flow-check — 篇章連貫檢查 (discourse coherence gate)

LLM-generated prose is locally fluent but globally unplanned: each sentence is conditioned on its neighbours, while "does this paragraph's role in the chapter's argument hold" is maintained by no one. Incremental revision makes it worse — every 入章-style insertion splices fluent text between two sentences that were originally written as one continuous move, leaving a seam that is invisible sentence-by-sentence. This skill finds those seams and ranks them; it changes nothing by itself.

The five checks

Run 1–3 per chapter, 4 once across chapters, 5 first whenever git history is available (cheapest, highest hit rate).

1. 逆向大綱 (reverse outline). Compress every paragraph to one sentence: "this paragraph's claim is X." Then read only the outline. Flag: adjacent claims that don't connect, a claim unrelated to its section heading, the same claim appearing twice.
2. 銜接命名 (connective naming). For each adjacent paragraph pair, force an answer: "the implied connective between ¶N's end and ¶N+1's start is ___ (因此／然而／此外／例如／回到⋯⋯)." Cannot name one → seam. This is objective where "does it read smoothly" is not.
3. 冷讀 (cold read). A fresh-context reviewer reads the chapter text and nothing else, marking: places it had to re-read, terms/entities appearing without introduction, abrupt topic shifts. Isolation is the invariant: the cold reader gets no outline, no other chapters, no conversation history — a reader with full context cannot feel the jumps because their brain silently supplies the missing bridges. (This is also why the author agent must not grade its own chapter.) "Fresh" means spawned empty: a context-inheriting subagent (a forked conversation, a resumed session, a teammate that saw the drafting) carries the bridges with it and is disqualified — use a plain new agent whose whole prompt is the chapter text plus the marking instructions. A bigger context window changes none of this: the point is what the reader has not seen. A separate clean session is equivalent to a fresh subagent. If neither is available, prepare a handoff containing only the chapter and marking instructions, run the other checks, and mark the cold read pending; never call a same-context re-read independent.
4. 跨章一致 (cross-chapter consistency). One agent over the per-chapter outlines from layer 1 (not full text): terminology drift on core concepts (the manuscript's own coinages — is the same mechanism named the same way in every chapter?), theory-chapter concepts actually used in the case chapters, comparison-chapter dimensions matching case-chapter section structure, and every 「如第X章所述」-style reference resolving to something that exists and says what is claimed.
5. git 接縫定位 (seam targeting). git log --stat -- <manuscript files> to list late insertions and heavily-patched spots; check the ±3 sentences around each insertion point before anything else. Seams concentrate here.

Output: 跳躍點清單

One severity-ranked list, 章:節/段 → 類型 → 說明 → 建議橋接句:

Each 建議橋接句 is a proposal for the author to confirm — same scan-and-propose discipline as fix-terms. A bridge sentence must not alter any cited claim (house-style reference/content-integrity.md): if the only way to connect two paragraphs changes what a source is made to say, report the gap as structural instead of papering over it.

Orchestration

Run the chapters sequentially or use parallel agents when available and authorized; isolation of layer 3 matters, not the orchestration API. Do layer 4 over the collected outlines, then deduplicate and rank. For a single-file paper, treat top-level sections as the chapters. Findings from layer 5 seed the other checks ("check these spots first"), but must not be included in the cold reader's prompt because they would prime that independent assessment. State which layers ran and which remain pending in the report.

What it does NOT do

No fact checking, no citation form (cite-check), no whether-a-source- backs-the-claim (source-kit § Source review), no cross-strait usage (fix-terms — that lints against a fixed external rule list; flow-check lints the manuscript's internal consistency with itself). Run order for a full lint pass: fix-terms → flow-check → cite-check (cite-check stays last as the mechanical gate).

Backs the /check_flow macro.