Autonovel consistency pass

Autonovel — Phase 3b.1 (Consistency pass)

Catch cross-chapter inconsistencies (character voice drift, plot holes, world-rule violations, repeated phrases) that the per-chapter drafting loop can't see.

When to use this skill

• State.json shows "phase": "consistency" (set by autonovel-revision when 3a complete)
• User says "do a consistency pass" / "check for plot holes" / "look for inconsistencies"

Prerequisites

!`cd "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" && ls chapters/*.md | wc -l`
!`cat "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/state.json" | python -c "import json,sys; s=json.load(sys.stdin); print(s.get('phase'), s.get('revision_cycle'))"`

Need ≥2 chapters drafted; otherwise consistency-pass is meaningless.

Workflow

Step 1 — Build rolling chapter summary

If story_state/facts.json exists, drafting has already been appending each accepted chapter's summary to consistency_state/rolling_summary.md live (see autonovel-drafting/references/fact-extraction.md). In that case, read the existing consistency_state/rolling_summary.md first, note which ## Chapter N headers are already present, and only fill gaps — chapters with no header (drafted before story-state was initialized, or the file is missing entirely). Do not regenerate a summary that already exists. If story_state/facts.json does not exist, or the summary file is missing entirely, build every chapter's summary from scratch as below.

For each chapter needing a summary:

1. Read chapters/ch_NN.md.
2. Generate a ~200-token summary covering: what happens, which characters appear, what state changes (relationship dynamics, possessions, knowledge), any world-rule invocations.
3. Append to consistency_state/rolling_summary.md with header ## Chapter N.

This is the rolling summary the windowed reviewer uses. Built once (or gap-filled); reused across windows.

Step 2 — Window the manuscript

Use lib/chunk_text.py:

import sys
sys.path.insert(0, "${HERMES_SKILL_DIR}/lib")
from chunk_text import chunk_by_chapters

# Read all chapter files in order
import os, glob
chapter_files = sorted(glob.glob("$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/chapters/ch_*.md"))
chapter_texts = [open(f).read() for f in chapter_files]
chunks = chunk_by_chapters(chapter_texts, group_size=4)

Each chunk holds 4 consecutive chapters (last chunk may be 1-3).

Step 3 — Per-window critique

For each chunk, build prompt:

• references/rolling-summary-template.md (this skill's references) — explains the task

• The chunk text (4 chapters)

• The rolling summary built in Step 1

• World, characters, canon references (small, just the rules-relevant parts)

• genre_context/story-contract.md — the novel's persistent reader promises, combination decisions, and required genre payoffs

• genre_context/compiled/evaluation-manuscript.md — the bounded selected-pack rules for manuscript-wide promises, guardrails, exclusions, rating, and content-tag coverage. Preserve each packet Source: citation in any finding based on a pack rule.

• The profile checklist from state.json: profile.tags' -tag exclusions, profile.rating, and profile.content_tags, so the reviewer can check excluded-tag content, rating exceedance, and content-tag coverage alongside the other categories (see ../autonovel/genres/README.md's -tag exclusion rule and "Content tags" rule, and ../autonovel/references/ratings.md). Content-tag coverage is manuscript-wide, not per-window — the reviewer notes per window which content tags a window delivers; the undelivered check itself runs once at merge (Step 4), against the outline's coverage map.

• STORY STATE for the window (if story_state/facts.json exists — see ../autonovel-drafting/references/fact-extraction.md for how the ledger is populated). For a window covering chapters a..b, run:

  !`python3 ${HERMES_SKILL_DIR}/lib/story_state.py query "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/story_state/facts.json" {a-1} <subjects>`
  

  For the first window (a=1), query chapter 0 — there are no facts stamped before chapter 1, so this returns the seed/empty state. <subjects> is the entity keys of every character appearing anywhere in the rolling summary for chapters a..b (fall back to the full cast if that's short). Budget: the ENTERING THE WINDOW block stays <= 1.5k tokens. If larger, restrict subjects to characters appearing in the window's rolling-summary entries only (never the full-cast fallback), and if still larger, drop facts for minor characters (fewest facts first) and note the omission in the window prompt. Never inject an unbounded ledger. Label this block "STORY STATE ENTERING THE WINDOW (as of end of chapter {a-1})" in the prompt.

