type::feature"]). Combined with batch-preloaded associations verified via ActiveRecord::QueryRecorder across both CE (FOSS_ONLY=1) and EE specs, this reduced the 100-MR payload back down to 80 KBβ€”cutting context-window consumption by ~40% while delivering zero-N+1 database execution.

The GitLab Hero Angle

GitLab's MCP server (MR !270521) treats context hydration as a first-class systems disciplineβ€”abstracting complex GraphQL and REST merge request graphs into token-dense, N+1-free payloads so GitLab Duo and MCP agents can reason over 100+ enterprise MRs in a single pass.

The Takeaway

Never treat an MCP tool as a raw pass-through proxy for your UI GraphQL schema; project only decision-critical scalars (detailedMergeStatus, diffHeadSha, reviewers) and flatten nested connection nodes into compact scalar arrays before hydrating the LLM context window.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.