NovaOS MCP
Field guide

Known NovaOS platform issues and gaps (Nov 2026)

Platform-side limitations and bugs observed in production apps, with workarounds. Check before reporting a "new" bug.

Runtime

  • A1 gateway_url missing from run context on some envs: use LLM_GATEWAY_URL env.
  • A2 dispatch failures are a generic 503 "Agent execution dispatch failed": retry once on a fresh thread.
  • A5 thread history stores only the first stream frame of a reply: keep your own history.
  • A6 30 s response-header window from pods: stream immediately.
  • A7 fleet pod mode does not serve streaming runs; execution_modes not in the publish schema.
  • A8 no shared warm pods for service-account runs; instances with config/secrets never use the warm pool.
  • A9 SA_EXECUTION_ENABLED="false" coerced to true on EU (service/runs behavior differs per env).
  • Streams cut ~300 s (all open streams were cut at once on 2026-10-11).
  • Prewarm on a chat thread replays the prefetch result as the next answer (ClickUp z95t6g7thr).
  • Multi-skill routing unreliable: one-skill dispatch is the workaround.
  • Pod reaper stopped 2026-10-25: retired pods can linger and hold quota.

MCP

  • Tools classified read/write by NAME, fail-closed; user runs refuse writes with -32042 even with SA grants or "Auto-confirm writes" (ClickUp z95t6g84f9). Workaround: returned writes.
  • An agent's tools/list shows every published tool in the org (not only declared dependencies).
  • ~1.2-1.5 s per MCP hop; sessions not reused by the gateway.
  • dependencies.mcp_servers stripped by CLI 2.2 manifest validation.

Knowledge / RAG

R1 query 504 at 30 s on S2S; R3 fixed query model; R4 sequential planner; R5 embedding model not validated at create; R6 flat RRF scores, no threshold; R8 planner drops token/model; R11 confidence floor 0.5; R12 7 MiB practical upload cap; R13 rerank maps by text (duplicates collapse); R15 weak keyword leg on codes; R16 non-atomic writes/deletes; R17 chunk sizes in characters; R18 silent lexical-only fallback; R19 duplicate upload -> new empty document; R24 streaming without sources; R25 tool errors as HTTP 200. XLSX truncated at 500 rows. Graph unqueryable after a new ingest until rebuilt. Collections shared across apps in an org.

Memory

M1-M6 (see 09-chat-memory-data): budgets ignored, no fact update, filters ignored, possible bugs, extraction outside the gateway. token_budget not enforced (ClickUp z95t6g7tx7). Memory 1.x lacks a proposals route.

Publish / install / secrets

  • requiredSecrets install fails on EU staging: SA_AGENT_INSTALL_FAILED (ClickUp z95t6g7xu2). Use config_schema.
  • No publisher-level secrets; secrets are per user and per instance; no API to delete one secret (rotation = reinstall).
  • Image scan / dependency audit "Not run" on FDE approvals.
  • novaos-cli init defaults to TypeScript whose SDK is unpublished; scaffold docs mention TestPyPI/--no-deps and ctx.tools.llm (stale).
  • mcp init Dockerfiles: curl health check on python:slim, root user, TS lockfile required.
  • sandbox needs a platform checkout's compose file.
  • No CLI command to install or run an agent (use marketplace API / this server).

Observability

L1/L4 agents cannot reach Langfuse; L2 gateway message logging off; L3 no run/session trace grouping; L8 platform evals unreliable; L9 logs only via novaos-cli agent logs; developer cost visibility limited (T11).