Changelog

5.1.0-beta.7 — 2026-10-26

• Native Codex and Claude Code marketplace manifests, a reproducible allowlisted ZIP and pinned Codex installation instructions.
• Desktop-only metadata, support/privacy/usage pages and public-directory submission materials.
• Causal scene contracts, character knowledge boundaries, arc changes and role-specific editorial guidance.
• Complete accepted-source provenance and literal chapter handoffs; provisional summaries stay explicitly provisional.
• Source-selected literal repair evidence, lossless catalog reuse and backward-compatible review verification.
• Read-only activity reporting and separate export/literary completion states.
• Bounded chapter-length recovery, continuity issue batching and the preceding recovery engine fixes.
• Historical beta6: 24 chapters / 83,517 words / automated literary score 72.52/100. No beta7 90+ literary or bestseller result is claimed. See the integration report and release validation for exact tests and boundaries.

Unreleased — 5.1 recovery build

• Preserve the failed original native run and isolated revision replay as explicit acceptance evidence. Add a regression rejecting old reads after an uncited passage changes, and verify that the new guard blocks the exact stale reports from the replay. All 294 Windows tests passed. Acceptance report and remaining work.

• Preserve exact Unicode passages when review receipts pass through Windows shells by emitting ASCII-escaped JSON. Reject corrupted rubric passage receipts on reuse. A real saved report reproduced 44 corrupt characters with the old command and none with the new command; all 30 references survived. The full suite passed 293 tests. Windows evidence.

• Verify and tally anonymous comparison votes with a deterministic helper instead of model-written aggregation code. Retain neutral signals, reject changed or reused evidence, and keep ties pending separate per-version defect evidence. Six regressions and a tally of four original real-model reports passed; the full suite passed 290 tests. Decision evidence.

• Extend paragraph-bound evidence to anonymous A/B comparisons, preserving counterbalanced order, original votes and complete shared context. Resolve citations separately without exposing revision identity, and reject altered sources, incomplete catalogs or swapped version mappings. Four real Luna/high readers validated all 37 references on their first calls; 284 automated tests passed. Comparison evidence.

• Separate strictly factual corrections against existing canon from literary preference. A fresh reviewer checks the complete staged change and dependencies; source-bound paragraph evidence and the original native receipt authorize the same guarded promotion without invented reader votes. Creative rewrites retain the four-reader rule. Real positive/negative controls accepted a surname repair and blocked a changed witness account; 278 automated tests passed. Evidence and limitations.

• Save a prepared headless Codex role command in each launch receipt, retaining the exact executable, model and effort. The call guard accepts that recipe with an optional response schema and bounded timeout, instead of asking the orchestrator to reconstruct interactive CLI flags. Two real Luna/high canaries passed on their first calls; 268 automated tests passed. See transport evidence.

• Honor autonomous mode after an audit's major_rewrite: retain the failed verdict, choose a repair inside the existing brief and three-pass budget, and continue through truthful final status and draft delivery. Check-in mode still asks for the author's choice. The original real run paused here after six chapters; a replay from that preserved checkpoint is required to validate the new branch.

• Stage an initial chapter before canonical copying, with a narrowly recorded H2-H6 to H1 heading transformation. Preserve source, title and body bytes, record separate hashes and visible counts, and refuse malformed prose or existing attempts. Two real heading failures were resolved without new model calls or narrative edits; 263 automated tests passed.

• Extend complete, frozen passage references to current-draft reader panels. Prepare four individual packets and structured schemas, resolve IDs in separate receipts, and verify the saved catalog when reusing reader evidence. Legacy quote reports and anonymous comparisons keep their existing format. Four fresh Luna/high readers validated on their first attempt while retaining concrete negative findings; 258 automated tests passed.

• Return the active phase's exact installed playbook and reference paths with CLI progress checks. Keep the manifest as the source of routing and identify count/state checks as distinct from editorial verification. A native run had advanced after omitting the required architecture review; supplying phase instructions addresses the omitted-context path but is not proof that a review ran.

