Noolog Docs

Quorum SDK

latest documentation.

Architecture

  • Architecture & NATS — How NSED uses NATS JetStream for orchestration, task distribution, and state.
  • Agent internals — The quorum-rs crate's library API and agent architecture for Rust developers.
  • Rooms & policies — Why a policy (who deliberates and how) is orthogonal to a room (who can watch).
  • Policy & sessions — Why the interactive client treats policy as the model and the session as the thread.
  • Provider registry — Why providers are a registry of factories a third-party crate can extend without forking.
  • Middleware — The pluggable pipeline for validation, transformation, and moderation at agent lifecycle points.
  • Persona layer stacking — Why personas stack in layers and why md paths resolve against process CWD.
  • Working dir override — How a before_prompt middleware picks the directory an agent's subprocess runs in.
  • read_file sandbox — Why HTTP-only agents get a sandboxed, read-only file tool scoped to operator roots.
  • Claim citations — How an evaluator's claim quote is grounded back to the exact proposal span it came from.
  • Compact history — Why agents fold older tool results into a summary to survive context-window pressure.
  • Context guards — Why the shrink-guard counts tool schemas alongside messages when fitting a call.
  • Agent ranking — How agents declare capabilities and earn competitive rank via floor-based promotion.
  • Telemetry design — Why agent telemetry is metrics-only, fire-and-forget, with deterministic trace ids.
  • Thread TUI roadmap — Roadmap for an email-style, branching, newest-on-top conversation thread view.
  • About provider-reported cost — Why the SDK carries the charge a provider reported alongside its own estimate, and why that figure is optional.
  • About model-down benching

Tutorials

How-to

  • Redeem invite — Bootstrap scoped NATS credentials by redeeming a single-use invite code on the agent host.
  • Register a device — Mint an operator token from a device key pair — self-serve, no invite code or admin needed.
  • Run an agent fleet — Boot a whole agent fleet from a single quorum.yml with one quorum serve command.
  • Smoke-test an agent — Verify one of your own agents in three escalating stages — chat, tool-calling, then full NSED.
  • Use the thread TUI — Hold a branching, email-style conversation whose replies are quorum deliberations.
  • Use Claude Code plugin — Expose every quorum lifecycle command as a slash command inside Claude Code.
  • Compose persona — Build an agent persona from reusable stacked layers of shared markdown prompt files.
  • Register a provider — Teach quorum serve a new provider.type via a ProviderFactory — no fork required.
  • Develop own agent — Build a deliberation agent in Rust and connect it to a quorum-rs orchestrator.
  • Limit job concurrency — Cap an agent to N concurrent jobs so tasks sharing mutable state can't race and corrupt it.
  • Expose dashboard (LAN) — Reach the per-agent control-plane dashboard from other hosts, and lock it down safely.
  • Dump prompts — Capture each agent's exact model payload to hunt prompt bloat, with no config change or rebuild.
  • Opt out telemetry — Disable all telemetry emission from your agent with a single config flag.
  • Verify telemetry — Subscribe to your agent's telemetry subtree to inspect the events it actually publishes.
  • Inspect telemetry (NATS CLI) — Live-tail and filter agent telemetry events straight from the NATS CLI, no forwarder needed.

Reference

  • Glossary — Key terms used across the NSED documentation and CLI output.
  • Chat Completions API — OpenAI-compatible /v1/chat/completions endpoint mapping model tags to NSED policies.
  • Exec agent protocol — Bridge protocol for external stdin/stdout subprocesses to act as NSED deliberation agents.
  • MCP agent protocol — MCP protocol giving external agents full tool access during NSED deliberation.
  • Telemetry catalog — Every event type published by quorum-rs telemetry, with fields and redaction rules.
  • Sandboxed tools — Sandboxed read_file and grep_search tools available to non-Claude agents.
  • Dashboard config — The knobs controlling whether the agent dashboard starts, where it binds, and how it authenticates.
  • openrouter config — Fields of the openrouter agent-config block for provider routing, ZDR, and web search.
  • persona yaml shapes — The two shapes the persona field on an agent entry accepts.
  • Task SKIP Protocol — How an agent deliberately sits out a task as an explicit answer instead of silence.
  • Thread TUI ref — Keys, line formats, and persisted data model of the quorum thread TUI.
  • Proposal Structure Metadata — The measured shape evaluators see on each candidate, and the tools that address it by line.
  • Sampling Parameters — Which request fields are sent, when they are omitted, and how to talk to a backend that fixes them server-side.
  • Provider-Executed Tools — Tools the backend runs itself, where our job is to hand the arguments back rather than do the work.
  • Task Clock — The date a task was issued, and why it is stated to the model rather than offered as a tool.
  • Model Health — How an agent decides its own model is unusable, and why only one kind of answer benches it.
  • Agent Identity Keys — The keys an agent declares, the signatures an operator attaches to a held response, and what is carried but not yet checked.