Go Further

Server API and GitHub automation

Use the existing Fabric Agents session APIs, and understand what a reliable GitHub ticket integration still needs.

Fabric Agents servers expose an authenticated API. Sessions created through the server's native API are the same conversations you see in Desktop and Web UI. Fabric Code has its own sessions and execution engine.

Start with the CLI

Connect fabric-cli to your server as described in Remote clients. You can then inspect workspaces, create a session and continue it from scripts:

fabric-cli --json workspaces
fabric-cli session create --name "GitHub issue 683" --mode ask
fabric-cli send SESSION_ID "Investigate the issue and propose a fix."
fabric-cli session messages SESSION_ID

Replace SESSION_ID with the ID returned by session creation. Choose an existing server-side connection/model when you need a particular account; see the CLI reference. Provider sign-in belongs to that server's profile. No additional Fabric cloud account is required.

Native APIs

Desktop, Web UI and CLI use authenticated WebSocket RPC for session creation, sending, history and events. The server also includes an HTTP coding-session adapter at /api/coding-runs/sessions. Availability depends on the deployment's server settings.

HTTP requestPurpose
GET /api/coding-runs/sessions/readinessCheck protocol capabilities and profile identity. This does not prove provider sign-in or a successful model call.
POST /api/coding-runs/sessions/readinessCheck a specific repository, connection/model and target.
POST /api/coding-runs/sessionsCreate a native coding session/worktree with a stable codingRunId.
POST /api/coding-runs/sessions/SESSION_ID/promptContinue that exact session with a stable promptId. Busy sessions return 409; they are not silently queued.
GET /api/coding-runs/sessions/SESSION_IDRead tracked execution/checkpoint state.
GET /api/coding-runs/sessions/SESSION_ID/messagesRead bounded conversation content for sessions owned by this adapter.
GET /api/coding-runs/sessions/SESSION_ID/artifactsInspect supported result artifacts.
POST /api/coding-runs/sessions/SESSION_ID/stopRequest and confirm cancellation of the owned run.

HTTP API clients authenticate with the server's existing bearer token or the supported signed internal authorization. A Web UI login cookie alone does not authenticate these API requests. Keep credentials in your integration's private configuration; do not put tokens in links, issue comments or webhook payloads.

A successful prompt acknowledgement means the message was persisted. It does not mean the model completed the task, pushed commits or opened a pull request. Reuse the same promptId and identical content when reconciling delivery. If acceptance is uncertain or the API requests recovery, inspect that state before sending anything again.

A GitHub Ready → In review → Rework workflow

This workflow fits the existing server APIs, but the server does not currently ship a complete GitHub ticket webhook receiver. An optional integration needs to:

  1. Verify GitHub's webhook signature and permitted installation/repository.
  2. Persist and deduplicate deliveries before acknowledging them.
  3. Bind each ticket to one server profile, workspace and native session.
  4. Queue that ticket's Ready/Rework events in order, using stable prompt IDs.
  5. Reconcile completion and verify the requested result artifacts.
  6. Reliably publish the result comment and In review status, retrying GitHub publication without running the model task again.

Rework continues the existing session; it does not create another session with changed creation arguments. Pin the intended workspace ID/slug in targetMachine on a multi-workspace server, rather than relying on the first workspace.

An administrator must explicitly choose the repository, provider account and unattended execution policy. Existing Ask sessions keep their normal approval behavior. Missing sign-in, permission requests, quota failures and interrupted runs need attention; they must not be reported as completed work.

For browser tests run by a headless server, keep a capable Desktop client connected to that workspace or configure a separate browser test runner. Opening the Web UI in an ordinary browser does not supply the Desktop browser-tools host.

On this page