• Clarify the native handoff: the agent is already the writing session, ordinary commentary is its live terminal output, and the launcher has already verified the installed snapshot. Reuse the saved install record instead of manually transcribing hashes or trying to control the agent's own window.

• Preserve a failed receipt and partial response on a caught keyboard interruption, stop and reap the owned child, and keep ownership when tree shutdown is unverified. Connectivity canaries now have a short explicit timeout and use the host's normal network-approval flow; force-killed attempts remain explicitly incomplete.

• Freeze complete manuscript passages for rubric calls and validate source-bound passage references. Critics no longer need to retype quoted prose; raw reports remain unchanged, stale or invented references fail, and legacy exact-quote reports remain supported. Two real Luna/high calls passed evidence validation on their first attempt; this is not a literary-quality pass.

• Validate rubric evidence before aggregating scores: all ten dimensions, actual chapter coverage, exact excerpts, finite scores and source-bound receipts. Missing assessments stay incomplete; altered reports or prose cannot reuse an earlier pass. This checks evidence, not literary merit.

• Make revision scope follow the defect: correct dependent factual expressions together, repair missing payoffs at scene level, and reject still-broken candidates before comparison. Re-audit current text before closing tickets; try one materially changed strategy within the original pass budget before declaring a plateau.

• Trace the central question through evidence, choice and consequence in the adversarial audit. A repeated confession or symbolic ending is not by itself a paid-off story.

• Replace raw status dumps with a readable phase, manuscript, delivery, reader-evidence and saved-quality summary, plus --json for full reports. Keep inspection available across snapshot changes while still blocking unsafe resume; preserve checker failures and diagnostics.

• Add an optional native CLI launcher for Codex and Claude Code. Prepare and verify a project, collect its author credit up front, preserve checkpoints on resume, and hand the terminal to the unmodified host's own sign-in, live UI and permissions. No second editorial engine or subscription-token proxy.

• Forward exact native model and effort selections. For Claude Code, use invocation-scoped effort settings so saved environment values cannot silently replace an explicit selection; leave global settings unchanged.

• Give each launcher one book ownership record and truthful exit receipts. On Windows, assign the suspended native child to a private kill-on-close job before it starts, then verify descendant shutdown before releasing ownership. Atomic preparation and scoped rollback preserve other work.

• Supply saved current-reader indexes to status verification without regenerating reviews. A supervised six-chapter Codex run delivered verified Markdown and EPUB within its word contract; literary gates failed. Unattended, full-length and complete Claude-book acceptance remain unverified. See native CLI validation.

• Restore the 5.0 prose-first drafting policy with compact writer packets and four creative reader checkpoints. Preserve the current portable installer, native-call guard, critic isolation and staged revision.

• Review causal mechanisms and repetitive middle-book explanations through a small continuity contract and targeted repairs.

• Bind current reader coverage to the original validation receipts and canonical source hashes; changed chapters cannot inherit old passes.

• Exclude standalone hidden production comments from reading formats and deterministic visible-prose counts; render section headings instead of raw Markdown markers. Preserve source hashes.

• Commit technical delivery status with exports and verify the source-bound receipt, keeping delivery, literary quality and publication approval separate.

• Repair architecture dependencies together before reverification. Keep exact schedules and premise triggers strict, while distinguishing chapter spans and ordinary approximate durations from invented event timestamps or exact anniversaries.

• Reconcile canonical chapter numbers and deterministic prose counts against saved progress before resuming and during complete-book export. A stale state or pending earlier phase blocks the complete-delivery claim while preserving the manuscript; honest failed literary gates remain reportable outcomes.

• Check the approved premise, reader promise, timeline and point of view against the outline and continuity ledger before drafting; repeat the comparison in the final audit. A self-consistent plan must still honor the author's brief.

