Summary

• Validate endpoint — POST /v1/graphs/{id}/validate checks schema without executing. Returns 200 with valid:true or 422 with error details and node refs. Uses FakeListChatModel to avoid requiring LLM API keys during validation.
• Export stub — GET /v1/graphs/{id}/export returns 501 until Phase 5 implements code generation.
• Run history — GET /v1/graphs/{id}/runs and GET /v1/runs with pagination (limit/offset), Literal status filter, and owner isolation.
• Cancel — POST /v1/runs/{id}/cancel returns 202. Handles both live runs (via RunManager) and stale DB runs (server restart recovery).
• Delete — DELETE /v1/runs/{id} returns 204. Rejects active runs with 409.
• Migration 003 — Composite indexes on runs table for history query performance.
• Bug fix — cancel_run() now also sets resume_event, fixing a hang when cancelling paused runs.

Test plan

Unit tests (240 total, all passing)

• 9 new CRUD tests: pagination, offset, status filter, owner isolation, cross-graph listing, delete success/wrong-owner
• 7 new validate/export route tests: valid schema, missing start, unknown tool with node_ref, not found, wrong owner, export 501, export not found
• 14 new history/cancel/delete route tests: empty list, paginated, status filter, not found, wrong owner, global list, graph_id filter, owner exclusion, cancel running, cancel completed (409), cancel stale DB, cancel not found, delete completed, delete active (409)
• 3 updated migration tests: version bumped to 3, bad migration test uses VERSION=4

Manual tests (19-29, all passing)

Notable findings during testing

• cancel_run() only set cancel_event but not resume_event, causing paused runs to hang on cancel — fixed by setting both
• Tool-only graphs complete too fast to test cancel mid-run — tests use human_input graphs that stay paused
• POST validate requires Content-Type: application/json due to middleware enforcement — tests send json={}

🤖 Generated with Claude Code