Continuous updates SITE / CHANGELOG

Website log

This site is part of an ongoing project, and the content is continuously revised as the project progresses. Every visible change is recorded here - only what actually happened, not the plan.

  • 2026-10-12
    14:09 JST
    Chinese demo repair reached production.
    • Fixed the stale redirect rules and shared-asset paths that broke the Chinese demo routes on mobile.
    • Re-checked five Chinese demos in an iPhone viewport: HTTP 200, zero console errors, zero failed resources and no horizontal overflow.
  • 2026-10-05
    17:32 JST
    Roguelike Your Life v5 full Skill package went live.
    • Published the complete 18-file package with manifest, full bundle, entrypoint, schemas, docs, tools, examples and a downloadable ZIP.
    • The launcher now uses a package-loading bootstrap and reports PACKAGE_LOAD_FAILED when the package cannot be read, instead of pretending a short runtime is the full Skill.
  • 2026-10-05
    17:04 JST
    Roguelike Skill transport v4 went live.
    • Added the compact transport contract, absolute runtime and manifest URLs, debug diagnostics, deep-link handling and a fetch-failure fallback.
    • This transport was later superseded by the v5 full package on the same day.
  • 2026-10-05
    16:28 JST
    English-default bilingual site went live.
    • The root route became the natural English version and /zh/ retained the Chinese site; copy, typography and spacing were redesigned for English instead of doing literal string replacement.
    • Added the EN / ZH switcher and completed desktop and 390px route checks for the published pages.
  • 2026-09-19
    • The three demos were changed to public shells of the active product front-end: Local Chat retains the four workbench and input interaction, Harness reuses the v0.0.20 Renderer bundle, and Studio retains training/testing/subpackaging/queuing/dubbing SPA; back-end connections, private data, and non-public content are all removed.
    • Local Chat public demo locks the active warm brown scroll theme and removes theme switching; the project details page is displayed in a fixed 1440×900 desktop viewport and is no longer mistakenly cut into a mobile layout along with the outer container.
    • The three core products on the homepage have added Project/Demo dual entrances; six canonical routes have been completed, and the old Local Chat and Studio screenshot demos have been migrated to the current capabilities page.
    • The Local Chat caliber is updated to the v1.8.226 source code baseline; LoRA / TTS Studio is updated to Edge app-mode, and generalized expressions are replaced with a real evidence chain of 101 → 237 → 2000 steps → 228 MB products.
    • The final published picture and corresponding plot of Local Chat are still selected by the author himself; the current page retains a clear placeholder and is not automatically selected from the historical gallery.
  • 2026-09-17
    • The website copy is aligned with the 2026-09 resume facts: the three core products are changed to the "positioning / key mechanism / verifiable results" structure; the homepage WHY is clearly stated as "allowing individuals to have the organizational capabilities to continuously promote complex products", and the front-end layout and core interactions remain unchanged.
    • Local Chat adds 48k tokens, 28 rounds of compression, SessionRoster, 36,672 role entries and asynchronous message anchors; LoRA & TTS Studio adds multi-index film development, manual review, Style-Bert-VITS2 and high-impact operation confirmation gate.
    • The model evaluation has been updated to include fixed character cards, first act, and original seven-round real session replays of DeepSeek, GLM 5.3 Flash, and Qwen 3.8 Flash; the five-round results of the old seven models have been reduced to historical experiments.
    • The homepage has added 4 publicly verifiable Open Design PRs, 4 public tools, and evidence of 4,034 image interactions in Civitai's first 11 days.
  • 2026-09-16
    • SENTIENCE, split-screen structure, fluidity, scrolling handover, page navigation and runtime switching interaction are retained, and only the content hierarchy is reorganized: AI Project Harness, Local Chat and LoRA & TTS Studio become three core products.
    • Added a new independent AI Project Harness v0.0.20 details page to clarify Project HEAD, 2+N Session, Mission Packet, audit and evidence write-back; the original agent-harness is retained as an independent support project for Claude/Codex contract, memory, skill and relay interoperability.
    • The Local Chat caliber is updated to v1.8.221, which is officially running, and local long-term status, asynchronous attribution, candidate isolation, atomic cutting, data reconciliation and failure rollback boundaries are added.
    • LoRA & TTS Studio has been upgraded from auxiliary capabilities to independent model adaptation and asset production products; MetaHuman, perception, character cards and emotion evaluation will continue to be used as sources of capabilities, product evidence or early exploration, and historical project pages will not be deleted.
    • The second screen of the homepage is rewritten as personal engineering judgment: the model is responsible for probabilistic candidates, and the product is responsible for advancing the real work through persistent state, explicit attribution, evidence promotion and failure recovery.
  • 2026-08-30
    • The homepage loading is upgraded to a minimalist ZC system interface: black field, acid green status signal and non-uniform radial emulsion particles jointly complete the development from the loading page to the text of the work.
    • The entrance is completed to skip, reduce dynamics, save traffic and WebGL failure downgrade; after development, it still enters the existing complete work structure. The movie quick cuts and face lock materials in the local experiment are not released this time.
  • 2026-08-29
    • lora-tts-studio has been adjusted from MVP to live: LoRA training, remote pod queue orchestration and TTS sound cloning have taken on daily tasks; automatic failure recovery and cross-model parameters are recommended to continue to be maintained as enhancements.
