Changelog

1.21.1 — 2026-10-20

• Rendered-site annotation review names Clear Assessment Guides for shareable assessments and decision boards, and Visual Version Review for choosing among implemented versions.

1.21.0 - 2026-10-20

• route-3d-work: add EL4CTEO/blender-skills as a trial Blender layer (read blender-core and blender-asset-qa first, install per project, keep the CLI build as source of truth). Name image-blaster and Scenario's skills (GameDev OS) as generator-first packs that need authorization; image-to-world output is a blockout, not a finished scene.
• Runtime-evidence gate: moving parts that turn about the wrong pivot.
• Motion explainers: add OpenDesign as a workspace lead that renders through HyperFrames; its generative-video presets are paid routes.

1.20.1 - 2026-10-20

• Fidelity loop: reject the generic flat, pastel, toy-like look agents fall back to, unless the brief asks for it; name the product's palette, materials, lighting and density in the target prompt.
• route-3d-work: state the medium-effort, 2–3 round default and defer increases to Work Router.
• Gaussian splats: list three.js r186's built-in GaussianSplat add-on for object- or room-scale captures in WebGPURenderer projects, with Spark for WebGL2, mobile and WebXR.

1.20.0 - 2026-10-20

• Name verification tools in the completion gates: the host-native browser first (Codex desktop Browser and Computer Use with optional CDP Developer mode; Claude in Chrome), then the Chrome DevTools MCP server when a host lacks a browser or needs performance and accessibility evidence, then one scripted browser per project.
• Point the visual build loop's capture step at that tool order.
• Add a word-budget test for route-ui-work/SKILL.md.

1.19.5 - 2026-10-18

• Add a motion-explainer route for conceptual animation and authored films, separating creative direction from rendering, voice and recorded UI evidence.
• Plan composition, typography, visual continuity and explanatory motion; prove one representative moving sequence and review transition midpoints before treating a technically valid export as aesthetically complete.
• Conditionally select official HyperFrames creative/animation skills, video-shotcraft with Remotion guidance, or a stated adaptation in an existing pipeline. Keep teaching pace, deliberate stillness and supplied references.
• Add a dated public-source review, nearby non-trigger cases and field-guide routing.

1.19.4 - 2026-10-18

• route-ui-work: add the MIT Shaders library as a conditional source for a decorative GPU effect. It is WebGPU only and leaves a transparent canvas where WebGPU is missing, so the page needs a fallback underneath. Components sample 5% of visitors for telemetry unless disableTelemetry is set. Canvas UI stays the choice when an effect needs a WebGL path.
• route-ui-work: add Anatomy as a conditional, opt-in catalog skill for interactive figures that explain a mechanism. Check its metaphor and every displayed number against the real system; gallery submission is a separate publishing step.
• Add applicable and nearby non-trigger policy cases and the October 18 UI leads review.

1.19.3 - 2026-10-18

• route-3d-work: when the camera moves, runtime evidence includes a short clip along the real path, reviewed with timestamps for flicker, z-fighting, pop-in and floating parts. Settled stills cannot show these.
• spatial-realism.md: bake code-built geometry or build it off the main thread when construction shows up as a startup long task; seat scanned and imported props in their surroundings; record how lightmaps travel (glTF has no ratified lightmap slot) and check UV set, color space and intensity after export; keep a switch that turns baked layers off for review.
• fidelity-loop.md: optional color statistics against the locked target support, but do not replace, the critique.
• runtime-and-slice.md: when the agent cannot listen, a local audio-analysis tool can screen sounds; acceptance still needs human listening, and cloud critique needs consent.
• Add applicable and nearby non-trigger policy cases and the October 18 three.js builder review.

1.19.2 - 2026-10-16

• Audit persistent copy in task-focused views. Remove subtitles and helper lines that only describe visible controls or repeat headings; keep concise context for non-obvious scope, excluded data, consequences, and recovery beside the relevant state.
• Add paired cases for a redundant process-table instruction and a useful excluded-row count with an action.

