Extract Concepts

# User-invoked prompt · extract-concepts (v1.8.2 · Obsidian-style)

> Replaces deleted `tools/extract.py` and `tools/lib/cortex/extraction.py`
> helpers. ROUTER reads this when user manually wants to run concept candidate
> extraction outside of archiver Phase 2.
>
> **v1.8.2 HARD RULE #11**: Concept candidates + extraction report render
> in Obsidian. Apply `references/obsidian-style.md` — concept entries use
> wiki frontmatter (kind: knowledge), report uses callouts + `[[wikilinks]]`.

## Trigger keywords

- `抽取概念候选` / `extract concept candidates`
- `跑 concept extraction` / `concept extraction`

## Goal

Tokenize a corpus of session text, count noun-phrase candidates, filter
stopwords + low-frequency, generate canonical slugs. This is the deterministic
first pass — the human or archiver subagent does LLM judgment on top.

## Steps

1. **Input**: ask user for source — typically:
   - "current session transcript"
   - "session file at meta/sessions/{sid}.md"
   - "all sessions in meta/sessions/{YYYY-MM}-*.md"
2. **Read + concatenate** all source text
3. **Tokenize**: split into 1-3 word noun-phrase candidates (alphanumeric + Chinese chars)
4. **Filter stopwords** (中英日 common words):
   - 中文: 是 / 的 / 了 / 在 / 有 / 我 / 你 / 他 / 这 / 那 / 都 / 也 / 就 / 还 / 但 / 等
   - English: the / a / an / and / or / but / is / are / was / were / be / been / has / have / had / for / of / with / this / that / it / on / in / to / from
   - 日本語: です / ます / これ / それ / あれ / という / こと / もの / ため
5. **Count occurrences** per candidate
6. **Filter `count >= 2`** (single-mention items are too transient)
7. **Generate slug** per candidate (deterministic, must be reproducible):
   - Lowercase
   - Replace spaces / non-alphanumeric with `-`
   - Chinese: pinyin transliteration if reliable; **fallback to SHA-1 hash of canonical name (first 10 chars)** to guarantee same name → same slug across runs
8. Sort by count desc, take top 20

## Output (no file write — print to user)