    • The homepage status timeline, project details page and original contribution matrix are updated simultaneously, retaining the true boundaries that still need to be manually confirmed, and do not write enhancement items as core links that are not completed.
    • The brand in the upper left corner of the entire site is simplified to "Zhou Cheng"; the browser icon is replaced with the ZC interlocking word mark in black, white and warm gray, and the misused LoRA Studio project icon is removed.
  • 2026-08-28
    • The Local Chat display has been updated to the official desktop version v1.8.31: the old PWA iframe is replaced with a desensitized interactive demonstration of warm brown and bright desktop, adding plot scrolling, three directions, character selection, reading, 8-category settings and multi-player wide-frame stage; the demonstration does not connect to the backend, and does not read real characters, sessions or keys.
    • The latest capability caliber of site-wide synchronization: OpenRouter SSE, network word collection/front-end display layering, asynchronous message position freezing, picture history recovery across restarts, card archiving/top/AI card change/version history, 2-4 independent scene members, and deterministic dress priority of "subsequent dress change > card-level basic dress > registry default dress".
    • Project narrative adjustment: Desktop Local Chat becomes the main display screen of talking-with, PWA remains as a browser/mobile lightweight branch; the homepage, project details page and original contribution matrix are updated simultaneously, and the overall visual system of the website remains unchanged.
    • Supplementary official version nodes: v1.8.27 two-person card drawing, v1.8.28 warm brown/black and purple plot scroll, v1.8.29 multi-person scene, v1.8.30 two-person/three-person wide format and v1.8.31 card-level outfit priority.
  • 2026-08-25
    • Secondary design of the site: On the premise of retaining the six components, four types of emotional evidence and all detailed entrances, reconstruct the first screen runtime main axis, chapter rhythm, acid signal system, loading animation and contact information final chapter; complete the return of low-height desktop and 320px mobile terminals.
    • Pre-launch text review: Update the Claude / Codex native adaptation of agent-harness with a single front-end entry caliber, remove the outdated memory quantity and taste routing; change the PWA link to the accurate HTTPS → Workers → LLM API upstream request description, and make it clear that pod file transfers are transferred through Hugging Face.
    • The content of personal works has been updated, and the visual system remains unchanged: two interactive surfaces of Local Chat v1.8.18 and talking-with PWA have been added, agent-harness task relay, CompactAnchor, Hermes and hierarchical knowledge routing have been rewritten, and 4 real Badcases have been published.
    • Added GitHub / Civitai public works entrance; corrected Emotion vector → Audio2Face / LiveLink existing prototypes and unified Runtime Trace unfinished project boundaries.
  • 2026-07-13
    • The Emotion Eval page was upgraded from the evaluation framework page to the first batch of Hengping results: "Seven-model emotional interaction Hengping (preliminary controlled)" - 7 models (GPT-5.4 Mini / GPT-5.6 Sol / Claude Sonnet 5 / Claude Fable 5 / Gemini 3 Flash / Grok 4.5 / DeepSeek V4 Pro) were connected to the same single model dual-call emotion runtime, and ran the same five-round multi-modal dialogue (35 final responses). The page contains: hard indicator table (JSON first-time success rate/first-round rewrite/number of scenario candidates intercepted by code/cost/latency median and range), three positioning of production adaptation (no total score ranking), original text comparison of the four models with the highest differentiation, six experimental restrictions and architectural neutrality statement.
    • The status copy of Emotion Eval is updated simultaneously on the homepage and related pages; the number of models and scenes in an outdated trailer copy on the homepage are corrected.
  • 2026-07-11
    • The state modeling page has been revised and launched: "Structured Emotional Representation and State Runtime" - Added an overview of emotional state representation (conceptual level structure), five-layer information flow architecture and four key designs, controlled smoke test indicators with sample denominators, and three real output cases; the bottom layer is the v8.4 prototype (single model dual call, model proposes candidates, deterministic code manages persistent state).
    • The character card page has been revised and launched: from the demo sample card to "Runtime Character Card: Joi Instance Disassembly" - the eight-layer anatomy content is replaced by an excerpt of the card used in actual operation (research adaptation, the copyright belongs to the original author), a 30-second summary, an excerpt of the main public basis and a skip scrolling entry are added; the scroll wheel takeover is cancelled, and the static degradation of the small-height screen is compensated.
    • The status and copy of the two cards in the emotional data section of the homepage are updated simultaneously.
  • 2026-07-07
    • Added "Emotional Data" independent section (homepage) and Emotion Eval evaluation framework page - the multi-model emotional scene evaluation is in progress, and the results will be released after manual review. The page currently only contains the target and framework, not the data.
    • Clean up outdated descriptions related to unconnected voice input (STT), and align the site-wide caliber with the current actual capabilities of PWA (Current state on the voice side: TTS output is online, STT is not yet connected).
  • 2026-07-01
    • Site internal metadata files (build scripts/screenshots/status files) are moved from the public directory to fix 404 rules.
  • 2026-06-28
    • Six component detail pages are online: talking-with / agent-harness / lora-tts-studio / evie-agent / perception / ue-realtime — each page includes system location, current status and known limitations, and iteration log.
    • Deployment process solidification: Cloudflare Pages Production branch + online verification after deployment.
  • 2026-06-26
    • The first version of the site is online: Homepage (Vision / Three-tier Architecture / Progress / Contact), deployed on Cloudflare Pages.