Glossary

Terms are listed alphabetically. Cross-references use italics.

Agent

A configured worker that does one piece of a workflow. An agent has a role, a set of skills, an output definition, instructions, and a provider + model. Agents run in waves and pass output to each other through connections.

Agent Palette

The left-edge dock from which you drag agent role cards onto the canvas.

AgentSA

Slice 2 of Situational Awareness. The harness's awareness of its own agents' recent uncertainty snapshots by role, the patterns it currently believes hold (with belief-source attribution), and the open blind spots it's tracking.

Approval Gate

A connection type that pauses the workflow when it's reached, requiring you to approve or reject before downstream agents run.

Audit log

An NDJSON file under ~/.cerevisor/audit-logs/ recording every tool call, permission decision, and event in a run. One file per session.

Canvas

The main editing surface where you place and connect agents.

Chat Builder (AI Chat Builder)

The floating chat window (open with Ctrl/Cmd+L) where you describe a workflow in natural language and Cerevisor creates or edits agents for you.

Claude Agent SDK

One of the six provider types. Anthropic's official agent SDK (the same harness that powers Claude Code) running in-process. Delegated loop: Cerevisor sets sandbox + approval policy up-front, then hands the entire agent run to the SDK. Authenticates with an Anthropic API key stored in a dedicated keychain slot. See Claude Agent SDK provider.

Codex CLI

OpenAI's command-line tool (@openai/codex). One of the six provider types Cerevisor supports. Authenticates against a ChatGPT subscription via device-auth.

Column

A vertical lane on the canvas representing one wave of execution. All agents in a column run in parallel.

Connection

A directed edge from one agent to another representing data handoff. Connection types: Always, Conditional, Approval Gate, Loop.

Control Group

A saved selection of agents you can recall instantly with Ctrl+Shift+1..9. Like RTS unit groups.

Cross-workflow link

A connection between agents in two different workflows on the same World View. Configured via the Cross-Workflow Link modal.

Cursor Agent

One of the six provider types. Uses the in-process Cursor SDK; supports local and cloud execution. Delegated loop.

Cursor cloud run

A workflow run that executes on Cursor's cloud infrastructure rather than your local machine. Survives Cerevisor restarts via the cloud-run tracker.

Custom agent

An agent built from the Custom role, or a template saved by you. Reusable across workflows.

Default provider

The provider in your library marked as default. Used by new workflows and any agent without an explicit override.

Draft (auto-save)

In-memory state saved every 30 seconds to a hidden file. Survives a crash; offered on next launch.

Embedded App

A native desktop application (Outlook, Excel, etc.) embedded as a node on the Cerevisor canvas. Windows only.

Execution Modal

The popup showing live status of a workflow run.

Failure Report

A structured summary of what went wrong in a run. Drives Loop & Enhance's decision to re-plan.

File intelligence

Harness-level scheduling logic that uses each agent's declared inputFiles[] and outputFiles[] to detect dependencies, avoid write conflicts, and inform wave scheduling, without injecting file content into agent context.

Flow type

One of four behaviors a connection can have: Always, Conditional, Approval Gate, Loop.

Gemini

One of the six provider types. Google's Gemini API (2.5 and 3.x families) accessed via the @google/genai SDK. Cerevisor owns the agent loop. Handles >200K-token tier pricing automatically and folds thinking tokens into the output billing total. See Gemini provider.

Handoff format

How an agent's output is packaged before the downstream agent reads it. Options: Summary, Structured JSON, File Reference, Full Output.

Harness

The orchestrator runtime that drives agents, manages waves, dispatches tool calls, enforces permissions, runs the agentic loop.

Harness self-memory

The harness's persisted retrospective on its own runs, per-role metrics, learned tips, corrections. Stored under ~/.cerevisor/memory/harness/.

Lemon Squeezy

The payment provider used for Cerevisor licenses. Singletons gate certain features (workflow caps, Loop & Enhance, MCP, cloud).

Loop (connection type)

A connection that re-runs the source agent until an exit condition is met or maxIterations is reached.

Loop & Enhance

A round-loop system that wraps a workflow run in detect-fix-rerun cycles. After each round, the harness analyzes failure, proposes workflow edits via an LLM, applies them, and runs again.

loop_fixation_warning

An orchestrator event emitted by Loop & Enhance when two consecutive rounds produce the same FailureReport fingerprint (the planner isn't making progress). Consumed by the Situational Awareness renderer to flip the run's state label to fixation-risk and surface a warning in the SA panel.

MCP (Model Context Protocol)

An open protocol for tool/context servers. Cerevisor is an MCP client; you can connect any MCP server (e.g. filesystem-mcp, github-mcp) and its tools become available to your agents.

Memory subsystem

Persistent storage under ~/.cerevisor/memory/ covering user profile, harness self-memory, meta-cognition observations, disagreements, freeform entries, and version snapshots.

Meta-cognition

The harness's diff-over-time analysis of its own retrospectives, producing observations like "role X has been failing more often lately" and an enduring self-portrait.

Native mode (Structured Orchestration)

A workflow setting that enriches the generated prompt with subagent types, tool scoping, and model hints. Independent of which provider executes the run.

