Simon Willison

There are some great ideas in here, sorry for not providing a full review!

Here's an example response: ```json { "id": "msg_01XQHWaBXzfy8jjfuSPSzXbA", "type": "message", "role": "assistant", "model": "claude-opus-4-20250514", "content": [ { "type": "text", "text": "I'll calculate the mean and standard deviation of the given...

Relevant schema: https://github.com/simonw/llm/blob/e18eb3a59574b46d57cc2f68423b3c44a3b37051/docs/logging.md?plain=1#L352-L380

I need to store this in the database such that `llm logs` can reproduce the output. This is hard because we have test both before and after the tool calls....

Currently returns no results at all against the `fixtures` database: https://latest.datasette.io/fixtures/-/query?sql=select%0D%0A++m.name+as+table_name%2C%0D%0A++json_group_array%28ii.name%29+AS+unique_column_names%0D%0AFROM%0D%0A++sqlite_master+AS+m%2C%0D%0A++pragma_index_list%28m.name%29+AS+il%2C%0D%0A++pragma_index_info%28il.name%29+AS+ii%0D%0AWHERE%0D%0A++m.type+%3D+%27table%27%0D%0A++AND+il.origin+%3D+%27u%27%0D%0AGROUP+BY%0D%0A++il.name%3B The compound unique column pattern used in Datasette `internal` isn't as common as I thought!

Relevant from the Click changelog: > Keep stdout and stderr streams independent in CliRunner. Always collect stderr output and never raise an exception. Add a new output stream to simulate...

Note that Click dropped support for Python 3.9 early - it's EOL in October: - https://github.com/pallets/click/pull/2893 My policy is to support until the EOL date. That's fine if Click still...

Imagine a tool that could take a 1000 page PDF and return 5 images of five pages from it, but then be called again on future rounds to return the...

```python import cairosvg svg_data = "…" png_bytes = cairosvg.svg2png(bytestring=svg_data) ``` Might work. https://chatgpt.com/share/682353b5-f200-8006-bb4b-6aed78513f3d

Attachments definitely make sense. I'm not sure about fragments, so I'm dropping that until it becomes obviously useful.