pre-publish

/pre-publish

You are running the Pre-Publish Gate. Every complaint a reader has ever logged in reviewer_complaints.md becomes a scan here. The book does not ship until this passes.

Input

• $ARGUMENTS: Optional — scope. Default is the whole manuscript (manuscript/chapters/*.md sorted).

Procedure

Step 1: Load the complaint registry

Read reviewer_complaints.md. For every entry that has a Scan method field, extract:

• Category
• Short complaint summary
• The exact scan to run

If there are fewer than 3 real complaints logged, note it to the user but proceed — this skill also runs the baseline scans.

Step 2: Enumerate chapter files

ls manuscript/chapters/*.md

If the list is empty, stop and ask the user where chapters live.

Step 3: Run baseline scans on every chapter

For each chapter:

1. /prose-scan — dialogue placeholder (CRITICAL), sentence-opener variance, burstiness, echo, sensory ratio, transition velocity, dialogue beats.

   • Invoke: bash .claude/scripts/prose-scan.sh manuscript/chapters/<file>
   • Record: CRITICAL / HIGH / MEDIUM counts per chapter.

2. /de-ai-audit — structural AI tells and vocabulary clusters (Claude-side, per chapter).

   • Record grade and any hits.

Step 4: Run manuscript-level scans (review-engine contract)

These run on the versioned review-engine contract (reference/MODEL_SELECTION_GUIDE.md §5), not the sunset Gemini CLI. Engine order: codex (PRIMARY, off-session-safe) → mcp__antigravity__agy_ask (in-session fast path) → agy CLI fed inline text (fallback) → loud SKIPPED/visible-fail per tier below.

Failure semantics are tiered — this is the gate's reliability contract:

• Continuity is REQUIRED-tier. It degrades codex → agy-inline → and if no engine can run, it is reported UNVERIFIED (BLOCKED), not SKIPPED-and-passed. A required full-manuscript check that cannot run is a visible non-pass — never a silent green.
• Voice and Logic are OPTIONAL-tier. If no engine runs, mark them SKIPPED and let the gate pass on the rest.

Preflight — probe in contract order:

bash -c 'source .claude/scripts/lib/cli-check.sh; cli_available codex || cli_available agy'

In-session, prefer the manuscript-review workflow (Workflow({ name: "manuscript-review", args: [<chapter paths>] }) after globbing manuscript/chapters/*.md) — it runs codex + agy (via the antigravity MCP) per chapter and aggregates with context/CITATIONS.md grounding flags. Off-session / CLI path, run the three checks against whichever engine probed available. Route output to a file, then Read it — large output gets truncated/swallowed inline:

• continuity-audit → reviews/continuity.md
• voice-lint → reviews/voice.md
• logic-check → reviews/logic.md

1. continuity-audit (REQUIRED) — state attributes, name introductions, knowledge graph.
   • codex: mkdir -p reviews && codex exec --sandbox read-only "<continuity prompt>; read manuscript/chapters/*.md in order" > reviews/continuity.md 2> reviews/continuity.err
   • agy fallback (inline only — never let agy open files): TEXT=$(cat manuscript/chapters/*.md); timeout 30m agy --dangerously-skip-permissions -p "Below is the manuscript inline (do not open files). <continuity prompt>. CHAPTER TEXT: $TEXT" → writes reviews/continuity.md.
2. voice-lint (optional) — character voice differentiation against docs/characters.md.
3. logic-check (optional) — plot logic and timeline.

Citation grounding (fail-closed). Cross-check every specific factual claim flagged by the workflow (needs_citation_check) against context/CITATIONS.md. Any release-critical specific claim left pending/disputed in the ledger is a Blocker — it halts the ship, it does not merely annotate the report. This is the gate against the twice-shipped fabrication bug; never trust a reviewer's "VERIFIED."

4. codex-review (optional editorial second opinion) — probe first with bash -c 'source .claude/scripts/lib/cli-check.sh; cli_available codex'. If available, run /codex-review for a full-book editorial pass (structure, pacing, voice, line-edit craft); its output is written to reviews/codex_review.md by the script. Codex's CRITICAL/HIGH findings feed the Blockers / High-priority buckets below; its MEDIUM/LOW feed the lower sections. If codex is unavailable, skip this step cleanly (note it as SKIPPED, same as the Gemini scans).

Step 4.4: Verify every AI-auditor finding before you trust it

This applies to every finding from the review engine (continuity/voice/logic) and codex.

