PW 3.0.258 – Core updates

[adrian]

Looking forwarding to trying out the memory feature @ryan 

Here's something of related interest: https://github.com/memvid/claude-brain

I do love this description of the issue:

You: "Remember that auth bug we fixed?"
Claude: "I don't have memory of previous conversations."
You: "We spent 3 hours on it yesterday"
Claude: "I'd be happy to help debug from scratch!"

[Soma]

I'm using github copilot now since 6 months to do all sort of testing and playing around with various projects. It offers a lot of the different agents, but it's mainly for use with IDE like VS code. I like it a lot and am always surprized at it's capability. Claude is able to help me A LOT updating a very old big online shop to the latest PW (currently local dev only) It has access to powershell to do db stuff and it understood PW very well without much help or special skills. Tho it's great to have it specialize for PW for sure.

So I can't use Claude models and some of the others with my github copilot API endpoint with AgentTools, but it allows me to use gpt-4.1 and gpt-4o for free which is cool for simple stuff like engineering and migrations. I'm testing it locally currently. I now also made an account at Antropic to use that as well in case I need it.

I have a local docker environment with php, mysql and apache running in a container. Claude had to adapt the pw-at.sh script quite a bit to make it work. I have no clue about this stuff, but it's working now. Thanks very much for the AgentTools module.

[ryan]

Support for sub-agents has been added in version 7, now posted to GitHub. Your primary agent can now delegate to other agents when it deems it worthwhile. They can be more instances of the same agent, or instance of other agent models you've defined in the module configuration. More details further down.

This version also adds a public API method for sharing agent configuration with other modules, which was requested by @psy. To get agent configuration use $at->getAgents(); (to get all) or $at->getPrimaryAgent() to get the primary agent. See the README file (near the bottom) for instructions on how to use it. Each of the returned $agent objects also includes an ask() method, for when you want AgentTools to handle the request/response process for you as well:

$agent = $at->getPrimaryAgent();
$answer = $agent->ask('What is the capital of France?'); 
echo $answer; 

Claude Code both requested and developed the sub-agents feature and I asked it to describe some examples of when/where it might be used:

Quote

 

TLDR: it turns the Engineer from a browser tool into a programmable ProcessWire specialist that any agent or script can call from the command line.    

The core use case is AI agent delegation — a calling agent (Claude Code, Codex, etc.) can spawn a ProcessWire-specialist sub-agent without having to explain ProcessWire itself. Instead of the calling agent needing to know ProcessWire's API, it just asks the Engineer and gets a concrete answer or a migration file back.                                                  

Some specific scenarios:
     
Answering ProcessWire-specific questions mid-task                                                         

An agent is working on a project and needs to know something about the site's structure before it can proceed — "does this site have a field called 'colors'?" or "what template does the homepage use?" Rather than running --at-eval and interpreting the raw PHP output itself, it delegates to the Engineer which gives a natural language answer.  

Creating migrations on demand                                                                                               

An agent is scaffolding a new feature and needs database changes — "add a text field called 'subtitle' to the 'blog-post' template." It runs --at-engineer-migrate "...", captures the migration file path from stdout, applies it, and continues. The calling agent never had to know how ProcessWire migrations work. 

Multi-agent pipelines                                                                                                                   

A higher-level orchestrator agent breaks a large task into steps and dispatches each ProcessWire-specific step to the Engineer as a sub-agent. The orchestrator doesn't need ProcessWire knowledge — it just knows how to call the Engineer and handle the output.                                                                                                 
CI/CD and automation scripts                                                                                                    

A deployment script that needs to apply site configuration changes as part of a release — the Engineer can be called programmatically to generate migrations based on a description, which then get applied automatically.