Changelog

All notable changes to NovelForge are recorded here.

The format follows Keep a Changelog, and this project adheres to Semantic Versioning.

[2.6.0] — 2026-10-01

Added

• Icons. logo.svg, logo.png, logo-dark.png, composer-icon.png, and composer-icon-dark.png — 512×512, well within the 5 MiB limit, referenced from both OpenAI manifests. Codex package validation requires logo and composerIcon, so without these the package could not be submitted to the ChatGPT/Codex directory despite being public on GitHub.
• PRIVACY.md and TERMS.md, referenced by the listing URLs the manifests declare. PRIVACY documents the session hook's exact behaviour and states that the plugin makes no network calls.
• Listing metadata: brandColor/brandColorDark, and websiteURL, supportURL, privacyPolicyURL, termsOfServiceURL.
• Validator now enforces the documented icon constraints: both fields present, ./-prefixed paths that resolve, square, at least 48×48, a supported format, within the size limit, and a viewBox for SVGs. It also checks that declared GitHub listing URLs point at files that exist. Two negative tests confirm the square and required-field checks fire.

Fixed

• A malformed extensions namespace in both OpenAI manifests. A script had walked ["com", "openai"] as two literal keys, producing extensions.com.openai alongside the correct extensions["com.openai"]. The onboardingSkill field added in 2.5.0 landed in the wrong object, so neither manifest declared it in a namespace any host reads. Both manifests were rebuilt with a single com.openai namespace. The validator did not catch this because extensions is an open map in the schema and nothing checked the nested shape.

Dogfooding

NovelForge v2.5.0 was run against an 18-chapter Arabic manuscript (الجرس تحت الماء الأسود, ~20,000 words) as a live test of onboarding → project-scaffold → manuscript-import. Results are in the working notes for that project; the substantive findings are recorded here because one of them concerns the system rather than the novel.

• manuscript-import produced a fabricated detail. Post-intake verification re-tested all 36 of intake's verbatim quotations against the source corpus: 32 verified, 4 missed. Three were transcription or normalisation errors in the derived records. One was content the manuscript does not contain — a third guard title, invented while compressing a passage that names only two. Nothing in the validator can detect this class of error, because a fabricated detail in a derived record is structurally indistinguishable from a correct one. It was caught only because every fact was cited to a chapter, which made verification possible at all.
• Intake otherwise behaved as designed: no invented canon, four genuinely unresolved questions surfaced, three of four leads recorded as UNRESOLVED rather than guessed, single POV and past tense confirmed across all 18 chapters, and the /novel create → scaffold path produced a minimal correct structure rather than the full template set.

[2.5.0] — 2026-10-01

Added

• context-budget — real token-budget discipline. Per-scope ceilings (scene 8,000, chapter 20,000, arc 60,000, novel 120,000), a dependency-free estimation method, an eleven-rank relevance ordering, and a trim-by-deferral sequence that names what is never sacrificed. Rationale for deferral is recorded, because an unexplained omission is indistinguishable from a mistake.
• project-scaffold — the skill /novel create lacked. Eight phases from author questions to a minimal default set, with a rule that a domain is created only when a template exists to populate it. Scaffolding is MEDIUM risk and may proceed without approval; filling canon afterwards is not scaffolding.
• project-settings — POV, tense, approval strictness, autonomy ceiling, and engine selection as project configuration rather than plan prose. Strictness governs MEDIUM risk only; HIGH-risk work always asks, and no setting waives that.
• cross-engine-audit — one call that selects the right engine subset per scope, sequences them so later stages consume earlier findings, and consolidates output into stable finding IDs. Records skipped engines and why, and treats engine disagreement as an authorial question rather than reconciling it automatically.
• onboarding — first-time orientation. Opens with one question, delivers a first artefact rather than an architecture tour, explains only three concepts before the first draft, and routes to scaffold, import, or series setup.
• examples/salt-and-ember/ — a complete worked project with 16 files, showing filled canon with dependencies, a power system with enforceable costs, a deliberately unsolved mystery, an accumulating findings ledger, a measured and trimmed context packet, and a chapter audit with an unresolved engine disagreement. Deliberately incomplete, with the omissions recorded.
• PROJECT_SETTINGS.md and audit/AUDIT_RUN.md templates.
• 3 new slash commands — audit, setup, budget.
• extensions.com.openai.onboardingSkill wired to the onboarding skill in both OpenAI manifests.