1.19.1 - 2026-10-15

• For requested overall UI ratings, define a task-specific acceptance matrix, scoring anchors and evidence before the first edit. Cover relevant visual-system consistency, typography, alignment, spacing, content and interactions without imposing a fixed checklist on small changes.
• Keep unverified dimensions and material acceptance failures visible; do not infer visual quality from technical checks or reshape a rubric to meet its requested target.
• Add a dense account-drawer policy case and a nearby small-copy-change case.

1.19.0 - 2026-10-11

• Keep the repo project-neutral: the paired eval and split review no longer name a consuming project; project task briefs and results live in that project's repo.
• install_optional_skills.py honors AGENT_HOST=codex for skills-CLI installs (previously always --agent claude-code).
• Add route-3d-work, a separate ~800-word skill for real-time 3D: project-as-lead, runtime → knowledge map, specialists by need, Dream Loop under the fidelity gate, tools and seven completion gates. route-ui-work stays a layer for HUD and overlays.
• Move interactive-worlds.md (split into runtime-and-slice.md and on-demand live-ai-and-hosts.md), spatial-realism.md, interaction-contracts.md and the Dream Loop/fidelity-gate and photoreal/playable sections of upstream-skills.md into route-3d-work/references/. Redirect stubs keep old anchors for one release.
• Host-native harness: Dream Loop runs in the active host's model (Codex → GPT-6 Astra, Claude Code → Claude Opus 5.5) unless the user names another; no cross-host moves.
• Add knowledge layers and tools: three.js llms.txt, Meng water/sky and optimize-threejs-games, 3dviz-pro-max as CLI retrieval only, two read-only threejs-game-skills references, Blender CLI/MCP split, Needle Inspector trial, conditional Spark. Exclude foreign-runtime and generator-first packs by default.
• optional-skills.json: claudedesignskills now installs only gsap-scrolltrigger, threejs-webgl and blender-web-pipeline; awwwards-3d leaves the creative profile; new meng-3d, meng-game-dev entries and a 3d profile.
• route-ui-work/SKILL.md: remove the model-policy block (Work Router owns model and effort), keep the three Claude agents, drop "interactive games and worlds" from the description. 2,866 → ~2,160 words.
• No-marketplace installer copies both routing skills.
• Field guide data (docs/router-data.json, docs/field-guide.html): the creative-expressive canonical chain reflects the ClaudeDesign/Awwwards demotion and the 3D handoff to route-3d-work.
• Add a paired-eval protocol (tests/route-3d-paired-eval.md) and the October 11 review.

1.18.11 - 2026-10-04

• Choose asset-authoring and runtime responsibilities per task; use a compatible hybrid Blender/Three.js pipeline without making either tool generate everything.
• Diagnose source, export, optimization and runtime defects before remodeling; inspect actual delivered attribute values and affected state variants.
• Check target presence in settled camera views as well as containment, and retain separate artifact, rendered-craft, UI/UX and performance verdicts after repairs.
• Add applicable and nearby non-trigger policy exercises; preserve the required fidelity execution gate and existing authorization boundaries.

1.18.10 - 2026-10-04

• Require an evidence-backed fidelity execution gate for substantial high-fidelity 3D work: locked target, running comparison, fresh-context critique, checked refinement and measured performance.
• Distinguish listed, loaded, executed and accepted workflows; keep unmet gates and material adaptations explicit without expanding model, delegation or spending access.
• Preserve direct handling of ordinary edits and separate UI/UX and interaction gates.
• Author missing targets with the host Imagegen tool from the existing product's baseline and real navigation constraints; retain prompt/output identity and keep unavailable or unauthorized generation visibly blocked.

1.18.9 - 2026-10-02

• Start Claude UI implementation at Opus 5.5 medium and preserve high for consequential review; verify actual provider alias resolution and qualify the old Sonnet effort ceiling.
• Account for the current WebDev evidence in open host selection while preserving Codex routes, concrete art direction, visual-input checks and rendered interaction validation.

1.18.8 - 2026-10-01

• Add scene-audio authoring checks, early proofs for specialist VR input, and a conditional OpenAI Decisions route. Distinguish MiniMax Design asset workflows from coding tools; retain access, rights and quality evidence boundaries.

