Loading Accelerated Logic…
GA Release

The Multi-Agent WebGPU Engine & Interactions API are now generally available.

Part 24: Keyboard Shortcuts, Agent-Directive Tags & Glossary

Every keyboard shortcut and slash command, the XML-style tags the model uses to create skills/memories and call tools, and a plain-language glossary of the whole platform.

Category: Reference • Read Time: 25 min read • Updated: 2026-11-10

1. Keyboard & Mouse Shortcuts

ActionShortcutWhere
Send messageEnterChat composer (Shift+Enter = new line)
Quick open / file searchCtrl / ⌘ + PGitHub Workspace command palette
Cancel / closeEscModals, menus, dialogs
Pause / resume generationPause / Break keyActive agent run
Navigate suggestion list↑ / ↓ then EnterComposer autocomplete
Slash commands/fix · /comments · /optimizeCode workspaces

2. AI Directive Tags

When you ask the assistant to remember something or create a skill, it can emit a structured proposal tag in its reply. The client validates creation proposals and shows a Keep or Discard card above the composer; creation still requires the matching setting and your explicit Keep action. Deletions only run if the corresponding permission setting is enabled.

<create_skill> — proposes a reusable persona/instruction set; saved only after Keep.

<create_memory> — proposes a long-term fact; saved only after Keep and skipped if a duplicate exists.

<delete_skill> / <delete_memory> — remove entries (gated; content_sub matches containing memories).

<mcp_tool_call> — invokes a built-in or MCP tool as JSON; subject to the permission level and confirmation gate.

<image_gen prompt="…"> — requests image generation; a nudge appears if the model mentions an image but omits the tag.

3. Multi-Agent Architectures (Quick Map)

Single AgentOne equipped model handles the whole request.
Standard (Sequential)Orchestrator → workers → finalizer pipeline; requires the fixed orch/finalizer roles.
Planning (Planner–Critic)A planner drafts steps and a critic reviews before execution.
Self-DebateOpposing-view agents challenge an answer to surface flaws.
Document AnalyzerTuned for ingested-file reasoning.
Auto RouterPicks the architecture and model tier from the prompt (unit-tested heuristics).

4. Glossary

WebGPU / WASM: browser APIs for GPU-accelerated and CPU execution; they let local models and transcription run without a server.

WebLLM / Transformers.js: engines that run open-weights LLMs in the browser.

Gemini Nano: Google's small on-device model available in supported Chrome builds.

Puter.js: a free gateway that provides inference without an API key.

Ollama: a local server that runs models on your machine; requires CORS permission (OLLAMA_ORIGINS) to be called from the browser.

MCP (Model Context Protocol): a standard for connecting external tools/servers; external servers are sandboxed to permission-gated calls.

Skill: a reusable, named instruction/persona you (or the model) can save and attach.

Memory: a persisted long-term fact that shapes future responses; semantically matched to relevant prompts.

Thinking traces: the visible per-agent reasoning log streamed during a run.

Infinity Mode: loops a request through critic self-correction up to a configured attempt count.

Prompt queue: messages entered while a run is active are buffered and processed in order with backoff.

SecretStore: the internal helper that centralizes credential reads/writes and keeps keys out of cloud sync.

5. The / Command Reference

Command What it does Where
/<skill-name>Filters and toggles skills/personas — equips on EnterChat composer (any message)
/securityExample: equips the Security Auditor skillChat composer
/persona-ish filtersType part of a persona name (e.g. /senior) to jump to itChat composer
/fix · /comments · /optimizeCode-edit directives applied to the workspace selectionGitHub / code workspaces
EscapeCloses the slash menu without equippingComposer, modals, palettes

The composer / menu is the skills registry, not a general shell — there are no slash commands for "export", "settings", or "clear". Those live in the sidebar, header, and Settings search (Part 23).

6. AI Directive Tags — Cheat Sheet with Trigger Phrasing

These tags appear only when the model decides to emit them. You rarely type them yourself — you steer toward them with phrasing:

Tag Effect Phrase that tends to trigger it
<create_skill>Proposes a reusable skill; saved after Keep (requires skill-creation permission)"Package that review checklist as a reusable skill"
<create_memory>Proposes a long-term fact; saved after Keep (requires memory-creation permission)"Remember that I always want accessibility notes in UI work"
<delete_skill>Removes a skill (requires delete permission)"Delete the outdated Hemingway persona skill"
<delete_memory>Removes memories matching content_sub (requires delete permission)"Forget that I lived in Chicago"
<mcp_tool_call>Invokes a built-in or MCP tool; subject to permission levels"Search the web for…", "use the files server to read…"
<image_gen>Requests image generation; renders inline on completion"Generate an image of…", "draw a …"
<canvas_preview>Emits a live preview instead of a plain code block"…and preview it in canvas"
<python_tool>Background WASM Python execution, summarized into the trace"Use Python to verify…"

7. Extended Glossary — The Rest of the Platform

AL Router: the built-in free route (Fast / Reviewed / Multi-agent) that fans requests across free no-credential services with automatic fallback (Part 2).

Canvas Preview: a live, sandboxed render of a single-file component embedded in the chat stream (Part 19).

Web Sandbox: the full multi-file virtual project environment with viewports and console drawer (Part 9).

WebLLM / Transformers.js: two distinct in-browser local inference engines with different size/quality trade-offs (Part 4).

Gemini Nano: Chrome's on-device model; appears only when the browser reports support (Part 4).

GGUF import: streaming a local .gguf file into the Ollama instance on the same machine with hash-verified dedup (Part 4).

Folder Memory: standing instructions scoped to one folder's chats (Parts 1, 8).

Model Organizer: the single hub where models are discovered, equipped, and managed (Parts 2, 23).

Fallback order (saved): an ordered preference list per equipped model — configuration data only, not applied at request time (Parts 8, 23).

Prompt Queue: buffered follow-ups with backoff and a recovery drawer (Part 8).

Fork: branching the conversation into a new thread at any message (Parts 2, 8).

Staging vs pinning: staged = this turn only; pinned = every turn in this chat (Part 5).

Provider Test Arena: concurrent-per-provider benchmark of equipped models (Part 23).

AppData vault: the AES-GCM encrypted credential backup inside Google Drive's hidden AppData folder (Parts 17, 22).

Context-optimization annotation: the marker telling the model that attached text was excerpt-selected to fit a window (Part 5).

8. Using This Page as a Wall Chart

Print this page (browser print dialog) and the two tables — shortcuts (Section 1) and directive tags (Section 6) — fit cleanly on one sheet each for a side monitor. Suggested desk setup:

  • Left sheet: shortcuts + / menu semantics
  • Right sheet: directive tags + troubleshooting first response (ask another model? new chat? Part 20 ladder)
  • Laminate both; the glossary (Sections 4 and 7) lives in the docs, not on the wall

Anything you genuinely memorize here should be the shortcut table — it is the only section that pays back on every single session.