feat: gate every agent prompt on projectType (fiction vs non-fiction)

[tdrml]

Summary

Gates all seven agent prompts on projectType so the entire pipeline adapts to NON_FICTION projects. Same shape across every gated builder: optional projectType param defaulting to fiction behavior, helper functions emit fiction or non-fiction blocks, shared scaffolding (header, file paths, completion, gh CLI, guard rails) stays identical.

This was driven by signal from the Phoenix Three book run (tdrml/auto-graph-books#30, repo tdrml/the-phoenix-three), where the fiction-tuned editor demanded scene-grade rendering on non-fiction prose constrained to public record — putting the reviser in an impossible bind that drove 4-5 review rounds per chapter. The editor gate (commit c0792c2 — first commit on this branch) addressed the immediate friction; this commit extends the same dispatch shape to outliner, writer, continuity-checker, auditor, fixer, reviser, and planner so non-fiction prose is generated and audited on its own terms from the start.

Per-agent changes (most → least)

• Outliner. Replaces the fictional Story Architecture (dramatic question + want/need/flaw/lie/ghost arcs + 8-point Climax Checklist + genre-aware predictability contracts + climax failure modes) with Reporting Architecture for non-fiction: central argument, structural anchors (founding decision / recognition payoff / pivot / near-failure averted), subject portraits drawn from public record only (no invented arcs on living people), research mandate (WebSearch/WebFetch), and non-fiction failure modes ("fabricated dramatization", "hagiography", "tell-all", "imagined interior on real people"). Character bible spec becomes a subject bible with public-record biographical facts and verbatim quotes; world bible becomes the institutional landscape; research notes spec demands a working bibliography and an "On Sourcing" front-matter draft.

• Writer. Replaces scene structure (objective/obstacle/escalation/turn) and "show pivotal moments / tell connective tissue" with paragraph-grade analytical structure (claim/evidence/implication) and explicit "where scene-grade rendering IS allowed" rules (only on documented moments). The AI-fiction-cliché watchlist swaps for the AI-non-fiction-tic watchlist actually observed in Phoenix Three editor reviews ("the kind of X", "What X taught/bought/added", inversion-aphorism patterns, hyphen-stack noun phrases, sourcing meta-commentary in third-limited, imagined interior on real people, invented named participants in real meetings). Climax handling → structural-anchor handling. WebSearch/WebFetch mandated in writing instructions.

• Continuity-checker. Issue types swap from CHARACTER / TIMELINE / WORLD_STATE / PLOT_THREAD to SUBJECT / TIMELINE / ENTITY / CITATION / SOURCING. Schema kept shared (the writer/editor pipeline reads a uniform shape) but field comments document the non-fiction mapping: characters → subjects, plotThreads → analytical threads, worldState → cited factual claims keyed by topic. Chapter summary becomes a factual summary keyed to claims rather than narrative.

• Auditor. Issue categories swap to CONTRADICTION / UNSUPPORTED_CLAIM / TIMELINE_ERROR / SUBJECT_INCONSISTENCY / ENTITY_ERROR / CITATION_ERROR. World-state schema example shows non-fiction semantics (publicTitle, source pointers in timeline entries, analytical threads).

• Continuity-fixer. Reconciliation principles shift: "the bibles win" → "the public record wins"; "never edit the climax for hygiene" → "never edit the structural-anchor chapter". Per-category fix guidance covers SUBJECT_INCONSISTENCY, ENTITY_ERROR, CITATION_ERROR (WebFetch verification required before applying), and UNSUPPORTED_CLAIM (hedge or cut; never fabricate provenance).

• Reviser (light). Reference materials reframe (subject bible, institutional landscape, working bibliography). Revision instructions add a SOURCING-aware diagnose-before-changing step distinguishing FABRICATION / CITATION / HEDGING / UNSOURCED SPECIFIC sub-classes. Self-check forbids new fabrication.

• Planner (light). "Climax convergence" dependency rule → "Structural-anchor convergence". Reference-materials and parallel-storyline rule reframe.

• Editor (already shipped in commit c0792c2 — first commit on this branch). Gate uses fiction criteria for NOVEL/NOVELLA/SHORT_STORY_COLLECTION; SOURCING-driven non-fiction criteria with FABRICATION / CITATION / HEDGING / UNSOURCED SPECIFIC sub-classes for NON_FICTION.

Handlers

Updated plan-chapters.ts, trigger-chapter.ts, revise-chapter.ts, continuity-check.ts, continuity-audit.ts, continuity-fixer.ts to pass project.config.projectType into their respective prompt builders. (generate-outline.ts already did; review-chapter.ts was updated in the editor-gate commit.)

Design choice: shared ContinuityState schema

Kept the ContinuityState JSON schema shared rather than forking. The non-fiction prompts use the same fields with documented non-fiction semantics in comments. Trade-off: field names (characters, plotThreads, worldState) read awkwardly for non-fiction, but forking the schema across writer/editor/reviser/auditor/fixer code that all read or write the state is a much larger coordinated change. Readability cost accepted; can revisit if a real run signals it's actually causing confusion.

Test plan

• pnpm vitest run packages/core/src/prompts/__tests__/prompts.test.ts — 203/203 pass (162 existing fiction + 41 new non-fiction-gate)
• pnpm vitest run — 1084/1084 actual tests pass
• pnpm --filter @auto-graph/core build — clean
• pnpm --filter @auto-graph/lambdas build — clean (only the pre-existing cloudwatch dep issue in reconcile-enqueuer.test.ts)
• CI green
• Post-merge CDK redeploy completes (no infra changes; pure prompt+handler updates, so should be a no-op deploy)
• Validate on a non-fiction kickoff (followup; not blocking on this PR)

Files changed

• 4 new commits' worth of changes in this branch:
  • c0792c2 — editor gate (3 files, 226 / 26 lines)
  • 2d6e3dc — outliner / writer / continuity-checker / auditor / fixer / reviser / planner gates + handler wiring + tests (14 files, 1685 / 256 lines)

🤖 Generated with Claude Code