Skip to content

Run analysis

Run Analysis starts an AI review. For a single change it runs the Improve workflow: a general review pass plus up to three review personas, with findings posted as inline comments on the PR/MR.

Open it from the Run Analysis button, or from Run analysis on any queue row (which pre-fills the change).

1. What are you analysing?

  • Single change: paste a GitLab MR or GitHub PR URL and click Fetch details. Any URL works, as long as your connected account (and the runner's credential) can read it; the change does not need to be in your queue.
  • Change group: pick one of your review groups to review several related changes together.

If the change was analysed before, earlier runs are listed with Open, Reuse settings (prefill the previous context, personas and model) or Start fresh.

2. Context for the review

Context for the review

Context makes the review check the change against what it is supposed to do:

Field Limit Use
RFC document 1 URL The design doc the implementation should follow
Jira tickets up to 10 Keys (PROJ-123) or URLs of tickets the change resolves
Extra context 4000 chars Requirements, risk areas, things to focus on or ignore

RFC and Jira references in the title, description or branch name are also detected automatically. Fetching RFC content needs RFC_CENTRAL_API_URL on the runner; Jira needs the Jira options in the runner (Runner CLI).

3. Configure the review

Configure the review

Review personas. Pick up to three. Each runs as a specialist reviewer next to the general pass, and its findings carry its badge. With none selected, standard Improve runs.

AI execution. Choose the harness (OpenCode or GitHub Copilot) and a model. Only available harnesses are offered: OpenCode when enabled by the administrator, Copilot when enabled and connected in your Settings. Refresh models reloads the list. Leave the model blank to use the deployment default.

Reviewers. Add or remove reviewers on the change. These are saved to the forge immediately.

4. Review and start

A summary of scope, workflow, harness, model, Jira tickets, notes, personas and reviewers. Start creates the run and dispatches a runner job.

What happens next

  1. The run appears in Analysis History as Running, and the change shows Arnold analysing….
  2. The runner checks out the change at the head commit, reviews it with the general pass and each persona, deduplicates findings, and filters out issues already raised in existing discussions.
  3. Findings are posted as inline comments, then reported back to Arnold with their exact comment ids and locations.
  4. The run completes (or Partial if some personas failed). You get an Analysis finished notification. Runs time out after 165 minutes.

API

POST /analyses/v1
{
  "mergeRequestId": "...",
  "rfcUrl": "https://docs.example.com/rfc/42",
  "jiraTickets": ["PROJ-123"],
  "reviewContext": "Focus on backwards compatibility of the public API.",
  "reviewPersonaIds": ["netra", "hayai"],
  "execution": { "harness": "opencode", "model": "openai/gpt-4.1" }
}

GET /analysis-execution/v1 lists available harnesses and models for the current user.