  This is the N-1 convention documented on story_state.py's current_facts(): a chapter's own transitions are stamped AT that chapter, so querying at a-1 gives the state the window's prose actually starts from — not a state that already bakes in the window's own changes. Tell the reviewer explicitly, in the prompt: transitions the window's own prose establishes will NOT appear in the ENTERING block (they're stamped inside [a, b]) — prose showing a change of state mid-window is the story moving, not a contradiction. Only flag a ledger contradiction (see the rubric's new category) when the prose conflicts with the ENTERING block and the window offers no transition explaining the change. If story_state/facts.json does not exist, skip this block entirely — the window is reviewed without ledger cross-checking (pre-Module-A-ledger novels).

The agent reads the chunk and emits a per-window critique JSON to critique/consistency_window_<start>_<end>.md. The critique has six categories per references/consistency-rubric.md:

• character-voice drifts (cite chapter:line)
• plot inconsistencies
• repeated phrases (citing both occurrences)
• world-rule violations
• profile and genre-contract violations (unresolved genre promises, required payoffs, selected-pack guardrails, excluded-tag content, or rating exceedance; cite chapter:line and the contract section or evaluation-packet Source:; plus, at merge, any profile.content_tags entry undelivered anywhere in the manuscript)
• ledger contradictions (chunk text contradicts an open fact from the STORY STATE ENTERING THE WINDOW block; cite fact id + chapter:line — see the rubric's N-1 reminder before flagging)

Step 4 — Merge windows

Combine all per-window critiques into a single critique/consistency.md. Deduplicate items mentioned in multiple windows (same character drift seen across chunks).

If profile.content_tags is non-empty, run the manuscript-wide content-tag coverage check here (see references/consistency-rubric.md's Profile violations): using the outline's content-tag coverage map as the source of truth, confirm every entry is delivered in at least one chapter (integrated, not merely named). Cite the delivering chapter for each; any entry with no delivering chapter is an undelivered-content-tag finding under profile violations, handed to autonovel-revision like any other.

At the same merge point, check every reader promise and required genre payoff in genre_context/story-contract.md against the whole rolling summary and all window findings. Record unresolved genre promises under profile violations, citing the contract section and the chapter where delivery was expected (or the manuscript ending if no delivery exists). Apply every manuscript rule in genre_context/compiled/evaluation-manuscript.md; when a finding comes from a selected pack, retain its packet Source: path alongside the chapter:line citation.

Emit relationship graphs for the user. If story_state/facts.json exists, run, for each act boundary in outline.md:

!`python3 ${HERMES_SKILL_DIR}/lib/story_state.py graph "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/story_state/facts.json" <last-chapter-of-act>`

and write the output to story_state/graph_act_N.md (N = 1, 2a, 2b, 3 for the outline's four acts — Act I, Act II Part 1, Act II Part 2, Act III), wrapped in a mermaid code fence: open with ```mermaid on its own line, close with ``` on its own line. Determine <last-chapter-of-act> from outline.md's ## Act Structure percentage marks and each chapter's % mark field: Act I's boundary is the last chapter with a % mark <= 23%, Act II Part 1's is the last chapter <= 50%, Act II Part 2's is the last chapter <= 77%, and Act III's is the manuscript's final chapter. If a chapter's % mark is missing or unparseable, approximate act boundaries by chapter count instead (act boundaries at 23%/50%/77% of chapters_total, rounded). These graphs are for the USER and human proofreaders to review relationship arcs at a glance — they are never injected into any prompt. If story_state/facts.json does not exist, skip this step silently (pre-Module-A-ledger novels).

Step 5 — Hand to autonovel-revision

Update state.json phase back to "revision", set revision_source: "consistency" so the revision skill knows where to target fixes. Invoke autonovel-revision.

Step 6 — Cycle limit

Increment consistency_cycle in state.json. If consistency_cycle >= 2, instead of looping, set phase = "export" and hand off to autonovel-export. (Hard limit prevents endless polishing.)

Handoff

autonovel-revision for fixes, then either back here (cycle < 2) or autonovel-export.

See also

• references/rolling-summary-template.md — the per-chapter summary prompt
• references/consistency-rubric.md — what counts as a finding, what doesn't