Summary
Cross-source linking between providers (Vercel, GitHub, Sentry) is non-functional. Three independent layers are broken. Not blocking — we're not live with multi-provider yet — but needs fixing before launch.
Layer 1: Entity Pipeline — Edge Rules Are Dead Code
GitHub declares a commit → deployment edge rule to link commits to Vercel deployments. It never fires because:
- Vercel transformer puts commit SHA in
attributes.gitCommitSha instead of relations[]
- GitHub has no
push event transformer; PR transformer puts headSha in attributes too
- Text-extracted SHAs get
category: "reference" which the edge resolver filters out (only accepts structural types: commit, branch, pr, issue, deployment)
- The
extractFromRelations function already has a case "commit" path — it just never executes because no relation with entityType: "commit" is ever produced
Fix: Both Vercel and GitHub transformers need to emit { entityType: "commit" } in their relations array. ~20 lines across 2 files.
Layer 2: Resource Linking — Zero Cross-Provider Awareness
orgIntegrations rows are fully independent — no FK or field links a Vercel project to its GitHub repoproviderConfig JSONB stores only { provider, type, sync } — no linkedRepoId, no rootDirectorybulkLink accepts resourceName in input but never writes it to DB- Installed sources UI resolves display names via live API call on every page load (no cached name)
Fix: Extend providerConfig Zod schemas with optional linkedRepo field (no DB migration — it's JSONB). Fetch cross-provider metadata inline during bulkLink using createMemoryCaller (already imported). Add resourceName column (one migration).
Layer 3: API Catalog — Missing Endpoints
Provider APIs expose rich cross-linking data that isn't registered in the catalog:
- Vercel: No
get-project endpoint (returns link.repoId, rootDirectory for monorepo awareness)
- Sentry: No
list-repos, list-code-mappings, or list-releases endpoints
- GitHub: No
deployments or deployment-statuses endpoints
vercelProjectsListSchema uses .loose() which silently drops link and rootDirectory from list responses
Fix: Register missing endpoints as Zod-typed catalog entries. ~30 lines of schemas.
TL;DR
The data exists in every provider's API. We just don't fetch it (Layer 3), don't store it (Layer 2), and don't produce the right entity types for the graph rules that already exist (Layer 1). Total fix is ~70 lines across ~6 files. Do it when we're ready to go live with multi-provider connections.
Summary
Cross-source linking between providers (Vercel, GitHub, Sentry) is non-functional. Three independent layers are broken. Not blocking — we're not live with multi-provider yet — but needs fixing before launch.
Layer 1: Entity Pipeline — Edge Rules Are Dead Code
GitHub declares a
commit → deploymentedge rule to link commits to Vercel deployments. It never fires because:attributes.gitCommitShainstead ofrelations[]pushevent transformer; PR transformer putsheadShain attributes toocategory: "reference"which the edge resolver filters out (only accepts structural types: commit, branch, pr, issue, deployment)extractFromRelationsfunction already has acase "commit"path — it just never executes because no relation withentityType: "commit"is ever producedFix: Both Vercel and GitHub transformers need to emit
{ entityType: "commit" }in theirrelationsarray. ~20 lines across 2 files.Layer 2: Resource Linking — Zero Cross-Provider Awareness
orgIntegrationsrows are fully independent — no FK or field links a Vercel project to its GitHub repoproviderConfigJSONB stores only{ provider, type, sync }— nolinkedRepoId, norootDirectorybulkLinkacceptsresourceNamein input but never writes it to DBFix: Extend
providerConfigZod schemas with optionallinkedRepofield (no DB migration — it's JSONB). Fetch cross-provider metadata inline duringbulkLinkusingcreateMemoryCaller(already imported). AddresourceNamecolumn (one migration).Layer 3: API Catalog — Missing Endpoints
Provider APIs expose rich cross-linking data that isn't registered in the catalog:
get-projectendpoint (returnslink.repoId,rootDirectoryfor monorepo awareness)list-repos,list-code-mappings, orlist-releasesendpointsdeploymentsordeployment-statusesendpointsvercelProjectsListSchemauses.loose()which silently dropslinkandrootDirectoryfrom list responsesFix: Register missing endpoints as Zod-typed catalog entries. ~30 lines of schemas.
TL;DR
The data exists in every provider's API. We just don't fetch it (Layer 3), don't store it (Layer 2), and don't produce the right entity types for the graph rules that already exist (Layer 1). Total fix is ~70 lines across ~6 files. Do it when we're ready to go live with multi-provider connections.