Changed

• novel-orchestrator gains Configuration and a context-budget routing rule, and reads approval strictness from settings. A duplicated Approval policy section was merged into one.
• novel-create routes through project-scaffold and records configuration.
• Validator checks the worked example: required files present, no stray skills/ directory, all three strictness levels documented, and open questions genuinely marked UNRESOLVED.

[2.4.0] — 2026-09-30

Added

• skills.sh support. NovelForge installs through the open agent skills registry, which covers 20-plus agents in one command. Verified against skills CLI v1.7.0: all 45 skills are discovered, and a scoped install lands correctly with a working skills-lock.json.
• skills.sh badge and installation section in the README, including telemetry opt-out and the distinction between installing the plugin and installing individual skills.
• gh skill install documented in the compatibility table.

Fixed

• .agents/skills is no longer a symlink into the canonical skills/ tree. That path is the install target for both the skills CLI and Codex CLI, so npx skills add <anything> wrote third-party skills directly into NovelForge's source, where they were indistinguishable from real engines and were picked up by the orchestrator routing table and the OpenCode catalog. Verified by installing an unrelated skill into a clone: it added a 46th engine to skills/. The directory is now real, git-ignored for installed skills, and documented in place.
• Validator now asserts .agents/skills is a real directory, so this cannot be reintroduced, and reports any third-party skills present there.

Changed

• Validator layout section distinguishes the two symlinked discovery paths (.opencode/skills, .gemini/skills) from the skills CLI install target (.agents/skills).

[2.3.1] — 2026-09-30

Fixed

• Invalid plugin category. category was set to Writing, which is not a title in the OpenAI dashboard's category list, so the plugin would have been skipped in the marketplace picker. Now Creativity, in all three places the value appears: plugin.json, .codex-plugin/plugin.json, and .agents/plugins/marketplace.json.
• Over-length defaultPrompt entries. All three starter prompts exceeded the documented 128-character submission limit in both manifests, at 161, 140, and 277 characters. Rewritten to 105, 102, and 93.

Added

• Validator now enforces the documented OpenAI listing limits: displayName and shortDescription at most 30 characters, longDescription at most 4000, developerName at most 80, at most three defaultPrompt entries of at most 128 characters each, at most 20 capabilities of at most 120 characters, description at most 4000, a recognised category, and required policy.installation, policy.authentication, and category on every marketplace entry.

[2.3.0] — 2026-09-30

Added

• manuscript-import — phased intake for an existing draft: inventory, structure, entity and relationship extraction, timeline reconstruction, knowledge states, Style DNA inference, and canon classification. Approval is required between phases, every extraction carries a source citation, and nothing is classified as permanent canon until the author confirms it.
• series-manager — canon deliberately shared across a series, with SERIES_CANON and NOVEL_CANON kept strictly separate, explicit inheritance per volume, and recorded divergence points. This is the sanctioned exception to workspace-manager's default isolation.
• 16 slash commands in commands/ for Claude Code: novel-create, novel-import, novel-status, novel-report, plan, write, critique, gate, continuity, canon-impact, branch, payoff, mystery, power, style, series.
• SessionStart hook shared by Codex and Claude Code. Detects a novel project, injects root, title, status, branch, open Critical findings, and open obligation count, and suppresses unfilled template placeholders rather than reporting them as values. Silent outside a project; never fails a session.
• Findings ledger (project/quality/FINDINGS.md) making critique stateful: stable finding IDs, resolved only once verified, author-accepted risks with reasons, and a regression watch list.
• Progressive disclosure — five engines gained a references/ file holding depth that is not needed until the skill is working on a problem: canon-manager/CHANGE_PROCEDURE.md, critique/CHECK_CATALOGUE.md, generic-writing-detector/PATTERNS.md, mystery-engine/FAIRNESS.md, and power-system/SCALING_AND_EXPLOITS.md.
• 5 new templates — import plan and extraction log, series bible and book entry, findings ledger.
• OpenCode HTTP catalog (catalog/) generated by scripts/build_catalog.py and kept current by a second workflow, allowing install without cloning.
• CONTRIBUTING.md and AGENTS.md.

