A GitLab merge request or a GitHub pull request. Arnold treats both the same way and identifies them by
provider, repository id and number (/review/GITHUB/<repoId>/<number>). The UI says "merge request" (MR) for both.
Forge
The source control host: GitLab (gitlab.com or self-managed) or GitHub (github.com or Enterprise Server).
Connection
A user's link between Arnold and a forge (GitLab OAuth or the GitHub App's user authorization). Arnold acts on
the forge as that user. Forge tokens are stored in Aerospike (Copilot tokens are additionally encrypted with
COPILOT_ENCRYPTION_KEY), so treat Aerospike as a secret store.
Repository sync
A periodic job that imports every open change request Arnold can see: repositories where the GitHub App is
installed, and GitLab projects visible to the service token. See Webhooks and sync.
Merge queue
Your personal list of change requests: ones you author, review or watch. You can reorder it by priority.
Analysis run
One execution of an AI workflow (usually Improve) on a change request or group. A run records its context,
personas, harness, model, status, findings and the comments it posted.
Workflow
What a runner does for a job: improve (inline findings), smart_diff (story mode), group_review
(cross-repository review), plus CLI-only review, describe, impact and sonarfix.
Harness
The agent program that drives the model inside the runner: OpenCode (any OpenAI-compatible endpoint) or the
opt-in GitHub Copilot harness. See LLM and harnesses.
Runner mode
Where jobs execute: LOCAL (workers poll Arnold), GITLAB (CI pipelines) or GITHUB (Actions). See
Runners.
Review persona
A versioned, publishable review policy (guidance, rules, prerequisites, exclusions) that runs as a specialist
sub-agent during Improve. Its findings are tagged with the persona. See Review personas.
Review group
A named set of related change requests, possibly across repositories, that can be analysed together. See
Review groups.
Story mode (smart diff)
An AI-generated review brief and ordered review steps for a change, with reviewer focus areas pinned to lines.
Acceptance status
Arnold's assessment of whether the author acted on an AI finding: applied, partially applied, ignored or
not applicable. It feeds review metrics.