Project status, user data, and model assets remain continuous across continuously running AI products.
AI PRODUCTS BUILT FOR WORK THAT CONTINUES
Build a sustainable desktop product around the three objects most likely to get out of control in a long-term AI product—project status, user data, and model assets. AI Project Harness, Local Chat and LoRA & TTS Studio have entered real projects, long-term daily use and training tasks respectively.
Project starting pointPersistent multimodal interaction is the earliest stress scenario, which simultaneously exposes the state fracture between long-term sessions, asynchronous generation, model assets, perception and expression, and thus splits three independent products.
Main line of projectAI Project Harness manages long-term projects and Agent division of labor, Local Chat manages local sessions and multi-modal tasks, and LoRA & TTS Studio manages training, quality screening, and model asset reflow. The dual-agent workbench, model horizontal evaluation, and early multi-modal experiments provide the basis for collaboration and verification samples.
The threshold for complex products is not just a certain technology, but the goals, context, quality standards and assets that must be maintained over the long term. AI provides more and more execution capabilities, but these capabilities are still fragmented across sessions, models, vendors, and tools; without organizational structure, they are difficult to accumulate into productive capabilities that individuals can use repeatedly.
Three products together complement this set of organizational capabilities: AI Project Harness allows multiple Agents to continue working on the same project, Local Chat precipitates long-term interaction and multi-modal creation into an environment that individuals can own, and LoRA & TTS Studio organizes training data, remote computing power, and model assets into a repeatable production chain. The ultimate goal is to increase the level of product complexity an individual can sustain over the long term.
The first three layers correspond to three independent products: long-term project continuity, local multi-modal interaction, model training and asset production. The fourth layer summarizes the Claude Code × Codex dual-engine workbench, evaluation methods, perception and embodied prototypes, as the basis for collaboration, verification methods and early application scenarios.
Windows local desktop MVP v0.0.20: Project/Objective/Task/Run/Session/Checkpoint hierarchical persistence, CTO, Review and task sessions collaborate through Mission Packet; only the checkpoint after user acceptance advances Project HEAD.
Current source code baseline v1.8.226: 48k token text budget, 28 rounds of post-compression, SessionRoster unifies current roles; 36,672 role entries enter the deduplication index, and image and video tasks write back the original session with sessionId and immutable message anchors.
Local model production workstation: The real DLRB job forms 237 train-ready images from 101 source images, and returns 228 MB products after 2000-step LoRA training; the TTS module retains independent real smoke boundaries.
Claude Code and Codex share the engineering constraints, self-developed Skills, Hooks, memory and relay status in the vault; environmental fingerprints and drift checks identify changes in code, dependencies and operating conditions. It is responsible for cross-Agent information exchange and remains independent from the project status management of AI Project Harness.
The early state hub of persistent multi-modal agents: MLLM router, persona state and cross-modal dispatch have formed prototypes to verify how complex scenes consume structured states; long-term memory and emotion model have not yet closed the loop.
Visual recognition, face following and VAD voice activity detection run in the local resident daemon as a capability experiment for active response and multi-modal input; the sensitivity and upper-level decision-making closed loop are still being polished.
Early Embodied Expression Prototype: Use Audio2Face, UDP LiveLink and TCP text links to verify the layered transmission of expressions, mouth shapes and dialogue, retained as a historical experiment on the multi-modal output side.
Evidence covers real session replays, badcase attribution, role and status specifications, public upstream PRs, operational tools, and content platform feedback. Each item can be linked back to a public link, version, sample, or specific product mechanism.
Fixed character cards, first scene and original seven rounds of input, real round-by-round playback of DeepSeek, GLM 5.3 Flash, Qwen 3.8 Flash; comparison of character consistency, language quality, fact drift, rejection, latency and interface stability.
Record forgotten settings, state string writing, output schema instability, and multi-modal failures from real product iterations, and press "Phenomena → Attribution → Repair → Real Test" to retain problems, changes, and regression results.
The role card writes the identity, tone, relationship boundaries and situational constraints into consumable specifications of the model and is a long-term state input for Local Chat; the page retains the desensitized excerpts and anti-homogenization methods of the real running card.
Use deterministic code to undertake model candidates: private and public layer isolation, life cycle advancement, behavioral intent compilation and schema verification, so that long-term state writing and changes can be checked.
Focusing on real issues such as manual context compression, same-session continuation, and false positives written outside the project, 4 public PRs were submitted, covering reproduction, interaction solutions, implementation, and regression testing.
Public tools:tribunal-skill、refute-everything、ZC-gh-state-diff、ZC-image-batch-audit. The Civitai account has accumulated 4,034 public picture interactions in the 11 days since its debut.
Let’s first look at the three core products that have already undertaken real tasks, and then look at the supporting projects that provide them with connection, verification, and stress scenarios.
live · Already running
AI Project Harness v0.0.20 — Hierarchical project status, CTO/Review/Task Session, Mission Packet, Post-review Checkpoint and Project HEAD have been run in three real projects.
Local Chat v1.8.226 — Current source code baseline: 48k token text budget, 28 rounds of post-compression, SessionRoster, 36,672 searchable role entries, and image and video tasks written back by sessionId + immutable message anchor.
LoRA & TTS Studio — DLRB LoRA has a complete evidence chain from 101 source images, 237 train-ready images to 228 MB product; Style-Bert-VITS2 is currently only marked by verified smoke.
agent-harness — Claude Code × Codex shares engineering constraints, Skills, Hooks, memory and relay states, and environment fingerprints and drift checks are used to identify context changes.
building · The skeleton is in place and is still being polished
evie-agent — MLLM router, persona state, and cross-modal dispatch continue to evolve as prototypes of complex Agent scenarios, and long-term memory and emotion models have not yet closed the loop.
perception — Vision, face following and VAD have been run through in the resident daemon, but sensitivity adjustment still needs to be completed to participate in the upper-level decision-making closed loop.
exploring · Still testing hypotheses
ue-realtime — MetaHuman Early prototypes of expressions and mouth shapes are retained as embodied expression experiments and are not currently on the near-term delivery path for the three core products.
Next step
The three products deal with different objects and form the same engineering main line: connecting the model to long-term projects, long-term use and long-term asset production, and constraining each change with status, ownership, acceptance and recovery mechanisms. Persistent multimodal Agent is the earliest application scenario of this main line.
Zhou ChengZHOU CHENG
SYSTEMS & AGENT ENGINEERINGEmail address copied
lucaszhouc@gmail.com
