🇬🇧United Kingdom @andrewbelcher

Account created on 21 November 2009, over 15 years ago
#

Merge Requests

More

Recent comments

🇬🇧United Kingdom andrewbelcher

Oh, and I also added the ability to apply restrictions via the tool explorer, in line with what can be configured on agents in AI Agents. That makes it a lot easier to test!

🇬🇧United Kingdom andrewbelcher

I believe that's all tests fixed - failure was a runner issue.

I've slightly tweaked it to not automatically modify the required property. That can still be done by things that modify the definition though.

🇬🇧United Kingdom andrewbelcher

I would actually describe this as a bug, and a have upped it to major as there is no reasonable workaround that I can see.

🇬🇧United Kingdom andrewbelcher

So to get the LLM to respect this, I had to put reiterate that it's a constant multiple times by:

  • Setting const
  • Setting enum with a single value
  • Adding a note to the description.

But it was consistently behaving after that!

I also adjusted the way required constraints are managed so it's computed at the point of render rather than adding the property, so it respects and changes. I've also made isRequired always FALSE if it's a constant.

🇬🇧United Kingdom andrewbelcher

Original issue in AI was a duplicate of 📌 Add module dependencies to tools Active which has been done!

🇬🇧United Kingdom andrewbelcher

Ah, just realised this is a duplicate of a ticket Marcus has already fixed - 📌 Add module dependencies to tools Active .

🇬🇧United Kingdom andrewbelcher

Steps given in the initial report.

The isAvailable method only appears to be called in ai_agents_extra's MCP plugin, and the ai_content_suggestions settings form.

That probably should be checked in other places, but is more suitable for "run time" checks, rather than dependency checks. For example, if a function call tried to inject a dependency from a module not enabled, you wouldn't be able to instantiate the plugin to call isAvailable.

🇬🇧United Kingdom andrewbelcher

PR also includes some fixes to type hints/usage etc.

🇬🇧United Kingdom andrewbelcher

andrewbelcher made their first commit to this issue’s fork.

🇬🇧United Kingdom andrewbelcher

andrewbelcher made their first commit to this issue’s fork.

🇬🇧United Kingdom andrewbelcher

I think this should actually be against the AI module. OpenAI provider (along with most other providers) do make use of key.

However AI, which is where the explicit dependency is, doesn't directly use it. In the case of the AWS Bedrock provider, key is never used as it makes use of the AWS module's profiles to manage the credentials.

🇬🇧United Kingdom andrewbelcher

PRs ready for both 1.0.x and 1.1.x.

🇬🇧United Kingdom andrewbelcher

Looks good now, thanks Marcus!

🇬🇧United Kingdom andrewbelcher

Initial prompt addition in MR, but I think we do need to look at the rest of the task, especially one shot learning examples.

🇬🇧United Kingdom andrewbelcher

Struggling to reproduce the error, but the PR looks good and should give safety against the error.

🇬🇧United Kingdom andrewbelcher

But how would I know how to do this as a site admin? During the process of installing drupal cms and going to install the ai recipe, I never see the place to manage my privacy settings.

The only way I knew to do this was by looking at the code. My privacy settings didn't appear on the admin pages, it didn't appear in my profile. I had to look quite hard to find it.

🇬🇧United Kingdom andrewbelcher

It looks like it might be related to how Klaro! handles whether consent for itself has been granted. When I manually open the Klaro! consent manager and save with the default settings:

Once that has been saved, a klaro cookie is set, and from that point on, the chat shows the Always option:

🇬🇧United Kingdom andrewbelcher

Can confirm this is the correct approach for 10.3 as well.

🇬🇧United Kingdom andrewbelcher

Still need to check D10 compatibility.

Production build 0.71.5 2024