Screens

Screens

Source: apps/web/src/screens — ChatScreen.tsx, EngineScreen.tsx (+ engine/ tabs: BasicsTab, PerformanceTab, AdvancedTab, ProfilesTab), ModelsScreen.tsx, SettingsScreen.tsx (+ settings/ tabs: EngineTab, SafetyTab, AgentTab, AboutTab), CoderScreen.tsx, LogScreen.tsx.

Chat (default)

Streaming chat over the engine's OpenAI-compatible API.

• Left rail — conversation list (persisted, newest first, title = first user line, model + message count, delete).
• Header — model select (the engine's public model_id), engine-ready indicator, active-param badges (thinking mode/effort, max tokens, greedy).
• Messages:
  • user bubble (right) with attachment chips;
  • assistant turn (left): collapsible thinking block (raw reasoning_content, accent icon), markdown body (GFM), streaming caret while in flight;
  • meta footer per message parsed from the engine's timings + usage: finish reason, TTFT, prompt tok/s, decode tok/s, cached tokens (prefix reuse), in/out tokens, thinking tokens, speculative draft accepted/total (%).
• Composer — auto-grow textarea (Enter send / Shift+Enter newline), image/video attachment (→ base64 data-URI parts; requires engine --vision), stop button (AbortController) while streaming, and a params popover: thinking on/off, reasoning effort (template default/low/medium/xhigh), max output tokens, temperature, top-p, top-k, min-p, presence/frequency penalties, seed, greedy, system prompt, preserve-thinking.
• Type /compact to ask the engine to condense the whole conversation into a structured checkpoint summary, which replaces the thread and becomes its starting context (handy before a context-limit warning).
• Offline banner — if the engine is down, chat shows a warning with a jump-to-Engine action instead of failing silently.
• Context-limit awareness — live "% of max-context" usage, warns before the engine would truncate (approaching >75%, near-full >90%).

Engine

Top: status row — engine state (stopped/starting/running/stopping/failed/external), loaded model + uptime + port, GPU memory bar + utilization + process list, Start/Stop action. Fail reasons surface inline (e.g. "weights require 18 GiB, only 3 GiB free" from the engine's own FATAL line). The status row and the Generated launch command card (below) stay visible across all tabs; everything else is split into four tabs — Basics / Performance / Advanced / Profiles — sharing one EngineProfile state object (no per-tab local state), so switching tabs never loses an in-progress edit.

Generated launch command card — exactly the argv passed to ninfer-serve (flags omitted when unset → executable defaults), copy button, "long-context preset" shortcut. The same profile→argv function drives the generated command and the real spawn, so what you review is what runs.

Basics tab — Presets, artifact pick + --model-id/--api-key/--host/--port/ --device, models directory.

Presets (one-click, replace the profile, review-before-start):

Preset	Purpose
Default	executable defaults (bf16 KV, no spec, C=1, 8k ctx)
Long context MTP3	240k ctx, FP8 KV, C=2, 2 dev + 8 host state slots, 8 GiB host KV, MTP3 + optimized head, vision, preserve-thinking
Chat MTP3	32k ctx, auto KV, MTP3, C=2
Vision multimodal	81,920 ctx, vision + media budgets, MTP3
Max concurrency (C=8)	8 lanes + matching state slots for aggregate decode
35B-A3B DFlash / 3.8-27B DFlash2	per-architecture speculative backends (7 drafts)

Plus RTX 5090 presets calibrated from the engine's own measured KV density (192k long-context MTP, 128k bf16, C=4 nvfp4 serving, 96k low-latency, 128k MoE, two 256k ultra-context presets at the variants' full 262,144-token native window), and a community-validated 262k fp8 profile ported from headpiece747/ninfer-5090-windows.

