Notes from integrating Jeff v1.2 into a coding-agent harness: template slots, consumer GPU, torchvision #8

Thanks for Jeff. Three small things came up while we wired it into a coding-agent harness. Each one is independent.

1. jeff-kit check-rows rejects real template files. We built evaluation rows from real tool output a coding agent had read. One was a blog template containing {{AUTHOR}} and {{BODY_HTML}}, and it fails with "unfilled template slots". That check is right for generated training data. For evaluation rows made of real content, slots can be legitimate. An opt-out would help, keeping the strict check as the default: either a flag such as --allow-template-slots, or a per-row marker in source.

2. Numbers from a consumer GPU.

• Setup: an RTX 3070 Ti Laptop GPU (8 GB) under WSL2, with torch 2.8.0+cu128, torchvision 0.23 and transformers 5.17, without flash-linear-attention.
• Memory: with the v1.2 base, guard, ground, and a small separate classifier server loaded, nvidia-smi showed 2,098 MiB in use out of 8,192.
• Latency: median per decision, measured from the client, was:
  • 144 ms for grounding rows with evidence capped at about 1,500 characters
  • 199 ms for ground adapter rows
  • 274 ms for guard on texts of up to about 4,000 characters
• Versions: pyproject.toml pins torch 2.14, but serving worked on 2.8. That may be useful for people reusing an existing environment.

3. Text-only serving still needs torchvision. If torchvision is missing, startup fails with:

Qwen3VLVideoProcessor requires the Torchvision library but it was not found in your environment.

The cause is that AutoProcessor loads the vision-language processor. torchvision is already in the serving dependencies, so this only affects people who reuse an environment. A one-line note in the docs might save someone a few minutes.

_{✓ Patchwork Harness approved: drafted from measured runs, and checked before posting by a cross-vendor review (GPT-6.1 Sol) in Patchwork Harness, which now uses Jeff for its prompt-injection guard.}