• Quote-verify before action [non-skippable]. An LLM auditor can fabricate a citation — flag a "contradiction" against a quote that doesn't exist in the chapter, or call an internally-consistent passage a continuity break. Before a finding goes in the Gate Report (and certainly before anyone fixes it), open the cited file and confirm the quoted text actually appears and means what the finding claims. A finding is a lead, not a fact. Drop or downgrade any finding that fails verification, and note it as "unverified — cited text not found."
• Two independent reviewers; divergence is signal. agy (Antigravity) and codex run on different engines and contexts on purpose — they routinely catch different real problems. When they disagree, investigate the delta; do not average it. A finding only one of them raised is not automatically noise (it is often the real catch); a finding only one of them cleared is not automatically safe.

Step 4.5: Mechanical artifact & integrity scans

These protect the shipped artifact and the manuscript-wide consistency that per-chapter prose scans never see. Each is backed by a script under .claude/scripts/; run it, route output to a file, and Read it.

1. Render check (the build artifact is the product). Build the epub and assert it renders correctly — prose scans pass green while the shipped file is silently broken.

   bash .claude/scripts/render-check.sh
   

   Asserts: TOC lists every chapter, 0 literal # Chapter headings leaked into the body, 0 stray <hr> rules, scene breaks present. Any failure is a Blocker (a buyer sees it even though the markdown looked clean).
2. Canon / number-name residue sweep. After any retcon or rename, the old values must be gone everywhere, not just in the chapter you rewrote.

   bash .claude/scripts/residue-sweep.sh   # reads build/residue_terms.txt: superseded -> replacement
   

   Expect zero hits on superseded strings and confirm the replacement strings appear where the topic is discussed. Any superseded-string hit is a Blocker.
3. Persona / editorial-token bleed sweep. Author-persona names and drafting tokens must never appear in the prose as in-world text.

   bash .claude/scripts/token-bleed-scan.sh   # reads build/persona_tokens.txt
   

   Greps prose for persona tokens and editorial markers (Scene N, Movement A/B, TODO, stray "since Ch 2"-style cross-refs). Any hit is a Blocker — it is a fourth-wall break, and a numeric/token sweep catches what prose-reading reviewers slide past.
4. Name-similarity check. Flag any two active cast names within a <3-letter edit distance (e.g. Brin vs Bren) — readers conflate them. Cross-check the cast list in context/FACTS_SHEET.md. Near-collisions are High-priority (resolve before publishing).

Step 5: Run registered complaint scans

For each complaint with a Scan method from Step 1:

• Execute the scan (regex, grep, review-engine prompt, or skill invocation as registered).
• Record: hit count, chapter locations, severity assessment.
• If the complaint says "Fixed in rev X", still run it — confirming the fix stuck is the whole point.

Step 6: Synthesize the Gate Report

Output this format:

# Pre-Publish Gate Report

**Manuscript:** [slug from book.yaml]
**Scan date:** [today]
**Chapters scanned:** [N]
**Total words:** [N]

## Gate Status
[PASSED / FAILED]

- Baseline scans: [PASSED / FAILED — details]
- Continuity (REQUIRED): [PASSED / FAILED / UNVERIFIED — no engine available = BLOCKER]
- Citation grounding (REQUIRED): [PASSED / FAILED — N ungrounded specific claims = BLOCKER]
- Voice: [PASSED / FAILED / SKIPPED — no engine available]
- Logic: [PASSED / FAILED / SKIPPED — no engine available]
- Editorial (codex): [PASSED / FAILED / SKIPPED — codex unavailable]
- Render check (artifact): [PASSED / FAILED]
- Residue sweep: [PASSED / FAILED — N superseded-string hits]
- Persona/token bleed: [PASSED / FAILED — N hits]
- Name-similarity: [PASSED / FLAGGED — pairs]
- Registered complaints: [X/Y passed]

## Blockers (must fix before shipping)

### CRITICAL from /prose-scan
[chapter, line, finding]

### Continuity failures
[from continuity-audit — verified per Step 4.4]

### Artifact / integrity failures
[render-check failures, residue-sweep superseded-string hits, persona/token-bleed hits — from Step 4.5]

### Logged-complaint regressions
[Any complaint from reviewer_complaints.md whose scan fired. These are things readers already caught once; shipping them again is worse than shipping a new bug.]

## High-priority (fix before publishing, not hard blockers)

[HIGH findings from /prose-scan, voice-lint drift, de-ai-audit C grades]

## Medium / Low (fix-if-you-have-time)

[Lower-severity findings, summarized]

## Numbers that changed since last publish

If previous run data is available in `build/pre_publish_log.json`, compare:
- CRITICAL count: [prev] → [now]
- Registered-complaint hits: [prev] → [now]
- Overall pass rate: [prev] → [now]

Step 7: Decision Gate

If any Blocker (CRITICAL, continuity failure, continuity UNVERIFIED with no engine available, ungrounded release-critical specific claim, or logged-complaint regression) is present:

**"Pre-publish gate FAILED. [N] blockers listed above. Fix them or e

---

Content truncated.