NovelForge Changelog

0.8.0 · Active pre-1.0 development baseline

0.8.0 is the current development identity for the architecture previously discussed as the “8.0” development line. NovelForge is still pre-1.0: latest main is the working implementation baseline, and this version does not promise a frozen 1.0 compatibility surface.

Version truth

• Current machine-facing version surfaces are normalized to 0.8.0: HARNESS_MANIFEST.yaml, SKILL.md, novelforge.py, the Project SDK default, exposed MCP server metadata, and documentation governance metadata.
• This replaces the previous fragmented 7.2 release-metadata / 7.3 implementation-metadata convention. Historical 7.x records below remain historical and are not rewritten.
• During active pre-1.0 development, justified breaking machine-contract cleanup may still land on main. Formal compatibility promises belong to a future intentionally frozen release contract.
• Documentation rebuilt against current main remains candidate_review unless its manifest lifecycle explicitly says otherwise; version normalization does not auto-promote semantic, native-copy, or visual review status.

Merged development changes

• The semantic runtime uses a small model-contract catalog with progressive disclosure of exact contract packs. Semantic fiction judgment belongs to the model; deterministic code owns authority, permissions, fingerprints, persistence, routing, hard budgets, typed validation, transactions, and reproducibility.
• PR #11 migrated the live machine namespace from NOVEL_OS_* / novel_os_* / .novel-os/ surfaces to NOVELFORGE_* / novelforge_* / .novelforge/ without compatibility aliases.
• PR #12 added task-aware context grounding with explicit question→evidence mapping and deterministic perspective/visibility filtering before model context is assembled.
• PR #13 added metadata-only novelforge_run_receipt_v1 observability without candidate prose, Canon authority, or a second state database.
• PR #18 made the Framework bundle release-complete by including the quality runtime and smoke-testing extracted bundles with novelforge.py doctor plus the model-free self-test.
• PR #19 merged Studio Phase 1: a read-only Run / Context Inspector driven by Run Receipts.
• PR #21 merged Studio Phase 2A: the portable one-product/many-host contract, safe Project Hub projection, synthetic project/scene fixtures, and a read-only Project Hub + Scene workspace prototype.
• PR #25 merged Studio Phase 2B: a versioned read-only host bridge plus a standards-compatible NovelForge Agent Skill. The bridge exposes an allowlisted read surface (bridge.describe, framework.doctor, project.inspect, capabilities.inspect, context.inspect, semantic.catalog), fails closed on unsupported operations, defaults away from host-private absolute paths, and keeps authority=false.
• PR #24 completed the remaining machine-contract rename: os_behavior_write → framework_behavior_write, semantic job/result IDs moved from novel-os-* to novelforge-*, the live .novel-os/ ignore surface was removed, and namespace hygiene now blocks those legacy machine identifiers from returning. No compatibility aliases were added.
• PR #27 made Context grounding story-order-aware and question-specific: future/incompatible evidence fails closed, including pinned evidence, and hard-budget drops can leave grounding explicitly incomplete rather than silently under-supported.
• PR #28 separated character epistemic status from acquisition mode and bound proposed character actions to story-ordered, character-visible evidence; future/unknown/invalid evidence cannot be used as positive support merely because the Framework stores it.
• PR #29 strengthened long-horizon reconciliation with story-ordered evidence, complete requirement coverage, explicit uncertainty, and separation between shared relationship state and per-character perceptions.
• PR #31 added novelforge_production_readiness_v1, a deterministic same-candidate-fingerprint conjunction gate for Surface, Reader Engagement, required Continuity, and required independent semantic review. Missing/pending/failing required gates block review readiness; RG-15 SAFE-BUT-FLAT cannot pass Reader Engagement; no numeric literary-score aggregation is introduced.
• PR #31 also added the minimum deterministic Publication core: manifest-discoverable novelforge_publication_ir_v1, publication/compiler.py, exact Accepted-text fingerprint preservation, and clean_text, web_reflow, print_book, and epub3 build profiles. Publication output remains derived/non-Canon. EPUB targets W3C EPUB 3.3; internal validation is deterministic, while release conformance requires an explicitly supplied external EPUBCheck command.
• PR #32 upgraded Story Loom to novelforge_brand_tokens_v2 and merged the WeiUI foundation contract. assets/brand/weiui.integration.json pins xiaooye/weiui at exact commit d84d1cd365fb5f90cbbab794d2358f7a13b29b79, allows only @weiui/tokens + @weiui/css, forbids @weiui/react / @weiui/headless as Studio runtime dependencies, and requires zero WeiUI runtime JavaScript. assets/brand/story-loom.weiui.css supplies the wui-theme Story Loom layer without forking WeiUI selectors.
• PR #32 also made mobile/i18n/accessibility/runtime-overhead expectations machine-checkable: mobile-first responsive rules, 44px minimum touch targets, en-US + zh-CN, logical properties, reduced motion, no idle decorative animation, no default polling, and deterministic light/dark contrast checks are enforced by scripts/design_system_quality.py and CI.
• The selected Phase 2C application framework is SolidJS + TypeScript + Vite + @solidjs/router, consuming WeiUI only as a zero-JS CSS/token foundation. Local Web remains first-class and preferred for minimum incremental runtime overhead; Tauri remains an optional/installable desktop host rather than the center of product architecture.
• The 0.8.0 normalization aligns the current machine/version identity instead of maintaining parallel “release” and “implementation” development numbers.
• Documentation governance tracks audience, tier, authority sources, freshness ownership, rewrite policy, lifecycle state, bilingual pairing, local-link integrity, version alignment, and deterministic visual/documentation checks. studio/ is included in bilingual manifest-coverage QA.

