Question-level checkpointing in the eval (--checkpoint-every, default 10)

Make the eval harness resume-safe (--out)

Log generation/request ids on eval failures for trace debugging

Add AGENT_MAX_RPM global throttle for free-tier benchmarking

Experiment: AGENT_SQL_ONLY mode (no templates, SQL for everything)

Add opponent/won to v_shots; fix benchmark label

Parallelize the eval harness + add per-model health table

Consolidate all eval questions into one document

Expand end-to-end benchmark to 109 Q/A pairs

Advanced analytics layer: make deep/compound questions first-class

eval: widen want_metric keywords to match rendered axis labels

Steer interpretive questions toward insightful metrics

Shot chart: color by any category (quarter, zone, opponent), not just made/miss

Expose court renderers as chart types for the flexible SQL path

Eval: correct two stale 'decline' expectations

Eval: provider/model health tracking, --runs, tougher interpretive questions

Distributions: drop mean line, fix theming, support player/team comparisons

Bias toward distributions: violin/box/histogram + stat_distribution template

Render multiple charts side by side + harden against model API errors

Make benchmark run() merge into existing results

Fix benchmark/eval to use auto tool_choice (forced function unsupported)

Pretty field labels + multi-model benchmark harness

Add flexible SQL query path + new templates + safe executor

Initial deploy: NL-to-chart NBA visualization app