Appendix A — Glossary

Purpose

This appendix is the canonical glossary. Every term used in the lab's documentation is defined here. The definitions are precise; when a term is used in a specific way on a specific page, that page is the authoritative reference.

The glossary is alphabetical. The placeholder definitions are at the end.

A

Adapter. A module that translates between a generic interface and a provider-specific format. Adapters appear in the research framework and the media tooling. See External Providers → Adapter contract.

ADR (Architecture Decision Record). A short document that captures a significant decision, the context in which it was made, and the consequences. The lab has 13 ADRs. See Architecture Decisions.

Agent. A specialized role with its own tools and its own lifecycle. The lab has 5 agents: Coordinator, Coding Assistant, Research Worker, Media Agent, Web Agent. See Agents and Roles.

API key. A credential used to authenticate with an external provider. The lab stores API keys in configuration files with owner-only permissions.

Artifact. A typed file produced by the research framework. Artifacts are validated by Pydantic schemas. See Artifacts.

Audit. A systematic review of the lab's state. The output is a report. Audits are performed weekly, monthly, quarterly, and ad-hoc. See Audit Model.

Audit trail. The complete record of the lab's activity. The trail is composed of the request journal, the mission states, the audit reports, and the configuration changes.

B

Backend. A playback target for the Media Agent. The lab supports three backends: Chromecast, VLC, and direct. See Media Playback.

Backup. A copy of the lab's data for disaster recovery. Backups are performed daily. The destination is the user's choice.

Browser automation. The capability of driving a browser through the OpenClaw browser tool. The capability is exposed through the Web Agent.

C

Candidate. A potential match for a research query. Candidates are produced by the F1 phase of the research framework. See Mission DAG → F1.

Cast protocol. Google's protocol for controlling Chromecast devices. The lab uses the protocol through the chromecast tool.

CDP (Chrome DevTools Protocol). The protocol used to control Chromium. The lab uses the protocol on port 18800. See Web Agent.

Checkpoint. A persistent state record of a mission. Checkpoints are written after every phase and after every domain tool step. See Mission Lifecycle → Checkpointing.

Chromecast. A Cast-capable display on the LAN. The Media Agent can discover and control Chromecast devices.

Coding Assistant. The lab's implementation arm. The Assistant delegates code-writing to a sandboxed sub-agent. See Coding Assistant.

Coding sub-agent. The external command-line tool (Codex, in the lab's case) that the Coding Assistant delegates to. The sub-agent is sandboxed.

Configuration. The settings that control the lab's behavior. Configuration is in files, not in code. See System Architecture → Configuration files.

Context. The information available to an agent at a given time. The context includes the session state, the memory, and the skill instructions.

Coordinator. The main control agent. The user's primary point of contact. See Coordinator.

Coverage. The fraction of the code that is exercised by tests. The lab's target is 70% minimum per project.

CWD (current working directory). The directory a process is currently in. The Coordinator's CWD is typically the workspace root.

D

DAG (Directed Acyclic Graph). A data structure that represents the dependencies between tasks. The research framework's phases are organized in a DAG. See Mission DAG.

Daily log. A markdown file (memory/YYYY-MM-DD.md) that records the day's activity. Daily logs are append-only. See Memory and Context.

Direct player. The Media Agent's simplest playback backend. Spawns the system's default media player. Used rarely.

Dispatch. The act of routing a request to an agent. The Coordinator is the only dispatcher.

Domain tool. A research framework tool that performs a specific domain task (hotel research, review research, etc.). The lab has 8 domain tools.

E

End-to-end test. A test that exercises the complete flow from input to output. The lab's research framework has end-to-end tests for each domain tool.

Environment variable. A configuration value passed to a process through the OS. The lab uses environment variables for sensitive configuration (secrets, paths).

Executor. The component that runs a phase or a tool. The orchestrator's executor is the PhaseRunner.

External provider. A service outside the lab that the lab talks to (model provider, search provider, streaming service).

F

F0, F1, F2, F4, F6. Phases of the research framework. F0 initializes the mission, F1 gathers candidates, F2 aggregates reviews, F4 computes prices, F6 writes the report.

Failure catalog. The list of known failure modes and their recovery procedures. See Failure Catalog.

Framework. A domain-specific orchestrator. The lab has two: the Research Framework and the Coding Sub-agent. See Framework.

