Fixed channel_plot(target = "agent") silently returning a blank image for plotly/htmlwidget plots. .channel_plot_agent() had no branch for interactive widgets, so the print step was skipped, the untouched graphics device was captured, and the agent received a plausible-looking but empty PNG with no error. Interactive plots now render at 900x650 through an isolated headless-Chrome subprocess and are encoded the same way as static plots; a zero-byte or oversized capture now aborts loudly instead of reaching the model silently.
Corrected agent-image token accounting from base64 text length (nchar(base64) / 4) to a pixel-based estimate (width * height / 750), which reflects what a vision model is actually charged for an image block rather than the length of its base64 encoding. The token limit and warning threshold moved to 3000/2000 to stay meaningful in those units; the previous 25000/20000 values were unreachable under the corrected formula.
channel_plot(target = "agent") now accepts only plotly among interactive widgets. Other htmlwidget types (leaflet, DT, visNetwork) abort with a message pointing at target = "user" instead of failing obscurely inside the renderer, which only knows how to locate a plotly plot.
Rendering interactive plots for agent analysis requires a Chromium-based browser (Chrome, Chromium, Brave, or Edge), discoverable via chromote::find_chrome() or the CHROMOTE_CHROME environment variable. This is optional — only used by channel_plot(target = "agent") on plotly/htmlwidget objects; static plots and target = "user" are unaffected.
