[Re-bin]

Summary

• replace raw, nested tool logs with a unified one-line activity language across reasoning, web, shell, files, memory, apps, MCP, images, subagents, automations, goals, and unknown plugin tools
• move streamed answers to Streamdown with resilient partial-Markdown repair, subtle native reveal animation, stable completed rendering, and accessible structured content
• make Apps discovery visually stable and cheaper to render without bundling third-party brand artwork
• add selected-answer follow-ups: highlight an assistant excerpt, send it to the composer, and pass it to the model as bounded, model-only runtime context
• tighten mobile interaction, background polling, long-thread rendering, remote assets, and development-mode stability

Activity behavior

• every default activity item is one line: action, useful target, and status
• reasoning is summarized inline instead of rendering a short heading with an indented paragraph
• web searches remain one row while fetched pages render as independent favicon, title, and compact URL rows
• CLI and MCP actions use product names and human verbs instead of raw function names or JSON
• repetitive file, search, read, and memory operations are grouped without losing useful state
• unknown plugin tools fall back to a safe human-readable action without exposing arguments
• truncated activity rows expose the full value on hover

Streamed Markdown

• Streamdown renders provider updates directly without a second UI timer or artificial text batching
• incomplete emphasis, links, fenced code, tables, and other partial Markdown remain readable while tokens arrive
• Streamdown's native word reveal replaces the custom per-character DOM transformer, including for long CJK answers
• syntax highlighting is deferred until streaming completes, reducing work on every provider delta
• a healthy renderer stays mounted when a response changes from streaming to complete; the error boundary still recovers on phase changes after a real failure
• completed answers contain no leftover animation wrappers
• tables have deliberate surrounding space, horizontal scrolling, keyboard focus, an accessible region label, and reduced-motion support
• existing code blocks, math, task lists, file references, previews, attachments, and safe HTML remain supported

Apps and performance

• deterministic initials render immediately before any remote logo is decoded
• remote logos and favicons load lazily, fade in only after decode, use no-referrer requests, and retain local fallbacks
• resolved and failed app logo URLs are cached for the browser session
• repository-host favicons remain distinct app initials instead of showing one GitHub mark for unrelated tools
• offscreen catalog rows and completed history defer layout and paint with content-visibility
• completed thread units are memoized so history does not rerender for every streamed token
• hidden-tab stream deltas are coalesced and polling/timers pause while the page is hidden
• Settings and secondary surfaces remain lazy-loaded and preload on user intent
• syntax highlighting stays Vite-prebundled so Refractor's CommonJS core does not crash the development WebUI

Selected-answer follow-up

• selecting text in one completed assistant answer offers a compact “Ask about this” action
• the action is clamped to the visual viewport and dismisses safely on scroll, resize, keyboard movement, or outside interaction
• the composer shows the selected excerpt before the user sends the follow-up
• only trusted WebUI envelopes can project quoted_context; ordinary WebSocket clients cannot inject it
• excerpts are bounded, control characters are normalized, runtime-context markers are escaped, and the context is excluded from public session history
• AgentLoop consumes a generic internal runtime-context block rather than depending on WebUI-specific metadata

Safety and compatibility

• credentials, authorization values, API keys, tokens, passwords, URL user credentials, and common provider secret formats are redacted from activity presentation
• absolute home paths and internal tool-result artifact paths are compacted
• private or local addresses are not exposed as clickable timeline targets
• link destinations remain functional while visible labels omit query noise
• existing channel, provider, tool, config, session, CLI, and gateway behavior is unchanged when the new optional WebUI quote field is absent
• the new quote field is additive and model-only; it is neither persisted in visible history nor accepted from non-WebUI WebSocket messages

Code structure

• activity sanitization, trace semantics, reasoning previews, URL safety, page visibility, and remote-logo state live in small focused modules/hooks
• live and completed events use the same presentation model
• streaming and completed Markdown remain behind the existing MarkdownText boundary
• channel-produced runtime context is normalized at the gateway boundary and consumed through a generic agent-core interface
• the final cleanup removes more custom streaming-renderer code than it adds

Scope

The product work is centered on the React WebUI. A narrow additive gateway/agent bridge supports selected-answer follow-ups by projecting trusted WebUI quotes into bounded runtime context. It does not change default agent execution, model/provider behavior, persisted config, existing channel setup, or CLI commands.

Licensing

• Streamdown 2.5.0 is Apache-2.0 licensed.
• no third-party brand SVGs are added to the repository, WebUI bundle, or Python wheel; remote logos remain optional presentation data with local text fallbacks
• product names, remote logos, and brands remain property of their respective owners and are used only for identification
• the implementation uses nanobot's own components and data model; it does not redistribute paid AIcss component source or assets

Verification

• synced with current origin/main
• full WebUI suite: 48 files, 688 tests passed
• focused gateway/agent runtime-context suite: 230 tests passed
• ESLint: passed with zero warnings
• Ruff across nanobot and tests: passed
• production TypeScript/Vite build: passed
• git diff whitespace check: passed
• browser smoke: long streamed Markdown, lists, blockquotes, tables, links, and activity rows rendered with zero console errors
• browser regression: fixed the dev-only Refractor/CommonJS crash that unit and production-build checks did not expose
• full local Python suite: 5,158 passed and 17 skipped; two workspace-scope tests could not invoke bare python under their deliberately restricted PATH on the local macOS runner