Simon Willison

It might be neater to have a dedicated `special` column, I'll see how I feel about it as I implement it. These are the three relevant tables: ```sql CREATE TABLE...

Actually no it's not the same as `input_schema`. `input_schema` is the thing you pass to the model and say "fill this out with data to call the tool". But.. in...

I think this actually depends on: - https://github.com/simonw/llm/issues/1111 Configuration like `display_width_x` already has a natural place to be stored: in the `tool_instances.arguments` column. That comes into play when you use...

Maybe there's a `supports_tool()` method that gets called for every tool, that way it can also handle things like models that need to reject tools with input schemas that include...

Here's the problem, illustrated using the default OpenAI plugin: https://github.com/simonw/llm/blob/fd2f0ffbcd93176811ccc6fc0897b67f6bc570f3/llm/default_plugins/openai_models.py#L751-L756 That's a blocking call to `build_messages()` which does this in two places: https://github.com/simonw/llm/blob/fd2f0ffbcd93176811ccc6fc0897b67f6bc570f3/llm/default_plugins/openai_models.py#L538-L539 https://github.com/simonw/llm/blob/fd2f0ffbcd93176811ccc6fc0897b67f6bc570f3/llm/default_plugins/openai_models.py#L581-L585 Which uses this: https://github.com/simonw/llm/blob/fd2f0ffbcd93176811ccc6fc0897b67f6bc570f3/llm/default_plugins/openai_models.py#L430-L435 Which calls...

I see two potential fixes here. 1. Go through that plugin and _all other_ plugins and find their equivalent of that `build_messages()` routine and find the bits where they do...

Maybe that second option could work with an extra rule: There's an upper size limit on the amount of content that can be pre-fetched for an asyncio attachment, potentially a...

We are already committed to sucking a LOT of information into memory anyway, for those APIs that require us to base64 encode the entire file and send that to the...

So the fix here could be to have the models themselves declare a `supports_attachment_urls = True` property which, if set, causes the asyncio mode to not bother trying to pre-download...

Getting a second opinion from Gemini 2.5 Flash: ```bash llm -f /tmp/llm.txt \ -m gemini-2.5-flash-preview-05-20 \ -f issue:https://github.com/simonw/llm/issues/1143 \ -s 'Think hard about this problem and if the proposed solution...