@local/dsh-skill-scout

Skill scout

A DSH plugin adding a Collection tab to Settings → Plugins, one step after Skills. It is the collection of skills already found and judged, and it is what makes the search accumulate: the scout reads it before it searches, so a skill found once is never hunted for again.

Nothing enters the store without a verdict. A row with no judgement on it would be trusted as though it had one, and the whole reason for keeping a store is that the scout can skip a search on the strength of what is in it.

Every entry carries the date it entered, because a store with no dates is a pile of things someone once thought were interesting.

The pipeline

a task        anyone — a person or an agent — says what to go and look for
   ↓
the scout     searches the skill roots and GitHub, ranks what comes back
   ↓
triage        cheap: structure, size, whether it can run here at all   ← most stop here
   ↓
the hearing   prosecutor, defence, judge — only for the survivors
   ↓
the catalogue one entry: link, description, two axes, verdict, rating, what was taken
   ↓
Add           a person decides. Nothing is adopted by the pipeline.

Why the tab does not search

The obvious design is for the panel to run the search when the button is pressed. It is the wrong one, for two reasons.

It would put a credential and an outbound call behind a browser button. Searching needs a GitHub token and the network. Both belong to the agent that already holds them; neither belongs in a route the page can reach.

And it would not finish the job. Judging needs a reader, and the reader is an agent. Half a pipeline behind a button and half in a person is worse than a queue.

So pressing search writes a task. An agent takes it, runs the scout, triages what comes back, puts the survivors on trial, and writes the catalogue. The tab is a board both sides can see and write to.

Two axes, not one score

axis	the question
Worth keeping	the verdict of the hearing
Runs here	whether it can be used in this Harness at all

They are separate because they diverge. One of the three skills judged so far was a gate that hooks a mechanism this Harness does not have: a good idea, competently made, and unusable here. A single score would have hidden exactly the distinction the catalogue exists to keep.

The rating carries its case count

Every entry shows the hearing's rating and how many hearings are behind it, because a bare number reads as calibrated. Three hearings is not a scale, and the entry says so rather than implying otherwise.

What it writes

Two files in the Harness home, both owned by this plugin:

file	what
skill-scout-queue.json	the tasks: what was asked, who asked, what came back, and the counts
skill-catalogue.json	the entries: what was found, what was decided, and on what evidence

The skills repository carries a rendered snapshot of the catalogue as Markdown, written by the agent when the catalogue changes. It is a record and is named as one, so there is no second live copy to drift.

Refreshing, and what needs what

what changes	what it needs
a task, a claim, a verdict, an entry, adoption	nothing — written and read live
installing this plugin for the first time	the page reloaded (F5)
updating an already-installed copy	dsh web restarted

What the tab does not do

• It does not search. An agent does, and that is a design decision rather than a missing feature — see above.
• It does not adopt. The Add button is a person's; a pipeline that adopts its own findings can install something nobody chose.
• It does not judge. The hearing is a procedure in a skill, not code. The host half only stores what the hearing produced.
• It does not touch the skills themselves. It is a catalogue of what was found elsewhere, not a manager of what is installed. That is the Skills tab next door.