{"@context":"https://schema.org","@type":"CreativeWork","@id":"https://forgecascade.org/public/capsules/68bd2179-e698-4ff8-a67a-87e81ca7077f","name":"Forge production hardening report April 26 2026","text":"On April 26 2026 Forge production was rebuilt and hardened after testing the MCP graph tools and the local inference path.\n\nThe MCP contradiction lookup now returns valid JSON instead of failing on graph timestamp values. PageRank now accepts the MCP requested count and falls back to a Cypher approximation when the Neo4j graph data science plugin is unavailable.\n\nThe production API image was rebuilt on Python 3.12 with the optional DeepLake and HiveMind requirement files installed. The image keeps persistent local data directories for both optional systems.\n\nOlares One is reachable from the production API container and is enabled for the self upgrade proposal path with one request at a time. A test showed that sending all routine robot language model traffic to Olares caused timeouts, so production now uses a safer split. Routine model calls stay on the smaller local Ollama model, while Olares handles heavier self upgrade reasoning with a strict concurrency cap. Local Ollama failures no longer fall through to cloud providers unless an operator explicitly enables that behavior.\n\nMarketplace backfill was fixed to connect to Neo4j before querying and to reuse that connected database client inside the marketplace service. It now truncates very long capsule titles before creating listings. After deployment the backfill found 94 unlisted capsules, created 94 listings, and failed 0.\n\nVerification results after deployment were healthy. The public API health endpoint returned healthy. Forge MCP health returned healthy for the database, schema, event system, overlay manager, circuit breakers, and anomaly detection. Neo4j reported 222 online indexes out of 222. The MCP contradiction query returned 2 unresolved contradictions with serialized timestamps. The MCP PageRank query returned exactly 10 ranked capsules using the Cypher backend in about 62 milliseconds. Local targeted tests passed with 35 passing tests.\n\nRemaining follow up work is mainly load tuning. Neo4j still consum","keywords":["forge","production","mcp","olares","deeplake","hivemind","ops"],"about":[{"@type":"Thing","name":"Metador"},{"@type":"Thing","name":"TEMP.Veles"},{"@type":"Thing","name":"Python"},{"@type":"Thing","name":"Container CLI/API"},{"@type":"Thing","name":"Python Startup Hooks"},{"@type":"Thing","name":"Small Sieve"},{"@type":"Thing","name":"IronNetInjector"},{"@type":"Thing","name":"Empire"}],"citation":[],"isPartOf":{"@type":"Dataset","name":"Forge Cascade Knowledge Graph","url":"https://forgecascade.org"},"publisher":{"@type":"Organization","name":"Forge Cascade","url":"https://forgecascade.org"}}