Skip to content

Glossary

Terms used across Cadence docs and the codebase.

Intended audience: Stakeholders, Business analysts, Solution architects, Developers, Testers

Learning outcomes by role

Stakeholders

  • Look up terms quickly when reviewing decks or contracts referencing Cadence.

Business analysts

  • Align vocabulary across requirements and support documentation.

Solution architects

  • Disambiguate integration jargon when comparing to other platforms.

Developers

  • Navigate to deep dives from glossary entries.

Testers

  • Build a shared lexicon for test case titles and defects.

Alphabetical reference for jargon used in this documentation. For behavior detail, follow the links to Concepts and Features. When a page cites internal module paths, those refer to the Python API service tree under src/cadence/. Documentation audiences describes how docs are structured and how optional repository maps are versioned.

One configured agent runtime for an org: framework, mode, plugins, and pool tier—identified by instance_id. In API paths and older code this is still named orchestrator (for example GET /api/orgs/{org_id}/orchestrators/{instance_id}). See AI Apps.

User-facing name for a plugin ZIP attached to an AI App: uploaded to an org or system catalog, referenced via active_plugin_ids, and loaded into the pool at runtime. The implementation artifact is still a plugin in APIs and code (/api/.../plugins, BasePlugin). See AI Agent system.

Long-lived automation secret stored hashed in PostgreSQL. Callers send it in X-API-KEY; effective access follows the key’s scopes and the owning user’s org memberships. See API keys.

The first node in the grounded mode graph. Loads anchor context for a specific resource_id before the router classifies intent. See Engine internals.

Platform abstraction for a stable routing entry (alias) to an AI App configuration, with optional public visibility—see Central Points.

SHA-256 digest of an orchestrator’s effective configuration. Used by the pool to detect stale loaded instances and deduplicate reload events. See Orchestration backends.

TTL-based cache for on-demand orchestrator instances. Instances are loaded from the database on first chat request and evicted after an idle timeout. See Hot-reload and AI App pool.

Stable PREFIX-NNNN identifier in API and engine error payloads (code in JSON). See Error codes.

RabbitMQ topic exchange (cadence.orchestrators) used for multi-node coordination of orchestrator lifecycle, settings changes, and plugin distribution. See Event system.

TypedDict that carries data between LangGraph nodes during graph execution. Contains message history, agent hops, tool rounds, routing decisions, and error state. See Engine internals.

Controls how eagerly an AI App stays in the process pool—hot instances load at startup; demand instances use the TTL-backed demand pool. See Hot-reload and AI App pool.

Permanent resident orchestrator instances loaded at startup. Hot-tier instances have the lowest latency because they are always in memory. See Hot-reload and AI App pool.

Access tokens for interactive users are JWTs. The claim jti (JWT ID) is the Redis key used to load the live session; revoking Redis invalidates the token even if the JWT string still verifies. See JWT sessions.

Per-node LLM model override settings within mode_config. Allows using different models for different graph nodes (e.g., a cheap model for routing, a powerful model for synthesis). See Engine internals.

An organization—the unit of isolation. Data and AI Apps are scoped by org_id. See Multi-tenancy.

Settings cascade flag. When false on a global setting, orgs cannot override it. When false on an org setting, instance config cannot override it. See Settings cascade.

Technical name for the ZIP package format: a BasePlugin subclass that supplies tools to the agent graph. In product documentation this is often called an AI agent; REST paths and code still use plugin (for example /api/orgs/{org_id}/plugins). See AI Agent system and AI Agent SDK.

Loaded plugin package containing tools, agents, and settings. Bundles are cached and loaded by the SDKPluginManager at orchestrator build time. See Orchestrator load and plugin settings.

Role-based access control: roles carry permission strings (cadence:…) evaluated per org. See Role-based access control.

Platform superuser flag on the session; bypasses normal org checks where implemented. See Security and access.

Four-tier configuration resolution: system defaults (Tier 1) → global settings (Tier 2) → org settings (Tier 3) → instance config (Tier 4). See Settings cascade.

Redis-backed view of the caller (interactive or API key): user id, permissions, org memberships, and related fields. See JWT sessions and Security and access.

Gathers all plugin tools at orchestrator build time. Deduplicates tool names and appends a synthetic call_synthesizer tool for the planner. See Engine internals.

HTTP header selecting which organization a request targets when the path does not embed {org_id}. See Multi-tenancy.