Glossary

An A-to-Z lookup for the words this documentation and the app itself use. Where a term has a page of its own, the entry says what it is in one or two sentences and links there.

If you are meeting the product for the first time, read Core Concepts instead: it explains how these fit together, which a glossary cannot.

A

Account. The organization tier above a workspace. Members, billing ceilings, the model access policy, and the audit log live here. See Workspaces and accounts.

Agent kind. What a step does and how it is prompted: Coder, Architect, Reviewer, Tester, Merger, and the rest, plus any custom kinds a deployment registers. A kind decides whether a step runs inline or in a container, and which default model it takes.

Agent tier. How much of the agent catalog a surface lists (basic through advanced). Separate from the interface mode: the mode decides which surfaces exist, the tier decides how much of one surface's catalog is offered.

B

Block. The unit on the board. A block is both plan and work item, at one of four levels: an epic, a service frame, a module inside a frame, or a task leaf. "Card" is the same thing as it is drawn on the board.

Board. The pannable canvas a workspace's blocks live on. See Design Your Board.

Budget. The monetary ceiling on metered model spend, tiered per workspace, account, and user. Subscription and local-runner runs are not metered against it. See Control Spend with Budgets.

Bug hunt. The interactive surface that rates a tracker board's open, unassigned bugs on impact against complexity and adopts the one you confirm as a task. Distinct from bug intake, the unattended step inside the recurring bug-triage pipeline, which claims the oldest matching issue on a schedule.

C

Companion. A rework partner paired with a producing agent kind. It loops that producer back on a bounded budget before a person is asked to intervene.

Consensus panel. Running one eligible step as a multi-model panel instead of a single model.

D

Decision. A point where a run parks and waits for a person: approving requirements, choosing between proposed implementations, resolving a judge's verdict, or acting on a merge review. A parked run waits indefinitely by design. See Decision prompts.

Document source. A connected system (Confluence, Notion, Linear Docs, a repository's own docs) the platform reads context from before a run. See Documents.

E

Ephemeral environment. A live preview stack provisioned for one run and torn down afterwards, so integration and end-to-end tests hit a running system. See Environments.

Epic. A block level above tasks, used to group work that ships together.

F

Foundational service. A shared capability your organization already runs (an auth service, a payments gateway) declared once so every agent is told it exists instead of reinventing it. See Register Foundational Services.

Frame. A service block on the board: the container a module or task sits inside. A task's repository is resolved by walking up to its enclosing frame, so a task outside one cannot run.

G

Gate. A step that checks something and only escalates when the check fails. The CI gate reads your host's real check runs; the conflicts gate watches mergeability; the post-release-health gate watches monitors after a merge. A failing gate dispatches a helper agent rather than stopping the run outright. Custom gates are registered like custom agents.

H

Harness. The fixed, trusted wrapper inside a run container that clones the repository, launches the model, commits, pushes, and opens the pull request. The model never does any of that itself. See Agent Isolation.

Headless run. A run started against a supplied brief with no board card and nothing pushed to a repository, driven through the public API.

I

Initiative. A larger body of work decomposed into blocks, optionally from a preset. See Plan an Initiative.

Interface mode. Basic (the shipped default) or advanced. Basic hides override controls whose default is already what the hidden field would show; nothing that is only reachable in advanced mode is needed for the everyday delivery loop.

J

Judge. An inline model that scores a step's output against a rubric and disposes: advance, park for a person, bounce the work back with findings, or fail. Distinct from a gate, whose check is programmatic and cheap.

M

Manifest. The declarative file a deployment supplies to wire an integration it owns (a runner pool, an environment provider, a code adapter). See Integration Manifests.

MCP. The Model Context Protocol, which appears here in two unrelated roles. A tool server is an MCP server an agent may call during a run. The MCP server is this platform's own public API exposed as MCP tools to an outside host. See MCP Server.

Model flavour, or route. One of the ways a single catalog model can be reached: a coding-plan subscription, a direct vendor key, an aggregator gateway, a residency-guaranteed route, or the Cloudflare fallback. A preset's route order decides which one a run takes.

Model preset. A named assignment of models to agent kinds: one base model plus optional per-kind overrides. Exactly one preset is a workspace's default. See Connect a Model Provider → Model presets.

Module. A block level between a service frame and its tasks, used to group tasks inside one service.

Mount. How one account-owned service appears on several boards as a single synced copy. Its position on each board belongs to the mount, not to the service. See Share Services Across Workspaces.

P

Pipeline. The ordered chain of steps a task runs: which agent kinds, in what order, with which gates and human decision points. See Run a Pipeline.

Prompt fragment. A reusable block of standing context or standards folded into agent prompts. See Apply Standards with Prompt Fragments.

R

Reusable operation. A custom task type that carries the whole bundle: a per-case form, its standing context, and its own canned pipeline. A custom task type carrying only presentation is a classification (a badge and a card), not an operation.

Run. One execution of a pipeline against one block. Runs are durable: they checkpoint, survive interruptions, and resume. See Runs.

Runner pool. Your own infrastructure executing agent containers instead of the platform's default. See Run Jobs on Your Own Runners.

S

Service. A top-level frame on the board, linked to exactly one repository (optionally a directory inside it, for a monorepo). The linkage is what lets a task resolve a repository at all.

Skill. A packaged instruction set attached to an agent kind, materialised into the run container for the agent to follow. See Skills.

Step. One unit of a pipeline: an agent, a gate, a judge, or a one-shot engine step such as the tracker or the disposer.

T

Task. A leaf block, and also the word used at the tracker boundary for a linked issue. On the board, "task" is a block level; the separate task type axis (bug, feature, and any custom types) decides a card's badge and which pipeline it starts with.

Task source. A connected issue tracker that seeds tasks and receives progress back. See Issue Sources.

Tool server. An MCP server an agent may call during a run, registered by the deployment and handed a credential resolved by name. See Add a Custom Agent Kind → Skills and tool servers.

U

Unattended run. A run nothing is watching: started over the public API, dispatched from a tracker issue, or fired by a schedule. It resolves the workspace's second default risk policy rather than its in-app one, and under an unattended policy it answers the checkpoints its own automatic loops raise instead of waiting for a person. Gates the pipeline asks for still stop it. See Runs nobody is watching.

W

Workspace. One board plus its members, settings, presets, and connections. A workspace has exactly one source-control installation and may hold many repositories.

Next: Core Concepts for how these fit together, or the Cookbook for a recipe per operation.