Managing rationale, confidence, and approvals for issues

Set up and manage rationale, confidence, and approvals for issues handled by Automatisation.

Qui peut utiliser cette fonctionnalité ?

Automations sont disponibles avec les offres GitHub Copilot Pro, GitHub Copilot Pro+, GitHub Copilot Max, GitHub Copilot Business et GitHub Copilot Enterprise. Automations sont disponibles uniquement dans les dépôts privés et internes. Ils ne sont pas disponibles lorsqu’ils ont été désactivés par un administrateur de référentiel, dans les référentiels détenus par comptes d’utilisateur managésou lorsqu’ils Agent cloud Copilot ne sont pas activés.
Sign up for Copilot

Dans cet article

Rationale and confidence need no extra configuration. When an Automatisation's agent changes a supported issue attribute, it attaches a rationale and a confidence level automatically. Whether that change applies directly or waits for review depends on your repository's automation level, or on an explicit request to suggest rather than apply the change.

Setting your repository's automation level

Remarque

Configuring an automation level is rolling out gradually and might not be available in your repository yet.

Your repository's automation level sets the confidence threshold used to route every supported change. Changes rated below the threshold are held as suggestions; changes at or above it apply automatically. The default level is Cautious, so only high-confidence changes apply automatically until you change it.

  1. Sur GitHub, accédez à la page principale du référentiel.

  2. Sous le nom de votre référentiel, cliquez sur Paramètres. Si vous ne voyez pas l’onglet « Paramètres », sélectionnez le menu déroulant , puis cliquez sur Paramètres.

    Capture d’écran d’un en-tête de dépôt montrant les onglets. L’onglet « Paramètres » est mis en évidence avec un encadré orange foncé.

  3. In the "Planning" section of the sidebar, click Agent suggestions for issues.

  4. Under "Automation level," select the level that fits your repository:

    • Full control: Every change is held for review. Nothing is applied automatically.
    • Cautious (default): Only high-confidence changes are applied automatically. Everything else is held for your review.
    • Balanced: Routine, clear-cut changes are applied automatically. Anything with ambiguity is held for review.
    • Full automation: Every change is applied automatically. The agent only holds a change back if it's flagged as uncertain.

  5. Click Save.

With Agent cloud Copilot

  1. Create an Automatisation that triages issues, following Création d’automatisations avec Copilot agent cloud. Use the When an issue is created trigger, or run the Automatisation on a schedule.
  2. Grant the Automatisation the issue tools for the attributes you want it to change, such as updating labels, setting the type, editing fields, assigning users, or closing issues.
  3. In the prompt, describe the triage task, for example: Triage this issue by setting a type, labels, and priority, and explain your reasoning for each. You don't need to mention rationale, confidence, or suggestions explicitly. To hold specific changes for review regardless of confidence, ask the agent to suggest those changes rather than apply them.
  4. Save the Automatisation. When it runs, its changes carry a rationale and a confidence level. Changes below your repository's automation level, and any changes you asked the agent to suggest, wait in the approvals panel on the issue.

With Flux de travail GitHub pilotés par des agents

Issue intents, the rationale and confidence metadata attached to a workflow's issue changes, are optional and enabled by default. Existing workflows keep working without changes.

  1. Update to the latest version of the Flux de travail GitHub pilotés par des agents extension by running gh aw upgrade.

  2. Add a workflow that triages issues to your repository's .github/workflows directory. You can start from an issue triage sample or author your own. See the Flux de travail GitHub pilotés par des agents documentation.

  3. Give the workflow the outputs for the attributes you want it to change by listing them under safe-outputs in the workflow's frontmatter. For example, add-labels to apply labels, set-issue-type to set the issue type, set-issue-field to edit fields, assign-to-user or assign-to-agent to assign issues, and close-issue to close them. Adding these outputs is enough for the workflow to make the corresponding issue changes with a rationale and confidence level.

  4. Optionally, control whether a safe output requires issue intent metadata by setting issue-intent on it:

    safe-outputs:
  add-labels:
    issue-intent: true
    • Omit issue-intent, or leave it unset, to keep it optional: the agent is encouraged to provide a rationale and confidence level, but the output still works without them.
    • Set issue-intent: true to require a rationale and confidence level for that output. The workflow fails if the agent omits them.
    • Set issue-intent: false to opt out entirely: the output never carries rationale or confidence metadata.

  5. In the workflow prompt, describe the triage task. To hold specific changes for review regardless of confidence, ask the agent to suggest those changes rather than apply them.

  6. Compile the workflow by running gh aw compile, then commit the workflow file and its generated lock file.

  7. When the workflow runs, its changes carry a rationale and a confidence level (unless you opted out). Changes below your repository's automation level, and any changes the agent suggested, wait in the approvals panel on the issue.

With the REST or GraphQL API

If you build your own integration instead of using an Automatisation, you can use the same issue-writing REST and GraphQL APIs to include a rationale and confidence level with a change, or mark a change as a suggestion to hold it for review. The change is routed the same way as a change from an Automatisation or a Flux de travail GitHub pilotés par des agents workflow: subject to your repository's automation level, or held for review if you marked it as a suggestion.

Finding issues with pending suggestions

To find issues that have suggestions waiting for review, search using the has:suggestions qualifier, for example is:issue is:open has:suggestions.

Reviewing suggestions in the approvals panel

When a change is held for review, it waits in an approvals panel on the issue instead of taking effect. Use the panel to inspect each suggestion and decide what to apply.

  1. Open the issue. Suggested changes appear in the approvals panel.
  2. Inspect each suggestion. It shows the proposed change, such as a label, type, field, assignee, or close, along with the rationale and confidence level.
  3. Act on the suggestions:
    • Select Accept to apply a suggestion, or Decline to dismiss it.
    • Select Accept all or Decline all to act on every pending suggestion at once.

Accepted changes take effect on the issue immediately. Declined suggestions are dismissed without changing the issue.

Ready to put this into practice? See Création d’automatisations avec Copilot agent cloud to create an Automatisation that triages issues for you.