nodecode-skill-matcher
Skill Matcher: routes a task to the few installed cells and skills that fit it, with the evidence for each match. A Nodecode rewrite of xXLODXx/hermes-skill-router by xXLODXx.
Get it
Tell the agent "add the skill-matcher cell", or run nodecode add nc://skill-matcher.
Set it up
Nothing is required. The cell reads the cell folders under this home and any
SKILL.md beneath them, and works on what is there.
In /settings, the skill-matcher section takes:
| Setting | Default | What it does |
|---|---|---|
dirs |
empty | extra folders to scan for cell .asd files and SKILL.md files; empty reads this home's cells/ folder |
max_matches |
3 | how many matches one routing names; the rest are reported as rejected |
min_score |
3 | evidence a match needs: a tag alone is 3, a description alone is not enough |
log |
true | append each routing to skill-matcher/decisions.jsonl under this home, never the task's words |
Use it
Ask something like "which of your skills should handle this invoice-OCR job?", or say /skill-matcher extract the tables from these pdfs. The model can call the
tool itself:
match_skills {"task": "extract the tables from these pdfs", "max_matches": 3}
and reads back a short block:
## Skill routing - 2 matches for this task - ocr-kit (required; confidence 0.99; evidence: description, name) - ocr-kit/pdf-extraction (required; confidence 0.99; evidence: description, name, subword, tag) Load one with (help 'name).
With no match it answers No high-confidence match. Continue with the normal index.
and nothing changes.
How it ranks
The task's words against each entry's name (5), tags (3) and description (1).
A stem or compound overlap -- excerpt against excerpts -- is supporting
evidence and never carries a match on its own, which is what keeps a coincidence
out of the prompt. Confidence is the score over ten, plus 0.12 when more than
one kind of evidence agrees, capped at 0.99. A match needs exact evidence and
min_score; the entries over max_matches come back as rejections rather than
crowding the answer. Words that describe everything here (nodecode, cell,
skill, task, the, and, ...) route to nothing.
A task routed twice in one session says unchanged since your last routing
rather than repeating the block -- the original's topic-change check.
The log
skill-matcher/decisions.jsonl under this home gets one line per routing: the
names matched, how many were rejected, the highest confidence and whether it
fell back. No word of the operator's task is in it. Delete the file to clear it.
What of the original it keeps, and what it leaves out
The scoring, the evidence gate, the three-candidate budget, the fallback line, the per-session topic check and the local decision log are the original's. The catalog is this organism's own instead of Hermes's skill folder.
Left out: the automatic per-turn injection (the original ran on four Hermes hooks
and pushed its block into the first turn of every session; here the model calls
the tool, or the operator types /skill-matcher), the keyword-learning store
with its lift gate, and the dashboard page.
Test it
Run .github/test-cell.sh tools/nodecode-skill-matcher from the cells repository.
The tests write a fixture inventory into the run's own scratch home, so they
read no skill of the machine and dial nothing.
License
MIT, the original's. See LICENSE and NOTICE.