1.18.7 - 2026-10-01

• Add scoped Meng game-design, camera and audio routes; connect playable-world choices to observable engagement; trial ChatGPT host delivery before platform commitment.

1.18.6 - 2026-10-01

• Verify intended audio in the final served video, review first paint through settled hero motion, and conditionally verify live CMS media, forms, cache, and public result.
• Start typical substantial UI work with GPT-6.1 Sol; retain Astra for exceptionally difficult unresolved quality-first direction, with dated Max-only evidence limits.

1.18.5 - 2026-09-28

• Correct the PreToolUse hook timeout from 5,000 seconds to 5 seconds. The hook schema uses seconds, so this keeps sudo blocking deterministic without allowing a stalled hook to run for more than an hour.

1.18.4 - 2026-09-28

• Fold the rendered-site UX annotation review procedure into a focused UI Router reference. Track exact routes and selected wording, verify interactions and media, and separate findings, source changes, saved retests and deployment.
• Keep small edits scoped and review-only work read-only; retain existing design leads, baseline guidance and completion gates without duplicating them.

1.18.3 - 2026-09-27

• Select Impeccable Critique, Audit and Polish by the actual review question and remaining findings, without a mandatory command stack.
• Label provider scores and project-specific visual ratings separately; do not exchange scales or use a technical score as an overall look-and-feel verdict.
• Add policy exercises for broad visual review, narrow technical audit and mixed scores.

1.18.2 - 2026-09-27

• Verify hosted-builder source rights and an editable export/rebuild when portability matters.
• Evaluate live conversation using complete lesson outcomes, audible response latency, recovery and comparable session-cost assumptions rather than provider billing units.
• Check decision services in the target language and calibrate confidence with fallback.
• Complete the supplied research coverage while preserving 1.18.1 model-routing policy, existing narration ownership and learning-mode boundaries.

1.18.1 - 2026-09-27

• Include Arena in model/design-routing reviews and inherit Work Router's regular benchmark sources with a standalone evidence-checking fallback.
• Prefer Astra for new substantial UI builds where visual quality is central to acceptance; retain Sol for implementation with settled design and explicit checks.
• Keep Max leaderboard evidence separate from high/xhigh recommendations and preserve rendered, responsive, accessibility and interaction verification.

1.18.0 - 2026-09-27

• Add conditional runtime and authoring selection for interactive worlds, with a playable slice before expansion or migration and no universal engine/model ranking.
• Separate authoritative state from optional dialogue/decision services; validate actions, recover delayed responses and assess the actual latency and session-cost needs.
• Add language-task contracts preserving comprehension or speaking goals, locale boundaries and honest speech/learning evidence; clarify generated-world and semantic-asset limits.
• Retain existing Meng, Dream Loop, asset, audio and model-routing owners. No new provider, engine, integration, paid generation or mandatory multi-agent workflow is installed.

1.17.1 - 2026-09-22

• Make high and xhigh choices explicit for Sol implementation and Astra creative/spatial judgment; inherit Work Router's timing and token policy without duplicate questions.

• Follow Work Router's four-model assessment, including Astra for the hardest creative/spatial judgment and quality-first direction, Sol for substantial builds, Luna for objective bounded work, and conditional Terra 5.6 evidence collection.

• Require rendered acceptance regardless of model; coding scores, prices and release demos do not establish superior visual taste or fastest accepted designs.

1.17.0 - 2026-09-22

• Align Codex UI execution with Work Router's GPT-6 Sol and GPT-6 Luna routes while keeping design-chain selection and rendered acceptance in UI Router.
• Use Sol for unresolved design synthesis and broad implementation, Luna only for bounded objectively verifiable work, and preserve explicit or well-progressing model choices.
• Record the joint release review without treating a model release as evidence of visual taste, working interaction, or a reason to add tools and subagents.

1.16.12 - 2026-09-22

• Require rendered reading and composition judgment alongside technical checks for broad reviews, including explicit Impeccable invocations.
• Check dense openings, repetition, long-list grouping, responsive layouts and every requested sibling page; preserve useful content and scoped technical audits.
• Block unsupported excellent verdicts and target-driven scoring; require version-specific evidence and reinspection. Add applicable and non-trigger policy exercises.