• Keep architecture-verifier packets neutral on retries: current constraints and story facts only, excluding earlier verdicts and repair history; record verification only after the reviewer actually finishes.

• Preserve full blind-reader instructions on retries and reject review evidence obtained by forcing empty defect arrays or positive votes. Structural JSON validity is not sufficient evidence of independent criticism.

• Guard existing native CLI calls with one active call per book, unique attempt directories, pending response publication and hashed completion receipts. A running call is waited on; it is not retried into the same output file.

• Add portable Markdown and EPUB export, optional DOCX/PDF reading proofs, deterministic prose counts, and complete-book artifact/length preflight. Exported files do not imply that literary gates or human publication approval passed.

• Preserve accepted chapters through hash-checked revision promotion and a shared cooperative book lock. Complete technical proofreading before final scoring; later corrections must refresh affected review evidence before export.

• Validate blind-reader JSON structure and exact quotations before counting votes, allowing only documented typography/inline-emphasis normalization. Propagate explicitly requested native model and reasoning settings.

• Separate evaluator criteria from orchestration targets; include install records in rollback and exclude Python bytecode caches from installed payloads.

• Add a recorded review-only continuity example, installer failure regressions, delivery integrity tests and native-call concurrency/termination tests. Complete-book validation remains a separate real-host acceptance test.

6.0.0-beta.1 (2026-09-22)

One skill, blind critics, bounded gates. Decisions in ADR 0012.

Added

• Blind reader, read-only auditor and a four-persona reader panel (primary, hostile, adjacent-genre, casual), with fixed verdict formats.
• Independence levels and grades for every verdict: a second tool from another model family, an isolated subagent, or the same context; a canary check where hosts do not document isolation; tested headless recipes for Claude Code, Codex, Hermes, Antigravity CLI and OpenCode.
• Two gates in the revision loop, the panel's and the rubric's, three passes each, with draft comparison: a revision is kept only when blind readers prefer it.
• Check-in with the author at every phase boundary, after chapter 1 and every fifth chapter; "go to the end" for autonomous mode.
• Patterns library inside the skill (references/patterns/), sourced, with machine-prose tells and ceilings in English and Brazilian Portuguese.
• Rhythm contract with one owner (resolves #35) and language-aware research (from #14, by Thinh Hoang).
• Continuity ledger with checks every five chapters and targeted repairs.
• Hook test in the editorial package; EPUB and PDF export when pandoc is present.
• Installer: targets and generate-agents commands; Claude Code subagents generated from role files and checked for drift; --agents-dest; copy-then-swap installs that leave the previous version intact on failure; backups that no host can load; automatic retirement of unchanged 5.x skills and V4 agents; clear errors instead of tracebacks; antigravity-cli target.
• Suite check that refuses skills pointing outside themselves, missing references, or naming retired skills.

Changed

• The core now ships specialists as references instead of sibling skills, so one folder is enough in any host.
• One PROJECT_STATE.yaml schema (schema_version: 6) and a renumbered artifact layout; the host contract migrates 5.x projects.
• Phase 5 is the Revision Loop; the unbounded Literary Barrier loop is gone.
• The writer never sees the rubric; critics never see targets or earlier scores.
• Home variables must be absolute; blank values fall back to the default for every target.
• README, landing page and docs rewritten around the goal, not a promise; the finished Vicente book is on the first screen.
• License holder aligned with the README: Felipe Lobo.

Removed

• book-genesis-full, book-genesis-codex, optional/, deprecated/, the eight V4 agents and knowledge/ (content folded into the core).
• humanizer and copy-editing, which were third-party copies without attribution.
• book-bestseller-studio, book-swarm-panel, book-researcher, book-editor, narrative-foundation, prose-craft, production-prep, series-architect, manuscript-manager as separate skills (now references).
• The unreachable phase runner in runner/filesystem.py and its tests; --include-legacy.
• Launch and marketing material, the social-posting scripts and sixteen outdated documents. Git history and the tag archive/imagination-edition-2026-09-22 keep everything.

