Add tool support and tests

[jamessanford]

• Add tool support for OpenRouter by propagating tool support from the supported_parameters field in the models API output
• Remove unnecessary second fetch of structured_outputs.
• Update recorded HTTP API interactions and snapshots.
• Add tool support test.

This resolves:

• support tool calling #38
• adds tool support #39

Note that streaming tool calls to certain models may still fail in version of llm up to 0.26 [edit: fixed in llm 0.27] due to:

• Streaming tool calls with no arguments don't work with Qwen 3 llm#1191

Note:

This is an alternate, smaller PR equivalent to #42. Compared to #42, the cassette recordings now only include a subset of models supported by OpenRouter, instead of the entire 400k response.

[simonw]

Testing this with:

llm -m openrouter/openai/gpt-5 'current version' -T llm_version --td

It output:

Tool call: llm_version({})
  0.27.1

LLM version: 0.27.1

[simonw]

And a more complex example:

llm install llm-tools-datasette
llm chat -m openrouter/openai/gpt-5 --td -T 'Datasette("https://datasette.io/content")'

Then in the chat:

Show tables

Then:

Show a count of rows in each of the views

Which worked - transcript here: https://gist.github.com/simonw/4b7c889ac6ed22b7d691f4981dedbe05

[simonw]

@jamessanford this is excellent, thanks very much for this.

[simonw]

Here's a model that does not support tools:

llm -m openrouter/nousresearch/hermes-2-pro-llama-3-8b hi -T llm_version

Output:

Error: OpenRouter: openrouter/nousresearch/hermes-2-pro-llama-3-8b does not support tools

So that works as it should.