1.16.11 - 2026-09-22

• Add focused spatial-realism checks for interface occlusion, complete placement chains, generated/imported asset integration, baked dependencies and comparable runtime evidence.
• Add Hyper3D/Rodin as a conditional connected asset route with live-schema, credit, asynchronous-job and destination-validation boundaries. No mandatory generation.
• Record applicable and nearby non-trigger cases; retain existing outcome-evidence rules.

1.16.10 - 2026-09-20

• Add conditional generated visual comparisons before costly implementation, with authentic-content and native-CMS validation boundaries.
• Check cross-page vocabulary, navigation, action hierarchy and shared forms; preserve intentional differences.
• Enforce explicit rejections without relocating them; separate acceptance dimensions and benchmark evidence from model-ranking assumptions.
• Add applicable and nearby non-trigger policy exercises.

1.16.9 - 2026-09-20

• Distinguish recognizing design language from naming a direction without examples.
• Calibrate terminology and explanations to the conversation; pair precise labels with visual choices without assuming novice status or requiring an expertise quiz.

1.16.8 - 2026-09-20

• Check rendered baseline/candidate isolation when CMS variants share templates or caches.
• Distinguish preserved source data from preserved navigation after alternating visits.
• Add an applicable cache-contamination case and a nearby single-edit non-trigger.

1.16.7 - 2026-09-20

• Add conditional Rive and DiffUI routes with explicit project-fit, access and validation boundaries; neither requires proactive account setup or purchase.
• Link Meng's complete playable-game testing workflow alongside the 3D specialists.
• Reconcile the field guide's visible version with its discovery metadata and record applicable and nearby non-trigger selection checks.

1.16.6 - 2026-09-20

• Distinguish free source, account-free tools, free account allowances, paid upgrades and separate execution costs when recommending UI resources.
• Correct 21st's free MCP/CLI search boundary and separate component retrieval from hosted AI access; recognize the supported authenticated CLI as a usable route.
• Add the official Motion specialist and account-free documentation MCP route, with live capability checks for premium source, generation and audit tools.
• Record focused access checks and applicable/non-trigger policy cases. Router updates and third-party installations remain separate operations.

1.16.5 - 2026-09-20

• Reconcile accumulated design feedback before broad revisions; distinguish optional experiments from an established brand and treat later corrections as scoped changes.
• Add a brief before-editing check and focused clarification for consequential gaps, without imposing questions or reference searches on routine edits.
• Check what contrast fixes obscure or displace, preserve detail behind summaries, and separate media approval from placement.
• Block overall-improvement claims when agreed requirements fail; diagnose context and implementation before blaming models. Existing Work Router policy is unchanged.
• Add ten applicable and nearby non-trigger policy cases.

1.16.4 - 2026-09-20

• Add conditional guidance for improving existing sites and apps: preserve useful components, original messaging, brand and successful interactions while addressing the largest observed visitor obstacle first.
• Preserve visual energy alongside utility; inspect image subjects across widths and motion, and exempt mechanical edits from the full routing workflow.
• Compare prominence and complete journeys, not just feature presence; apply scoring anchors to evidence and correct unsupported recommendations.
• Reuse existing CMS/editor validation and Astra-routing guidance rather than adding duplicate workflows. No new third-party skill or runtime dependency is installed.

1.16.3 - 2026-09-20

• Add conditional task/asset contracts for interactive scenes: visible meaning, category versus instance, complete actions and settled placement/reset evidence.
• Check shared input ownership across activity switches and prevent stale media completion from advancing or crediting a different activity.
• Separate catalog size, assisted performance and design rubrics from measured outcomes. Retain the existing spatial specialist and fidelity requirements.
• Align the stale Codex marketplace version pin and cover it with a parity test. No third-party installation or application runtime is changed by this release.

