[MODEL] Safety classifier withholds responses during open-source automotive (openpilot fork) work

What happened

During a long Claude Code session working on an open-source driver-assistance project (a sunnypilot/openpilot fork), the safety classifier withheld two consecutive assistant responses, with no explanation shown:

Your response above was stopped by a safety classifier — this is not a tool or API error.

It happened right after I asked Claude to start reworking a working personal patch into the upstream project's conventions, so it could be offered as a pull request.

Context

• The patch is a comfort feature for my own car (2022 Subaru Outback). It remembers Auto Vehicle Hold and Auto Start Stop across restarts, which the car otherwise resets on every start. It does this by sending copies of the car's own button messages. The panda (comma's firmware) safety rules only allow those button presses, and only on one bus.
• Everything earlier in the same session went through: decoding the CAN messages from my logs, a parked injection test, rebuilding the firmware, deploying to my device, and setting up a GitHub fork with a CI workflow. The feature is running on my car.
• The step that was stopped was the conventional part of contributing upstream: a default-off settings toggle, settings storage and UI. The car-facing behaviour was unchanged.

Expected

Either the work continues, or the stop includes some indication of what was considered unsafe, so the request can be scoped differently. As it is, it isn't clear whether open-source automotive work is allowed at all.

Details

• Claude Code 2.1.299, macOS, model Opus 5.5
• /feedback receipt: 6588f180-4574-47a9-9bb0-6f3fb28b2801
• Possibly related: #100212 (classifier withholding responses during legitimate work in another domain)