Concepts

The role of Copilot

Copilot is optional. It suggests. It never records the team’s decision, and it does not run an assessment unless you ask it to.

When it is off

Set nabsConfidence.useAI to false, or use the extension without GitHub Copilot Chat. New retros still gather the deterministic insights: commits, authors, lines, hot files, and, with a GitHub sign-in, issues, pull requests and discussions. Summarise Feedback then groups entries by type instead of asking a model. You can still write evidence and a team verdict. Home shows No Copilot in Set-up, and the rest of the page still works.

When it is on

The extension uses GitHub Copilot through VS Code’s language model API. It needs the GitHub Copilot Chat extension, a sign-in, and permission to use the language model. nabsConfidence.modelFamily can name a preferred family. Leave it empty to use the first Copilot model VS Code offers.

Three jobs ask Copilot something:

• Summary of the period, when you create a retro or run Analyse / Refresh Insights. With objectives chosen, it is asked for a few bullets on progress towards each one, then one “Outside focus” line, then questions for the team.
• Feedback summary, when you run Summarise Feedback into Retro. It is asked for themes about the chosen objectives, and for a few action items that name an objective with #obj:.
• Objective assessment, when you run Assess Objectives with Copilot. For each chosen objective it is asked for one of four verdicts: on track, at risk, off track, or not enough evidence, plus its reasoning, the evidence it relied on, and what would change its mind.

A fourth way to ask is Copilot Chat, with @retro. That does not write the retro unless you use the button that offers to write a summary in.

What it is told, and what you should still check

The assessment prompt tells the model to use only the evidence it was given, not to invent numbers or facts, and to answer “not enough evidence” when the evidence does not support a judgement. Answers that do not match the expected shape are dropped, and you see that Copilot did not assess that part.

That is an instruction and a parser. It is not a guarantee. Read the reasoning, and keep the team’s verdict when you disagree. If the team’s verdict differs from Copilot’s, a reason is required. Both are kept in the retro.

Copilot does not assess objectives on its own when you open Home or the dashboard. Someone runs Assess Objectives with Copilot.