G

Gateway. A long-running process that accepts requests and dispatches them to agents. The lab has two gateways: the Coordinator (port 18789) and the Scout (port 18790).

Git. The version control system. The Coding Assistant uses Git through the user's shell, but never commits, pushes, or pulls.

H

Health check. A verification that a component is working correctly. Health checks fall into three categories: liveness, correctness, performance. See Health Checks.

Heartbeat. A periodic probe that verifies the gateways and the in-flight missions are healthy. The default interval is 30 minutes.

HTTP API. The interface exposed by the gateways. The Coordinator's API is on port 18789; the Scout's API is on port 18790.

I

Incident. An unexpected event that requires intervention. Incidents are recorded in post-mortem documents under security/incidents/.

Intent. The user's goal, as resolved by the Coordinator from the natural-language request. The intent is mapped to one or more agents.

Invocation. A single call to a tool or a sub-agent. The Coding Assistant's invocations are tracked in the telemetry.

J

Journal. The append-only log of external calls. The journal is the foundation of the lab's audit trail for external API usage. See Request Journal.

K

Keyword. A word or phrase that the Coordinator uses to resolve the user's intent. The keywords are configured per agent.

L

LaunchAgent (macOS) / systemd unit (Linux). The supervisor that keeps a process running. The Coordinator and the Scout are supervised by launchd (macOS) or systemd (Linux).

Layer. A level in the lab's 8-layer architecture. Each layer has a single responsibility. See Framework.

Lifespan. The duration a component runs. The gateways are long-running; the tool processes are per-call; the sub-agent is per-invocation.

Liveness probe. A simple check that verifies a component is up. The lab has liveness probes for every component.

Logger. The component that writes log messages. The lab uses structured logging (JSON Lines) for the runtime logs.

M

Main session. The user's primary session. The MEMORY.md file is loaded only in the main session. See Memory and Context.

Media Agent. The lab's entertainment specialist. The agent handles media search and playback. See Media Agent.

Media Lab Server. The local HTTP server that exposes image and music generation. The server runs on 127.0.0.1:8765. See Media Lab Server.

Memory. The lab's persistent knowledge. The memory is organized in three layers: long-term (MEMORY.md), daily logs, and session state. See Memory and Context.

Mission. A research task. A mission is identified by a mission_id and tracked in phase.json. See Mission Lifecycle.

Mission DAG. The directed acyclic graph of phases. The DAG defines the dependencies and the parallel groups. See Mission DAG.

M3. The current model. Replaced M2.7 in June 2026. The lab uses M3 as the default and M2.7-highspeed as the fast variant.

N

NL (natural language). The user's input language. The Coordinator parses natural language and routes to agents.

Non-goal. An explicit statement of what an agent does NOT do. The non-goals are the boundaries that prevent an agent from drifting outside its role. See Agents and Roles.

O

Orchestrator. The component that drives a phase DAG. The research framework's orchestrator is in scout/orchestrator/.

Output directory. The directory where the Media Lab Server writes its outputs. Default location: ~/.openclaw/media/outbound/.

P

P3. The optional phase of the research framework that uses the Perplexity adapter for synthesis. P3 is opt-in because it costs API credits.

Parallel group. A set of phases that can run concurrently. F2, F4, and Y1 are in parallel group 1.

ParsedQuery. The structured representation of a research query. Produced by F0. See Artifacts → task.json.

Phase. A single step in a research mission. The lab has 8 phases: F0, F1, F2, F4, P3, Y1, Y2, F6.

Phase status. The lifecycle status of a phase. The lab has 9 phase statuses. See Mission DAG → Failure Status.

Playwright. The library that drives the browser. The lab uses Playwright through the OpenClaw browser tool.

Pool. The token pool for paid APIs. The pool is configured in token_pool_config.json and writes its state to token_pool_state.json. See External Providers → Token pool.

Post-mortem. A document that records an incident. The post-mortem is stored under security/incidents/.

Progressive prompt. The technique of splitting a large task into smaller prompts. The Coding Assistant uses progressive prompts to avoid the sub-agent being killed by the OS.

Q

Quorum. The minimum number of adapters that must succeed for a phase to be considered succeeded_partial. The default is "at least one".

Quota. The rate limit imposed by an external provider. The lab tracks the quota and waits for the reset when it is exhausted.

