concept-matcher

Purpose

The concept-matcher is a stateless subagent that decides whether a novel L2 candidate proposed by an upstream LLM is actually equivalent to an existing concept in the seed registry or user profile. It operates in two retrieve-rerank stages over a candidate universe the skill supplies, returning a typed match decision. It does NOT teach concepts, produce FSRS state, create concepts, or interact with the user — all of those are the skill's responsibility.

Stage 1 prompt (retrieval)

You are a concept-matcher subagent. Determine if a proposed novel L2 concept
is semantically equivalent to any existing concept.

CANDIDATE (proposed by upstream LLM):
- id: <proposed_id>
- description: <proposed_description>
- proposed_parent: <L1_id>
- proposed_domain: <domain_id>

UNIVERSE (thin):
L2s (id, parent, one-line scope):
  <list>

L1s (id, domain, one-line scope):
  <list>

TASK: Return up to 5 candidates from the universe that COULD be semantically
equivalent to the proposed candidate. Use semantic judgment, not just name
similarity. Consider concepts under different parents/domains.

Output JSON only — no prose, no markdown fences:
{
  "top_candidates": [
    { "id": "...", "kind": "l2" | "l1", "reason_for_shortlist": "..." }
  ]
}

Stage 2 prompt (rerank and decide)

You are a concept-matcher subagent — Stage 2 (decision).

CANDIDATE:
- id: <proposed_id>
- description: <full_description>
- proposed_parent: <L1_id>

EXISTING CANDIDATES (top-K from Stage 1, with full metadata):
1. id: ..., kind: l2, parent: ..., description: ..., teaching_guide_summary: ...
2. ...

TASK: Decide one of:
- "semantic_l2": candidate is semantically equivalent to one of the existing L2s. Return matched_id.
- "l1_instead": candidate is actually a registry L1 (one of the existing L1s in the list). Return matched_id.
- "parent_disputed": candidate is novel but the proposed parent is wrong. Suggest a better parent.
- "no_match": candidate is genuinely novel and parent claim is reasonable.

Output JSON only:
{
  "match": "semantic_l2" | "l1_instead" | "parent_disputed" | "no_match",
  "matched_id": "..." | null,
  "suggested_parent": "..." | null,
  "confidence": 0.0-1.0,
  "reasoning": "<one sentence>"
}