Multi-graph: follow-up items from PR #156

[charlie83Gs]

Context

PR #156 introduced the multi-graph foundation (schema isolation, RBAC, session resolver, sync fan-out, MCP graph scoping, frontend graph picker). This issue tracks remaining work needed before the feature is production-ready.

Backend: Graph-Scoped Endpoints

The frontend now routes all graph-scoped data calls through graphRequest() which prefixes /graphs/{slug}/... for non-default graphs. However, only /graphs/{slug}/nodes has a backend implementation (graph_nodes.py). All other graph-scoped resource endpoints need to be added:

• /graphs/{slug}/nodes/{id}/dimensions
• /graphs/{slug}/nodes/{id}/facts
• /graphs/{slug}/nodes/{id}/edges
• /graphs/{slug}/nodes/{id}/history
• /graphs/{slug}/nodes/{id}/convergence
• /graphs/{slug}/facts
• /graphs/{slug}/edges
• /graphs/{slug}/sources
• /graphs/{slug}/seeds
• /graphs/{slug}/edge-candidates
• /graphs/{slug}/conversations
• /graphs/{slug}/syntheses
• /graphs/{slug}/research

Strategy: Refactor existing endpoint handlers to accept session as a parameter, then create thin graph-scoped wrappers that use GraphContext for session routing. Same handler logic, different session source.

Backend: Worker Graph-ID Propagation

Workers now resolve graph_id to per-graph session factories, but the API dispatch points don't include graph_id in the workflow input yet. When dispatching from graph-scoped endpoints:

• Include graph_id in workflow input dict for all dispatch calls from graph-scoped API endpoints
• Verify child workflow dispatch propagates graph_id (e.g., bottom_up_wf → node_pipeline_wf)

MCP: Per-User Graph Membership Check

MCP tools check OAuth token graph:{slug} scopes but don't verify the user's GraphMember membership. A user with an unrestricted token (no graph scopes) can access any graph by slug. Needs:

• Thread authenticated user through MCP tool context
• Check GraphMember role for non-default graphs in _get_graph_factory

Frontend: Graph Picker UX

• Show loading/error state when switching graphs fails
• Add graph indicator to page header showing active graph name
• Handle 404s from graph-scoped endpoints gracefully (show "endpoint not yet available for non-default graphs" instead of generic error)

Operational

• Startup cleanup for graphs stuck in provisioning status (crash recovery)
• list_graphs node_count — currently returns 0 for all graphs (requires cross-schema count query per graph)
• Qdrant collection auto-recovery for non-default graphs (currently only ensured during provisioning, not worker startup)
• Integration tests for: provisioning flow, schema isolation, multi-graph sync, RBAC enforcement

Architecture Notes

• graphRequest() is wired but will 404 on non-default graphs until backend endpoints are added — this is expected and documented
• GraphAwareMixin is on all Hatchet input models — workers resolve sessions but API dispatch needs to populate graph_id
• Schema-mode graph pool sizes are configurable via graph_pool_size / graph_max_overflow settings
• Sync fan-out (sync_dispatch_wf + sync_graph_wf) prevents cross-graph starvation

🤖 Generated with Claude Code

[charlie83Gs]

Status update (2026-04-06)

Completed

• Foundation (PR feat: multi-graph support with schema isolation and RBAC #156) — schema isolation, RBAC, session resolver, sync fan-out, MCP graph scoping, frontend graph picker
• Worker graph_id propagation (PR feat(api,workers): graph_id propagation for multi-graph workflows #164) — dispatch_with_graph helper, all 16 API dispatch points wired, child workflow propagation (GatherFactsPipeline → seed_dedup_batch, bottom-up scope → workflow)
• Alembic head merge (PR fix(kt-db): merge alembic heads after multi-graph PR #163)

Remaining

Backend — graph-scoped endpoints (highest priority):
Only /graphs/{slug}/nodes exists. The remaining ~13 endpoints still 404 for non-default graphs:

• Node sub-resources: dimensions, facts, edges, history, convergence
• Top-level resources: facts, edges, sources, seeds, edge-candidates, conversations, syntheses, research

MCP — per-user graph membership check:
Token scope check exists but GraphMember role verification is missing.

Frontend UX:

• Loading/error state for graph switching
• Active graph indicator in header
• Graceful 404 handling for unavailable graph-scoped endpoints

Operational:

• Crash recovery for provisioning-stuck graphs
• list_graphs node_count (currently always 0)
• Qdrant collection auto-recovery for non-default graphs
• Integration tests for provisioning, schema isolation, multi-graph sync, RBAC

Summary

Infrastructure works and default-graph behavior is unaffected. Non-default graphs are not yet usable end-to-end — the graph-scoped endpoints are the critical blocker.

[charlie83Gs]

Final status — all backend items complete

Completed (PR #166)

Backend — graph-scoped endpoints:
All 13 resource groups now have graph-scoped versions at /graphs/{slug}/...:

• ✅ Node sub-resources: dimensions, facts, edges, history, convergence, perspectives, children
• ✅ Top-level: facts, edges, sources, seeds, edge-candidates, conversations, syntheses
• ✅ Research: prepare, confirm, decompose, proposals, build, bottom-up, agent-select, summary
• ✅ All endpoints use _impl extraction pattern — zero logic duplication between default and graph-scoped routers

Backend — worker graph_id propagation:

• ✅ All workflow dispatches include graph_id (PR feat(api,workers): graph_id propagation for multi-graph workflows #164 + PR feat: multi-graph follow-up — graph-scoped endpoints, MCP membership, UX, ops #166)
• ✅ Super-synthesis sub_configs include graph_id for child workflows

MCP — per-user graph membership check:

• ✅ user_id and is_superuser threaded through token claims
• ✅ _get_graph_factory verifies GraphMember role for non-default graphs
• ✅ Fails closed — missing user_id denies access

Frontend:

• ✅ GraphProvider with error, switchGeneration state
• ✅ GraphPicker with loading spinner, active graph indicator
• ✅ graphRequest() transparent error passthrough

Operational:

• ✅ Crash recovery for provisioning-stuck graphs (API startup)
• ✅ list_graphs returns real node_count (concurrent cross-schema queries)
• ✅ Qdrant auto-recovery for per-graph collections (worker startup)

Remaining — tracked in new issues

• Graph management UI — Graph management UI — admin page, improved picker, org-style switcher #168 (admin page improvements, org-style picker, delete confirmation, member invite by email)
• Integration tests — need running infra, will track separately

🤖 Generated with Claude Code

[charlie83Gs]

All backend items implemented in PR #166. UI improvements tracked in #168.