- multiple parsers per domain (start with HN threads)
- test duckdb_utils
- reuse duckdb connection
- better model config options?
- organize prompts, always use dedent
- CI tests
- sometimes selector agent gets "lost in the sauce" with tool attempts
- parser codegen could be higher quality (twitter followers example)
- better process for iterating on codegen parsers? (e.g. using IDE)
- many tool calls in selector proposal can result in exceeding context length
- trimmed down tool results so might be better (see selector_tools constants to tune)