Migration

• Reinstall with python runner/installer.py install <target>. Unchanged 5.x skills and V4 agents are moved to the backup folder; anything you changed is left in place with a warning.
• Open a 5.x book in a new session: Book Genesis renames the old state to PROJECT_STATE.v5.yaml, maps the artifacts to the new numbers, and shows you the mapping before continuing.

2026-09-10 — DeepSeek Harness and wider agent distribution

• Add DeepSeek Harness, Cursor, GitHub Copilot, Qwen Code, Pi, and Windsurf/Cascade: 15 install targets, comprising 14 named hosts and Shared.
• Respect DSH_HOME (including its blank-value fallback) and PI_CODING_AGENT_DIR; retain custom destinations, reference integrity, and conflict backups.
• Extend install/reinstall and wrapper verification to all targets; document official discovery conventions separately from native writing evidence.
• Add an English 30-day launch kit, platform drafts, demo scripts, measurement templates, and evidence gates. These are launch materials, not claims of a published video or social campaign.

2026-09-10 — More native hosts and recovery checks

• Add OpenCode, Antigravity, and Gemini CLI installer targets; handle OpenCode custom/XDG configuration and Gemini substitute home correctly.
• Add verify-install for installed bundle and reference integrity.
• Reject empty chapters, unmet recorded length plans, skipped prerequisites, and installer/source overlap; replace project state atomically with section-aware updates.
• Bundle native startup/recovery instructions and initial state; require evidence labels for offline market planning.
• Test nine installation targets and add macOS to Windows/Linux CI. Compatibility observations and architecture audit record measured results and limitations.

Agent-native restoration — 2026-09-10

Restored the portable-skills foundation from c974486 as the public product. Added OpenClaw and Hermes installation targets to the existing installer, kept skill references and conflict backups, and added portable-install CI. The standalone app remains in Git history. See verification and limitations.

---

Changelog

5.0.0 — 2026-09-10

Changed

• Recenter the public product on native agent skills: the host owns model access, tools, permissions, and progress.
• Remove the interactive book-genesis book-generation CLI from the release surface. The maintainer helper is now runner/installer.py and only installs or verifies skills.
• Add DeepSeek Harness, Cursor, GitHub Copilot, Qwen Code, Pi, and Windsurf/Cascade targets.
• Replace legacy top-level launch copy with English release documentation and a factual launch kit.
• Update the Remotion demo to show idea → saved chapter → review → next step without an autonomous-book or bestseller claim.

Removed

• Old top-level V4 marketing drafts, duplicate Portuguese README/card, and stale standalone-app promotional copy.

Known limits

• Native writing acceptance and repeated complete-book reliability remain host-specific evidence tasks. Installation checks do not establish reader preference, publication readiness, or sales.

Unreleased

Added

• Safe portable installer for Claude Code, Codex, Kimi Code, and shared Agent Skills directories.
• Distribution manifest, dependency validation, checksums, dry-run, conflict detection, and recoverable forced replacement.
• Independent evaluator protocol with Grade A/B/C isolation reporting.
• Blind literary critic role that receives no target score or previous evaluation.
• Commercial-length fields and Literary Barrier phase in canonical runner state.
• Portable-suite and installer regression tests.

Changed

• skills/book-genesis/ is now canonical universal core.
• Platform-specific agents consume shared packets and gates instead of redefining editorial logic.
• Public links now point to felipelobomotta-blip/book-genesis-studio.
• Legacy V4 skills, plus native Claude agents/knowledge, remain opt-in through --include-legacy.

V4.2 — 2026-06-10

Added: Premise Forge (Phase 1.5)

The pipeline now transforms raw ideas into structurally sound premises before writing starts.

Previously, the system executed your idea as-is. If your shower thought lacked an irony engine — the structural contradiction that makes readers unable to put a book down — the best prose in the world couldn't fix it.