R

Rate limit. The maximum number of requests a provider allows per unit of time. The lab respects rate limits through the adapter's per-minute cap.

Report. The final output of a research mission. The report is a markdown file that aggregates all the artifacts.

Request journal. The append-only log of external calls. See Request Journal.

Research Framework. The lab's research orchestrator. See Research Framework.

Research Worker. The lab's research specialist. See Research Worker.

Resume. The act of continuing a paused mission. The orchestrator loads the state from phase.json and picks up where it left off.

Retry policy. The rules for retrying a failed operation. The lab's default is 3 attempts with exponential backoff.

Root cause. The underlying cause of an incident. The post-mortem identifies the root cause.

Router. The Coordinator's component that dispatches requests to agents.

S

Sandbox. A restricted environment that limits what a process can do. The coding sub-agent runs in a sandbox. The sandbox has three levels: read-only, workspace-write, full.

Scout. The gateway process that hosts the Research Worker and the Media Agent. Runs on port 18790.

Session. A continuous conversation between the user and the Coordinator. Sessions have a timeout and are stored in the session store.

Skill. A markdown file (SKILL.md) that describes a capability. The Coordinator activates a skill when the request matches the skill's trigger phrases.

SKILL.md. The file that defines a skill. See Tool Spec.

SLA (Service Level Agreement). A commitment about the system's performance. The lab's SLAs are documented in Health Checks.

Snapshot. The accessibility tree of a web page. The Web Agent uses snapshots for actions.

Source. A reference to a URL with metadata. Sources are produced by adapters and aggregated in artifacts.

State. The data the lab persists. The state includes the long-term memory, the daily logs, the session state, the skill artifacts, the research artifacts, and the audit trail.

State machine. A model of a component's lifecycle. The mission has a state machine; the phases have a state machine.

Strategy. The algorithm used by the token pool to select a token. The lab supports round-robin, weighted-round-robin, least-used, and failover.

Sub-agent. The external coding CLI that the Coding Assistant delegates to. The sub-agent is sandboxed.

Systemd unit. The Linux supervisor that keeps a process running. The lab uses systemd on Linux and launchd on macOS.

T

Task. The input to a research mission. A Task is a Pydantic model with a mission_id, a raw_query, and a parsed structure.

Telemetry. The metrics, logs, and traces the lab produces. See Telemetry.

Template. A pre-configured project structure that a new project can be copied from. The lab has 4 templates: Next.js, CLI Python, research framework, agent skill. See Project Templates.

Token. A credential for a paid API. The lab uses a token pool for paid APIs.

Token pool. The mechanism that manages multiple API tokens. See Token pool.

Tool. A typed, deterministic capability. The lab has dozens of tools, organized by category.

Tooling Layer. The layer that wraps deterministic capabilities. See Tooling Layer.

TOOL.md. The file that defines a tool's specification. See Tool Spec.

Trust boundary. The line between trusted and untrusted code. The lab has multiple trust boundaries. See Security Principles → Trust boundaries.

Trust level. The sandbox level for the coding sub-agent. The lab's default is workspace-write.

TTL (Time to Live). The duration a value is valid. The lab uses TTLs for caches and rotated files.

Typed artifact. A file that is validated by a schema. The research framework writes typed artifacts.

U

Unit test. A test that exercises a single function or module. The lab's projects have unit tests for every module.

User authorization. The user's explicit approval for a high-risk operation. The Coordinator surfaces the authorization request and waits for the response.

V

Validation. The act of checking a value against a schema. The lab validates every typed artifact on read and on write.

VLC. A media player that the lab uses as a playback backend. VLC is controlled through its RC (TCP) or HTTP interface. See Media Playback → VLC.

W

Web Agent. The lab's browser automation capability. The agent drives a Chromium-based browser through the OpenClaw browser tool. See Web Agent.

Workspace. The root directory of the lab. The workspace contains the agent state, the project directories, and the configuration.

Y

Y1, Y2. Phases of the research framework that handle YouTube. Y1 searches for videos; Y2 extracts transcripts.

Placeholders

The documentation uses placeholders to anonymize real values. The placeholders are:

The placeholders are used in commands, paths, and configuration. They are replaced with real values when the operator applies the documentation to their environment.

See also

• All other pages in the documentation.