1.16.2 - 2026-09-20

• Add exact Meng specialists for detailed 3D materials, geometry, virtual tours and HiDPI rendering; choose camera and control behavior from the requested experience.
• Distinguish composed imagery, fixed-camera scenes, object inspection, guided tours, free walking and XR without imposing a scroll-world workflow on a single scene.
• Require relevant asset, lighting, rendered-result and runtime evidence for realism; label estimates, unavailable measurements and untested devices explicitly.
• Consolidate the spatial guidance and add applicable/non-trigger policy cases. No third-party skill, runtime or paid service is installed by this release.

1.16.1 - 2026-09-20

• Diagnose embedded media with the same source across the relevant browser/origin; avoid treating a single error or successful direct link as universal embed evidence.
• Check player readiness separately from API loading, with timeout recovery and late callback protection. Verify paused seeks, return state, completion and replay.
• Preserve origin-scoped preview data and verify the loaded entrypoint/dependencies.
• Add applicable and nearby non-trigger policy cases. No new runtime, external service, model rule or third-party skill is introduced.

1.16.0 - 2026-09-20

• Prefer the complete selected upstream specialist over a shortened router summary when its distinctive workflow fits; add exact Meng and Dream Loop entrypoints.
• Distinguish loading a workflow, installing selected complete folders and adapting shared principles; preserve required dependencies and source/version provenance.
• Make Dream Loop an explicit optional specialist and label the lighter visual loop as an unbenchmarked adaptation, without imposing its pass limits on the original.
• Preserve existing selective Meng bundles, user constraints and current host rules; no new third-party skills installed by this release.

1.15.1 - 2026-09-20

• Add final acceptance on saved CMS/page-builder pages, including representative editor save/reload persistence when native editability is required.
• Tie review evidence to the saved version and relevant dependencies; reassess only affected flows after material changes. Preserve live-site authorization boundaries.
• Add applicable and nearby non-trigger policy cases; no new CMS runtime trial claimed.

1.15.0 - 2026-09-20

• Clarify routing for public pages that are also the working experience.
• Add conditional audio/video control, synchronization and recovery evidence.
• Add focused question examples for unresolved purpose, navigation, sound entry and physical use; retain settled preferences and avoid routine approval rounds.
• Record selected X references with observed-versus-claimed limits and manual policy cases. No external code, assets, paid service or additional skill installed.

1.14.0 - 2026-09-19

• Add a conditional visual-target/build/critique method informed by Dream Loop, with bounded refinement, optional independent review and working-interaction evidence.
• Keep image generation, Blender and paid asset services conditional; retain user constraints and Work Router ownership of models, effort and delegation.
• Remove the separate Fable-only-for-long-transformations restriction.
• Strengthen interactive app checks for real state changes and requested exports.
• Align field-guide versions and stale writing-route data with the current skill.
• Record source review and manual applicability cases; no external package installed.

1.13.0 - 2026-09-17

• Move ui-router:ui-builder from Sonnet high to Opus low. Sonnet high is off the efficient frontier; Opus low reaches higher quality for comparable burn.
• Escalate a stalling builder to Opus medium rather than raising Sonnet effort, and keep Sonnet low/medium for the mechanical tail.
• Warn that Opus low keeps replies long and delegates readily, so the builder needs an exact file map and the sole-writer rule held.

1.12.0 - 2026-09-17

• Add conditional Neuform and Aura routes for visual-direction exploration, HTML prototypes and reusable handoff material; distinguish those outputs from CMS or page-builder integration.
• Add ThreeUI Pro's entitled-source route alongside Community, with compatibility, fallback and destination-runtime checks for spatial enhancements.
• Keep vendor capability, current entitlement, successful retrieval or export, destination integration and observed behavior as separate evidence states. Retry a relevant authorized connected service when prior access results may be stale.
• Align UI Router manifests, marketplace metadata and field-guide version data.

1.11.0 - 2026-09-06

• Make concrete composition decisions and bounded visual refinement part of design steering, preserving accepted direction and purposeful expression.
• Check coherence across the requested surface and use a contextual removal test instead of blanket bans on visual techniques.
• Add conditional custom-motion checks for input alternatives, semantic text, property ownership, lifecycle, quality budgets and observed performance.
• Strengthen full-page capture and reference-adaptation checks when relevant.
• Record selected practices from Meng To's Skills, threeui, kage and sketchbook repositories, with pinned sources and manual applicability cases. No external skill pack, code or preferred graphics stack is installed by this update.

