Demonstration Evidence

This page is the boundary between the architecture and a claim that the runner pool works in a real cluster.

Current status

No terminal recording is published yet. The unraid-lab smoke test must pass first.

The diagrams elsewhere on this site explain the intended design and source contracts. They do not prove that a GitHub Actions job completed on an Imp runner.

What the recording must prove

flowchart LR
  queued[Harmless job queued] --> vm[Fresh ImpVM created]
  vm --> ready[Guest ready over VSOCK]
  ready --> assigned[Runner registers and receives job]
  assigned --> completed[Job completes]
  completed --> cleanup[Secret and VM clean up]
  cleanup --> capacity[Pool restores capacity]

Plain explanation: the recording needs to show the whole loop, not only that a VM booted. A working runner pool starts a job, finishes it, removes the one-time setup, and restores the pool.

Recording rules

The recording may show:

• a harmless queued GitHub Actions job;
• ImpVM metadata and lifecycle status;
• runner registration and job assignment;
• cleanup events and replacement capacity.

The recording must not show:

• GitHub App keys;
• installation tokens or JIT payloads;
• Secret data;
• full logs that might contain credentials.

Evidence checklist

Check	Required evidence
Job selection	The job is assigned to the intended Imp runner group.
VM creation	A new ImpVM is created for that job.
Guest handoff	The guest becomes ready and receives its setup through VSOCK.
Completion	GitHub reports the harmless job as completed.
Cleanup	The VM-owned JIT Secret and VM are removed after the runner exits.
Replenishment	The pool returns to its configured idle capacity.

Related material

• Under the Hood explains the source-backed control path.
• Architecture visual aids provides the presentation diagrams.
• Runner Pool E2E Runbook is the operator runbook.