Version 0.2 adds real local accounts, encrypted locked history/providers, per-session key proof, recovery, help and public downloadable releases. Current measured results supersede the older preview evidence. Public repository visibility is approved; paid/public account hosting remains out of scope. Live model quality, signing and universal integrations remain unverified.
This is a usable development preview, not a claim that the entire universal-language/universal-testing vision is complete. The original workflow informed the design; private group source was not redistributed.
| Requested capability | Current implementation | Remaining acceptance work |
|---|---|---|
| Desktop, browser and push usage | Shared TypeScript/Monaco editor, packaged Electron/frozen service, loopback web API and push analysis | Signed/notarized releases, clean-machine installation testing, secure remote/mobile hosting and native mobile packaging |
| Code/data inputs in any language or format | Twelve fixed CLI adapters and bounded inert binary fixtures | Framework/package managers, multi-directory projects, arbitrary toolchain adapters, document extraction/OCR and format-specific semantic harnesses |
| Original families and all 48 supplied categories | Approved CLI contracts, reviewed pytest harnesses, property tools, coverage/mutation, fuzz engines and a 48-category applicability catalog | Dedicated arbitrary-application GUI/API/browser/load/chaos adapters; authorized external hardware, human and security-test evidence |
| Generate and improve code; real-time assistance | Consented task-routed schema-validated proposals; reviewed application to drafts; independent retesting and correctness-gated comparison | Owner-selected model configuration and held-out live quality/cost/latency evaluation; no live model is configured for this preview |
| Multiple models/APIs/agents | OpenAI-compatible, Anthropic and Ollama task routes; custom models through the local inference API | Task specialization evaluations, additional provider protocols and controlled multi-agent orchestration; arbitrary agent tools are not accepted as trusted host execution |
| Efficiency, readability and safety priorities | Runtime, peak child RSS, footprint, long-line proxy and strict-warning review preferences | Input-size scaling studies, richer readability/security tools and independent hardware/toolchain fingerprints; no inferred universal Big-O or security score |
| Iteration reports and source/test artifacts | Immutable revisions, retained attempts, HTML/JSON/ZIP exports and bounded metric timelines | Model-proposal provenance persistence, larger-history archival, independently supervised measurements for hostile targets and richer comparison visualizations |
| Train new models and serve APIs | Actual parameterized causal-model creation/local fine-tuning, rights/split checks, safe-tensor output and authenticated local inference | Rights-cleared substantial datasets, near-duplicate/PII review, GPU-scale training and evaluation demonstrating useful coding ability |
- Select local/API provider models and repository visibility. Configure credentials privately; measure generated code on unseen projects with reviewed oracles before claiming AI effectiveness.
- Add one framework/compiler/format integration at a time, with positive and negative fixtures, resource bounds, missing-prerequisite behavior and CI evidence. Unsupported inputs must remain unsupported rather than silently reported as successful.
- Add requirements-document extraction for explicitly supported text formats, then dedicated application test environments. Human/hardware categories cannot be satisfied by generated placeholder results.
- Harden distribution and hosting independently: sign packages, verify clean-machine updates/uninstall/history preservation, use real remote identity/quotas and disposable VM workers for hostile workloads.
The owner requested leaving Docker Desktop unchanged. Local compiler execution consequently remains blocked; GitHub Actions supplies genuine isolated integration evidence. There is no permissive host fallback.