Changelog

2.1.0 — 2026-09-18

Changed

• rebranded the public product from Mixmatter / Mixmatter to Mixmatter
• retained the technical repository, package, Skill, and web slugs as mixmatter for compatibility
• replaced the previous offset-sheet badge with a layered material icon built from source, photographic, halftone, tape, and marker elements
• added coordinated light and dark icon variants with identical geometry and corrected depth ordering
• shifted the active identity to a cool blue / cyan / lavender material palette
• updated plugin-facing identity and public documentation while keeping the 2.0 reconstruction behavior intact

Added

• assets/mixmatter-icon-light.svg
• assets/mixmatter-icon-dark.svg

2.0.0 — 2026-09-11

Added

• seven research-derived canonical system documents covering runtime, art direction, core constitution, regression benchmarking, shared schema, capability contracts, and visual grammar
• staged reasoning model: READ → UNDERSTAND → PROTECT → DIRECT → RECONSTRUCT → MATERIALIZE → CRITIQUE → REVISE
• source-specific preservation contracts using must preserve / should preserve / may transform / may remove
• semantic-anchor, visual-anchor, structural-relation, identity-invariant, low-information-field, and transformation-budget concepts
• source-specific direction hypotheses rather than generic style presets
• cause-level critique and revision continuity
• source-type preservation defaults and a 40-image regression benchmark design
• natural-language Direction, Structure, and Intensity semantics

Changed

• upgraded Mixmatter from a reconstruction prompt system into a research-backed AI art-direction and visual-reconstruction system
• vague requests now default to autonomous source-specific judgment; clarification or alternatives are reserved for genuine ambiguity
• revisions preserve successful crop, hierarchy, locks, identity, and material decisions instead of behaving like random rerolls
• alternative directions return to the original source image unless the user explicitly asks to build on a prior result
• material effects are now explicitly causal and bounded rather than mandatory style signatures
• strengthened the default requirement for visibly reconstructed composition so restraint does not collapse into ordinary photo styling
• retained the established v1 visual language as rendering authority while adding 2.0 reasoning and preservation intelligence
• retained strict source-text protection and zero-new-text default
• public product name standardized as Mixmatter while package/repository identifiers may continue to use mixmatter

Architecture

• remains a Skills-only Agent Skill / OpenAI Plugin architecture
• no MCP server is required
• no custom ChatGPT host UI is required
• no external Mixmatter image-generation backend is required
• host-native image understanding and generation/editing capabilities are used when available
• future web or backend runtimes may implement the same six capability contracts without redefining the core product

Removed

• ChatGPT creation-card and result-card implementation
• MCP widget server and host bridge
• UI-specific runtime dependencies, state assumptions, and host UI specifications
• prompt behavior that required a card or widget for vague requests and revisions

Preserved from the abandoned UI experiment

• Direction / Structure / Intensity as product semantics
• revision continuity
• distinction between revision-from-result and alternative-from-source
• Typography, Restore, and Custom as interpretable direction families where relevant
• the final Mixmatter product logo

1.0.2 — 2026-09-10

Changed

• rebuilt the master prompt around stronger flat, planar, and designed-2D-surface composition
• added the default zero-new-text rule for sources with and without visible text
• prohibited translation, parallel bilingual versions, and second-language equivalents of monolingual source text
• added explicit handling for source content that is already bilingual
• added the exact-user-wording exception without permitting extra copy, translation, or unrequested bilingual variants
• added text-footprint guidance of about 15% total readable area and about 8% for any single readable block unless the user explicitly requests larger treatment of their exact supplied text
• added high text-density scene guidance that preserves typographic density rather than typographic completeness
• strengthened halftone as a clearly visible, intentionally varied structural device
• strengthened torn-paper collage through visible tears, cut-paper overlaps, rough interruptions, pasted fragments, and deliberate paper-layer transitions
• synchronized the universal Skill, OpenAI-packaged Skill, master prompt, packaged prompt reference, quality rubric, manifest, listing copy, and review tests
• replaced the truncated plugin icon with a safely decodable 1024×1024 RGBA PNG exported from the canonical dark SVG mark
• aligned the plugin manifest with the current ingestion schema by adding capabilities and removing the unsupported supportURL field
• bumped OpenAI plugin metadata to 1.0.2

Text-control contract

A compliant v1.0.2 result adds zero new text by default. Source text remains in its original language and may be retained, cropped, obscured, fragmented, or reduced. When a user explicitly supplies text to add, only that exact wording may appear; no translation, bilingual duplication, subtitle, caption, label, or filler copy may be added.

1.0.1 — 2026-09-10

Changed

• replaced the ambiguous text-free wording with an explicit no-new-typography rule
• added a four-part source-text policy for incidental, scene-identifying, identity-critical, and text-dominant sources
• added the governing principle: Source text is content. New typography is generation.
• added a fallback rule: if exact source text cannot be preserved reliably, crop, obscure, simplify, or retain it as photographic texture rather than hallucinating a replacement
• strengthened hard failures for pseudo-text, filler editorial copy, translated/rewritten/duplicated source text, enlarged source-text headlines, and approximate hallucinated replacements
• reduced prompt language that strongly implied typography-led poster or magazine-cover design
• clarified that Mixmatter hierarchy must come from crop, scale, color, texture, source-derived geometry, overlap, and negative space rather than generated headlines or body copy
• synchronized the universal Skill, OpenAI-packaged Skill, master prompt, packaged reference, and quality rubric around the same text policy
• updated OpenAI starter prompts to explicitly prohibit new text while allowing source text to survive selectively
• bumped OpenAI plugin metadata to 1.0.1
• aligned manifest developer name with the verified public developer identity Fan Jiale

Regression target

The v1.0.1 hotfix specifically addressed source photographs containing signage triggering newly invented editorial headlines, filler copy, duplicated source words, or pseudo-text.

OpenAI Plugin packaging — pre-2.0

Added

• .codex-plugin/plugin.json for OpenAI plugin packaging
• skills/mixmatter/SKILL.md as the OpenAI-packaged Skill entry point
• self-contained bundled references for the master prompt and quality rubric
• host behavior for missing-image handling and direct image generation/editing when available
• brand, privacy, terms, support, evaluation, and listing materials
• repo-scoped development marketplace metadata

Architecture

• Skills-only plugin with no MCP server, authentication layer, or external runtime dependency
• root SKILL.md remains the canonical universal Agent Skill entry point
• behavioral changes require a versioned release and changelog entry

1.0.0 — 2026-09-09

Initial Mixmatter image-only release.

Added

• source-aware reconstruction workflow
• structural anchor preservation
• Disassemble / Recompose / Reassign / Reduce / Hierarchize method
• selective photographic, printed, graphic, and collaged visual states
• source-derived graphics rule
• limited-palette print language
• selective halftone guidance
• collage discipline rules
• explicit failure modes
• 100-point quality rubric
• 10 canonical high-resolution before/after showcase plates across multiple source categories

Baseline status

• v1.0.0 established the initial image-only core baseline
• platform-specific packaging may add manifests, metadata, assets, or synchronized Skill copies
• behavioral changes require a versioned release and a changelog entry

Intentionally excluded

• generated typography
• inferred place names
• captions and metadata
• typography-led poster layouts