Active gaps and compatibility notes

• Run Receipt still has Core-owned consumer/read-surface work: stable manifest discoverability, event-schema alignment for run.receipt_recorded, and a stable query/projection boundary instead of persistence-internal access. PR #25 deliberately does not expose unsafe Control Plane or Run Receipt reads through the host bridge.
• The Publication core is real, but broader Issue #16 scope remains open: richer semantic structure/profile contracts, print-PDF via a paged-media engine, broader typesetting controls, visual-regression hooks, and Studio publication preview/authoring are not implied by the minimum compiler.
• Story Loom/WeiUI foundation integration is now real, but the Phase 2C SolidJS application itself is not implied by those token/CSS artifacts. App-shell code, route implementation, host lifecycle, and measured idle CPU/RAM still need their own implementation evidence.
• Existing projects that intentionally pin an older Framework revision remain bound to that revision until explicitly upgraded. Generic Framework development itself follows latest main rather than an internal development lock.
• A future stable migration guide must be generated from frozen contracts and the final bundle, not inferred from issue descriptions or intermediate commits.

Product / publication status

• Studio currently has merged read-only Phase 1, Phase 2A, and Phase 2B product slices: observability, Project Hub, Scene workspace, portable host boundary, and Agent Skill delivery are real on main.
• Story Loom v2 + the exact-pinned zero-JS WeiUI CSS/token foundation are also real on main. This is a product-foundation implementation, not a completed Studio app.
• Phase 2C is directed toward SolidJS + TypeScript + Vite + @solidjs/router. @weiui/react is explicitly not a planned runtime dependency. Tauri remains an optional installable host; the Local Web surface remains first-class.
• These Studio slices do not make Studio a write-capable, collaborative, authenticated, or production-hosted application. Generic invoke/write, project mutation, resume, and broader Control Plane reads remain outside the current bridge contract.
• Publication now has a minimum deterministic Core implementation: exact-text Publication IR plus clean-text, Web HTML, print-oriented HTML/CSS, and EPUB 3.3 outputs. This does not mean the full Typesetting Toolkit or Studio Publish experience is complete; Issue #16 remains open for the broader scope.
• Broader MCP registry/management and later write-capable Studio operations remain deferred to their owning workstreams.
• UI, host-bridge, Agent Skill, production-readiness receipts, and publication outputs never become Canon, Memory, semantic truth, settlement truth, or workflow authority merely by existing.

