Spaces:
Paused
Paused
Commit History
-
Persist chat tools and stream artifact drafts
Verified signatureStores the tools toggle per chat, keeps warm prefill telemetry live before native batch metrics arrive, and samples partial artifact files into a transient tree and sandboxed preview during generation.179a355 -
Make reasoning preview behave like a terminal line
Verified signatureLeft-aligns the collapsed reasoning preview, allows natural wrapping, pins the full trace to the bottom of a one-line clipped viewport, and removes the centered ticker animation.4daed10 -
Collapse live reasoning and suppress warm prepare
Verified signatureKeeps reasoning collapsed by default with a rolling one-line live preview. Warm consecutive requests remain Ready until real prefill/decode instead of showing a synthetic graph-preparation phase.750f06c -
Clarify warm inference preparation
Verified signatureTracks whether the first inference graph has already produced progress so later requests no longer claim to be recompiling the first graph after Stop.08ae3c0 -
Flatten artifact file tool arguments
Verified signatureReplaces nested and bundled artifact source arguments with flat file_N_path/file_N_content pairs so Qwen tool parsing does not cross nested JSON boundaries. Keeps legacy chat compatibility.af9299f -
Keep model resident on cancel and harden artifact tools
Verified signatureCancel now waits for native abort without restarting the worker. Adds explicit unload, loaded-model tab title, raw file bundles for symbol-safe artifacts, and disables prompt-cache reuse for Qwen3.5 tool runs.643dfc6 -
Persist tool logs and recover chats with invalid artifacts
Verified signatureStore tool events with each IndexedDB chat session, reconstruct logs for older records, and keep a malformed legacy artifact in failed code-only mode instead of dropping the whole conversation during hydration.b6aacab -
Fix artifact tool termination and dynamic transcript scrolling
Verified signatureReject structurally truncated HTML before browser evaluation, stop the agent tool loop after a passing artifact, and make the transcript own dynamic expanded-content scrolling across desktop, mobile, and 4K layouts.328256a -
Remove duplicate shard verification pass
Verified signatureSkip the redundant full-Blob SHA-256 scan before mandatory llama.cpp GGUF loading; retain pinned manifest metadata and exact byte-length checks.dd0c917 -
Warm the active chat prompt shape
Verified signaturePass the current chat system instructions into the WebGPU warm-up. Empty and custom instructions now use their matching chat template; changing instructions invalidates the loaded configuration and requires an explicit reload.45c238a -
Warm WebGPU before first prompt
Verified signatureExpose model warm-up and first-inference preparation in live telemetry. Prime long prefill and short prompt-cache batch paths on Chrome JSPI while preserving Safari compat behavior.3349d47 -
Use manifest context limits and handle memory exhaustion safely
Verified signatureDefault context and completion to 8192, expose each model contextLength as the maximum for both controls, and turn context-full, allocation failure, and WebGPU OOM signals into recoverable UI errors.fa59dd6 -
Set 8K context defaults and add artifact ZIP export
Verified signatureUse an 8192-token default context across every model, expose model-specific context ceilings, and let users download the current multi-file OPFS artifact as a ZIP archive.d0973a0 -
Add persistent multi-file artifact workspaces
Verified signatureAdd OPFS-backed multi-file artifact deployments, file tools and repair loop, opaque client-side VFS previews, per-chat restoration, Files/Rendered UX, and deterministic browser security/persistence coverage. WebGPU runtimes and model assets are unchanged.33e51e5
-
Render ordered agent flow and edit messages
Verified signaturePreserve reasoning, text, tool calls, structured results, and artifacts in chronological inline segments. Reconstruct persisted multi-round tool history and let users edit a turn, truncate its branch, and regenerate.f3ad26c -
Accept expression input in js_eval
Verified signatureKeep the disposable worker sandbox while accepting either a single JavaScript expression or an explicit-return function body, matching real small-model tool output.399c79e -
Add agentic artifact tools and live diagnostics
Verified signatureExecute HTML artifacts in an isolated iframe, return console diagnostics, support stable-ID inspect/update/retest loops, add chart and confirmed Wikipedia search tools, and expose a tool reference in the composer.897170b -
Add resizable sidebars and inline artifacts
Verified signaturePersist draggable sidebar widths with snap collapse and reopen. Move artifact rendering into chat with HTML and Rendered tabs, fullscreen preview, and sidebar navigation. Includes component and headless UX regression coverage.7142bcd -
Ship validated WebGPU Q1 turbo runtime
Verified signatureChrome exact-token A/B: raw Q1 vec4 +18.9% median on 1.7B; accepted 27B optimization sequence about +26% GPU throughput-equivalent. Final trace-free Chrome and physical Safari 26 gates passed on 1.7B and 27B with one graph split, zero CPU ops, and full WebGPU layer placement.3d95ffd
-
Promote optimized Chrome WebGPU runtime
Valeriy Selitskiyc56db82 -
Update production bundle for 4096-token defaults
Valeriy Selitskiy1d8700e -
Update production bundle for tool-call hotfix
Valeriy Selitskiy432e8f2 -
Release v1.2.2 interactive chat workspace
Valeriy Selitskiy21ad36a -
Release v1.2.1 adaptive chat workspace
Valeriy Selitskiy6c62075 -
Release v1.2.0 live model telemetry
Valeriy Selitskiyc3633b1 -
Release v1.1.4 manifest progress fix
Verified signatureCorrect manifest progress accounting and document verified Safari and Firefox gates; Windows hardware gates remain open.20134b6 -
Release v1.1.3 scored WebGPU evidence
Verified signatureAdds browser state-drift scoring and release evidence, refreshes the pinned wllama artifacts, and keeps DFlash inspection-only and disabled by default.3fd9ca4 -
Release v1.1.2 benchmark runtime evidence
Verified signatureAdd fail-loud Flash/KV benchmark controls, requested-versus-observed runtime reporting, controlled JSPI/compat selection, per-tier DFlash disposition, tests, provenance, and a prebuilt static bundle.55c2c6a -
Remove superseded application bundle
Verified signatureeb4b56a -
Publish 27B-default static bundle
Verified signature82eefac -
Remove stale multipage benchmark bundle
Verified signature2b228a6 -
Publish single-entry benchmark-enabled static bundle
Verified signatureb23c0fa -
Publish Phase 5 benchmark bundle
Verified signaturebeac9c3 -
Publish verified static production bundle
Verified signature5062906