plot_reviewer

Under construction

Parked / experimental — not currently dispatched by @cosmotron. Requires a vision-enabled model endpoint; the recommended local Ollama build is text-only (no image input). Until a vision endpoint is wired back in, structural plot-health (missing/empty/blank) is covered deterministically by build_review_evidence instead. Treat the rest of this page as a design target, not current behaviour.

mode: subagent — advisory visual QA, runs alongside @reviewer.

What it does

Looks at the session's result PNGs with the model's vision capability and reports structural/qualitative VISUAL_FINDINGS — things like an unexpected feature in a Cℓ comparison plot or a covariance matrix that looks visually off. Never gates the run — the deterministic @reviewer (see reviewer) is the hard PASS/FAIL gate; this agent's output is advisory context alongside it.

Model requirement

Relies on the chosen local model's vision capability — the recommended Qwen3 model is vision-capable, so no per-agent model: override is needed here; it inherits the roster default. See Configuration.