Memory
Sphere (second-brain-v3) W4P adaptation decisions (Aug 2026): prefix stripped once at entry (src/worker.ts), zero route-file edits; outbound fix = origin + first path segment across 12 frontend files; Browser Run code deleted (7 REST calls, BROWSER binding, CRAWL_API_TOKEN, CLOUDFLARE_ACCOUNT_ID). IngestionAgent DO (idFromName('ingestion')) orchestrates pipeline via ingestion service; DO alarms only, never cron; local queue, default concurrency 3, service 429+Retry-After as backpressure. Originals stored in customer's own DOCUMENTS R2 (r2_key kept — needed for downloads and the rebuild-vectors utility). Embedding model from env EMBEDDING_MODEL/EMBEDDING_DIMS injected per plan by platform, never hardcoded. cleanupOrphanedDocuments moves to a daily DO alarm. crawl_jobs kept as UI projection of service-side jobs. Bundle artifact = worker.js + manifest.json + migrations/*.sql; platform applies migrations via D1 API, expand-migrate-contract. INGESTION_CONTRACT revision 2: result = { chunks: [{ content, metadata? }] }, 10-min job timeout, 50MB file cap on the base64 path (128MB Worker memory: raw bytes + base64 string coexist at both hops).
Tags: kind:semantic, status:canonical — Source: mcp — 2026-08-20 11:28:08 UTC