7.0.0 · Adaptive Fiction Framework

Architecture

• Repositioned the repository as a fully project-agnostic fiction agent framework.
• Established one-way dependency: consuming Project → NovelForge Framework.
• Added standard and mapped Project Adapter support so mature fiction repositories can migrate without destructive directory rewrites.
• Added novelforge.toml project manifest and novelforge.lock.json framework dependency lock contract.

Fiction Core

• Added generic Story Architecture, Character/Relationship, Canon/State, dependency, settlement, and continuity contracts.
• Promoted recurring anti-AI prose corrections into framework-level Surface Fundamentals (HF-01..HF-29).
• Added generic Reader Engagement positive quality model (RG-01..RG-15), including SAFE-BUT-FLAT detection.

Runtime

• Preserved session-native Harness with manager/specialist/reviewer separation.
• Added durable SQLite Control Plane for sessions, events, handoffs, leases, result hashes, and logical consume-once receipts.
• Added provider-neutral runtime routing across chat sessions, local Codex/Claude, MCP, provider APIs, GitHub/service jobs, local models, and humans.
• Kept mandatory independent semantic judgment fingerprint-bound and fresh-per-fingerprint by default.
• Added typed GitHub event ingress and a no-API peer-chat semantic bridge.
• Added optional manually dispatched provider-backed semantic eval workflow; it requires an explicit secret and is never part of normal CI.
• Added weekly deterministic maintenance that observes/tests/queues work without LLM execution or automatic Framework promotion.

Adaptive Learning

• Added durable Learning Store for preference evidence, revisable hypotheses, contradictions, Corpus gaps, promotion candidates, and rollback records.
• Added autonomous preference-dimension discovery and Corpus-gap generation.
• Enforced evidence hierarchy: model inference alone cannot become durable user taste or General Craft.

Corpus Intelligence

• Added provider-neutral Corpus Scout and rights/storage gate.
• Added rights classes redistributable | analysis_only | unknown.
• Added question-bounded analysis, counterexample search, cross-work generalization, and named-author imitation boundaries.
• Migrated eight generic cross-work mechanism benchmark seeds without raw source text or consumer-project facts.
• Scheduled maintenance can generate typed Corpus discovery queues, while actual Web/GitHub/MCP discovery still requires an authorized host connector and is never fabricated.

Evals

• Added generic deterministic + semantic eval runner.
• Added blind semantic queue builder that strips expected/gold/release labels before reviewer dispatch.
• Added v7 Surface/Reader/Character/Canon/Corpus fixture suite.
• Normal CI reports semantic cases as PENDING_MODEL when no independent judgment exists rather than fabricating PASS.

Project Engineering

• Added executable Project SDK: init, validate, spec-new, build, self-test.
• Added generic mapped Project Adapter for legacy/mature repositories.
• Adopted engineering discipline for structural changes: spec → plan → tasks → implementation → verification → acceptance.
• Added deterministic compact project bundle/fingerprint build model.

Documentation / Repository Quality

• Added paired English and Simplified Chinese authoritative documentation.
• Added Mermaid architecture/learning/runtime/project diagrams.
• Added agent-framework adopt/adapt/reject research matrix.
• Added CI hard gates for consumer-project leakage, bilingual pairing, relative links, manifests, Project SDK, Learning/Corpus, Runtime/MCP, Semantic transport, Evals, and authority boundaries.
• Normal CI does not invoke paid/login-bound model inference.

Migration note

• Old internal compatibility identifiers may still use pre-v7 novel_os_* names inside stable executable schemas/environment variables. They are implementation compatibility details, not consumer-project dependencies. A future schema migration can rename them behind a dedicated structural-change spec without destabilizing v7 release behavior.