Output definition

The contract for what an agent's output should look like. Drives validation, the agent output gate, and downstream context injection.

Output Verifier

An LLM-as-judge pass over an agent's output to check it matches its output definition.

Pipeline

A chained run of multiple workflows. Configure via the Pipeline modal.

Placement mode

A canvas state where your cursor becomes a ghost of an agent about to be placed. Entered by pressing 0–9 or dragging from the palette.

Provider

A backend that runs your agents. Six types are supported: Anthropic, Gemini, Claude Agent SDK, OpenAI-compatible endpoint (Ollama, OpenRouter, etc.), OpenAI Codex CLI, and Cursor Agent. Three of those (Anthropic, Gemini, OpenAI-compatible) are Cerevisor-loop providers where Cerevisor owns the agent loop; the other three are delegated-loop providers where the upstream runtime owns the loop. See Provider overview.

Provider Library

The collection of all configured providers. One is marked as the default.

Provider override (per-agent)

Setting a specific provider on a single agent, overriding the workflow's default.

Role

The starting profile for an agent, Researcher, Architect, Developer, etc. Sets smart defaults for subagent type, model preference, and accent color.

Run finalization

End-of-run cleanup: persists the session, fires retrospectives, updates harness memory, writes meta-cognition observations.

SA Directive

The always-on situational-awareness framing appended to every agent's system prompt (alongside the Wow Directive). Toggle and edit text live at Settings -> Workflow. Independent of the Wow Directive: Wow asks "is this output worth showing off?"; SA asks "does this output honestly surface what the agent doesn't know?". See SA configure and run.

SAStateLabel

The coarse discrete label summarizing the harness's Situational Awareness posture for a run. One of: baseline, high-confidence-known-domain, low-confidence-new-domain, fixation-risk, context-stale, unknown.

Scheduled run

A workflow configured to run automatically on a cron schedule. Configured via the Schedule Config popup; managed under Settings → Automation.

situation_reporter

One of the 17 built-in roles. Composes a structured handoff (default output file wave-situation.md) after a wave completes, so downstream waves inherit one canonical read of what's done, what's missing, and what's still unknown. Defaults: subagent type plan, model sonnet. Pairs with the Situational Awareness pillar.

Situational Awareness (SA pillar)

Cerevisor's continuously-updated, three-level picture of a workflow run, composed from memory + orchestrator + persistence signals into typed snapshots. Three slices: TaskworkSA, AgentSA, TeamworkSA. Surfaced as state labels, uncertainty badges, and per-wave reports. See Situational Awareness overview.

Skill

A reusable, focused capability. Lives as a folder containing a SKILL.md plus optional helper scripts. Cerevisor merges ~/.claude/skills/ and ~/.cerevisor/skills/ on launch.

Skill compatibility (provider compatibility)

Whether a skill works with a given provider. The skill-preflight gate checks this before any wave runs.

Skill Workshop

The in-app editor for creating, editing, and managing skills. Switch to it from the title bar.

Subagent type

One of four execution modes: explore (read-only), plan (planning), bash (shell), general-purpose (full toolset).

TaskworkSA

Slice 1 of Situational Awareness. The harness's awareness of the task itself: what changed since the user last interacted (the sinceLastSession delta), which sibling workflows in the workspace have been touched recently, and the current workflow's identity.

TeamworkSA

Slice 3 of Situational Awareness. The harness's awareness of the working relationship: open disagreements (where the user's claim and the harness's observation diverge), and the current chat-builder session's accumulator of accepted / rejected / clarified items.

Telemetry

The product is no-telemetry by default. Nothing leaves your device unless you explicitly opted into a feature that requires network access (a provider's API, a web-search backend, an MCP server you added).

Template

A reusable starting point; either a workflow template (on the Home Screen) or an agent template (Custom Agents).

Trust Profile

A user-chosen risk appetite (Cautious / Standard / Trusted Developer) that controls how often the permission system prompts vs. auto-approves.

UncertaintySnapshot

A bundled per-agent-run uncertainty signal composed from the agent's self-reported confidence marker (if any), the semantic verifier's verdict, and the missing-marker signal. Mirrored in the renderer and shown as the per-agent uncertainty badge in Situational Awareness.

Wave

A column on the canvas. Agents in the same wave run in parallel. Later waves wait for earlier ones to finish.

WaveSituationReport

A structured report assembled by the Situational Awareness pillar after each wave completes. Lists per-agent status + uncertainty, files produced this wave, files the wave was expected to produce but didn't, and free-form notes. Injected into the next wave's downstream agents as a contextNotes segment so the harness's SA carries forward across waves.

Workflow

A single document containing agents, columns, connections, and settings. Saved as a .cerevisor (or .opaal) file.

Workspace

A multi-workflow document containing N workflows arranged on the World View. Saved as a .cerevisor-world file.

World View

The multi-workflow canvas where each workflow appears as a framed area. Add cross-workflow links, see all your work at once.

Wow Directive

A global quality bar appended to every prompt and injected into every agent's system message. Editable in Settings → Workflow.