Performance tab — Context & memory (--max-context, --kv-capacity [follow-ctx / auto / fixed], --prefill-chunk, --default-max-tokens, --default-thinking-budget), Scheduling (--max-concurrency 1–8, --max-pending-requests, --pending-timeout-ms), KV cache & context cache (--kv-dtype bf16/int8/fp8/nvfp4/k8v4, --no-prefix-reuse, --no-cuda-graph, device/host state slots, --host-kv-mib, max private continuations / shared prefixes / long anchors — auto-disabled with no-prefix-reuse), Speculative decoding (--spec off/MTP/DFlash/DFlash2, --draft-tokens range-aware: 1–5 for MTP, 1–15 otherwise, --lm-head-draft; per-artifact capability check computed from the catalog), reasoning-effort dropdown.

Advanced tab — Vision & media (--vision, residency frozen at startup, --media-cache-mib, --media-live-mib, --media-preprocess-threads, --max-request-mib), Sampling defaults (--no-thinking, --preserve-thinking, --greedy, --temperature/--top-p/--top-k/--min-p/--presence-penalty/ --frequency-penalty/--seed — process-level; request fields override), Logging, storage & misc (--log-level, --log-stats-interval-ms, --request-log-jsonl, --response-store-max-records/-mib, --context-cost-presets, --cors). Every control carries a tooltip quoting the engine docs; every omitted value means "engine executable default."

Profiles tab — save/load named profiles, 1:1 with flags. Also downloads (HF repo + file → hf download … --local-dir <modelsDir> with live output tail), engine log (live tail, 2 s poll, auto-stick scroll, severity coloring — the startup sequence: weights → host state/KV pinning → CUDA graphs → engine ready → listening — is readable in-app), and engine source updates (drives POST /api/engine/update: pull runs git pull --ff-only; build stops a running engine first, then runs the configured buildCommand, one job at a time, streaming stdout/stderr into a live log pane with a running/done/failed badge).

Models

• Downloaded artifacts cards — filename, model/weights badges, size, mtime, public id, speculation support + vision, HF source link, "loaded" highlight when the running engine uses it, spec support line; unregistered files flagged.
• Registered catalog table (the five artifact identities) — local/downloaded/ download states with live download progress tail, launch/download actions, per-row spec support (mtp 1–5; dflash 1–15 on 35B-A3B; dflash2 1–15 on 3.8-27B with companion weights).
• Models directory card — path + copy, with the validation caveat (partial artifacts fail readiness).

Settings

Four tabs — Engine / Safety & Permissions / Agent / About:

• Engine tab — engine paths (ninfer-serve, ninfer CLI, models dir, hf), default port, API key (Studio injects it on proxied requests), Coder lint/test commands, engine source pull/rebuild. Persisted to config.json via the control plane — see Configuration.
• Safety & Permissions tab — Safe Mode, Sandbox, and Commit Approval toggles (moved here from the Coder sidebar — see below), plus a default permissions template (tool allow/ask/deny tiers + denied path prefixes) that seeds a brand-new Coder workspace's permissions the first time it's created. Changing the template never touches an existing workspace's own permissions.
• Agent tab — reserved for future autonomy/tool-access configuration; currently a placeholder.
• About tab — architecture note, version.

All three screens (Chat, Coder, Settings) stay permanently mounted (see Architecture), so Safe Mode/Sandbox/Commit Approval are backed by a shared React context (CoderSafetyProvider) — toggling one in Settings is reflected in Coder immediately, with no reload.

Coder mode

An agentic coding harness over the same engine: plan/act loop with file read/write/ edit, grep/glob, shell exec with sessions and background jobs, git integration, Scout / Verify / Critic harness passes (opt-in, labeled, collapsible), worker/subagent delegation with depth and step budgets, conversation checkpoints with restore, human-approval (HITL) gates, sandboxed vs. live execution modes, and a per-workspace memory bank — with a live prefill/decode indicator so the model thinking is never invisible.

Safe Mode, Sandbox, and Commit Approval now live in Settings > Safety & Permissions (see above) rather than the Coder sidebar. The sidebar still holds per-workspace tool permissions (allow/ask/deny tiers + denied paths) and approved risky commands — a separate, per-workspace concept from the Settings-level default template.

Backed by desktop/control/src/coder/ (exec.rs, fs.rs, grep.rs, search.rs, memory.rs, web.rs, workspace.rs) via the /api/coder/* endpoints. See Security for the confinement model of exec vs. the file tools, and the SSRF protections on web_fetch.