A
Agent. The coding model that edits your code in a run. In Bellows it is chosen through an executor. Attachment. A screenshot, PDF or other file you add to a task or follow-up so the agent can work from it. See Attachments.B
Base branch. The branch a task starts from and its pull request targets. You can pick it in the task composer. Bellows reads.bellows.yaml from its tip. See Policy and reviewers.
C
Connection. A stored third-party credential, such as Jira, that you attach to a task so the agent can use it. See Connections.D
Driver. The part of Bellows that picks up queued jobs and starts a runner for each one. See Introduction.E
Evidence. What a policy requires before Bellows publishes work: a passing full gate run, a clean review, or both. Without it the task fails withevidence required.
Executor. The coding agent that does the work: Claude Code, OpenCode, or a custom profile built on them. See Executors.
Executor profile. A saved executor setup with a name, an agent, a model and optional configuration. It is personal (only you) or organization-wide (every member). See Executors.
F
Failure kind. The reason shown on a failed run, such astimed out, gate failed or config error. Each one is explained in Task lifecycle.
Follow-up. Your next message on a finished task. It starts a new job on the same checkout and session. See Run tasks.
G
Gate. A named check command in.bellows.yaml, such as build, test or lint. A run succeeds only if its gates pass. See Gates.
Gate filter. A filter: true gate that lets the agent run part of a suite. The value reaches your command as $GATE_FILTER. A filtered run never counts as proof. See Gates.
Gate repair. The automatic rounds in which the agent gets a failing gate’s output and tries to fix it. The limit is the executor’s Gate repair rounds setting. See Gates.
I
Inbox. The Tasks page. Its Needs attention tab holds everything waiting on you: answers, reviews and failures. See Run tasks.J
Job. One attempt by the agent inside a task. Also called a run. Follow-ups, answers, retries and edits each start a new job. See Task lifecycle.M
Mark done. Closes a task in Bellows. It does not merge or close the pull request. Reopen brings the task back. MCP. Model Context Protocol. The Bellows MCP server lets an assistant such as Claude Code create and follow tasks for you. See Connect an MCP client.N
Needs attention. The inbox tab for tasks waiting on you. See Inbox.O
On-demand gate. A gate markedon_demand: true. The agent can run it while it works. The end-of-run pass skips it, and it never fails the task. See Gates.
Organization. One Bellows GitHub App installation. It owns the repository list and keeps its data apart from other organizations. See Core concepts.
Organization access token. A token with the oat_ prefix. It is read-only and meant for automation that should outlive any member. See Install and sign in.
P
Personal access token. A token with thefat_ prefix that acts as you in the CLI. You create it on your Account page. See Install and sign in.
Policy. The policy block in .bellows.yaml. It requires passing gates, a review, or both before work is published. See Policy and reviewers.
Preflight. An advisory check of a task selection (repository, executor, skills, connections) that creates nothing. Use bellows job preflight or the preflight_task MCP tool. The server checks again when you create the task.
Q
Question. The agent asking you something mid-run. The task shows Needs answer until you answer. See Run tasks.R
Repository. A GitHub repository tracked by Bellows. You choose which ones under Settings → Repositories. See Configure your repo. Reviewer. A separate agent run, declared underreviewers in .bellows.yaml, that reviews the work and returns a verdict of clean or blockers. See Policy and reviewers.
Run. See Job.
Runner. The container your .bellows.yaml describes. The agent works and the gates run there. See Introduction.
S
Secret. An environment value that is write-only after you save it. The opposite is a variable, which stays readable. See Environment variables and secrets. Service. A sidecar container, such as Postgres or Redis, that runs next to the agent so tests have something to talk to. See Services. Setup. Thesetup command in .bellows.yaml. It runs before the agent starts and is retried while it fails. See Configure your repo.
Skill. A versioned bundle of instructions (SKILL.md) the agent can load. Call one from a task with a /. See Skills.