| name | ac-harness |
| description | Launch the autonomous AC drive harness to test in-sim — car setups, dashboards, voice coaching, telemetry capture. Use whenever a task needs Assetto Corsa driven and evidence collected; never write a throwaway .scratch driver. |
AC autonomous harness (pointer skill — Cursor mirror)
Mirror of .claude/skills/ac-harness/SKILL.md (kept in sync per docs/00_Core/TOOLCHAIN.md).
The canonical procedure lives in
docs/10_Development/18_Autonomous_Harness.md
— read it now; this skill deliberately does not re-inline it (prose-drift pitfall:
skill-reinlines-canonical-procedure).
What you get from the runbook:
- The one command (
python -m tools.ac_harness.auto_drive --car … --track … [--setup …])
that owns preflight → sidecar → deterministic launch → setup apply+verify → drive →
WS assert → evidence bundle.
- The evidence contract (
report.json, hud.png, lap-archive paths) your task cites as
proof — a run without its bundle is not verified.
- The troubleshooting table of rig lore, and the honest known limitation (setup runs vs.
the autonomous drive don't yet compose in one command — tracked as a follow-up).
Hard rules for agents:
- Issue
--preflight-only first when unsure the rig is ready; fix what it names.
- Never bypass a
stage="setup" FAIL by dropping the flag — that failure is the harness
doing its job.
- One rig, one driver: check for a live
acs.exe / peer autonomous session and yield.
- Downstream tasks compose on the evidence bundle; promote anything durable out of
.scratch/.
