fix: implement server instructions tool

[ant-ms]

The README reads as follows:

DJANGO_MCP_GET_SERVER_INSTRUCTIONS_TOOL (default=True) if true a tool will be offered to obtain global instruction and tools will instruct the agent to use it, as agents do not always have the MCP server glboal instructions incldued in their system prompt.

However, I could not find any such functionality in the whole codebase. Maybe it was forgotten? If I just missed, it somehow, please let me know and close this PR.

This PR fixes this by checking the DJANGO_MCP_GET_SERVER_INSTRUCTIONS_TOOL variable and if it is set to True (which it is by default) it will create a "get_server_instructions" tool which returns the server instructions, just like the README indicates.

As a result the agents will check this tool before doing anything:

PS. I could not find any contribution guidelines in this project, let me know if I missed something and I'll adapt.

[nejch]

Hey @omarbenhamid just checking in on this PR as it would help us get server-wide instructions deployed. Is this project still actively maintained after the initial implementation? Thanks and sorry for the ping 🙇

[omarbenhamid]

Hello,

Thanks for the reminder.

The global instructions have been abandoned as most agents (by the time) needed to be explicitly hinted to read them, and they did not do by default.

@ant-ms the PR seems ok except for the error handling : by default the good practice is to let the exception raise up.

Sorry for the random maintainance issue : it is easier when client projects give me time on, but it gets tighter when no projects have requirements on the topic.

[ant-ms]

@omarbenhamid Thanks a ton for your time, I appreciate it.

The global instructions have been abandoned as most agents (by the time) needed to be explicitly hinted to read them, and they did not do by default.

I guess some of the models have gotten better. Testing this locally in our projects, the newer/smarter models seem to, mostly reliably (around 80% of the time), call the tool as instructed, and most of the ones that didn't succeed at first, query it after failing another tool call.

This is still not a perfect score of course, but I suspect, that as models continue to get better this will improve. The alternative would be to say that this score is not good enough to be upstream, in which case I would suggest removing this feature from the docs too to avoid confusion.