mcp --profile full
🛰️ Swarm/Hive Tool V&Vagent-driven tools over the encrypted daemon
🗄️ Data Tier V&VSQLite singleton + PostgreSQL·AGE·pgvector
🌐 Federation V&Vcertified enterprise federation — cross-host mTLS
native certified data-tier nodes · live health
certified data tier at cloud scale · live droplets
the certified data tier at the hub · each GLM-5.3-Flash NHI a node sized by its signed writes · links pulse on live writes over mTLS
ai-memory operations / second — live stream
every ai-memory data path encrypted in transit — agent→daemon (mTLS), daemon→PostgreSQL (verify-full), node↔node (mTLS federation)
agent-to-agent coordination primitives — live rate & mix
MCP-ready time — the #3329 non-blocking-boot guarantee (backfill never blocks init)
per-op p50 / p95 / p99 over the certified data tier
Every ai-memory data path is encrypted — customers encrypt all data in transit, so the test fabric does too. Full-spectrum latency across the 3 encrypted legs plus the OpenRouter API.
concurrent NHIs vs sustained throughput — the knee = agents one certified module handles
Ed25519-attested stores with synchronous remote embed (OpenRouter) — the honest write ceiling
what the runs cost in model activity — account snapshots per run; per-generation tokens via #3338
p50 / p95 / p99 against PERFORMANCE.md
% of the drivable tool surface exercised end-to-end
sustained agents as certified modules are added (federation)
adversarial "bad memories" injected vs infections that spread — red-team
kill cycles → recovered mission·goals·objectives, resumed exactly where it left off
Every defect the swarm uncovers is opened 1:1 by Fable, fixed by Codex hard-coders (Opus-class), then Fable code-reviews + security-reviews and merges into release/v1.0.0 — after which ai-memory is recompiled, redeployed to the test envs, and re-tested. GLM-5.3-Flash executes tests only — it never writes code, submits PRs, or touches the codebase.
Fleet-standard ai-memory configuration — byte-identical on every node & agent wired into ai-memory (only db and key-file paths differ per node). Everything hits OpenRouter; Ollama is not used on any node.
—