Glossary¶
Scope: both products; product-specific definitions are identified below.
These definitions follow the current configuration, routes, and types in the source tree.
| Term | Definition |
|---|---|
| A2A | Agent-to-agent task protocol using JSON-RPC; the current implementation provides an inbound service. |
| Agent Loop | Processing loop from inbound event to response. |
| Approval | Review before a tool call, controlled by permissions.approval.mode. |
| Channel | Message adapter for endpoints such as Telegram, Discord, and CLI. |
| Checkpoint | Shadow Git snapshot of workspace files; excludes the database, sessions, memory, and logs, so it is not a full backup. |
| Clarification | Flow in which the agent asks the user for missing information. |
| Compression | Compaction of older session messages to control model-context length. |
| Cron Job | Job triggered by the scheduler. |
| Dashboard | Web interface using the Gateway API and /ws/dashboard. |
| Execution Backend | Backend selected by tools.exec.host for execution tools; sandbox copies the working directory but does not provide OS-level isolation. |
| Gateway | Service exposing HTTP APIs, WebSockets, and Dashboard assets. |
| Knowledge | Indexed documents for retrieval, stored separately from agent memory. |
| Memory | Persistent memory organized by type and scope, with retrieval and forgetting policies. |
| Migration | SQLite schema migration runs automatically during database initialization; echo-agent migrate only updates USER memory ownership and imports legacy MEMORY.*.md shards. |
| Multi-Agent | Task execution involving multiple collaborating agents. |
| OpenTelemetry | Optional trace and metric export framework; current code creates spans but not the formerly documented named business metric instruments. |
| Pairing | Gateway's short-lived code authorization for a platform/user identity; it does not issue an API token. |
| Plugin | Loadable Python extension that can register tools and lifecycle hooks; separate channel packages are not automatically wired. |
| Risk Category | Tool risk levels are READ_ONLY, WRITE, EXEC, and DANGEROUS. Names such as MINIMAL_TOOLS identify tool-profile sets, not risk levels. |
| Security Profile | Deployment profile: personal_cli, daemon, or public_gateway. |
| Session | Conversation context determined by channel, user, chat, and optional thread. |
| Skill | Knowledge or workflow described by SKILL.md; some candidate skills require approval. |
| Spill | Storage for oversized tool output, referenced in conversation and retrievable with read_spill. |
| Tool | Callable operation exposed to the agent; execution tools include exec, execute_code, and process. |
| Tools Profile | Tool admission set: minimal, messaging, coding, or full. |
| Workspace | Directory containing configuration, runtime data, and working files. |
Product and workflow terms¶
| Term | Definition |
|---|---|
| Desktop / Client | Local desktop workspace with an embedded Rust execution engine. |
| Standalone Runtime | Python Agent service used through terminals, channels, or Gateway. |
| ACP | Agent Client Protocol; Desktop uses typed in-process channels for host-engine interaction. |
| Host | The component owning UI, resource management, local services, and engine lifecycle. |
| Artifact | An actual task output file whose contents and saved location can be verified. |
| Expert group | Desktop collaboration with persistent members, work, messages, and decisions. |
| Provenance | The origin and evidence supporting a fact, memory, or result. |
| Acceptance criteria | Observable conditions used to decide whether a task is complete. |