wiki-ideate

Run the paper-wiki wiki-ideate action. This file adapts the action to a Claude Code slash command.

1. Read WIKI.md, then research.md and the current wiki/{papers,concepts,gaps} inventory.

2. Stop for a course wiki; novelty-gap ideation is research-only.

3. Resolve the argument to one of three paths — goal-focused, gap-focused, or exploratory. Ask the user when a name matches more than one file.

   • goal-focused: the argument names a wiki/field/problems.md candidate line, or an already-adopted goal card under wiki/field/problems/<goal-id>.md.
   • gap-focused (existing): the argument names a wiki/gaps/<id>.md file.
   • exploratory (existing): no argument.

   The goal-focused / exploratory split is a written-down criterion, not a feeling: attempt Q1 (extensional identity — can the opponent's already-reported result satisfy one instantiation of our main facet?) and Q4 (F11 standard — does the opponent's explicit conclusion sentence claim to have solved it?) against the candidate's problem statement, after a real retrieval-and-adjudication attempt. If both questions return a determinate result (problem-open, problem-solved, or partially-solved with criteria_delta), the statement "can be adjudicated for openness" — take the goal-focused path (step 5 below). If neither question can return a determinate result (e.g. the problem is too vague to name what "one instantiation of the main facet" even is), the statement "cannot be written out" — take the exploratory path instead: open a direction card, let a mechanism trace emerge from probing, and come back later.

   Late admission (补票) is explicitly legal: an exploratory direction card may, once a mechanism trace appears, retroactively adopt a goal. But late admission must run the full four-way diagnosis + openness adjudication in step 5 below — it is a hard constraint, not a suggestion, that it is forbidden to shortcut or waive any step because probe data already exists（不得因已有探针数据减免任何一步）; probe data is reference-only and never itself constitutes adjudication evidence. This is the path that gives E3 ("idea 起于发现") its legitimate standing — it is no longer a residual quota, it is a named route.

4. If fewer than three papers are compiled, warn that coverage is too thin and ask before continuing. Suggest wiki-compile or wiki-search-latest first. In an auto run (/wiki-auto-driven), do not ask: stop this ideation cleanly and append one INBOX row (coverage too thin).

5. Goal-focused adoption flow (runs only when step 3 resolved to the goal-focused path; a late-admission/补票 candidate re-enters here and must run every sub-step below in full — see step 3's 补票 clause).

   i. Four-way diagnosis (四分诊断) — judge why the gap exists: gold-mine / dead-end / tool-missing / already-reworded (换词已做). The four categories are non-exclusive; multiple tags plus one primary cause are allowed. Check hidden-done first — semantic search across vocabulary; search_record is upgraded from "log what was searched" to search-then-record（先搜后记）: the search runs first, the record follows, and even a zero-hit round still lands its query list as a search_record. Check tool-missing against the candidate's potential attack surface. The worker produces only a symmetric two-sided evidence brief (one column "why it matters", one column "why it isn't worth it" — never a single-sided brief): hiding a taste judgment inside an enum is the least honest move this design could make — the same architectural discipline as "no kill verdict" in the review layer and "a human reads the criterion directly" in the probe layer.

   ii. Human signs gold-mine or dead-end. In attended mode this verdict belongs to a human, never to the agent. In an auto run (/wiki-auto-driven), the agent renders the verdict itself and fills the column, signed 机签 (auto + 日期), with one wiki/INBOX.md row — a human may overrule it afterwards by normal editing.