Assistant: Add 'Agent' (or non-Ask) mode; don't enable code execution for Ask

[jmcphers]

Currently, Agent's Ask mode has access to a code execution tool. Especially when used with Claude, the model loves this tool and will use it aggressively to answer your questions, even if it's unnecessary to do so.

This makes Ask work well if you are using it to ask questions about your data, such as what are the biggest 3 suppliers by volume? but it also makes it difficult to use Ask to generate code or edits, because it's so execution-happy.

Proposal:

• Remove the code execution tool when using Ask or Edit modes. Users can use these tools without worrying about the model trying to run code or mutate data.
• Add a new Agent mode that includes the Execute Code tool. Switching to this mode indicates intent to let the model run code and mutate data.
• (in the future) provide tools that the model can use to query state and run canned, safe R/Python code, even in Ask mode, so that some data questions can be asked without resorting to Agent.

Existing Ask/Edit switcher, where we'd add Agent:

[seeM]

There is already upstream support for an "agent" chat mode so this might work:

1. Add a new chatParticipant to the positron-assistant package.json with modes = ["agent"].
2. Add a corresponding implementation in positron-assistant participants.ts.
3. Update execute code tool filtering in participants.ts based on participant ID.

[jonvanausdeln]

Using 2025.07.0-120 on Windows 11, with Anthropic Claude 4 Sonnet.

• I was not able to get it to execute code in Ask nor Edit modes. It was very eager to execute code in Agent mode.
• It was not very eager to actually edit my file in Agent mode. I had to ask it explicitly to add the code to my file. This ultimately might confuse users who are used to Agent mode in VS Code, as it seems to behave differently.

I'm closing this issue as it meets all the criteria. As the new mode is used more "in the wild", individual issues can be opened.