The short answer
Action routing is a small component that looks at retrieved passages and suggests which Paqad workflow might apply. It only ever suggests, and every suggestion is marked as needing your approval. In paqad-ai 1.91.1 it is off by default, has no setting to switch it on, and is not called by any command or hook. It is not how Paqad decides which workflow handles your request. That is the router's job, described in Workflow routing.
What the component does
Retrieval (RAG, short for retrieval-augmented generation) finds passages in your project that relate to a task. See the RAG overview. Action routing takes those passages and the list of workflows your project has, then:
- Splits each workflow name into keywords. For example,
root-cause-analysisgivesroot,causeandanalysis. Words shorter than three letters are dropped. - Adds
securityandvulnerabilityas keywords for a pentest workflow when the task is rated high risk. - Counts how many passages contain any of those keywords, in their text or in the names of the functions and classes they export.
- Suggests each workflow that matched at least one passage, with a confidence equal to the share of passages that matched, highest first.
This is plain keyword matching done by a script. No AI model is involved.
How it guards against misleading text
Retrieved passages come from files in your repository, and a file could contain text written to steer an AI. The component is built so that retrieved text cannot take control:
- Every suggestion carries
requires_user_approval: true. Nothing is run automatically. - Suggestions refer to passages by ID, never by copying their text.
- The explanation attached to a suggestion is built from workflow names only.
- Only workflows that exist in your project's workflow list can be suggested.
Status in 1.91.1
The component sits inside a library-level context loader. In 1.91.1 no Paqad command or hook calls that loader, and the action_routing switch resolves to its default, off, because the configuration layer does not expose it. So on a normal installation it has no effect on your AI tool.
Older documentation described intent types called code_lookup, doc_lookup and cross_cutting, and a rag_intent override for workflow steps. None of these exist in paqad-ai 1.91.1.
What actually shapes retrieval per request
What does change retrieval from one request to the next is the router's decision, made when you send a prompt:
- A request routed to no workflow (small talk) gets no retrieval.
- Any other workflow gets retrieval when it is switched on, seeded with your prompt text.
- Only feature development, or a first prompt that has not been routed yet, gets the rule sections of the context file.
The search width then follows adaptive retrieval depth.