Phase 1.5 introduces a forge step between research and foundation:

• book-architect is dispatched in "forge mode" (dispatch 0)
• 5 premise variants are generated: Variant 1 is your raw idea scored honestly; Variants 2–5 use different irony engines (inverted protagonist, relocated stakes, market gap fusion, weaponized reader frustrations from negative comp reviews, collapsed character contradiction)
• Each variant is scored across 6 dimensions: hook, irony engine, native escalation, the central question, gap fit, retellability
• The floor IS the score — a 9/9/9/9/9/4 is a 4
• Winner needs floor ≥ 8.0; one re-forge round allowed; if still below 8.0, the system proceeds with the best variant and flags it at Checkpoint 1
• Hard rule: ELEVATE, DON'T REPLACE — you must recognize your idea inside the winning premise
• Output: premise.md — binding for the entire foundation downstream

Checkpoint 1 now opens by showing you the transformation: raw idea → forged pitch sentence → "what changed and why" in plain language.

Added: 8.5 Gate System

The quality gate was previously a target without enforcement. V4.2 makes it a hard gate:

• Per-chapter gate: chapters below 8.5 on Genesis Floor or Casual score cannot advance
• Polish Loop: the evaluator produces a "Path to 8.5" work order; the editor targets only the blocking dimensions; max 5 iterations per chapter
• Anti-inflation: each polish cycle can add +0.5 maximum — jumping from 7.5 to 8.5 requires ≥ 2 real improvement cycles
• 3-state verdict: PASS (Floor ≥ 8.5 AND Casual ≥ 8.5) / POLISH (above hard floor but below 8.5) / FAIL (below hard floor)
• Exit gate: manuscript only moves to packaging when ALL chapters ≥ 8.5 AND CVI-Launch ≥ 9.0

The hard floor is genre-adjusted: literary fiction and memoir set it at 7.5; commercial fiction, thriller, and prescriptive nonfiction at 7.0.

Added: Phase 5 exit criteria

Phase 5 (full manuscript revision) now has a concrete exit condition:

• Every chapter must hold Genesis Floor ≥ 8.5
• CVI-Launch must reach ≥ 9.0 ("breakout potential, Gone Girl tier")
• Up to 3 full revision cycles
• If still below after 3 cycles: explicit decision at Checkpoint 2 (proceed anyway, keep revising, or abort)

Changed: book-architect

• Role updated from 2 dispatches to 3 (dispatch 0 = forge mode → premise.md, dispatch 1 = foundation + outline, dispatch 2 = voice DNA)
• New section: PREMISE FORGE MODE — full scoring rules, variant generation instructions, elevate-don't-replace constraint, genre shift detection
• Foundation section now reads premise.md first; the winning variant's irony engine, escalation ladder, and central question are binding inputs
• Beat Subversion marked MANDATORY in outline section

Changed: book-orchestrator

• Pipeline diagram updated with Phase 1.5
• PHASE 1.5 block added with full dispatch instructions, after-return state update, and genre-shift delta research trigger
• CHECKPOINT 1 updated to show premise transformation first
• Parallelism chain updated: research → premise forge → foundation → voice DNA → entity build → outline continuity
• Project init tree updated with premise.md

Changed: README

• Full rewrite. Previous README described V3 architecture with different agent names.
• New README explains the pipeline in plain English, describes all 3 gates with concrete scoring examples, and lists all 8 agents with plain-English descriptions.

Added: all 8 agents to agents/ folder

Previous repo only contained book-orchestrator.md. All agents now included:

• book-orchestrator.md
• book-architect.md
• book-researcher.md
• book-writer.md
• book-evaluator.md
• book-editor.md
• book-disruptor.md
• book-packager.md

---

V4.0 — 2026-05-26 (initial public release)

• First public release of Book Genesis V4
• 8.5+ editorial quality target (not yet enforced as a hard gate)
• 9-skill book studio documentation
• MIT licensed