Conversation
added 30 commits
March 31, 2026 14:37
…-tension Implements buildMoodDirective — detects when 3+ recent chapters share high-tension mood keywords (紧张/冷硬/压抑/tense/grim/etc.) and emits a structured directive forcing the writer to downshift to a quieter scene. Completes the moodDirective slot that was declared but left as undefined in the previous structured-directives commit.
Three changes: 1. FanficCanonImporter now extracts a writing_style section from source material — narration voice, sentence rhythm, dialogue markers, imagery preferences, emotional expression patterns, with source-text citations. 2. initFanficBook calls generateStyleGuide after canon import so fanfic books start with both style_profile.json and style_guide.md. 3. importCanon reads up to 5 parent chapters and generates a style guide for the target book. importChapters does the same on first run. These ensure continuation and fanfic books have style reference data available to the writer from the first chapter onward.
Adds StateManager.rollbackToChapter(bookId, targetChapter) which: - Restores the snapshot at targetChapter - Deletes chapter files, snapshots, and runtime artifacts beyond it - Updates the chapter index to remove discarded entries review reject now calls rollbackToChapter(chapterNum - 1) by default, undoing the rejected chapter's state mutations and any subsequent chapters that depended on them. --keep-subsequent preserves legacy behavior. review approve now re-snapshots the approved chapter to ensure a clean rollback target exists. This prevents bad drafts from permanently polluting truth files — the core "audit-then-commit" invariant that was missing.
Some API providers (e.g. openclaudecode.cn) require a specific User-Agent header for external clients. Without it requests get 403. Adds INKOS_LLM_HEADERS (JSON object) which is injected as defaultHeaders into the OpenAI SDK client. No effect when unset. Also adds an optional `headers` field to LLMConfigSchema.
added 24 commits
April 1, 2026 15:44
Three tests expected the old 7-column hook table format but the schema now includes a payoff_timing column between expected_payoff and notes.
Reverts 822708e, 4248e1d, eb288b3 which: - Softened all directive language from "must/do not" to "try/prefer" - Compressed writer context into compact Story Briefs - Removed selectedEvidenceBlock from governed mode - Extended compact briefs to auditor/reviser Control experiment showed these 3 commits are the primary regression suspects: iter-001 build + gpt-5.4 scored 75.5/66.5, while iter-002 (with these commits) scored 71.5/37.5 on the same model. Restores imperative directive language and full context sections.
…ount Long-span-fatigue warnings (mood monotony, title clustering, pacing repetition, opening/ending pattern similarity) are sequence-level properties that cannot be fixed by revising the current chapter's text. Previously these warnings inflated blockingCount in evaluateMergedAudit, causing the revision gate to trigger or reject revisions for issues that no single-chapter edit can resolve. Now blockingCount only counts chapter-level issues (LLM audit, AI tells, sensitive words). Sequence-level warnings still appear in the audit report but no longer drive revision decisions.
The settler may extract a chapter number from story content rather than using the pipeline's authoritative value (e.g. parsing year "1988" from a contemporary fiction novel as "第1988章"). This corrupts the manifest's lastAppliedChapter, causing all subsequent chapters to use absurd numbers. buildRuntimeStateArtifactsIfPresent now accepts an authoritative chapter number and replaces the delta's chapter field when it disagrees. Root cause of 旧城暗号v4's ch4→ch1988 numbering jump.
…oach
Removes the hook lifecycle/pressure scoring system (5-level timing,
pressure formulas, pressure map tables) and replaces it with:
1. Simple stalest-first sorting for mustAdvance (same as iter-001)
2. Hook seed excerpts in composer — for each hookAgenda target, finds
the original chapter where it was planted and the latest advancement
chapter, extracts the hookActivity text, and injects it as an
evidence block the writer can directly build on.
The writer now sees concrete narrative material ("original seed (ch2):
萧炎右手碰到戒面时，有一丝温热渗出") instead of abstract metadata
("pressure: high, movement: partial-payoff, reason: stale-promise").
This addresses the hook execution problem: writer knew WHICH hooks to
advance but not HOW. Seed excerpts provide the "how" material.
Combines 43 commits of autoresearch pipeline improvements with master's chapter-number anchoring fixes. Key features merged: - Hook seed excerpt system (replacing lifecycle pressure) - Audit drift isolation (audit_drift.md) - Cadence analysis (mood/scene/title fatigue detection) - State validation retry + degraded recovery - Review reject rollback - Title collapse auto-repair - Chapter number anchoring (contiguous prefix) - Sequence-level fatigue excluded from revision blockers
Post-merge regression: bootstrapStructuredStateFromMarkdown still included currentState.chapter in lastAppliedChapter, allowing a hallucinated chapter number in current_state.md to override the contiguous durable progress. Changes: - state-bootstrap.ts: derivedProgress and lastAppliedChapter now come exclusively from durableStoryProgress (contiguous artifact chain), currentState.chapter is no longer trusted - manager.ts: getNextChapterNumber returns durableChapter + 1 without consulting manifest - story-markdown.ts: hook startChapter/lastAdvancedChapter parsed with parseStrictChapterInteger (rejects "第141号文明" style prose) - Tests updated to provide durable chapter artifacts matching the chapter numbers they expect
Hook budget: planner intent now shows "N/12 active hooks" in the Hook Agenda section. When N >= 10, adds explicit guidance to prioritize resolving old debt over opening new threads. Soft hint only — no hard cap. Ending trail: composer now extracts the last meaningful sentence from the 3 most recent chapter files and injects it as context, so the writer can see "ch3: buried under rubble | ch4: buried under rubble" and know to vary the chapter ending structure.
Three changes: 1. Architect fanfic prompt now requires a "new spacetime": - Explicit divergence point from source material - Original core conflict (not retelling original plot) - 5-chapter ignition constraint - 50% fresh scene locations 2. New FoundationReviewerAgent reviews architect output before writing begins. Scores 5 dimensions (source DNA, new space, core conflict, opening momentum, pacing feasibility). Threshold: 80 total, 60 floor per dimension. Rejects with specific feedback, architect regenerates. Max 2 retries before accepting best effort. 3. Review loop integrated into initBook (original mode) and initFanficBook (fanfic mode). Both now go through generate → review → (reject → regenerate) cycle. This addresses the #2 problem (斗破同人 75.5): fanfic architect was retelling original plot because nothing required it to create new narrative space.
Writing pipeline overhaul driven by Meta-Harness autoresearch: - Foundation Reviewer with 5-dimension scoring and reject/retry loop - New-spacetime requirements for fanfic mode - Hook seed excerpt system (replacing lifecycle pressure) - Review reject state rollback - State validation recovery with settler retry - Chapter number anchoring to contiguous durable artifacts - Audit drift isolation - Title collapse auto-repair - Hook budget hints and chapter ending trail - Mood/title/scene fatigue detection Benchmark: from-scratch novels improved from 75 to 92/100.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Writing pipeline overhaul. 53 commits. Foundation Reviewer, Hook Seed Excerpt, Review Reject Rollback, Chapter Number Fix. Benchmark: from-scratch 75→92, fanfic 39→82+. See CHANGELOG.md for details.