Changed

• Validator now verifies that references/, scripts/, and assets/ paths named inside a skill resolve to real, non-empty files.
• Validator now checks catalog freshness against the canonical skills tree.
• CI additionally smoke-tests the session hook for silence outside a project and tolerance of malformed stdin.

Removed

• The five compatibility aliases — trope-management, versioning, memory-governance, workspace-isolation, prose-style. They existed to protect installs that do not exist: the plugin had no users when they were added, and four of the five were referenced by nothing at all. Removing them takes the surface from 48 skills to 45 and eliminates a class of ambiguity where two skills claimed one job. Use trope-manager, version-control, memory-manager, workspace-manager, and narrative-style.

[2.2.0] — 2026-09-30

Added

• scripts/validate.py — repository validator covering skill spec compliance, skill substance floors, cross-reference integrity, orchestrator routing coverage, alias target validity, manifest version agreement, live Agent Plugins schema conformance, licence/author agreement, en-GB spelling, stale version labels, and discovery-symlink integrity.
• .github/workflows/validate.yml — runs the validator on every push and pull request with --schema --strict, plus a skills-tree integrity check that tolerates symlinks not materialising on Windows checkouts.
• Substance checks for skills. Format-valid-but-empty skills are now a build failure. Both historical regressions — the v1.5 one-line pointers and the v2.0.1 description restatements — were well-formed files carrying no instructions, so they passed every spec rule.

[2.1.0] — 2026-09-30

Added

• Skill expansion. All 43 working engines rewritten from 90–120 byte restatements into substantive procedure, failure modes, and guardrails. Average body 2,911 bytes, up from roughly 120.
• Depth recovery. Craft procedures lost in v1.0→v1.5 restored into the writing engines; v1.5 engine depth restored into the analysis engines; v2.0.0 phase-2 content extended.
• Orchestrator state machine and a routing table covering all 42 engines.
• 12 new templates for engines that previously had no on-disk artefact: scenes, relationships, factions, battles, pacing, tension, causal chains, consequences, cliché reviews, voice profiles, glossary, reader state.
• 18 templates expanded, including three still headed "v1.5" inside a v2.0.1 package and twelve restored as verbatim copies of older versions.
• 5 compatibility aliases rewritten as real routing documents, replacing conflicting duplicate skill pairs.

[2.0.1] — 2026-09-30

Changed

• Restored the 12 skills deleted in v2.0.0 and the 25 missing templates, but at placeholder depth. Documented in 2.1.0.

[2.0.0] — 2026-09-30

Added

• Phase-2 engines: style drift, generic-writing detection, anti-cliché, trope management, theme and symbolism, research provenance, author profile, memory governance, workspace isolation, quality gate.

Removed

• Twelve craft skills and six templates. This removed the writing half of the system and left only analysis engines; the omission was not documented at the time and was reverted in 2.0.1.

[1.5.0] — 2026-09-30

Added

• Twelve analysis engines: narrative graph, causality, consequence, pacing, tension, payoff debt, reader knowledge, character simulation and arcs, power exploit, battle choreography, faction simulation.

Changed

• Twelve Phase-0 skills were replaced with one-line pointers to the v1.0 specialists rather than extended. This was undetected until 2.1.0.

[1.0.0] — 2026-09-30

Added

• Initial system: 19 craft skills, 22 templates, Novel Bible source of truth, targeted context retrieval, risk-based autonomy, branching.