1.10.0 - 2026-09-06

• Show concrete references and component samples at useful decision points, including focused comparisons during iterations and a working slice before costly expansion.
• Infer audience and expression from each project; remove audience-only creative mismatch gates and distinguish AI-assisted builds from AI-facing experiences.
• Add Refero, 21st.dev, Beautiful UI, AI Elements, Agent Elements, PAIR and Transitions.dev to the conditional resource map with access and reuse boundaries.
• Check selected component code, adapt it to the existing system, and verify it in context. Preserve accepted decisions and distinguish samples from working results.
• Align the field guide, discovery descriptions and manifests; preserve the 1.9.0 art-direction and skill-identity improvements.

1.9.0 - 2026-09-06

• Resolve same-named design skills by provider and exact path; do not silently equate originals with opinionated derivatives.
• Add a focused art-direction check for aesthetic-rule conflicts, repeated palettes, and purpose-versus-styling preservation.
• Keep expressive editorial work eligible for a compatible expressive lead, with separate checks for useful content and requested visual character.
• Require reference traits to appear in the implemented experience; verify meaningful motion rather than accepting a reference list or working controls as sufficient.
• Preserve explicit muted/earthy designs and small accepted-system edits; add policy review cases without claiming automated behavioral evaluation.

1.8.0 - 2026-09-06

• Make Mobbin a primary inspiration source and check exposed reference tools before assuming access is missing. Distinguish tool availability from successful retrieval.
• Expand motion inspiration by behavior with direct creator work, experimental web galleries, showcases and technical examples; retain separate curator attributions.
• Require observed motion evidence for timing claims; keep research proportional and preserve free-only constraints without treating every paid tool as unconnected.

1.7.0 - 2026-09-05

• Add conditional design steering: infer the brief, show concrete alternatives when direction is uncertain, translate plain-language reactions, and establish a representative slice before expansion. Retain decisions in project notes.
• Keep small edits and clear references direct; preserve delegated taste and avoid mandatory questionnaires, repeated approval rounds, or global preference writes.

1.6.0 - 2026-09-05

• Delegate general discovery to Work Router when available; retain UI-specific checks and a standalone fallback. Align the field guide and preserve existing install authorization.

1.5.0 - 2026-09-05

• Add conditional product-decision diagnosis informed by the reviewed Layers intro and orientation skills; keep defined visual edits direct and assumptions explicit.
• Add HyperFrames as a free local-video resource, with separate hosted-service costs and export validation. Do not install overlapping animation packs by default.
• Strengthen source-advice evaluation and phase completion evidence without adding another broad design lead or treating a source demo as a verified implementation.

1.4.0 - 2026-09-05

• Added task-specific free visual resources and reference workflows, with public-source attribution and separate paid/unverified candidates.
• Routed ThreeUI Community, Toolcraft and Canvas UI by output; documented asset licenses, browser support and project-local installation boundaries.
• Added Component Gallery, 60fps public examples, UNCUT and website galleries to reference selection.
• Made host inventory guidance explicit for Codex and Claude Code; distinguish skill files, application installs and live MCP access.

1.2.0 - 2026-08-22

• Mirrored Codex Design Router 1.3.1's active Natural Writing route.
• Reclassified Natural Writing as a companion whose availability must be verified, not an assumed install.
• Added trigger and explicit-invocation verification guidance while preserving exactly one final editor.

1.1.0 - 2026-08-19

• Replaced the default three-humanizer sequence with one Natural Writing final editor.
• Added product/UX and marketing copy handoff rules, author-voice preservation, factual-integrity gates, and Claude-pattern cleanup checks.
• Kept the original humanizer sources as explicit fallback references.

1.0.0 - 2026-08-18

• Initial Claude Code-native release.
• Preserved the complete researched UI skill directory and named chains.
• Added Claude-specific discovery, routing agents, and optional installation flow.
• Added deterministic sudo blocking hook.